Wolfgang E.

Nagel
Dietmar H. Kröner
Michael M. Resch Editors

High Performance
Computing
in Science and
Engineering ’21

1 23
High Performance Computing in Science
and Engineering ’21
Wolfgang E. Nagel • Dietmar H. Kröner
Michael M. Resch
Editors

High Performance
Computing in Science
and Engineering ’21
Transactions of the High Performance
Computing Center, Stuttgart (HLRS) 2021

123
Editors
Wolfgang E. Nagel Dietmar H. Kröner
Zentrum für Informationsdienste Abteilung für Angewandte
und Hochleistungsrechnen (ZIH) Mathematik
Technische Universität Dresden Universität Freiburg
Dresden, Germany Freiburg, Germany

Michael M. Resch
Höchstleistungsrechenzentrum
Stuttgart (HLRS)
Universität Stuttgart
Stuttgart, Germany

ISBN 978-3-031-17936-5 ISBN 978-3-031-17937-2 (eBook)

https://doi.org/10.1007/978-3-031-17937-2

Mathematics Subject Classification (2020): 65Cxx, 65C99, 68U20, 90-08, 97M50

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2023
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

Cover illustration: The image shows the resulting polycrystalline microstructure of a martensitic phase
transformation predicted by a multiphase-field simulation performed with PACE3D. The regions shown
in shades of gray represent retained austenite, while the regions depicted in color are martensitic variants.
Details can be found in “High-performance multiphase-field simulations of solid-state phase transformations
using PACE3D” by E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl, and B. Nestler, Institute of
Applied Materials (IAM), Karlsruhe Institute of Technology (KIT), Straße am Forum 7, 76131 Karlsruhe,
Germany, on pages 167ff.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Contents

Physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Ligand-induced protein stabilization and enhanced molecular dynamics

sampling techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Timo Schäfer, Andreas C. Joerger, John Spencer, Friederike Schmid, and
Giovanni Settanni
Hadronic contributions to the anomalous magnetic moment of the muon
from Lattice QCD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
M. Cè, E.-H. Chao, A. Gérardin, J.R. Green, G. von Hippel, B. Hörz,
R.J. Hudspith, H.B. Meyer, K. Miura, D. Mohler, K. Ottnad, S. Paul, A. Risch,
T. San José, and H. Wittig
Quantum simulators, phase transitions, resonant tunneling, and
variances: A many-body perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
A.U.J. Lode, O.E. Alon, J. Arnold, A. Bhowmik, M. Büttner, L.S. Cederbaum,
B. Chatterjee, R. Chitra, S. Dutta, C. Georges, A. Hemmerich, H. Keßler,
J. Klinder, C. Lévêque, R. Lin, P. Molignini, F. Schäfer, J. Schmiedmayer, and
M. Žonda

Molecules, Interfaces and Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Molecular dynamics simulation of selective laser melting . . . . . . . . . . . . . . 65
Fabio Oelschläger, Dominic Klein, Sarah Müller, and Johannes Roth
Molecular dynamics investigations on the influence of solutes on the
tensile behavior of Polyamide6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Wolfgang Verestek, Johannes Kaiser, Christian Bonten, and Siegfried
Schmauder

v
vi Contents

Dynamical properties of the Si(553)-Au nanowire system . . . . . . . . . . . . . . 97

Mike N. Pionteck, Felix Bernhardt, Johannes Bilk, Christof Dues, Kevin
Eberheim, Christa Fink, Kris Holtgrewe, Niklas Jöckel, Brendan Muscutt,
Florian A. Pfeiffer, Ferdinand Ziese, and Simone Sanna

Reactivity of organic molecules on semiconductor surfaces revealed by

density functional theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Fabian Pieck, Jan-Niclas Luy, Florian Kreuter, Badal Mondal, and Ralf
Tonner-Zech
Electro-catalysis for H2 O oxidation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Travis Jones

Materials Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

Production of defects in two-dimensional materials under ion and electron
irradiation: insights from advanced first-principles calculations . . . . . . . . 151
S. Kretschmer, T. Joseph, S. Ghaderzadeh, Y. Wei, M. Ghorbani-Asl, and
A.V. Krasheninnikov

High-performance multiphase-field simulations of solid-state phase

transformations using Pace3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl, and B. Nestler
Bridging scales with volume coupling — Scalable simulations of muscle
contraction and electromyography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Benjamin Maier, David Schneider, Miriam Schulte, and Benjamin Uekermann

Computational Fluid Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

Analysis of the hot gas ingress into the wheel space of an axial turbine stage207
Jannik Borgelt, Thomas Hösgen, Matthias Meinke, and Wolfgang Schröder
Direct numerical simulation of bypass transition under free-stream
turbulence for compressible flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Duncan Ohno, Björn Selent, Markus J. Kloker, and Ulrich Rist
Direct numerical simulation of a disintegrating liquid rivulet at a trailing
edge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Adrian Schlottke, Matthias Ibach, Jonas Steigerwald, and Bernhard Weigand

Numerical Investigation of the Flow and Heat Transfer in Convergent

Swirl Chambers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Florian Seibold and Bernhard Weigand
Contents vii

On the validity of the linear Boussinesq hypothesis for selected internal

cooling features of gas turbine blades . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Philipp Wellinger and Bernhard Weigand

Development of turbulent inflow methods for the high order HPC

framework FLEXI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
Daniel Kempf, Min Gao, Andrea Beck, Marcel Blind, Patrick Kopper, Thomas
Kuhn, Marius Kurz, Anna Schwarz, and Claus-Dieter Munz
A narrow band-based dynamic load balancing scheme for the level-set
ghost-fluid method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Daniel Appel, Steven Jöns, Jens Keim, Christoph Müller, Jonas Zeifang, and
Claus-Dieter Munz
Numerical simulation of flake orientation during droplet impact on
substrates in spray painting processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Qiaoyan Ye and Martin Dreher, and Bo Shen

A low-pass filter for linear forcing in the open-source code OpenFOAM –

Implementation and numerical performance . . . . . . . . . . . . . . . . . . . . . . . . 339
Jordan A. Denev, Thorsten Zirwes, Feichi Zhang, and Henning Bockhorn
Numerical simulation of vortex induced pressure fluctuations in the
runner of a Francis turbine at deep part load conditions . . . . . . . . . . . . . . 353
Jonas Wack, Marco Zorn, and Stefan Riedelbauch
Validation of ACD and ACL propeller simulation using blade element
method based on airfoil characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
Michael Schollenberger, Mário Firnhaber Beckers, and Thorsten Lutz

Transport and Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381

The HPC project GLOMIR+ (GLObal MUSICA IASI Retrievals - plus) . 383
Matthias Schneider, Benjamin Ertl, Christopher Diekmann, and Farahnaz
Khosrawi
Global long-term MIPAS data processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
Michael Kiefer, Bernd Funke, Maya García-Comas, Udo Grabowski, Andrea
Linden, Axel Murk, and Gerald E. Nedoluha
WRF simulations to investigate processes across scales (WRFSCALE) . . . 411
Hans-Stefan Bauer, Thomas Schwitalla, Oliver Branch, and Rohith Thundathil
viii Contents

Computer Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423

Dynamic molecular dynamics ensembles for multiscale simulation coupling425
Philipp Neumann, Niklas Wittmer, Vahid Jafari, Steffen Seckler, and Matthias
Heinen
Scalable discrete algorithms for big data applications . . . . . . . . . . . . . . . . . 439
Demian Hespe, Lukas Hübner, Lorenz Hübschle-Schneider, Peter Sanders, and
Dominik Schreiber

Miscellaneous Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451

Large scale simulations of partial differential equation models on the
Hawk supercomputer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
Ruben Buijse, Martin Parnet, and Arne Nägel
Scaling in the context of molecular dynamics simulations with ms2 and ls1
mardyn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Simon Homes, Robin Fingerhut, Gabriela Guevara-Carrion, Matthias Heinen,
and Jadran Vrabec
Scalable multigrid algorithm for fluid dynamic shape optimization . . . . . . 481
Jose Pinzon, Martin Siebenborn, and Andreas Vogel
Numerical calculation of the lean-blow-out in a multi-jet burner . . . . . . . . 497
Alexander Schwagerus, Peter Habisreuther, and Nikolaos Zarzalis
Data-driven multiscale modeling of self-assembly and hierarchical
structural formation in biological macro-molecular systems . . . . . . . . . . . 513
Philipp Nicolas Depta, Maksym Dosta, and Stefan Heinrich
 Physics

In this part, three physics projects are presented, which achieved important scientific
results in 2020/21 by using Hawk/Hazel Hen at the HLRS and ForHLR II of the
Steinbuch Center.
Fascinating new results are being presented in the following pages on soft mat-
ter/biochemical systems (ligand-induced protein stabilization) and on quantum sys-
tems (anomalous magnetic moment of the muon, ultracold-boson quantum simulators,
phase transitions, resonant tunneling, and variances).
Studies of the soft matter/biochemical systems have focused on ligand-induced
protein stabilization.
T. Schäfer, A.C. Joerger, J. Spencer, F. Schmid, and G. Settanni from Mainz (T.S.,
F.S., G.S.), Frankfurt (A.C.J.), and Sussex (J.S.) present interesting new results on
ligand-induced protein stabilization in their project Flexadfg. The authors show how
Molecular Dynamics simulations of several cancer mutants of the DNA-binding
domain of the tumor suppressor protein p53 allowed to establish the destabilizing effect
of the mutations as well as the stabilizing effects of bound ligands. In addition, the
authors report on the development of a new reweighting technique for metadynamics
simulations that speeds up convergence and may provide an advantage in the case of
simulations of large systems.
Studies of the quantum systems have focused on the anomalous magnetic moment
of the muon, and on ultracold-boson quantum simulators, phase transitions, resonant
tunneling, and variances.
M. Cè, E. Chao, A. Gérardin, J.R. Green, G. von Hippel, B. Hörz, R.J. Hudspith,
H.B. Meyer, K. Miura, D. Mohler, K. Ottnad, S. Paul, A. Risch, T. San José, and
H. Wittig from Mainz (E.C., G.v.H., R.J.H., H.B.M., K.M., K.O., S.P., T.S.J., H.W.),
Darmstadt (K.M., D.M., T.S.J.), Zeuthen (A.R.), Geneva (M.C., J.R.G.), Marseille
(A.G.), and Berkeley (B.H.) present interesting results obtained by their lattice QCD
Monte Carlo simulations on Hawk/Hazel Hen in their project GCS-HQCD on leading
hadronic contributions to the anomalous magnetic moment of the muon, on the
energy dependence of the electromagnetic coupling, on the electroweak mixing angle,
and on the hadronic vacuum polarisation and light-by-light scattering contributions.

1
2 Physics

The authors focus will turn to increasing the overall precision of their determination
of the hadronic vacuum polarization contribution to the muon anomalous magnetic
moment to the sub-percent level.
A.U.J. Lode, O.E. Alon, J. Arnold, A. Bhowmik, M. Büttner, L.S. Cederbaum,
B. Chatterjee, R. Chitra, S. Dutta, C. Georges, A. Hemmerich, H. Keßler, J. Klinder,
C. Lévêque, R. Lin, P. Molignini, F. Schäfer, J. Schmiedmayer, and M. Žonda from
Freiburg (A.U.J.L., M.B. ), Haifa (O.E.A., A.B., S.D.), Basel (J.A., F.S.), Heidelberg
(L.S.C.), Kanpur (B.C.), Zürich (R.C., R.L.), Hamburg (C.G., A.H., H.K., J.K.), Wien
(C.L., J.S.), Oxford (P.M.), and Prague (M.Z.) present interesting results obtained
in their project MCTDHB with their multiconfigurational time-dependent Hartree
method for indistinguishable particles (MCTDH-X) on Hazel Hen and Hawk. In
the past the authors have implemented their method to solve the many-particle
Schrödinger equation for time-dependent and time-independent systems in various
software packages. The authors present interesting new results of their investigations
on ultracold boson quantum simulators for crystallization and superconductors in
a magnetic field, on phase transitions of ultracold bosons interacting with a cavity,
and of charged fermions in lattices described by the Falicov–Kimball model. In
addition, the authors report on new results on the many-body dynamics of tunneling
and variances, in two- and three-dimensional ultracold-boson systems.

Fachbereich Physik, Peter Nielaba

Universität Konstanz,
78457 Konstanz,
Germany,
e-mail: [email protected]
Ligand-induced protein stabilization and
enhanced molecular dynamics sampling
techniques

Timo Schäfer, Andreas C. Joerger, John Spencer, Friederike Schmid and Giovanni
Settanni

Abstract Molecular dynamics (MD) simulations provide an increasingly important

instrument to study protein-materials interaction phenomena, thanks to both the
constant improvement of the available computational resources and the refinement of
the modeling methods. Here, we summarize the results obtained along two different
research directions within our project. First, we show how MD simulations of several
cancer mutants of the DNA-binding domain of the tumor suppressor protein p53
allowed to establish the destabilizing effect of the mutations as well as the stabilizing
effects of bound ligands. Second, we report on the development of a new reweighting
technique for metadynamics simulations that speeds up convergence and may provide
an advantage in the case of simulation of large systems.

1 Introduction

Molecular dynamics (MD) simulations provide a way to observe the motions of

molecular objects at the atomic scale. In classical MD simulations, each system is
composed of a set of particles and is represented by a Hamiltonian energy function
of the coordinates and momenta of the particles. By numerical integration of the
Hamiltonian equations of motion, MD allows to follow the trajectories of all the
particles as a function of time and, in this way, to extrapolate its static and dynamic
properties. A crucial step to guarantee the accuracy of the observations is the

Timo Schäfer, Friederike Schmid and Giovanni Settanni

Department of Physics, Johannes Gutenberg University, Mainz, Germany,
e-mail: [email protected]
Andreas C. Joerger
Institute of Pharmaceutical Chemistry, Johann Wolfgang Goethe University, Frankfurt am Main,
Germany
John Spencer
Department of Chemistry, School of Life Sciences, University of Sussex, United Kingdom

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 3

W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_1
4 T. Schäfer, A.C. Joerger, J. Spencer, F. Schmid and G. Settanni

availability of highly optimized force fields, i.e., the classical Hamiltonian function,
which approximates the underlying quantum chemical nature of the system. During the
course of the last 40 years, MD force fields for the simulations of biomolecules have
been dramatically improved in terms of accuracy to the point that it is now possible
to simulate phenomena like protein folding, at least for some small proteins [1], as
well as protein-ligand and protein-materials interactions [2–11]. In this report, we
will show how we have been able to use classical MD simulations to characterize the
stability of several cancer mutants of p53, a protein that is found mutated in about
50% of the cancer cases diagnosed every year [12]. In that study simulations were
used also to asses the effect of small ligands which, by binding to a mutation-induced
pocket on the protein surface, are capable of stabilizing the protein, thus representing
possible lead compounds for the development of cancer therapeutics. In addition to
protein folding, many biological phenomena, however, occur on time scales that are
still inaccessible to standard classical MD. For this reason, a wide range of enhanced
sampling techniques have been proposed. Metadynamics [13–15] is one of the most
popular techniques. In metadynamics, a time-dependent energy term is added to the
Hamiltonian, to drive the system away from regions of conformational space already
sampled. We started to use this method in the past to characterize the conformational
properties of a large protein complex, fibrinogen [9]. We soon discovered that in
this case, the available methods to unbias the metadynamics sampling and obtain
an equilibrium distribution of the most significant observables of the system were
inadequate. We have then developed a new method, which is more accurate than
those previously available in the limit of short trajectories [16]. In what follows, we
also review these findings.

2 Methods

GROMACS[17] was used to perform the MD simulations. GROMACS exploits intra-

node OpenMP parallelization and inter-node MPI parallelization and was optimally
compiled to run efficiently both on Hazelhen and on Hawk HPC infrastructures at
HLRS. Unless stated differently, in what follows, the simulations were performed
with the CHARMM36m force field [18], while the ligands were modeled using the
Charmm Generalized force field (CGenFF)[19]. The time step for the simulations
was set to 2 fs. The LINCS [20] algorithm was used to constrain the length of bonds
involving hydrogen atoms. A cut-off of 1.2 nm with a switch function starting at
1.0 nm was used for direct non-bonded interactions. A cell-list like algorithm [21] was
used. Periodic boundary conditions were adopted along all directions and a smooth
particle-mesh Ewald (sPME) approach [22] was used for long-range electrostatics.
The water was modeled explicitly using a modified TIP3P model[23]. Pressure and
temperature were regulated at 1 atm and 300 K, using the Parrinello–Rahman [24] and
Nose–Hoover [25, 26] algorithms, respectively. Few water molecules were replaced
by sodium and chlorine ions during simulation setup to neutralize the charge of the
simulation box and to achieve the physiological ion concentration (0.15 M).
Ligand-induced protein stabilization 5

In the case of the p53 DNA-binding domain, the simulations were based on the
structure of a stabilized pseudo-wild-type (pdb id 1UOL) [27] and mutant structures
determined by X-ray crystallography or modeled in ref. [28]. The systems were
minimized for a max 50’000 steps, then equilibrated in the NVT ensemble for 1 ns
with positional restraints on the heavy atoms of the protein. Then, they were further
equilibrated for 1 ns in the NPT ensemble with no restraints. Four production runs
were started for each mutant. Each run was 200 ns long. Further methodological
details are provided in the original publication ref. [28]. System sizes ranged from
37000 to 43000 atoms, which are sufficient to ensure good scaling on Hazelhen
and Hawk running on 100 and 64 nodes, respectively. Simulations were set up
as job chains with each job not exceeding 3 hours length, writing a single restart
file at the end. The trajectories were analyzed using the program VMD [29] and
WORDOM [30].

3 Simulations of p53 cancer mutants and stabilization by ligand

binding

The tumor suppressor protein p53 is involved in several processes protecting the
human genome, including activation of DNA repair mechanisms or induction of
apoptosis (cell death) in case of extensive DNA damage. Mutations of this protein
can often result in cancer and, indeed, mutations in this protein occur in half of
the diagnosed cancers [12]. Among the most frequently found cancer mutations
are those of residue 220 in the DNA-binding domain of the protein, with the most
abundant being Y220C, found in about 100’000 new cancer cases each year. The
wild-type protein (WT) (i.e., the one without mutations) is only marginally stable,
and many cancer mutations induce a loss of stability, which reduces the folding
transition temperature of the protein, leading to unfolding and, consequently, to a loss
of function at body temperature. In some cases, such as Y220C, the mutation creates a
crevice on the protein surface. This offers a possible strategy to reactivate the mutant
protein: a drug that binds to the mutation-induced pocket may actually stabilize the
protein and thereby rescue its tumor suppressor function. This strategy has already
been successfully used to rescue the Y220C mutant [2, 31]. In collaboration with our
experimental partners, we investigated other frequent cancer mutants with a mutation
of Y220.
In ref. [28], we analyzed the cancer associated mutants Y220H, Y220N, Y220S,
and Y220C using experimental biophysical techniques including, among others, X-ray
crystallography as well as MD simulations. The latter have been used to estimate
the effects of mutations on protein stability by monitoring the root-mean-square
fluctuations (RMSF) [32], that is the average amplitude of the fluctuations of the
atomic positions of the protein around the average structure. We used the same
approach in the context of the p53 DNA-binding domain mutants, and we verified that
the RMSF measured on the simulations correlated with experimentally determined
differences in melting temperature of the mutant proteins (Tab. 1). Our simulations
6 T. Schäfer, A.C. Joerger, J. Spencer, F. Schmid and G. Settanni

also provided additional information complementing the structural data obtained by

X-ray crystallography, for example, by sampling the dynamics of different regions
of the protein. We monitored several atomic distances characterizing the mutation-
induced crevice on the surface of the protein, which can be split into a central crevice
and different subsites (Fig. 1). These distances did not show significant fluctuations
in the WT, whereas in the cancer mutants, they fluctuated between two states which
can be associated with open and collapsed conformations of the mutation-induced
crevice (Fig. 1c).
Several carbazole-based compounds were shown to bind to the mutation-induced
crevice in Y220C and stabilize the protein structure, thus increasing the folding
transition temperature [2, 31]. Given the similarity between the crevices generated by
the various mutations at site 220, in ref. [28] the effects of some of these carbazole
compounds on the Y220S and Y220C mutants were analyzed. It was shown that
several compounds that bind Y220C in the mutation-induced crevice can also bind
to the equivalent pocket in Y220S, to a lesser extent also in the Y220N pocket.
The binding of the compounds resulted in a considerable increase in the melting
temperature of the mutants. MD simulations of the mutants in the presence of the
carbazole compounds in the crevice revealed that the ligands dramatically reduced
the RMSF of the protein, which correlated with the binding constant measured
experimentally and the ligand-induced increase of the folding transition temperature
(Tab. 1). In particular, the simulations showed that the collapsed state of the crevice
is completely absent in the presence of the ligands (Fig. 2).
Summarizing this part of the report, the simulations of p53 DNA-binding domain
mutants, with and without ligands, provided important insights into the molecular
basis of the experimentally observed stabilizing effect upon binding. This information
will aid the development of more potent small-molecule stabilizers and molecules
targeting more than one mutant.

Codon 220 mutation Average RMSF (Å) Tm (◦ C)

Native structure
WT 1.46 51.5
Y220H 1.56 45.1
Y220C 1.61 43.8
Y220N 1.64 39.9
Y220S 1.67 39.4
Ligand complexes
Y220C-PK9323 1.35
Y220S-PK9301 1.42
Y220S-PK9323 1.58

Table 1: Root-mean-square fluctuations (RMSF) of the simulated p53 DNA-binding

domain constructs averaged over residues 97-289 and the experimentally determined
folding transition temperature. Adapted from ref. [28], Bauer et al. ©2020 licensed
under CC-BY.
Ligand-induced protein stabilization 7

Fig. 1: MD simulation of p53 cancer mutants. (A) Representative C 𝛼 -trace structures

of the open and closed states from the simulations of the p53 mutant Y220S
superimposed onto the crystal structure (green). Highlighted residues are represented
as sticks. Distances defining the size of the central cavity (d1) and the subsite 2 (d2)
are represented with magenta dashed lines. (B) Same as (A) but for the Y220H mutant.
The closed state is determined by the H220 side chain swinging out of the pocket. (C)
Distribution of the d1 and d2 distances in the simulations of the p53 variants under
consideration. The green line represent the distance in the crystal structure. Only WT
shows an unimodal distribution. All mutants show the presence of a collapsed state of
the crevice beside the open state. Adapted from ref. [28], Bauer et al. ©2020 licensed
under CC-BY.
8 T. Schäfer, A.C. Joerger, J. Spencer, F. Schmid and G. Settanni

1.4
1.2 WT
1
0.8
0.6
0.4
1.4
1.2 Y220C
1
0.8
0.6
0.4
1.4
1.2 Y220S
1
0.8
0.6
0.4
1.4
1.2 Y220N
1
0.8
0.6
0.4
1.4
Distance (nm)

1.2 Y220H
1
0.8
0.6
0.4
1.4
1.2 Y220C-PK9323
1
0.8
0.6
0.4
1.4
1.2 Y220S-PK9323
1
0.8
0.6
0.4
1.4
1.2 Y220S-PK9301
1
0.8
0.6
0.4
0 50 100 150 200
0 50 100 150 200
0 50 100 150 200
0 50 100 150 200 2k 4k 6k 8k 10k

Time (ns) Counts

Fig. 2: Time series and resulting distributions of the d1 distance (see fig. 1) in all the
simulated constructs. The presence of the ligands PK9323 and PK9301 in the crevice
dramatically reduces structural fluctuations and prevents the population of collapsed
states of the mutation-induced pocket. Adapted from ref. [28], Bauer et al. ©2020
licensed under CC-BY.

3.1 Data reweighting in metadynamics simulations

In this section, we review the results presented in ref. [16]. Metadynamics introduces
a time-dependent bias potential in the classical energy of the system. The bias is
dependent on selected collective variables 𝒔(𝒓) of the system’s coordinates, and it is
built as a sum of Gaussian functions that are deposited at regular time intervals on
the points reached by the trajectory:
Ligand-induced protein stabilization 9
𝑡
∑︁ 𝑡
∑︁
(𝑠𝑖 (𝒓)−𝑠𝑖 (𝒓 (𝑡 ′ ))) 2 /2𝜎𝑖2
Í𝑑
𝑉 (𝒔(𝒓), 𝑡) = 𝑊e− 𝑖 = 𝑔𝑡 ′ (𝒓, 𝒔(𝑡 ′)) (1)
𝑡 ′ =Δ𝑡 ,2Δ𝑡 ,... 𝑡 ′ =Δ𝑡 ,2Δ𝑡 ,...

Í𝑑 ′
where the Gaussian hills are 𝑔𝑡 ′ (𝒔(𝒓)) = 𝑊e− 𝑖 (𝑠𝑖 (𝒓)−𝑠𝑖 (𝒓 (𝑡 ))) /2𝜎𝑖 . This results in
2 2

pushing the system away from those values of the collective variables that have been
already sampled.
In order to obtain equilibrium properties from the trajectories sampled using
metadynamics, a reweighting procedure is necessary that takes into account the
time-dependent influence of the bias. In other words, a weight 𝑤(𝒔, 𝑡) has to be
assigned to each sampled conformation of the system, which is dependent on the
time evolution of the bias. Once we have that, we can measure equilibrium properties
from the biased simulations:

⟨𝐴⟩0 = ⟨𝐴𝑤⟩b /⟨𝑤⟩b (2)

where 𝐴 is any observables defined on the system, the subscript 0 indicates the
unbiased (i.e. equilibrium) average and the subscript 𝑏 indicates the average done
over the biased simulations.
Several reweighting techniques have been proposed for metadynamics simula-
tion [33–37], which come with some limitations, for example, some of them are
specifically suited for well-tempered metadynamics, which is a modified version
of the original algorithm where the height of the Gaussians decreases during the
simulation. The simulations we ran on Hazelhen to determine the conformational
states of the fibrinogen complex [9] showed however that the popular reweighting
scheme from Tiwary and Parrinello [37], when used to build the free energy landscape
as a function of the collective variables, did not match exactly the negative bias
potential, which also represents an estimate of the free energy of the system. Tiwary’s
method is based on the assumption that in between two Gaussian depositions, the
system samples a biased energy function 𝑈 + 𝑉 where 𝑈 is the unbiased energy
function of the system and 𝑉 is the metadynamics bias. Thus the conformations of
the system follow a canonical distribution with biased energy.

e−𝛽 (𝑈 (𝒓)+𝑉 (𝒔 (𝒓),𝑡))

𝑝 𝑏 (𝒓, 𝑡) = ∫ (3)
d𝒓e−𝛽 (𝑈 (𝒓)+𝑉 (𝒔 (𝒓),𝑡))

For the unbiased system, however, we would have the following distribution:

e−𝛽𝑈 (𝒓)
𝑝 0 (𝒓) = ∫ (4)
d𝒓e−𝛽𝑈 (𝒓)

We can then obtain the equilibrium distribution from the biased distribution by means
of the following reweighting factor [34, 37]:
10 T. Schäfer, A.C. Joerger, J. Spencer, F. Schmid and G. Settanni

𝑝 0 (𝒓)
𝑤(𝒓, 𝑡) = (5)
𝑝 𝑏 (𝒓, 𝑡)
∫
d𝒓e−𝛽 (𝑈 (𝒓 ,𝑡)+𝑉 (𝒔 (𝒓),𝑡))
= e𝛽𝑉 (𝒔 (𝒓),𝑡)) ∫ (6)
d𝒓e−𝛽𝑈 (𝒓 ,𝑡)

which can be further rewritten as [34, 37]:

∫ ∫
𝛽𝑉 (𝒔 (𝒓),𝑡))
d𝒓 d𝒔 𝛿(𝒔 − 𝒔(𝒓))e−𝛽 (𝑈 (𝒓 ,𝑡)+𝑉 (𝒔 (𝒓),𝑡))
𝑤(𝒓, 𝑡) = e ∫ ∫ (7)
d𝒓 d𝒔 𝛿(𝒔 − 𝒔(𝒓))e−𝛽𝑈 (𝒓 ,𝑡)
∫
𝛽𝑉 (𝒔 (𝒓),𝑡))
d𝒔 𝑝 0 (𝒔)e−𝛽𝑉 (𝒔,𝑡)
=e ∫ (8)
d𝒔 𝑝 0 (𝒔)
∫
d𝒔 e −𝛽𝐹 (𝒔) e−𝛽𝑉 (𝒔,𝑡)
= e𝛽𝑉 (𝒔 (𝒓),𝑡)) ∫ (9)
d𝒔 e−𝛽𝐹 (𝒔)

where
e−𝛽𝐹 (𝒔)
∫
𝑝 0 (𝒔) = d𝒓𝛿(𝒔 − 𝒔(𝒓)) 𝑝 0 (𝒓) = ∫ (10)
d𝒔 e−𝛽𝐹 (𝒔)
is the unbiased distribution projected on the low-dimensional CV space Ω and 𝐹 (𝒔)
is the unbiased free energy.
Eq. 9, although formally exact, contains the term 𝐹 (𝒔), which is not known a priori
(it is actually often the main aim of the simulation). In metadynamics the negative bias
potential asymptotically approximates the free energy of the system [14, 15, 36, 38]:
𝛾
𝐹 (𝒔) ≈ − 𝑉 (𝑠, 𝑡) + 𝑐(𝑡) (11)
𝛾−1
where 𝛾 is the so-called bias factor of well-tempered metadynamics, which goes
to infinity in the case of standard metadynamics, and 𝑐(𝑡) is a time-dependent
offset of the free energy profile. In the Tiwary and Parrinello method [37], 𝐹 (𝒔) is
approximated using eq. 11, leading to the following expression for the weight (for
standard metadynamics):
∫
d𝒔 1
𝑤 𝑡 𝑤 (𝒓, 𝑡) = e𝛽𝑉 (𝒔 (𝒓) ,𝑡) ∫ = e𝛽𝑉 (𝒔 (𝒓),𝑡) 𝛽𝑉 (𝒔,𝑡) (12)
d𝒔e 𝛽𝑉 (𝒔,𝑡) ⟨e ⟩𝒔

The approximation eq. 11, however, is only valid asymptotically, thus at short time
scales the reweighting scheme may not provide accurate results.
Alternatively, we suggested [16] to approximate 𝐹 (𝒔) with the value of the negative
potential at the end of the simulation, which is supposedly more accurate:
∫
d𝒔 e𝛽 (𝑉 (𝒔,𝑡 𝑓 )−𝑉 (𝒔,𝑡)
𝑤(𝒓, 𝑡) = e𝛽𝑉 (𝒔 (𝒓),𝑡)) ∫ (13)
d𝒔 e𝛽𝑉 (𝒔,𝑡 𝑓 )
Ligand-induced protein stabilization 11

where 𝑡 𝑓 corresponds to the time at the end of the simulation. Eq. 13 would converge
to eq. 12 at large 𝑡, but it will behave differently at small 𝑡. At short simulation times,
a simple exponential reweighting would be more accurate than the Tiwary’s eq. 12. A
simple exponential reweighting, however, in the standard metadynamics setting where
the bias increases with time would result in an underweighting of the initial part of
the simulation. In ref. [16] we propose a simple way to correct it by subtracting the
average value of the bias at every time step. This results in the balanced exponential
reweighting:
′ (𝒔 (𝒓),𝑡) 1
𝑤 𝑏𝑒𝑥 (𝒓, 𝑡) ∝ e𝛽𝑉 = e𝛽 ( 𝑉 (𝒔 (𝒓),𝑡)−𝑉𝑎 (𝑡) ) = e𝛽𝑉 (𝒔 (𝒓),𝑡) . (14)
e𝛽 ⟨𝑉 (𝒔,𝑡) ⟩𝒔
The scheme proposed above differs from the previously proposed scheme of eq. 12
by the normalization factor of the exponential weight: in the case of eq. 12 the average
value over Ω of the exponential of the bias potential is used, whereas in the new
scheme eq. 14 we propose to use the exponential of the average bias. The average
of the exponential (eq. 12) is very sensitive to small changes in the upper tail of the
distribution of the bias potential and is, therefore, less robust in the initial part of the
trajectory where the global free energy minimum of the system may not have been
reached, yet.
The newly proposed scheme can be implemented without changes to the output of
the popular metadynamics software PLUMED [39]. To demonstrate its advantages,
we have tested it in several different scenarios and compared it with existing schemes.
The standard mean of comparison that we have adopted consists in recovering the
free energy landscape of a given system as a function of the collective variables
by reweighting the conformations of the system sampled along the metadynamics
trajectory. We did that for a series of systems for which an accurate free energy
landscape is accessible and can be used as reference.
The first system we studied is a particle in a uni-dimensional double-well potential
of the form 𝑈 (𝑥) = (𝑥 2 − 1) 2 . Simulations are performed at a temperature such that
𝑘 𝑏 𝑇 is 1/10 of the barrier separating the two energy wells. The system is discretized
along the 𝑥 direction into equally sized bins, and pseudo standard metadynamics
simulation are performed by moving the particle to the left or right bin using a
Metropolis criterion for accepting the move. The energy function for the Metropolis
criterion is 𝑈 (𝑥) + 𝑉 (𝑥, 𝑡), 𝑉 (𝑥, 𝑡), where:
𝑡
∑︁ 𝑣 ′
e ( −( 𝑥−𝑥 (𝑡 )) )
2 /2𝜎
𝑉 (𝑥, 𝑡) = √ (15)
𝑡 ′ =Δ𝑡 ,2Δ𝑡 ,... 2𝜋𝜎

In the above
√ expression, the metadynamics Gaussian hills have volume 𝑣 (that is
height 𝑣/ 2𝜋𝜎) and width 𝜎. The deposition period is Δ𝑡. Several simulations were
run using different metadynamics parameters but keeping the length and step size
(bin width) fixed. We then estimated the free energy landscape of the system using
eq. 12 and 14:
12 T. Schäfer, A.C. Joerger, J. Spencer, F. Schmid and G. Settanni
!
𝛿(𝑥 − 𝑥(𝑡 ′))𝑤(𝑥(𝑡 ′), 𝑡 ′)
Í𝑡
𝑡 ′ =0
𝐹𝑒𝑠𝑡 (𝑥, 𝑡) = −𝛽−1 log Í𝑡 ′ ′
(16)
𝑡 ′ =0 𝑤(𝑥(𝑡 ), 𝑡 )

where the 𝛿 functions are the characteristic functions of the discrete bins along the 𝑥
axis.
In Fig. 3a we report the estimated and reference free energy landscape of the
system as a function of the simulated time. We repeated the simulations 72 times
with different initial conditions and measured the error of the estimated free energy
with respect to the reference in each of the simulations (the error is computed as the
root-mean-square deviation of the two free energy profiles limited to the interval
(-2,2) after subtracting the average). The data (Fig. 3b) show that the estimate obtained
with the newly proposed eq. 14 converges faster than the other tested methods to the
reference landscape. Also other estimates of the quality of the free energy landscape,
like the estimate of the free energy difference between the minima (Fig. 3c) and
the estimate of the height of the barrier (Fig. 3d) reveal a similar picture. Another
advantage of the balanced exponential reweighting is the low run-to-run variability
reported by the error bar in Fig.3b-d. A detailed look at the weights of the sampled
conformations (Fig. 3) shows that while the balanced exponential weights are generally
constant along the simulation, they are smaller than average in the very initial part of
the simulation when the system has not yet explored both minima. On the other hand,
Tiwary’s method produces weights that are larger than average in the initial part of
the simulation. This overestimate reduces the quality of the free energy profile for the
early part of the run.
Although the uni-dimensional system offers already a good overview of the
advantages of the newly proposed reweighting scheme, a test with a more realistic
system is necessary to assess the performance in a normal-use scenario. For that,
in the same work [16], we used an alanine dipeptide, which represents a standard
benchmark of enhanced sampling techniques. The alanine dipeptide can be considered
as the smallest protein-like unit as its structure can be characterized by the two protein
backbone dihedral angles 𝜙 and 𝜓. We performed the metadynamics simulations
using GROMACS [17] with the PLUMED [39] plugin. The system was simulated
for short trajectories (8 ns) in vacuum using the standard force field AMBER03 [40]
and standard values for time step and non-bonded interaction cutoff. The backbone
dihedral angles 𝜙 and 𝜓 were biased during the metadynamics simulations (for all
the details of the simulations and the set of metadynamics parameters used please
refer to our original work [16]). Adopting a strategy similar to the uni-dimensional
case, we estimated the free energy landscape of the system as a function of 𝜙 and 𝜓
using several different reweighting schemes including the newly proposed balanced
exponential. We did that at several different time points along the simulation. We
then compared the free energy estimates with the reference obtained by running an
extremely long well-tempered metadynamics simulations (Fig. 4).
Ligand-induced protein stabilization 13

8
a b

0.04 0.08 0.12

RMSDFES /10 kbT

0.12
RMSDFES steps0.5
0.08
0.04
0 5 10 15 20
steps /106
2·107

0
c

0.04 0.08 0.12

|ΔF| /10 kbT
F(x)/10kbT
4

0
d

0.04 0.08 0.12

4·106 |ΔF‡ ⎯ 1| /10 kbT

3·106
2

e
2·106
0.04 0.08 0.12
RMSDV /10 kbT

106
0

-1.5 -1 -0.5 0 0.5 1 1.5 4 8 12 16 20

x steps /106
2

f g 0
log(w/<w>)
1

-10
0
x

-20
-2 -1

-30
0 4 8 12 16 0 4 8 12 16
step / 106 step / 106

Fig. 3: (a) FES obtained along a single run at different trajectory lengths using
balanced exponential reweighting, Tiwary reweighting and negative bias potential
(red, blue and green points, respectively). The reference
√︁ FES is plotted in purple. (b-e)
Time series of (b) the RMSD (in inset is the RMSD· (𝑡)), (c) estimated free energy
difference between the two minima (absolute value), (d) estimated error on height
of free energy barrier, and (e) the RMSD between reweighted FES and negative
bias. Same color scheme as in part a. The solid lines represent the average values of
the quantities across the 72 independent runs. Shaded bands indicate the standard
deviations. (f-g) Time series of the position of the particle along one run where
balanced exponential (f) and Tiwary (g) weights are reported according to a color
scale. Adapted with permission from ref. [16], ©2020 American Chemical Society.
14 T. Schäfer, A.C. Joerger, J. Spencer, F. Schmid and G. Settanni

a b

π
0 π/2
ψ /rad
-π/2
-π

c d
π
0 π/2
ψ /rad
-π/2
-π

e f
π
0 π/2
ψ /rad
-π/2
-π

-π -π/2 0 π/2 π -π -π/2 0 π/2 π

φ /rad φ /rad

Fig. 4: (a) Reference FES of the alanine dipeptide as a function of the CV 𝜙 and
𝜓. Contour levels are plotted every 1 kbT. The black contours indicate the region
within 10 kbT from the global minimum, which is used for the RMSD calculations.
The red triangle, square and circle indicate the position of the minima C7𝑒𝑞 and
C𝑎𝑥 and the transition state ‡, respectively. The color shades are guides for the eye.
(b-f) Difference between the reference FES and those estimated using the balanced
exponential, Tiwary’s, Branduardi’s, and Bonomi’s reweighting, and the negative
bias, respectively, after 1.4 ns in one of the runs. Contour levels are plotted every
0.25 kbT, and the range of color shades is 4 times smaller than that in part a. Adapted
with permission from ref. [16], ©2020 American Chemical Society.
Ligand-induced protein stabilization 15

3
a b

RMSDFES t1/2
0.2 0.6 1

|ΔΔFC7eq-Cax| /kbT
RMSDFES / kbT
0.6

2
0 2 4 6 8 1012
Time / ns

1
3 0.2

1 0
c d
|ΔΔFCax-‡| / kbT

RMSDV / kbT
2

0.5
0 1

0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Time / ns Time / ns
e f 0
-π-π/2 0 π/2 π

log(w/<w>)
φ /rad

-10

0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 8
Time / ns Time / ns

Fig. 5: (a), (b) and (c) provide the time series of the 𝑅𝑀𝑆𝐷 from reference FES, and
the error on the Δ𝐹𝐶7𝑒𝑞 −𝐶𝑎𝑥 and Δ𝐹𝐶𝑎𝑥 −‡ , respectively, for the balanced exponential
(red), Tiwary’s (blue), Branduardi’s (purple), Bonomi’s (orange) reweighting and
negative bias (green) estimate of the FES. The data are averaged over 96 runs. Shaded
bands indicate standard deviations. The black line at 1.4 ns indicates the √ time point
where the FES in Fig. 4 have been extracted. (a inset) the 𝑅𝑀𝑆𝐷 𝐹𝐸𝑆 𝑡 shows to
approximately reach a plateau for balanced exponential, Tiwary’s and Bonomi’s
reweighting schemes. (d) 𝑅𝑀𝑆𝐷 𝑉 between the different reweighting schemes and
negative bias (same color scheme as above). (e), (f) Time series of 𝜓 when 𝜙=-1.88 rad
along one selected run. Balanced exponential (e) and Tiwary (f) weights are reported
according to the color scale. ⟨𝑤⟩ is the average of the weights along the run. Adapted
with permission from ref. [16], ©2020 American Chemical Society.
16 T. Schäfer, A.C. Joerger, J. Spencer, F. Schmid and G. Settanni

The results obtained in this test confirmed the observations made in the uni-
dimensional case (Fig. 5): the balanced exponential reweighting scheme converges
faster than most of the other methods to the reference free energy landscape, with the
exception of the method by Bonomi et.al [34], which, as we demonstrate in ref. [16],
within some limits, may provide similar but not better results.
We also tested the newly developed algorithm in other scenarios: the reweighting
of observables not biased in the metadynamics simulation and in well-tempered
metadynamics. In all tested scenarios the balanced exponential reweighting provided
similar or faster convergence than the other methods and, in addition, lower run-to-run
fluctuations. We refer the interested reader to our original publication [16] for further
details.

4 Conclusions

In this report, we have summarized the results obtained for our projects using the
computational resources made available by the HLRS Stuttgart with the Hazelhen
and Hawk HPC infrastructure. We show how molecular dynamics simulations of
the tumor suppressor protein p53, an important target in cancer research, have been
used to understand the effect of cancer mutations on the stability of the protein as
well as the stabilizing effect of ligands binding to a mutation-induced pocket on the
protein surface. In addition, we proposed an improved reweighting method for the
analysis of metadynamics simulations, which is particularly useful in the context of
large MD simulations of complex systems where, due to the high computational cost,
fast convergence to the underlying free energy landscape is essential.

Acknowledgements This work was partly funded by the German Research Foundation (DFG)
grant SFB1066 project Q1 and SFB TRR 146 as well as DFG grant JO 1473/1-1 and Worldwide
Cancer Research (grants 14-1002, 18-0043). We gratefully acknowledge support with computing
time from the HPC facility Hazelhen and Hawk at the High Performance Computing Center Stuttgart
(HLRS) project Flexadfg and the HPC facility Mogon at the University of Mainz.

References

1. K. Lindorff-Larsen, S. Piana, R.O. Dror, D.E. Shaw, Science 334(6055), 517 (2011)
2. N. Basse, J. Kaar, G. Settanni, A. Joerger, T. Rutherford, A. Fersht, Chemistry and Biology
17(1) (2010)
3. S. Köhler, F. Schmid, G. Settanni, PLoS Computational Biology 11(9) (2015)
4. S. Köhler, F. Schmid, G. Settanni, Langmuir 31(48), 13180 (2015)
5. G. Settanni, J. Zhou, T. Suo, S. Schöttler, K. Landfester, F. Schmid, V. Mailänder, Nanoscale
9(6) (2017)
6. G. Settanni, J. Zhou, F. Schmid, Journal of Physics: Conference Series 921(1), 012002 (2017)
7. G. Settanni, T. Schäfer, C. Muhl, M. Barz, F. Schmid, Computational and Structural Biotech-
nology Journal 16, 543 (2018)
Ligand-induced protein stabilization 17

8. S. Köhler, F. Schmid, G. Settanni, in High Performance Computing in Science and Engineering

’16: Transactions of the High Performance Computing Center Stuttgart (HLRS) 2016 (Springer
International Publishing, 2017), pp. 61–78
9. T. Schäfer, J. Zhou, F. Schmid, G. Settanni, Blood proteins and their interactions with
nanoparticles investigated using molecular dynamics simulations (2018)
10. T. Schafer, C. Muhl, M. Barz, F. Schmid, G. Settanni, in High Performance Computing in
Science and Engineering ’ 18 (Springer International Publishing, 2019), pp. 63–74
11. T. Schäfer, C. Muhl, M. Barz, F. Schmid, G. Settanni, in High Performance Computing in
Science and Engineering ’19 (Springer International Publishing, Cham, 2021), pp. 65–76
12. A.C. Joerger, A.R. Fersht, Annual Review of Biochemistry 85, 375 (2016)
13. A. Laio, M. Parrinello, Proceedings of the National Academy of Sciences of the United States
of America 99(20), 12562 (2002)
14. G. Bussi, A. Laio, M. Parrinello, Physical review letters 96(9), 090601 (2006)
15. A. Barducci, G. Bussi, M. Parrinello, Physical Review Letters (2008)
16. T.M. Schäfer, G. Settanni, Journal of Chemical Theory and Computation 16(4), 2042 (2020)
17. H.J.C. Berendsen, D. van der Spoel, R. van Drunen, Computer Physics Communications
91(1–3), 43 (1995)
18. J. Huang, S. Rauscher, G. Nawrocki, T. Ran, M. Feig, B.L. De Groot, H. Grubmüller, A.D.
MacKerell, Nature Methods 14(1), 71 (2016)
19. K. Vanommeslaeghe, E. Hatcher, C. Acharya, S. Kundu, S. Zhong, J. Shim, E. Darian,
O. Guvench, P. Lopes, I. Vorobyov, A.D. Mackerell, J. Comput. Chem. 31(4), 671 (2010)
20. B. Hess, H. Bekker, H.J.C. Berendsen, J.G.E.M. Fraaije, LINCS: A Linear Constraint Solver
for Molecular Simulations. Tech. rep. (1997)
21. M.P. Allen, D.J. Tildesley, Computer Simulation of Liquids (Clarendon Press., 1987)
22. U. Essmann, L. Perera, M.L. Berkowitz, T. Darden, H. Lee, L.G. Pedersen, The Journal of
Chemical Physics 103(19), 8577 (1995)
23. W.L. Jorgensen, J. Chandrasekhar, J.D. Madura, R.W. Impey, M.L. Klein, Journal of Chemical
Physics 79(2), 926 (1983)
24. M. Parrinello, A. Rahman, Physical Review Letters 45(14), 1196 (1980)
25. S. Nosé, The Journal of Chemical Physics 81(1), 511 (1984)
26. W.G. Hoover, Physical Review A 31(3), 1695 (1985)
27. A.C. Joerger, M.D. Allen, A.R. Fersht, Journal of Biological Chemistry 279(2), 1291 (2004)
28. M.R. Bauer, A. Krämer, G. Settanni, R.N. Jones, X. Ni, R. Khan Tareque, A.R. Fersht, J. Spencer,
A.C. Joerger, ACS Chemical Biology 15(3), 657 (2020)
29. W. Humphrey, A. Dalke, K. Schulten, Journal of Molecular Graphics 14, 33 (1996)
30. M. Seeber, M. Cecchini, F. Rao, G. Settanni, A. Caflisch, Bioinformatics 23(19), 2625 (2007)
31. M.R. Bauer, R.N. Jones, R.K. Tareque, B. Springett, F.A. Dingler, L. Verduci, K.J. Patel, A.R.
Fersht, A.C. Joerger, A.C. Joerger, Future Medicinal Chemistry 11(19), 2491 (2019)
32. G. Interlandi, S.K. Wetzel, G. Settanni, A. Plückthun, A. Caflisch, Journal of Molecular Biology
375(3), 837 (2008)
33. G. Tiana, European Physical Journal B 63(2), 235 (2008)
34. M. Bonomi, A. Barducci, M. Parrinello, Journal of Computational Chemistry 30(11), 1615
(2009)
35. J. Smiatek, A. Heuer, Journal of Computational Chemistry 32(10), 2084 (2011)
36. D. Branduardi, G. Bussi, M. Parrinello, Journal of Chemical Theory and Computation (2012)
37. P. Tiwary, M. Parrinello, The Journal of Physical Chemistry B 119(3), 736 (2015)
38. J.F. Dama, M. Parrinello, G.A. Voth, Physical Review Letters 112(24), 240602 (2014)
39. G.A. Tribello, M. Bonomi, D. Branduardi, C. Camilloni, G. Bussi, Computer Physics Commu-
nications 185(2), 604 (2014)
40. D.A. Case, T.E. Cheatham, T. Darden, H. Gohlke, R. Luo, K.M. Merz, A. Onufriev, C. Simmer-
ling, B. Wang, R.J. Woods, Journal of Computational Chemistry 26(16), 1668 (2005)
Hadronic contributions to the anomalous
magnetic moment of the muon from Lattice QCD

M. Cè, E.-H. Chao, A. Gérardin, J.R. Green, G. von Hippel, B. Hörz, R.J. Hudspith,
H.B. Meyer, K. Miura, D. Mohler, K. Ottnad, S. Paul, A. Risch, T. San José and
H. Wittig

Abstract The recently reported new measurement of the anomalous magnetic moment
of the muon, 𝑎 𝜇 , by the E989 collaboration at Fermilab has increased the tension
with the Standard Model (SM) prediction to 4.2 standard deviations. In order to
increase the sensitivity of SM tests, the precision of the theoretical prediction, which
is limited by the strong interaction, must be further improved. In our project we
employ lattice QCD to compute the leading hadronic contributions to 𝑎 𝜇 and various
other precision observables, such as the energy dependence (“running”) of the
electromagnetic coupling, 𝛼, and the electroweak mixing angle, sin2 𝜃 W . Here we
report on the performance of our simulation codes used for the generation of gauge
ensembles at (near-)physical pion masses and fine lattice spacings. Furthermore, we
present results for the hadronic running of 𝛼, the electroweak mixing angle, as well

H.B. Meyer, K. Miura, T. San José and H. Wittig

Helmholtz Institut Mainz, Johannes Gutenberg Universität, 55099 Mainz, Germany
K. Miura, D. Mohler and T. San José
GSI Helmholtzzentrum für Schwerionenforschung, Darmstadt, Germany
M. Cè and J. R. Green
Theoretical Physics Department, CERN, Geneva, Switzerland
E.-H. Chao, G. von Hippel, R.J. Hudspith, H.B. Meyer, K. Ottnad, S. Paul, T. San José and H. Wittig
Institut für Kernphysik and PRISMA+ Cluster of Excellence, Universität Mainz, Johann-Joachim-
Becher-Weg 45, D-55099 Mainz, Germany
A. Risch
John von Neumann Institute for Computing NIC, Deutsches Elektronen-Synchrotron DESY,
Platanenallee 6, 15738 Zeuthen, Germany
A. Gérardin
Aix Marseille Univ, Université de Toulon, CNRS, CPT, Marseille, France
B. Hörz
Nuclear Science Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA
K. Miura
Kobayashi-Maskawa Institute for the Origin of Particles and the Universe, Nagoya University,
Nagoya 464-8602, Japan

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 19

W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_2
20 M. Cè et al.

as the hadronic vacuum polarisation and light-by-light scattering contributions to 𝑎 𝜇 .

Results from an ancillary calculation of the spectrum in the isovector channel are
crucial in order to further increase the precision of our determination of the hadronic
vacuum polarisation contribution.

1 Introduction

The Standard Model of Particle Physics provides a quantitative and precise description
of the properties of the known constituents of matter in terms of a uniform theoretical
formalism. However, despite its enormous success, the Standard Model (SM) does
not explain some of the most pressing problems in particle physics, such as the nature
of dark matter or the asymmetry between matter and antimatter. The world-wide
quest for discovering physics beyond the SM involves several different strategies,
namely (1) the search for new particles and interactions that are not described by the
SM, (2) the search for the enhancement of rare processes by new interactions, and (3)
the comparison of precision measurements with theoretical, SM-based predictions of
the same quantity. These complementary activities form an integral part of the future
European strategy for particle physics [1].
Precision observables, such as the anomalous magnetic moment of the muon,
𝑎 𝜇 , have provided intriguing hints for the possible existence of “new physics”. The
longstanding tension between the direct measurement of 𝑎 𝜇 and its theoretical
prediction has recently increased to 4.2 standard deviations, following the publication
of the first result from the E989 experiment at Fermilab [2]. As E989 prepares to
improve the experimental precision further, it is clear that the theoretical prediction
must be pushed to a higher level of accuracy as well, in order to increase the sensitivity
of the SM test. Since the main uncertainties of the SM prediction arise from strong
interaction effects, current efforts are focussed on quantifying the contributions from
hadronic vacuum polarisation (HVP) and hadronic light-by-light scattering (HLbL).
This has also been emphasised in the 2020 White Paper [3] in which the status of the
theoretical prediction is reviewed.
Our project is focussed on calculations of the hadronic contributions to the muon
anomalous magnetic moment from first principles, using the methodology of Lattice
QCD. To this end, we perform calculations of the HVP contribution at the physical
value of the pion mass, in order to reduce systematic errors. Another highly important
ingredient of our calculation is the determination of the spectrum in the isovector
channel of the electromagnetic current correlator, which constrains the long-distance
contribution to the HVP. Our group has also developed a new formalism for the direct
calculation of the HLbL contribution, which has produced the most precise estimate
from first principles so far [4].
The HVP contribution to the muon anomalous magnetic moment is closely linked
to the hadronic effects that modify the value of the electromagnetic coupling, Δ𝛼.
Since Δ𝛼 depends on other SM parameters such as the mass of the 𝑊-boson, a precise
determination provides important information for precision tests of the SM. Finally,
The muon anomaly from Lattice QCD 21

we also compute the hadronic contributions to the “running” of the electroweak

mixing angle, a precision observable which is particularly sensitive to the effects of
physics beyond the SM in the regime of low energies.

2 Computational setup

One of the major computational tasks of our project is the generation of gauge-
field ensembles at (close to) physical light-quark masses. For a particular challenge
encountered in these simulations please refer to [5]. The generation of a gauge field
ensemble dubbed E250 at physical pion and kaon masses has been a long standing
goal of our programme on Hazel Hen and Hawk, and has been finalized since the last
report. Furthermore we recently produced two somewhat coarser lattices named D452
and D1521 at light pion masses. For both ensembles, the generation of 500 gauge field
configurations, corresponding to 2000 molecular dynamics units (MDU) had been
proposed. As it turned out, the rather coarse lattice spacing of D152 lead to sustained
algorithmic problems, hence the run was stopped after 275 gauge configurations
(1100 MDU). For D452 no such issues were observed, and we were able to produce
1000 gauge configurations (4000 MDU) due to better than expected performance for
this run. Preliminary results for observables suggest that ensemble D452 will play a
vital role for obtaining more precise results for the observables in our project.
Figure 1 shows the Hamiltonian deficits Δ𝐻 as well as the Monte Carlo history of
the topological charge for ensembles E250 and D452. The acceptance rate resulting
from the history of Δ𝐻 is (87.1 ± 1.0)% for E250 and (91.5 ± 0.7)% for D452. The
generation of these chains is now complete and the calculation of physics observables
has been started on compute clusters operated by JGU Mainz.
The openQCD code used in our calculations exhibits excellent scaling properties
over a wide range of problem sizes. Figure 3 shows the strong-scaling behaviour
for the system size corresponding to ensemble D452, i.e. for a 128 × 643 lattice, as
measured on Hawk (left pane). The timings refer to the application of the even-odd
preconditioned O(𝑎) improved Wilson–Dirac operator 𝐷ˆ w to a spinor field which
accounts for the largest fraction of the total computing time for several of our projects.
For the gauge field generation runs on Hawk and for the HLbL runs on lattices of
size 128 × 643 we used the following setup:
A Local lattice volume of size 84 per MPI rank with 8192 MPI ranks on 64 nodes.
In addition to this setup, we performed spectroscopy runs on J303 (192 × 643 ) with
setup B and on E250 192 × 963 with setup C:
B Local lattice of size 12 × 83 per MPI rank with 8192 MPI ranks on 64 nodes.
C Local lattice volume of size 12 × 62 × 12 per MPI rank with 32768 MPI ranks on
256 nodes.

1 This ensemble was called D151 in the proposal.

22 M. Cè et al.

Fig. 1: Monte-Carlo histories of the Hamiltonian deficit Δ𝐻 (left) and the total
topological charge (right) for E250 (top) and D452 (bottom).

When possible a hypercube of size 22 × 42 processes was grouped onto a single

CPU, to minimise off-CPU communication.
For the current allocation year we received a total of 100 MCore hours. It was
foreseen that 32% / 40% / 27% would be spent on the gauge field generation / HVP
spectroscopy / HLbL respectively, while the actual percentages of computing time
spent by the time of writing this report are given by 30% / 44% / 26%.

3 The hadronic running of the electroweak couplings

We start our discussion of precision observables with the hadronic contributions

to the energy dependence (“running”) of the electromagnetic coupling, 𝛼, since its
calculation shares many features and definitions with the determination of the HVP
contribution to the muon anomalous moment.
The electromagnetic coupling in the Thomson limit is one of the most precisely
known quantities, 𝛼 = 1/137.035999084(21) [6]. However, for scales above a few
hundred MeV, the low-energy hadronic contribution to the vacuum polarization
The muon anomaly from Lattice QCD 23
128
128
120
112
112
104
96
96
88
80
80

Speedup
Speedup

72
64 64
56
48 48
40
32 32
D200 on Hawk
"ideal" scaling 24
16 16
8
0 0
0 4096 8192 12288 16384 20480 24576 28672 32768 0 4096 8192 12288 16384 20480 24576 28672 32768
# cores # cores

Fig. 2: Left: Strong-scaling behaviour of the openQCD code on Hawk. The plot
shows the application of the even-odd preconditioned Wilson–Dirac operator on a
128 × 643 lattice. Speedup factors are defined relative to 256 cores. A clear indication
of hyperscaling is seen in this regime. Right: Weak scaling with a local volume of 84
on Hawk.

induces a theoretical uncertainty. If one parameterizes the energy dependence, or

“running”, of the coupling in the so-called on-shell scheme,
𝛼
𝛼(𝑞 2 ) = , (1)
1 − Δ𝛼(𝑞 2 )
(5)
the five flavour quark contribution at the Z-pole is found to be Δ𝛼had (𝑀𝑍2 ) =
0.02766(7) [6]. Its poorer precision is a limiting factor for the global fit of the
electroweak sector of the Standard Model (SM). Thus, the goal is to improve the
precision of this estimate, by reducing the uncertainties associated with the strong
interaction. We concentrate on the quark contribution at low energies, which is given
by the one-subtracted dispersion relation
h i
Δ𝛼had (𝑞 2 ) = 4𝜋𝛼 Π̄ 𝛾𝛾 (𝑞 2 ), Π̄ 𝛾𝛾 (𝑞 2 ) = Re Π 𝛾𝛾 (𝑞 2 ) − Π 𝛾𝛾 (0) . (2)

The standard method to obtain eq. (2) employs the optical theorem, which relates
the HVP function with the so-called R-ratio, i.e. the total hadronic cross section
𝜎(𝑒 + 𝑒 − → hadrons) normalized by 𝜎(𝑒 + 𝑒 − → 𝜇+ 𝜇− ), via a dispersion integral.
While the integral can be evaluated using experimental data for the 𝑅-ratio in the low-
energy domain, this procedure introduces experimental uncertainties into a theoretical
prediction. Therefore, lattice computations in the space-like region 𝑄 2 = −𝑞 2 provide
(5)
a valuable ab initio crosscheck. In order to estimate Δ𝛼had (𝑀𝑍2 ), we use the so-called
Euclidean split technique [7, 8]
24 M. Cè et al.
(5) (5)
Δ𝛼had (𝑀𝑍2 ) = Δ𝛼had (−𝑄 20 )
h i
(5) (5)
+ Δ𝛼had (−𝑀𝑍2 ) − Δ𝛼had (−𝑄 20 ) (3)
h i
(5) (5)
+ Δ𝛼had (𝑀𝑍2 ) − Δ𝛼had (−𝑀𝑍2 ) .

The first term is the result of this project at the threshold energy 𝑄 20 = 5 GeV2 , while
the second and third terms can be evaluated in perturbative QCD, with or without the
help of experimental data.
There is growing interest to probe electroweak precision observables such as
sin2 𝜃 W (𝑞 2 ) at momentum transfers 𝑞 2 ≪ 𝑀𝑍2 in parity-violating lepton scattering
experiments. Such measurements are sensitive to modifications of the running of the
mixing angle due to beyond the Standard Model (BSM) physics. At leading order,
the hadronic contribution to the running of sin2 𝜃 W is given by [9]
4𝜋𝛼
Δhad sin2 𝜃 W (𝑞 2 ) = − Π̄ 𝑍 𝛾 (𝑞 2 ). (4)
sin2 𝜃 W
We employ the time-momentum representation (TMR) [10, 11] to compute
the vacuum polarization functions Π̄ 𝛾𝛾 and Π̄ 𝑍 𝛾 . This allows us to represent our
observables in terms of integrals over two-point functions with a known 𝑄 2 -dependent
kernel 𝐾 (𝑡, 𝑄 2 ) = 𝑡 2 − 4/𝑄 2 sin2 (𝑄𝑡/2),
∫ ∞
Π̄(−𝑄 2 ) = d𝑡 𝐺 (𝑡)𝐾 (𝑡, 𝑄 2 ), (5)
0

where 𝐺 (𝑡) is the correlator, projected to zero momentum and averaged over the three
spatial directions to improve the signal,
3
1 ∑︁ ∑︁
𝐺 (𝑡) = − 𝑗 𝑘 (𝑥) 𝑗 𝑘 (0) . (6)
3 𝑘=1 x

𝛾
The two currents 𝑗 𝑘 are either the electromagnetic 𝑗 𝜇 or the vector part of the
weak neutral current 𝑗 𝜇𝑍 ,

𝛾 1 2
𝑗 𝜇 = 𝑗 𝜇3 + √ 𝑗 𝜇8 + 𝑗 𝜇𝑐 , (7a)
3 3
 
1 𝛾 1 1 𝑐
𝑗 𝜇𝑍 | vector = − sin2 𝜃 W 𝑗 𝜇 − 𝑗 𝜇0 − 𝑗 . (7b)
2 6 12 𝜇

Figure 3 shows the hadronic contribution to the running of 𝛼 and sin2 (𝜃 𝑊 ) for a
domain of low 𝑄 2 . The error bands represent the total uncertainties affecting these
observables: statistical, scale setting, extrapolation to the physical point and isospin
breaking corrections.
The muon anomaly from Lattice QCD 25

Fig. 3: Hadronic contributions to the running of 𝛼 (left) and sin2 𝜃 W (right) with
the energy. We depict both the total running as well as the different components
according to the SU(3)-isospin decomposition and the charm quark contribution.

Fig. 4: Comparison of the hadronic contribution to the running of 𝛼 at several energies.

The top row and band show our result. The horizontal line separates lattice from
phenomenological results. We find agreement with BMW [12], slight tensions with
KNT [13] and bigger discrepancy with DHMZ [14] and Jegerlehner [15].
26 M. Cè et al.

Several other groups have determined Δ𝛼(𝑄 2 ). In Figure 4, we compare our

results (top row) with another lattice calculation [12] and the findings of three groups
[13–15] that employ the data-driven approach based on the R-ratio. Our results are
consistent with the lattice calculation by BMW, while a slight tension is observed
with respect to phenomenological determinations.

4 Hadronic vacuum polarization contribution to 𝒂 𝝁

The hadronic vacuum polarization is the single largest contributor to the error budget
of the Standard Model prediction for the anomalous magnetic moment of the muon [3]
𝑎 𝜇 . Recently, the BMW collaboration [16] has published a value for this contribution
hvp
𝑎 𝜇 , which is larger than the value quoted in the White Paper [3]; the latter is based
on a dispersive relation with input from experimentally measured 𝑒 + 𝑒 − → hadrons
cross-sections. Taken at face value, the BMW results would strongly reduce the
current tension with the experimental world average for 𝑎 𝜇 [17]. Therefore, it is vital
for this important test of the Standard Model to resolve the tension between the BMW
hvp
and the dispersive result for 𝑎 𝜇 .

4.1 The intermediate-distance window contribution

hvp
Current lattice calculations of 𝑎 𝜇 are performed in the ‘time-momentum’ represen-
tation [10],
  2 ∑︁
hvp 𝛼
𝑎𝜇 = 𝑎 𝐺 (𝑡) 𝐾
e(𝑡), (8)
𝜋 𝑡 ≥0

where 𝐾 e(𝑡) is a positive, exactly known kernel [18] and 𝐺 (𝑡) is the vector correlator
hvp
introduced in Eq. (6). It has proved useful to decompose 𝑎 𝜇 = 𝑎 winSD
𝜇 +𝑎 winID
𝜇 +𝑎 winLD
𝜇
into a short-distance, intermediate-distance and a long-distance contribution,
  2 ∑︁
𝛼
𝑎 winSD
𝜇 = 𝑎 𝐺 (𝑡) 𝐾
e(𝑡) [1 − Θ(𝑡, 𝑡0 , Δ)], (9)
𝜋 𝑡 ≥0
  2 ∑︁
𝛼
𝑎 winID
𝜇 = 𝑎 𝐺 (𝑡) 𝐾
e(𝑡) [Θ(𝑡, 𝑡0 , Δ) − Θ(𝑡, 𝑡1 , Δ)], (10)
𝜋 𝑡 ≥0
  2 ∑︁
𝛼
𝑎 winLD
𝜇 = 𝑎 𝐺 (𝑡) 𝐾
e(𝑡) Θ(𝑡, 𝑡1 , Δ). (11)
𝜋 𝑡 ≥0

Here Θ(𝑡, 𝑡¯, Δ) = (1 + tanh[(𝑡 − 𝑡¯)/Δ])/2 is a smoothened step function, and the
standard choice of parameters is 𝑡 0 = 0.4 fm, 𝑡1 = 1.0 fm and Δ = 0.15 fm. Particularly
the intermediate-distance contribution 𝑎 winID
𝜇 can be computed on the lattice with
The muon anomaly from Lattice QCD 27

hvp
smaller (relative) statistical and systematic uncertainty than the total 𝑎 𝜇 . Therefore
𝑎 winID
𝜇 has emerged as an excellent benchmark quantity to first test the consistency
of different lattice QCD calculations, and in a second step their consistency with
the data-driven evaluation based on the 𝑅-ratio. Furthermore, the contributions of
different quark-flavour combinations can be compared separately between lattice
calculations.

Fig. 5: Left: chiral and continuum extrapolation of the connected charm contribu-
tion to 𝑎 winID
𝜇 , with 𝑦˜ = (𝑚 𝜋 /(4𝜋 𝑓 𝜋 )) 2 , 𝑓 𝜋 being the pion decay constant. Right:
Extrapolation in the SU(3)f -breaking difference of squared kaon and pion masses
of the 𝑢, 𝑑, 𝑠 quark-disconnected contribution to 𝑎 winID 𝜇 . In both panels, the target
extrapolation point is marked by a dashed vertical line. The 𝛽 values are in one-to-one
correspondence with the lattice spacing 𝑎.

As sample results, we present the charm-quark as well as the quark-disconnected

contributions to 𝑎 winID
𝜇 in Figure 5. The results are

𝑎 winID,charm
𝜇 = (+2.85 ± 0.12) × 10−10 , (12)
𝑎 winID,disc
𝜇 = (−0.87 ± 0.03stat ± 0.07syst ) × 10−10 . (13)

In the case of the disconnected contribution, the systematic error is estimated as

the difference between the linear fit in (𝑚 2𝐾 − 𝑚 2𝜋 ) 2 of all data and our most chiral
ensemble E250. We also point out that the physical values of the lattice spacings
𝑎 were determined using the results of Ref. [19]. The analysis of the light-quark
contributions are in the final stage, and we provide the relative-error budget for the
isovector contribution, which represents about 80% of the total 𝑎 winID𝜇 (Table 1).
The dominance of the chiral and continuum extrapolation error highlights the
importance of the physical-quark-mass ensemble E250.
28 M. Cè et al.

Type %
Extrapolation 0.6
Scale setting 0.4
Statistical error of 𝐺 (𝑡) 0.2
Finite-size effect ≤ 0.2
Renormalization < 0.01
Total 0.77

Table 1

4.2 Scattering phase shift and the timelike pion form factor

The computation of 𝐼 = 1 p-wave 𝜋𝜋 scattering can be used to reduce the uncertainty

on the long-distance contribution to the integrand for the light-quark HVP [20], in
particular on 𝑎 winLD
𝜇 defined in Eq. (11). The fact that this issue has not been addressed
in our discussion of the running of 𝛼 is related to the kernel function 𝐾 e(𝑡) in the
hvp
integral representation of 𝑎 𝜇 in (8), which gives a higher weight to the long-distance
regime.
In addition, our related determination of the timelike pion form factor [21, 22],
can be used to access infinite-volume physics from the lattice calculation in a finite
volume, providing a valuable cross-check of finite-volume effects encountered in
lattice calculations of the HVP [23]. To this end we proposed the determination of
the 𝑙 = 1 pion-pion scattering phase shift 𝛿1 and of the pion timelike form factor on
ensembles J303 and E250. During the current reporting period we finished the full
statistics for J303 and obtained first results for E250. Figure 6 shows 𝛿1 for ensemble
J303.
The next steps will include the reconstruction of the isovector part of the correlator
𝐺 (𝑡) for this ensemble, from which an accurate determination of the long-distance
regime is possible. The corresponding analysis for ensemble E250 is in progress.
The precise determination of 𝐺 (𝑡) at large distances will be crucial for reducing the
hvp
statistical error on 𝑎 𝜇 below the percent level.

5 Hadronic light-by-light scattering contribution to 𝒂 𝝁

Although sub-leading in the electromagnetic coupling (i.e. 𝑂 (𝛼3 ) compared to

𝑂 (𝛼2 )) and therefore a smaller contribution to 𝑎 𝜇 than the HVP, the hadronic light-
by-light contribution has a large associated theoretical error. This is mostly due to
the difficulty in computing such a complicated contribution. To match the upcoming
expected experimental precision from E989 and J-PARC, it has been suggested
that the theory error should be around 10 × 10−11 [3], whereas the currently best
estimate included in the White Paper estimate is 92(19) × 10−11 [3]. Following
our infinite-volume QED, position-space methodology outlined in [24], we have
The muon anomaly from Lattice QCD 29

Fig. 6: Isospin 1, pion-pion p-wave phase shift 𝛿1 on ensemble J303 as a function

of the center of momentum energy 𝐸 𝑐𝑚 in lattice units. The lattice data is shown in
blue, the red curve shows the results of a parameterization with a “pole plus constant”
form for the scattering K-matrix.

determined the HLbL contribution at the physical pion mass [4]. The crucial results
on the two largest-volume, lowest-pion mass results (i.e. ensembles D200 and D450,
the left-most blue and green points in the left plot of Fig. 8) were computed on Hawk;
without the resources granted to us, two of our near-physical determinations would
have been absent and likely our approach to the physical pion mass would be less
well-constrained.

Fig. 7: Different quark Wick-contraction classes appearing in the computation of the

QCD four-point correlation function. The straight horizontal lines represent muon
propagators, wavy lines represent photon propagators. From left to right, top to
bottom, they are the fully-connected, (2 + 2), (3 + 1), (2 + 1 + 1) and (1 + 1 + 1 + 1).
Each class contains the diagrams obtained from all the possible permutations of the
four points attached to photons.
30 M. Cè et al.

Fig. 8: (Left) The combined chiral/finite-volume/continuum extrapolation of the two

leading diagrammatic contributions to the HLbL. (Right) The chiral extrapolation for
the (3 + 1) contribution to the HLbL.

As is expected by the charge factors and large-𝑁 𝑐 arguments, only the leading
light-quark connected and (2 + 2)-disconnected diagrams contribute significantly
to hadronic light-by-light scattering. However, there are three additional topologies,
whose magnitude is argued to be small but have not been directly computed so far,
those are the (3 + 1), the (2 + 1 + 1), and the (1 + 1 + 1 + 1) contributions, shown
in Fig. 7. Our work [4] was the first to determine all of these contributions (which
all turn out to be consistent with zero within our statistical precision), and the data
for D200 and D450 generated on Hawk were used in the extrapolation of the (3 + 1)
light-quark topology (the right-most blue and green points of Fig. 8).

Fig. 9: A comparison of our HLbL result with the literature. The results in circles
are the two available lattice determinations (this work and [25], above the horizontal
dashed line). The results in squares are phenomenological predictions from [3],[26],
[27, 28], and [29].
The muon anomaly from Lattice QCD 31

Following our precise determination of all relevant parts to this contribution our
final result (including systematic effects) compares favourably to a previous lattice
determination and to phenomenological predictions in general, as is illustrated in
Figure 9. Our result is the most precise determination thus far, and including it in a
global average with other independent determinations gives 𝑎 HLbL 𝜇 = 97.5(11.6) ×
10−11 . This number is precise enough to meet the expected precision of the experiments
and brings the focus back to the precision determination of the HVP.

6 Summary and outlook

We have reported on the current status of our determination of a set of precision

observables that are crucial for exploring the limits of the Standard Model, using the
ab initio formalism of lattice QCD. The continued production of gauge ensembles
close to the physical pion mass is crucial for achieving full control over systematic
effects at the desired overall accuracy.
Our results for the hadronic running of the electromagnetic coupling and weak
mixing angle are already at the same level of precision as reached in conventional
data-driven approaches and can be readily applied in phenomenological studies.
hvp
A full account will be published soon. Regarding the HVP contribution 𝑎 𝜇 , we
have focussed mostly on the intermediate-window quantity which is emerging as a
benchmark quantity that facilitates the comparison between different calculations. For
the other most important hadronic contribution, arising from light-by-light scattering,
we have published the most accurate result so far.
hvp
Our focus will turn to increasing the overall precision of our determination of 𝑎 𝜇
to the sub-percent level.

Acknowledgements This work is partly supported by the Deutsche Forschungsgemeinschaft (DFG,

German Research Foundation) grant HI 2048/1-2 (Project ID 399400745) and by the DFG-funded
Collaborative Research Centre SFB 1044 The low-energy frontier of the Standard Model. The Mainz
(𝑔 − 2) 𝜇 project is also supported by the Cluster of Excellence Precision Physics, Fundamental
Interactions, and Structure of Matter (PRISMA+ EXC 2118/1) funded by the DFG within the German
Excellence Strategy (Project ID 39083149). This work is supported by the European Research
Council (ERC) under the European Union’s Horizon 2020 research and innovation programme
through grant agreement 771971-SIMDAMA. The work of M.C. is supported by the European
Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie Grant
Agreement No. 843134. The authors gratefully acknowledge the Gauss Centre for Supercomputing
e.V. (www.gauss-centre.eu) for supporting this project by providing computing time on the GCS
Supercomputer Hazel Hen and Hawk at Höchstleistungsrechenzentrum Stuttgart (www.hlrs.de)
under project GCS-HQCD. Additional calculations were performed on the HPC clusters “Clover”
and “HIMster II” at the Helmholtz-Institut Mainz and “MOGON II” at JGU Mainz. We are grateful
to our colleagues in the CLS initiative for sharing ensembles.
32 M. Cè et al.

References

1. R.K. Ellis, et al., Physics Briefing Book: Input for the European Strategy for Particle Physics
Update 2020 (2019)
2. B. Abi, et al., Measurement of the Positive Muon Anomalous Magnetic Moment to 0.46 ppm,
Phys. Rev. Lett. 126(14), 141801 (2021). DOI 10.1103/PhysRevLett.126.141801
3. T. Aoyama, et al., The anomalous magnetic moment of the muon in the Standard Model, Phys.
Rept. 887, 1 (2020). DOI 10.1016/j.physrep.2020.07.006
4. E.H. Chao, R.J. Hudspith, A. Gérardin, J.R. Green, H.B. Meyer, K. Ottnad, Hadronic light-by-
light contribution to (𝑔 − 2) 𝜇 from lattice QCD: a complete calculation, Eur. Phys. J. C 81(7),
651 (2021). DOI 10.1140/epjc/s10052-021-09455-4
5. D. Mohler, S. Schaefer, Remarks on strange-quark simulations with Wilson fermions, Phys.
Rev. D 102(7), 074506 (2020). DOI 10.1103/PhysRevD.102.074506
6. P. Zyla, et al., Review of Particle Physics, PTEP 2020(8), 083C01 (2020). DOI 10.1093/ptep/p
taa104
7. S. Eidelman, F. Jegerlehner, A.L. Kataev, O. Veretin, Testing nonperturbative strong interaction
effects via the Adler function, Phys. Lett. B 454, 369 (1999). DOI 10.1016/S0370-2693(99)003
89-5
8. F. Jegerlehner, The Running fine structure constant alpha(E) via the Adler function, Nucl. Phys.
B Proc. Suppl. 181-182, 135 (2008). DOI 10.1016/j.nuclphysbps.2008.09.010
9. F. Burger, K. Jansen, M. Petschlies, G. Pientka, Leading hadronic contributions to the running
of the electroweak coupling constants from lattice QCD, JHEP 11, 215 (2015). DOI
10.1007/JHEP11(2015)215
10. D. Bernecker, H.B. Meyer, Vector Correlators in Lattice QCD: Methods and applications, Eur.
Phys. J. A 47, 148 (2011). DOI 10.1140/epja/i2011-11148-6
11. A. Francis, B. Jaeger, H.B. Meyer, H. Wittig, A new representation of the Adler function for
lattice QCD, Phys. Rev. D 88, 054502 (2013). DOI 10.1103/PhysRevD.88.054502
12. S. Borsanyi, et al., Hadronic vacuum polarization contribution to the anomalous magnetic
moments of leptons from first principles, Phys. Rev. Lett. 121(2), 022002 (2018). DOI
10.1103/PhysRevLett.121.022002
13. A. Keshavarzi, D. Nomura, T. Teubner, 𝑔 − 2 of charged leptons, 𝛼( 𝑀𝑍2 ) , and the hyperfine
splitting of muonium, Phys. Rev. D 101(1), 014029 (2020). DOI 10.1103/PhysRevD.101.014029
14. M. Davier, A. Hoecker, B. Malaescu, Z. Zhang, A new evaluation of the hadronic vacuum
polarisation contributions to the muon anomalous magnetic moment and to 𝜶 (m2Z ), Eur. Phys.
J. C 80(3), 241 (2020). DOI 10.1140/epjc/s10052-020-7792-2. [Erratum: Eur.Phys.J.C 80, 410
(2020)]
15. F. Jegerlehner, 𝛼QED,eff (s) for precision physics at the FCC-ee/ILC, CERN Yellow Reports:
Monographs 3, 9 (2020). DOI 10.23731/CYRM-2020-003.9
16. S. Borsanyi, et al., Leading hadronic contribution to the muon magnetic moment from lattice
QCD, Nature 593(7857), 51 (2021). DOI 10.1038/s41586-021-03418-1
17. B. Abi, et al., Measurement of the Positive Muon Anomalous Magnetic Moment to 0.46 ppm,
Phys. Rev. Lett. 126(14), 141801 (2021). DOI 10.1103/PhysRevLett.126.141801
18. M. Della Morte, A. Francis, V. Gülpers, G. Herdoíza, G. von Hippel, H. Horch, B. Jäger, H.B.
Meyer, A. Nyffeler, H. Wittig, The hadronic vacuum polarization contribution to the muon
𝑔 − 2 from lattice QCD, JHEP 10, 020 (2017). DOI 10.1007/JHEP10(2017)020
19. M. Bruno, T. Korzec, S. Schaefer, Setting the scale for the CLS 2 + 1 flavor ensembles, Phys.
Rev. D 95(7), 074504 (2017). DOI 10.1103/PhysRevD.95.074504
20. A. Gérardin, M. Cè, G. von Hippel, B. Hörz, H.B. Meyer, D. Mohler, K. Ottnad, J. Wilhelm,
H. Wittig, The leading hadronic contribution to (𝑔 − 2) 𝜇 from lattice QCD with 𝑁f = 2 + 1
flavours of O(𝑎) improved Wilson quarks, Phys. Rev. D 100(1), 014510 (2019). DOI
10.1103/PhysRevD.100.014510
21. H.B. Meyer, Lattice QCD and the Timelike Pion Form Factor, Phys. Rev. Lett. 107, 072002
(2011). DOI 10.1103/PhysRevLett.107.072002
The muon anomaly from Lattice QCD 33

22. C. Andersen, J. Bulava, B. Hörz, C. Morningstar, The 𝐼 = 1 pion-pion scattering amplitude and
timelike pion form factor from 𝑁f = 2 + 1 lattice QCD, Nucl. Phys. B 939, 145 (2019). DOI
10.1016/j.nuclphysb.2018.12.018
23. F. Erben, J.R. Green, D. Mohler, H. Wittig, Rho resonance, timelike pion form factor, and
implications for lattice studies of the hadronic vacuum polarization, Phys. Rev. D 101(5),
054504 (2020). DOI 10.1103/PhysRevD.101.054504
24. E.H. Chao, A. Gérardin, J.R. Green, R.J. Hudspith, H.B. Meyer, Hadronic light-by-light
contribution to (𝑔 − 2) 𝜇 from lattice QCD with SU(3) flavor symmetry, Eur. Phys. J. C 80(9),
869 (2020). DOI 10.1140/epjc/s10052-020-08444-3
25. T. Blum, N. Christ, M. Hayakawa, T. Izubuchi, L. Jin, C. Jung, C. Lehner, Hadronic Light-by-
Light Scattering Contribution to the Muon Anomalous Magnetic Moment from Lattice QCD,
Phys. Rev. Lett. 124(13), 132002 (2020). DOI 10.1103/PhysRevLett.124.132002
26. F. Jegerlehner, The Anomalous Magnetic Moment of the Muon, vol. 274 (Springer, Cham, 2017).
DOI 10.1007/978-3-319-63577-4
27. A. Nyffeler, Hadronic light-by-light scattering in the muon g-2: A New short-distance constraint
on pion-exchange, Phys. Rev. D 79, 073012 (2009). DOI 10.1103/PhysRevD.79.073012
28. F. Jegerlehner, A. Nyffeler, The Muon g-2, Phys. Rept. 477, 1 (2009). DOI 10.1016/j.physrep.
2009.04.003
29. J. Prades, E. de Rafael, A. Vainshtein, The Hadronic Light-by-Light Scattering Contribution to
the Muon and Electron Anomalous Magnetic Moments, Adv. Ser. Direct. High Energy Phys.
20, 303 (2009). DOI 10.1142/9789814271844\_0009
Quantum simulators, phase transitions, resonant
tunneling, and variances: A many-body
perspective

A.U.J. Lode, O.E. Alon, J. Arnold, A. Bhowmik, M. Büttner, L.S. Cederbaum,

B. Chatterjee, R. Chitra, S. Dutta, C. Georges, A. Hemmerich, H. Keßler, J. Klinder,
C. Lévêque, R. Lin, P. Molignini, F. Schäfer, J. Schmiedmayer and M. Žonda

A.U.J. Lode
Institute of Physics, Albert-Ludwigs-Universität Freiburg, Hermann-Herder-Str. 3, 79104 Freiburg,
Germany
e-mail: [email protected]
O.E. Alon
Department of Mathematics, University of Haifa, Haifa 3498838, Israel
Haifa Research Center for Theoretical Physics and Astrophysics, University of Haifa, Haifa 3498838,
Israel
e-mail: [email protected]
J. Arnold
Department of Physics, University of Basel, Klingelbergstrasse 82, CH-4056 Basel, Switzerland
A. Bhowmik
Department of Mathematics, University of Haifa, Haifa 3498838, Israel
Haifa Research Center for Theoretical Physics and Astrophysics, University of Haifa, Haifa 3498838,
Israel
M. Büttner
Institute of Physics, Albert-Ludwigs-Universität Freiburg, Hermann-Herder-Str. 3, 79104 Freiburg,
Germany
L.S. Cederbaum
Theoretische Chemie, Physikalisch-Chemisches Institut, Universität Heidelberg, Im Neuenheimer
Feld 229, D-69120 Heidelberg, Germany
e-mail: [email protected]
B. Chatterjee
Department of Physics, Indian Institute of Technology-Kanpur, Kanpur 208016, India
R. Chitra
Institute for Theoretical Physics, ETH Zürich, 8093 Zürich, Switzerland
S. Dutta
Department of Mathematics, University of Haifa, Haifa 3498838, Israel
Haifa Research Center for Theoretical Physics and Astrophysics, University of Haifa, Haifa 3498838,
Israel
C. Georges
The Hamburg Center for Ultrafast Imaging, Luruper Chaussee 149, 22761 Hamburg, Germany
A. Hemmerich
The Hamburg Center for Ultrafast Imaging, Luruper Chaussee 149, 22761 Hamburg, Germany
Zentrum für Optische Quantentechnologien and Institut für Laser-Physik, Universität Hamburg,
22761 Hamburg, Germany

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 35

W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_3
36 A.U.J. Lode et al.

Abstract This 2021 report summarizes our activities at the HLRS facilities Hawk and
Hazel Hen in the framework of the multiconfigurational time-dependent Hartree for
indistinguishable particles (MCTDH-X) high-performance computation project. Our
results are a bottom-up investigation into exciting and intriguing many-body physics
and phase diagrams obtained via the direct solution of the Schrödinger equation and
its comparison to experiments, and via machine learning approaches. We investigated
ultracold-boson quantum simulators for crystallization and superconductors in a
magnetic field, the phase transitions of ultracold bosons interacting with a cavity and
of charged fermions in lattices described by the Falicov–Kimball model. Moreover,
we report exciting findings on the many-body dynamics of tunneling and variances,
in two and three-dimensional ultracold-boson systems, respectively.

1 Introduction

Quantum many-body physics represents a demanding field of extraordinary depth

featuring an intriguing and physically rich set of phenomena. Here, we report our
activities exploring these phenomena within the framework of our multiconfigurational
time-dependent Hartree for indistinguishable particles high-performance computation
project at the HLRS. In our quest to explore many-body phenomena and in continuation
of our fruitful research in the past years [1–8] we developed and applied computational
methods to solve the Schrödinger equation numerically: we investigated ultracold-

H. Keßler
The Hamburg Center for Ultrafast Imaging, Luruper Chaussee 149, 22761 Hamburg, Germany
J. Klinder
The Hamburg Center for Ultrafast Imaging, Luruper Chaussee 149, 22761 Hamburg, Germany
C. Lévêque
Vienna Center for Quantum Science and Technology, Atominstitut, TU Wien, Stadionallee 2, 1020
Vienna, Austria
Wolfgang Pauli Institute c/o Faculty of Mathematics, University of Vienna, Oskar-Morgenstern Platz
1, 1090 Vienna, Austria
R. Lin
Institute for Theoretical Physics, ETH Zürich, 8093 Zürich, Switzerland
P. Molignini
Clarendon Laboratory, Department of Physics, University of Oxford, OX1 3PU, United Kingdom
Institute for Theoretical Physics, ETH Zürich, 8093 Zürich, Switzerland
F. Schäfer
Department of Physics, University of Basel, Klingelbergstrasse 82, CH-4056 Basel, Switzerland
J. Schmiedmayer
Vienna Center for Quantum Science and Technology, Atominstitut, TU Wien, Stadionallee 2, 1020
Vienna, Austria
M. Žonda
Department of Condensed Matter Physics, Charles University in Prague, Ke Karlovu 5, 121 16
Prague 2, Czech Republic
Quantum simulators, phase transitions, resonant tunneling, . . . 37

boson quantum simulators for i) crystallization [9] and ii) superconductors in a

magnetic field [10], phase transitions of iii) bosons in a cavity [11] and iv) mixtures
of heavy and light fermions in a lattice [12], v) many-boson dynamics of vi) two-
dimensional resonant tunneling [13] and vii) three-dimensional variance of position,
momentum, and angular momentum [14]. The works i) – vii) report new fundamental
insights into many-body-land and were obtained via software implementations [15–17]
of the MCTDH-X family of methods [18–25] as well as novel computational methods
for the simulation of the single-shot detection of quantum many-body systems [26–28]
and phase diagrams [12, 29].
We base this report on the findings in Refs. [9–14] and structure it as follows. In
Sec. 2, we provide an introduction to the MCTDH-X theory, in Sec. 3, we discuss the
observables we used in this work, and in Sections 4,5,6,7,8, and 9 we – respectively –
discuss the findings in Refs. [9–14]. We conclude with an outlook in Sec. 10.

2 Theory

Here we describe the theoretical method we chiefly used to solve the many-body
Schrödinger equation, MCTDH-X, the multiconfigurational time-dependent Hartree
method for indistinguishable particles [23–25].
MCTDH-X can be applied to a variety of systems – bosons [24, 25], fermions [19,
20], particles with internal degrees of freedom [30] or in a cavity [27]. The MCTDH-X
comes from the MCTDH [31] family of methods for indistinguishable particles which
also includes multilayering MCTDH [32, 33], restricted active space truncation [34–
37], and MCTDH-X approaches [38, 39]. MCTDH-X has been reviewed [8, 18] and
compared directly to experimental observations [11, 28].
The MCTDH-X method uses a time-dependent variational principle [40] applied to
the Schrödinger equation in combination with the following ansatz that represents the
state |Ψ⟩ as a time-dependent superposition of time-dependent symmetric (bosons)
or anti-symmetric (fermions) configurations:
h i 𝑛𝑗
∑︁ Ö𝑀 𝑏ˆ †𝑗 (𝑡)
|Ψ⟩ = 𝐶n (𝑡)|n; 𝑡⟩; |n; 𝑡⟩ = √︁ |𝑣𝑎𝑐⟩ = |𝑛1 , ..., 𝑛 𝑀 ; 𝑡⟩. (1)
n 𝑗=1 𝑛𝑗!

Here, n = (𝑛1 , ..., 𝑛 𝑀 ) is an occupation-number state with fixed particle number

Í
𝑁 = 𝑗 𝑛 𝑗 and {n} is a shorthand for the set of all possible such occupation-number
vectors. The coefficients 𝐶n (𝑡) are the complex-valued weights of the time-dependent
configurations |n; 𝑡⟩ at time 𝑡. These time-dependent configurations are built by
(anti-)symmetrizing a product of 𝑀 orthonormal time-dependent single-particle
states {𝜙1 (r; 𝑡), ..., 𝜙 𝑀 (r; 𝑡)}.
38 A.U.J. Lode et al.

In the variational derivation of the MCTDH-X working equations the coefficients

|n; 𝑡⟩ and the orbitals {𝜙1 (r; 𝑡), ..., 𝜙 𝑀 (r; 𝑡)} are considered as variational parameters.
The result of this variational derivation is a set of coupled non-linear integro-
differential MCTDH-X working equations [23–25], see the review [18] and References
therein for more details.

3 Quantities of interest

In this and the following sections, we will omit the dependency of quantities on time
where we consider it notationally convenient.

3.1 One- and two-body densities

The one-body density matrix 𝜌 (1) (the 1-RDM) is a standard probe for analyzing
𝑁-particle quantum states; it is defined as follows:

𝜌 (1) (r, r′) = N⟨𝚿| 𝚿

ˆ † (r′) 𝚿(r)|𝚿⟩.
ˆ (2)

Here, Ψ̂ (†) represents the field annihilation (creation) operator.

3.2 Fragmentation and condensation

The eigenvalues 𝜆 𝑗 and eigenfunctions 𝜙 (𝑗 𝑁 𝑂) of the 1-RDM [Eq. (2)], “(𝑁𝑂)”

being a shorthand for natural orbitals] determine the degree of fragmentation or
condensation of the state |Ψ⟩; we may rewrite
∑︁
𝜌 (1) (r, r′) = 𝜆j 𝜙j(NO),∗ (r′)𝜙j(NO) (r). (3)
j

If only a single eigenvalue and eigenfunction are significant, the state |Ψ⟩ represents
a coherent condensate [41–43]. If, in contrast to the condensed case, two or more
eigenvalues and eigenfunctions contribute to the 1-RDM, the state |Ψ⟩ represents a
fragmented condensate [44–46] that has lost some of its coherence [47].
Quantum simulators, phase transitions, resonant tunneling, . . . 39

3.3 Single-shot measurements

A single-shot measurement s is a projective measurement of a quantum state |Ψ⟩.

Single-shot measurements are random samples drawn from the probability |Ψ| 2 . For
ultracold atoms, single-shot measurements correspond to absorption images in real
or momentum space. Ideally, such a single-shot image,

s = (s1 , ..., sN ) , (4)

contains the positions or momenta s1 , ..., sN of all 𝑁 particles in the state |Ψ⟩.

3.3.1 Full distribution functions

Single-shot images allow for the extraction of the full distribution function 𝑃𝑛 (r) of
the particle number at position r. From a set of 𝑁 𝑠 single-shot images, {s1 , ..., sNs }
one can find 𝑃𝑛 (r) by determining the relative abundance of the detection of 𝑛
particles at position r by counting how many of the 𝑁 𝑠 single-shot images contain
exactly 𝑛 particles at position r. Doing this analysis for 𝑛 = 1, 𝑛 = 2, ..., 𝑛 = 𝑁 and
for all positions r yields the full distribution function 𝑃𝑛 (r).

4 Quantum simulation of crystallization

The quantum properties underlying crystal formation can be replicated and investigated
with the help of ultracold atoms. In this work [9], how the use of dipolar atoms
enables even the realization and precise measurement of structures that have not yet
been observed in any material.
Crystals are ubiquitous in nature. They are formed by many different materi-
als—from mineral salts to heavy metals like bismuth. Their structures emerge because
a particular regular ordering of atoms or molecules is favorable, because it requires
the smallest amount of energy. A cube with one constituent on each of its eight
corners, for instance, is a crystal structure that is very common in nature. A crystal’s
structure determines many of its physical properties, such as how well it conducts a
current or heat or how it cracks and behaves when it is illuminated by light. But what
determines these crystal structures? They emerge as a consequence of the quantum
properties of and the interactions between their constituents, which, however, are
often scientifically hard to understand and also hard measure.
To nevertheless get to the bottom of the quantum properties of the formation of
crystal structures, scientists can simulate the process using Bose–Einstein conden-
sates—trapped ultracold atoms cooled down to temperatures close to absolute zero or
minus 273.15 degrees Celsius. The atoms in these highly artificial and highly fragile
systems are extremely well under control.
40 A.U.J. Lode et al.

With careful tuning, the ultracold atoms behave exactly as if they were the
constituents forming a crystal. Although building and running such a quantum
simulator is a more demanding task than just growing a crystal from a certain material,
the method offers two main advantages: First, scientists can tune the properties for
the quantum simulator almost at will, which is not possible for conventional crystals.
Second, the standard readout of cold-atom quantum simulators are images containing
information about all crystal particles. For a conventional crystal, by contrast, only
the exterior is visible, while the interior—and in particular its quantum properties—is
difficult to observe.
In our work [9], we demonstrate that a flexible quantum simulator for crystal
formation can be built using ultracold dipolar quantum particles. Dipolar quantum
particles make it possible to realize and investigate not just conventional crystal
structures, but also arrangements and many-body physics that were hitherto not
seen for any material. The study explains how these crystal orders emerge from an
intriguing competition between kinetic, potential, and interaction energy and how the
structures and properties of the resulting crystals can be gauged in unprecedented
detail, see Fig. 1.

Fig. 1: Quantum simulation of crystallization with ultracold dipolar atoms: imbalance,

order parameter, and full distribution functions. Panels (a) and (b) demonstrate that
the superfluid/Mott-insulating (SMI), kinetic crystal (KCS) and density-wave crystal
(DWCS) states can be mapped out with ultracold dipolar bosons in lattices and
Í h i2
their crystal order parameter Δ = 𝑘 𝜆𝑁𝑘 and the imbalance of particles in even
Í
and odd sites, Θ = 𝑒,𝑜 ⟨𝑛𝑜 ⟩ − ⟨𝑛𝑒 ⟩. Alternatively, (c),(d),(e) show that the full
distribution functions 𝑃𝑛 (𝑥) with 𝑛 = 0, 1, 2 [cf. Sec. 3.3.1] can be used to identify
the SMI, KCS, and DWCS. Since Θ and 𝑃𝑛 (𝑥) are extracted from single-shot images,
our results demonstrate a viable experimental realization and readout of a tunable
dipolar-ultracold-atom quantum simulator of crystallization. Figure adapted from [9].
Quantum simulators, phase transitions, resonant tunneling, . . . 41

5 Dynamics of ultracold bosons in artificial gauge fields: Angular

momentum, fragmentation, and the variance of entropy

Fig. 2: Left: MCTDH-X simulations of one-body density [diagonal of Eq. (2)], natural
orbitals [cf. Eq. (3)], and natural orbital phases, 𝛽 𝑗 , for two regimes of artificial gauge
field strengths; weak in a)–i) and strong in j)–r) at time 𝑡 = 50.0. Right: We show
 2 ∫
the variance of image entropy 𝜎𝜁 = 𝑁1𝑠 𝜁 − 𝜁 𝑗 , where 𝜁 𝑗 = dr𝑠 𝑗 (r) ln 𝑠 𝑗 (r)
and 𝜁 = − 𝑁1𝑠 𝑁𝑗=1
Í 𝑠 𝑗
𝜁 as a function of propagation time and strength of the artificial
gauge field. The variance of the image entropy is a sensitive probe of the evolution of
fragmentation and angular momentum, see Ref. [10] for details and explanations on
𝜎𝜁 . Figure adapted from Ref. [10].

The charge neutrality inherent to ultracold trapped atoms introduces a constraint

to explore many-body physics in presence of an external magnetic field. Magnetism
in ultracold atoms can, however, effectively be realized via artificial gauge fields
[48]. These artificial gauge fields can make the neutral ultracold atoms behave as
if they were charged particles experiencing a magnetic field. In recent times, the
successful realizations of artificial gauge fields in, both, continuum [49–51] and
lattice potentials [52–54] have spearheaded novel ways for investigating the physics
of charged particles with neutral atoms.
In this work [10], we analyze the quantum dynamics of two-dimensional weakly
interacting ultracold bosons trapped in a harmonic potential triggered by an abrupt
quench of an artificial gauge field. For a quench to a sufficiently strong artificial gauge
field ghost vortices [55] at the edge of the density and phantom vortices [56] in the
natural orbitals, accompanied by topological defects in their phase do emerge. These
vortices and phase defects are hallmarks of the deposition of angular momentum in
the quantum state as time proceeds, see Fig. 2(Left). We find that the quench in the
artificial gauge field also triggers fragmentation and the emergence of non-trivial
many-body correlations together with the dynamics of angular momentum. To test
the possibility of the experimental detection of the emergent angular momentum
42 A.U.J. Lode et al.

and fragmentation in the BEC, we simulated single-shot images. We demonstrate

that it is feasible to detect, angular momentum, fragmentation, and correlations via
statistically analyzing the variance of the entropy of single-shot images, Fig. 2(Right).

6 Cavity-induced Mott transition in a Bose–Einstein condensate

At the absolute temperature of a thousandth of a Kelvin, all atoms in a cloud of

rubidium gas condense into exactly the same quantum state. This collection of ultracold
atoms is a new state of matter called a Bose–Einstein condensate (BEC) [57, 58]. In
recent years, technological advances in the preparation of BEC have turned them into
a versatile platform for the investigation of different and exotic physical phenomena
occurring at very low temperatures, where quantum physics dominates over thermal
noise. In one interesting scenario, the BEC is excited with an external driving laser,
and coupled to an optical cavity. This cavity can be realized as two mirrors facing
each other, where photons from the laser can bounce back and forth and interact
with the atoms. As the driving laser becomes stronger, the whole cavity-BEC system
undergoes two consecutive phase transitions [59, 60]. In the first one, the atoms
spontaneously self-organize into a lattice compatible with the profile generated by
the cavity light-field [59]. Immediately after the self-organization the atoms are in
coherence, or in other words, are still sharing one single quantum state as they do in
the BEC. As a result, all atoms can flow freely in the whole lattice system, showing
superfluidity. By pushing the laser intensity further, the atoms undergo the second
phase transition, where they localize even more into the lattice structure and lose
coherence [60], turning into a so-called Mott insulator state. Mott insulators are of
great interest to physicists because it is proposed to underpin many condensed matter
phenomena like high-temperature superconductivity [61,62], and could be applied for
innovative electronics [63] like Mott transition field effect transistor (Mott-FET) [64].
The realization with a BEC in a cavity offers an extremely clean and controlled
environment for the study of these phenomena. Although the Mott transition of an
optical lattice system has been studied extensively over the past decades [27,60,65–69],
it is still a difficult task to precisely determine its phase boundary. In the present
work [11], we propose a new scheme to predict this phase boundary using MCTDH-X
in a numerically exact fashion, and compare it quantitatively to experimental results.
This is mainly performed by tracking the experimental and simulated momentum
space density distributions 𝜌( 𝒌) [diagonal of the momentum-space version of Eq. (2)]
at various driving intensities, as presented in Fig. 3(a). Our main results are then
presented in Fig. 3(b), where we compare the experimental and simulated phase
diagrams consisting of the normal BEC phase (NP), the self-organized superfluid
phase (SSF) and the self-organized Mott insulating phase (SMI). For a large range
of parameter regimes, we see a remarkably good agreement in the SSF–SMI phase
boundary between the experiments and the MCTDH-X simulations. Our study thus
brings experimental realizations and theoretical predictions of Mott insulating physics
one step closer to each other.
Quantum simulators, phase transitions, resonant tunneling, . . . 43

Fig. 3: (a) The momentum density distributions 𝜌(𝑘) = ⟨Ψ𝑘† Ψ𝑘 ⟩/𝑁 from (first row)
experiments and (second row) MCTDH-X simulations for six different states, which
are measured or simulated at various driving intensities. (b) The phase diagram
delineating the normal BEC phase (NP), self-organized superfluid phase (SSF), and
self-organized Mott insulator phase (SMI) from experiments and simulations. The
brown crosses and the black circles are respectively the experimental NP–SSF and
SSF–SMI boundaries; whereas the black diamonds and the blue squares are the
experimental NP–SSF and SSF–SMI boundaries.

7 Interpretable and unsupervised phase classification

Phase diagrams and phase transitions are of paramount importance to physics [70–72].
While typical many-body systems have a large number of degrees of freedom,
their phases are usually characterized by a small set of physical quantities like
response functions or order parameters. For instance, the thermal phase transition
in the celebrated two-dimensional classical Ising model [73] is revealed by the
magnetization. However, in general the identification of phases and their order
parameters is a complex problem involving a large state space [74, 75]. Machine
learning methods are apt for this task [72, 76–85] as they can deal with large data
sets and efficiently extract information from them. Ideally, such machine learning
methods should not require prior knowledge about the phases, e.g., in the form of
samples that are labelled by their correct phase, or even the number of distinct phases.
That is, the methods should be unsupervised [77, 79, 86–103].
Yet, they should also allow for a straightforward physical insight into the character
of phases. That is, we desire interpretable methods [104,105] for which we understand
why they yield a given phase classification, i.e., whose decision making is fully
explainable. Significant progress in this direction has been made recently [98–103],
but some open issues remain regarding the interpretability of phase classification
methods. Many state-of-the-art phase classification methods rely on highly expressive
machine learning models, such as deep neural networks [106] (DNNs), for which it
is difficult to interpret the underlying functional dependence between their output
and the input data [100, 101]. That is, these models are black-boxes which allow for a
given phase classification task to be solved, but whose internal workings remain a
mystery to the user.
44 A.U.J. Lode et al.

In our contribution [12], we have proposed a novel interpretable and unsupervised

phase classification method referred to as the mean-based method. The method
requires input data that characterizes the state of a physical system for various choices
of system parameters in a suitable symmetry-adapted representation (see Fig. 4).
For each sampled point in parameter space, one calculates the magnitude of the
difference between mean input features of neighbouring points in parameter space.
This quantity serves as an indicator (label 𝐼 in Fig. 4) of phase transitions and reveals
the underlying phase diagram, because it is largest at phase boundaries. As such, the
mean-based method is conceptually simple, computationally cheap, and generic in
nature. In particular, the computation of its indicator does not rely on a black-box
predictive model and allows for this data-driven method to be directly explainable.

Fig. 4: Our workflow to predict a phase diagram with indicators 𝐼 for phase transitions.
Here, we illustrate the procedure for a two-dimensional parameter space: The
parameter space is sampled on a grid which yields a set of points { 𝒑 𝑖 } of fixed system
parameters. At each such point 𝒑 𝑖 a set of samples {𝑺𝑖
} is generated. Based on these
samples, a scalar indicator for phase transitions, 𝐼 𝒑 𝑖 , is calculated. This indicator
highlights the boundaries (red) between phases (grey). Different unsupervised phase
classification schemes are established via different indicators. Figure reprinted
from [12].

As a physical system, we considered the two-dimensional spinless Falicov–Kimball

model (FKM) [107–109]. This simple model of heavy and light fermions on a lattice
is used to address a broad range of contemporary physical problems [110–115] and
also serves as a standard test bed for the development of methods in the context of
strongly correlated systems [109, 116–121]. The FKM ground-state phase diagram
features a large number of different phases [122–124]. Figure 5(a) shows a sketch
of the expected phase diagram in two-dimensional parameter space. It highlights
the regions of stability of three main types of orderings, namely, (1) segregated, (2)
diagonal, and (3) axial orderings. Hitherto, the classification of ground-state phases
in the FKM was a manual and – due to the richness of the phase diagram [122–124] –
lengthy and cumbersome task. The complexity of the FKM phase diagram makes it a
challenging example for phase classification methods.
Quantum simulators, phase transitions, resonant tunneling, . . . 45

The probing of correlations is a standard procedure for studying phase transitions.

This motivates the analysis of the FKM phase diagram with the set of correlation
functions that measure square, axial, and diagonal order [Fig. 5(c)] as a suitable
choice of representation for the input data. The resulting phase diagram is shown in
Fig. 5(b). The mean-based method reproduces the main characteristics of the FKM
phase diagram [compare Fig. 5(a),(b)]. Moreover, we obtain a detailed, physical
subdivision of the phase diagram – see the identified orderings in Fig. 5 (top) and
labels (1)–(9) in Fig. 5(b). In particular, the method is able to distinguish between
dimer structures (7), different stable tile patterns (8),(9), or orderings exhibiting a
complicated phase separation (4).
These results demonstrate that the mean-based method with correlation functions
as input is an excellent tool to detect phase transitions. Moreover, the method
can straightforwardly be extended to other types of inputs, such as the Fourier-
transformed heavy-particle configurations, and higher-dimensional parameter spaces.
This establishes the mean-based method as a general phase classification method:
we infer that applications to arbitrary phase diagrams featuring, e.g., quantum or
topological phase transitions are feasible. Specifically, applications to quantum-
classical systems such as the FKM and its numerous generalizations [111, 116, 124–
126] are now straightforward.

8 Longitudinal and transversal many-boson resonant tunneling in

two-dimensional Josephson junctions

Going beyond our previous work on bosonic Josephson junctions [127–130], we

investigate the out-of-equilibrium quantum dynamics of a few interacting bosonic
clouds in a two-dimensional asymmetric double-well potential at the resonant-
tunneling scenarios [13]. Recall that at the single-particle level of resonant tunneling,
particles tunnel under the potential barrier from, typically, the ground-state in one
well to an excited state in the other well, that is, states of distinct shapes and properties
are coupled when their one-particle energies coincide. In two spatial dimensions, two
kinds of resonant tunneling processes become possible, to which we call longitudinal
and transversal resonant tunneling. Longitudinal resonant tunneling implies that the
state in, say, the right well is longitudinally excited with respect to the state in the
left well, whereas transversal resonant tunneling means that the state in the right
well is transversely excited with respect to the state in the left well. We demonstrate
that the inter-boson interaction makes resonant tunneling phenomena in two spatial
dimensions profoundly rich. To this end, we analyze resonant-tunneling phenomena in
terms of the loss of coherence of the junction and development of fragmentation, and
the coupling between transversal and longitudinal degrees-of-freedom and excitations.
In [13], a comprehensive analysis of the tunneling dynamics is performed by
exploring the time evolution of a few physical quantities, namely, the survival
probability, occupation numbers of the reduced one-particle density matrix, and the
many-particle position, momentum, and angular-momentum variances. To accurately
46 A.U.J. Lode et al.

Fig. 5: (a) Sketch of the ground-state phase diagram of the spinless FKM. A multitude
of other phases with smaller stability regions are expected to be present in the full
diagram [122, 123]. Red-dashed lines highlight the boundaries of the phases with
(1) segregated, (2) diagonal, and (3) axial orderings. For each ordering (1)–(3), an
example of a typical ground-state heavy-particle configuration 𝒘 0 on a square lattice
with linear size 𝐿 = 20 is shown on top. Here, the absence (𝑤 0,𝑖 = 0) and presence
(𝑤 0,𝑖 = 1) of a heavy particle at lattice site 𝑖 is denoted by a white or black square,
respectively. (b) Indicator for phase transitions obtained with the mean-based method
and correlation functions as input. Representative configurations (1)–(9) for some
of the largest inferred regions of stability (connected regions marked in blue by the
indicator), i.e., phases, are shown on top: These regions connect configurations of the
sq
same character. (c) Illustration of the correlation functions that measure square (𝜅 𝑛 ),
ax di
axial (𝜅 𝑛 ), and diagonal (𝜅 𝑛 ) correlation at a distance 𝑛 from the origin. Blue squares
denote the lattice sites marked by the corresponding stencil, where red denotes the
origin. Figure adapted from [12].

calculate these physical quantities from the time-dependent many-boson wavefunction,

we employ the MCTDHB method [18, 23–25, 30, 131, 132]. We use the numerical
implementation in [15, 16]. By comparing the survival probabilities and variances at
the many-body and mean-field levels of theory, and by following the development
of fragmentation, we unravel the detailed mechanisms of many-body longitudinal
as well as transversal resonant-tunneling processes in two-dimensional asymmetric
double-well potentials. In particular, we show that the position and momentum
variances along the transversal direction are almost negligible at the longitudinal
resonant-tunneling conditions, whereas they are considerable at the transversal
resonant-tunneling conditions which is administrated by the combination of density-
and breathing-mode oscillations. We demonstrate explicitly that the range of the inter-
boson interaction (width 𝜎 of the Gaussian interaction potential) does qualitatively not
affect neither the mean-field nor the many-body physics of two-dimensional resonant-
tunneling dynamics, see Fig. 6. All in all, we characterize the impact of the transversal
and longitudinal degrees-of-freedom in the many-boson tunneling dynamics at the
Quantum simulators, phase transitions, resonant tunneling, . . . 47

resonant-tunneling scenarios. The results are computed highly accurately in two

spatial dimensions at the many-body level of theory using MCTDHB, see Fig. 7 for a
selected few converged sensitive quantities taken from [13].

9 Morphology of a three-dimensional trapped Bose–Einstein

condensate from position, momentum, and angular-momentum
many-particle variances

The theory and properties of trapped BECs at the limit of an infinite number of
particles have attracted much interest [133–144]. In [14], we analyze the many-particle
position, momentum, and angular-momentum variances of a three-dimensional
anisotropic trapped BEC at the limit of an infinite number of particles, addressing
three-dimensional scenarios that have neither one-dimensional nor two-dimensional
analogs [138, 140, 145–150]. The variance of the position operator is associated with
the width of a wave-packet in position space, the variance of the momentum operator
is similarly related to the width of wave-packet in momentum space, and the variance
of the angular-momentum operator is correlated with how much a wave-packet is
non-spherically symmetric.
To this end, we compute the variances of the three Cartesian components of
the position, momentum, and angular-momentum operators of an interacting three-
dimensional trapped BEC at the limit of an infinite-particle-number limit, and
investigate their respective anisotropies [13]. We examine simple scenarios and show
that the anisotropy of a BEC can be different at the many-body and mean-field levels
of theory, although the BEC has identical many-body and mean-field densities per
particle. The analysis offers a geometrical-based picture to classify correlations via the
morphology of 100% condensed bosons in a three-dimensional trap at the limit of an
infinite number of particles. Fig. 8 presents results for the out-of-equilibrium quench
ˆ 𝑌ˆ , and 𝑍)
dynamics of the position ( 𝑋, ˆ and momentum (𝑃ˆ 𝑋 , 𝑃ˆ𝑌 , and 𝑃ˆ 𝑍 ) variances per
particle, and Fig. 9 depicts results for the ground-state angular-momentum ( 𝐿ˆ 𝑋 , 𝐿ˆ 𝑌 ,
and 𝐿ˆ 𝑍 ) variances per particle with and without spatial translations. The position and
momentum variances are given analytically within many-body theory and computed
numerically within mean-field theory, whereas the angular-momentum variances are
computed analytically both at the many-body and mean-field levels of theory using
the anisotropic three-dimensional harmonic-interaction model, see in this context
[151–166].

10 Conclusions and outlook

This 2021 report documents the substantial scientific activity in the framework of
the MCTDHB project spurred by the computational resources at the HLRS: our
high-performance computations resulted in significant contributions to the literature
48 A.U.J. Lode et al.

30 25 (b)
σ=0.25/√
π (a) σ=0.25/√
π
σ=0.25 σ=0.25
25
X position variance

X position variance
σ=0.25√π 20 σ=0.25√π
20
15
15
10
10
5 5

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
t/tRabi t/tRabi

1 30
σ=0.25/√
π (d)
σ=0.25
25

X position variance
0.8 σ=0.25√π
20
0.6
n1(t)/N

15
0.4
10
0.2 σ=0.25/√
π 5
σ=0.25
σ=0.25√π (c)
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
t/tRabi t/tRabi

1 25 (f)
σ=0.25/√
π
σ=0.25
X position variance

0.9 20 σ=0.25√π

0.8 15
n1(t)/N

0.7 10

0.6 σ=0.25/√
π 5
σ=0.25 (e)
σ=0.25√π
0.5 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
t/tRabi t/tRabi

Fig. 6: Time-dependent mean-field variance per particle of the 𝑋ˆ component of

the position operator for the ground state at the (a) longitudinal (𝑐 = 0.25) and (b)
transversal (𝜔 𝑛 = 0.19) resonant-tunneling scenarios.√ The three different widths√of
the inter-boson interaction potential are 𝜎 = 0.25 𝜋, 𝜎 = 0.25, and 𝜎 = 0.25/ 𝜋.
The interaction parameter is Λ = 0.01𝜋. The mean-field results are found to be
visibly independent of 𝜎. Many-body results for the longitudinal [panels (c) and (d)]
and transversal [panels (e) and (f)] dynamics. The dynamics are shown for the first
occupation number per particle 𝑛1𝑁(𝑡) [cf. 𝜆1 in Eq. (3)], see panels (c) and (e), and the
𝑋ˆ position variance per particle, see panels (d) and (f), respectively. The many-body
dynamics are computed with 𝑁 = 10 bosons, 𝑀 = 6 time-adaptive orbitals, and the
same interaction parameter. The quantitative many-body results are found to be weakly
dependent on the inter-boson interaction width 𝜎. All in all, we find that neither the
mean-field nor the many-body physics of the two-dimensional resonant-tunneling
dynamics qualitatively depend on 𝜎. The color codes are explained in each panel.
See [13] for more details and the definition of the variances shown in (b),(d),(f). The
quantities shown are dimensionless. Figure panels reprinted from [13].
Quantum simulators, phase transitions, resonant tunneling, . . . 49

Angular−momentum variance
1 7 ΨG, M=6, 64 X 64
ΨG, M=6, 128 X 128
(b)
6
0.8
5
0.6
n1(t)/N

4
0.4 c=0, M=6
3
c=0, M=10
c=0.25, M=6 2
0.2 c=0.25, M=10
c=0.5, M=6 1
c=0.5, M=10 (a)
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
t/tRabi t/tRabi

Angular−momentum variance
18 12
Ψ , M=6 (c) ΨG, M=6, 128 X 128 (d)
16 Ψ G ΨG, M=6, 256 X 256
G, M=10 10
Y position variance

14
12 8
10
6
8
6 4
4
2
2
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
t/tRabi t/tRabi

Fig. 7: Convergence of time-dependent many-body quantities with computational

resources for the initial state Ψ𝐺 for the longitudinally-asymmetric [panels (a) and
(b)] and transversely-asymmetric [panels (c) and (d)] two-dimensional double-well
potentials at the many-particle resonant-tunneling conditions. The bosonic cloud
consists of 𝑁 = 10 bosons with the interaction parameter Λ = 0.01𝜋. (a) Convergence
of the first occupation number per particle 𝑛1𝑁(𝑡) [cf. 𝜆1 in Eq. (3)] with the number of
time-dependent orbitals. The asymmetry parameters are 𝑐 = 0, 𝑐 = 0.25, and 𝑐 = 0.5.
The convergence is verified with 𝑀 = 6 and 𝑀 = 10 time-adaptive orbitals. (b) The
variance per particle of the 𝐿ˆ 𝑍 component of the angular-momentum operator with
the number of grid points. The asymmetry parameter is 𝑐 = 0.25. The convergence
is verified with 64 × 64 and 128 × 128 grid points. (c) The variance per particle of
the 𝑌ˆ component of the position operator with the number of time-adaptive orbitals.
The frequency is 𝜔 𝑛 = 0.19. (d) The variance per particle of the 𝐿ˆ 𝑍 component of
the angular-momentum operator with the number of grid points. The frequency is
𝜔 𝑛 = 0.19. The convergence is verified with 128 × 128 and 256 × 256 grid points. The
color codes are explained in each panel. See [13] for more details and the definition
of the variances shown in (b)–(d). The quantities shown are dimensionless. Figure
panels reprinted from [13].
50 A.U.J. Lode et al.

0.7 0.7
(a) (b)

0.6 0.6

Momentum variance
Position variance

0.5 0.5

0.4 0.4
Var X (GP)
Var Y (GP)
Var Z (GP) g=0.18 Var PX (GP) Var PX (MB)
Var X (MB) Var PY (GP) Var PY (MB)
Var Y (MB)
Var Z (MB)
Var PZ (GP) Var PZ (MB) g=0.18
0.3 0.3
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

1.8 1.1
Var X (GP) Var PX (GP) Var PX (MB)
(c) Var Y (GP) g=27.0 (d) Var PY (GP) Var PY (MB) g=27.0
Var Z (GP) Var PZ (GP) Var PZ (MB)
Var X (MB)
1.5 Var Y (MB) 0.9
Var Z (MB)
Momentum variance
Position variance

1.2 0.7

0.9 0.5

0.6 0.3

0.3
0.1
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

Fig. 8: Many-particle position [ 𝑋, ˆ 𝑌ˆ , and 𝑍;

ˆ panels (a) and (c) in red, green, and
blue] and momentum [𝑃ˆ 𝑋 , 𝑃ˆ𝑌 , and 𝑃ˆ 𝑍 ; panels (b) and (d) in red, green, and blue]
variances per particle as a function of time computed at the limit of an infinite number
of particles within many-body (dashed lines) and mean-field (solid lines) levels of
theory in an interaction-quench scenario. The harmonic trap is 20% anisotropic.
The coupling constant 𝑔 in the Gross–Pitaevskii equation is indicated in each panel.
Different anisotropy classes of the position variances and of the momentum variances
emerge with time. See [14] for more details and the definition of the variances shown
in (a)–(d). Figure panels reprinted from [14].

on many-body physics [9–14] as well as the implementation and distribution of an

increasing amount of computational software [15–17, 29, 167] and datasets [168, 169].
We unveiled exciting fundamental many-body physics and novel computational tools
and thus underpin the versatility and prosperity of the MCTDHB project at the
HLRS. We look forward to continue our quest to compute the quantum many-body
properties for ever larger systems and envision large-scale computations with, both,
large configuration spaces and large spatial two- and three-dimensional spatial grids.
Quantum simulators, phase transitions, resonant tunneling, . . . 51

1.0 1.0
(a) Var LX (GP) (b) Var LX (GP)
Var LY (GP) Var LY (GP)
Var LZ (GP) Var LZ (GP)
Var LX (MB) Var LX (MB)
Angular-momentum variance

Angular-momentum variance
Var LY (MB) Var LY (MB)
Var LZ (MB) Var LZ (MB)
0.75 0.75

0.5 0.5

0.25 0.25

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Interaction parameter Interaction parameter

1.0 1.0
(c) Var LX (GP) (d) Var LX (GP)
Var LY (GP) Var LY (GP)
Var LZ (GP) Var LZ (GP)
Var LX (MB) Var LX (MB)
Angular-momentum variance

Angular-momentum variance
Var LY (MB) Var LY (MB)
Var LZ (MB) Var LZ (MB)
0.75 0.75

0.5 0.5

0.25 0.25

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Interaction parameter Interaction parameter

Fig. 9: Many-particle angular-momentum ( 𝐿ˆ 𝑋 , 𝐿ˆ 𝑌 , and 𝐿ˆ 𝑍 ; in red, green, and blue)

variances per particle as a function of the interaction parameter Λ computed at
the limit of an infinite number of particles within many-body (dashed lines) and
mean-field (solid lines) levels of theory for the ground state of the three-dimensional
anisotropic harmonic-interaction model. The frequencies of the trap are 𝜔 𝑥 = 0.7,
𝜔 𝑦 = 5.0, and 𝜔 𝑧 = 10.5. Results for several translations r0 of the center of the trap are
shown in the panels: (a) r0 = (0, 0, 0); (b) r0 = (0.25, 0, 0); (c) r0 = (0, 0.25, 0); (d)
r0 = (0.25, 0.25, 0). Different anisotropy classes of the angular-momentum variances
emerge with the interaction parameter. See [14] for more details and the definition
of the variances shown in (a)–(d). The quantities shown are dimensionless. Figure
panels reprinted from [14].

Acknowledgements Financial support by the Deutsche Forschungsgemeinschaft (DFG) is gratefully

acknowledged. OEA acknowledges funding by the Israel Science Foundation (Grants No. 600/15
and No. 1516/19). We acknowledge financial support by the Austrian Science Foundation (FWF)
under grants P32033 and M2653. We acknowledge financial support by the Swiss National Science
Foundation (SNSF), the NCCR Quantum Science and Technology, and Mr. G. Anderheggen.
Computation time on the HPC2013 cluster of the IIT Kanpur and on the HazelHen and Hawk
clusters at the HLRS Stuttgart is gratefully acknowledged.
52 A.U.J. Lode et al.

References

1. A.U.J. Lode, K. Sakmann, R.A. Doganov, J. Grond, O.E. Alon, A.I. Streltsov, L.S. Cederbaum,
in High Perform. Comput. Sci. Eng. ’13 Trans. High Perform. Comput. Center, Stuttgart 2013
(Springer International Publishing, 2013), pp. 81–92. DOI 10.1007/978-3-319-02165-2_7
2. S. Klaiman, A.U.J. Lode, K. Sakmann, O.I. Streltsova, O.E. Alon, L.S. Cederbaum, A.I.
Streltsov, in High Perform. Comput. Sci. Eng. ’14 Trans. High Perform. Comput. Center,
Stuttgart 2014 (Springer International Publishing, 2015), pp. 63–86. DOI 10.1007/978-3-319
-10810-0_5
3. O.E. Alon, V.S. Bagnato, R. Beinke, I. Brouzos, T. Calarco, T. Caneva, L.S. Cederbaum, M.A.
Kasevich, S. Klaiman, A.U.J. Lode, S. Montangero, A. Negretti, R.S. Said, K. Sakmann,
O.I. Streltsova, M. Theisen, M.C. Tsatsos, S.E. Weiner, T. Wells, A.I. Streltsov, in High
Perform. Comput. Sci. Eng. ’15 Trans. High Perform. Comput. Center, Stuttgart 2015 (Springer
International Publishing, 2016), pp. 23–49. DOI 10.1007/978-3-319-24633-8_3
4. O.E. Alon, R. Beinke, L.S. Cederbaum, M.J. Edmonds, E. Fasshauer, M.A. Kasevich,
S. Klaiman, A.U.J. Lode, N.G. Parker, K. Sakmann, M.C. Tsatsos, A.I. Streltsov, in High
Perform. Comput. Sci. Eng. ’16 Trans. High Perform. Comput. Cent. Stuttgart 2016 (Springer
International Publishing, 2017), pp. 79–96. DOI 10.1007/978-3-319-47066-5_6
5. O.E. Alon, R. Beinke, C. Bruder, L.S. Cederbaum, S. Klaiman, A.U.J. Lode, K. Sakmann,
M. Theisen, M.C. Tsatsos, S.E. Weiner, A.I. Streltsov, in High Perform. Comput. Sci. Eng.
17 Trans. High Perform. Comput. Center, Stuttgart 2017 (Springer International Publishing,
2018), pp. 93–115. DOI 10.1007/978-3-319-68394-2_6
6. O.E. Alon, V.S. Bagnato, R. Beinke, S. Basu, L.S. Cederbaum, B. Chakrabarti, B. Chatterjee,
R. Chitra, F.S. Diorico, S. Dutta, L. Exl, A. Gammal, S.K. Haldar, S. Klaiman, C. Lévêque,
R. Lin, N.J. Mauser, P. Molignini, L. Papariello, R. Roy, K. Sakmann, A.I. Streltsov, G.D.
Telles, M.C. Tsatsos, R. Wu, A.U.J. Lode, in High Perform. Comput. Sci. Eng. ’ 18 (Springer
International Publishing, 2019), pp. 89–110. DOI 10.1007/978-3-030-13325-2_6
7. A.U.J. Lode, O.E. Alon, L.S. Cederbaum, B. Chakrabarti, B. Chatterjee, R. Chitra, A. Gammal,
S.K. Haldar, M.L. Lekala, C. Lévêque, R. Lin, P. Molignini, L. Papariello, M.C. Tsatsos, in
High Perform. Comput. Sci. Eng. ’ 19, ed. by W.E. Nagel, D.H. Kröner, M.M. Resch (Springer
International Publishing, Cham, 2021), pp. 77–87. DOI 10.1007/978-3-030-66792-4. URL
https://link.springer.com/10.1007/978-3-030-66792-4
8. A. Lode, A. U. J. , Alon, O. E., Bastarrachea-Magnani, M. A. , Bhowmik, d.F. de Buchleitner,
A., Cederbaum, L. S., Chitra, R., Fasshauer, E., S.E. Parny, L., Haldar, S. K., Lévêque, C., Lin,
R., Madsen, L. B., Molignini, P., Papariello, L., Schäfer, F., Strelstov, A. I., Tsatsos, M. C.,
and Weiner, in High Perform. Comput. Sci. Eng. ’20 Trans. High Perform. Comput. Center,
Stuttgart 2020 (2022), p. (in press)
9. B. Chatterjee, C. Lévêque, J. Schmiedmayer, A.U.J. Lode, Phys. Rev. Lett. 125(9), 093602
(2020). DOI 10.1103/PhysRevLett.125.093602. URL https://link.aps.org/doi/10.
1103/PhysRevLett.125.093602
10. A.U.J. Lode, S. Dutta, C. Lévêque, Entropy 23(4), 392 (2021). DOI 10.3390/e23040392.
URL https://www.mdpi.com/1099-4300/23/4/392
11. R. Lin, C. Georges, J. Klinder, P. Molignini, M. Büttner, A.U.J. Lode, R. Chitra, A. Hemmerich,
H. Keßler, SciPost Phys. 11, 30 (2021). DOI 10.21468/SciPostPhys.11.2.030. URL
https://scipost.org/10.21468/SciPostPhys.11.2.030
12. J. Arnold, F. Schäfer, M. Žonda, A.U.J. Lode, Phys. Rev. Research 3, 033052 (2021). DOI
10.1103/PhysRevResearch.3.033052. URL https://link.aps.org/doi/10.1103/Phy
sRevResearch.3.033052
13. A. Bhowmik, O.E. Alon, (2021). URL http://arxiv.org/abs/2101.04959
14. O.E. Alon, Symmetry 13(7) (2021). DOI 10.3390/sym13071237. URL https://www.mdpi
.com/2073-8994/13/7/1237
15. R. Streltsov, A. I., Cederbaum, L. S., Alon, O. E., Sakmann, K., Lode, A. U. J., Grond, J.,
Streltsova, O. I., Klaiman, S., Beinke. The Multiconfigurational Time-Dependent Hartree
for Bosons Package, version 3.x, http://mctdhb.org, Heidelberg/Kassel (2006-Present). URL
http://mctdhb.org
Quantum simulators, phase transitions, resonant tunneling, . . . 53

16. O.I. Streltsov, A. I., Streltsova. The multiconfigurational time-dependent Hartree for bosons
laboratory, version 1.5, http://MCTDHB-lab.com. URL http://mctdhb-lab.com
17. A.U.J. Lode, M.C. Tsatsos, E. Fasshauer, R. Lin, L. Papariello, P. Molignini, S.E. Weiner,
C. Lévêque. MCTDH-X: The multiconfigurational time-dependent Hartree for indistinguish-
able particles software, http://ultracold.org (2020). URL http://ultracold.org
18. A.U.J. Lode, C. Lévêque, L.B. Madsen, A.I. Streltsov, O.E. Alon, Rev. Mod. Phys. 92, 011001
(2020). DOI 10.1103/RevModPhys.92.011001. URL https://link.aps.org/doi/10.11
03/RevModPhys.92.011001
19. R. Lin, P. Molignini, L. Papariello, M.C. Tsatsos, C. Lévêque, S.E. Weiner, E. Fasshauer,
R. Chitra, A.U.J. Lode, Quantum Sci. Technol. 5, 024004 (2020). DOI 10.1088/2058-9565/ab
788b. URL https://iopscience.iop.org/article/10.1088/2058-9565/ab788b
20. E. Fasshauer, A.U.J. Lode, Phys. Rev. A 93(3), 033635 (2016). DOI 10.1103/PhysRevA.93.03
3635. URL https://link.aps.org/doi/10.1103/PhysRevA.93.033635
21. A.U.J. Lode, C. Bruder, Phys. Rev. A 94(1), 013616 (2016). DOI 10.1103/PhysRevA.94.013616
22. A.U.J. Lode, Phys. Rev. A 93(6), 063601 (2016). DOI 10.1103/PhysRevA.93.063601. URL
https://link.aps.org/doi/10.1103/PhysRevA.93.063601
23. O.E. Alon, A.I. Streltsov, L.S. Cederbaum, J. Chem. Phys. 127(15), 154103 (2007). DOI
10.1063/1.2771159
24. A.I. Streltsov, O.E. Alon, L.S. Cederbaum, Phys. Rev. Lett. 99(3), 030402 (2007). DOI
10.1103/PhysRevLett.99.030402
25. O.E. Alon, A.I. Streltsov, L.S. Cederbaum, Phys. Rev. A 77(3), 033613 (2008). DOI
10.1103/PhysRevA.77.033613
26. K. Sakmann, M. Kasevich, Nat. Phys. 12(5), 451 (2016). DOI 10.1038/nphys3631. URL
http://www.nature.com/articles/nphys3631
27. A.U.J. Lode, C. Bruder, Phys. Rev. Lett. 118(1), 013603 (2017). DOI 10.1103/PhysRevLett.
118.013603
28. J.H.V. Nguyen, M.C. Tsatsos, D. Luo, A.U.J. Lode, G.D. Telles, V.S. Bagnato, R.G. Hulet,
Phys. Rev. X 9(1), 011052 (2019). DOI 10.1103/PhysRevX.9.011052. URL https:
//link.aps.org/doi/10.1103/PhysRevX.9.011052
29. Arnold, Julian, Schäfer, Frank, Zonda, Martin, Lode, A. U. J. Interpretable-and-unsupervised-
phase-classification. URL Interpretable-and-unsupervised-phase-classificat
ion
30. A.U.J. Lode, Phys. Rev. A 93(6), 063601 (2016). DOI 10.1103/PhysRevA.93.063601. URL
https://link.aps.org/doi/10.1103/PhysRevA.93.063601
31. M.H. Beck, A. Jäckle, G.A. Worth, H.D. Meyer, Phys. Rep. 324(1), 1 (2000). DOI
10.1016/S0370-1573(99)00047-2
32. U. Manthe, J. Phys.: Condens. Matter 29(25), 253001 (2017). DOI 10.1088/1361-648X/aa6e96.
URL https://iopscience.iop.org/article/10.1088/1361-648X/aa6e96
33. H. Wang, M. Thoss, J. Chem. Phys. 119, 1289 (2003). DOI 10.1063/1.1580111. URL
http://aip.scitation.org/doi/10.1063/1.1580111
34. H. Miyagi, L.B. Madsen, Phys. Rev. A 87(6), 062511 (2013). DOI 10.1103/PhysRevA.87.06
2511. URL https://link.aps.org/doi/10.1103/PhysRevA.87.062511
35. H. Miyagi, L.B. Madsen, Phys. Rev. A 89(6), 063416 (2014). DOI 10.1103/PhysRevA.89.06
3416. URL https://link.aps.org/doi/10.1103/PhysRevA.89.063416
36. C. Lévêque, L.B. Madsen, New J. Phys. 19, 043007 (2017). DOI 10.1088/1367-2630/aa6319.
URL http://stacks.iop.org/1367-2630/19/i=4/a=043007
37. C. Lévêque, L.B. Madsen, J. Phys. B: At., Mol. Opt. Phys. 51, 155302 (2018). DOI
10.1088/1361-6455/aacac6. URL https://iopscience.iop.org/article/10.1088/1
361-6455/aacac6/pdf
38. L. Cao, S. Krönke, O. Vendrell, P. Schmelcher, J. Chem. Phys. 139(13), 134103 (2013). DOI
10.1063/1.4821350
39. L. Cao, V. Bolsinger, S.I. Mistakidis, G.M. Koutentakis, S. Krönke, J.M. Schurer, P. Schmelcher,
J. Chem. Phys. 147(4), 044106 (2017). DOI 10.1063/1.4993512. URL http://aip.scitat
ion.org/doi/10.1063/1.4993512
54 A.U.J. Lode et al.

40. P. Kramer, M. Saraceno, Geometry of the time-dependent variational principle in quantum

mechanics (Springer, Lecture Notes in Physics, 2007). DOI 10.1007/3-540-10271-x_317
41. O. Penrose, L. Onsager, Phys. Rev. 104(3), 576 (1956). DOI 10.1103/PhysRev.104.576
42. E.P. Gross, Nuovo Cim. 20(3), 454 (1961). DOI 10.1007/BF02731494. URL http:
//link.springer.com/10.1007/BF02731494
43. L.P. Pitaevskii, Sov. Phys. JETP 13(2), 451 (1961). URL http://jetp.ac.ru/cgi-bin/d
n/e{_}013{_}02{_}0451.pdf
44. E.J. Mueller, T.L. Ho, M. Ueda, G. Baym, Phys. Rev. A 74, 033612 (2006)
45. P. Nozieres, D. St. James, J. Phys. Paris 43(7), 1133 (1982). DOI 10.1051/jphys:019820043070
113300. URL http://www.edpsciences.org/10.1051/jphys:019820043070113300
46. R.W. Spekkens, J.E. Sipe, Phys. Rev. A 59(5), 3868 (1999). DOI 10.1103/PhysRevA.59.3868
47. K. Sakmann, A.I. Streltsov, O.E. Alon, L.S. Cederbaum, Phys. Rev. A 78(2), 023615 (2008).
DOI 10.1103/PhysRevA.78.023615
48. J. Dalibard, F. Gerbier, G. Juzeliūnas, P. Öhberg, Rev. Mod. Phys. 83, 1523 (2011). DOI
10.1103/RevModPhys.83.1523. URL https://link.aps.org/doi/10.1103/RevModP
hys.83.1523
49. I.B. Spielman, Phys. Rev. A 79(6), 063613 (2009). DOI 10.1103/PhysRevA.79.063613. URL
https://link.aps.org/doi/10.1103/PhysRevA.79.063613
50. Y.J. Lin, R.L. Compton, K. Jiménez-García, J.V. Porto, I.B. Spielman, Nature 462(7273), 628
(2009). DOI 10.1038/nature08609. URL http://www.nature.com/articles/nature08
609
51. Y.J. Lin, R. Compton, A. Perry, W. Phillips, J. Porto, I. Spielman, Phys. Rev. Lett. 102(13),
130401 (2009). DOI 10.1103/PhysRevLett.102.130401. URL https://link.aps.org/d
oi/10.1103/PhysRevLett.102.130401
52. M. Aidelsburger, Artificial Gauge Fields with Ultracold Atoms in Optical Lattices. Springer
Theses (Springer International Publishing, Cham, 2016). DOI 10.1007/978-3-319-25829-4.
URL http://link.springer.com/10.1007/978-3-319-25829-4
53. M. Aidelsburger, M. Atala, S. Nascimbène, S. Trotzky, Y.A. Chen, I. Bloch, Phys. Rev.
Lett. 107(25), 255301 (2011). DOI 10.1103/PhysRevLett.107.255301. URL https:
//link.aps.org/doi/10.1103/PhysRevLett.107.255301
54. L.K. Lim, C.M. Smith, A. Hemmerich, Phys. Rev. Lett. 100(13), 130402 (2008). DOI
10.1103/PhysRevLett.100.130402. URL https://link.aps.org/doi/10.1103/PhysR
evLett.100.130402
55. M. Tsubota, K. Kasamatsu, M. Ueda, Phys. Rev. A 65, 023603 (2002). DOI 10.1103/PhysRe
vA.65.023603. URL https://link.aps.org/doi/10.1103/PhysRevA.65.023603
56. S.E. Weiner, M.C. Tsatsos, L.S. Cederbaum, A.U.J. Lode, Sci. Reports 7, 40122 (2017). DOI
10.1038/srep40122
57. M.H. Anderson, J.R. Ensher, M.R. Matthews, C.E. Wieman, E.A. Cornell, Collect. Pap. Carl
Wieman 269, 453 (2008). DOI 10.1142/9789812813787_0062
58. K.B. Davis, M.O. Mewes, M.R. Andrews, N.J. Van Druten, D.S. Durfee, D.M. Kurn, W. Ketterle,
Phys. Rev. Lett. 75(22), 3969 (1995). DOI 10.1103/PhysRevLett.75.3969
59. K. Baumann, C. Guerlin, F. Brennecke, T. Esslinger, Nature 464, 1301 (2010). DOI
10.1038/nature09009. URL http://www.nature.com/articles/nature09009
60. J. Klinder, H. Keßler, M.R. Bakhtiari, M. Thorwart, A. Hemmerich, Phys. Rev. Lett. 115(23),
230403 (2015). DOI 10.1103/PhysRevLett.115.230403. URL https://link.aps.org/d
oi/10.1103/PhysRevLett.115.230403
61. P.A. Lee, N. Nagaosa, X.G. Wen, Rev. Mod. Phys. 78, 17 (2006)
62. T. Yanagisawa, M. Miyazaki, EPL (Europhysics Lett. 107, 27004 (2014). DOI 10.1209/0295
-5075/107/27004. URL https://iopscience.iop.org/article/10.1209/0295-507
5/107/27004
63. You Zhou, S. Ramanathan, Proc. IEEE 103, 1289 (2015). DOI 10.1109/JPROC.2015.2431914.
URL http://ieeexplore.ieee.org/document/7137616/
64. D. Newns, T. Doderer, C. Tsuei, W. Donath, J. Misewich, A. Gupta, B. Grossman, A. Schrott,
B. Scott, P. Pattnaik, R. von Gutfeld, J. Sun, J. Electroceramics 4, 339 (2000). DOI 10.1023/A:
1009914609532. URL http://link.springer.com/10.1023/A:1009914609532
Quantum simulators, phase transitions, resonant tunneling, . . . 55

65. M.R. Bakhtiari, A. Hemmerich, H. Ritsch, M. Thorwart, Phys. Rev. Lett. 114, 123601 (2015).
DOI 10.1103/PhysRevLett.114.123601. URL https://link.aps.org/doi/10.1103/P
hysRevLett.114.123601
66. P. Molignini, L. Papariello, A.U.J. Lode, R. Chitra, Phys. Rev. A 98, 053620 (2018). DOI
10.1103/PhysRevA.98.053620. URL https://link.aps.org/doi/10.1103/PhysRevA.
98.053620
67. A.U.J. Lode, F.S. Diorico, R. Wu, P. Molignini, L. Papariello, R. Lin, C. Lévêque, L. Exl,
M.C. Tsatsos, R. Chitra, N.J. Mauser, New J. Phys. 20(5), 055006 (2018). DOI 10.1088/1367
-2630/aabc3a. URL https://doi.org/10.1088/1367-2630/aabc3a
68. R. Lin, L. Papariello, P. Molignini, R. Chitra, A.U.J. Lode, Phys. Rev. A 100(1), 013611
(2019). DOI 10.1103/PhysRevA.100.013611. URL https://link.aps.org/doi/10.11
03/PhysRevA.100.013611
69. R. Lin, P. Molignini, A.U.J. Lode, R. Chitra, Phys. Rev. A 101, 061602 (2020). DOI
10.1103/PhysRevA.101.061602. URL https://link.aps.org/doi/10.1103/PhysRev
A.101.061602
70. S. Sachdev, Quantum Phase Transitions (Cambridge University Press, Cambridge, 2011).
DOI 10.1017/CBO9780511973765. URL https://doi.org/10.1017{%}2Fcbo978051
1973765
71. N. Goldenfeld, Lectures on Phase Transitions and the Renormalization Group (CRC Press,
2018). DOI 10.1201/9780429493492. URL https://www.taylorfrancis.com/books/
9780429962042
72. G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, L. Vogt-Maranto, L. Zde-
borová, Rev. Mod. Phys. 91(4), 045002 (2019). DOI 10.1103/RevModPhys.91.045002. URL
https://link.aps.org/doi/10.1103/RevModPhys.91.045002
73. L. Onsager, Phys. Rev. 65(3-4), 117 (1944). DOI 10.1103/PhysRev.65.117. URL https:
//link.aps.org/doi/10.1103/PhysRev.65.117
74. J.P. Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity (Oxford
University Press, 2021). DOI 10.1093/oso/9780198865247.001.0001. URL https:
//oxford.universitypressscholarship.com/view/10.1093/oso/9780198865247
.001.0001/oso-9780198865247
75. P.M. Chaikin, T.C. Lubensky, Principles of Condensed Matter Physics (Cambridge University
Press, 1995). DOI 10.1017/CBO9780511813467. URL https://www.cambridge.org/co
re/product/identifier/9780511813467/type/book
76. J. Carrasquilla, R.G. Melko, Nat. Phys. 13(5), 431 (2017). DOI 10.1038/nphys4035. URL
http://www.nature.com/articles/nphys4035
77. E.P.L. van Nieuwenburg, Y.H. Liu, S.D. Huber, Nat. Phys. 13(5), 435 (2017). DOI 10.1038/
nphys4037. URL http://www.nature.com/articles/nphys4037
78. K. Ch’ng, J. Carrasquilla, R.G. Melko, E. Khatami, Phys. Rev. X 7(3), 031038 (2017). DOI
10.1103/PhysRevX.7.031038. URL https://link.aps.org/doi/10.1103/PhysRevX.
7.031038
79. L. Wang, Phys. Rev. B 94(19), 195105 (2016). DOI 10.1103/PhysRevB.94.195105. URL
https://link.aps.org/doi/10.1103/PhysRevB.94.195105
80. B.S. Rem, N. Käming, M. Tarnowski, L. Asteria, N. Fläschner, C. Becker, K. Sengstock,
C. Weitenberg, Nat. Phys. 15(9), 917 (2019). DOI 10.1038/s41567-019-0554-0. URL
http://www.nature.com/articles/s41567-019-0554-0
81. A. Bohrdt, C.S. Chiu, G. Ji, M. Xu, D. Greif, M. Greiner, E. Demler, F. Grusdt, M. Knap, Nat.
Phys. 15(9), 921 (2019). DOI 10.1038/s41567-019-0565-x. URL http://www.nature.c
om/articles/s41567-019-0565-x
82. V. Dunjko, H.J. Briegel, Rep. Prog. Phys. 81(7), 074001 (2018). DOI 10.1088/1361-6633/aa
b406. URL https://iopscience.iop.org/article/10.1088/1361-6633/aab406
83. T. Ohtsuki, T. Ohtsuki, J. Phys. Soc. Jpn. 86(4), 044708 (2017). DOI 10.7566/JPSJ.86.044708.
URL https://journals.jps.jp/doi/10.7566/JPSJ.86.044708
84. J. Carrasquilla, Adv. Phys. X 5(1), 1797528 (2020). DOI 10.1080/23746149.2020.1797528.
URL https://www.tandfonline.com/doi/full/10.1080/23746149.2020.1797528
56 A.U.J. Lode et al.

85. A. Bohrdt, S. Kim, A. Lukin, M. Rispoli, R. Schittko, M. Knap, M. Greiner, J. Léonard, (2020).
URL http://arxiv.org/abs/2012.11586
86. S.J. Wetzel, Phys. Rev. E 96(2), 022140 (2017). DOI 10.1103/PhysRevE.96.022140. URL
https://link.aps.org/doi/10.1103/PhysRevE.96.022140
87. Y.H. Liu, E.P.L. van Nieuwenburg, Phys. Rev. Lett. 120(17), 176401 (2018). DOI 10.1103/Ph
ysRevLett.120.176401. URL https://link.aps.org/doi/10.1103/PhysRevLett.120
.176401
88. P. Huembeli, A. Dauphin, P. Wittek, Phys. Rev. B 97(13), 134109 (2018). DOI 10.1103/Phys
RevB.97.134109. URL https://link.aps.org/doi/10.1103/PhysRevB.97.134109
89. J.F. Rodriguez-Nieva, M.S. Scheurer, Nat. Phys. 15(8), 790 (2019). DOI 10.1038/s41567-019
-0512-x. URL http://www.nature.com/articles/s41567-019-0512-x
90. K. Liu, J. Greitemann, L. Pollet, Phys. Rev. B 99(10), 104410 (2019). DOI 10.1103/PhysRe
vB.99.104410. URL https://link.aps.org/doi/10.1103/PhysRevB.99.104410
91. F. Schäfer, N. Lörch, Phys. Rev. E 99(6), 062107 (2019). DOI 10.1103/PhysRevE.99.062107.
URL https://link.aps.org/doi/10.1103/PhysRevE.99.062107
92. E. Greplova, A. Valenti, G. Boschung, F. Schäfer, N. Lörch, S.D. Huber, New J. Phys. 22(4),
045003 (2020). DOI 10.1088/1367-2630/ab7771. URL https://iopscience.iop.org/a
rticle/10.1088/1367-2630/ab7771
93. Y. Che, C. Gneiting, T. Liu, F. Nori, Phys. Rev. B 102(13), 134213 (2020). DOI 10.1103/PhysRe
vB.102.134213. URL https://link.aps.org/doi/10.1103/PhysRevB.102.134213
94. M.S. Scheurer, R.J. Slager, Phys. Rev. Lett. 124(22), 226401 (2020). DOI 10.1103/PhysRevLet
t.124.226401. URL https://link.aps.org/doi/10.1103/PhysRevLett.124.226401
95. O. Balabanov, M. Granath, Phys. Rev. Res. 2(1), 013354 (2020). DOI 10.1103/PhysRevResea
rch.2.013354. URL https://link.aps.org/doi/10.1103/PhysRevResearch.2.013
354
96. Y. Long, J. Ren, H. Chen, Phys. Rev. Lett. 124(18), 185501 (2020). DOI 10.1103/PhysRevLet
t.124.185501. URL https://link.aps.org/doi/10.1103/PhysRevLett.124.185501
97. N. Käming, A. Dawid, K. Kottmann, M. Lewenstein, K. Sengstock, A. Dauphin, C. Weitenberg,
Mach. Learn. Sci. Technol. (2021). DOI 10.1088/2632- 2153/abffe7. URL https:
//iopscience.iop.org/article/10.1088/2632-2153/abffe7
98. C. Casert, T. Vieijra, J. Nys, J. Ryckebusch, Phys. Rev. E 99(2), 023304 (2019). DOI
10.1103/PhysRevE.99.023304. URL https://link.aps.org/doi/10.1103/PhysRevE.
99.023304
99. S. Blücher, L. Kades, J.M. Pawlowski, N. Strodthoff, J.M. Urban, Phys. Rev. D 101(9), 094507
(2020). DOI 10.1103/PhysRevD.101.094507. URL https://link.aps.org/doi/10.11
03/PhysRevD.101.094507
100. Y. Zhang, P. Ginsparg, E.A. Kim, Phys. Rev. Res. 2(2), 023283 (2020). DOI 10.1103/PhysRe
vResearch.2.023283. URL https://link.aps.org/doi/10.1103/PhysRevResearch
.2.023283
101. A. Dawid, P. Huembeli, M. Tomza, M. Lewenstein, A. Dauphin, New J. Phys. 22(11), 115001
(2020). DOI 10.1088/1367-2630/abc463. URL https://iopscience.iop.org/article
/10.1088/1367-2630/abc463
102. A. Cole, G.J. Loges, G. Shiu, (2020). DOI https://arxiv.org/abs/2009.14231. URL
http://arxiv.org/abs/2009.14231
103. N. Rao, K. Liu, M. Machaczek, L. Pollet, (2021). URL https://arxiv.org/abs/2102.0
1103http://arxiv.org/abs/2102.01103
104. R. Guidotti, A. Monreale, S. Ruggieri, F. Turini, F. Giannotti, D. Pedreschi, ACM Comput.
Surv. 51(5), 1 (2019). DOI 10.1145/3236009. URL https://dl.acm.org/doi/10.1145
/3236009
105. C. Molnar, Interpretable Machine Learning. URL https://christophm.github.io/int
erpretable-ml-book/
106. I. Goodfellow, Y. Bengio, A. Courville, Deep Learning (MIT Press, 2016). URL http:
//www.deeplearningbook.org
Quantum simulators, phase transitions, resonant tunneling, . . . 57

107. L.M. Falicov, J.C. Kimball, Phys. Rev. Lett. 22(19), 997 (1969). DOI 10.1103/PhysRevLet
t.22.997. URL https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.
22.997
108. J. Hubbard, Proc. R. Soc. London. Ser. A. Math. Phys. Sci. 276(1365), 238 (1963). DOI
10.1098/rspa.1963.0204. URL https://royalsocietypublishing.org/doi/10.1098
/rspa.1963.0204
109. J.K. Freericks, V. Zlatić, Rev. Mod. Phys. 75(4), 1333 (2003). DOI 10.1103/RevModPhys.75.
1333. URL https://link.aps.org/doi/10.1103/RevModPhys.75.1333
110. M. Hohenadler, F.F. Assaad, Phys. Rev. Lett. 121(8), 086601 (2018). DOI 10.1103/PhysRevLet
t.121.086601. URL https://link.aps.org/doi/10.1103/PhysRevLett.121.086601
111. M. Gonçalves, P. Ribeiro, R. Mondaini, E.V. Castro, Phys. Rev. Lett. 122(12), 126601 (2019).
DOI 10.1103/PhysRevLett.122.126601. URL https://link.aps.org/doi/10.1103/P
hysRevLett.122.126601
112. M. Eckstein, M. Kollar, Phys. Rev. Lett. 100(12), 120404 (2008). DOI 10.1103/PhysRevLett.
100.120404. URL https://link.aps.org/doi/10.1103/PhysRevLett.100.120404
113. M.M. Oliveira, P. Ribeiro, S. Kirchner, Phys. Rev. Lett. 122(19), 197601 (2019). DOI
10.1103/PhysRevLett.122.197601. URL https://link.aps.org/doi/10.1103/PhysR
evLett.122.197601
114. C. Prosko, S.P. Lee, J. Maciejko, Phys. Rev. B 96(20), 205104 (2017). DOI 10.1103/PhysRe
vB.96.205104. URL https://link.aps.org/doi/10.1103/PhysRevB.96.205104
115. A. Kauch, P. Pudleiner, K. Astleithner, P. Thunström, T. Ribic, K. Held, Phys. Rev. Lett. 124(4),
047401 (2020). DOI 10.1103/PhysRevLett.124.047401. URL https://doi.org/10.110
3/PhysRevLett.124.047401https://link.aps.org/doi/10.1103/PhysRevLett.1
24.047401
116. J.K. Freericks, V.M. Turkowski, V. Zlatić, Phys. Rev. Lett. 97(26), 266408 (2006). DOI
10.1103/PhysRevLett.97.266408. URL https://link.aps.org/doi/10.1103/PhysRev
Lett.97.266408
117. H. Aoki, N. Tsuji, M. Eckstein, M. Kollar, T. Oka, P. Werner, Rev. Mod. Phys. 86(2), 779
(2014). DOI 10.1103/RevModPhys.86.779. URL https://link.aps.org/doi/10.1103
/RevModPhys.86.779
118. T. Maier, M. Jarrell, T. Pruschke, M.H. Hettler, Rev. Mod. Phys. 77(3), 1027 (2005). DOI
10.1103/RevModPhys.77.1027. URL https://link.aps.org/doi/10.1103/RevModP
hys.77.1027
119. V. Turkowski, J.K. Freericks, Phys. Rev. B 75(12), 125110 (2007). DOI 10.1103/PhysRevB.7
5.125110. URL https://link.aps.org/doi/10.1103/PhysRevB.75.125110
120. J. Kaye, D. Golez, SciPost Phys. 10(4), 091 (2021). DOI 10.21468/SciPostPhys.10.4.091.
URL https://scipost.org/10.21468/SciPostPhys.10.4.091
121. L. Huang, L. Wang, Phys. Rev. B 95(3), 035105 (2017). DOI 10.1103/PhysRevB.95.035105.
URL https://link.aps.org/doi/10.1103/PhysRevB.95.035105
122. R. Lemański, J.K. Freericks, G. Banach, Phys. Rev. Lett. 89(19), 196403 (2002). DOI
10.1103/PhysRevLett.89.196403. URL https://link.aps.org/doi/10.1103/PhysRev
Lett.89.196403
123. R. Lemański, J.K. Freericks, G. Banach, J. Stat. Phys. 116(1-4), 699 (2004). DOI 10.1023/B:
JOSS.0000037213.25834.33. URL http://link.springer.com/10.1023/B:JOSS.000
0037213.25834.33
124. Čenčariková, Farkašovský, Condens. Matter Phys. 14(4), 42701 (2011). DOI 10.5488/CMP.
14.42701. URL http://www.icmp.lviv.ua/journal/zbirnyk.68/42701/abstract
.html
125. M.D. Petrović, B.S. Popescu, U. Bajpai, P. Plecháč, B.K. Nikolić, Phys. Rev. Appl. 10(5),
054038 (2018). DOI 10.1103/PhysRevApplied.10.054038. URL https://link.aps.org
/doi/10.1103/PhysRevApplied.10.054038
126. X.H. Li, Z. Chen, T.K. Ng, Phys. Rev. B 100(9), 094519 (2019). DOI 10.1103/PhysRevB.100
.094519. URL https://link.aps.org/doi/10.1103/PhysRevB.100.094519
58 A.U.J. Lode et al.

127. K. Sakmann, A.I. Streltsov, O.E. Alon, L.S. Cederbaum, Phys. Rev. Lett. 103(22), 220601
(2009). DOI 10.1103/PhysRevLett.103.220601. URL https://link.aps.org/doi/10.
1103/PhysRevLett.103.220601
128. K. Sakmann, A.I. Streltsov, O.E. Alon, L.S. Cederbaum, Phys. Rev. A 89(2), 023602 (2014).
DOI 10.1103/PhysRevA.89.023602. URL https://link.aps.org/doi/10.1103/PhysR
evA.89.023602
129. S.K. Haldar, O.E. Alon, New J. Phys. 21(10), 103037 (2019). DOI 10.1088/1367-2630/ab4315.
URL https://iopscience.iop.org/article/10.1088/1367-2630/ab4315
130. A. Bhowmik, S.K. Haldar, O.E. Alon, Sci. Reports 10, 21476 (2020). DOI 10.1038/s41598-0
20-78173-w. URL http://www.nature.com/articles/s41598-020-78173-w
131. A.I. Streltsov, O.E. Alon, L.S. Cederbaum, Phys. Rev. A 73(6), 063626 (2006). DOI
10.1103/PhysRevA.73.063626
132. A.U.J. Lode, K. Sakmann, O.E. Alon, L.S. Cederbaum, A.I. Streltsov, Phys. Rev. A 86(6),
063606 (2012). DOI 10.1103/PhysRevA.86.063606
133. Y. Castin, R. Dum, Low-temperature Bose-Einstein condensates in time-dependent traps:
Beyond the U(1) symmetry-breaking approach, vol. 57 (American Physical Society, 1998).
DOI 10.1103/PhysRevA.57.3008. URL https://link.aps.org/doi/10.1103/PhysRev
A.57.3008
134. E.H. Lieb, R. Seiringer, J. Yngvason, Phys. Rev. A 61(4), 043602 (2000). DOI 10.1103/Phys
RevA.61.043602. URL https://link.aps.org/doi/10.1103/PhysRevA.61.043602
135. E.H. Lieb, R. Seiringer, Phys. Rev. Lett. 88(17), 170409 (2002). DOI 10.1103/PhysRevLett.
88.170409. URL https://link.aps.org/doi/10.1103/PhysRevLett.88.170409
136. L. Erdős, B. Schlein, H.T. Yau, Invent. Math. 167(3), 515 (2007). DOI 10.1007/s00222-006-0
022-1. URL http://link.springer.com/10.1007/s00222-006-0022-1
137. L. Erdős, B. Schlein, H.T. Yau, Phys. Rev. Lett. 98(4), 359 (2007). DOI 10.1103/PhysRevLett.
98.040404. URL https://link.aps.org/doi/10.1103/PhysRevLett.98.040404
138. S. Klaiman, O.E. Alon, Phys. Rev. A 91(6), 063613 (2015). DOI 10.1103/PhysRevA.91.063613.
URL https://link.aps.org/doi/10.1103/PhysRevA.91.063613
139. S. Klaiman, L.S. Cederbaum, Phys. Rev. A 94(6), 063648 (2016). DOI 10.1103/PhysRevA.94.
063648. URL https://link.aps.org/doi/10.1103/PhysRevA.94.063648
140. S. Klaiman, A.I. Streltsov, O.E. Alon, Uncertainty product of an out-of-equilibrium many-
particle system, vol. 93 (American Physical Society, 2016). DOI 10.1103/PhysRevA.93.023605.
URL https://link.aps.org/doi/10.1103/PhysRevA.93.023605
141. I. Anapolitanos, M. Hott, D. Hundertmark, Rev. Math. Phys. 29, 1750022 (2017). DOI
10.1142/S0129055X17500222. URL https://www.worldscientific.com/doi/abs/10
.1142/S0129055X17500222
142. A. Michelangeli, A. Olgiati, Anal. Math. Phys. 7, 377 (2017). DOI 10.1007/s13324-016-0147-3.
URL http://link.springer.com/10.1007/s13324-016-0147-3
143. O.E. Alon, J. Phys. A Math. Theor. 50, 295002 (2017). DOI 10.1088/1751-8121/aa78ad.
URL https://iopscience.iop.org/article/10.1088/1751-8121/aa78ad
144. L.S. Cederbaum, Phys. Rev. A 96, 013615 (2017). DOI 10.1103/PhysRevA.96.013615. URL
http://link.aps.org/doi/10.1103/PhysRevA.96.013615
145. S. Klaiman, R. Beinke, L.S. Cederbaum, A.I. Streltsov, O.E. Alon, Chem. Phys. 509, 45 (2018).
DOI 10.1016/j.chemphys.2018.02.016. URL https://www.sciencedirect.com/scienc
e/article/abs/pii/S0301010417307668?via{%}3Dihub
146. K. Sakmann, J. Schmiedmayer, (2018). URL http://arxiv.org/abs/1802.03746
147. O.E. Alon, L.S. Cederbaum, Chem. Phys. 515, 287 (2018). DOI 10.1016/j.chemphys.2018.09
.029. URL https://www.sciencedirect.com/science/article/pii/S030101041
8307183?via{%}3Dihub{#}b0210
148. O.E. Alon, Mol. Phys. 117(15-16), 2108 (2019). DOI 10.1080/00268976.2019.1587533. URL
https://www.tandfonline.com/doi/full/10.1080/00268976.2019.1587533
149. O.E. Alon, J. Phys. Conf. Ser. 1206, 012009 (2019). DOI 10.1088/1742-6596/1206/1/012009.
URL https://iopscience.iop.org/article/10.1088/1742-6596/1206/1/012009
150. O.E. Alon, Symmetry 11, 1344 (2019). DOI 10.3390/sym11111344. URL https:
//www.mdpi.com/2073-8994/11/11/1344
Quantum simulators, phase transitions, resonant tunneling, . . . 59

151. P.D. Robinson, J. Chem. Phys. 66, 3307 (1977). DOI 10.1063/1.434310. URL http:
//aip.scitation.org/doi/10.1063/1.434310
152. R.L. Hall, J. Phys. A: Math. Gen. 11, 1235 (1978). DOI 10.1088/0305-4470/11/7/011. URL
https://iopscience.iop.org/article/10.1088/0305-4470/11/7/011/meta
153. L. Cohen, C. Lee, J. Math. Phys. 26(12), 3105 (1985). DOI 10.1063/1.526688
154. M.S. Osadchii, V.V. Murakhtanov, Int. J. Quantum Chem. 39, 173 (1991). DOI 10.1002/qua.
560390207. URL http://doi.wiley.com/10.1002/qua.560390207
155. M.A. Załuska-Kotur, M. Gajda, A. Orłowski, J. Mostowski, Phys. Rev. A 61, 8 (2000). DOI
10.1103/PhysRevA.61.033613. URL https://journals.aps.org/pra/abstract/10.1
103/PhysRevA.61.033613
156. J. Yan, J. Stat. Phys. 113, 623 (2003). DOI 10.1023/A:1026029104217
157. M. Gajda, Criterion for Bose-Einstein condensation in a harmonic trap in the case with
attractive interactions, vol. 73 (American Physical Society, 2006). DOI 10.1103/PhysRevA.7
3.023603. URL https://link.aps.org/doi/10.1103/PhysRevA.73.023603
158. J.R. Armstrong, N.T. Zinner, D.V. Fedorov, A.S. Jensen, J. Phys. B: At., Mol. Opt. Phys. 44(5),
055303 (2011). DOI 10.1088/0953-4075/44/5/055303. URL http://stacks.iop.org/095
3-4075/44/i=5/a=055303?key=crossref.aa534c8a7543acdd895681648ff1992e
159. J.R. Armstrong, N.T. Zinner, D.V. Fedorov, A.S. Jensen, Phys. Rev. E 86, 021115 (2012). DOI
10.1103/PhysRevE.86.021115. URL https://journals.aps.org/pre/abstract/10.1
103/PhysRevE.86.021115
160. C. Schilling, Phys. Rev. A 88, 042105 (2013). DOI 10.1103/PhysRevA.88.042105. URL
https://journals.aps.org/pra/abstract/10.1103/PhysRevA.88.042105
161. C.L. Benavides-Riveros, I.V. Toranzo, J.S. Dehesa, J. Phys. B: At., Mol. Opt. Phys. 47, 195503
(2014). DOI 10.1088/0953-4075/47/19/195503. URL https://iopscience.iop.org/a
rticle/10.1088/0953-4075/47/19/195503
162. P.A. Bouvrie, A.P. Majtey, M.C. Tichy, J.S. Dehesa, A.R. Plastino, Eur. Phys. J. D 68, 1 (2014).
DOI 10.1140/epjd/e2014-50349-2. URL https://link.springer.com/article/10.1
140/epjd/e2014-50349-2
163. J.R. Armstrong, A.G. Volosniev, D.V. Fedorov, A.S. Jensen, N.T. Zinner, J. Phys. A Math.
Theor. 48(8), 085301 (2015). DOI 10.1088/1751- 8113/48/8/085301. URL https:
//iopscience.iop.org/article/10.1088/1751-8113/48/8/085301
164. C. Schilling, R. Schilling, Phys. Rev. A 93, 021601 (2016). DOI 10.1103/PhysRevA.93.021601.
URL https://journals.aps.org/pra/abstract/10.1103/PhysRevA.93.021601
165. S. Klaiman, A.I. Streltsov, O.E. Alon, Chem. Phys. 482, 362 (2017). DOI 10.1016/j.chemphys
.2016.07.011
166. S. Klaiman, A.I. Streltsov, O.E. Alon, J. Phys. Conf. Ser. 999, 12013 (2018). DOI 10.1088/17
42-6596/999/1/012013. URL https://iopscience.iop.org/article/10.1088/174
2-6596/999/1/012013
167. A.U.J. Lode, P. Molignini, R. Lin, M. Büttner, P. Rembold, C. Lévêque, M.C. Tsatsos,
L. Papariello. UNIQORN:Universal Neural-network Interface for Quantum Observable
Readout from {$N$}-body wavefunctions, https://gitlab.com/UNIQORN/uniqorn (2021).
URL https://gitlab.com/UNIQORN/uniqorn
168. A.U.J. Lode, R. Lin, M. Büttner, P. Rembold, C. Lévêque, M.C. Tsatsos, P. Molignini,
L. Papariello. Uniqorn data set (2021). URL https://drive.google.com/file/d/1Du8
KRhsITezlMVWEBrLOnDIAfFZOPcEj/view?usp=sharing
169. A.U.J. Lode, R. Lin, M. Büttner, P. Rembold, C. Lévêque, M.C. Tsatsos, P. Molignini,
L. Papariello. Uniqorn triple-well data set (2021). URL https://drive.google.com/fil
e/d/1Zqc8wyzeqWrna-7uMJ9XI{_}WFreQBJzDu/view?usp=sharing
Molecules, Interfaces and Solids

In this funding period, the field of molecules, interfaces and solids benefits enormously
from the computational resources provided by the High Performance Computing
Center Stuttgart and the Steinbuch Centre for Computing Karlsruhe. In what follows,
we selected some projects in this area to demonstrate the impact of high performance
computing in physics, chemistry, and material science.
The collaborative work by Oelschläger, Klein, Müller and Roth from the Institute
of Functional Matter and Quantum Technology and the Graduate School of Excellence
advance Manufacturing Engineering at the University Stuttgart is an outstanding
example in this respect. The project comprises numerical simulations of selective
laser melting with a focus on the additive manufacturing of products by a printer. The
main challenge here is to close the gap between manageable system sizes and the
industrial scales which still differ by about two orders of magnitude. Therefore the
authors perform a proof of principle study to validate the applicability of atomistic
molecular dynamic (MD) simulations (based on the classical Newton equation of
motions with interactions modelled by embedded-atom potentials) to the problem of
powder bed fusion using a laser beam, and demonstrate that all components of the
model work quite well. For the time being a mechanism for pore formation could be
identified. Furthermore, it could be shown that lower laser velocities favour droplet
formation which may cause balling and splashing. Further studies are definitely
worthwhile, for example, in order to clarify the effects of different packing densities,
of recrystallisation, or of a variable floor.
Another MD study is devoted to the influence of solutes on the tensile behaviour of
polyamide6. The simulations performed by Verestek and Schmauder from the IMWF
University of Stuttgart mainly address the mechanical properties of the dry as well as
the saturated (with water, methanol or ethanol) bulk material. Here, a clear trend with
increased weakening is observed from dry to the solutes, where the impact increases
when passing over from water to methanol and ethanol. Thereby the weakening is not
caused by the growing distance between the amide groups. Tensile test simulations in
the triaxial state show that an increased crystallinity comes from the reorganisation
of fibril-like structures. Moreover differences in the Young’s modulus are observed,
where methanol has the highest reduction.

61
62 Molecules, Interfaces and Solids

The Theoretical Physics and Materials Research groups at the Justus-Liebig

University Gießen led by S. Sanna investigated the lattice dynamics of a Si(553)-Au
surface from first principles by means of a DFT-based atomistic simulation using
the double-Au strand model and the rehybridised model. In particular, a theoretical
characterisation of the surface-localised phonon-mode is given both above and below
the phase transition temperature. Thereby the comparison with the experiments
allows to assign the spectral features to the calculated displacement pattern. The
strong coupling between the electronic and phononic systems observable in Raman
scattering could be related to the phonon activated charge transfer between the Au
chain and the Si step edge. Modifying the Si(553)-Au nanowire system by atomic
deposition or doping would be an interesting direction of future studies.
Chemical functionalisation of semiconductor surfaces, that is, combining estab-
lished silicon-based technology with the structural richness of organic chemistry, is a
promising route to widen the application range of semiconducting devices. This is
why the group of Ralf Tonner at the university of Leipzig investigates the reactivity
of organic molecules on semiconductor surfaces, using density functional theory as
implemented in the VASP (Vienna Ab Initio Simulation Package) code, which makes
very efficient use of HLRS’ computational resources. The contribution investigates
how pyrazine, methyl benzylazide and a cyclooctyne derivative adsorb on and react
with a silicon or germanium surface. Furthermore, computational protocols for even
more efficient treatments have been explored, namely how far one can reduce a
model system that still leads to meaningful results. Reduced model systems allow for
more accurate quantum chemical methods beyond density functional theory, so the
reduction of the model is especially important if there is more than one organic layer
on top of the silicon surface.
The transformation of our industry from fossil to renewable energy sources
will require efficient conversion of solar/wind to chemical energy. Splitting water
into hydrogen and oxygen by electric energy will be highly important, since „green”
hydrogen is the basis for all sorts of industrial processes and heavy-load transportation.
Although the oxygen produced during water splitting is a side product disposed into
the atmosphere, oxygen formation is a complicated process and thus the bottleneck
for large-scale water electrolysis as well as related processes like carbon dioxide
or nitrogen reduction. Phenomenologically this means that one needs significant
over-potentials to make oxygen formation sufficiently fast. Among the best catalysts for
dioxygen formation are iridium dioxide based materials. While iridium is admittedly a
too precious metal to be used massively on an industrial scale, an understanding why
it is one of the best catalysts for oxygen formation will possibly produce new ideas
for catalysts based on earth-abundant elements. Travis Jones from the Berlin-based
Fritz Haber institute of the Max Planck society presents large-scale simulations
of dioxygen formation on IrO2 surfaces, which is a multi-step and multi-electron
reaction at an electrochemical interface. As in the previous contribution, density
functional methods are being used that can properly describe the bond-breaking
and bond-making that occur in chemical reactions. Here, the Quantum ESPRESSO
package is used, that also can make good used of the huge amounts of CPUs offered
by the HLRS computer. One of the important results is that the oxygen formation
Molecules, Interfaces and Solids 63

rate is determined by the amount of surface oxidation, and that the “kink” (or bend)
observed experimentally in the logarithmic Tafel plot (reaction rate as a function of
the applied potential) arises from the response of the surface oxidation to the potential
rather than from a change in the reaction mechanism. This means the “chemistry”
(oxygen-oxygen bond formation) as such does not require a high over-potential, which
is instead used to produce the necessary amount of surface oxidation.
We wish to stress that almost all projects supported in this field are of high
scientific quality, also those which could not be included in report because of space
limitations. This underlines once again the strong need of supercomputing facilities
in modern condensed matter based science and technology.

Institut für Physik, Holger Fehske

Lehrstuhl Komplexe Quantensysteme,
Ernst-Moritz-Arndt-Universität Greifswald,
Felix-Hausdorff-Str. 6,
17489 Greifswald,
Germany,
e-mail: [email protected]

Fachbereich Chemie, Christoph van Wüllen

Technische Universität Kaiserslautern,
Erwin-Schrödinger-Str. 52,
67663 Kaiserslautern,
Germany,
e-mail: [email protected]
Molecular dynamics simulation of selective laser
melting

Fabio Oelschläger, Dominic Klein, Sarah Müller and Johannes Roth

Abstract This report deals with the atomistic simulation of selective laser melting
(SLM) used to produce additive manufactured objects. After a short introduction into
the subject, modifications to basic molecular dynamics simulation code are described
which are required to simulate the annealing process. Although the sample sizes
studied in this report are already impressively large, scaling of system parameters
are required to connect simulation and experiment. First results will be reported and
further developments and improvements be described.

1 Introduction

Additive manufacturing (AM) of products with a printer plays an increasing role today.
AM means adding layers of material in the manufacturing process while traditionally
material is removed from a workpiece by subtractive technologies such as turning
or milling. Though there are many different methods for AM, all work according
to a similar basic principle: a three-dimensional digital model is dissected by a
slicer into layers which are subsequently produced and stacked. In contrast to public

Fabio Oelschläger
Institut für Funktionelle Materie und Quantentechnologien, Universität Stuttgart,
e-mail: [email protected]
Dominic Klein
Institut für Funktionelle Materie und Quantentechnologien, Universität Stuttgart,
e-mail: [email protected]
Sarah Müller
Graduate School of Excellence advanced Manufacturing Engineering, GSaME, Universität Stuttgart,
e-mail: [email protected]
Johannes Roth
Institut für Funktionelle Materie und Quantentechnologien, Universität Stuttgart,
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 65

W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_4
66 F. Oelschläger, D. Klein, S. Müller and J. Roth

belief where simple plastic parts are printed on commodity printers with no special
requirements for precision and stability, we are dealing here with industrial printing
of load bearing, metallic power parts. Here the additive manufacturing is far away
from working perfectly and being competitive with other methods. Deviation of size
and defects are frequent. To understand the defects and failures on an atomistic level,
large scale molecular dynamics (MD) simulations on supercomputers are carried out
with the simulation code IMD[14] which is well suited for this purpose and has been
demonstrated to run effectively with billions of atoms (See Sec. 5). For the simulation
of additive manufacturing presented here only a few modifications have to be made,
for example the setup of the moving laser beam and the addition of gravitational
force.
A main challenge is the gap between manageable simulation sizes and industrial
scales which typically differ by one to two orders of magnitude in the current case.
In principle the supercomputers are large enough to reach industrial scales but such
simulations would require the whole machine for weeks together with sufficient
resources for storage, analysis, and visualization. For the time being we have to resort
to running smaller simulations and scaling the results suitably.
In the present proof of principle study we demonstrate that all components of the
model work quite well. First quantitative results for selective laser melting (SLM) of
a single sphere of aluminum are also given.
The report is organized as follows: after an introduction into additive manufacturing
by SLM we discuss the problems occurring. Then we present the first results of the
simulations. The report ends with performance data collect for our simulations on
HLRS Hawk.

2 Additive manufacturing by selective laser melting (SLM)

2.1 Description of the method

The method studied in this work is powder bed fusion using a laser beam, often
referred to as Selective Laser melting (SLM, SLM Solutions Group). Fig. 1 shows a
printer working according to this principle. The printer consists of two containers
with movable bases. A feed container is filled with powder material while the other
printing container is more or less empty. The base of the feed container is lifted up if
a new layer is to be printed. A drum shifts the dosed material to the printing container
while the base of the latter is lowered. Then, a mobile laser scans the desired printing
plain and fuses the powder into a solid shape[9]. The printed chamber is typically
lowered by 30 𝜇m to 100 𝜇m [18] but thicker layers are also common [17]. The
most common materials applied in the SLM method are Ti compounds, especially
Ti6Al4V in aerospace industry [1, 17, 18]. Al compounds are used in automobile
industry since they are suitable in light-way construction [20, 22].
Selective laser melting 67

Fig. 1: Schematic presentation of the SLM method. The mobile mirror takes care for
the precise positioning of the laser spot. The base of the feed container (left) is lifted
while the base of the printing container (center) is lowered.

2.1.1 Defects in manufactured objects

Several defect types can occur during printing: Pores are gas inclusions with sizes
less that 100 𝜇m[21]. The source may be evaporation of the low-melting metal or
protective gas used to avoid oxidation. Without gas a low packing density can be the
origin of the pores. The pores are spherical and equidistributed in the sample. The
size and shape of the pores has considerable influence on the quality and especially
the density of the product.
Lack-of-fusion (LOF) originates predominantly through insufficient energy supply
during melting. The width of the molten lane is too small which causes bonding
defects and powder inclusions and the different layers are not fused. Consequently
LOF occurs predominantly along the lanes and between the layers. The defects
increase roughness and disturb the matter flow for the next layer. Thus, they can
gradually propagate through several layers. The problem may also be caused by
surface cooling which reduces the wettability[21].
After melting, cooling rates reach 108 K/s [21]. The temperature gradient and the
related strong thermal expansion gradient lead to high stress in the material. Cracks
are created and propagate through the material. Cracks originate at the surface and
propagate into the bulk. Non-spherical pores could also generate cracks[6].
68 F. Oelschläger, D. Klein, S. Müller and J. Roth

2.1.2 Important parameters and influence of the laser velocity

Many parameters influence the quality of the product in SLM (Fig. 2). Adjustable
parameters (Fig. 3) which have great influence on quality are scanning speed, laser
power, hatch distance or spacing, laser wavelength, spot size, and layer thickness[15].
Since up to now the studied objects are single powder particles, hatch distance and
layer thickness will no longer be taken into account. The laser wavelength plays only
an implicit role through its connection to power and reflectivity.

Fig. 2: An overview of the parameters controlling the quality of SLM.

Following the defects discussed in Sec.2.1.1 it is important to melt the powder

completely but to avoid supplying too much energy. An optimal matching is required
for an optimal outcome. Higher laser velocities for example often require increased
laser power in order to guarantee the melting of the material. Sadali et al[15] observe
that the pore sizes and the splashing effect are reduced with increasing speed. Spherical
material agglomeration at the surface are caused by the separation of instable liquid
traces through insufficient wetting (balling effect). They get fewer and shrink with
increasing velocity. Nonetheless, a faster laser velocity does not automatically mean
a better result. The number of microcracks for example increase. While these can
still be minimized by Hot Isostatic Pressing (HIP), further undesirable effects occur
which cannot be removed in post processing[15].
Selective laser melting 69

Fig. 3: The most important SLM parameters and the typical orders of magnitude [16].

2.2 Modeling of SLM by molecular dynamics Simulations

In molecular dynamics simulations (MD) the classical Newtonian equations of

motions are integrated at discrete time steps, typically fractions of femto-seconds for
atomistic systems. Interactions in metals are modeled by embedded-atom potentials
(EAM) which incorporate the properties of the free electron gas. To represent a
specific material these effective interactions are fitted to the quantum mechanical
properties. A linked-cell algorithm can be applied to find interacting atoms since
EAM interactions are effectively pairwise and short-ranged. A domain decomposition
of the simulation box and message passing can be implemented. Thus, MD ideally
scales linearly up to huge sample sizes. The principle setup for the SLM simulations
is as follows: the objects which are to be irradiated are placed on a fixed plate which
is modeled by immobile atoms. Then a laser beam is moving across the objects with
a fixed velocity.

2.2.1 Laser absorption

If matter is irradiated by a laser, beam energy will be absorbed according to the law
of Lambert-Beer:

𝐼 (𝑥) = 𝐼 (0) · exp −𝜇𝑥 (1)
where 𝐼 (𝑥) is the intensity along 𝑥 and 𝜇 is the inverse absorption length. The laser
will excite the electrons and the energy will eventually be transformed into thermal
energy of the atoms. In the present work the rescaling method has been applied since
the time scales of the thermalization processes are shorter than the interaction time
scales of the laser. The velocities of the atoms 𝑣(𝑡 + d𝑡) = 𝑎(𝑥)𝑣(𝑡) are scaled with
70 F. Oelschläger, D. Klein, S. Müller and J. Roth

respect to the Lambert law by

𝐸 kin (𝑡 + d𝑡)
𝑎(𝑥) 2 = (2)
𝐸 kin (𝑡)
where 𝑥 is the position of the atom with respect to the surface. The kinetic energy
𝐸 kin (𝑡) can be summed up over all atoms and results in the absorbed energy d𝐸 per
volume d𝑉 and time d𝑡 via
!

𝑃tot 𝑥2 + 𝑦2
d𝐸 = (1 − 𝑅) · 𝜇 · exp −𝜇𝑧 · · exp − · d𝑉 · d𝑡 (3)
𝜋𝜎 2 𝜎2

where 𝑅 is the reflectivity of the surface and 𝑃tot the absorbed power if a circular
Gaussian beam of width 𝜎 is assumed.
The derivation here is certainly only correct for an object of constant height. The
modification for a sphere and a fixed laser beam could be derived easily, at least
numerically. In the present case, however, we are interested in moving laser beams
and thus the correct absorption of objects of variable height caused by a moving laser
beam together with changing shape during the process can only be modeled by an
accompanying ray tracing calculations. Currently, addressing this problem has been
postponed since it would complicate the modeling considerably and distract from the
proof of principle.

2.2.2 Sample sizes

The size of MD simulation samples have increased tremendously in recent years up

to about 1013 atoms[2]. But such runs would require full-time simulations on entire
top-level supercomputers.
The sizes of powder particle in the SLM process in the range of a few 𝜇m [8] is
at the limit of current production simulations. For the intended proof of principle
simulations the size of the system has been scaled down to a few hundred Å, but will
be increased subsequently in the production simulations.
Scale factors have been introduced to connect simulation and experiment. The
procedure can be motivated by the work of Glosli et al [7]. This group could
demonstrate that the Kelvin-Helmholtz instability in fluids occurs at relatively small
particle numbers already. Selective laser melting depends on the behavior of fluids in
a similar way. The scaling also includes an increase of the gravitational force.
Selective laser melting 71

3 Simulation of SLM

3.1 Preparation of the sample

The initial sample is a crystal with periodic boundary conditions which is equilibrated
at 300 K. Then a sphere is cut out of the material and equilibrated under the influence
of gravity which causes the sphere to settle down slightly. Thus, the contact plane
between the fixed ground and the sphere is enlarged.

3.2 Variation of the laser power at fixed laser velocity

The approximations applied largely exclude the direct transfer of the simulation data
to experiment as discussed in Sec. 2.2.1. Therefore the model has been verified
predominantly in a qualitative comparison.
The laser velocity is set to a fixed value of 1Å per fs. A number of power values
are simulated to obtain a reference simulation for further studies. The first goal is to
relate each laser power at this velocity to a melting capability. The criterion is the
amount of material molten during simulation.

100
100

80
molten fraction [%]

molten fraction [%]

90

60
80

40
70

20 P = 6 eV/fs P = 6 eV/fs
P = 8 eV/fs P = 8 eV/fs
P = 10 eV/fs 60 P = 10 eV/fs
P = 16 eV/fs P = 16 eV/fs
0
0 10 20 30 40 50 0 50 100 150 200
time t [ps] time t [ps]

(a) Initial range (b) Long time development

Fig. 4: Number of particles in the molten phase vs. simulation time. The laser beam
with velocity 𝑣 laser = 1Å per fs has passed the sphere completely after 10 ps.

Fig. 4 shows the molten fraction of the material vs. time for different laser powers
𝑃. The short time behavior has been zoomed in on the left (Part (a)). Obviously the
laser melts the sample instantaneously and the molten fraction stays constant for a
short time. Fitting a constant in this range determines the molten fraction at this laser
power.
72 F. Oelschläger, D. Klein, S. Müller and J. Roth

Table 1: Molten fraction of material.

power
 𝑃 molten increase
 
eV/fs %/fs
6 68.54% 8.5·10−5
8 79.49% 8.6·10−5
10 86.95% 8.2·10−5
16 99.26% 8.4·10−5

The data in Tab. 1 show that the sphere is completely molten at a power of 𝑃 = 16
eV/fs. The missing part of ≈ 0.8% results from fluctuations at the beginning. A
closer look at Fig. 4 (a) shows that 100% are finally reached. The complete melting
is displayed in Fig. 5. The side view in Fig. 5 (b) indicates that the floor is wetted
completely as observed in experiment[5]. This is a strong hint that the model is
suitable.

(b) Side view

(a) Perspective view

Fig. 5: The completely molten sphere at a power of 𝑃 = 16 eV/fs after 210 ps.

An interesting phenomenon occurs at low powers (Fig. 4(b)). A nearly linear

increase of the molten fraction is visible for a certain period while the molten fraction
directly after laser impact is constant up to fluctuations until saturation sets in finally.
The rate of increase in this region is nearly the same for all laser powers. The increase
of molten material on average is 8.4·10−5 %/fs. The reason is that the fraction molten
directly by the laser transfers the heat to the still solid part by heat conduction which
is a material property only.
A further phenomenon which is due to the current model limitations is visible in
Fig. 6. Shown is a cut through the center of the sphere in the 𝑧-𝑥-plane. Red atoms are
in the molten part, white atoms are solid. The sphere is not yet completely molten at
𝑃 = 10 eV/fs, but the solid rest shrinks subsequently. Interesting are the voids formed
at the bottom in the left figure. The melt from above closes the voids laterally and
prevents a spreading of the un-molten nucleus as in Fig. 5. Thus, the partially solid
Selective laser melting 73

(a) 𝑡 = 112.5 ps (b) 𝑡 = 140 ps

Fig. 6: Cuts through the sphere. Voids disappear due to the surrounding vacuum.
Inclusions would be visible if the voids were filled with protective gas.

structures lead to void formation. The voids vanish some time later since they are not
filled with a gas. The surrounding material has penetrated into the void and closed it.
With protective gas the closed void could have formed a pore.

3.3 Behavior at reduced laser velocity

The goal here is to achieve the same melting volume with different parameters. The
laser power has to be reduced to obtain the same amount of molten material at lower
laser velocity since laser power is applied energy per time. The connection is not a
simple proportionality as exemplified by Fig. 7. Obviously, the curves at half power
and half velocity do not coincide. The parameter pairs 𝑣 = 1 Å/fs, 𝑃 = 8 eV/fs and
𝑣 = 0.5 Å/fs, 𝑃 = 5 eV/fs on the other hand show a rather similar behavior at the
beginning.
Fig. 8 shows a simulation with higher velocity at the top and one with lower
velocity at the bottom. To the right are the same samples at a later time. Details
are different although the simulations look rather similar at first. In the shorter time
pictures much less small droplets have been formed at lower laser velocities, and this
behavior continues until later times. The droplets can be responsible for a number of
defects including and splashing or balling. Defects are reduced at higher velocities as
mentioned in Sec. 2.1.2 already[15]. The fact is taken as a further sign that the model
is suitable for an accurate description of the problem.

4 Summary and outlook

The goal of this work is to validate the applicability of atomistic MD simulations

to SLM. The crux is the fact that the scale of the powder particles are a few 𝜇m
which is a challenge even for big supercomputers. Currently the model is scaled
74 F. Oelschläger, D. Klein, S. Müller and J. Roth

100

80

molten fraction [%]

60

40

v = 10 A/fs, P = 6 eV/fs
20 v = 10 A/fs, P = 8 eV/fs
v = 5 A/fs, P = 3 eV/fs
v = 5 A/fs, P = 4 eV/fs
v = 5 A/fs, P = 5 eV/fs
0
0 10 20 30 40 50
time t [ps]

Fig. 7: Fraction of particles belonging to the molten material vs. simulation time.
Like Fig. 4 (a) for different velocities.

(a) 𝑡 = 12.5 ps (a) 𝑡 = 30.0 ps

Fig. 8: Two parameter pairs leading to a similar result. Top: 𝑣 = 1 Å/fs, 𝑃 = 8 eV/fs,
bottom: 𝑣 = 0.5 Å/fs, 𝑃 = 5 eV/fs

down to several hundred Å. The results of these proof of principle simulations have
demonstrated that this project should be continued. Hints have been found that the
mechanisms at the basis of the observed defects can be studied.
Selective laser melting 75

The modeling is not yet complete, but the basic requirements for a successful
description of SLM are fulfilled. A mechanism for pore formation could be identified
and the disappearance of the pores without protective gas could be observed. It could
be shown that lower laser velocities favor droplet formation which has been named as
a reason for balling and splashing[15].
Pure aluminum in vacuum has been studied, a material which is rarely used
but contained in many compounds as a majority element. Pure metals simplify the
identification and discussion of the many open questions.
More realistic simulations are necessary. They include:
• The parameter space of sphere size, laser velocity and laser power should be
studied in more detail.
• Several spheres in different configurations should be simulated to observe their
interaction. Different packing densities are of highest interest since they may
lead to pore formation.
• Recrystallization should also be studied in this context. Then other defects like
microcracks also come into play.
• Protective gas like argon should be added to modify the behavior of the pores.
• Cooling process of the fixed floor should be taken into account. Currently the
simulations are run in an isolated NVE ensemble which eventually leads to
complete melting of the whole setup late after irradiation. This can be avoided
by applying an isothermal ensemble which leads to active cooling.
• Up to now the sphere has been placed on solid ground. But typically the un-molten
powder lies above an already processed layer and the interaction with a variable
floor should also be studied.
• The simulation can be extended to other metals and alloys as desired.
Mesoscopic simulations have been carried out at true sample sizes[13] already.
Obviously, they cannot give information about atomistic processes. At the HLRS it
should be possible so scale up the simulations to the 𝜇m regime where the number of
atoms reaches orders of several billions. Ablation simulations of Al[3] and Si[10]
have demonstrated already that this is possible for production runs with IMD. And if
the whole supercomputer HAWK is available, demonstration runs with even larger
sizes should be feasible.

5 Performance of simulation code

In the current reporting period we have successfully carried out molecular dynamics
simulations laser ablation simulations on Hawk with
• quasi 1D silicon samples (Si1D),
• quasi 2D silicon samples of small (Si2Dsmall) and big (Si2Dbig) sizes, as well as
• simulations of the selective laser melting of Al samples (SLM).
76 F. Oelschläger, D. Klein, S. Müller and J. Roth

Further simulation tasks of minor size and already terminated tasks on Hazelhen will
not be reported.

5.1 Typical number of applied processors

The typical number (#) of applied processors is displayed in Tab. 2 according to

simulation type. Please consider that the SLM simulations have not yet been run in
full productive mode.

simulation wall time # atoms 𝑛 # nodes # processors performance [CPUs/𝑛Δ𝑡 ]

Si1D 24h 18,121,592 64 8,192 5.344767 · 10−5
Si2Dsmall 24h 67,108,864 128 16,384 1.342413 · 10−4
Si2Dbig 24h 536,870,912 1024 65,536 5.174720 · 10−4
SLM 24h 27,680,214 64 8,192 1.374238 · 10−4

Table 2: Typical resources applied and performance achieved in CPU seconds (cpus
per atom per simulation-step Δ𝑡).

5.2 Parallelization degree

IMD is a typical molecular dynamics simulation code. As such the main computation
load is generated in a single main loop where the classical interactions between atoms
are calculated. The program achieves parallelization degrees of the order of 90 and
more percent, depending on the simulated experimental setup. The main loop can
be fully parallelized with MPI due to its short range nature and distributed to any
number of processors. It has been demonstrated (for example on JuGene) that IMD
shows weak scaling up to the full machine size as long as the system is sufficiently
homogeneous.

5.3 Scaling

In “real” simulations the samples are not homogeneously filled with atoms but
contain empty space around spheres for example. The simulation box is spatially
decomposed and distributed to various nodes and thus some processors may idle.
Together with the unavoidable communication load between increasing numbers of
Selective laser melting 77

processor some performance loss and scaling decay occurs (Fig. 9) especially for large
simulations. Applying dynamical load balancing implemented in IMD, performance
can be improved to some extent.

Si1D Si2Dbig
Si2Dsmall SLM
0.00055
0.0005
0.00045
0.0004
0.00035
CPUs/n∆t

0.0003
0.00025
0.0002
0.00015
0.0001
5x10-5
0
16 32 64 128 512 1024
# nodes

Fig. 9: Scaling of the benchmarked simulation types.

5.4 Consumed computing resources

Since July 2020 a total of 452,758 node-hours have been spent on HPE Apollo (Hawk).
The largest share has been used for the ongoing PhD thesis of D. Klein [10, 11]
studying laser ablation of silicon (about 407,948 node-hours, including a share for
the SLM study). F. Oelschläger [12] has spent 14,738 node-hours for the study
of SLM presented here. K. Vietz [19] used 19,685 node-hours for the study of
Al-Ni-alloys under laser ablation. The remaining part was spent by E. Eisfeld [3, 4] at
the completion of his thesis on the molecular dynamics simulation of Al with plasma
effects.

Acknowledgements Dominic Klein is supported by the Hans-Böckler-Stiftung. Sarah Müller is

supported by the Deutsche Forschungsgemeinschaft DFG.
78 F. Oelschläger, D. Klein, S. Müller and J. Roth

References

1. E. Brandl, A. Schoberth, and C. Leyens. Morphology, microstructure, and hardness of titanium

(Ti-6Al-4V) blocks deposited by wire-feed additive layer manufacturing (ALM). Materials
Science and Engineering: A, 532:295–307, 2012.
2. W. Eckhardt. Scientists Achieve World’s Largest Molecular Dynamics Simulation, 2013.
https://gauss-centre.eu/results/materials-science-and-chemistry/artic
le/scientists-achieve-worlds-largest-molecular-dynamics-simulation/,
abgerufen am 28.04.2021.
3. E. Eisfeld. Molekulardynamische Simulationen der Laserablation an Aluminium unter Ein-
beziehung von Plasmaeffekten. PhD thesis, Universität Stuttgart, 2020.
4. E. Eisfeld, D. Förster, D. Klein, and J. Roth. Laser ablation of al under short and ultrashort
pulses: Insights from atomistic simulations. J. Phys. D, 2022. submitted.
5. H. Eskandari Sabzi and P. E. J. Rivera-Díaz-del Castillo. Defect Prevention in Selective Laser
Melting Components: Compositional and Process Effects. Materials, 12(22), 2019.
6. C. Galy, E. Le Guen, E. Lacoste, and C. Arvieu. Main defects observed in aluminum alloy parts
produced by SLM: From causes to consequences. Additive Manufacturing, 22:165–175, 2018.
7. J. N. Glosli, D. F. Richards, K. J. Caspersen, R. E. Rudd, J. A. Gunnels, and F. H. Streitz. Extend-
ing stability beyond CPU millennium: a micron-scale atomistic simulation of Kelvin-Helmholtz
instability. In SC ’07: Proceedings of the 2007 ACM/IEEE Conference on Supercomputing,
pages 1–11, 2007.
8. J. Hajnys, M. Pagac, J. Mesicek, J. Petru, and F. Spalek. Research of 316l metallic powder for
use in SLM 3d printing. Advances in Materials Science, 20(1):5–15, 2020.
9. F. Horsch. 3D-Druck für alle. Carl Hanser Verlag GmbH & Co. KG, 2 edition, 2014. pp.
114,119.
10. D. Klein. Computer simulations of laser ablation from simple metals to complex metallic alloys.
PhD thesis, Universität Stuttgart, 2021.
11. D. Klein, E. Eisfeld, and J. Roth. Molecular dynamics simulations of the laser ablation of
silicon with the thermal spike model. J. Phys. D, 54, 2021. submitted.
12. F. Oelschläger. Molekulardynamische Simulation von selektivem Laserschmelzen. Bachelorar-
beit, Universität Stuttgart, 2021.
13. C. Panwisawas, C. Qiu, M. J. Anderson, Y. Sovani, R. P. Turner, M. M. Attallah, J. W. Brooks,
and H. C. Basoalto. Mesoscale modelling of selective laser melting: Thermal fluid dynamics
and microstructural evolution. Computational Materials Science, 126:479–490, 2017.
14. J. Roth, E. Eisfeld, D. Klein, S. Hocker, H. Lipp, and H.-R. Trebin. IMD – the ITAP
molecular dynamics simulation package. The European Physical Journal Special Topics,
227(14):1831–1836, 2019.
15. M. F. Sadali, M. Z. Hassan, F. Ahmad, H. Yahaya, and Z. A. Rasid. Influence of selective
laser melting scanning speed parameter on the surface morphology, surface roughness, and
micropores for manufactured Ti6Al4V parts. Journal of Materials Research, 35(15):2025–2035,
2020.
16. M. Saunders. X marks the spot - find ideal process parameters for your metal AM parts, 2017.
https://www.linkedin.com/pulse/x-marks-spot-find-ideal-process-paramet
ers-your-metal-marc-saunders, abgerufen am 24.04.2021.
17. X. Shi, S. Ma, C. Liu, C. Chen, Q. Wu, X. Chen, and J. Lu. Performance of High Layer
Thickness in Selective Laser Melting of Ti6Al4V. Materials, 9(12), 2016.
18. B. Song, S. Dong, B. Zhang, H. Liao, and C. Coddet. Effects of processing parameters on
microstructure and mechanical property of selective laser melted Ti6Al4V. Materials & Design,
35:120–125, 2012.
19. K. Vietz. Simulation der Laserbearbeitung von Aluminium-Nickel-Materialien. Bachelorarbeit,
Universität Stuttgart, 2021.
20. Q. Yan, B. Song, and Y. Shi. Comparative study of performance comparison of AlSi10Mg alloy
prepared by selective laser melting and casting. Journal of Materials Science & Technology,
41:199–208, 2020.
Selective laser melting 79

21. B. Zhang, Y. Li, and Q. Bai. Defect Formation Mechanisms in Selective Laser Melting: A
Review. Chinese Journal of Mechanical Engineering, 30(3):515–527, 2017.
22. J. Zou, Y. Zhu, M. Pan, T. Xie, X. Chen, and H. Yang. A study on cavitation erosion behavior
of AlSi10Mg fabricated by selective laser melting (SLM). Wear, 376-377:496–506, 2017.
Molecular dynamics investigations on the
influence of solutes on the tensile behavior of
Polyamide6

Wolfgang Verestek, Johannes Kaiser, Christian Bonten and Siegfried Schmauder

Abstract Although PA6 shows remarkable properties, when in contact with specific
media, the strength is reduced. Here we report about molecular dynamics (MD)
simulations of PA6 with solutes and their influence on the mechanical properties
of the bulk material. For this, four selected scenarios have been simulated. More
precisely, four cases are defined: the “dry” case without any solute, “saturated with
water” at 5.8 mass %, “saturated with ethanol” at 12.4 mass % and “saturated with
ethanol” at 12.3 mass %. The simulated nano tensile tests show differences in the
Young’s modulus and the ultimate tensile stress, where methanol shows the highest
reduction.

Wolfgang Verestek
IMWF University of Stuttgart, Pfaffenwaldring 32, 70569 Stuttgart, e-mail: wolfgang.verestek@
imwf.uni-stuttgart.de
Johannes Kaiser
IKT University of Stuttgart, Pfaffenwaldring 32, 70569 Stuttgart, e-mail: johannes.kaiser@ikt.
uni-stuttgart.de
Christian Bonten
IKT University of Stuttgart, Pfaffenwaldring 32, 70569 Stuttgart, e-mail: christian.bonten@ikt
.uni-stuttgart.de
Siegfried Schmauder
IMWF University of Stuttgart, Pfaffenwaldring 32, 70569 Stuttgart e-mail: siegfried.schmaude
[email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 81

W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_5
82 W. Verestek, J. Kaiser, C. Bonten and S. Schmauder

1 Introduction

Polyamide 6 (PA6), among other Polymers, has sparked scientific and industrial
interest in commercial research and academic laboratories due to its remarkable
properties. Strengthening can happen at relatively low filler concentrations, for
example with 4.7 mass %-layered silicates the Young’s modulus and the strength of
the material can be doubled [1].
Although PA6 shows remarkable properties, when in contact with specific media,
the strength is reduced. Here we report about molecular dynamics (MD) simulations
of PA6 with solutes and their influence on the mechanical properties of the bulk
material. For this, four selected scenarios have been simulated. More precisely, the
following four, experimentally based [2] cases are defined:
1. Dry: no solute
2. Saturated with water: 5.8 mass % water
3. Saturated with methanol: 12.4 mass % methanol
4. Saturated with ethanol: 12.3 mass % ethanol

2 Model

In the following the model creation, equilibration and the performed simulations will
be described.

2.1 Model creation

Polyamide 6 can be produced by a condensation reaction of monomers or by a ring

opening polymerization. For PA6 this results in amide groups (-N(H)C(=O)-) with
five CH2 s between each of them. The resulting Polymer with polymerization grade n
is depicted in Fig. 1. The solutes investigated in this contribution are shown in Fig. 2.

Fig. 1: Polyamide 6 with polymerization grade n. Nitrogen is depicted in Blue, Oxygen

in Red. Hydrogens attached to nitrogen and Oxygen are shown, Hydrogens attached
to Carbon are not shown for clarity.
Influence of solutes in PA6 on the tensile behavior of PA6 83

Fig. 2: Solutes. From left to right: Water, Methanol, Ethanol

The three-dimensional periodic models for the tensile test simulations were created
with EMC [3] and contain approx. 0.1 mio. atoms. For all models 20 molecules
of PA6 were created with a polymerization grade n (see Fig. 1) ranging from 210
to 300 in steps of 10. This results in two molecules for each polymerization grade.
The resulting models are summarized in table 1. All models have an initial density
of 1.05 g/dm3 and are thermalized and relaxed during equilibration, see Sec. 2.2.
To produce the pure PA6 models without solute the corresponding molecules were
simply removed, resulting in a lower initial density, and the equilibration procedure
was applied.

Table 1: Model setup

Solute Mass fraction Solute molecules

None 0.0 mass % 0

Water 5.8 mass % 1985

Methanol 12.4 mass % 2561
Ethanol 12.3 mass % 1769

2.2 Equilibration

For simulation the open source MD code LAMMPS [4] was used and the 2nd
generation Polymer Consistent Force Field (PCFF) [5] with an cutoff of 9.5 Å
was applied. The equilibration consists of an initial energy minimization with the
conjugate gradient method as well damped dynamics to relax overlapping and very
close atoms. This is followed by an NVE ensemble for 10.000 steps with a limited
displacement of 0.1 Å/step and a time step width of 0.25 fs to allow for further
relaxation.
To allow higher chain mobility a soft-modified version of the PCFF potential was
utilized [6] with n = 2, 𝛼 𝐿 𝐽 = 0.5 and 𝛼𝐶 = 10. Two simulations with each 50.000 steps
in the NVT ensemble with a starting temperature of 700K and a final temperature
relaxed the chain morphology further. The first simulation used 𝜆 = 0.8 and the
84 W. Verestek, J. Kaiser, C. Bonten and S. Schmauder

second one used 𝜆 = 0.9 [6]. To allow larger time steps the rRESPA-Algorithm [7]
was employed with 3 levels (bonds, pair and long range/kspace) and scaling factors 2
and 2 resulting in an inner time step of 0.25 fs for bond, 0.5 fs for pair and an outer
time step of 1.0 fs for long range interactions.
Finally a sequence of NPT and NVT simulations, each 100.000 steps, was applied.
The NVT simulations had a starting temperature of 700 K and an end temperature
of 300 K. The NPT simulations had starting and end temperature set to 300 K and
a target pressure of 0 bar and isotropic volume scaling. Again, for larger time steps
the rRESPA-Algorithm was employed with the same settings as before, resulting in
a simulated time of 0.1 ns per simulation run. The final, equilibrated structures are
shown for the dry case in Fig. 3 as well as in Fig. 4 for models with solutes.

Fig. 3: Equilibrated structure for PA6 without solutes. On the left side, the typical color
coding is applied to the Atoms (H: White, C: Grey, N: Blue, O: Red). On the right
side the color coding is due to the Molecule ID and allows an better understanding of
the molecular morphology.

Fig. 4: Equilibrated structures of PA6 with solutes. The typical atomic color coding is
applied (H: White, C: Grey, N: Blue, O: Red) plus additionally the solute molecules
are colorized in Orange. From left to right the solute is water, methanol and ethanol
with mass fractions of approx. 6 and 12 mass % respectively, see Tab. 1.
Influence of solutes in PA6 on the tensile behavior of PA6 85

2.3 Tensile test simulation

As before, the tensile test simulations for a temperature of 300 K were performed
with the polymer consistent force field, but the cutoff was increased to 12 Å. Tensile
test simulations were done in two different setups. The first setup was simulated in the
NPT ensemble with a target pressure of 0 bar in lateral direction and a relaxation time
of 0.1 ps for the temperature and 1 ps for the pressure. A second set of simulations was
performed in the NVT ensemble with the same settings. Due to the NVT ensemble
lateral contraction is not allowed and results effectively in a triaxial stress state. Again,
the rRESPA-scheme was employed to allow larger time steps with an inner time
step (bonds) of 0.375 fs, a middle time step (pair) of 0.75 fs and an outer time step
(long range) of 1.5 fs. For each equilibrated model, tensile tests were simulated.
Deformation was applied each time step with true strain rates of 1.0e-5 and 1.0e-6
1/fs. To reduce the noise of the stress-strain curves, the pressure was averaged over
100 time steps (150 fs) before becoming a data point for the plot. To get better, less
morphology dependent stress-strain curves, each model was deformed in x, y and z
direction, see Fig. 5, and an average resulting stress is computed.

Uniaxial tensile test Triaxial tensile test

700
x 1.0e-6 1/fs
x 1.0e-6 1/fs
600 y 1.0e-6 1/fs
y 1.0e-6 1/fs
500 z 1.0e-6 1/fs
z 1.0e-6 1/fs
Stess [MPa]

400 ave 1.0e-6 1/fs

ave 1.0e-6 1/fs
300

200
x 1.0e-6 1/fs
100 y 1.0e-6 1/fs
z 1.0e-6 1/fs
ave 1.0e-6 1/fs
0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3 3.5
Strain [-] Strain [-]

Fig. 5: Stress-strain curves for PA6 without solute for the uniaxial stress state (NPT)
and the Triaxial stress state for a true strain rate of 1.0e-6 1/fs for deformation in
x, y and z direction and their average. The lower diagram for triaxial stress state
additionally shows the averaged stresses orthogonal to the loading direction with
dashed lines.

For the uniaxial case the stress goes up nearly linear in the beginning until plastic
deformation starts and reaching the ultimate tensile stress. After reaching the a
plateau strain hardening takes place and leads to higher stresses than the ultimate
tensile stress. One should consider here, that the plots show true stress-strain curves.
Transitioning to engineering stress-strain curves would show lower stresses for higher
deformation grades as the contraction for the uniaxial case is neglected. It is obvious
that the triaxial stress state results in a somewhat higher ultimate tensile stress. But
86 W. Verestek, J. Kaiser, C. Bonten and S. Schmauder

after reaching the ultimate tensile stress the orthogonal stresses decay nearly to zero
whereas the stress in loading direction reaches a plateau at which the disentanglement
of the molecular chains happens.

3 Results

The uniaxial model, Fig. 6, stays more compact compared to the triaxial case where
a big void is already seen in the first picture of Fig. 7. Due to additional degree of
freedom in the lateral direction the molecules in the uniaxial model get aligned in
loading direction. This is not only true for the voids at higher strains, but also for the
bulk phase at lower strains. The molecules align in both cases at higher strains when
spanning a void and look like “nano” fibrils.
In Fig. 8 to 11 the simulated stress-strain curves are shown. For the case of the
dry PA6 without any solute distinct ultimate tensile stress points can be seen for the
uniaxial case for all simulated strain rates. This is followed by a little bit lower yield
plateau before strain hardening takes place and reaching a stress of approximately
490 MPa at a strain of 350 %. It is worth mentioning, that the stress-strain curves for
both strain rates nearly lie on top of each other in the strain hardening regime after
leaving the plateau.
For the triaxial case the ultimate tensile stresses are also clearly visible. This is
true for the higher stresses in loading direction as well as for the lower stresses in
lateral direction. The ultimate tensile stresses in loading direction are separated by
approx. 60 and 100 MPa, respectively, where the strain rate is lower by a factor of 10
and 100. In contrast to this, the maximum stresses in transversal direction show only
a very slight reduction due to the lower strain rate. After reaching the ultimate tensile
stress the lateral stresses decay towards very low values whereas the stress in loading
direction decays only to a certain, nearly horizontal level, at which disentanglement
of the molecules takes place.
The tensile test simulation for PA6 with Water are shown in Fig. 9. Compared to
the dry, uniaxial case three differences are noticeable. Firstly, for the higher strain
rate of 1.5e-5 1/fs the ultimate tensile stress is clearly visible followed by a plateau,
whereas for the lower strain rates the plateau vanishes and the ultimate tensile stress
is at about 227 MPa and 221 MPa. Secondly, the ultimate tensile stress for all strain
rates are lower than those for the dry case. And lastly, the slope of the strain hardening
is higher compared to the uniaxial case in Fig. 8 and shows a higher stress, more than
700 MPa at 350 % strain compared to approx. 500 MPa for the dry case. For the
triaxial load case the behavior is comparable at a load level approx. 30 MPa lower.
The tensile test simulation for PA6 with methanol are shown in Fig. 10. For
the uniaxial case the slope in the hardening regime is higher than in the dry case,
comparable to the water loaded case. But the stress level is lower than that for water
with approx. 600 and 650 MPA for the two strain rates at 350 % strain. Also the
Influence of solutes in PA6 on the tensile behavior of PA6 87

Fig. 6: Simulated tensile test for dry PA6 with lateral contraction for a true strain rate
of 1.0e-6 1/fs after 25, 50 and 100 % simulation time and 45.5 %, 111.7 %, 208.0 %
and 348.2 % strain respectively. The color coding is the same as before (H: White, C:
Grey, N: Blue, O: Red). Visualization with Ovito [8].

plateau for the higher strain rate of 1.5e-5 1/fs is less pronounced. In general the
ultimate tensile stresses are lower than those for the dry case and the water case for
all stress states and strain rates.
The uniaxial stress-strain curves for ethanol, Fig. 11 look very similar to those of
methanol. But the plateau for the strain rate of 1.0e-5 1/fs is a little bit more distinct
than that for methanol. In contrast the slope of the strain hardening regime is less
steep and results in lower stresses of approx. 580 MPa and 550 MP. For the triaxial
case, the curves are very similar to the methanol case. Generally, the ultimate tensile
stresses are slightly higher for ethanol than for methanol, but still lower than for the
water case.
88 W. Verestek, J. Kaiser, C. Bonten and S. Schmauder

Fig. 7: Simulated tensile test for dry PA6 without lateral contraction for a true strain
rate of 1.0e-6 1/fs after 25, 50 and 100 % simulation time and 45.5 %, 111.7 %,
208.0 % and 348.2 % strain respectively. The color coding is the same as before (H:
White, C: Grey, N: Blue, O: Red). Visualization with Ovito [8].

Uniaxial tensile test Triaxial tensile test

700
ave 1.0e-5 1/fs
ave 1.0e-5 1/fs
600 ave 1.0e-6 1/fs
ave 1.0e-6 1/fs
500 ave 1.0e-7 1/fs
Stress [MPa]

ave 1.0e-7 1/fs

400

300

200

100 ave 1.0e-5 1/fs

ave 1.0e-6 1/fs
ave 1.0e-7 1/fs
0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3 3.5
Strain [-] Strain [-]

Fig. 8: Simulated tensile test for dry PA6 with and without lateral contraction for true
strain rates 1.0e-5, 1.0e-6 and 1.0e-7 1/fs.
Influence of solutes in PA6 on the tensile behavior of PA6 89

Uniaxial tensile test Triaxial tensile test

700
ave 1.0e-5 1/fs
ave 1.0e-5 1/fs
600 ave 1.0e-6 1/fs
ave 1.0e-6 1/fs
500 ave 1.0e-7 1/fs
Stress [MPa]

ave 1.0e-7 1/fs

400

300

200

100 ave 1.0e-5 1/fs

ave 1.0e-6 1/fs
ave 1.0e-7 1/fs
0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3 3.5
Strain [-] Strain [-]

Fig. 9: Simulated tensile test for PA6 with 5.8 mass % water with and without lateral
contraction for true strain rates 1.0e-5, 1.0 e-6 and 1.0 e-7 1/fs.

Uniaxial tensile test Triaxial tensile test

700
ave 1.0e-5 1/fs
ave 1.0e-5 1/fs
600 ave 1.0e-6 1/fs
ave 1.0e-6 1/fs
500 ave 1.0e-7 1/fs
Stress [MPa]

ave 1.0e-7 1/fs

400

300

200

100 ave 1.0e-5 1/fs

ave 1.0e-6 1/fs
ave 1.0e-7 1/fs
0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3 3.5
Strain [-] Strain [-]

Fig. 10: Simulated tensile test for PA6 with 12.4 mass % methanol with and without
lateral contraction for true strain rates 1.0e-5, 1.0 e-6 and 1.0 e-7 1/fs.

Uniaxial tensile test Triaxial tensile test

700
ave 1.0e-5 1/fs
ave 1.0e-5 1/fs
600 ave 1.0e-6 1/fs
ave 1.0e-6 1/fs
500 ave 1.0e-7 1/fs
Stress [MPa]

ave 1.0e-7 1/fs

400

300

200

100 ave 1.0e-5 1/fs

ave 1.0e-6 1/fs
ave 1.0e-7 1/fs
0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3 3.5
Strain [-] Strain [-]

Fig. 11: Simulated tensile test for PA6 with 12.3 mass % ethanol with and without
lateral contraction for true strain rates 1.0e-5, 1.0 e-6 and 1.0 e-7 1/fs.
90 W. Verestek, J. Kaiser, C. Bonten and S. Schmauder

Figure 12 shows a summary for the dry case and all solutes with and without
contraction for a strain rate of 1.0e-7 1/fs. To enable a more detailed representation the
stress-strain curves are only shown up to 50% strain. The aforementioned observations
can also be seen here.

Uniaxial tensile test Triaxial tensile test

350

300

250
Stress [MPa]

200

150

100

50
dry methanol dry methanol
water ethanol water ethanol
0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4 0.5
Strain [-] Strain [-]

Fig. 12: Simulated tensile test for PA6 for the dry case as well as water, methanol and
ethanol with and without lateral contraction for a true strain rate of 1.0 e-7 1/fs.

Starting from here the Young’s modulus for each averaged stress-strain curve has
been calculated and is shown in Tab. 2. The previously shown curves are somewhat
fluctuating, therefor the Young’s modulus was computed from the Bézier smoothed
curves at 1% strain. It can be seen, that the triaxial load case leads to a stiffer behavior
than the uniaxial one. Furthermore a clear trend among the solutes can be observed.
Regardless of the load case, dry PA6 has the highest Young’s modulus followed
by water and ethanol. Methanol shows the lowest Young’s modulus. Additionally,
experimental values from [2] are shown for the quasi static case. It should be stated
here, that in the experiments different compositions of the constituents led to various
grades of crystallinity in the PA6 compounds and thereby also to a certain range.
Therefor only max and mean values of one composition (90% Ultramid B40 (BASF),
10% SelarPA3426 (DuPont)) are reported here. It is interesting to note, that the
Young’s modulus of dry PA6 is very close to the experimental one, whereas the
Young’s moduli with solutes differ by a factor of approx. 2.5-4.5 with respect to the
experimental maximum values. Part of this might be subjected to the well known
strain rate and size effect. Also the diffusion speed of the solutes in PA6 might play a
role.
Having calculated the Young’s modulus the next step would be the yield stress. As
there is no clear and pronounced yield stress visible in the plots above 𝑅 𝑝0.2 as well
as 𝑅 𝑝2 have been computed and are listed in Tab. 3
The ultimate tensile stresses for the different simulations are summarized in Tab. 4
and Fig. 13. From Fig. 13 a clear trend is visible. Each solute lowers the yield stress.
water has the least effect. Methanol and ethanol have a similar influence whereas
methanol reduces the yield stress a little bit more than ethanol.
Influence of solutes in PA6 on the tensile behavior of PA6 91

Table 2: Young’s modulus

strain rate Dry Water Methanol Ethanol

1/fs MPa MPa MPa MPa

uniaxial
1e-5 4730 4450 3730 3800
1e-6 3930 3470 2840 3110
1e-7 3000 2550 1890 2130

triaxial
1e-5 7040 7170 5730 6100
1e-6 6150 6010 4950 5070
1e-7 5530 5320 4290 4410

Experiment 2700/2530 985/480 434/398 664/558

Uniaxial tensile test Triaxial tensile test

400 400

350 350
Stress [MPa]

Stress [MPa]

300 300

250 250

200 200

150 150 dry

water
methanol
ethanol
100 100
1e-07 1e-06 1e-05 1e-07 1e-06 1e-05
Strain rate [1/fs] Strain rate [1/fs]

Fig. 13: Ultimate tensile stresses for different solutes, stress states and strain rates.

The reduction of the yield stress is caused by the solute molecules diffusing
through the polymer and being attracted by the amide groups. Most of the solute
molecules were found to be close to an amide functional group, either the negatively
charged oxygen of the solute aiming at the hydrogen of the nitrogen or the doubly
bonded oxygen of the amide group attracting a hydrogen of the hydroxyl group of the
solute. Due to these hydrogen bridges between the amide group and the solute the
chain mobility is increased. There are two reasons for the increased chain mobility.
Firstly, the solute effectively shields some amide groups inside the PA6 and prevent
these amide groups from building hydrogen bridges with other amide groups. This
shielding of the amide groups allow PA6 molecules slide along each other more
easily without being pinpointed by hydrogen bridges between amide groups. One
92 W. Verestek, J. Kaiser, C. Bonten and S. Schmauder

Table 3: Yield stress

strain rate Dry Water Methanol Ethanol

1/fs MPa MPa MPa MPa

𝑅 𝑝0.2 uniaxial
1e-5 131.8 101.4 88.0 87.2
1e-6 87.1 74.9 62.4 61.3
1e-7 64.8 49.6 47.0 40.5

𝑅 𝑝0.2 triaxial
1e-5 171.4 149.2 134.1 130.4
1e-6 153.2 122.4 106.6 111.8
1e-7 111.4 104.0 82.4 88.3

𝑅 𝑝2 uniaxial
1e-5 253.4 222.1 194.3 200.6
1e-6 195.2 152.8 130.2 130.4
1e-7 143.2 97.8 81.6 86.5

𝑅 𝑝2 triaxial
1e-5 334.5 303.9 269.9 265.6
1e-6 277.2 242.5 208.7 208.5
1e-7 224.0 203.1 159.9 171.0

could assume that this shielding happens by the solute molecules getting close to
the amide groups and by widening the average distance of the hydrogen bonds
between the amide groups. However, the partial radial distribution function (pRDF)
for the O-H hydrogen bond of the amide group, see Fig. 4 does not show a strong
influence. Two points are worth mentioning. Firstly, for the first peak around 1.8 Å
one can see that water leads to a slightly lower probability in the pRDF. Secondly, the
second peak at approx. 3.2 Å has two nearly overlapping groups, namely dry/water
and methanol/ethanol, whereas the second group with methanol and ethanol shows
a slightly higher probability. This is counter intuitive regarding the results of the
Young’s modulus or the ultimate tensile stress and leads to the conclusion that the
shielding is not a simple widening of the hydrogen bonds between the amide groups.
Finally one other interesting observation can be seen when comparing the pRDF
for the amide groups over the course of the tensile test simulation for the two different
stress states, see Fig. 15. While for the uniaxial stress state the pRDF increases just
Influence of solutes in PA6 on the tensile behavior of PA6 93

Table 4: Ultimate tensile stresses

strain rate Dry Water Methanol Ethanol

1/fs MPa MPa MPa MPa

uniaxial
@ 20% strain
1e-5 345.5 312.3 270.6 280.4
1e-6 273.6 227.9 190.7 202.6
1e-7 221.4 182.9 134.4 155.5

triaxial
@ 13% strain
1e-5 407.2 383.4 329.3 331.1
1e-6 338.4 311.4 257.5 262.7
1e-7 295.3 269.3 213.9 220.4

Uniaxial tensile test

14
dry water methanol ethanol
12

10
RDF Amide O−H

0
1.5 2 2.5 3
Distance [Å]

Fig. 14: Partial radial distribution function for O-H in the amide group.

slightly, the pRDF increases nearly linearly for the triaxial stress state and indicates
an increasing crystallinity. When examining Fig. 7 one can easily see, that during the
later stages of the tensile test, the fibril like structures reorganize in a crystalline way.
This explains the increase in the pRDF.
94 W. Verestek, J. Kaiser, C. Bonten and S. Schmauder

Uniaxial tensile test Triaxial tensile test

50 50
40 40
RDF O−H

RDF O−H
30 30
20 20
10 4 10 4
3.5 3.5
3 3
0 2.5 0 2.5
1.5 2 1.5 2
2 1.5 Strain [−] 2 1.5 Strain [−]
2.5 1 2.5 1
3 0.5 3 0.5
Distance [Å] 0 Distance [Å] 0

Fig. 15: Partial radial distribution function for O-H in the amide group during the
tensile test simulation for the uniaxial and triaxial stress state.

4 Summary

In this work amorphous, dry PA6 as well as PA6 saturated with water, methanol and
ethanol have been investigated by means of molecular dynamics. For this, nano tensile
tests have been simulated for uniaxial and triaxial stress states and the corresponding
Young’s moduli, 𝑅 𝑝0.2 , 𝑅 𝑝2 and ultimate tensile stresses have been investigated. A
clear trend with increased weakening could be seen from dry to the solutes, where
water had the least impact followed by ethanol and methanol. Finally a partial radial
distribution analysis of the hydrogen bond between amide groups showed, that the
weakening mechanism is not based on increasing the distance between amide groups.
Furthermore it was revealed, that the increasing crystallinity that could be observed
during the tensile test simulation in the triaxial state stems from the reorganization of
fibril like structures.

Acknowledgements The authors want to thank the DFG (German Research Foundation) for funding
within the project Schm 746/186-1 and BO 1600/30-1. The simulations were performed at the
computer clusters Hawk and Hazelhen at the HLRS, University of Stuttgart.

References

1. Usuki, A., Kojima, Y., Kawasumi, M., Okada, A., Fukushima, Y., Kurauchi T. and Kamigaito,
O.: Synthesis of nylon 6-clay hybrid. In: Journal of Materials Research 8 5, 1179–1184 (1993)
doi: 10.1557/JMR.1993.1179
2. Schubert, M.: "Einfluss der Blendmorphologie auf das Bruchverhalten von konditionierten
PA6-Blends" In: Master Thesis, 2019, IKT Univeristät Stuttgart
3. in’t Veld, P. J.: EMC: Enhanced Monte Carlo - A multi-purpose modular and easily extendable
solution to molecular and mesoscale simulations. http://montecarlo.sourceforge.net
/emc/Welcome.htmlCited26June2020
4. Plimpton, S.: Fast Parallel Algorithms for Short-Range Molecular Dynamics. In: Journal of
Computational Physics 117 1, 1–19 (1995) doi: 10.1006/jcph.1995.1039
Influence of solutes in PA6 on the tensile behavior of PA6 95

5. Sun, H., Mumby, S. J., Maple, J. R. and Hagler, A. T.: An ab Initio CFF93 All-Atom Force
Field for Polycarbonates. In: Journal of the American Chemical Society 116 7, 2978–2987
(1994) doi: 10.1021/ja00086a030
6. Beutler, T. C., Mark, A. E., van Schaik, R. C., Gerber, P. R., van Gunsteren, W. F.: Avoiding sin-
gularities and numerical instabilities in free energy calculations based on molecular simulations.
In: Chemical Physics Letters 222 6, 529–539 (1994) doi: 10.1016/0009-2614(94)00397-1
7. Tuckerman, M. and Berne, B.J.: Reversible multiple time scale molecular dynamics. In: Journal
of Chemical Physics 97 3, 1990–2001 (1992). doi: 10.1063/1.463137
8. Stukowski, A.: Visualization and analysis of atomistic simulation data with OVITO - the Open
Visualization Tool. In: Modeling and Simulation in Materials Science and Engineering 18 1,
015012 (2010). doi: 10.1088/0965-0393/18/1/015012
Dynamical properties of the Si(553)-Au nanowire
system

Mike N. Pionteck, Felix Bernhardt, Johannes Bilk, Christof Dues, Kevin Eberheim,
Christa Fink, Kris Holtgrewe, Niklas Jöckel, Brendan Muscutt, Florian A. Pfeiffer,
Ferdinand Ziese and Simone Sanna

Abstract The lattice dynamics of the Si(553)-Au surface is modeled from first
principles according to different structural models. A multitude of surface-localized
phonon modes is predicted. As a general rule, low-energy modes are associated to
vibrations within the Au chain, while high-energy modes are mostly localized at
the Si step edge. The presence of model specific displacement patterns allows to
identify the structural models compatible with the measured spectra at low and at
room temperature. Our atomistic models within density functional theory allow to
assign spectroscopic signatures available from the literature to displacement patterns,
and to explain the activity of nominally Raman silent modes.

1 Introduction

One-dimensional (1D) physics is characterized by a multitude of exotic phenomena

such as spin-charge separation, triplet superconductivity, and Luttinger liquid behavior
[1–5]. In order to study such phenomena, 1D systems have to be realized. In this
respect, self assembled metallic atomic chains on stepped semiconductor surfaces are
regarded as the most promising quasi-1D model systems [6–12]. Different realizations
of atomic nanowires have been reported in the last decade, that demonstrate their
high reproducibility, stability and tunability [9, 13–22].

C. Dues, K. Holtgrewe, S. Sanna and F. Ziese,

Institut für Theoretische Physik and Center for Materials Research (LaMa), Justus-Liebig-Universität
Gießen, Heinrich-Buff-Ring 16, 35392 Gießen, Germany,
e-mail: [email protected]
F. Bernhardt, J. Bilk, K. Eberheim, C. Fink, N. Jöckel, B. Muscutt, F.A. Pfeiffer, M.N. Pionteck
Institut für Theoretische Physik, Justus-Liebig-Universität Gießen, Heinrich-Buff-Ring 16, 35392
Gießen, Germany

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 97

W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_6
98 M.N. Pionteck et al.

(a) (b) (c) (d)

D SC CSC

Fig. 1: Structural models of the Si(553)-Au surface. (a) Double Au strand model
(D) [16], (b) Spin-chain model (SC) [9], (c) centered spin-chain (CSC) [25]. Spin-
polarized atoms are represented in blue (spin up) and red (spin down). (d) Rehybridized
model (R) [26]. Si atoms vertically displaced from the step edge are shown in black.
Surface unit cells are highlighted.

Among the atomic chains, the Si(553)-Au surface is one of the most studied
systems [9, 15, 16, 23–27], as it was proposed to feature an antiferromagnetic spin
ordering [9] and to undergo an order-disorder type phase transition below 100 K
[28], which is argument of debate [27, 29, 30]. The Si(553) surface is tilted by 12.5◦
in [112̄] direction with respect to the (111) plane. It features 14.8 Å wide (111)
nanoterraces, separated by steps of double atomic height along the [11̄0] direction.
Evaporation of 0.48 ML Au onto this surface generates one Au chain per terrace as
shown in Fig. 1. However, the exact structure of this system has been controversely
discussed. Double Au strands and a Si honeycomb chain as shown in Fig. 1 (a), are
common to all models. However, while the Krawiec model features a planar Si step
edge, in the model proposed by Erwin et al. [9] and refined by Hafke et al. [25] [Fig. 1
(b) and (c)], the step edge is spin polarized, and every third step edge atom is shifted
downwards by 0.3 Å. Recently, Braun et al. [26] proposed a diamagnetic, 𝑠𝑝 2 + 𝑝
rehybridized model [Fig. 1 (d)], in which every third edge atom is 0.8 Å lower than
the others.
In this manuscript, we provide a thorough theoretical characterization of the
surface-localized phonon modes of the Si(553)-Au system, above and below the
phase transition temperature. Thereby, the vibrational properties of the system are
calculated from DFT-based atomistic simulations for all the proposed structural
models. This requires a computational power beyond that of local compute cluster.
The comparison with available experimental data [30] allows to assign the measured
spectral features to the calculated displacement patterns. The spectra calculated for
the double Au strand [16] and the 𝑠𝑝 2 + 𝑝 rehybridized [26] models are compatible
with the measured spectra, at 30 K and 300 K, respectively.
Dynamical properties of Si(553)-Au nanowires 99

2 Methodology

The VASP package [31, 32] implementing the DFT offers highly customizable paral-
lelization schemes to exploit the full capability of massively parallel supercomputers
such as the HPE Apollo 9000 system (Hawk) at the HLRS. In order to separate the
discussion about the computational setup and the discussion about the algorithm
performances, we divide the present section in two parts.

2.1 Computational details

DFT calculations have been performed with the Vienna ab initio Simulation Package
(VASP) [31, 32]. Projector augmented waves (PAW) potentials [33] with projectors
up to 𝑙 = 1 for H, 𝑙 = 2 for Si and 𝑙 = 3 for Au are used. A number of 1 (1s1 ), 4
(3s2 3p2 ), and 11 (5d10 6𝑠1 ) valence electrons is employed for H, Si, and Au atoms,
respectively. Plane waves up to an energy cutoff of 410 eV build up the basis for the
expansion of the electronic wave functions. The silicon surfaces are modeled with
asymmetric slabs consisting of 6 Si bilayers stacked along the [111] crystallographic
direction, the surface termination including the Au chains, and a vacuum region of
about 20 Å. H atoms saturate the dangling bonds at the opposite face of the slabs.
These atoms as well as the three lowest Si bilayers are frozen at their bulk position in
order to model the substrate, while the remaining atoms are free to relax. The atomic
positions are relaxed until the residual Hellmann–Feynman forces are lower than
0.005 eV/Å. 4×9×1 (4×27×1) Monkhorst–Pack 𝑘-point meshes [34] are employed to
perform the energy integration in the Brillouin zone of the supercell of 5×6 (5×2)
periodicity.
The calculated phonon frequencies depend strongly on the computational approach.
Indeed, they depend both directly and indirectly (through the resulting structural
differences) on the employed xc-functional. In order to estimate the dependence
of the phonon eigenvalues on the computational approach, we have calculated the
vibrational frequencies of the double strand model as proposed by Krawiec [16]
within the PBEsol [35], LDA [36] and GGA-PBE [37] approach.
According to the actual knowledge of the Si(553)-Au system, the D model is
considered as a candidate for the description of the RT structure, while the SC, CSC,
and R models are considered for the description of the LT phase. However, all the
structures are modeled within DFT at 0 K and thermal lattice expansion is neglected.
While a limited number of surface localized phonons are experimentally detected,
the frozen-phonon slab calculations lead to 246 vibrational modes for the double
Au chain model by Krawiec [16] and 738 modes for the spin-polarized [9, 25] and
rehybridized models [26]. In order to achieve a comparison with the experimental
results, we discard the phonons, whose atomic displacement vectors are localized by
less than an arbitrarily chosen threshold of 40% in the two topmost atomic layers. Yet,
the number of calculated phonon modes is still too high to allow for a frequency-based
assignment of the calculated eigenmodes to the measured spectral features. To achieve
100 M.N. Pionteck et al.

this task, we compute the Raman scattering efficiency. Raman spectra are generated
as described in detail in Refs. [38, 39]. As the electronic structure is self-consistently
calculated according to the phonon displacement, phonon induced charge transfer and
modifications of the electronic structure are both accounted for. Theoretical spectra are
constructed adding Lorentz functions centered at the calculated phonon frequencies,
with height corresponding to the calculated Raman efficiency and experimental
width. Spectra are calculated for an experimental laser frequency of 647 nm, after
consideration of the DFT underestimation of 0.5 eV of the fundamental band gap of
the Si bulk.

2.2 Computational performance

The algorithm implemented in VASP can be parallelized with MPI, OpenMP and a
hybrid of MPI and OpenMP. However, our tests show that the code performs best
with pure MPI parallelization, i.e. one MPI task for each physical core. In addition,
VASP offers several parallelization levels that can be specified by VASP-internal input
parameters. The main parallelization levels are determined by KPAR, i.e. the number
of 𝑘-points treated in parallel, and then NPAR, i.e. the number of bands treated in
parallel in the matrix diagonalization. According to our tests, these two parameters
influence massively the walltime of a calculation. In particular, the parallelization with
respect to the 𝑘-points improves the performance because it divides the problem into
independent chunks, whose demand for inter-communication is restricted. Therefore,
performance can be maximized by setting KPAR as high as possible, i.e. setting
KPAR to the number of 𝑘-points or the number of nodes, whichever is smaller. In
general, performance is improved when KPAR is a divisor of both the number of
𝑘-points and the number of nodes. On the other hand, 𝑘-point parallelization is
memory intensive, since the whole problem must be copied to each 𝑘-point group.
This results in a proportional behaviour between RAM usage and the number of
𝑘-points treated in parallel. Based on shared memory, NPAR requires less RAM,
however it influences the walltime less than KPAR. NPAR should also be a divisor of
bands and number of cores to avoid idle time. The best configuration of KPAR and
NPAR depends on the host system. Hence, we need to test the best configuration for
HPE Apollo 9000 (Hawk) and to formulate a rule for setting KPAR and NPAR on
Hawk. Then we need to check the performance of the code with the best setting.
Since the scaling increases with the problem size, we choose a model system rep-
resenting the LiNbO3 (0001) surface (𝑧-cut), modeled by a large supercell containing
512 atoms, 3072 bands and 32 𝑘-points for the performance test. Self-consistent
minimization of the electronic energy is performed, which is the main framework of
the DFT. The calculations are terminated after six minimization steps. Since the first
electronic step contains the non-scalable setup of the initial charge, it is excluded from
the timing considerations. The loop times for all other steps (step 2 to 6) are averaged
to the target value 𝑇LOOP . This procedure makes the performance test extrapolatable
to real calculations, which perform 30 to 50 electronic steps and, hence, in which
Dynamical properties of Si(553)-Au nanowires 101

the first step with non-scalable portion carries only a small weight. For the tests, we
use one MPI task for each physical core, i.e. 128 MPI tasks per node. The left panel
of Fig. 2 shows the average walltimes 𝑇LOOP of the performance tests with different
configurations of KPAR and NPAR as a function of the number of MPI tasks. The
green dots denote the average walltime of the best configuration for the respective
number of MPI tasks. The tests are in agreement with the experience that setups with
higher KPAR require less wall time. However, because of the high memory demand
of the problem, each 𝑘-point group requires two nodes at the minimum, which limits
KPAR to the half of the total number of nodes for this problem. Furthermore, the
performance tests yield the rule that NPAR is approximately best set to the square
root of the number of MPI tasks employed for each 𝑘-point group.

12
2400

2100
10

1800
8
T2048 / TLOOP

1500
TLOOP [s]

1200 6

900
4

600

2
300

0 0
0 2048 4096 8192 12288 16384 0 2048 4096 8192 12288 16384
#cores #cores

Fig. 2: Results of the performance test; left panel: test results for different paralleliza-
tion setups, right panel: speed-up of the best guess and Amdahl fits (“Amdahl fit 1”:
fit for 2048, 4094, 8192, 12288 and 16384 cores; “Amdahl fit 2”: fit for 2048, 4094
and 8192 cores).

Subsequently, we evaluate the scalability of the VASP code on Hawk. As we

use 16 nodes at 128 cores at the minimum, we calculate the speed-up as the ration
between the loop walltime for 2048 and 𝑁 employed cores
𝑇2048
𝑆2048 (𝑁) = . (1)
𝑇LOOP (𝑁)
The right panel of Fig. 2 shows the speed-up as a function of the number of employed
cores. The green dots again indicate the best configuration of KPAR and NPAR for
the respective number of cores and are fitted to the speed-up by an analytic function.
This function is derived from Amdahl’s law, which is defined as the speed-up with
102 M.N. Pionteck et al.

respect to one MPI task

𝑇1 1
𝑆(𝑁) = = , (2)
𝑇LOOP (𝑁) 𝑠 + 𝑁𝑝

where 𝑠, 𝑝 ∈ (0, 1) with 𝑝 = 𝑠 − 1 denote the serial and parallel portions of the code.
Thus the speed-up in (1) can be fit with the serial part 𝑠 as parameter in accordance
with Amdahl’s law
1−𝑠
𝑇2048 𝑇2048 𝑇1 𝑆(𝑁) 𝑠 + 2048
𝑆2048 (𝑁) = = = = . (3)
𝑇LOOP (𝑁) 𝑇1 𝑇LOOP (𝑁) 𝑆(2048) 𝑠 + 1−𝑠
𝑁

At first, the fit is done for the speed-up of the best configurations at 2048, 4094,
8192, 12288 and 16384 cores and plotted as a green line labelled “Amdahl fit 1” in the
right panel of Fig. 2. The obtained serial portion of “Amdahl fit 1” is at 𝑠 = 0.0000443.
This yields a scale efficiency of 73 % for 8192 cores, i.e. 64 nodes. We observe that
the speed-up flattens abruptly for more than 8192 cores. This can be explained by our
test conditions. Since we use 32 𝑘-points for our tests, the parallelization of 𝑘-points
is limited to 32 groups. Up to 64 nodes, KPAR can be set to half the number of nodes
(the maximum due to memory constraints). Consequently, the problem can scale
with 2048, 4094 and 8192 cores. However, for more than 8192 cores, the number of
𝑘-point groups cannot be increased further, resulting in an abrupt flattening of the
speed-up. For calculations featuring a larger number of 𝑘-points, a linear speedup is
expected even for a larger number of cores.
To investigate the speed-up in the domain with optimal KPAR-parallelization, we
fit also Eq. (3) with the best configurations at 2048, 4094 and 8192 cores as the blue
line labeled “Amdahl fit 2” in the right panel of Fig. 2. The serial portion drops then to
𝑠 = 0.0000042 translating to a scaling efficiency of 96 % for 8192 cores. This shows
that we are able to scale our code efficiently if the necessary memory is available
and enough 𝑘-points can be used. Thus, optical calculations with a high number of
𝑘-points are particularly well suited to exploit the full capability of Hawk.
According to our experience, the queue time is significantly longer for jobs longer
than 4 hours. Therefore, we have optimized our jobs for walltimes of maximum 4
hours.

3 Results

3.1 Phonon modes at room temperature

The vibrational properties of the Si(553)-Au system at 300 K are computed with the
structural model proposed by Krawiec [16] [see Fig. 1 (a)]. Although, according to
recent investigations, the Si(553)-Au surface at 300 K oscillates between the D, R, and
Dynamical properties of Si(553)-Au nanowires 103

SC/CSC phases, the system is for the vast majority of the time in the D configuration
[27], which is therefore employed to describe the high temperature phase of the
Si(553)-Au surface.
Among the calculated phonon modes a considerable fraction is Raman silent. The
calculated phonon spectra shown in Fig. 3 closely reproduce the measured spectra
[30]. Solely the peaks of moderate intensity predicted at about 150 cm−1 in the
crossed polarization are not experimentally observed. However, the assignment of
phonon modes in this spectral region is difficult because of the overlap with broad
and intense bulk phonons. Generally, both the frequency and the relative intensity of
the spectral features are in satisfactorily agreement with the experiment, although the
most intense vibrational signatures in the (yx) crossed configuration are somewhat
red shifted within PBEsol in comparison with the experiment. The good agreement

Table 1: Raman frequencies (in cm−1 ) measured at 300 K [30] and calculated
(0 K frozen phonon calculations performed with the D model). PBEsol calculated
frequencies are listed (Theo.), along with the highest and lowest frequency calculated
with other XC-functionals. Char. and Loc. indicate whether the phonon has Au or Si
character, and the surface localization of the atomic displacement vectors, respectively.
Modes with calculated Raman efficiency below 1% of the main peak are not listed.

z(yy)-z z(yx)-z
Exp. Theo. Theo. Min-Max Char. Loc. Exp. Theo. Theo. Min-Max Char. Loc.

37.5±0.3 35.4 35.1−36.4 Au 75% 43.3±0.9 44.2 44.2−46.1 Au 58%

 48.9 47.7−52.9 Au+Si 93%
48.3±0.1 49.7±0.4 52.7 52.7−59.8 Au 91%
52.7 52.7−59.8 Au 91%
60.1±0.1 64.9 64.9−66.5 Au+Si 68% 43.3±0.9 44.2 44.2−46.1 Au 58%
68.4±0.3 69.8 66.1−72.9 Au 92% 49.7±0.4 52.7 52.7−59.8 Au 91%
85.5±0.1 84.2 84.1−86.2 Au+Si 88% 61.6±0.5 62.4 61.7−64.6 Au 91%
100.8±0.2 109.4 109.4−113.1 Si 74% 73.9±1.1 65.4 65.4−67.6 Si 47%
121.3±0.4 122.0 122.0−129.4 Si 55% 87.4±0.9 79.7 79.7−86.2 Si 57%
 82.1 82.1−86.1 Au 77%
134.1±1.4 131.9 131.8−141.1 Si 50% 94.0±0.2
84.4 84.4−94.9 Au 77%
147.4±0.7 140.2 140.2−144.0 Si 47% 101.7±0.2 96.8 96.2−98.8 Au 72%
158.4±4.7 165.1 165.1−167.5 Si 55% 113.5±0.5 111.4 111.4−115.3 Si 52%
172.6±1.9 169.8 169.7−171.3 Si 51%
392.2±0.1 386.2 379.8−388.3 Si 40%
413.2±0.1 411.5 410.1−415.7 Si 50%
 104 M.N. Pionteck et al.

(yy)
1.0 123.2
84.2 386.2
35.4
48.9
52.7 64.9 411.5
69.8 109.4
Intensity (a.u.)

(yx)
Lorentz function
Sum
82.1
65.4
52.7

44.2

50 100 150 200 400

-1
Raman shift (cm )

Fig. 3: Raman spectra of the Si(553)-Au surface calculated within DFT-PBEsol for
the (yy) and (yx) polarization with the structural model –D– by Krawiec et al. [16].

Fig. 4: Schematic representation of the displacement pattern of selected phonon

modes that are Raman active in the crossed polarization according to the double Au
strand model [16] of the RT phase.

between measured and calculated spectra allows to assign the calculated phonon
displacement patterns on the basis of energy, symmetry, and Raman intensity. The
result of this procedure is shown in Tab. 1.
Displacement patterns of Raman active modes are represented in Figs. 4−5, in
which upwards and downwards vertical displacements are represented by dots and
crosses, respectively. The calculated displacement patterns reflect the polarization
dependence of the Raman spectra. The modes detectable in parallel polarization
Dynamical properties of Si(553)-Au nanowires 105

are symmetric with respect to the mirror symmetry plane perpendicular to the Au
wires shown in Fig. 4 (d) and Fig. 5 (d). The modes that are Raman active in crossed
polarization do not possess this symmetry. This relationship between selection rules
and phonon symmetry is valid, assuming deformation potentials as the origin of the
Raman scattering.
More in detail, in the crossed configuration six modes are clearly visible, along
with low intensity signatures. All the corresponding displacement patterns (shown
in Fig. 4) break the local symmetry of the terrace, as expected for modes that are
Raman active in crossed polarization. Further modes of moderate intensities at higher
frequencies are predicted but not experimentally observed (due to the overlap with
bulk resonances) and will not be discussed in detail.
In parallel polarization, more modes are Raman active. While at lower frequencies
the Au related modes dominate, above 100 cm−1 only Si modes are found. This is
expected, due to the much higher mass of the Au atoms with respect to Si. Among
this modes, the overlap of two close phonons at 48.9 and 52.7 cm−1 , with similar
displacement pattern, represented in Fig. 5 (b), is experimentally observed as a single
peak measured at 48.3 cm−1 . Both modes perform the same seesaw movement within
the Au-chain, however, the first one features a more pronounced in-phase movement
of the step edge, which increases the phonon effective mass and lowers its frequency.
These two modes break the local terrace symmetry and should not be Raman
active in parallel polarization, assuming deformation potentials as the origin of the
Raman scattering. However, a further Raman mechanism −scattering at charge density
fluctuations− must be considered for these modes. Charge fluctuations between the
step edge and the Au chain occur when atomic displacements modify the Au-Au
bond length and the relative height of the step edge atoms, as revealed by previous
studies [20, 21, 27]. Scattering at charge density fluctuations, which is well known,

Fig. 5: Schematic representation of the displacement pattern of phonon modes that

are Raman active in the parallel polarization according to the double Au strand model
[16] of the RT phase.
106 M.N. Pionteck et al.

e.g., for highly doped semiconductors [40], is traced to the modifications of the
adsorption edge due to the phonon-induced charge redistribution and had not been
previously observed for quasi 1D systems. This mechanism contributes to the Raman
scattering only in parallel configuration, and occurs for all modes associated with
strong charge fluctuations. Yet, it is generally difficult to distinguish and quantify the
relative contributions of scattering at charge density fluctuations and deformation
potentials to the total Raman intensity, as both occur in parallel polarization. However,
for the mode in Fig. 5 (b), no deformation potential scattering can take place in
parallel polarization, due to the phonon symmetry. Therefore only scattering at charge
density fluctuations can be responsible for the Raman intensity of this mode.
The close agreement between the theoretical and experimental results strongly
suggests that the double Au strand model [16] correctly describes the Si(553)-Au
system.

3.2 Phonon modes below and above the phase transition

The Si(553)-Au system is supposed to undergo an order-disorder type structural

transition starting below 100 K [27]. According to the actual knowledge of the system,
the double Au strand model of (5×2) periodicity [16] describes the high symmetric
RT phase, while the (centered) spin-chain [9, 25] and rehybridized [26] models of
(5×6) periodicity have been proposed for the description of the lower symmetric
LT phase. Between RT and LT the morphology of the step edge of the Si terraces
fluctuates among the configurations of the two phases, establishing an interplay (via
charge transfer) with the Au chain that continuously enhances the chain dimerization
from RT to LT [27].
The different symmetry of the two phases must be mirrored in their vibrational
properties. In particular, modes associated with the Au chain are expected to modify
their frequency from RT to LT, while modes localized at the Si step edge cannot exist
in the same form in the two phases. To verify this assumption, we have calculated the
vibrational properties of the Si(553)-Au system at LT with all the structural models
that have been proposed in the literature for the description of this phase, namely
the spin-chain [9] and centered spin-chain model [25], as well as the rehybridized
model [26]. Unfortunately, the calculation of the Raman scattering efficiency for the
LT structural models is an exceptionally demanding task, due to both system size
and number of phonons (396 atoms and 738 vibrational modes for each structure).
However, the knowledge of the calculated Raman frequencies and displacement
patterns can still be used to interpret the experimental data.
The comparison of the displacement patterns and corresponding eigenfrequencies
calculated with the RT and LT models allows to discriminate three categories of
phonons: modes which are common to the RT and LT phases, modes that occur
both in the RT and in the LT phase, however with a different frequency, and modes
that exist either only in the RT or only in the LT structure. Most of the frequencies
Dynamical properties of Si(553)-Au nanowires 107

Fig. 6: Schematic representation of the displacement pattern of the dimerization mode

and of the honeycomb translation mode. The displacement patterns are calculated
according to the rehybridized model[26] of the LT phase.

calculated with the RT model can be identified in all three candidate models within a
few cm−1 . However, important exceptions are found, corresponding exactly to the
measured LT-RT differences [30].
The Au dimerization mode shown in Fig. 6 (a), exists in both the RT and LT
structure, yet at different frequencies. This mode is predicted at 18.8±5 cm−1 with
the RT model, and is thus not experimentally accessible. However, it becomes much
harder (42.0 cm−1 ) within the rehybridized model, which features a more pronounced
dimerization. As this mode shortens the Au-Au bond length, it requires more energy
for a strongly dimerized Au chain. This suggests that the peak observed at LT at
40.5 cm−1 is the overlap of a weakly temperature dependent mode at 37.5 cm−1 and
the dimerization mode at about 42 cm−1 . This would both explain the observed
modifications in intensity and frequency, and also be in agreement with a previous
interpretation [27].
The second exception is represented by a further mode with strong frequency shift,
the mode calculated at 69.8 cm−1 shown in Fig. 5 (d). Similarly to the dimerization
mode, this mode shortens the Au-Au bond length and becomes much harder by about
8 cm−1 for the rehybridized model.
Another difference between the HT and LT phase is the mode associated to the
displacement pattern calculated with the rehybridized model and displayed in Fig. 6
(b). This mode is not existent in the RT structure, but closely related to the chain
translation modes predicted at 65.4 cm−1 and at 79.7 cm−1 [Fig. 4 (d) and (e)] at RT.
This mode is a translation of the HC chain, which is strongly hindered by the LT
modification of the step edge pinning the outer atoms and therefore making this mode
much harder.
Finally, the mode measured at 413 cm−1 and associated to the theoretically
predicted step edge mode at 411.5 cm−1 [see Fig. 5 (h)] is only a phonon mode of the
RT phase. Due to the different symmetry of the Si step edge in the structural models
associated to the LT and RT phases, this lattice vibration has no low temperature
counterpart. In the rehybridized model, this mode is decomposed into local vibrations
of the step edge, as previously pointed out by Braun et al. [27].
108 M.N. Pionteck et al.

Table 2: Raman frequencies of selected modes discussed in the text calculated within
PBEsol according to different structural models representing the low temperature
structure.

Exp. R SC CSC

40.5±0.2 42.0 8.7 16.7

82.7±1.4 77.3 69.4 70.7
99.1±1.5 91.0 65.0 64.8

To summarize, the rehybridized model can well explain the observed temperature
shifts. On the contrary, in the (centered) spin-chain model [9, 25] the dimerization
is not as pronounced as in the rehybridized model [26] and therefore the frequency
shifts with respect to the high temperature model cannot be reproduced. For example,
the dimerization mode [see Fig. 6 (a), and Tab. 2] is predicted by DFT-PBEsol at
8.7 and 16.7 cm−1 , for the spin-chain and centered spin-chain models, respectively.
This value is far from the value of 42.0 cm−1 calculated with the rehybridized model
and assigned to the peak measured at 40.5 cm−1 . Similarly, the mode that couples
the Au dimers [see Fig. 5 (d)] calculated at 68.9 cm−1 for the RT structure, does not
significantly shift at LT in the SC and CSC model (69.4 and 70.7 cm−1 , respectively)
and cannot explain the spectral feature measured at 82.7 cm−1 . Thus, the comparison
of the calculated vibrational properties with the measured spectra yields a strong
argument for the rehybridized model for the description of the low-temperature phase.

4 Conclusions

The dynamical properties of the Si(553)-Au surface have been studied by DFT. The
atomistic calculations satisfactorily reproduced the measured spectra, if the double
Au strand model and the rehybridized model are used for the simulation of the RT
and LT phase, respectively. We propose an assignment of the mode specific phonon
displacement patterns to the measured spectral signatures. Furthermore, a pronounced
temperature dependence of different modes in the LT phase and in the RT phase
is interpreted as a signature of a structural phase transition. As a general feature,
modes that modify the Au-Au bond length and modes localized at the Si step edge
are significantly influenced by the phase transition.
Modes which involve the Si step edge atoms either disappear or are shifted to much
higher energies at LT. In particular, the mode predicted at 411.5 cm−1 , associated to
transversal vibrations of the Si step edge, disappears at LT, where thermal fluctuations
are frozen and the surface structure shows a higher order.
Dynamical properties of Si(553)-Au nanowires 109

The phonon activated charge transfer between the Au chain and the Si step edge,
which is responsible for the observed order-disorder phase transition [27], leads to
Raman scattering by charge density fluctuations and allows for a direct observation
of the strong coupling between electronic and phononic systems on the Si(553)-Au
surface.
The knowledge of the vibrational properties of the Si(553)-Au surface is moreover
a prerequisite for future investigations of the Si(553)-Au system modified by atomic
deposition or doping, which is often performed to control the coupling with higher
dimensions.

Acknowledgements We gratefully acknowledge financial support from the Deutsche Forschungsge-

meinschaft in the research unit FOR1700 (project SA 1948/1-2) FOR2824, project SA 1948/2-1 and
FOR5044, (project SA 1948/3-1). The Höchstleistungrechenzentrum Stuttgart (HLRS) is gratefully
acknowledged for grants of high-performance computer time.

References

1. S. Kagoshima, H. Nagasawa, T. Sambongi, Electron Correlation and Magnetism in Narrow-Band

Systems, Springer Series in Solid-State Sciences, vol. 72 (Springer, Berlin, Heidelberg, 1988).
DOI 10.1007/978-3-642-83179-9. URL http://dx.doi.org/10.1007/978-3-642-8
1639-0{_}15{%}0Ahttp://link.springer.com/10.1007/978-3-642-83179-9http:
//link.springer.com/10.1007/978-3-642-83179-9
2. G. Grüner, Rev. Mod. Phys. 60(4), 1129 (1988). DOI 10.1103/RevModPhys.60.1129. URL
https://link.aps.org/doi/10.1103/RevModPhys.60.1129
3. J.M. Luttinger, J. Math. Phys. 4(9), 1154 (1963). DOI 10.1063/1.1704046. URL http:
//link.aip.org/link/JMAPAQ/v4/i9/p1154/s1{&}Agg=doihttp://scitation.ai
p.org/content/aip/journal/jmp/4/9/10.1063/1.1704046
4. K. Schönhammer, Luttinger Liquids: basic concepts (Springer Netherlands, Dordrecht, 2004),
Physics and Chemistry of Materials with Low-Dimensional Structures, vol. 25, chap. 4, pp.
93–136. DOI 10.1007/978-1-4020-3463-3_4. URL http://www.springerlink.com/c
ontent/p6ru1315m33g7685http://link.springer.com/10.1007/978-1-4020-346
3-3{_}4
5. T. Giamarchi, Quantum Physics in One Dimension (Clarendon Press, Oxford, 2007)
6. C. Zeng, P. Kent, T.H. Kim, A.P. Li, H.H. Weitering, Nature Materials 7(7), 539 (2008). URL
https://www.nature.com/articles/nmat2209
7. P.C. Snijders, H.H. Weitering, Rev. Mod. Phys. 82, 307 (2010). DOI 10.1103/RevModPhys.82.
307
8. H. Weitering, Nat. Phys. 7(10), 744 (2011). DOI 10.1038/nphys2074. URL http://dx.doi
.org/10.1038/nphys2074
9. S.C. Erwin, F.J. Himpsel, Nat. Commun. 1(5), 58 (2010). DOI 10.1038/ncomms1056. URL
http://www.ncbi.nlm.nih.gov/pubmed/20975712
10. J. Aulbach, J. Schäfer, S.C. Erwin, S. Meyer, C. Loho, J. Settelein, R. Claessen, Phys. Rev.
Lett. 111(13), 137203 (2013). DOI 10.1103/PhysRevLett.111.137203. URL http:
//link.aps.org/doi/10.1103/PhysRevLett.111.137203
11. C. Brand, H. Pfnür, G. Landolt, S. Muff, J.H. Dil, T. Das, C. Tegenkamp, Nat. Commun. 6,
8118 (2015). DOI 10.1038/ncomms9118
12. K. Holtgrewe, S. Appelfeller, M. Franz, M. Dähne, S. Sanna, Phys. Rev. B 99, 214104 (2019).
DOI 10.1103/PhysRevB.99.214104. URL https://link.aps.org/doi/10.1103/PhysR
evB.99.214104
110 M.N. Pionteck et al.

13. J.N. Crain, J. McChesney, F. Zheng, M.C. Gallagher, P. Snijders, M. Bissen, C. Gundelach, S.C.
Erwin, F.J. Himpsel, Phys. Rev. B 69(12), 125401 (2004). DOI 10.1103/PhysRevB.69.125401
14. I. Barke, F. Zheng, T.K. Rügheimer, F.J. Himpsel, Phys. Rev. Lett. 97(22), 226405 (2006). DOI
10.1103/PhysRevLett.97.226405. URL http://link.aps.org/doi/10.1103/PhysRevLe
tt.97.226405
15. J. Aulbach, S.C. Erwin, R. Claessen, J. Schäfer, Nano Lett. 16(4), 2698 (2016). DOI
10.1021/acs.nanolett.6b00354. URL http://pubs.acs.org/doi/abs/10.1021/acs.nan
olett.6b00354
16. M. Krawiec, Phys. Rev. B 81(11), 115436 (2010). DOI 10.1103/PhysRevB.81.115436. URL
http://link.aps.org/doi/10.1103/PhysRevB.81.115436
17. I. Miccoli, F. Edler, H. Pfnür, S. Appelfeller, M. Dähne, K. Holtgrewe, S. Sanna, W.G. Schmidt,
C. Tegenkamp, Phys. Rev. B 93, 125412 (2016). DOI 10.1103/PhysRevB.93.125412. URL
https://link.aps.org/doi/10.1103/PhysRevB.93.125412
18. C. Braun, C. Hogan, S. Chandola, N. Esser, S. Sanna, W.G. Schmidt, Phys. Rev. Materials 1,
055002 (2017). DOI 10.1103/PhysRevMaterials.1.055002
19. F. Edler, I. Miccoli, J.P. Stöckmann, H. Pfnür, C. Braun, S. Neufeld, S. Sanna, W.G. Schmidt,
C. Tegenkamp, Phys. Rev. B 95(12), 125409 (2017). DOI 10.1103/PhysRevB.95.125409. URL
http://link.aps.org/doi/10.1103/PhysRevB.95.125409
20. Z. Mamiyev, S. Sanna, C. Lichtenstein, T. and. Tegenkamp, H. Pfnür, Phys. Rev. B 98(24),
245414 (2018). DOI 10.1103/PhysRevB.98.245414
21. C. Hogan, E. Speiser, S. Chandola, S. Suchkova, J. Aulbach, J. Schäfer, S. Meyer, R. Claessen,
N. Esser, Phys. Rev. Lett. 120, 166801 (2018). DOI 10.1103/PhysRevLett.120.166801. URL
https://link.aps.org/doi/10.1103/PhysRevLett.120.166801
22. M. Tzschoppe, C. Huck, F. Hötzel, B. Günther, Z. Mamiyev, A. Butkevich, C. Ulrich, L. Gade,
A. Pucci, J. Phys. Condens. Matter 31(19), 195001 (2018). DOI 10.1088/1361-648X/ab0710
23. S. Polei, P.C. Snijders, S.C. Erwin, F.J. Himpsel, K.H. Meiwes-Broer, I. Barke, Phys. Rev.
Lett. 111(15), 156801 (2013). DOI 10.1103/PhysRevLett.111.156801. URL http:
//link.aps.org/doi/10.1103/PhysRevLett.111.156801
24. S. Polei, P.C. Snijders, K.H. Meiwes-Broer, I. Barke, Phys. Rev. B 89(20), 205420 (2014). DOI
10.1103/PhysRevB.89.205420. URL http://link.aps.org/doi/10.1103/PhysRevB.89
.205420
25. B. Hafke, T. Frigge, T. Witte, B. Krenzer, J. Aulbach, J. Schäfer, R. Claessen, S.C. Erwin,
M. Horn-von Hoegen, Phys. Rev. B 94(16), 161403 (2016). DOI 10.1103/PhysRevB.94.161403.
URL http://link.aps.org/doi/10.1103/PhysRevB.94.161403
26. C. Braun, U. Gerstmann, W.G. Schmidt, Phys. Rev. B 98, 121402 (2018). DOI 10.1103/PhysRe
vB.98.121402. URL https://journals.aps.org/prb/pdf/10.1103/PhysRevB.98.1
21402
27. C. Braun, S. Neufeld, U. Gerstmann, S. Sanna, J. Plaickner, E. Speiser, N. Esser, W.G.
Schmidt, Phys. Rev. Lett. 124, 146802 (2020). DOI 10.1103/PhysRevLett.124.146802. URL
https://link.aps.org/doi/10.1103/PhysRevLett.124.146802
28. F. Edler, I. Miccoli, H. Pfnür, C. Tegenkamp, Phys. Rev. B 100, 045419 (2019). DOI
10.1103/PhysRevB.100.045419. URL https://link.aps.org/doi/10.1103/PhysRevB.
100.045419
29. B. Hafke, C. Brand, T. Witte, B. Sothmann, M. Horn-von Hoegen, S.C. Erwin, Phys. Rev. Lett.
124, 016102 (2020). DOI 10.1103/PhysRevLett.124.016102. URL https://link.aps.org
/doi/10.1103/PhysRevLett.124.016102
30. J. Plaickner, E. Speiser, C. Braun, W.G. Schmidt, N. Esser, S. Sanna, Phys. Rev. B 103, 115441
(2021). DOI 10.1103/PhysRevB.103.115441. URL https://link.aps.org/doi/10.1103
/PhysRevB.103.115441
31. G. Kresse, J. Furthmüller, Computational Materials Science 6(1), 15 (1996). DOI https:
//doi.org/10.1016/0927-0256(96)00008-0
32. G. Kresse, J. Furthmüller, Phys. Rev. B 54, 11169 (1996). DOI 10.1103/PhysRevB.54.11169
33. P.E. Blöchl, Phys. Rev. B 50, 17953 (1994). DOI 10.1103/PhysRevB.50.17953
34. H.J. Monkhorst, J.D. Pack, Phys. Rev. B 13, 5188 (1976). DOI 10.1103/PhysRevB.13.5188
Dynamical properties of Si(553)-Au nanowires 111

35. J.P. Perdew, A. Ruzsinszky, G.I. Csonka, O.A. Vydrov, G.E. Scuseria, L.A. Constantin, X. Zhou,
K. Burke, Phys. Rev. Lett. 100, 136406 (2008). DOI 10.1103/PhysRevLett.100.136406. URL
https://link.aps.org/doi/10.1103/PhysRevLett.100.136406
36. D.M. Ceperley, B.J. Alder, Phys. Rev. Lett. 45, 566 (1980). DOI 10.1103/PhysRevLett.45.566.
URL https://link.aps.org/doi/10.1103/PhysRevLett.45.566
37. J.P. Perdew, K. Burke, M. Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996). DOI 10.1103/PhysRevL
ett.77.3865. URL https://link.aps.org/doi/10.1103/PhysRevLett.77.3865
38. E. Speiser, N. Esser, B. Halbig, J. Geurts, W.G. Schmidt, S. Sanna, Surface Science Reports
75(1), 100480 (2020). DOI https://doi.org/10.1016/j.surfrep.2020.100480. URL http:
//www.sciencedirect.com/science/article/pii/S0167572920300017
39. S. Sanna, S. Neufeld, M. Rüsing, G. Berth, A. Zrenner, W.G. Schmidt, Phys. Rev. B 91, 224302
(2015). DOI 10.1103/PhysRevB.91.224302. URL https://link.aps.org/doi/10.1103
/PhysRevB.91.224302
40. G. Abstreiter, M. Cardona, A. Pinzcuk, Light Scattering by free carrier excitations in semi-
conductors, in: Light scattering in Solids IV: Electronic Scattering, Spin Effects, SERS and
Morphic Effects, Topics in Applied Physics, vol. 54 (Springer Verlag, 1984)
Reactivity of organic molecules on semiconductor
surfaces revealed by density functional theory

Fabian Pieck, Jan-Niclas Luy, Florian Kreuter, Badal Mondal and Ralf Tonner-Zech

Abstract The present report discusses the reactivity and properties of organic
molecules on mainly semiconductor surfaces by means of density functional theory.
For pyrazine on Ge(001), a benzylazide on Si(001) and a cyclooctyne derivate on
Si(001) adsorption modes and possible reaction paths are presented. The charge
transfer effect between the inorganic and organic interface is presented with the
example of corroles on Ag(111). Approaches towards more realistic and efficient
models are discussed in the third chapter, while the scaling of large simulations on
the resources of the High-Performance Computing Center Stuttgart is addressed in a
final chapter.

1 Introduction

The increasing demand for consumer electronics is driving the development of

cheaper and increasingly powerful microelectronic devices. To satisfy these demands
device miniaturization is a core strategy, new device architectures are developed and
even the use of germanium as alternative elemental semiconductor material, besides
silicon is promoted [1].
To enlarge the property and application range of semiconducting devices the
formation of organic-semiconductor and metal-organic interfaces needs to be under-
stood. The basic idea for the first approach is to combine the advantages of the mature
silicon-based technology with the chemical richness and diverse electronic properties
of organic compounds [43, 71, 84]. While Si(001) is still the technologically most
important semiconductor surface for the organic functionalization [33, 46, 53, 58],
also the Ge(001) surface is considered [7, 37]. Both surfaces have in common the
formation of surface dimers, that act as the fundamental reactive unit of the surface

Ralf Tonner-Zech
Wilhelm-Ostwald-Institut für Physikalische und Theoretische Chemie, Linnéstr. 2, 04103 Leipzig,
Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 113
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_7
114 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

[7, 37, 75]. Since these pristine semiconductor surfaces are generally highly reactive
[78] the critical step is the attachment of the first organic layer to obtain uniform
and defect-free interfaces. A promising approach is the use of tailored bifunctional
molecular building blocks [17]. These building blocks can be attached to the substrate
by 1,3-dipolar [36] and [2 + 2] cycloaddition reactions [44, 52], while a second
functional group is reserved for the attachment of further organic molecules.
To obtain well-defined structures, bifunctional building blocks showing orthogonal
reactivity for attachment and further functionalization are required. So far, only a
small number of bifunctional molecules, often restricted to a fixed combination of the
two functional groups, are promising for the attachment on Si(001) [14, 28, 66, 68, 83].
Here, cyclooctyne has been found to be an excellent platform molecule [44, 52].
While the attachment to the Si(001) surface is accomplished by the strained triple
bond, additional organic layers can be grown on the first layer [38] when bifunctional
anchor molecules, such as ethynyl-cyclopropyl-cyclooctyne [33] or methyl-enol-
ether-functionalized cyclooctyne [19] are used. The approach of combining organic
molecules with semiconducting surfaces opens a huge range of possible applications
in the area of biosensors [12, 74, 76], microelectronics [2, 22, 25, 56, 74, 77, 79],
batteries [8], organic thin film growth [10, 62, 84], organic solar cells and organic
light emitting diodes [45, 85].
Enabling the organic functionalization in these applications, requires an understand-
ing of the fundamental reactivity of organic molecules on semiconducting surfaces
and efficient screening approaches for potential new molecules. Here, computational
approaches by means of density functional theory (DFT) can significantly contribute.
Still, modelling of organic layers on semiconducting surfaces remains challenging
due to large system sizes often with large conformational freedom, various types of
bonding interactions involved and complex chemical environment including solvents
and defects. Therefore, developing small and efficient models is a desirable step.
The next sections address various aspects of our theoretical work regarding
the functionalization of semiconductor surfaces. In the first section we answer
fundamental questions concerning the reactivity and selectivity of organic molecules
on semiconducting surfaces at the examples of pyrazine on Ge(001) [63] and benzyl-
azide [24] as well as a cyclooctyne derivate [19] on Si(001). For a full understanding
of devices, we also need an in-depth understanding of the metal-organic interfaces that
appear for example when electrodes are attached to organic semiconductors. We are
thus also extending our investigations in this direction. With the electronic properties
of a corrole on Ag(111) [86], a metal-organic interface is studied in the second
section to understand charge transfer effects, which also occur on semiconducting
surfaces. The third section discusses developments towards more realistic and efficient
surface models [38, 39], while in a final section our scaling benchmark for HAWK is
presented.
Reactivity of organic molecules on semiconductor surfaces 115

2 Reactivity and selectivity of organic molecules on Si(001) and

Ge(001)

2.1 Pyrazine on Ge(001)

The behavior of heteroaromatic molecules on semiconductors is of particular interest

to understand the role of aromaticity on surface reactions, which can serve as the
guiding principle for surface modification [29, 72, 73]. While pyrazine was already
studied on a silicon surface [30, 31, 34, 48, 49, 67, 82], only the related molecule
pyridine was studied on a germanium surface [5] and no knowledge of pyrazine on
Ge(001) was available.
Figure 1 shows the most stable adsorption structures for pyrazine on Ge(001).
Interestingly, structures where the molecule bonds to the surface with one or two
nitrogen atoms (1N, 2N, Figure 1a and 1b) as well as with two carbon atoms (2C,
Figure 1c) are observed. In terms of bonding energies 1N and 2N are with −120
and −115 kJ·mol−1 favored over 2C with −82 kJ·mol−1 . Several organic compounds,
e.g. benzene, are known to be able to form covalent bonds with carbon atoms to
semiconductor surfaces [20,26]. Additional structures were found, but with a bonding
energy of up to −13 kJ·mol−1 they are significantly less stable. In addition, 1N, 2N
and 2C adsorption modes could be verified by comparing calculated XPS peaks with
experimental results [63].
Besides the most stable adsorption structure, also their occurrence dependent on
temperature or coverage is of interest. To derive this, adsorption paths and reaction
paths for the conversion of 2N to 1N and 1N to 2C were calculated (Figure 2). In
agreement with experimental data, the pyrazine molecule is initially able to adsorb

Fig. 1 Top and side view of

the optimized structures for the
three most stable adsorption
structures of pyrazine on
Ge(001). 1N (a), 2N (b) and
2C (c). Bond length and out of
plane displacement (Δ𝑅) in Å.
Reprint from reference [63].
116 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

to the 1N and 2N structure. While the reaction barrier between 1N and 2N is with
7 kJ·mol−1 negligible, the barrier between 1N and 2C is with 89 kJ·mol−1 more
pronounced. Consequently, pyrazine can constantly interconvert between 1N and 2N
even at lower temperatures (180 K), while a larger fraction of molecules in the 2C
state is only observed at elevated temperatures (423 K) [63].
For the study of multiadsorption effects, the surface model had to be doubled
in size. Several arrangements of pyrazine in the 1N and 2N mode were calculated.
Based on the change in bonding energy, it was visible that high surface coverage
can only be achieved with adsorbates in the 1N structure, while the intermolecular
repulsion for molecules in the 2N structure is too large to form densely packed layers.
To deepen our understanding of the reactivity of pyrazine, the loss of aromaticity
was further studied by means of nuclear-independent chemical shifts (NICS) and the
harmonic oscillator model of aromaticity (HOMA), while the bonding situation was
determined by the calculation of partial charges and the application of the energy
decomposition method for extended systems (pEDA). Here, in agreement with the
out-of-plane displacement (Δ𝑅) a significant loss of aromaticity could be observed
for the 2N and 2C structure [63]. Analysis of the partial charges in combination with
the pEDA could identify the N-Ge bonds in the 1N and 2N structure as dative bonds

Fig. 2: Reaction path from 1N to 2N and 1N to 2C with the transition states

TS(1N→2N) and TS(1N→2C) and a shallow intermediate (SI) on the energy plateau
between 1N and TS(1N→2C). Stated structures are indicated along the path with
colored circles. In addition, a blue circle indicates that a barrierless adsorption for
pyrazine from the gas phase to that structure was found. The bond length of the
formed bonds is stated in Å. Bonding energy are given relative to the 1N structure.
Reprint from reference [63].
Reactivity of organic molecules on semiconductor surfaces 117

resulting in a minor charge transfer to the surface. The 2C structure showed a stronger
surface to molecule donation and could be identified as the product of an inverse
electron demand Diels–Alder reaction.
Overall, these results showed that pyrazine adsorbs on Ge(001) forming a mix of
reaction products. We could explain the experimentally observed temperature and
coverage dependence, while thoroughly analyzing the bonding situation.

2.2 Methyl-substituted benzylazide on Si(001)

For the organic functionalization of silicon, molecules easily reacting with the first
organic layer are needed. Here, azides are often used mainly due to reactions such as
alkyne-azide couplings used in click chemistry [32]. While the intention is to use
azides as a building block for the second organic layer, also their behavior towards a
clean Si(001) surface has to be understood since it poses an important side reaction.
Studies of the unsubstituted benzylazide on Si(001) revealed an intermediate state
involving the intact molecule and the elimination of N2 to reach the final state with
the molecule attached to silicon via the remaining nitrogen atom [6, 36].
For the methyl-substituted benzylazide the computed adsorption and reaction
profile on Si(001) is shown in Figure 3. The molecule initially adsorbs without a barrier
into a stable [3 + 2] cycloaddition product with a bonding energy of −263 kJ·mol−1 .

Fig. 3: Adsorption and reaction profile of the methyl-substituted benzylazide on

Si(001). Gibbs energies (electronic energies) relative to the free molecule and surface.
HSE06 energies with thermodynamic corrections based on PBE-D3. Reprint from
reference [24].
118 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

The structure consists of a five-membered ring with nearly symmetric Si-N bonds.
This intermediate state (IM) can further decompose via N-N bond cleavage and a
high barrier of Δ𝐺 ‡ = 142 kJ·mol−1 toward the first adsorption configuration (C1).
This reaction is thermodynamically favored by −83 kJ·mol−1 and driven by the loss
of N2 . A second structure (C2) is accessible via a low-lying transition state structure
with a barrier of 29 kJ·mol−1 . C2 is more stable by 50 kJ·mol−1 in comparison to
C1. The gain in stability stems from an additional interaction of the benzyl moiety
with the silicon dimer row, similar to the butterfly structure observed for benzene on
Si(001) [65, 69].
Based on calculated STM images and experimental STM and XPS data, the IM
state could be identified at low temperatures (50 K), while the C1 and C2 states are
observed at 300 K. In agreement with the calculated barrier the adsorbate can switch
between these states at 300 K [24].
Besides the structures discussed, additional structures and reaction paths were
investigated including other intermediate states showing a single N atom bound to
the electrophilic surface silicon atom or structures being bonded only via the tolyl
ring. Also, a final state comprising of a SiDimer -N-SiSubsurface ring and the additional
dissociative adsorption of the methyl group [11] were investigated. However, these
possibilities could be rejected due to too high barriers or unfavorable thermodynamics.
From the DFT investigations, we thus conclude that the [3 + 2] cycloaddition
product IM is the most likely species observed at low temperatures, which converts
to adsorption configurations C1 and C2 at higher temperatures, exhibiting three
membered Si-N-Si rings.

2.3 Methyl enol ether functionalized cyclooctyne on Si(001)

While azides are well suited as a building block for the second organic layer,
cyclooctyne derivates selectively react on Si(001) over the strained triple bond of
the cyclooctyne ring [33, 53, 58] and are therefore an ideal building block for the
first layer. This chemoselectivity is traced back to the direct adsorption pathway on
Si(001) [44,51,58], which is in contrast to almost all other organic adsorbates [15,35].
The unused second functional group of the substituted cyclooctynes can thus be used
for further reactions [33, 53, 58]. A promising cyclooctyne derivate is the methyl
enol ether functionalized cyclooctyne (MEECO), since the enol ether group could be
employed in click reactions with tetrazine derivatives [42].
Figure 4 shows the four most stable adsorption structures found for the cyclooctyne
derivate. Shown are only structures of the Z-isomer of MEECO, since this isomer
is found to bind more strongly to the surface than the E-isomer, which is the more
stable isomer in the gas phase. For both isomers the same adsorption structures were
found and we therefore assume that both react in the same manner with a Si(001)
surface [19].
Reactivity of organic molecules on semiconductor surfaces 119

Figure 4a shows the only structure, where the adsorbate is not cleaved into two
fragments (intact). This structure resembles the on-top structure of cyclooctyne [52]
showing the same bonding energy using the PBE functional. This supports previous
findings that exchanging the functional group averted from the triple bond does not
alter the bonding to the substrate [53].
The structure in Figure 4b (ether-cleavage) is obtained by breaking the C-O bond
at the ether group. The mechanism of this reaction strongly resembles a previously
observed SN 2-type attack [50]. This breaking of the C-O bond proceeds via an
intermediate state with the intact ether group being datively bonded to the surface
dimer [19]. The bonding energy is with −599 kJ·mol−1 considerable larger than for
the intact structure and similar to the energy of an ether-functionalized cyclooctyne
[53]. Alternatively, the ether-cleavage reaction can be performed across the dimer
row leading to structure in Figure 4c (methoxy dissociation). This structure shows
with −637 kJ·mol−1 the largest bonding energy.
The formation of a methylene group (Figure 4d, aldehyde) leads to the fourth
relevant adsorption mode. With a bonding energy of −440 kJ·mol−1 this structure
is also thermodynamically favored over the intact structure. These four adsorption
modes agree with experimental XPS data, which furthermore show, that most of the
methyl enol ether groups stay intact and are therefore available for a reaction with a
second organic building block [19]. Additional adsorption modes like an adsorption
via the ether oxygen atom, the enol double bond or both the double and triple bond
were also investigated but clearly less stable.

Fig. 4: Most stable adsorption modes of the cyclooctyne derivate. Bonding energies
(Gibbs energies) at HSE06-D3 level of theory in kJ·mol−1 . Bond length in Å. Reprint
from reference [19].
120 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

Finally, also the reaction barriers for the conversion intact to ether-cleavage and
the formation of methoxy dissociation were calculated. The obtained barriers of 79
and 102 kJ·mol−1 , respectively, agree well with experimental observations that for low
temperatures a reduced reactivity for the enol ether group of the MEECO molecules
is observed. In addition, blocking neighboring surface dimers at high coverage can
further reduce the reactivity of the second functional group.
In conclusion, MEECO molecules adsorb chemoselective via the strained triple
bond of cyclooctyne at 150 K on Si(001). Increasing the substrate temperature
promotes the formation of additional products via reactions of the enol ether group.
Therefore, low temperature and high coverages enable MEECO as an excellent
molecular building block for the first organic layer.

3 Interface formation of corroles on Ag(111)

While the previous sections focused on the reactivity and chemoselectivity of organic
molecules on semiconducting surfaces, this section presents intriguing electronic
effects observed in the interface formation of a corrole on Ag(111).
Corroles belong to the class of cyclic tetrapyrroles and are structurally closely re-
lated to porphyrins and phthalocyanines [47]. The interface chemistry of tetrapyrroles
is of fundamental interest [3, 21, 80], especially their possibility to coordinate metal
atoms [9, 13, 21, 40] and the role of aromaticity [54, 70, 81]. In contrast to porphyrins,
corroles contain three methine bridges and one direct pyrrole-pyrrole bond (see Figure
5). As a result of this structure, corroles form neutral complexes with trivalent metal

Fig. 5: Adsorption structures of 3H-HEDMC (a,b) and 2H-HEDMC (c,d) on Ag(111).

Cavg and Agavg indicate the average high of carbon atoms in the corrole ring and silver
atoms of the first layer. Color scheme in (b) and (d) indicates the vertical displacement
of surface atoms due to the presence of an adsorbate. Reprint from reference [86].
Reactivity of organic molecules on semiconductor surfaces 121

ions. Furthermore, the contracted corrole macrocycle provides a tighter coordination

environment for the metal ion compared to porphyrins, resulting in an additional
stabilization of higher oxidation states [4, 18].
In this study 2,3,8,12,17,18-hexaethyl-7,13-dimethylcorrole (3H-HEDMC, Figure
5) and the parent molecule 3H-C (see Figure 6), in which all ethyl groups are
replaced by hydrogen atoms, were investigated. In all calculations only the most
stable tautomer of the corroles has been used. In the gas phase 3H-HEDMC and 3H-C
show a nonplanar structure, in which also the hydrogen atoms of the N-H groups are
pointing out of the plane formed by the nitrogen atoms. In Figure 5 the on-surface
structure of 3H-HEDMC is shown. Here, all alkyl groups are pointing away from the
surface and the out-of-plane angle for one hydrogen atom is even more pronounced.
For both, 3H-HEDMC and 3H-C, large bonding energies are obtained with −403
and −267 kJ·mol−1 , respectively. The interaction with the surface solely stems from
dispersion interactions, while the electronic contributions are even repulsive. This
nicely explains why different adsorption sites (bridge, hollow, on-top) show similar
adsorption energies [86].
The hydrogen atom of the N-H group, which is already pointing towards the
surface, is easily transferred to the surface indicated by the small reaction barrier
of 76 kJ·mol−1 . This barrier nicely agrees with the experimentally observed onset
temperature of 180 K [86]. The N-H dissociation relieves the steric stress in the
corroles leading to the planar structures 2H-HEDMC (see Figure 5) and 2H-C.
Furthermore, the adsorption heights decrease by around 0.2 Å while the bonding
energies increase to −337 (2H-C) and −457 kJ·mol−1 (2H-HEDMC). Noteworthy,
this increase in the bonding energies is also based on now attractive electronic
interactions. Furthermore, this reaction is driven by the increase in entropy due to the
combination and release of H2 .
Experimental data indicate a substantial charge transfer contributing to the
stabilization of 2H-HEDMC [86]. Charge density difference plots for 3H-C and 2H-C
are shown in Figure 6, whereby charge accumulation is shown in blue and charge
depletion in red. For 3H-C the most prominent feature is a charge accumulation

Fig. 6 Charge density dif-

ference plots of 3H-C (a)
and 2H-C (b) on Ag(111).
Charge accumulation shown
in blue and charge depletion
in red. An isosurface value
−3
of 0.0001 𝑒·Å was used.
Adapted reprint from reference
[86].
122 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

directly below the hydrogen atom, which is already directed towards the surface.
Larger changes in the electron density are observed for 2H-C. Here, the molecule
receives a significant amount of charge from the surface. The distribution of this
charge at the molecule resembles the shape of the SOMO of the gas phase molecule.
The charge flow to the adsorbate can be quantified by Hirshfeld partitioning to be
−0.08 𝑒 for 3H-C and −0.43 𝑒 for 2H-C. Furthermore, an analysis of the spin density
indicates that the charge transfer from the surface into the singly occupied molecular
orbital molecule almost completely quenches the spin [86]. Finally, the effect on the
aromaticity was studied by means of NICS and HOMA. Here, a significant higher
aromatic character was observed for the 2H-C anion in comparison to the neutral
molecule.
This suggests that the charge transfer from the surface to the adsorbed 2H-HEDMC
can be described as aromaticity-driven; by accepting electron density from the metal
surface, the molecule gains aromatic character and thus aromatic stabilization. Such
interfacial charge transfer effects play a prominent role in the context of charge
injection in organic electronic devices [16, 27, 64].

4 Developing accurate and efficient computational models

4.1 The introduction of a hierarchical model system

Studying several organic adsorbates, or assuming a complete first layer and modelling
the reactions resulting in the formation of the second layer is a computationally highly
expensive task. Therefore, we aim at deriving an accurate and efficient model to study
the formation of hybrid interfaces and their growth with DFT.
For our benchmark, three model reactions of interest for the formation of the second
layer were selected: With a variant of the azide-alkyne 1,3-dipolar cycloaddition
(AAC) [60] a prototypical click reaction with excellent performance [46] was selected.

Fig. 7 Hierarchical models to

study an interface between an
organic layer and a semicon-
ductor surface. For details see
text. Reprint from reference
[38].
Reactivity of organic molecules on semiconductor surfaces 123

A possible alternative to the AAC reaction is the acyl-chloride mediated esterification

(ACE) [61] due to its smaller reaction barriers. As a third reaction an inverse electron
demand Diels–Alder (IEDDA) reaction utilizing a tetrazine derivative as the diene
[42] was selected. Here, we considered only the first reaction step.
For our model systems we assume an ideal first layer of organic molecules (details
see [38]) on the Si(001) surface. Our approach assumes that for the reaction towards
the second layer, bond cleavage and formation are predominantly local phenomena.
Therefore, the idea is to remove parts of the system, which are not directly involved
in the reactions of interest. In Figure 7, the four proposed hierarchical models are
sketched. The complete system composed of the surface slab and the complete first
organic layer is used as the reference model (rank 1 model). Removing the substrate
leads to the rank 2 model. If necessary, dangling bonds are saturated with hydrogen
atoms. This model is lacking electronic interactions with the substrate, while template
effects and intra-layer-effects are still included. In comparison to the rank 1 model
speedups by a factor of 2 to 6 are observed for the AAC reaction.
Instead of removing the substrate, one could remove all molecules of the first
organic layer, which do not participate in the reaction with the new adsorbate. This
leads to the rank 3 model. Here, electronic interactions with the substrate are included,
while the intra-layer-effects are neglected. Still, a minor speedup of a factor 1.5 to 2
is achieved. For the simplest possible system, the substrate and the first organic layer
is removed (rank 4 model). While this model lacks all environmental influences, it
shows with a factor of 50 the greatest speedup, can be carried out with molecular
quantum chemistry methods and might still be accurate enough to sample the reaction
between two organic molecules.
All model systems were tested with the three reactions of interest. Furthermore,
different density functionals and even CCSD(T) and MP2 calculations were performed
for the rank 4 model [38]. Figure 8 shows exemplary results for the AAC reaction.
It shows that reaction energies and barriers are only slightly affected. Furthermore,
two dominant structural effects can be identified: As shown by the inset in Figure

Fig. 8 Reaction energies of

the azide-alkyne cycloaddition
(AAC). Reprint from reference
[38].
124 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

8, a clustering of the values for rank 1, 2 and rank 3, 4 are observed. Removing the
organic layer and therefore the intra-layer effects leads to the large change from rank
1, 2 to rank 3, 4. Ranks 3 and 4 are higher in energy since they lack the attractive
dispersion interactions with the organic layer. Removing the surface results in only a
minor change in energy due to the presence of steric repulsion between the silicon
surface and the first layer of molecules.
The benchmark of different density functionals for all reaction energies, reaction
barriers and models revealed that rank 2 always performs very well (mean error
smaller than 5 kJ·mol−1 ). In contrast, the model ranks 3 and 4 showed larger deviations
and for optB88 even larger than 10 kJ·mol−1 . Also including solvent effects, does not
alter the overall picture.
In conclusion, in terms of efficiency a rank 1 model should only be used in case
highly accurate results are necessary. A rank 2 model is commonly suited to study
the chemistry within the organic layers. The rank 3 model is a good choice to extract
adsorption and reaction energies for the adsorbate-surface interactions. Finally, the
gas phase model (rank 4) can be used for screening molecular building blocks and
benchmark of DFT approaches.

4.2 Reactivity of a Si(001) dimer vacancy

The previous sections worked with ideal Si(001) or Ge(001) surfaces. In the literature,
some studies concerning the influence of step edges exists [41, 57]. Although, the
formation of defects can be suppressed [23], we believe that they could be used
to increase selectivity of surface reactions. By incorporating surface defects in our
models, we are able to quantify and understand changes in reactivity with respect to
the pristine surface. Furthermore, the investigation of nonideal surfaces will help to
create more realistic computational models of organic-semiconductor interfaces.
In Figure 9 the structure and the formation energy of common defects of a Si(001)
surface are shown. The presence of defects allows for additional structural motive and
thereby increase the richness of the adsorption chemistry. In our study, we focused on
the most favorable point defect, the bonded dimer vacancy (DV) [59]. To study this
defect the surface model was increased to a 8 x 4 supercell with eight silicon layers
where one of the dimers is missing.
The introduction of the dimer vacancy has some consequences on the atomic
and electronic structure of the surface [39]. The defect itself relaxes by forming
bonds between the silicon defect atoms as indicated by the bond length of 2.88 Å.
Furthermore, the neighboring dimer is flipped. By calculating and visualizing the
band structure of this slab, a strongly localized defect state is observed inside the
former band gap. Consequently, the band gap is narrowed by 0.15 eV.
To reveal the influence of the defect on the reactivity of the surface, the adsorption
of cyclooctyne, acetylene and ethylene was studied. Interestingly, acetylene and
ethylene prefer to bond directly to the defect atoms leading to bonding energies
of −302 and −321 kJ·mol−1 , respectively. However, the most reactive molecule
Reactivity of organic molecules on semiconductor surfaces 125

cyclooctyne prefers to adsorb on-top of a silicon dimer leading to a bonding energy

of −307 kJ·mol−1 (healing the defect: −250 kJ·mol−1 ). As supported by pEDA, an
adsorption of cyclooctyne in the tight space of the DV results in a larger Pauli
repulsion and smaller orbital interactions between the adsorbate and surface, which
significantly lowers the bonding energy [39].
For acetylene and ethylene we also searched for direct adsorption paths resulting
in a healing of the defect. Since direct adsorption paths to these final states were not
found, also adsorption paths to other structures and paths connecting the different
adsorption modes were calculated. The complete reaction network for acetylene is
shown in Figure 10. Following the reactions with the lowest barriers (highlighted
in red), the most likely steps to reach the thermodynamically most stable structure
(pedestal) would start with a direct adsorption in the pre-bridge-E structure. This
structure can easily overcome the barrier of 5 kJ·mol−1 to reach the bridge-E structure.
Although a direct connection to the pedestal structure exists, another intermediate step
to the bridge structure is more likely due to the smaller barrier of 154 kJ·mol−1 . The
last step shows with 170 kJ·mol−1 the largest barrier and therefore the rate determining

Fig. 9 Common defects of a

Si(001) surface. Bond length
in Å. Reprint from reference
[39].
126 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

step. Overall, due to the missing direct adsorption and pronounced barriers for a
reaction with the defect, the DV can be concluded to be less reactive than silicon
dimers on a pristine surface.

5 Scaling of computational methods

Structural optimizations and reaction path calculations were performed with the
software package "Vienna Ab initio Simulation Package" (VASP). The Software
is licensed by the VASP Software GmbH located in Vienna, Austria (https:
//www.vasp.at/, [email protected]). The best code performance on HAWK is
obtained by using the latest version of the Intel Fortran compiler together with the
HPE MPI. Basic numeric libraries like (SCA)LAPACK, BLAS or FFTW are also
used to push the performance.
VASP comprises three levels of parallelization. The least intense level in terms
of MPI communication is the parallel calculation of the wave function at different
k-points. The second level is composed of the calculation of bands, while in the third

Fig. 10: Reaction network of acetylene on the DV-defect. Bonding energies are stated
relative to the isolated surface and molecule, while reaction barriers (in brackets) are
given to the previous minimum. C-C bond length given in Å. The most likely path is
highlighted in red. Reprint from reference [39].
Reactivity of organic molecules on semiconductor surfaces 127

level the calculation is parallelized over the basis set (plane waves). Due to the high
MPI communication in the treatment of the plane wave, this parallelization is usually
restricted to the physical socket, i.e. 64 cores for HAWK. For molecular dynamics
(MD) and nudged-elastic band (NEB) calculations a very good code scaling was
already shown on Cray XC40 (Hazel Hen) at HLRS for up to 1536 and 3072 cores,
respectively [55].
In Figure 11 the scaling benchmark of our automatization routine, which is
further developed in this project, is shown. The surface reaction network of GaH3 on
GaP(001) is used as a benchmark system. Every calculation was restricted to a wall
clock time of 10 minutes, where the first 40 seconds were consumed for initialization.
Every instance of VASP had to deal with a system containing 120 atoms in the unit
cell (H32 Ga50 P34 GaH3 ), 366 explicitly treated electrons, 5 k-points, 260 bands per
k-point and ca. 107,000 plane waves per band. Consequently, the VASP parallelization
over k-points was enabled where possible. Requirements for file IO and memory per
core are negligible.
As expected, our automatization routine shows a nearly linear scaling behavior
up to 29 nodes due to the trivial parallelization over multiple VASP instances. The
usage of more and more nodes for a single VASP instance, more than 29 total used
nodes, leads only to a minor decrease in the scaling compared to the ideal scaling.
This trend will persist even for larger numbers of cores due to the very good scaling
behavior of VASP [55]. Furthermore, a superlinear scalability is observed for 88 used
nodes. Here, the reduced memory requirements due to the separation of the bands
to multiple nodes overcompensates the increased MPI communication. All in all,
with the combination of a high-performing DFT code and our automatization routine
we are now able to efficiently use in a single calculation significant amounts of the
computational resources provided by HAWK in a massively parallel fashion.

Fig. 11: Scaling behavior of our automatization routine using VASP 5.4.4 on HAWK
at HLRS. For details on the calculation see text.
128 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

References

1. International Roadmap for Devices and Systems (IRDS) 2020. https://irds.ieee.org, Last
accessed June 2021.
2. S. V. Aradhya and L. Venkataraman. Single-molecule junctions beyond electronic transport.
Nat. Nanotechnol., 8(6):399–410, 2013.
3. W. Auwärter, D. Écija, F. Klappenberger, and J. V. Barth. Porphyrins at interfaces. Nat. Chem.,
7(2):105–120, 2015.
4. I. Aviv-Harel and Z. Gross. Aura of Corroles. Chem. - A Eur. J., 15(34):8382–8394, 2009.
5. S.-S. Bae, S. Kim, and J. Won Kim. Adsorption Configuration Change of Pyridine on Ge(100):
Dependence on Exposure Amount. Langmuir, 25(1):275–279, 2009.
6. S. Bocharov, O. Dmitrenko, L. P. Méndez De Leo, and A. V. Teplyakov. Azide Reactions for
Controlling Clean Silicon Surface Chemistry: Benzylazide on Si(100)-2 x 1. J. Am. Chem. Soc.,
128(29):9300–9301, 2006.
7. J. M. Buriak. Organometallic Chemistry on Silicon and Germanium Surfaces. Chem. Rev.,
102(5):1271–1308, 2002.
8. Q. Cai, B. Xu, L. Ye, Z. Di, S. Huang, X. Du, J. Zhang, Q. Jin, and J. Zhao. 1-Dodecanethiol
based highly stable self-assembled monolayers for germanium passivation. Appl. Surf. Sci.,
353:890–901, 2015.
9. M. Chen, H. Zhou, B. P. Klein, M. Zugermeier, C. K. Krug, H.-J. Drescher, M. Gorgoi,
M. Schmid, and J. M. Gottfried. Formation of an interphase layer during deposition of cobalt
onto tetraphenylporphyrin: a hard X-ray photoelectron spectroscopy (HAXPES) study. Phys.
Chem. Chem. Phys., 18(44):30643–30651, 2016.
10. R. G. Closser, D. S. Bergsman, L. Ruelas, F. S. M. Hashemi, and S. F. Bent. Correcting defects
in area selective molecular layer deposition. J. Vac. Sci. Technol. A Vacuum, Surfaces, Film.,
35(3):031509, 2017.
11. F. Costanzo, C. Sbraccia, P. Luigi Silvestrelli, and F. Ancilotto. Proton-transfer reaction of
toluene on Si(100) surface. Surf. Sci., 566-568:971–976, 2004.
12. Y. Cui, Q. Wei, H. Park, and C. M. Lieber. Nanowire Nanosensors for Highly Sensitive and
Selective Detection of Biological and Chemical Species. Science, 293(5533):1289–1292, 2001.
13. K. Diller, A. C. Papageorgiou, F. Klappenberger, F. Allegretti, J. V. Barth, and W. Auwärter. In
vacuo interfacial tetrapyrrole metallation. Chem. Soc. Rev., 45(6):1629–1656, 2016.
14. M. Ebrahimi and K. Leung. Selective surface chemistry of allyl alcohol and allyl aldehyde on
Si(100)2x1: Competition of [2+2] CC cycloaddition with O-H dissociation and with [2+ 2] CO
cycloaddition in bifunctional molecules. Surf. Sci., 603(9):1203–1211, 2009.
15. M. A. Filler and S. F. Bent. The surface as molecular reagent: organic chemistry at the
semiconductor interface. Prog. Surf. Sci., 73(1-3):1–56, 2003.
16. Y. Gao, Y. Shao, L. Yan, H. Li, Y. Su, H. Meng, and X. Wang. Efficient Charge Injection in
Organic Field-Effect Transistors Enabled by Low-Temperature Atomic Layer Deposition of
Ultrathin VOx Interlayer. Adv. Funct. Mater., 26(25):4456–4463, 2016.
17. S. M. George, B. Yoon, and A. A. Dameron. Surface Chemistry for Molecular Layer Deposition
of Organic and Hybrid Organic-Inorganic Polymers. Acc. Chem. Res., 42(4):498–508, 2009.
18. A. Ghosh. Electronic Structure of Corrole Derivatives: Insights from Molecular Structures,
Spectroscopy, Electrochemistry, and Quantum Chemical Calculations. Chem. Rev., 117(4):3798–
3881, 2017.
19. T. Glaser, J. Meinecke, C. Länger, J.-N. Luy, R. Tonner, U. Koert, and M. Dürr. Combined XPS
and DFT investigation of the adsorption modes of methyl enol ether functionalized cyclooctyne
on Si(001). ChemPhysChem, 22(4):404–409, 2021.
20. S. Gokhale, P. Trischberger, D. Menzel, W. Widdra, H. Dröge, H.-P. Steinrück, U. Birkenheuer,
U. Gutdeutsch, and N. Rösch. Electronic structure of benzene adsorbed on single-domain
Si(001)-(2x1): A combined experimental and theoretical study. J. Chem. Phys., 108(13):5554–
5564, 1998.
21. J. M. Gottfried. Surface chemistry of porphyrins and phthalocyanines. Surf. Sci. Rep.,
70(3):259–379, 2015.
Reactivity of organic molecules on semiconductor surfaces 129

22. R. Har-Lavan, O. Yaffe, P. Joshi, R. Kazaz, H. Cohen, and D. Cahen. Ambient organic molecular
passivation of Si yields near-ideal, Schottky-Mott limited, junctions. AIP Adv., 2(1):012164,
2012.
23. K. Hata, T. Kimura, S. Ozawa, and H. Shigekawa. How to fabricate a defect free Si(001) surface.
J. Vac. Sci. Technol. A Vacuum, Surfaces, Film., 18(4):1933–1936, 2000.
24. J. Heep, J.-N. Luy, C. Länger, J. Meinecke, U. Koert, R. Tonner, and M. Dürr. Adsorption of
Methyl-Substituted Benzylazide on Si(001): Reaction Channels and Final Configurations. J.
Phys. Chem. C, 124(18):9940–9946, 2020.
25. G. Hills, C. Lau, A. Wright, S. Fuller, M. D. Bishop, T. Srimani, P. Kanhaiya, R. Ho, A. Amer,
Y. Stein, D. Murphy, Arvind, A. Chandrakasan, and M. M. Shulaker. Modern microprocessor
built from complementary carbon nanotube transistors. Nature, 572(7771):595–602, 2019.
26. W. A. Hofer, A. J. Fisher, G. P. Lopinski, and R. A. Wolkow. Adsorption of benzene on
Si(100)-(2x1): Adsorption energies and STM image analysis by ab initio methods. Phys. Rev.
B, 63(8):085314, 2001.
27. M. Hollerer, D. Lüftner, P. Hurdax, T. Ules, S. Soubatch, F. S. Tautz, G. Koller, P. Puschnig,
M. Sterrer, and M. G. Ramsey. Charge Transfer and Orbital Level Alignment at Inorganic/Organic
Interfaces: The Role of Dielectric Interlayers. ACS Nano, 11(6):6252–6260, 2017.
28. M. Hossain, Y. Yamashita, K. Mukai, and J. Yoshinobu. Selective functionalization of the
Si(100) surface by switching the adsorption linkage of a bifunctional organic molecule. Chem.
Phys. Lett., 388(1-3):27–30, 2004.
29. H. G. Huang, J. Y. Huang, Z. H. Wang, Y. S. Ning, F. Tao, Y. P. Zhang, Y. H. Cai, H. H. Tang,
and G. Q. Xu. Adsorption of nitrogen-containing aromatic molecules on Si(111)-7x7. Surf.
Sci., 601(5):1184–1192, 2007.
30. H. G. Huang, Z. H. Wang, and G. Q. Xu. The Selective Formation of Di-𝜎 N-Si Linkages in
Pyrazine Binding on Si(111)-7x7. J. Phys. Chem. B, 108(33):12560–12567, 2004.
31. S. C. Jung and M. H. Kang. Adsorption structure of pyrazine on Si(100): Density-functional
calculations. Phys. Rev. B, 80(23):235312, 2009.
32. H. C. Kolb, M. G. Finn, and K. B. Sharpless. Click Chemistry: Diverse Chemical Function
from a Few Good Reactions. Angew. Chemie Int. Ed., 40(11):2004–2021, 2001.
33. C. Länger, J. Heep, P. Nikodemiak, T. Bohamud, P. Kirsten, U. Höfer, U. Koert, and M. Dürr.
Formation of Si/organic interfaces using alkyne-functionalized cyclooctynes-precursor-mediated
adsorption of linear alkynes versus direct adsorption of cyclooctyne on Si(0 0 1). J. Phys.
Condens. Matter, 31(3):034001, 2019.
34. H.-K. Lee, J. Park, I. Kim, H.-D. Kim, B.-G. Park, H.-J. Shin, I.-J. Lee, A. P. Singh, A. Thakur,
and J.-Y. Kim. Selective Reactions and Adsorption Structure of Pyrazine on Si(100): HRPES
and NEXAFS Study. J. Phys. Chem. C, 116(1):722–725, 2012.
35. T. R. Leftwich and A. V. Teplyakov. Chemical manipulation of multifunctional hydrocarbons
on silicon surfaces. Surf. Sci. Rep., 63(1):1–71, 2008.
36. T. R. Leftwich and A. V. Teplyakov. Cycloaddition Reactions of Phenylazide and Benzylazide
on a Si(100)-2 x 1 Surface. J. Phys. Chem. C, 112(11):4297–4303, 2008.
37. P. W. Loscutoff and S. F. Bent. REACTIVITY OF THE GERMANIUM SURFACE: Chemical
Passivation and Functionalization. Annu. Rev. Phys. Chem., 57(1):467–495, 2006.
38. J.-N. Luy, M. Molla, L. Pecher, and R. Tonner. Efficient hierarchical models for reactivity of
organic layers on semiconductor surfaces. J. Comput. Chem., 42(12):827–839, 2021.
39. J.-N. Luy and R. Tonner. Organic Functionalization at the Si(001) Dimer Vacancy Defect-
Structure, Bonding, and Reactivity. J. Phys. Chem. C, 125(10):5635–5646, 2021.
40. H. Marbach. Surface-Mediated in Situ Metalation of Porphyrins at the Solid-Vacuum Interface.
Acc. Chem. Res., 48(9):2649–2658, 2015.
41. A. Mazzone. Acetylene adsorption onto Si(100): a study of adsorption dynamics and of surface
steps. Comput. Mater. Sci., 35(1):6–12, 2006.
42. J. Meinecke and U. Koert. Copper-Free Click Reaction Sequence: A Chemoselective Layer-by-
Layer Approach. Org. Lett., 21(18):7609–7612, 2019.
43. X. Meng. An overview of molecular layer deposition for organic and organic-inorganic hybrid
materials: mechanisms, growth characteristics, and promising applications. J. Mater. Chem. A,
5(35):18326–18378, 2017.
130 F. Pieck, J.-N. Luy, F. Kreuter, B. Mondal and R. Tonner-Zech

44. G. Mette, M. Dürr, R. Bartholomäus, U. Koert, and U. Höfer. Real-space adsorption studies of
cyclooctyne on Si(001). Chem. Phys. Lett., 556:70–76, 2013.
45. L. Miozzo, A. Yassar, and G. Horowitz. Surface engineering for high performance organic
electronic devices: the chemical approach. J. Mater. Chem., 20(13):2513, 2010.
46. N. Münster, P. Nikodemiak, and U. Koert. Chemoselective Layer-by-Layer Approach Utilizing
Click Reactions with Ethynylcyclooctynes and Diazides. Org. Lett., 18(17):4296–4299, 2016.
47. S. Nardis, F. Mandoj, M. Stefanelli, and R. Paolesse. Metal complexes of corrole. Coord. Chem.
Rev., 388:360–405, 2019.
48. W. K. H. Ng, S. T. Sun, J. W. Liu, and Z. F. Liu. The Mechanism for the Thermally Driven Self-
Assembly of Pyrazine into Ordered Lines on Si(100). J. Phys. Chem. C, 117(30):15749–15753,
2013.
49. T. Omiya, H. Yokohara, and M. Shimomura. Well-Oriented Pyrazine Lines and Arrays on
Si(001) Formed by Thermal Activation of Substrate. J. Phys. Chem. C, 116(18):9980–9984,
2012.
50. L. Pecher, S. Laref, M. Raupach, and R. Tonner. Ether auf Si(001): Ein Paradebeispiel für
die Gemeinsamkeiten zwischen Oberflächenwissenschaften und organischer Molekülchemie.
Angew. Chemie, 129(47):15347–15351, 2017.
51. L. Pecher, S. Schmidt, and R. Tonner. Modeling the Complex Adsorption Dynamics of Large
Organic Molecules: Cyclooctyne on Si(001). J. Phys. Chem. C, 121(48):26840–26850, 2017.
52. L. Pecher, C. Schober, and R. Tonner. Chemisorption of a Strained but Flexible Molecule:
Cyclooctyne on Si(001). Chem. - A Eur. J., 23(23):5459–5466, 2017.
53. L. Pecher and R. Tonner. Computational analysis of the competitive bonding and reactivity
pattern of a bifunctional cyclooctyne on Si(001). Theor. Chem. Acc., 137(4):48, 2018.
54. M. D. Peeks, T. D. W. Claridge, and H. L. Anderson. Aromatic and antiaromatic ring currents
in a molecular nanoring. Nature, 541(7636):200–203, 2017.
55. F. Pieck, J.-N. Luy, N. L. Zaitsev, L. Pecher, and R. Tonner. Chemistry at Surfaces Modeled by
Ab Initio Methods. HLRS project report, 2019.
56. S. R. Puniredd, S. Jayaraman, S. H. Yeong, C. Troadec, and M. P. Srinivasan. Stable Organic
Monolayers on Oxide-Free Silicon/Germanium in a Supercritical Medium: A New Route to
Molecular Electronics. J. Phys. Chem. Lett., 4(9):1397–1403, 2013.
57. M. Raschke and U. Höfer. Influence of steps and defects on the dissociative adsorption of
molecular hydrogen on silicon surfaces. Appl. Phys. B Lasers Opt., 68(3):649–655, 1999.
58. M. Reutzel, N. Münster, M. A. Lipponer, C. Länger, U. Höfer, U. Koert, and M. Dürr.
Chemoselective Reactivity of Bifunctional Cyclooctynes on Si(001). J. Phys. Chem. C,
120(46):26284–26289, 2016.
59. N. Roberts and R. J. Needs. Total energy calculations of missing dimer reconstructions on the
silicon (001) surface. J. Phys. Condens. Matter, 1(19):3139–3143, 1989.
60. V. V. Rostovtsev, L. G. Green, V. V. Fokin, and K. B. Sharpless. A Stepwise Huisgen
Cycloaddition Process: Copper(I)-Catalyzed Regioselective Ligation of Azides and Terminal
Alkynes. Angew. Chemie Int. Ed., 41(14):2596–2599, 2002.
61. F. Ruff and Ö. Farkas. Concerted SN 2 mechanism for the hydrolysis of acid chlorides:
comparisons of reactivities calculated by the density functional theory with experimental data.
J. Phys. Org. Chem., 24(6):480–491, 2011.
62. T. Sandoval and S. Bent. Adsorption of Multifunctional Organic Molecules at a Surface: First
Step in Molecular Layer Deposition. In Encycl. Interfacial Chem., pages 523–537. Elsevier,
2018.
63. T. E. Sandoval, F. Pieck, R. Tonner, and S. F. Bent. Effect of Heteroaromaticity on Adsorption
of Pyrazine on the Ge(100)-2x1 Surface. J. Phys. Chem. C, 124(40):22055–22068, 2020.
64. J. C. Scott. Metal-organic interface and charge injection in organic electronic devices. J. Vac.
Sci. Technol. A Vacuum, Surfaces, Film., 21(3):521–531, 2003.
65. K. W. Self, R. I. Pelzel, J. H. G. Owen, C. Yan, W. Widdra, and W. H. Weinberg. Scanning
tunneling microscopy study of benzene adsorption on Si(100)-(2x1). J. Vac. Sci. Technol. A
Vacuum, Surfaces, Film., 16(3):1031–1036, 1998.
66. Y. X. Shao, Y. H. Cai, D. Dong, S. Wang, S. G. Ang, and G. Q. Xu. Spectroscopic study of
propargyl chloride attachment on Si(100)-2x1. Chem. Phys. Lett., 482(1-3):77–80, 2009.
Reactivity of organic molecules on semiconductor surfaces 131

67. M. Shimomura, D. Ichikawa, Y. Fukuda, T. Abukawa, T. Aoyama, and S. Kono. Formation

of one-dimensional molecular chains on a solid surface: Pyrazine/Si(001). Phys. Rev. B,
72(3):033303, 2005.
68. B. Shong, T. E. Sandoval, A. M. Crow, and S. F. Bent. Unidirectional Adsorption of Bifunctional
1,4-Phenylene Diisocyanide on the Ge(100)-2x1 Surface. J. Phys. Chem. Lett., 6(6):1037–1041,
2015.
69. P. L. Silvestrelli, F. Ancilotto, and F. Toigo. Adsorption of benzene on Si(100) from first
principles. Phys. Rev. B, 62(3):1596–1599, 2000.
70. M. Stȩpień, N. Sprutta, and L. Latos-Grażyński. Figure Eights, Möbius Bands, and More:
Conformation and Aromaticity of Porphyrinoids. Angew. Chemie Int. Ed., 50(19):4288–4340,
2011.
71. P. Sundberg and M. Karppinen. Organic and inorganic-organic thin film structures by molecular
layer deposition: A review. Beilstein J. Nanotechnol., 5:1104–1136, 2014.
72. F. Tao, S. L. Bernasek, and G.-Q. Xu. Electronic and Structural Factors in Modification and
Functionalization of Clean and Passivated Semiconductor Surfaces with Aromatic Systems.
Chem. Rev., 109(9):3991–4024, 2009.
73. F. F. Tao and S. L. Bernasek. Chemical Binding of Five-Membered and Six-Membered
Aromatic Molecules. In Funct. Semicond. Surfaces, pages 89 – 104. John Wiley & Sons, Inc.:
Hoboken, NJ, USA, 2012.
74. A. V. Teplyakov and S. F. Bent. Semiconductor surface functionalization for advances in
electronics, energy conversion, and dynamic systems. J. Vac. Sci. Technol. A Vacuum, Surfaces,
Film., 31(5):050810, 2013.
75. P. Thissen, O. Seitz, and Y. J. Chabal. Wet chemical surface functionalization of oxide-free
silicon. Prog. Surf. Sci., 87(9-12):272–290, 2012.
76. B. Wang, J. C. Cancilla, J. S. Torrecilla, and H. Haick. Artificial Sensing Intelligence with
Silicon Nanowires for Ultraselective Detection in the Gas Phase. Nano Lett., 14(2):933–938,
2014.
77. R. A. Wolkow. CONTROLLED MOLECULAR ADSORPTION ON SILICON: Laying a
Foundation for Molecular Devices. Annu. Rev. Phys. Chem., 50(1):413–441, 1999.
78. J. T. Yates. Surface chemistry of silicon-the behaviour of dangling bonds. J. Phys. Condens.
Matter, 3(S):S143–S156, 1991.
79. J. T. Yates Jr. A New Opportunity in Silicon-Based Microelectronics. Science, 279(5349):335–
336, 1998.
80. T. Yokoyama, S. Yokoyama, T. Kamikado, Y. Okuno, and S. Mashiko. Selective assembly on a
surface of supramolecular aggregates with controlled size and shape. Nature, 413(6856):619–
621, 2001.
81. Z. S. Yoon, A. Osuka, and D. Kim. Möbius aromaticity and antiaromaticity in expanded
porphyrins. Nat. Chem., 1(2):113–122, 2009.
82. A. Yu, Y. Qu, K. Han, and G. He. DFT Investigations About Pyrazine Molecules on Si(100)-2x1
Surface. Chem. Res. Chinese Univ., 23(4):444–451, 2007.
83. Y. P. Zhang, J. H. He, G. Q. Xu, and E. S. Tok. Selective Attachment of 4-Bromostyrene on the
Si(111)-(7x7) Surface. J. Phys. Chem. C, 115(31):15496–15501, 2011.
84. H. Zhou and S. F. Bent. Molecular Layer Deposition of Functional Thin Films for Advanced
Lithographic Patterning. ACS Appl. Mater. Interfaces, 3(2):505–511, 2011.
85. H. Zhou and S. F. Bent. Fabrication of organic interfacial layers by molecular layer deposition:
Present status and future opportunities. J. Vac. Sci. Technol. A Vacuum, Surfaces, Film.,
31(4):040801, 2013.
86. M. Zugermeier, J. Herritsch, J.-N. Luy, M. Chen, B. P. Klein, F. Niefind, P. Schweyen, M. Bröring,
M. Schmid, R. Tonner, and J. M. Gottfried. On-Surface Formation of a Transient Corrole Radical
and Aromaticity-Driven Interfacial Electron Transfer. J. Phys. Chem. C, 124(25):13825–13836,
2020.
Electro-catalysis for H2 O oxidation

Travis Jones

Abstract Electrocatalysts facilitate two fundamental processes – electron transfer

and chemical bond formation/rupture – to convert renewable electrical energy into
chemical fuels. Doing so requires the protons and electrons supplied by the oxygen
evolution reaction. The reaction steps constituting the oxygen evolution reaction
are assumed to result in an electrochemical mechanism qualitatively distinct from
the purely chemical ones familiar from thermal catalysis. Such an electrocatalytic
mechanism is often thought to be well-described by an exponential dependence of
rate on applied overpotential, Tafel’s law, requiring the electrochemical bias to act
directly on the reaction coordinate. The aim of the ECHO project has been to test this
assumption for the oxygen evolution reaction by combining experimental efforts with
density functional theory based modeling of the electrified solid/liquid interface for
an important class of electrocatalysts, iridium (di)oxide. Leveraging the computing
power of Hawk enabled the use of computational models with explicit solvent to
compute realistic surface phase diagrams of IrO2 and test its role on the kinetics of
the oxygen evolution reaction on those surfaces under fixed bias and fixed charge
conditions. The results suggest the oxygen evolution reaction is mediated by oxidative
charge rather than the action of the electrochemical bias on the reaction coordinate.

Travis Jones
Department of Inorganic Chemistry, Fritz-Haber-Institut der Max-Planck-Gesellschaft, Faradayweg
4-6, 14195, Berlin, Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 133
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_8
134 Travis Jones

1 Introduction

Increasing renewable energy deployment is creating new techno-economic challenges

for chemical and energy sectors. Among these challenges is the fact that wind and
solar are non-dispatchable sources, that is, they cannot continuously meet energy
demand owing to their intermittent nature. Energy storage is then required to ensure
a continuous supply of electricity from such sources. Chemical energy conversion is
poised to play a prominent role in overcoming this challenge, while simultaneously
opening new avenues for sustainable chemical production, by transforming the
green electricity from intermittent sources into chemical fuels [4, 8, 11, 31]. Such
technologies rely on electrocatalysts to facilitate the transformation of electrical
energy into chemical bonds. And while a range of electrocatalytic processes are being
explored to produce chemical fuels and feedstocks – e.g. CO2 reduction, N2 reduction,
and water splitting – these transformations require the protons and electrons supplied
by the oxygen evolution reaction (OER) [4, 8, 13, 27, 31, 33, 34, 36, 37]. Catalyzing the
OER is, however, a significant bottleneck. Even with high-performance catalysts the
slow electrochemical kinetics of the OER limit electrochemical conversion efficiencies,
and under the acidic working conditions of the electrolyzers capable of coping with
the intermittency of renewable sources, only iridium-based materials combine OER
activity and operational stability [4, 5, 20, 33]. This has made understanding the OER
mechanism, especially over iridium-based materials, a central focus in electrocatalysis
research [20, 27, 33, 36, 37].
In the simplest and most general description, the OER involves the transfer of four
electrons across the electrode/electrolyte interface, the breaking of O-H bonds, and
the formation of an O-O bond, giving the following stoichiometry:

2H2 O → 4H+ + 4e− + O2 . (1)

Thermodynamically the reaction proceeds at 1.23 V on the standard hydrogen
electrode (SHE) scale, though appreciable OER rates require applied potentials in
excess of this limiting value. Understanding the origin of such overpotentials (𝜂),
and finding ways to minimize them, has been a major driver of research into the
OER. The complexity of the electrified solid/liquid interface together with the large
variety of mechanistic possibilities associated with the transfer of four electrons [16]
and the chemical bond making/breaking steps [15] involved in the OER, however,
has severely impeded progress towards this goal. This complexity is illustrated
schematically in the exemplary reaction pathway shown in Figure 1. As a multistep
heterogeneously-catalyzed reaction, several intermediates and transition states should
be expected along the reaction coordinate. The energies of these states are likely to
be closely linked to the nature of the surface, much like traditional thermal catalysis.
Unlike thermal catalysis, however, the reactant and products, including the protons
released during the OER, are solvated in the electrochemical double layer. Mass
transport and concentration effects can also ad to the complexity of the OER. While
these later aspects can be minimized experimentally, the intrinsic complexity of the
electrocatalytic kinetics and the corresponding (unknown) double layer structure
Electro-catalysis for H2 O oxidation 135

cannot. To avoid a complete description of this complexity, most efforts to understand

OER kinetics rely on the validity of the phenomenological Butler–Volmer kinetic
theory [18, 27, 33, 36, 37], which requires the applied electrochemical potential to
act directly on the reaction coordinate in accord with the well-developed theory of
outer-sphere electron transfer [1, 32]

H2O H+
H2O
H+ H+ H O2
H H O
H H O
O O O O O O

1 2 3 4 5

Fig. 1: Schematic reaction pathway of the OER showing water adsorption (1), and
its oxidation through OH𝑎𝑑𝑠 formation (2), O𝑎𝑑𝑠 (3), OOH𝑎𝑑𝑠 formation via water
nucleophilic attack of O𝑎𝑑𝑠 (4), and the formation of adsorbed O2 (5).

When Butler–Volmer kinetic theory holds, the OER rate depends exponentially
on the applied overpotential. This dependence can be expressed using the slope
of the plot of 𝜂 versus the log of the OER current, a Tafel plot. An example of a
measurement on IrO2 calcined at 450 ◦ C is shown in Figure 2. Here, the plot of 𝜂 = 𝑎
log 𝑖 + 𝑏 produces straight lines, in agreement with the exponential dependence of
rate on 𝜂 predicted by Butler–Volmer theory. Within this picture, the Tafel slope, 𝑎,
takes on characteristic values linked to the nature of the rate determining step under
the measurement conditions and can thus be used to derive mechanistic insights
[1, 18, 32, 35]. The change in slope from 46 mV/dec to 75 mV/dev in Figure 2 would
then indicate a change in mechanism when crossing from the low to high current
regime. This type of Tafel slope analysis forms much the experimental foundations
of our mechanistic understanding of the OER [18, 27, 28, 33, 36, 37], though there
is little experimental evidence demonstrating the validity of Butler–Volmer theory
during the OER.
The experimental challenges associated with probing the electronic and atomic
structure of the electrified solid/liquid interface have led to a crucial reliance on ab
initiosimulations to address mechanistic questions related to the OER [33, 36, 37].
And while early computational studies have helped shape the current field, these
studies focused on the thermodynamics of possible reaction intermediates in vacuum
[29, 33, 36, 37], which cannot address the relevant electrocatalytic performance at
non-zero overpotentials owing to the lack of a double layer and electrochemical
bias. More recent studies have aimed to improve the thermodynamic description by
including solvent effects using ice-like models [9], though it is unclear if such models
accurately capture the solvation energy of the solid/liquid interface, and a purely
136 Travis Jones

Fig. 2: Experimental Tafel plot of IrO2 calcined at 450 ◦ C, data from Ref. [21].
The Tafel slope can be seen to change from 46 mV/dec at low overpotentials to 75
mv/dec at high overpotentials, consistent with a change in reaction mechanism within
a Butler–Volmer picture.

thermodynamic description does not directly address kinetics. Further efforts have
aimed to identify kinetic effects [19, 25], though these too avoid the use of explicit
solvent and therefor do not have an atomic scale description of the double layer, which
can be circumvented by assuming Butler–Volmer theory holds.
In the ECHO project we aim to move beyond a Butler–Volmer description of
the OER by combining density functional theory (DFT) molecular dynamics using
explicit solvent with synchrotron based X-ray spectroscopy to identify how the actives
phase(s) of OER electrocatalysts evolve as a function of electrochemical bias and by
modeling the OER kinetics on these surfaces under constant potential and constant
charge conditions. Over this reporting period we have succeeded in developing
models of the surface phases present on IrO2 from the OER onset up to 𝜂 ≈ 400
mV. We have also been able to compute the activation energies associated with O-H
bond breaking and O-O bond formation on these surfaces under constant charge and
constant electrochemical bias conditions. Through these efforts were able to show
the principle role of the applied overpotential is to modify the surface chemistry of
the electrocatalyst rather than to accelerate the OER by direct action on the reaction
coordinate. This new picture of the OER is described in Ref. [21] together with the
experimental verification. Herein the focus will be on the main computational results.

2 Methods

DFT calculations were performed with the Quantum ESPRESSO package [10] using
the PBE exchange and correlation potential [24] and pseudopotentials from the
standard solid-state pseudopotentials database [26]. In the case of iridium, several
pseudopotentials show similar performance, and the projected augmented wave
dataset from the PSLibrary [6] was used owing to improved SCF convergence of
Electro-catalysis for H2 O oxidation 137

the projected augmented wave dataset. All simulations were performed with spin
polarization using a wave function cutoff of 60 Ry and a charge density cutoff of
540 Ry. To compute surface phase diagrams symmetric slabs of (110) terminated
rutile-type IrO2 were employed using a (1 × 2) supercell. The two slides of the
slab were separated by 18 water molecules. The phase diagram was constructed by
considering the possible permutations of O or OH on the 4 total oxygen sites on the
(110) (1 × 2) surface, 2 𝜇2 sites and 2 𝜇1 sites. Of these, only five terminations were
found to be stable. Denoting each oxygen site on one side of the symmetric slabs
the stable surfaces include: i) 𝜇2 -OH/𝜇2 -OH and 𝜇1 -OH/𝜇1 -OH; ii) 𝜇2 -O/𝜇2 -OH
and 𝜇1 -OH/𝜇1 -OH; iii) 𝜇2 -O/𝜇2 -O and 𝜇1 -OH/𝜇1 -OH; iv) 𝜇2 -O/𝜇2 -O and 𝜇1 -O/𝜇1 -
OH; v) 𝜇2 -O/𝜇2 -O and 𝜇1 -O/𝜇1 -O (see Figure 4 in Results). Each surface was
equilibrated in water for 20 ps using a 1 fs timestep at 350 K (the elevated temperature
was used to account for the PBE induced over structuring of water). A Berendsen
thermostat controlled the ionic temperature with dt/d𝜏 = 1/50 during the equilibration.
The surface (electrochemical) potential vs. pH phase diagram was generated using
the computational hydrogen electrode [22]. Configurational entropy for equivalent
adsorption sites and the zero-point energies associated with 𝜇1 -OH and 𝜇2 -OH,
taken from calculations on surfaces relaxed in vacuum, were included. The surface
energies were computed using molecular dynamics snapshots. In an effort to reduce
the influence of the error in the energy of bulk water in these snapshots [14] on the
computational phase diagram, we defined an interfacial energy following Ref. [38].
That is:
𝑁
1 ∑︁

⟨𝐸 𝑖𝑛𝑡 ⟩ = 𝐸 𝑡𝑜𝑡 ,𝑖 − 𝐸 𝐻2 𝑂,𝑖 − 𝐸 𝑠𝑢𝑟 𝑓 ,𝑖 . (2)
2𝑁 𝑖=1
Here the factor of 1/2 accounts for the two sides of slab and 𝑁 is the number of
snapshots considered in the sum. 𝐸 𝑡𝑜𝑡 ,𝑖 represents the total energy of the solvated
slab in the 𝑖th snapshot. 𝐸 𝐻2 𝑂,𝑖 is the total energy of the water in the 𝑖th snapshot
in the absence of the slab. 𝐸 𝑠𝑢𝑟 𝑓 ,𝑖 is the energy of the slab in the 𝑖th snapshot
without water. These energies were all computed using a (4 × 4) k-point mesh with
Marzari–Vanderbilt smearing (𝜎 = 0.02 Ry) [17]. The sum in the equation above was
continued until the energy convergence of the interfacial energy contribution was
better than 0.1 eV. This required less than 32 snapshots per hydrogen coverage. The
hydrogen coverage dependence on applied potential was computed by interpolating
between the computed hydrogen coverages using a quadratic function to capture
the solvent induced Frumkin behavior we found to be associated with hydrogen
adsorption.
Minimum energy paths were computed by starting with an molecular dynamics
snapshot of the 𝜇2 -O/𝜇2 -O and 𝜇1 -O/𝜇1 -OH surface and retaining 2 water bilayers on
one side of the slab while introducing ≈ 15 Å of vacuum to separate periodic images.
Note, while this is not required for constant charge simulations, the constant bias
simulations require vacuum; to facilitate comparison between the two approaches we
employed the same cell dimensions for both constant charge and bias simulations.
After introducing the vacuum layer, 𝜇1 -OH was removed from the non-solvated side
138 Travis Jones

of the slab and the bottom two layers of atoms on this side were fixed to their bulk
coordinates. The fixed charge minimum energy paths were computed using zero
net charge in the simulation cell along the entire length of the path. The fixed bias
minimum energy path simulations were performed using the effective screening
medium method [2, 23], with the potential of the neutral interface in the initial
state and biases 0.1 V to 0.5 V anodic of this zero charge potential. The climbing
image nudged elastic band method [12] was used with 8 images per path to locate
transition states with a single climbing image/transition state per path. The paths
were considered to be converged when the force on each image was below 0.05 eV/Å.
The possible role of finite temperature effects was explored using metadynamics
performed at 350 K [3]. As with the minimum energy path simulations, these
metadynamics simulations were performed under both constant charge and constant
bias conditions on asymmetric slabs with a solvation layer on one side. Here, however,
the full solvation layer was preserved during the molecular dynamics simulations,
which used the parameters noted above, and ≈ 15 Å of vacuum to separate periodic
images. To keep the water confined near the surface during the simulations, a potential
wall was placed at about ≈ 12 Å above the surface. Free energy barriers were
computed by biasing the O—O distance of oxygen involved in O-O coupling using
Gaussian kernels stored on a grid. Each free energy barrier was computed through
three separate simulations, each of which employed 8 parallel walkers to explore the
free energy landscape.
The metadynamics and minimum energy path calculations proved to be the
most computationally demanding aspects of the work. A similar parallelization
strategy was used for each. In the metadynamics simulations DFT based molecular
dynamics was performed while biasing a collective variable describing the reaction
coordinate, where multiple replicas of the systems, in this work 8, with different
starting configurations were used to explore the free energy surface. To allow sufficient
time for the simulation to sample the free energy landscape the system must was
biased slowly. In this work the total simulation time exceeded 10 ps in each case,
which required tens of thousands of DFT force evaluations. Similarly, for minimum
energy path simulations using the nudged elastic band algorithm, a minimum energy
path was first discretized into a series of images along the path by interpolating the
atomic rearrangement between reactant and product. The forces on the atoms in
each of the (8) images were then computed. The atoms were allowed to relax using
the computed forces that are tangent to the initial path, whereas fictitious spring
forces were applied along the path to keep the images from collapsing into a common
minimum [12]. As the images relaxed to the minimum energy path, the highest energy
image was allowed to feel the full force; however, the force component normal to the
path was inverted to drive the highest energy replica towards the transition state. To
make these problems computationally tractable efficient parallelization was required
In this work the systems of interest are solvated IrO2 surfaces containing approx-
imately 110 atoms, 300 Kohn–Sham states, and 16 total k-points and 8 replicas.
Figure 3 shows scaling data for this system. Owing to the low communication between
replicas in both metadynamics and minimum energy path calculations, parallelization
over replicas results in trivial linear scaling, though limited by the small number of
Electro-catalysis for H2 O oxidation 139

Fig. 3: Scaling data for a single replica of the simulation cell used in this work. As
indicated in the figure legend, data is shown for both a coarse and fine k-point mesh
together with the effect of different parallelization strategies.

replicas. It is then essential to develop a parallelization for each replica. To test the
scalability of one replica of the system, a 4-layer oxygen terminated IrO2 slab was
used with 18 water molecules as a solvent layer. This system was tested with both a
coarse (4 × 4) k-point mesh and a fine (8 × 8) k-point mesh. In both cases the k-points
were distributed over 𝑛 𝑝𝑜𝑜𝑙 pools of CPUs. The communication requirement between
these pools is very low and leads to trivial linear scaling that is limited by the small
number of k-points. Within each pool, 𝑛 𝑃𝑊 CPUs work on a single k-point, and the
wave-function coefficients are distributed over the 𝑛 𝑃𝑊 CPUs. From previous work it
was found this level of parallelization is limited. To further extend the scaling, linear
algebra operations needed for subspace diagonalization were also distributed. These
operations are performed on ≈ 𝑁 𝑏 × 𝑁 𝑏 square matrices, where 𝑁 𝑏 is the number
of Kohn–Sham states, and were thus distributed on a square grid of 𝑛2𝑑 cores, with
𝑛2𝑑 ≤ 𝑛 𝑃𝑊 . Figure 3 shows this parallelization strategy is acceptable up to 1024 cores
per replica, where the efficiency is 88%. This efficiency can be improved slightly by
employing 2 OpenMP threads to give an efficiency of 92%. For the same system,
scaling can be furthered by dividing the 𝑛 𝑃𝑊 cores into 𝑛𝑡 𝑎𝑠𝑘 task groups, where
each task group handles the FFT over 𝑁 𝑏 /𝑛𝑡 𝑎𝑠𝑘 states. As seen in Figure 3, using two
task groups allows the scaling to be extended to at least 2048 cores per replica, though
this comes with a drop in efficiency to 71%. As all problems require 8 replicas, 1024
cores per replica is preferred as each problem can then be fit on 64 nodes. Increasing
the k-point mesh to improve the energy convergence results in a larger number of
CPU pools, which increases the scaling when following the aforementioned approach.
As with the coarse mesh, 1024 cores yields the most efficient approach and allows the
use of 64 nodes for both the minimum energy path and metadynamics simulations.
With this parallelization strategy it became possible to explore the surface phases
of IrO2 and the OER mechanism on Hawk using implicit solvent, as discussed below.
140 Travis Jones

3 Results and discussion

Figure 4a shows an example of the symmetric slab used to compute the surface
energies needed to construct the surface (electrochemical) potential vs. pH phase
diagram, while Figure 4b-f shows examples the low energy surface terminations
found by equilibration using DFT molecular dynamic; solvent has been removed for
clarity. Going from Figure 4b to Figure 4f represents an increase in oxidative charge
coverage (𝜃 ℎ+ ) from 0 monolayer (ML) to 1 ML, respectively. While we considered
all hydrogen coverages, those shown in Figure 4 represent the lowest energy states.
It can be seen that hydrogen is first removed from 𝜇2 -O sites in Figure 4c and d,
before hydrogen is lost from 𝜇1 -O sites, Figure 4e and f. Oxidation beyond the 1 ML
𝜃 ℎ+ surface shown in Figure 4f leads to 𝜇1 -OOH and 𝜇1 -OO, for which we have no
experimental evidence [21] and will thus not be considered further in this work.

a) c) d)

2-OH
b)

e) f)

1-OH

Fig. 4: a) Example of cell used for thermal equilibration. Ir atoms are grey, O
as red, and H as white. Low energy terminations of the IrO2 (110) surface b) 𝜇2 -
OH/𝜇2 -OH and 𝜇1 -OH/𝜇1 -OH, c) 𝜇2 -O/𝜇2 -OH and 𝜇1 -OH/𝜇1 -OH, d) 𝜇2 -O/𝜇2 -O
and 𝜇1 -OH/𝜇1 -OH, e) 𝜇2 -O/𝜇2 -O and 𝜇1 -O/𝜇1 -OH, f) 𝜇2 -O/𝜇2 -O and 𝜇1 -O/𝜇1 -O.

The surface energies of the terminations shown in Figure 4 can be used to construct
a (electrochemical) potential vs. pH phase diagram following the procedure outlined
in the Methods section. Figure 5 shows equilibration in water introduces Frumkin
behavior and broadens the surface oxidation window such that it extends over ≈ 1
V. Here it is worth noting only 1/4 ML intervals are shown, though 𝜇2 -OH to
𝜇2 -O oxidation is predicted to begin at ≈ 1 V vs SHE. This onset and the wide
surface oxidation window are in good agreement with experimental findings [21].
By comparison, in the absence of explicit solvent surface oxidation is complete by
1.5 V [25], in contrast to experimental findings. From the phase diagram computed
with explicit solvent, the active surface under acidic conditions can be expected to
Electro-catalysis for H2 O oxidation 141

θh+~1

θh+=
θh+= /4
θh+=

3/4
pH

1/2
1
θh+=0

Potential / NHE (V)

Fig. 5: Computed surface pH vs. potential (normal hydrogen electrode, NHE) phase
diagram. Constructed for rutile-type IrO2 (110) surface. The total oxidative hole
coverage with respect to the fully protonated surface indicated.

have an oxidative charge coverage ranging from 1/4 ML to 1 ML depending on the

applied bias. Importantly, the oxidative charge coverage is predicted to change over
the relevant operating potentials (see Figure 2), and this change in surface oxidation
state may be expected to influence the electrocatalytic reaction rate.
Minimum energy path calculations were used to test the effect of oxidative charge
on the OER rate by first assuming the elementary steps under acidic conditions can
be described as four proton-coupled electron transfers [25, 30]:

S + H2 O → S − OH + e− + H+ (3)
− +
S − OH → S − O + e + H (4)
S − O + H2 O → S − OOH + e− + H+ (5)
− +
S − OOH → S − OO + e + H , (6)
where S is an empty 𝜇1 site on the surface.
Of these steps, analysis of experimental Tafel slopes has been used to suggest the
second (Equation 4) is rate-limiting [7]. Computational studies using implicit solvent
have suggested step 3 (Equation 5) is the rate determining step and the remaining
proton-coupled electron transfers are barrierless [25]. In this work, water nucleophilic
attack (step 3) was found to be rate-limiting in the presence of the explicit solvent and
will be the focus of the remainder of the discussion, see Ref. [21] for more details.
With the rate determining step identified in a realistic model of the electrified
solid/liquid interface it becomes possible to explore the validity of Butler–Volmer
theory for the OER by focusing on water nucleophilic attack. First consider the
activation energy (𝐸 𝑎 ) for this oxyl-water coupling step on a surface with 𝜃 ℎ+ =3/4
142 Travis Jones

ML (Figure 4e) at the surface’s potential of zero charge (pzc). From Figure 5, this
surface oxidation states is consistent with the surface at high overpotential. Computing
the minimum energy path for O-O coupling with the electrochemical bias fixed to the
surface’s pzc shows O-O coupling occurs with the concerted transfer of hydrogen to
a 𝜇2 -O site, see Figure 6a. The activation energy for this elementary step is 0.63 eV,
and the heat of reaction (Δ𝐻rxn ) is near zero. This data is denoted in the plot of 𝐸 𝑎 vs.
Δ𝐻rxn by a green triangle in Figure 6b. The two extra green triangles at Δ𝐻rxn ≈ 0 eV
in Figure 6b show that increasing the bias by 0.1 V and 0.5 V to capacitively charge
the system while holding the oxidative charge constant at 𝜃 ℎ+ = 3/4 ML has little
impact on 𝐸 𝑎 . The change introduced by the capacitive charge is indicated by the
arrow labeled QC .
From Figure 5 it can be seen that constraining 𝜃 ℎ+ to 3/4 ML to capacitively
charge the system by 0.5 V is not realistic. Under such a bias the surface is predicted
to oxidize to 𝜃 ℎ+ ≈ 1 ML. Including this oxidative charge yields the total charge,
QT , which, as shown in Figure 6b, reduces Δ𝐻rxn to ≈-0.2 eV. This oxidation charge
also lowers 𝐸 𝑎 from ≈0.60 eV to 0.23 eV. Similarly, reducing 𝜃 ℎ+ from 3/4 to 1/2
ML increases 𝐸 𝑎 to 0.78 eV, see QT′ in Figure 6b; this latter change cannot be
compensated by pure capacitive charging (QC′ ). Thus, oxidative charge appears to
have a stronger influence on the activation energy than capacitive charge.
Another way of viewing the data in Figure 6 is through the Brønsted–Evans–Polanyi
(BEP) relationship. The BEP relationship is a linear correlation between 𝐸 𝑎 and Δ𝐻rxn
often observed for chemical bond making and breaking steps [15], and in thermal
catalysis in particular. From Figure 6 we can see such a correlation exists for O-O
coupling obeys the Brønsted–Evans–Polanyi (BEP) relationship. This relationship
holds as changes in Δ𝐻rxn appears to be dominated by oxidative charge rather than
purely capacitive charge. Such behavior would be expected when considering classical
electrochemical theory, as inner-sphere reaction kinetics are thought to be insensitive
to building up the double layer. To test this concept explicitly, minimum energy path
calculations were performed without holding the electrochemical potential fixed, that
is, a second set of minimum energy path calculations simulations were performed
under a fixed charge condition.
The degree to which 𝜃 ℎ+ alone mediates the BEP relationship can be found by
computing the minimum energy paths for O-O coupling using a fixed number of
electrons rather than fixing the electrochemical bias as this does not allow capacitive
charging. While this approach does not capture the experimental conditions, it
does offer a computational means of testing the hypothesis that charge mediates
the observed BEP relationship. As expected from classical electrochemical theory,
fixing the charge during the minimum energy path calculations does not change
the mechanism or break the BEP relationship, see the filled squares in Figure 6b.
Thus, the BEP relationship holds and 𝐸 𝑎 = 𝐸 0 + 𝛼Δ𝐻rxn even in the absence of
electron transfer. Moreover, replacing the spectator O/OH species on the surface with
adsorbed Cl (an experimentally testable situation) reveals the BEP slope 𝛼 is primarily
controlled by the charge on the ligand rather than chemical nature of the ligand. These
results lead to a situation familiar from traditional thermal catalysis: as the surface
becomes more reduced or oxidized, 𝐸 𝑎 increases or decreases, respectively. Thus,
Electro-catalysis for H2 O oxidation 143

a)

b) c)

QC’
QC QT’

QT

Fig. 6: a) The initial and final states for water-oxyl coupling on the IrO2 (110) surface.
During the reaction a water molecule near the oxyl forms an O-O bond with 𝜇1 -O.
During this step hydrogen is transferred from water to a surface 𝑚𝑢 2 -O through a
Zundel-like species. The final state of the elementary step has adsorbed 𝜇1 -OOH on
the surface. b) The activation energy computed for O-O coupling along the minimum
energy path plotted as a function of the corresponding heat of reaction. The heat
of reaction can be seen to become more negative as the oxidative surface charge
increases. The green triangles show 𝐸 𝑎 computed from the minimum energy paths
for oxyl-water coupling on surfaces with 𝜃 ℎ+ =1/2 ML, 3/4 ML, and 1ML (in order
of increasing exothermicity) computed under constant potential conditions. In all
cases the computed activation energies include reactions on the surfaces with no
net charge in the initial state. For 𝜃 ℎ+ of 1/2 ML and 3/4 ML the green triangles
also include results with capacitive charging from 0.1 V to 0.5 V. These results are
denoted by the label QC for the 𝜃 ℎ+ = 3/4 ML and the label QC′ for the 𝜃 ℎ+ = 1/2
ML surface. The small solid arrows under the QC and QC′ labels show how the
activation energy drops marginally with capacitive charging even as high as 0.5 V.
The dashed arrow labeled QT shows the effect of allowing the complete charging
of the surface/double layer by including oxidative charging, rather than capacitive
charging alone, when the 𝜃 ℎ+ = 3/4 ML surface is biased by an additional 0.5 V.
Similarly, QT′ shows the effect of reducing the 𝜃 ℎ+ = 3/4 ML surface to 𝜃 ℎ+ = 1/2
ML. By way of comparison, the squares show the computed activation energies as
a function of Δ𝐻rxn under constant charge conditions. The unfilled squares show
surfaces with adsorbed Cl, which were used to investigate a non-reducible ligand.
Circles show the results of minimum energy path calculations without solvent. c) The
activation energies computed along the 0 K minimum energy path (labeled MEP)
and the activation free energies computed using metadynamics (labeled MD) plotted
against the coverage of oxidative charge.
144 Travis Jones

the activation energy is linearly dependent not only on Δ𝐻rxn , but on 𝜃 ℎ+ , as shown
in Figure 6c. In this case, the activation energy can be written as 𝐸 𝑎 = 𝜁 𝜃 ℎ+ + 𝜅 in
close analogy with the tradition BEP relationship, where the constants 𝜁 and 𝜅 now
describe the linear dependence of 𝐸 𝑎 on the total oxidative charge. This relationship
also holds for the activation free energy from metadynamics and so does not appear
to be an artifact of the 0 K minimum energy path calculation, see Figure 6c.
Identifying the linear dependence of the activation energy for O-O coupling on
oxidative charge allows the per-site electrocatalytic response of IrO2 to be expressed
though an Eyring-like equation for the OER current:
 
𝑘 B𝑇 𝜁 𝜃 ℎ+ + 𝜅
𝑖 = 4|𝑒| 𝜃 𝜇1 exp − , (7)
ℎ 𝑘 B𝑇
In the above equation, 𝑘 B is the Boltzmann constant; ℎ is the Planck constant; 𝑇 is
the temperature; 𝜃 𝜇1 is the coverage 𝜇1 -O species; 𝜃 ℎ+ is the total oxidative charge
coverage. The factor of 4 in the equation accounts for the 4 electrons transferred
during the OER. Note that the OER rate response to the electrochemical bias in
Equation 7 is captured in the exponential dependence on the oxidative charge rather
than the exponential dependence directly
on the overpotential, as predicted by
Butler–Volmer theory;

e.g. 𝑖 ∝ exp 𝜂/𝑘 B𝑇 in Butler–Volmer theory rather than the
𝑖 ∝ exp 𝜃 ℎ+ /𝑘 B𝑇 found in this work.
It is now possible to predict the electrocatalytic response of IrO2 , in particular the
Tafel plot, using Equation 7 together with the computed phase diagram. To do so,
we can approximate the activation free energy in the Eyring-like equation using the
computed activation energies or we can take the computed activation free energies
from the metadynamics simulations. Using the BEP relationship computed from the
minimum energy reveals good agreement with experiment, see the blue and red fit
lines in Figure 7. From these, the computed Tafel plot has a Tafel slope of 39 mV/dec
up to 1.58 V, at which point the slope increases to 77 mV/dec. By way of comparison,
crystalline IrO2 has Tafel slopes ranging from 43-47 mV/dec and 71-76 mV/dec over
the same potential windows [21], see Figure 2. Using the computed activation free
energies also results in good agreement with experiment, with Tafel slopes of 36
mV/dec up to 1.58 V and 69 V/dec above 1.58 V, see the brown and purple fit lines
in Figure 7 and the Tafel equations labeled 𝜂MD . Now from the calculations, however,
the bend in the Tafel slope, irrespective of how the BEP relationship is approximated,
can be ascribed to a change in the response of 𝜃 ℎ+ to the applied electrochemical bias
instead of the change in mechanism suggested by Butler–Volmer theory [7]. Here
the 𝜇1 -OH species are become depleted at 1.58 V, and, as a result, further increases
in the potential beyond this value result in less oxidative charge storage. Beyond
describing this behavior, the simulations also reveal the log OER current is linear in
𝜃 ℎ+ , Figure 7b. This prediction is consistent with experiment, with Figure 7c showing
an experimental example of the log OER current response of crystalline IrO2 as
function of the total charge stored in the electrocatalyst. As noted above, the total
charge is dominated by the oxidative contribution, 𝜃 ℎ+ ; note, however, experimentally
the active surface area of this sample is unknown so the measured charge cannot be
converted directly into 𝜃 ℎ+ . Regardless, the good agreement between the computed
Electro-catalysis for H2 O oxidation 145

and measured trends suggests bond rupture/formation is slow compared to electron

transfer. As a result, the OER rate displays BEP behavior and falls outside the scope
of Butler–Volmer theory.

a) b)

c)

Fig. 7: a) The computed Tafel plot of IrO2 . The Tafel plots computed using the
activation energies from the minimum energy path calculation are fit with blue and
red fit lines and the corresponding Tafel equations are shown in the lower right. The
Tafel plots computed using the activation free energies are fit with brown and purple
fit lines and the corresponding Tafel equations are shown in the upper left and denoted
𝜂MD . b) The computed 𝜃 ℎ+ vs. log OER current plot of IrO2 . c) The measured total
charge vs. log OER current plot of IrO2 calcined at 450 ◦ C, data from Ref. [21].

4 Summary

By leveraging the excellent scalability of Quantum ESPRESSO on Hawk it was

possible to calculate the surface (electrochemical) potential vs. pH phase diagram
of the (110) surface of rutile-type IrO2 in the presence of explicit solvent. Doing
so revealed solvent induced Frumkin behavior that broadens the surface oxidation
windows across the range of potentials relevant for the OER under acidic conditions.
Minimum energy path calculations and metadynamics performed in the presence
of explicit solvent under constant bias and constant charge conditions revealed this
oxidative charging controls the OER rate, rather than the direct action of the bias on
the reaction coordinate commonly assumed. These results predict bi-linear Tafel plots
and a linear response of log OER current to total charge, both of which are confirmed
by experiment. Recognizing this role of oxidative charge in OER electrocatalysis
146 Travis Jones

demonstrates inner-sphere chemistry controls electrocatalytic OER and provides a

fundamental link between thermal- and electrocatalysis. This will enable the tools
and concepts developed to understand and engineer traditional catalysts to be applied
to their electrochemical counterparts.

References

1. A. J. Bard and L. R. Faulkner. Electrochemical Methods: Fundamentals and Applications.

Wiley, 2nd edition, 2001.
2. N. Bonnet, T. Morishita, O. Sugino, and M. Otani. First-principles molecular dynamics at a
constant electrode potential. Phys. Rev. Lett., 109:266101, 2012.
3. G. Bussi and A. Laio. Using metadynamics to explore complex free-energy landscapes. Nat.
Rev. Phys, 2:200–212, 2020.
4. A. Buttler and H. Spliethoff. Current status of water electrolysis for energy storage, grid
balancing and sector coupling via power-to-gas and power-to-liquids: A review. Renew. Sus.
Energy Rev., 82:2440–2454, 2018.
5. M. Carmo, D. L. Fritz, J. Mergel, and D. Stolten. A comprehensive review on PEM water
electrolysis. Inter. J. Hydrogen Energy, 38:4901–4934, 2013.
6. A. D. Corso. Pseudopotentials periodic table: From H to Pu . Comput. Mater. Sci., 95:337 –
350, 2014.
7. L. De Faria, J. Boodts, and S. Trasatti. Electrocatalytic properties of ternary oxide mixtures of
composition Ru0 .3Ti (0.7?𝑥) Ce 𝑥 O2 : oxygen evolution from acidic solution. J. Appl. Electrochem.,
26:1195–1199, 1996.
8. P. De Luna, C. Hahn, D. Higgins, S. A. Jaffer, T. F. Jaramillo, and E. H. Sargent. What would it
take for renewably powered electrosynthesis to displace petrochemical processes? Science, 364,
2019.
9. J. A. Gauthier, C. F. Dickens, L. D. Chen, A. D. Doyle, and J. K. Nørskov. Solvation Effects for
Oxygen Evolution Reaction Catalysis on IrO2 (110). J. Phys. Chem. C, 121:11455–11463, 2017.
10. P. Giannozzi, S. Baroni, N. Bonini, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, G. L.
Chiarotti, M. Cococcioni, I. Dabo, A. Dal Corso, S. de Gironcoli, S. Fabris, G. Fratesi,
R. Gebauer, U. Gerstmann, C. Gougoussis, A. Kokalj, M. Lazzeri, L. Martin-Samos, N. Marzari,
F. Mauri, R. Mazzarello, S. Paolini, A. Pasquarello, L. Paulatto, C. Sbraccia, S. Scandolo,
G. Sclauzero, A. P. Seitsonen, A. Smogunov, P. Umari, and R. M. Wentzcovitch. QUANTUM
ESPRESSO: a modular and open-source software project for quantum simulations of materials.
J. Phys.: Condens. Matter, 21(39):395502, 2009.
11. T. M. Gür. Review of electrical energy storage technologies, materials and systems: challenges
and prospects for large-scale grid storage. Energy Environ. Sci., 11:2696–2767, 2018.
12. G. Henkelman, B. P. Uberuaga, and H. Jónsson. A climbing image nudged elastic band method
for finding saddle points and minimum energy paths. J. Chem. Phys., 113(22):9901–9904,
2000.
13. C. Hu, L. Zhang, and J. Gong. Recent progress made in the mechanism comprehension and
design of electrocatalysts for alkaline water splitting. Energy Environ. Sci., 12:2620–2645,
2019.
14. D. T. Limmer, A. P. Willard, P. Madden, and D. Chandler. Hydration of metal surfaces can be
dynamically heterogeneous and hydrophobic. PNAS, 110(11):4200–4205, 2013.
15. R. A. Marcus. Theoretical relations among rate constants, barriers, and Broensted slopes of
chemical reactions. J. Phys. Chem., 72:891–899, 1968.
16. R. A. Marcus. Electron transfer reactions in chemistry. Theory and experiment. Rev. Mod.
Phys., 65:599–610, 1993.
17. N. Marzari, D. Vanderbilt, A. De Vita, and M. C. Payne. Thermal contraction and disordering
of the Al(110) surface. Phys. Rev. Lett., 82:3296–3299, 1999.
Electro-catalysis for H2 O oxidation 147

18. Y. Matsumoto and E. Sato. Electrocatalytic properties of transition metal oxides for oxygen
evolution reaction. Mater. Chem. Phys., 14:397–426, 1986.
19. J. T. Mefford, Z. Zhao, M. Bajdich, and W. C. Chueh. Interpreting Tafel behavior of consecutive
electrochemical reactions through combined thermodynamic and steady state microkinetic
approaches. Energy Environ. Sci., 13:622–634, 2020.
20. T. Naito, T. Shinagawa, T. Nishimoto, and K. Takanabe. Recent advances in understanding
oxygen evolution reaction mechanisms over iridium oxide. Inorg. Chem. Front., 8:2900–2917,
2021.
21. H. N. Nong, L. J. Falling, A. Bergmann, M. Klingenhof, H. P. Tran, C. Spöri, R. Mom,
J. Timoshenko, G. Zichittella, A. Knop-Gericke, S. Piccinin, J. Pérez-Ramírez, B. R. Cuenya,
P. S. Robert Schlögl, D. Teschner, and T. E. Jones. Key role of chemistry versus bias in
electrocatalytic oxygen evolution. Nature, 587:408–413, 2020.
22. J. K. Nørskov, J. Rossmeisl, A. Logadottir, L. Lindqvist, J. R. Kitchin, T. Bligaard, and
H. Jónsson. Origin of the overpotential for oxygen reduction at a fuel-cell cathode. J. Phys.
Chem B, 108(46):17886–17892, 2004.
23. M. Otani and O. Sugino. First-principles calculations of charged surfaces and interfaces: A
plane-wave nonrepeated slab approach. Phys. Rev. B, 73:115407, 2006.
24. J. P. Perdew, K. Burke, and M. Ernzerhof. Generalized gradient approximation made simple.
Phys. Rev. Lett., 77:3865–3868, 1996.
25. Y. Ping, R. J. Nielsen, and W. A. Goddard. The Reaction Mechanism with Free Energy Barriers
at Constant Potentials for the Oxygen Evolution Reaction at the IrO2 (110) Surface. J. Am.
Chem. Soc., 139:149–155, 2017.
26. G. Prandini, A. Marrazzo, I. Castelli, N. Mounet, and N. Marzari. Precision and efficiency in
solid-state pseudopotential calculation. npj Comput. Mater., 4:72, 2018.
27. T. Reier, H. N. Nong, D. Teschner, R. Schlögl, and P. Strasser. Electrocatalytic oxygen evolution
reaction in acidic environments: Reaction mechanisms and catalysts. Adv. Energy Mater.,
7:1601275, 2017.
28. T. Reier, M. Oezaslan, and P. Strasser. Electrocatalytic Oxygen Evolution Reaction (OER) on
Ru, Ir, and Pt Catalysts: A Comparative Study of Nanoparticles and Bulk Materials. ACS Catal.,
2:1765–1772, 2012.
29. J. Rossmeisl, Z.-W. Qu, H. Zhu, G.-J. Kroes, and J. Nørskov. Electrolysis of water on oxide
surfaces. J. Electroanal. Chem., 607:83–89, 2007.
30. J. Rossmeisl, Z.-W. Qu, H. Zhu, G.-J. Kroes, and J. Nørskov. Electrolysis of water on oxide
surfaces. J. Electroanal. Chem., 607:83–89, 2007.
31. R. S. Schlögl. Sustainable Energy Systems: The Strategic Role of Chemical Energy Conversion.
Topic. Catal., 59:772–786, 2016.
32. W. Schmickler and Santos. Interfacial Electrochemistry. Springer, 2010.
33. Z. W. Seh, J. Kibsgaard, C. F. Dickens, I. Chorkendorff, J. K. Nørskov, and T. F. Jaramillo.
Combining theory and experiment in electrocatalysis: Insights into materials design. Science,
355, 2017.
34. L. C. Seitz, C. F. Dickens, K. Nishio, Y. Hikita, J. Montoya, A. Doyle, C. Kirk, A. Vojvodic,
H. Y. Hwang, J. K. Nørskov, and T. F. Jaramillo. A highly active and stable IrO 𝑥 /SrIrO3 catalyst
for the oxygen evolution reaction. Science, 353:1011–1014, 2016.
35. T. Shinagawa, A. Garcia-Esparza, and K. Takanabe. Insight on Tafel slopes from a microkinetic
analysis of aqueous electrocatalysis for energy conversion. Sci. Rep., 5:13801, 2015.
36. J. Song, C. Wei, Z.-F. Huang, C. Liu, L. Zeng, X. Wang, and Z. J. Xu. A review on fundamentals
for designing oxygen evolution electrocatalysts. Chem. Soc. Rev., 49:2196–2214, 2020.
37. N.-T. Suen, S.-F. Hung, Q. Quan, N. Zhang, Y.-J. Xu, and H. M. Chen. Electrocatalysis for
the oxygen evolution reaction: recent development and future perspectives. Chem. Soc. Rev.,
46:337–365, 2017.
38. D. Wang, T. Sheng, J. Chen, H.-F. Wang, and P. Hu. Identifying the key obstacle in photocatalytic
oxygen evolution on rutile TiO2 . Nat. Catal., 1:291–299, 2018.
 Materials Science

In materials science or material science challenges, computational methods require

extensive computational capacity. Modern simulation methods from this research area
are efficiently parallelised and show a good scaling behaviour on high-performance
computers. This part presents a selection of materials science projects (Def-2-Dim
and Pace3D), from numerous research activities, that provide new insights into
the change of materials on an atomic and microstructural scale, by describing the
performance of phase formation and damage simulations using HPC systems. The
results can be used to derive design instructions for data-driven, accelerated material
development.
The first project (Def-2-Dim) elaborates material simulation methods, such as first
principal and molecular dynamics calculations, that operate on an atomic scale, so as to
investigate the formation of defects in two-dimensional materials, under the influence
of electron and ion irradiation. The quantification of defects, obtained through the
investment of extensive computational resources, is important to determine their
effect on the electronic, optical and catalytic material properties and provides insights
into the post-synthetical doping of two-dimensional materials, using transition metal
atoms and a dislocation-mediated mechanism.
The second project (Pace3D) introduces a selected and dedicated application of
the high-performance materials simulation framework Pace3D to three-dimensional
microstructure simulations on a mesoscopic scale. Large-scale simulations on high-
performance clusters are performed with a convincing scalability behaviour that
enables the resolution of large-scale microstructure sections. Examples of phase
evolution in solid-state phase transformations, involving phase and grain interfaces,
are presented and discussed.
The third project (DiHu) chosen for this part illustrates the application of materials
science simulation methods to describe the multi-scale, biomechanical phenomena
of contracting muscles and highlights the demand of high-performance computing
techniques and resources. To study the combined process of muscle contraction
and electromyography, the research is conducted using the open-source software
OpenDiHu, which shows that the new multi-scale approach is highly scalable.

149
150 Materials Science

To conclude this part, simulations for the prediction of muscle contractions

(DiHu) are presented as a cross-material application. The results demonstrate the
computational effort required to capture multi-scale processes and show the transfer
of materials science and solid mechanics methods to life science topics.

Institute for Applied Materials – Britta Nestler

Computational Materials Science (IAM-CMS),
Karlsruhe Institute of Technology (KIT),
Straße am Forum 7, 76131 Karlsruhe, Germany,
e-mail: [email protected]
Production of defects in two-dimensional
materials under ion and electron irradiation:
insights from advanced first-principles
calculations
Annual report on ‘Bundesprojekt’ ACID 44172
(Def-2-Dim)

S. Kretschmer, T. Joseph, S. Ghaderzadeh, Y. Wei, M. Ghorbani-Asl and

A.V. Krasheninnikov

Abstract The main goal of the research project with the extensive use of GCS
computational resources was to study the formation of defects in two-dimensional
(2D) materials under electron and ion irradiation using atomistic simulations. The
influence of defects on the electronic, optical and catalytic properties of 2D materials
has also been investigated. Specifically, the role of electronic excitations in the
production of defects under electron irradiation was elucidated, and the types of defect
formed in 2D MoS2 sheets upon cluster impacts were identified. The first-principles
calculations provided insights into the post-synthesis doping of 2D materials with
transition metal atoms through dislocation-mediated mechanism, and also allowed for
the understanding of how the implanted species (e.g., Cl atoms) affect the electronic
properties of 2D transition metal dichalcogenides. Also analytical potential molecular
dynamics simulations were used to study the behavior of 2D materials under ion and
cluster irradiation. The role of adatoms and surface reconstructions in the novel 2D
material hematene was addressed as well. In addition to publications in peer-refereed
scientific journals (8 papers published, see Refs. [1–8] and 4 manuscripts are currently
under review), the obtained results were also disseminated through popular articles at
different internet resources, including the GCS webpage.

A.V. Krasheninnikov
Institute of Ion Beam Physics and Materials Research, Helmholtz-Zentrum Dresden-Rossendorf,
Bautzner Landstraße 400, 01328 Dresden,
Tel.: +49 351 260 3148
Fax: +49 351 260 0461
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 151
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_9
152 S. Kretschmer et al.

1 Introduction

Since the beginning of the 20th century [9, 10], the interaction of energetic particles –
electrons and ions – with matter has been the subject of intensive studies. The research
in this area has been motivated first by the necessity to assess the irradiation-induced
damage in the materials in radiation-harsh environments such as fission/fusion reactors
or cosmic space. It was also realized later on that in spite of the damage, irradiation may
have overall beneficial effects on the target [11,12]. A good example is the industrially
important ion implantation into semiconductors [11]. An additional motivation to
study effects of electron irradiation on materials comes from transmission electron
microscopy (TEM). Every day hundreds, if not thousands, of microscopists in the
world get insights into material structure and cope with the beam-induced damage: it
is often not clear if the observed defects were present in the sample before it was put
into the TEM column or whether they are artefacts of the imaging technique.
The past decade has also witnessed an enormous interest in nanosystems and
specifically two-dimensional (2D) materials [13] such as graphene or transition
metal dichalcogenides (TMDs), which have already shown good potentials [14–16]
for nanoelectronics, photonics, catalysis, and energy applications due to a unique
combination of electronic, optical, and mechanical properties. It has also been
demonstrated that irradiation, especially when combined with heat treatment, can
have beneficial effects on 2D systems. For example, the bombardment of graphene with
low-energy ions can be used for increasing the Young’s modulus of the system [17],
implanting dopants [18], or adding new functionalities, like magnetism[19]. Beams of
energetic electrons, which nowadays can be focused to sub-Å areas, have been shown to
work as cutting and welding tools on the nanoscale [20,21] or stimulate local chemical
reactions. Further developments of irradiation-assisted methods of 2D materials
treatment require complete microscopic theory of defect production in these systems.
Moreover, 2D materials give a unique opportunity to understand at the fundamental
level the energy deposition and neutralization of the highly-charged ions [22, 23] by
measuring their initial/final charge states and kinetic energy. On the theory side, the
reduced dimensionality makes it possible to carry out computationally very expensive
non-adiabatic first-principles calculations [24–26] to assess the interaction of the ions
with the target.
However, a growing body of experimental facts indicate that many concepts
of energetic particle-solid interaction are not applicable to these systems, as the
conventional approaches based on averaging over many scattering events do not work
due to their very geometry, as, e.g., in graphene – a membrane just one atom thick
– or require substantial modifications. It is intuitively clear that the impact of an
energetic particle will not give rise to a collisional cascade, as in bulk solids, but
cause sputtering. A different electronic structure and high surface-to-volume ratio
also affect the redistribution of the energy deposited by the energetic particle in the
system and thus influence defect production. As for TMDs, experiments indicate
that defects are produced at electron energies well below the knock-on threshold
(the minimum electron energy required to ballistically displace an atom from the
material), and the mechanism of damage production (electronic excitations, etching)
Annual report on Def-2-Dim project 153

in these systems is not fully understood at the moment. The conventional theory based
on the charge-state-dependent empirical potentials [27] cannot adequately describe
the interaction of highly-charged ions with TMDs, as evident, e.g., from Fig. 3 in Ref.
[22].
In this project, by combining first-principles calculations with analytical potential
molecular dynamics simulations, we studied the formation of defects in 2D materials
under electron and ion irradiation using atomistic simulations. The influence of
defects on the electronic, optical and catalytic properties of 2D materials has also
been investigated. The highlights from our research are presented below.

2 Results

2.1 Production of defects in two-dimensional materials under electron

irradiation

We investigated [1] the role of electronic excitations in defect production by using

advanced first-principles simulation techniques based on the Ehrenfest dynamics
combined with time-dependent density-functional theory and demonstrate that
a combination of excitations and knock-on damage in 2D MoS2 under electron
beam can give rise to the formation of vacancies and explain the experimental
observations of defects production far below the knock-on threshold. As our first-
principles calculations showed, electronic excitations will quickly delocalize in an
otherwise unperturbed periodic structure. The situation changes dramatically when
the translational symmetry is broken, as is the case for the displacement of one target
atom upon momentum transfer from an impinging electron.
The results of Ehrenfest dynamics (ED simulations of a high-energy electron
impact into MoS2 sheet are presented in Fig. 1. The electron was modeled as a
classical particle with a precisely defined trajectory, which can give rise to electronic
excitations in the target material, as schematically illustrated in Fig. 1(a). The spatial
extent of the electronic excitation created in the system immediately after the impact
is illustrated in Fig. 1(b). Simulations where exactly one electron is excited with the
excitation initially being localized on a sulfur atom are shown in Fig. 1(c) and panel
(d) illustrates the spatial extent of the excitation after 1.6 fs as described within the
framework of ED.
Based on the results of our simulations, we proposed a possible mechanism for
the observed sub-knock-on threshold damage which involves the localization of the
excitation at an emergent vacancy site. According to our calculations, this localized
electronic excitation then gives rise to a decrease in the displacement threshold as
anti-bonding states are occupied at the emergent defect site. Consequently, beam
damage may be expected for voltages below the knock-on threshold as observed
by the Sub Angstrom Low Voltage Electron (SALVE) microscope operated by our
collaborators at Ulm University.
154 S. Kretschmer et al.

Fig. 1: ED simulations of a high-energy electron impact into MoS2 sheet. The electron
is modeled as a classical particle with a precisely defined trajectory, which can give
rise to electronic excitations in the target material, as schematically illustrated in (a).
(b) The spatial extent of the electronic excitation created in the system immediately
after the impact. (c) Simulations where exactly one electron is excited with the
excitation initially being localized on a sulfur atom. (d) The spatial extent of the
excitation after 1.6 fs as described within the framework of ED. Reprinted with
permission from Ref. [1], Copyright (2020) American Chemical Society.

2.2 Response of supported two-dimensional materials to ion and cluster

irradiation

2D materials with nanometer-size holes are promising systems for DNA sequencing,
water purification, and molecule selection/separation. However, controllable creation
of holes with uniform sizes and shapes is still a challenge, especially when the 2D
material consists of several atomic layers as, e.g., MoS2 , the archetypical transition
metal dichalcogenide. We used [2] analytical potential molecular dynamics (MD)
simulations to study the response of 2D MoS2 to cluster irradiation. We modelled
both freestanding and supported sheets and assess the amount of damage created in
MoS2 by the impacts of noble gas clusters in a wide range of cluster energies and
incident angles. We showed that cluster irradiation can be used to produce uniform
holes in 2D MoS2 with the diameter being dependent on cluster size and energy.
Energetic clusters can also be used to displace sulfur atoms preferentially from
either top or bottom layers of S atoms in MoS2 and also clean the surface of MoS2
Annual report on Def-2-Dim project 155

sheets from adsorbents. Our results for MoS2 , which should be relevant to other 2D
transition metal dichalcogenides, suggest new routes toward cluster beam engineering
of devices based on 2D inorganic materials.

Fig. 2: (a) Simulation setup for irradiation of freestanding MoS2 monolayers with
various noble gas clusters. (b) Defect production in free-standing MoS2 monolayers
after impacts of the Xe79 cluster with different initial kinetic energies under normal
incidence. (c) Radius of the pore created by the impact as a function of cluster
energy. (d) Snapshots from MD simulations of a free-standing and supported MoS2
monolayer on a SiO2 substrate.

The effects of ion irradiation on 2D materials have further been studied [8] in
the context of Cl ion implantation onto 2D MoSe2 , one of the prominent members
of the transition metal dichalcogenide materials family. The efficient integration
of transition metal dichalcogenides into the current electronic device technology
requires mastering the techniques of effective tuning of their optoelectronic properties.
Specifically, controllable doping is essential. For conventional bulk semiconductors,
ion implantation is the most developed method offering stable and tunable doping. The
n-type doping in MoSe2 flakes was experimentally realized by our coworkers through
low-energy ion implantation of Cl+ ions followed by millisecond-range flash lamp
annealing. Atomistic simulations at the kinetic Monte Carlo level made it possible
to assess the distribution of impurities in the irradiated samples. Density-functional
theory calculations were carried out to understand the atomic structure of the irradiated
material and assess the effects on the electronic properties. A comparison of the results
of the density functional theory calculations and experimental temperature-dependent
micro-Raman spectroscopy data indicates that Cl atoms are incorporated into the
atomic network of MoSe2 as substitutional donor impurities.
156 S. Kretschmer et al.

Controlled production of defects in hexagonal boron nitride (h-BN) through ion

irradiation has recently been demonstrated to be an effective tool for adding new
functionalities to this material such as single photon generation and for developing
optical quantum applications. Using atomistic simulations, we studied [7, 28] the
response of single-layer and multi-layer h-BN to noble-gas cluster irradiation. Our
results quantify the densities of defects produced by noble gas ions in a wide range
of ion energies and elucidate the types and distribution of defects in the target. The
simulation data can directly be used to guide the experiment aimed at the creation of
defects of particular types in h-BN targets for single-photon emission, spin-selective
optical transitions and other applications by using beams of energetic ions.

2.3 Influence of defects and impurities on the electronic and chemical

properties of two-dimensional materials: insights from
density-functional-theory calculations

Highly-doped TMDs. Doping of materials beyond the dopant solubility limit remains
a challenge, especially when spatially nonuniform doping is required. In 2D materials
with a high surface-to-volume ratio, such as transition metal dichalcogenides, various
post-synthesis approaches to doping have been demonstrated, but full control over
the spatial distribution of dopants remains a challenge. Post-growth doping of single
layers of WSe2 was performed by our coworkers through adding transition metal
(TM) atoms in a two-step process, which includes annealing followed by deposition of
dopants together with Se or S. The Ti, V, Cr, and Fe impurities at W sites are identified
by using transmission electron microscopy and electron energy loss spectroscopy.
The dopants are revealed to be largely confined within nanostripes embedded in
the otherwise pristine WSe2 . Density functional theory calculations [6] showed that
the dislocations assist the incorporation of the dopant during their climb and give
rise to stripes of TM dopant atoms. This work demonstrated a possible spatially
controllable doping strategy to achieve the desired local electronic, magnetic, and
optical properties in 2D materials.
Exotic, novel 2D Material: hematene. Exfoliation of atomically thin layers from non-
van der Waals bulk solids gave rise to the emergence of a new class of 2D materials,
such as hematene (Hm), a structure just a few atoms thick obtained from hematite.
Due to a large number of unsaturated sites, the Hm surface can be passivated under
ambient conditions. Using density functional theory calculations, we investigated [3]
the effects of surface passivation with H and OH groups on Hm properties and
demonstrate that the passivated surfaces are energetically favorable under oxygen-rich
conditions. Although the bare sheet is antiferromagnetic and possesses an indirect
band gap of 0.93 eV, the hydrogenated sheets are half-metallic with a ferromagnetic
ground state, and the fully hydroxylated sheets are antiferromagnetic with a larger
band gap as compared to the bare system. The electronic structure of Hm can be
Annual report on Def-2-Dim project 157

Fig. 3: (a) Atomic model of WSe2 sheet with a dislocation line created by removing
W and Se atoms, as schematically illustrated in the inset, and the resulting strain
distribution (orange shading). The green balls represent Se and the gray balls W
atoms. b) The energies of transition metal (TM) atom in substitutional (W) positions
labeled in (a). It is evident that TM atoms in the substitutional positions prefer to
be in the strained areas next to the dislocation cores. c) The relative energies of
TM adatoms placed at various positions. The inset shows a configuration when the
adatom (red ball) takes the position in the middle of the hollow area, but below Se
atoms. Adatoms prefer to be in the strained area and ultimately take the position
in the heptagon, with the atomic configuration being shown in (a) in the red frame.
Reprinted with permission from Ref. [6], Copyright (2020) Wiley.

further tuned by mechanical deformations. The band gap of fully passivated Hm

increases monotonically with biaxial strain, hinting at the potential applications of
Hm in electromechanical devices.
Role of mirror twin boundaries in water adsorption. Water adsorption on TMDs
and other 2D materials is generally governed by weak van der Waals interactions.
This results in a hydrophobic character of the basal planes, and defects may play a
significant role in water adsorption and water cluster nucleation. However, there is
a lack of detailed experimental investigations on water adsorption on defective 2D
materials. By combining low-temperature scanning tunneling microscopy (STM)
158 S. Kretschmer et al.

experiments and DFT calculations, we studied [5] in that context the well-defined
mirror twin boundary (MTB) networks separating mirror-grains in 2D MoSe2 . These
MTBs are dangling bond-free extended crystal modifications with metallic electronic
states embedded in the 2D semiconducting matrix of MoSe2 . Our DFT calculations
indicate that molecular water also interacts similarly weak with these MTBs as with
the defect-free basal plane of MoSe2 . However, in low temperature STM experiments,
nanoscopic water structures are observed that selectively decorate the MTB network.
This localized adsorption of water is facilitated by the functionalization of the MTBs
by hydroxyls formed by dissociated water. Hydroxyls may form by dissociating water
at undercoordinated defects or adsorbing radicals from the gas phase in the UHV
chamber. Our DFT analysis indicates that the metallic MTBs adsorb these radicals
much stronger than on the basal plane due to charge transfer from the metallic states
into the molecular orbitals of the OH groups. Once the MTBs are functionalized with
hydroxyls, molecular water can attach to them, forming water channels along the
MTBs. This study demonstrated the role metallic defect states play in the adsorption
of water even in the absence of unsaturated bonds that have been so far considered to
be crucial for adsorption of hydroxyls or water.

3 Implementation of the project and technical data

3.1 Codes employed and their scaling

Three different codes were used to carry out the simulations, GPAW [29], VASP [30,31]
and LAMMPS [32]. All these codes are widely used (thousand of users) in atomistic
simulations on various platforms with massive parallel architecture, and a good
scaling behavior has been demonstrated.

3.1.1 GPAW code

For the Ehrenfest dynamics simulations, we employed the GPAW code (released
under GNU Public licence version 3, see https://wiki.fysik.dtu.dk/gpaw/)
[29]. GPAW is implemented in the combination of Python and C programming
languages, where high-level algorithms are implemented in Python, and numerically
intensive kernels in C and in numerical libraries. Python has only a small overhead
for the calculations, and for example in Cray XT5 GPAW has been measured to
execute 4.8 TFLOP/s with 2048 cores, which is 25% of the peak performance. The
library requirements for GPAW are NumPy (a fast array interface to Python), BLAS,
LAPACK and ScaLAPACK (for DFT calculations only). We modified the GPAW code
and the accompanying PAW potentials to enable direct first-principles simulations of
the electron impact in the 2D materials. The scaling behavior of the code is presented
Annual report on Def-2-Dim project 159

(c)

(b) (d)

Fig. 4: DFT simulations of water and hydroxyl adsorption on pristine MoSe2 and
that with mirror twin boundaries (MTBs). (a–c) Atomic structure of MoSe2 with
MTB and adsorbed water molecules, hydroxyl groups and water hexamers attached
to the sheet (a) and OH group (b and c). (d) Formation energy of water clusters on
top of MoSe2 in the pristine area, next to MTB and OH group. (e) The dependence
of adsorption energy of OH group on distance to the MTB defined as the separation
between the MTB and the coordinates of the Se atom the OH group is attached to.
The plot also shows charge transfer from the MTB into the empty molecular orbitals
of the hydroxyl group, as schematically illustrated in panel (f), which presents the
electronic structure of MoSe2 with MTB and isolated OH group.

in Fig. 5. Here all the scaling tests were carried out at HLRS facilities (Cray XT5
HazelHen). Typically, these time-dependent DFT calculations were run on 1024 cores.

3.1.2 VASP code

The VASP code [30, 31], which requires a software licence agreement with the
University of Vienna, Austria, is currently used by more than 1400 research groups
in academia and industry worldwide. At the moment it is the ‘standard accuracy
reference’ in the DFT calculations. We have employed the code for the calculations of
the ground state properties of various defective systems and for Born–Oppenheimer
molecular dynamics simulations of ion impacts into 2D materials. The scaling
behavior of the code is presented in Fig. 6. For these DFT calculations between 128
and 512 cores were used.
160 S. Kretschmer et al.

Fig. 5: Scaling behavior of GPAW when running Ehrenfest dynamics simulations

on CPUs at HLRS. This data was obtained for a system composed of 36 atoms, 5
K-points and 160 bands were used.

Fig. 6: Scaling behavior of the VASP code when running Ehrenfest dynamics
simulations on CPUs at HLRS. This data was obtained for a system composed of 108
atoms, Cut-off energy of 400 eV and 5 K-points.
Annual report on Def-2-Dim project 161

3.1.3 LAMMPS code

Analytical potential molecular dynamics were carried for large systems comprising
a 2D material on a substrate under ion and cluster irradiation using the LAMMPS
package [32]. The code has been shown to perform well in several architectures,
ranging from x86 clusters to Blue Gene P and Cray supercomputers. LAMMPS
has very nice feature called “partition” command which allows it to run replicates
under one mpirun. When LAMMPS is run on P processors and this command is not
used, LAMMPS runs in one partition, i.e. all P processors run a single simulation.
If this command is used, the P processors are split into separate partitions and each
partition runs its own simulation. The arguments to the switch specify the number
of processors in each partition. Arguments of the form 𝑀 × 𝑁 mean 𝑀 partitions,
each with 𝑁 processors. We have used a partition command to invoke each impact
point to a single processor via the partition command-line switch in the script. The
benchmark results are shown in Fig. 7. The figure illustrates the good scalability of
ion irradiation simulations for graphene on SiO2 substrate containing 5101 atoms.
The timings given below are for 3000 molecular dynamics steps of each system and
1200 impact points. The LAMPPS simulations were typically run on 512 to 1024
cores.

Fig. 7: Scaling behavior of the LAMMPS code when running ion irradiation simulation
of graphene deposited on substrate. The speedup is normalized so that at the first
data point (1200 processing elements.
162 S. Kretschmer et al.

3.2 CPU time used

The group has used 28.8 million CPU hours in an extended time frame from March
2020 to April 2021 (total granted CPU time from March 2020 to Feb 2021: 28.2
million CPU hours). The time frame extension was necessary due to delays of some
parts of the project which required the collaboration with experimental groups due to
the Corona pandemic. Besides, the hiring process of new members of the simulation
team took much longer than planned due to delays with getting the visa/work permits.
Detailed statistics of the used CPU time down to the level of each sub-project are
presented in Fig.8. Note that over one-forth of CPU time is invested in on-going
research whose results still need to be published.

0.3 0.5 CPU time [Mio CPU-h] per sub-project

time period: March 2020 - April 2021
total used CPU time: 28.8 Mio CPU-h

6.5 Nanoletters, Kretschmer & Krasheninnikov

8.1
ACS AMI, Ghaderzadeh
JPCC, Wei
Nanoscale, Joseph
Nanomaterials, Ghaderzadeh
Nanoscale, Ghorbani-Asl
AdvMat., Kretschmer & Krasheninnikov
JPCC, Ghorbani-Asl & Kretschmer
1.6
on-going research
6.9 0.9
1.9
2.1

Fig. 8: Detailed CPU time statistics (Mio CPU-h) for the published sub-projects (see
Refs. [1–8]) from March 2020 to April 2021. One-forth of the used CPU time is spent
on on-going projects whose results still need to be published.

3.3 Popular internet articles about our computational research

Supercomputing Helps Study Two-Dimensional Materials

https://www.gauss-centre.eu/news/article/supercomputing-helps-s
tudy-two-dimensional-materials/
Natrium statt Lithium: Die Akkus der Zukunft
https://www.dw.com/de/natrium-statt-lithium-die-akkus-der-zuk
unft/a-54512116
Annual report on Def-2-Dim project 163

Natrium statt Lithium: Die Akkus der Zukunft

https://www.focus.de/perspektiven/nachhaltigkeit/natrium-statt
-lithium-natrium-statt-lithium-die-akkus-der-zukunft_id_123060
25.html
Simulations Help To Tailor The Properties Of 2D Materials
https://prace-ri.eu/simulations-help-to-tailor-the-propertie
s-of-2d-materials/

Conflict of interest

The authors declare that they have no conflict of interest.

References

1. Silvan Kretschmer, Tibor Lehnert, Ute Kaiser, and Arkady V. Krasheninnikov. Formation
of Defects in Two-Dimensional MoS2 in the Transmission Electron Microscope at Electron
Energies below the Knock-on Threshold: The Role of Electronic Excitations. Nano Letters,
20:2865–2870, 2020.
2. Sadegh Ghaderzadeh, Vladimir Ladygin, Mahdi Ghorbani-Asl, Gregor Hlawacek, Marika
Schleberger, and Arkady V Krasheninnikov. Freestanding and Supported MoS 2 Monolayers
under Cluster Irradiation: Insights from Molecular Dynamics Simulations. ACS Applied
Materials Interfaces, 12:37454–37463, 2020.
3. Yidan Wei, Mahdi Ghorbani-Asl, and Arkady V. Krasheninnikov. Tailoring the Electronic
and Magnetic Properties of Hematene by Surface Passivation: Insights from First-Principles
Calculations. The Journal of Physical Chemistry C, 124(41):22784–22792, oct 2020.
4. Janis Köster, Mahdi Ghorbani-Asl, Hannu-pekka Komsa, Tibor Lehnert, Silvan Kretschmer,
Arkady V Krasheninnikov, and Ute Kaiser. Defect Agglomeration and Electron-Beam-Induced
Local-Phase Transformations in Single-Layer MoTe 2. The Journal of Physical Chemistry C,
125:13601–13609, 2021.
5. Jingfeng Li, Thomas Joseph, Mahdi Ghorbani-Asl, Sadhu Kolekar, Arkady V. Krasheninnikov,
and Matthias Batzill. Mirror twin boundaries in MoSe2monolayers as one dimensional
nanotemplates for selective water adsorption. Nanoscale, 13:1038–1047, 2021.
6. Yung-Chang Lin, Jeyakumar Karthikeyan, Yao-Pang Chang, Shisheng Li, Silvan Kretschmer,
Hannu-Pekka Komsa, Po-Wen Chiu, Arkady V. Krasheninnikov, and Kazu Suenaga. Formation
of Highly Doped Nanostripes in 2D Transition Metal Dichalcogenides via a Dislocation Climb
Mechanism. Advanced Materials, 33:2007819, 2021.
7. Sadegh Ghaderzadeh, Silvan Kretschmer, Mahdi Ghorbani-Asl, Gregor Hlawacek, and Arkady V
Krasheninnikov. Atomistic Simulations of Defect Production in Monolayer and Bulk Hexagonal
Boron Nitride under Low- and High-Fluence Ion Irradiation. Nanomaterials, 11:1214, 2021.
8. Slawomir Prucnal, Arsalan Hashemi, Mahdi Ghorbani-Asl, René Hübner, Juanmei Duan, Yidan
Wei, Divanshu Sharma, Dietrich R T Zahn, René Ziegenrücker, Ulrich Kentsch, Arkady V
Krasheninnikov, Manfred Helm, and Shengqiang Zhou. Chlorine doping of MoSe 2 flakes by
ion implantation. Nanoscale, 13:5834–5846, 2021.
9. E. Rutherford. The scattering of 𝛼 and 𝛽 particles by matter and the structure of the atom.
Philos. Mag, 21(125):669–688, 1911.
10. N Bohr. On the constitution of atoms and molecules. Philos. Mag, 26:1–24, 1913.
164 S. Kretschmer et al.

11. Michael Nastasi, James Mayer, and James K Hirvonen. Ion-Solid Interactions: Fundamentals and
Applications. Cambridge Solid State Science Series. Cambridge University Press, Cambridge,
1996.
12. Roger Smith, editor. Atomic and Ion Collisions in Solids and at Surfaces: Theory, Simulation
and Applications. Cambridge University Press, Cambridge, 1997.
13. K S Novoselov, D Jiang, F Schedin, T J Booth, V V Khotkevich, S V Morozov, and A K Geim.
Two-dimensional atomic crystals. Proceedings of the National Academy of Sciences of the
United States of America, 102(30):10451–10453, 2005.
14. J N Coleman, M Lotya, A O’Neill, S D Bergin, P J King, U Khan, K Young, A Gaucher, S De,
R J Smith, I V Shvets, s. K Arora, G Stanton, H.-Y. Kim, K Lee, G T Kim, G S Duesberg,
T Hallam, J J Boland, J J Wang, J F Donegan, J C Grunlan, G Moriarty, A Shmeliov, R J Nicholls,
J M Perkins, E M Grieveson, K Theuwissen, D W McComb, P D Nellist, and V Nicolosi.
Two-Dimensional Nanosheets Produced by Liquid Exfoliation of Layered Materials. Science,
331:568–571, 2011.
15. B Radisavljevic, A Radenovic, J Brivio, V Giacometti, and A Kis. Single-layer MoS2 transistors.
Nature Nanotechnology, 6:147–150, 2011.
16. Manish Chhowalla, Hyeon Suk Shin, Goki Eda, Lain-Jong Li, Kian Ping Loh, and Hua Zhang.
The chemistry of two-dimensional layered transition metal dichalcogenide nanosheets. Nature
Chemistry, 5(4):263–275, 2013.
17. G López-Polín, C Gómez-Navarro, V Parente, F Guinea, MI Katsnelson, F Pérez-Murano, and
J Gómez-Herrero. Increasing the Elastic Modulus of Graphene by Controlled Defect Creation.
Nat. Phys., 11(1):26–31, 2014.
18. U Bangert, W Pierce, D M Kepaptsoglou, Q Ramasse, R Zan, M H Gass, J A den Berg, C B
Boothroyd, J Amani, and H Hofsäss. Ion Implantation of Graphene – Toward IC Compatible
Technologies. Nano Lett., 13:4902–4907, 2013.
19. R. Nair, M. Sepioni, I-Ling Tsai, O. Lehtinen, J Keinonen, Arkady V. Krasheninnikov,
T. Thomson, a. K. Geim, and I. V. Grigorieva. Spin-half paramagnetism in graphene induced
by point defects. Nature Physics, 8(3):199–202, 2012.
20. A. V. Krasheninnikov and F. Banhart. Engineering of nanostructured carbon materials with
electron or ion beams. Nature Materials, 6(10):723–733, 2007.
21. A V Krasheninnikov and K Nordlund. Ion and electron irradiation-induced effects in nanostruc-
tured materials. Journal of Applied Physics, 107(7):071301, 2010.
22. Roland Kozubek, Mukesh Tripathi, Mahdi Ghorbani-Asl, Silvan Kretschmer, Lukas Madauß,
Erik Pollmann, Maria O’Brien, Niall McEvoy, Ursula Ludacka, Toma Susi, Georg S. Duesberg,
Richard A. Wilhelm, Arkady V. Krasheninnikov, Jani Kotakoski, and Marika Schleberger.
Perforating Freestanding Molybdenum Disulfide Monolayers with Highly Charged Ions. Journal
of Physical Chemistry Letters, 10(5):904–910, 2019.
23. Richard A. Wilhelm, Elisabeth Gruber, Janine Schwestka, Roland Kozubek, Teresa I. Madeira,
José P. Marques, Jacek Kobus, Arkady V. Krasheninnikov, Marika Schleberger, and Friedrich
Aumayr. Interatomic coulombic decay: The mechanism for rapid deexcitation of hollow atoms.
Physical Review Letters, 119:103401, 2017.
24. A. V. Krasheninnikov, Y. Miyamoto, and D. Tománek. Role of electronic excitations in ion
collisions with carbon nanostructures. Phys. Rev. Lett., 99:016104, 2007.
25. M. Caro, A. A. Correa, E. Artacho, and A. Caro. Stopping power beyond the adiabatic
approximation. Scientific Reports, 7(1):2618, 2017.
26. M. Ahsan Zeb, J. Kohanoff, D. Sánchez-Portal, A. Arnau, J. I. Juaristi, and Emilio Artacho.
Electronic stopping power in gold: The role of d electrons and the H/He anomaly. Physical
Review Letters, 108(22):225504, 2012.
27. Richard A Wilhelm and Wolfhard Möller. Charge-state-dependent energy loss of slow ions. II.
Statistical atom model. Physical Review A, 93:052709, 2016.
28. M Fischer, J M Caridad, A Sajid, S Ghaderzadeh, M. Ghorbani-Asl, L Gammelgaard, P Bøggild,
K S Thygesen, A V Krasheninnikov, S Xiao, M Wubs, and N Stenger. Controlled generation of
luminescent centers in hexagonal boron nitride by irradiation engineering. Science Advances,
7:eabe7138, 2021.
Annual report on Def-2-Dim project 165

29. J Enkovaara, C Rostgaard, J J Mortensen, J Chen, M Dułak, L Ferrighi, J Gavnholt, C Glinsvad,

V Haikola, H A Hansen, H H Kristoffersen, M Kuisma, A H Larsen, L Lehtovaara, M Ljungberg,
O Lopez-Acevedo, P G Moses, J Ojanen, T Olsen, V Petzold, N A Romero, J Stausholm-Møller,
M Strange, G A Tritsaris, M Vanin, M Walter, B Hammer, H Häkkinen, G K H Madsen, R M
Nieminen, J K Nørskov, M Puska, T T Rantala, J Schiøtz, K S Thygesen, and K W Jacobsen.
Electronic structure calculations with GPAW: a real-space implementation of the projector
augmented-wave method. J. Phys. Condens. Matter, 22:253202, 2010.
30. G. Kresse and J. Hafner. Ab initio molecular dynamics for liquid metals. Physical Review B,
47:558–561, 1993.
31. G. Kresse and J. Furthmüller. Efficiency of ab-initio total energy calculations for metals and
semiconductors using a plane-wave basis set. Computational Materials Science, 6:15–50, 1996.
32. Steve Plimpton. Fast parallel algorithms for short-range molecular dynamics. J. Comput. Phys.,
117:1–19, 1995.
High-performance multiphase-field simulations of
solid-state phase transformations using Pace3D

E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

Abstract Computational materials science contributes to the accelerated development

of new or optimized materials. The phase-field method has established itself as
a powerful tool to describe the temporal microstructure evolution during solid-
phase transformations. The use of high performance computers allows studying
the evolution of large, three-dimensional microstructures incorporating phase- and
grain boundary specific behaviors as well as phase transitions. This allows a more
realistic representation of the phenomena investigated and lead to more reliable
predictions of the microstructural evolution. In this work, current applications of the
phase-field method are presented using the Pace3D software package and applied
at the ForHLR II supercomputer. Additionally, the scaling behavior is shown when
using up to 5041 cores.

1 Introduction

In materials science, understanding the conditions under which certain microstructures

are formed during manufacturing is crucial to improving the properties of materials
or developing new classes of materials. Complementary to experiments, simulations
can help to investigate specific process conditions on the resulting microstructures
and thus on the behavior of the material in the application. The phase-field method
is an elegant method to describe interface motions numerically, since no explicit
interface tracking is necessary [1]. With increasing understanding of this method, it
is used for material discovery [2]. Particularly through the possibilities of modern

E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

Institute of Applied Materials (IAM), Karlsruhe Institute of Technology (KIT), Straße am Forum 7,
76131 Karlsruhe, Germany
B. Nestler
Institute of Digital Materials, Hochschule Karlsruhe Technik und Wirtschaft, Moltkestr. 30, 76131
Karlsruhe, Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 167
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_10
168 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

high-performance computing, realistic processes in the microstructure formation

during the production of materials can be investigated [3] and the understanding
of critical process parameters can be expanded. The aim of this paper is to present
selected results in the simulation of solid-phase transformations made possible by the
use of high-performance computers. Three applications are discussed, namely the
martensitic transformation, shape-instabilities of finite three-dimensional cementite
rods embedded in a ferrite matrix, as well as coarsening and grain growth in solid
oxide fuel cells anodes.
This work is structured as follows: In section 2 the basic model of the used multiphase-
field framework in presented. The Pace3D framework is described including its
performance and scaling behavior in section 3. In section 3 the results of applying
the multiphase-field method to simulate solid-state phase transformations and the
used cores are discussed.

2 Core phase-field model

The free energy of a system with volume 𝑉 is generally composed of energy

contributions due to interfaces separating physically distinguishable regions and
energy contributions present within a region. The latter are referred to as volumetric
energy density contributions. In order to parametrize a system, each physically
distinguishable region is assigned an order parameter 𝜙 𝛼 (x, 𝑡), which is called phase-
field parameter and which is part of the 𝑁-tuple 𝜙(x, 𝑡), where 𝑁 is the number of all
order parameters, 𝑡 is the time and x is the location vector. In the multiphase-field
context, the free energy can be expressed by
∫
F (𝝓, ∇𝝓, . . . ) = 𝑊intf (𝝓, ∇𝝓) + 𝑊¯ bulk (𝝓, . . . )d𝑉 . (1)
𝑉

Here, the term 𝑊intf (𝝓, ∇𝝓) contains the energy contributions of all interfaces and
the term 𝑊¯ bulk (𝜙, . . . ), which can depend on arbitrary quantities like the chemical
concentration, contains the volumetric energy contributions. The phase-dependent
interpolated volumetric energy density is given by
∑︁
𝑊¯ bulk (𝜙, . . . ) = 𝛼
ℎ 𝛼 (𝝓)𝑊bulk (. . . ), (2)
𝛼

whereas ℎ 𝛼 (𝝓) is a normalized interpolation function, which reduces in the simplest

case to ℎ 𝛼 (𝝓) = 𝜙 𝛼 [4]. The effort of a system to minimize the free energy is
modeled with a phase-field evolution equation. Two approaches to the formulate the
Allen–Cahn evolution equation in the multiphase-field context can be distinguished.
Í
Nestler et al. [5] use a Lagrangian parameter 𝜆 to ensure the sum constraint 𝛼 𝜙 𝛼 = 1:
Multiphase-field simulations of solid-state phase transformations 169

𝜕𝜙 𝛼 𝛿F
𝜏𝜀 =− −𝜆 (3)
𝜕𝑡 𝛿𝜙 𝛼
𝛿F 1 ∑︁ 𝛿F
=− + , ∀𝜙 𝛼 , 𝛼 = 0, . . . , 𝑁. (4)
𝛿𝜙 𝛼 𝑁 e 𝛽 𝛿𝜙 𝛽
(5)

Here, 𝑁
e is the number of locally active phases,
 
𝛿F (𝝓, ∇𝝓, . . . ) 𝜕 𝜕
= −∇· 𝑊 (𝝓, ∇𝝓, . . . ) (6)
𝛿𝜙 𝛼 𝜕𝜙 𝛼 𝜕∇𝜙 𝛼

is the variational derivative of the energy functional F with respect
to the order
parameter 𝜙 𝛼 and the divergence operator is represented by ∇ · (. . . ) . The kinetics
of the phase transformations is determined by the relaxation parameter 𝜏 and the
width of the diffuse interface is set using the parameter 𝜀. The relaxation parameter
is weighted locally using the relaxation parameter between two phases 𝜏𝛼𝛽 and their
volume fractions
Í
𝛼<𝛽 𝜏𝛼𝛽 𝜙 𝛼 𝜙 𝛽
𝜏(𝜙) = Í . (7)
𝛼<𝛽 𝜙 𝛼 𝜙 𝛽

Another approach is based on Steinbach [6] and reads

e "
𝑁
𝜕𝜙 𝛼 1 ∑︁ 𝛼𝛽
= 𝑀 𝛼𝛽 Δ𝑊intf
𝜕𝑡 𝑁𝜀 𝛽≠𝛼
e
√︁ !# (8)
8 𝜙 𝛼 𝜙𝛽 𝛼𝛽
+ Δ𝑊bulk , ∀𝜙 𝛼 , 𝛼 = 0, . . . , 𝑁,
𝜋

𝛼𝛽
with Δ𝑊intf = 𝛿Fintf /𝛿𝜙 𝛽 − 𝛿Fintf /𝛿𝜙 𝛼 as the contribution the due to curvature
𝛼𝛽
minimizing and Δ𝑊bulk = 𝛿Fbulk /𝛿𝜙 𝛽 − 𝛿Fbulk /𝛿𝜙 𝛼 as the driving force due a
difference in the volumetric energy densities. Here, the mobility between two
phases 𝑀 𝛼𝛽 is directly used in the evolution equation. According to Nestler et al. [5],
the interfacial energy density
1
𝑊intf (𝝓, ∇𝝓) = 𝜀𝑎(𝜙, ∇𝜙) + 𝜔(𝜙) (9)
𝜀
consists of the gradient energy density 𝜀𝑎(𝜙, ∇𝜙) and the potential energy den-
sity 𝜔(𝜙)/𝜀 of obstacle type [5, 7]. The interaction of both contributions leads to
the formation of a diffuse interface as well as to the representation of the interfacial
energy over a volumetric range. This core multiphase-field model can be coupled
with additional volumetric energy contribution for variable application scenarios
which is detailed in Section 3.
170 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

3 The Pace3D framework: Performance and scaling

The models used in this work are implemented in the multiphysics massive parallel
Pace3D framework (“Parallel Algorithms for Crystal Evolution in 3D”)[8]. To calcu-
late the evolution of the phase-fields and the chemical potentials, the corresponding
equations are discretized on a uniform grid with finite differences and an explicit Euler
method is employed for the time integration. The mechanical equilibrium condition
is solved implicitly to update the stresses and strains using a finite element scheme.
Parallelization is integrated employing the Message Passing Interface (MPI) and
assigning spatial subdomains to each MPI process (spatial domain decomposition).
Investigations concerning the performance and scaling on the ForHLR II were
conducted. Specifically, weak scaling was performed for a baseline solver and an
optimized solver employing a thin abstraction layer over single instruction, multiple
data (SIMD) intrinsics as well as various other optimizations detailed in [8, 9], e.g.
a buffer ensuring that expensive gradients are only calculated once or using the
update pattern of the phase-field equation for better vectorization. The domain size
as well as the initial and boundary conditions were chosen to be representative of
typical simulation domains employed in this work. The results up to 5041 cores are
shown in Fig. 3. The plot shows how many lattice updates per second per core were

Fig. 1: Weak scaling on the ForHLR II for an optimized solver employing SIMD
intrinsics and a baseline solver. While the baseline solver scales almost ideally, its
performance makes it unattractive. Compared to this the optimized solver shows
reasonable percentages of peak performance with a somewhat worse scalability.
Multiphase-field simulations of solid-state phase transformations 171

calculated on a whole-application basis, with higher values corresponding to smaller

time-to-solution. The optimized version likely became memory-bound within the
confines of a single node, as the baseline solver showed almost no drop in MLUP/s
whereas the optimized version shows a drop. Scaling beyond one node (starting from
64 cores) for the baseline solver shows almost no drop since the time for a single
time step is significantly larger than the time taken for the necessary communication
between cores. However, as the optimized solver is about 30 times faster than the
baseline solver, it shows a significant efficiency drop as more and more core are
added. Strictly speaking there should only be a significant drop between in-node and
out-of-node scaling as the network latency is triggered between workers. We suspect
the additional drop is due to the network architecture of the ForHLR II since the same
scaling studies on the Hazel Hen as well as SuperMUC did not show a significant
out-of-node drop in efficiency.
The performance was investigated by counting the floating-point operations
executed and correlating these with the runtimes, yielding the floating point operations
per second (FLOP/s on a whole-application basis. Comparing the obtained FLOP/s
to the values for the Intel Xeon E5-2660 shows that 19 % peak performance was
reached for a single core and 14 % for 16 cores.

4 Modeling and simulating solid-state phase transformations

using Pace3D

4.1 Martensitic transformation

To precisely control the thermal and mechanical properties of the material, steel
is subjected to a defined heat treatment. Depending on the cooling temperature,
cooling rate and chemical composition, different microstructures are obtained starting
in the (partly) austenitic region. At a rapid cooling rate under the martensite start
temperature, martensite is formed from austenite by a diffusionless transformation
mechanism which leads to a strengthening of steel.
To model the martensitic transformation in a polycrystalline microstructure,
interfacial, chemical and elastic contributions are considered in free energy of the
system resulting in
∫
F (𝜙, ∇𝜙, 𝜖)
¯ = ¯ + 𝑊¯ chem (𝜙, c, 𝑇)d𝑉 .
𝑊intf (𝜙, ∇𝜙) + 𝑊¯ elast (𝜙, 𝜖) (10)
𝑉

Here, 𝑊¯ elast is the elastic free energy which is given as a function of the phase-field 𝝓
and the elastic strain 𝜖. ¯ The chemical free energy 𝑊¯ chem in the context of a constant
concentration and temperature depends only on 𝝓. Since the transformation rate of
martensite is high, the concentration field is considered to the constant over time. The
temporal evolution of the stresses and strains is implicitly calculated based of the
assumption of mechanical equilibrium in each discrete time step. The multiphase-field
172 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

evolution equation. 8 is applied to solve the temporal evolution of the martensitic

phases, whereas an additional term is included to ensure a stable interface in the
context of high driving forces [10]. For a more detailed description of the underlying
models, the reader is referred to [10, 11].
In Fig. 2, results of simulating the martensitic transformation using Pace3D
are displayed. Different application scenarios have been investigated, ranging from
three-dimensional elastic martensitic transformation in a single-crystal to a two-
dimensional elasto-plastic martensitic transformation in a dual-phase microstructure
using a phase-dependent plasticity formulation. Starting from a martensitic-free
microstructure or a setting with an initial martensitic nucleus within an austenitic
grain, martensitic variants nucleate at structural defects like grain boundaries and
spread into the austenitic regions. In this process, different variants are formed, which
minimize the energy of the system by alternating occurrence of martensitic variants,
so that the typical microstructures for lath martensite are formed.
The necessary computational costs depend, among other things, on the number
of discretization points, the physical models used, and how far the system is from
equilibrium. An overview of typically used cores in different application scenarios
can be found in Tab. 4.1. The jump from an elastic to an elasto-plastic simulation
reflecting the difference between case IV and case V requires an increase in the
necessary resources, since additionally plastic strains have to be calculated. In case
IV, the model according to [4] is applied which uses the same plastic fields for all
phases. In case V, the models according to [15, 16] are applied which are capable
of calculating phase-dependent plastic strains using different nonlinear mechanical
constitutive equations for each phase separately. Consequently, the calculation of
phase-dependent plastic strains require higher computational costs compared to
case IV.

4.2 Shape-instabilities of finite three-dimensional rods

The macroscopic behavior of any metallic material is highly influenced by its

microstructural patterns. Rod-shaped patterns, which can be often observed in
mesoscopic regions wherein multiple phases are systematically aligned, therefore are
of interest in material investigations for many years. To this end, the morphological
behavior of three-dimensional finite cementite rods is considered in this work.
To model the diffusion- and curvature-driven shape-instabilities of finite three-
dimensional rods, interfacial and chemical contributions are considered in the energy
functional of the system resulting in
∫
F (𝝓, ∇𝝓, 𝒄, 𝑇) = 𝑊intf (𝝓, ∇𝝓) + 𝑊¯ chem (𝜙, c, 𝑇)d𝑉, (11)
𝑉
Multiphase-field simulations of solid-state phase transformations 173

(a) (b)

(c) (d)

Fig. 2: Resulting microstructures of the martensitic phase transformation us-

ing Pace3D. a) Three-dimensional elastic martensitic transformation in a single-
crystal [12]. b) Elastic martensitic transformation in a microstructure consisting
of austenitic grains and lamellar graphite [13]. Lamellar graphite is shown in
black. c) Elastic martensitic transformation in a polycrystalline microstructure [14].
d) Elasto-plastic martensitic transformation in a dual-phase microstructure without
phase-dependent plasticity [10]. Ferritic grains are shown in gray scale. In a)-d),
Martensitic variants are displayed with colors. In b) and c), austenitic grains are
shown in gray scale.
174 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

Table 1: Typical used computational resources to simulation the martensitic phase

transformation in different application scenarios.

Case Resolution in Used cores Duration in Results pub-

voxel cells hours lished in

I: Three-dimensional elastic 150 × 150 × 150 16 18 [12]

martensitic transformation in
a single-crystal
II: Elastic martensitic trans- 900 × 900 55 48 [13]
formation in a microstructure
consisting of austenitic grains
and lamellar graphite
III: Elastic martensitic trans- 800 × 800 196 10 [14]
formation in a polycrystalline
microstructure
IV: Elastic martensitic trans- 4000 × 4000 381 15 [11]
formation in a dual-phase mi-
crostructure
V: Elasto-plastic martensitic 4000 × 4000 784 38 [10]
transformation in a dual-phase
microstructure without phase-
dependent plasticity
VI: Elasto-plastic martensitic 4000 × 4000 1191 72 [13]
transformation in a dual-phase
microstructure with phase-
dependent plasticity

where 𝑊¯ chem is dependent of concentration, temperature and the phase-field variable.

All the required thermodynamic data leading to diffusion-driven phase evolution are
included in that energetic contribution. For an in-depth understanding of the used
model, readers are directed to Refs. [5, 17, 18].
Single finite rods have already been analyzed in [19, 20]. To capture the effect
of the presence of neighboring rods on the carbon redistribution mechanisms those
studies were extended in [21]. Besides an in-depth analysis of the underlying carbon
diffusion paths, a critical aspect ratio has been found above which the cementite rods
in the lamellar arrangement spheroidize into 2 particles instead of only 1 particle for
lower aspect ratios, which is called ovulation. For single isolated rods, the critical
aspect ratio is found to be 11, whereas for multiple lamellar arranged rods the critical
value alters to 12. Since those previous investigations have been conducted with a
fixed interlamellar spacing 𝜆, a variation of that quantity, apart from varying the rods’
aspect ratios, is the main object of interest for the present work.
Multiphase-field simulations of solid-state phase transformations 175

order parameter (φ)

z

cementite (θ)

ferrite (α)
x
(a)
y
λ

0
internal
rods
terminal rods

early stages

(b)

AR 11 AR 12 AR 13

(c)

Fig. 3: Temporal morphological evolution. a) Starting simulation setup showing

precipitate cementite rods embedded into ferritic matrix in a lamellar arrangement.
A cut through the 𝑥-𝑧-plane in the middle of the domain is performed to show the
longitudinal cross-section of the rods. b) Mass transfer in early stages due to chemical
potential fields. c) Final stage morphologies for aspect ratios 11, 12 and 13.

The variation of the interlamellar spacing 𝜆 reveals another interesting change

in the microstructural evolution, which is shown in Fig. 3c). Hence, apart from the
critical aspect ratio obtained in [21], a critical value for the interlamellar spacing
is found as well. In this context critical means that there is a particular distance in
between the rods in which the shape-instability events of the internal rods highly
differ from the ones of the terminal rods. Whilst the internal rods undergo ovulation
events, i.e. a break-up of rods, the terminal rods solely experience spheroidization
176 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

events. This fundamental difference in shape alteration leads to the different amounts
of remaining particles for the aspect ratio 12, which represents an intermediate stage
between the findings made in [20,21]. Eventually, a small increase of the interlamellar
spacing results into a shift of the critical aspect ratio from 12 to 13 for lamellar
arranged cementite rods. To conclude, this work figures out an intermediate stage of
transformation mechanisms apart from revealing a slightly shifted critical value of
the aspect ratio for a variation of the interlamellar spacing.

Table 2: Parameter set and core usage for the simulations of shape-instabilities.

Name Symbol Value Unit

Number of cores per simulation - 24-36 -

Duration in Hours per simulation - 70-75 h
Amount of simulations - 12 -
Resolution in voxel cells 𝑁𝑥 × 𝑁 𝑦 (130-160) × 45 × (180-200) -
× 𝑁𝑧
Discretization spacing Δ𝑥 = Δ𝑦 = Δ𝑧 1.0 × 10−9 m
Interlamellar spacing 𝜆 (30-40) × Δ𝑥 m
Interface width parameter 𝜖 3.0 × Δ𝑥 m
Temperature 𝑇 973 K
Interfacial Energy 𝛾𝑠 0.49 Jm−2 [22]
Diffusivity 𝐷 2 × 10−9 m2 s−1 [22]
Molar volume Ω̃ 7 × 10−6 m3 /mole
eq
Equilibrium concentration 𝑐𝜃 0.25 mole fraction
eq
Equilibrium concentration 𝑐𝛼 0.00067 mole fraction

4.3 Coarsening and grain growth in SOFC anodes

Solid oxide fuel cells (SOFCs) are a prominent technology to establish the efficient
and environmentally friendly utilization of chemically stored energy. A high durability
of fuel cells is key to make this technology also economically feasible by reducing
maintenance costs and guaranteeing a long lifetime. One of the most important
factors which limits the lifetime of SOFCs is the degradation of the anode. Here,
the coarsening of nickel at the relatively high operating temperatures is one of
the processes responsible for the degradation [23, Section 2.3]. To gain a deeper
Multiphase-field simulations of solid-state phase transformations 177

understanding of the underlying mechanisms, which govern the coarsening of nickel,

we employ multiphase-field simulations to study such process on the micro-scale.
An initial polycrystalline microstructure has been artificially generated by a Voronoi
diagram, thereby assigning particles statistically uniform to one of the three phases
(Nickel, Yttria-stabilized Zirconia (YSZ) and Pores) which results in volume fractions
of 34 %, 34 % and 32 % for each of these phases, respectively. Furthermore, the
nickel phase is additionally decomposed by a second Voronoi construction using 480
separate order parameters, which results initially in more than 500 nickel grains. The
decomposition into grains is done to incorporate the possibility of grain-boundary
diffusion and secondly to allow for concomitant thermal grooving, which, albeit
subtly, leads to a perturbation of the nickel surface [24, 25]. The corresponding initial
three-dimensional microstructure is shown in Fig. 4a, which is subject to periodic
boundary conditions at either side of the domain.
Please note the clear distinction in the terminology here between particles and
grains of nickel. Whenever the term particle is used, we refer to the geometry of the
nickel phase as if no information about orientations and GBs was available. Contrarily,
a grain is a connected region where the orientation does not change, here described
as a connected bulk region of a single order parameter. Furthermore, by coarsening
we mean a relative increase in mean particle size of nickel (again without knowledge
about orientations), whereas grain growth describes the average increase in grain size
due to the annihilation of grains.
We utilize a recently developed extension [25] of the grand-potential model [17,26]
which allows to quantitatively render interface diffusion phenomena.
Therefore, the grand-potential functional
∫
F (𝝓, ∇𝝓, 𝝁, 𝑇) = 𝑊intf (𝝓, ∇𝝓) + 𝑊¯ chem (𝜙, c(𝜇), 𝑇)d𝑉, (12)
𝑉

is utilized. It includes an additional dependence on the chemical potentials 𝝁 =

{𝜇Ni , 𝜇YSZ }. Note that an isothermal domain is assumed and thus the dependence
on temperature 𝑇 is dropped henceforth. The relaxation parameters are interpolated
by Eq. 7. We utilize a set of order parameters 𝝓 = {𝜙Ni1 , . . . , 𝜙Ni480 , 𝜙YSZ , 𝜙Pore } to
distinguish between a number of nickel grains as well as YSZ and Pores, which are
rendered by only a single order parameter each. The chemical contribution to the
functional can be written as
" #
∑︁ ∑︁
𝛼 𝛼 𝛼
𝑊chem (𝜙, c(𝜇)) =
¯ ℎ 𝛼 (𝝓) 𝑓 (c ( 𝝁)) − 𝜇𝑖 𝑐 𝑖 (𝜇𝑖 )
𝛼 𝑖 (13)
𝛼 ∈ {Ni1 , . . . , Ni480 , YSZ, Pore} , 𝑖 ∈ {Ni, YSZ} .

Simple parabolic forms of the bulk free energy densities 𝑓 𝛼 are assumed, which read

𝑓 𝛼 (c 𝛼 ( 𝝁)) = 𝐴(𝑐 Ni
𝛼 𝛼
(𝜇Ni ) − 𝑐 Ni,eq ) 2 + 𝐴(𝑐 YSZ
𝛼 𝛼
(𝜇YSZ ) − 𝑐 YSZ,eq )2 . (14)
178 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

The flux density which governs the coarsening of nickel is expressed as

1 ∑︁ ¯ 𝛼𝛽
jNi = − 𝑀 𝜙 𝛼 𝜙 𝛽 ∇𝜇Ni (15)
𝜖 𝛼<𝛽 Ni

𝛼𝛽
such that only interface diffusion is considered. The mobilities 𝑀¯ Ni quantify the
kinetics of the nickel coarsening. For nickel, only grain-boundary and surface diffusion
¯ Ni𝑘 -YSZ = 0 for all 𝑘 ∈ {1, . . . , 480}. Diffusion of YSZ
YSZ-Pore = 𝑀
is assumed, i.e. 𝑀¯ Ni Ni
is neglected at operating temperature (jYSZ = 0), in accordance with experimental
findings [27].
The corresponding model parameters are listed in Table 4.3, where the number 𝑘
in the subscript is omitted from now on since nickel grains share identical properties.
We assume equal isotropic interfacial energies between the three phases. The nickel
grain-boundaries are assigned a relatively smaller value, the ratio 𝛾Ni−Ni /𝛾Ni-Pore is
close to the ratio for high-angle GBs determined in a recent and extensive experimental
study [24]. The grain-boundary mobility 1/𝜏Ni-Ni is responsible for the rate at which
grain growth occurs. Other boundary relaxation parameters are chosen according
to [25, p. 12, below Eq. (99)], such that the length scale of attachment kinetics is
on the order of the interface width and thus interface diffusion becomes dominant
on the length √︃scale of the problem. For the Ni-surface we obtain the characteristic
length 𝑙 𝑐 = 𝑀¯ Ni-Pore 𝜋 2 /(32 · 0.82 ) ≈ 0.44𝑢 𝑙 . The equilibrium compositions 𝑐 𝛼
Ni 𝑖,eq
with 𝑖 ∈ {Ni, YSZ} and the thermodynamic prefactor 𝐴 are so chosen to guarantee
sufficient conservation of volume. The parameters are represented in a unit system of
the model, where 𝑢 𝑙 , 𝑢 𝑡 and 𝑢 𝐸 are the units of length, time and energy, respectively.
The corresponding values in SI-units can be affixed by assuming a surface diffusivity,
surface energy, temperature and initial particle size for the system.

Ni YSZ Pore Ni YSZ Pore

(a) (b)

Fig. 4: Initial (a) and final (b) three-dimensional microstructures of the SOFC anode.
Multiphase-field simulations of solid-state phase transformations 179

Table 3: Parameter set and core usage for the simulation of the SOFC anode.

Name Symbol Value Unit

Number of cores - 216 -

Duration in Hours - 144 h
Resolution in voxel cells 𝑁 𝑥 × 𝑁 𝑦 × 𝑁𝑧 192 × 192 × 192 -
Discretization spacing Δ𝑥 = Δ𝑦 = Δ𝑧 1.0 𝑢𝑙
Interface width parameter 𝜖 4.0 𝑢𝑙
Relaxation parameter Ni-YSZ 𝜏Ni-YSZ 1.0 𝑢𝐸 𝑢𝑡 /𝑢𝑙4
Relaxation parameter Ni-Pore 𝜏Ni-Pore 1.0 𝑢𝐸 𝑢𝑡 /𝑢𝑙4
Relaxation parameter YSZ-Pore 𝜏YSZ-Pore 1.0 𝑢𝐸 𝑢𝑡 /𝑢𝑙4
Relaxation parameter Ni-GBs 𝜏Ni-Ni 100.0 𝑢𝐸 𝑢𝑡 /𝑢𝑙4
Surface energy Ni 𝛾Ni-Pore 1.0 𝑢𝐸 /𝑢𝑙2
Interfacial energy Ni-YSZ 𝛾Ni-YSZ 1.0 𝑢𝐸 /𝑢𝑙2
Surface Energy YSZ 𝛾YSZ-Pore 1.0 𝑢𝐸 /𝑢𝑙2
Grain boundary energy Ni 𝛾Ni-Ni 0.43 𝑢𝐸 /𝑢𝑙2
Equilibrium composition Ni-Ni Ni
𝑐Ni,eq 0.9 -
Equilibrium composition Pore-Ni Pore
𝑐Ni,eq 0.1 -
Equilibrium composition YSZ-Ni YSZ
𝑐Ni,eq 0.1 -
Equilibrium composition Ni-YSZ Ni
𝑐YSZ,eq 0.1 -
Equilibrium composition Pore-YSZ Pore
𝑐YSZ,eq 0.1 -
Equilibrium composition YSZ-YSZ YSZ
𝑐YSZ,eq 0.9 -
Surface diffusion mobility ¯ Ni-Pore
𝑀 Ni 0.4 𝑢𝑙6 /(𝑢𝐸 𝑢𝑡 )
GB diffusion mobility ¯ Ni-Ni
𝑀 Ni 0.4 𝑢𝑙6 /(𝑢𝐸 𝑢𝑡 )
Thermodynamic prefactor 𝐴 5.0 𝑢𝐸 /𝑢𝑙3

During the simulation the microstructure undergoes coarsening of Nickel, which

leads to larger and smoother Nickel particles (see Fig. 4b), while the YSZ structure
remains invariant. The coarsening of nickel is quantified by means of continuous
particle size distributions [28] which were successfully applied in experimental
studies on Ni-coarsening in SOFCs [29].
From the individual particle size distributions at different times we calculate the
mean particle diameter 𝑑50 at each time 𝑡, which corresponds to the intersection of
the distribution with a volume fraction of 50 %. The evolution of the mean particle
diameter of nickel is shown in Fig. 7. It is observed, that the particle size increases
180 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

Number of Ni-Grains NNi

Ni-Diameter d50 (ul )

24 500

400
23
300

22 200

0.0 0.5 1.0 1.5

Time t (105 ut ) 0.0 0.5 1.0 1.5
Time t (105 ut )

(b)
(a)

Fig. 5: Evolution of the mean particle diameter of nickel 𝑑50 , obtained by neglecting the
grain structure and calculated by means of continuous particle size distributions [28]
(a) and the number of grains 𝑁Ni (b) in the SOFC anode.

monotonically with time, while the coarsening rate is largest at early time. The overall
increase in particle diameter is about 13 % over the whole simulation time. Inside the
nickel particles, the individual grains of the polycrystal are concomitantly growing
which causes the number of grains to decay with time (Fig. 5b). From the initial
number of 564 grains, only 137 grains persist at the end of the simulation.

Ni YSZ Pore Ni YSZ Pore

(a) (b)

Fig. 6: Tortuous electron pathways inside the nickel structure under an applied voltage
difference from top to bottom, where the color of the lines indicates the magnitude of
the electric flux density (a) and triple-phase boundaries with each individual segment
colored separately (b) in the final state.
Multiphase-field simulations of solid-state phase transformations 181

4.0 1.6

TPB-Length lTPB (104 ul )

Nickel-Tortuosity TNi

3.5 1.4

3.0 1.2

2.5 1.0

2.0 0.8
0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5
Time t (105 ut ) Time t (105 ut )

(a) (b)

Fig. 7: Evolution of the nickel tortuosity 𝑇Ni (a) and triple-phase boundary length 𝑙 TPB
(b) in the SOFC anode.

To gain information about the effective conductivity of the nickel phase, which is
relevant for the electron transport, the tortuosity of nickel 𝑇Ni has been calculated
by applying a voltage difference (Fig. 6a) with respect to two sides of the domain,
as in [30]. The tortuosity of the initial microstructure is about 𝑇Ni (𝑡 = 0) = 3.5 and
does not change significantly with time (see Fig. 7a), apart from an initial drop,
down to a value of about 𝑇Ni ≈ 3.25, which corresponds to a reduction of not more
than 8 %. Consequently, the effective conductivity of the nickel network is about 31 %
compared to the ideal case of straight transport pathways (corresponding to 𝑇Ni = 1)
which is related to the tortuous nature of the streamlines and in addition to the various
bottlenecks where the flux density becomes locally larger in magnitude (see Fig. 6a).
On the contrary, the length of the triple-phase boundary (visualized in Fig. 6b) 𝑙 TPB
decreases in a pronounced fashion with time (Fig. 7b). Here, the decrease in 𝑙 TPB
is fastest at early times. Therefore, both the evolution of the particle diameter of
Ni as well as the dynamics of TPB length are in agreement with experimental
measurements in Ni-YSZ [31] and Ni-CGO anodes [29], at least qualitatively. The
overall loss in TPB length, comparing final and initial states of the simulation run,
constitutes about 20 %. Since the TPB is the region in the anode at which all of the
transport mechanisms (electron, ion and gas transport) are available, a reduced 𝑙 TPB
is unfavorable as it suppresses the oxidization reaction of the fuel gas. Thus, the
reduction in 𝑙TPB constitutes one of the factors which affect the degradation of the
anode material. Please note however, that the assumption of equal interfacial energies
(Table 4.3) may not be very accurate. Recent findings [32] show that the interfacial
energy between Nickel and YSZ is usually higher than that of the surface energy of
YSZ [33]. Together with the surface energies of nickel [34, 35] it is expected that the
nickel particles should show dewetting from the YSZ substrate, which is obviously
not rendered here. The wetting condition is likely to affect the performance of the
anode [36, 37]. Consequently it is probable, that the degradation is underestimated in
the current work.
182 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

5 Conclusion

In this work, we presented simulation result of applying the multiphase-field model

to simulate different solid-phase transformation processes using Pace3D and high-
performance facilities. The scaling and performance of a base solver and an optimized
version were investigated, showing a factor of 30 speedup after optimization on a
single core. However, this performance increase came at a decline in scalability, at
least on the network of the ForHLR II, but even at the worst investigated point a 20-fold
improvement remained. In the case of modelling martensitic phase transformation,
the resulting microstructure including the corresponding stress and strain state can be
analyzed whereas the number of used cores increases when more complex models are
applied. With regards to the modelling of shape-instabilities, the computational effort
in terms of necessary amount of cores also increases slightly with growing aspect
ratios and interlamellar spacings. Apart from the computational effort, the simulations
of the shape-instabilities for finite cementite rods reveal an intermediate stage for a
particular interlamellar spacing in between pure spheroidization or pure ovulation
mechanisms in which both modes occur simultaneously. Besides, a shift of the critical
value of the aspect ratio is observed when the aforementioned intermediate stage
has occurred for the corresponding interlamellar spacing. Furthermore, simulations
of an artificially generated three-dimensional and polycrystalline Ni-YSZ SOFC
anode are presented. The chosen model is suitable to efficiently delineate herein
coarsening by surface and grain-boundary diffusion with concomitant grain growth.
The simulations reveal that the Ni coarsening rates in the SOFC anode are fastest at
early time and slow down at later times, while, as desired the YSZ-substrate remains
invariant. It was shown, that for the assumed parameters, especially equal interfacial
energies (wetting angle of 90◦ ), no significant change in the tortuosity of nickel occurs.
However, the reduction in triple-phase boundary length, which is also enhanced at
early times, might contribute to the degradation of the anode. Further simulations
for more realistic wetting angles and experimental microstructures are reserved for a
future publication on the coarsening of Nickel in Ni-YSZ SOFC anodes.

Acknowledgements This work was performed on the supercomputer ForHLR II funded by the
Ministry of Science, Research and the Arts Baden-Wuerttemberg and by the Federal Ministry of
Education and Research. The authors gratefully acknowledge financial support of the parallel code
development of the PACE3D package by the Deutsche Forschungsgemeinschaft (DFG) under the
grant numbers NE 822/31-1 (Gottfried-Wilhelm Leibniz prize). Furthermore, funding through
the coordinated research programme “Virtual Materials Design (VirtMat), project No. 9” and
“KNMFi” within the Helmholtz programme “Material System Engineering (MSE), No. 43.31.01”
is granted. Research on SOFC was supported by the Federal Ministry for Economic Affairs and
Energy (BMWi) within the KerSOLife100 project (Funding No.: 03ET6101A) and the HGF Impuls-
und Vernetzungsfonds “Electro-chemical materials for high temperature ion conductors” within
the programme “MTET, No. 38.02.01” Part of the work contributes to the research performed
at CELEST (Center for Electrochemical Energy Storage Ulm-Karlsruhe) and was funded by the
German Research Foundation (DFG) under Project ID 390874152 (POLiS Cluster of Excellence).
Multiphase-field simulations of solid-state phase transformations 183

References

1. N. Moelans, B. Blanpain, and P. Wollants. An introduction to phase-field modeling of

microstructure evolution. Calphad, 32(2):268–294, June 2008.
2. M.R. Tonks and L.K. Aagesen. The phase field method: mesoscale simulation aiding material
discovery. Annual Review of Materials Research, 49:79–102, 2019.
3. J. Hötzer, M. Jainta, P. Steinmetz, B. Nestler, A. Dennstedt, A. Genau, M. Bauer, H. Köstler,
and U. Rüde. Large scale phase-field simulations of directional ternary eutectic solidification.
93(0):194 – 204, 2015.
4. D. Schneider, E. Schoof, O. Tschukin, A. Reiter, C. Herrmann, F. Schwab, M. Selzer, and
B. Nestler. Small strain multiphase-field model accounting for configurational forces and
mechanical jump conditions. Computational Mechanics, 61(3):277–295, 2018.
5. B. Nestler, H. Garcke, and B. Stinner. Multicomponent alloy solidification: Phase-field modeling
and simulations. Phys. Rev. E 71, 041609, 2005.
6. I. Steinbach. Phase-field models in materials science. Modelling and simulation in materials
science and engineering, 17(7):073001, 2009.
7. J. Hötzer, O. Tschukin, M. Ben Said, M. Berghoff, M. Jainta, G. Barthelemy, N. Smorchkov,
D. Schneider, M. Selzer, and B. Nestler. Calibration of a multi-phase field model with
quantitative angle measurement. Journal of materials science, 51(4):1788–1797, 2016.
8. J. Hötzer, A. Reiter, H. Hierl, P. Steinmetz, M. Selzer, and B. Nestler. The parallel multi-physics
phase-field framework Pace3D. Journal of Computational Science, 26:1–12, 2018.
9. J. Hötzer. Massiv-parallele und großskalige Phasenfeldsimulationen zur Untersuchung der
Mikrostrukturentwicklung. PhD thesis, 2017.
10. E. Schoof, C. Herrmann, N. Streichhan, M. Selzer, D. Schneider, and B. Nestler. On the
multiphase-field modeling of martensitic phase transformation in dual-phase steel using 𝐽2 -
viscoplasticity. Modelling and Simulation in Materials Science and Engineering, 27(2):025010,
2019.
11. E. Schoof, D. Schneider, N. Streichhan, T. Mittnacht, M. Selzer, and B. Nestler. Multiphase-field
modeling of martensitic phase transformation in a dual-phase microstructure. International
Journal of Solids and Structures, 134:181–194, 2018.
12. E. Schoof. Chemomechanische Modellierung der Wärmebehandlung von Stählen mit der
Phasenfeldmethode. PhD thesis, Karlsruher Institut für Technologie (KIT), 2020.
13. E. Schoof, C. Herrmann, D. Schneider, J. Hötzer, and B. Nestler. Multiphase-field modeling and
simulation of martensitic phase transformation in heterogeneous materials. In High Performance
Computing in Science and Engineering’18, pages 475–488. Springer, 2019.
14. P. G. Kubendran Amos, E. Schoof, N. Streichan, D. Schneider, and B. Nestler. Phase-field
analysis of quenching and partitioning in a polycrystalline fe-c system under constrained-carbon
equilibrium condition. Computational Materials Science, 159:281–296, 2019.
15. D. Schneider, F. Schwab, E. Schoof, A. Reiter, C. Herrmann, M. Selzer, T. Böhlke, and B. Nestler.
On the stress calculation within phase-field approaches: a model for finite deformations.
Computational Mechanics, 60(2):203–217, 2017.
16. C. Herrmann, E. Schoof, D. Schneider, F. Schwab, A. Reiter, M. Selzer, and B. Nestler.
Multiphase-field model of small strain elasto-plasticity according to the mechanical jump
conditions. Computational Mechanics, 62(6):1399–1412, 2018.
17. A. Choudhury and B. Nestler. Grand-potential formulation for multicomponent phase transfor-
mations combined with thin-interface asymptotics of the double-obstacle potential. Phys. Rev.
E, 85, 2012.
18. T. Mittnacht, P.G. Kubendran Amos, D. Schneider, and B. Nestler. Morphological stability of
three-dimensional cementite rods in polycrystalline system: A phase-field analysis. Journal of
Materials Science and Technology, 77:252–268, 2021.
19. F.A. Nichols. On the spheroidization of rod-shaped particles of finite length. Journal of
Materials Science, 11(6):1077–1082, 1976.
20. P.G. Kubendran Amos, L.T. Mushongera, Tobias Mittnacht, and Britta Nestler. Phase-field
analysis of volume-diffusion controlled shape-instabilities in metallic systems-II: Finite 3-
dimensional rods. Computational Materials Science, 144:374–385, mar 2018.
184 E. Schoof, T. Mittnacht, M. Seiz, P. Hoffrogge, H. Hierl and B. Nestler

21. T. Mittnacht, P.G. Kubendran Amos, D. Schneider, and B. Nestler. Understanding the influence
of neighbours on the spheroidization of finite 3-dimensional rods in a lamellar arrangement:
Insights from phase-field simulations. In Numerical Modelling in Engineering, pages 290–299.
Springer, 2018.
22. K. Ankit, A. Choudhury, C. Qin, S. Schulz, M. McDaniel, and B. Nestler. Theoretical and
numerical study of lamellar eutectoid growth influenced by volume diffusion. Acta Materialia,
2013.
23. San Ping Jiang and Siew Hwa Chan. A review of anode materials development in solid oxide
fuel cells. Journal of Materials Science, 39(14):4405–4439, 2004.
24. P. Haremski, L. Epple, M. Wieler, P. Lupetin, R. Thelen, and M.J. Hoffmann. A Thermal
Grooving Study of Relative Grain Boundary Energies of Nickel in Polycrystalline Ni and in a
Ni/YSZ Anode Measured by Atomic Force Microscopy. SSRN Electronic Journal, 214, 2020.
25. P.W. Hoffrogge, A. Mukherjee, E.S. Nani, P.G. Kubendran Amos, F. Wang, D. Schneider,
and B. Nestler. Multiphase-field model for surface diffusion and attachment kinetics in the
grand-potential framework. Physical Review E, 103(3):033307, mar 2021.
26. M. Plapp. Unified derivation of phase-field models for alloy solidification from a grand-potential
functional. Phys. Rev. E, 84:031601, Sep 2011.
27. M. Trini, P.S. Jørgensen, A. Hauch, J.J. Bentzen, P.V. Hendriksen, and M. Chen. 3D Microstruc-
tural Characterization of Ni/YSZ Electrodes Exposed to 1 Year of Electrolysis Testing. Journal
of The Electrochemical Society, 166(2):F158–F167, feb 2019.
28. B. Münch and L. Holzer. Contradicting geometrical concepts in pore size analysis attained
with electron microscopy and mercury intrusion. Journal of the American Ceramic Society,
91(12):4059–4067, 2008.
29. L. Holzer, B. Iwanschitz, T. Hocker, B. Münch, M. Prestat, D. Wiedenmann, U. Vogt, P. Holtap-
pels, J. Sfeir, A. Mai, and T. Graule. Microstructure degradation of cermet anodes for solid
oxide fuel cells: Quantification of nickel grain growth in dry and in humid atmospheres. Journal
of Power Sources, 196(3):1279–1294, 2011.
30. J. Joos, T. Carraro, A. Weber, and E. Ivers-Tiffée. Reconstruction of porous electrodes by
FIB/SEM for detailed microstructure modeling. Journal of Power Sources, 196(17):7302–7307,
2011.
31. A. Faes, A. Hessler-Wyser, D. Presvytes, C. G. Vayenas, and J. Van herle. Nickel-Zirconia
Anode Degradation and Triple Phase Boundary Quantification from Microstructural Analysis.
Fuel Cells, 9(6):841–851, dec 2009.
32. H. Nahor, H. Meltzman, and W.D. Kaplan. Ni–YSZ(111) solid–solid interfacial energy. Journal
of Materials Science, 49(11):3943–3950, jun 2014.
33. A. Tsoga and P. Nikolopoulos. Surface and grain-boundary energies in yttria-stabilized zirconia
(YSZ-8 mol%). Journal of Materials Science, 31(20):5409–5413, 1996.
34. R. Tran, Z. Xu, B. Radhakrishnan, D. Winston, W. Sun, K.A. Persson, and S.P. Ong. Surface
energies of elemental crystals. Scientific Data, 3(1):160080, dec 2016.
35. M. Kappeler, A. Marusczyk, and B. Ziebarth. Simulation of nickel surfaces using ab-initio and
empirical methods. Materialia, 12(March):100675, aug 2020.
36. R. Davis, F. Abdeljawad, J. Lillibridge, and M. Haataja. Phase wettability and microstructural
evolution in solid oxide fuel cell anode materials. Acta Materialia, 78:271–281, 2014.
37. Z. Jiao and N. Shikazono. Prediction of Nickel Morphological Evolution in Composite Solid
Oxide Fuel Cell Anode Using Modified Phase Field Model. Journal of The Electrochemical
Society, 165(2):F55–F63, 2018.
Bridging scales with volume coupling —
Scalable simulations of muscle contraction and
electromyography

Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

Abstract Measuring the electric signals of a contracting muscle, called electromyog-

raphy (EMG), is a valuable experimental tool for biomechanics researchers. Detailed
simulations of this process can deliver insights on a new level. However, the solution
of appropriate biophysically based multi-scale models exhibits high computational
loads, which demands for High Performance Computing. At the same time, separate
parallelization strategies are required for different spatial scales. Currently, no open-
source software is capable of efficiently exploiting supercomputers and handling
the complexity of relevant state-of-the-art multi-scale muscle models. We employ a
volume coupling scheme to augment the open-source software OpenDiHu and enable
coupled simulations of muscle contraction and EMG. We investigate the weak scaling
behavior of different components of the multi-scale solver and find that the new
approach does not hinder scalability. As a result, we are able to simulate a respective
highly resolved scenario with 100 million degrees of freedom on the supercomputer
Hawk at the High Performance Computing Center Stuttgart.

1 Introduction

The musculoskeletal system enables humans to perform a variety of vital tasks in

everyday life. The capabilities of skeletal muscles satisfy versatile requirements,
ranging from forceful to finely controlled actions and from fast and impulsive ones
to long, enduring tasks. The functioning of this complex system is governed by the
interplay of processes on several scales, including molecular, biochemical processes

Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

Institute for Parallel and Distributed Systems, Universität Stuttgart, Universitätsstraße 38, 70569
Stuttgart, Germany,
e-mail: [email protected],
[email protected],
[email protected],
[email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 185
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_11
186 Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

at the muscle fiber membranes, macroscopic electrical and mechanical properties on

the organ scale, and the neural pathway connecting numerous motor neurons in the
spinal cord with hundreds of thousands of muscle fibers in a skeletal muscle.
Detailed, biophysically-based simulations of the neuromuscular system can foster
physiological understanding, provide an in-silico laboratory to complement the limited
in-vivo studies, and ultimately help to diagnose and treat neuromuscular disorders. For
this purpose, we develop simulation software for state-of-the-art multi-scale models
of the neuromuscular system. As such detailed models with high temporal and spatial
resolutions pose high computational loads, we focus on efficient parallelization and
coupling schemes for these multi-physics models in order to exploit the computational
power of compute clusters and supercomputers.

1.1 Related work

Relevant multi-scale approaches have been proposed in the literature and combine
models of electrophysiology on the one hand, i.e., propagation of electric stimuli
along muscle fibers activating the muscle, and continuum mechanics models of
muscle contraction on the other hand [11, 13, 22, 23]. They have been implemented
using computational software frameworks such as Chaste [19] and OpenCMISS [6].
We base our computations on the multi-scale chemo-electro-mechanical model
formulated by the authors of [11, 20, 23] and initially implemented in OpenCMISS.
In previous work, we enhanced the software and investigated domain decomposition
strategies, which allowed a highly parallel execution of the electrophysiology part of
the model [7]. In our new codebase OpenDiHu, we developed tailored solution schemes
for the overall model exploiting instruction-level and distributed memory parallelism
[15, 17, 18]. OpenDiHu is capable of simulating the multi-scale electrophysiology
model with a realistic number of 270,000 muscle fibers for the biceps brachii muscle.

1.2 Research targets

The measurement of electric signals on the skin surface over a contracting muscle,
called electromyography (EMG), is an important, non-invasive window to gain
insights into the muscle’s functioning. However, besides our OpenDiHu code, no
open-source software currently exist that are capable of simulating both EMG and
muscle contraction using detailed biophysical models. In this work, we present
our numeric and algorithmic setup to enable this simulation and demonstrate that
our software OpenDiHu is able to conduct the respective detailed simulations by
exploiting compute power of the supercomputer Hawk at the High Performance
Computing Center Stuttgart.
Scalable simulations of muscle contraction and EMG 187

In such simulations, models of two different scales use discretizations of the

three-dimensional (3D) muscle domain: First, the electrophysiological descriptions
rely on highly-resolved 3D meshes to accurately represent the finely dosed muscle
activation and high-density EMG signals. Second, the solver of the nonlinear solid
mechanics formulation usually uses a coarser 3D mesh.
In parallel execution using a domain decomposition approach, the possible number
of subdomains in this setup is limited by the small number of elements in the coarser
mechanics mesh. Usually, identical partitionings for the fine and coarse 3D meshes
are beneficial as the two corresponding models are strongly linked and need to
frequently exchange large amounts of data. However, this limits the possible degree
of parallelism and prevents High Performance Computing. Therefore, we want to
use different partitionings and equip our solid mechanics solver with a coarse mesh
and the multi-scale electrophysiology solver with a fine mesh. The high-performance
numerical coupling library preCICE [8] is used to glue both partitionings together
into a single simulation setup.

1.3 Main contributions of this work

In this work, we evaluate the weak scaling performance of our software OpenDiHu
in highly resolved multi-scale simulations of muscular EMG. Further, we present
a scheme to bridge the scales between a detailed multi-scale electrophysiology
model and an organ-level muscle contraction model, which is also applicable in the
area of High Performance Computing. This scheme enables simulations of EMG
measurements on a contracting muscle with an unprecedented level of detail.
The remainder of this paper is structured as follows. Section 2 presents the
used models and their discretization and solution schemes. Section 2 describes
our developed partitioning method to enable the full model for High Performance
Computing. Section 3 presents performance results and Sect. 4 concludes this work.

2 Models and solvers

We describe the multi-scale model using the illustration in fig. 1. Figure 1 (a) shows
the hierarchical structure of a skeletal muscle. A tendon connects the skeleton and
the muscle belly, which consists of dozens of fascicles. Each fascicle contains tens of
thousands of muscle fibers and, in each muscle fiber, strings of sarcomeres form the
molecular motor, which generates the muscle force.
Figure 1 (b) shows the corresponding representation in the discretized domain,
using a 3D finite element mesh (green color) for the macroscopic muscle, numerous
embedded 1D meshes (red color) for the muscle fibers, and (0D) points on every fiber
(yellow color) to describe the sarcomeres.
188 Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

(a) (b) (c)

Fig. 1: Components of the implemented multi-scale model and its discretization,

comprising the muscle belly, muscle fibers and sarcomeres. The schematic in (c)
illustrates the coupled quantities between the four model parts given by different
colors.

Figure 1 (c) summarizes the four parts of the multi-scale model and the coupled
quantities: (i) the 3D continuum mechanics model (green), (ii) the 3D model of the
electric potential (blue color) responsible for the EMG, (iii) numerous instances of the
1D model of electric activation on the muscle fibers (red color), and (iv) numerous
instances of the 0D force generation model of the sarcomere (yellow color). The
0D and 1D models are bidirectionally coupled by the trans-membrane voltage 𝑉𝑚 ,
the 3D continuum mechanics model influences all other models by deforming the
computational domains, and, furthermore, the 0D model is bidirectionally coupled to
the 3D model over the activation 𝛾 and the contraction velocity 𝑙¤𝐻𝑆 . In the following
sections, we formulate the model equations for the four outlined components of the
multi-scale model.

2.1 Continuum mechanics model (3D)

Contraction and deformation of the muscle domain Ω 𝑀 are described by a dynamic,

incompressible, nonlinear solid mechanics formulation. The quantities of interest
¤ and stresses in terms of the second Piola–
are the displacements u, velocities v = u,
Kirchhoff tensor S. The underlying equations are given by the following balance
principles:

gradX (P) + B = 𝜌0 v¤ in Ω 𝑀 (balance of linear momentum), (1)

S = S⊤ in Ω 𝑀 (balance of angular momentum), (2)
divX (v) = 0 in Ω 𝑀 (incompressibility). (3)

Here, eq. (1) is the balance of linear momentum with body forces B in reference
configuration and the first Piola–Kirchhoff stress tensor P = F S with the deformation
gradient F with respect to referential coordinates X. The symmetry of the stress
tensor S in eq. (2) follows from conservation of angular momentum. Equation (3)
Scalable simulations of muscle contraction and EMG 189

follows from conservation of mass assuming a constant density 𝜌0 throughout the

muscle domain. We additionally specify Dirichlet and Neumann boundary conditions
to fix certain displacements and prescribe surface traction forces, respectively.
The constitutive relation between stresses and strains is based on the work of
Heidlauf et al. [12]. The model uses a hyperelastic, transversely-isotropic material
with an additional active stress term modeling force production in the sarcomeres,
which we omit here for the sake of brevity. This term depends on a spatially varying
muscle activation level 𝛾 ∈ [0, 1], which is computed in the 0D model and then
homogenized to the 3D domain. The resulting displacements of the mechanics model
deform the 3D domain and the embedded 1D fiber domains such that all other models
are influenced by this model of muscle contraction.

2.2 Model of the electric potential (3D)

In the 3D muscle domain Ω 𝑀 , extracellular and intracellular spaces are assumed to

coexist at every point. Electric conduction in the extracellular space is described by
the bidomain equation,



div (𝝈 𝑒 + 𝝈 𝑖 ) grad(𝜙𝑒 ) + div 𝝈 𝑖 grad(𝑉𝑚 ) = 0 in Ω 𝑀 . (4)

Here, 𝝈 𝑖 and 𝝈 𝑒 are the conductivity tensors in the intra and extracellular spaces, 𝜙𝑒 is
the electric potential in the extracellular space, and 𝑉𝑚 is the trans-membrane voltage,
measured between the electric potentials of intra and extracellular spaces. If 𝑉𝑚 is
known at every point, we can compute 𝜙𝑒 by solving the given equation. The resulting
value of 𝜙𝑒 corresponds to the electric signals measured during intramuscular EMG.

2.3 Model of muscle fiber activation (1D)

The muscle fibers are frequently stimulated by the nervous system at points around
𝑗
their centers. Upon stimulation of a fiber Ω 𝑓 , an electric spike, called action potential,
propagates towards both ends of the fiber. This is described by the monodomain
equation, which is a 1D diffusion-reaction equation:
!
𝜕𝑉𝑚 1 𝜕 2𝑉𝑚 𝑗 𝑗
= 𝑗 𝑗 𝜎eff 2
− 𝐴𝑚 𝐼ion (𝑉𝑚 , y) on Ω 𝑓 . (5)
𝜕𝑡 𝐴 𝑚 𝐶𝑚 𝜕𝑠

𝑗 𝑗
Here, 𝑉𝑚 is again the trans-membrane voltage, 𝐴𝑚 and 𝐶𝑚 are the surface-to-volume-
ratio and the electric capacitance of the membrane of fiber 𝑗, respectively, 𝜎eff is the
effective conductivity, 𝑠 is the spatial coordinate along the fiber, and 𝐼ion is the current
over the membrane. 𝐼ion acts as the reaction term in this equation and is computed by
the 0D model.
190 Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

After solving this equation for 𝑉𝑚 , we prolong the 𝑉𝑚 values from the 1D fiber
𝑗
domains Ω 𝑓 to the 3D muscle domain Ω 𝑀 using tri-linear interpolation and can
subsequently compute the model of the 3D electric potential given in eq. (4).

2.4 Subcellular force generation model (0D)

𝑗
On spatial “0D” points on every muscle fiber Ω 𝑓 , we solve the dynamics of opening
and closing ion channels in the fiber membranes, intracellular calcium dynamics
and cross-bridge cycling leading to force production. These dynamic processes are
described by a system of differential-algebraic equations:

𝜕y
= 𝐺 (𝑉𝑚 , y), 𝐼ion = 𝐼ion (𝑉𝑚 , y), 𝛾 = 𝐻 (y, 𝜆¤ 𝑓 ) on Ω𝑖𝑠 (6)
𝜕𝑡
Here, y is the vector of evolving internal states, from which quantities such as the
ionic current 𝐼ion and the muscle activation 𝛾 can be derived. 𝐼ion and 𝛾 additionally
depend on the trans-membrane voltage 𝑉𝑚 given by the 1D model in eq. (5) and the
contraction velocity 𝜆¤ 𝑓 in fiber direction computed by the 3D continuum mechanics
model, eqs. (1) to (3).
Models with different degrees of detail exist, e.g., the model of membrane potential
depolarization by Hodgkin and Huxley [14] with y ∈ R3 or the model proposed by
Shorten et al. [24] with y ∈ R56 . Subcellular models can be conveniently stored in
CellML format [9], an XML-based description language, and can directly be loaded
and solved in OpenDiHu. Such CellML models are exchanged among bioengineering
researchers via an open-source online model repository1 established by the Physiome
project [21], which hosts a multitude of curated models.

2.5 Discretization and solution

All spatial derivatives in the presented models are discretized using the finite element
method and the meshes introduced in fig. 1. The 3D mechanics model uses Taylor-
Hood elements, where the displacements u and velocities v are discretized using
quadratic ansatz functions and a Lagrange multiplier 𝑝 that enforces incompressibility,
identified as the hydrostatic pressure, is discretized using linear ansatz functions. The
transmembrane voltage 𝑉𝑚 on the 1D fiber domains is also described using linear
ansatz functions. The nodes of the 1D meshes are the locations where the 0D model
instances are solved. Thus, we need to solve as many 0D model instances as there are
fibers multiplied by the number of nodes in each 1D fiber mesh.

1 https://www.cellml.org/
Scalable simulations of muscle contraction and EMG 191

We solve the overall multi-scale model by a subcycling scheme. After solving one
timestep of the two 3D models, we proceed to solve several smaller timesteps of the
0D and 1D models using the Strang operator splitting [25], before we continue to
again solve the 3D models. The 3D mechanics model is discretized in time using the
implicit Euler method, the 1D model is solved using the Crank–Nicolson method,
and, for the 0D model, we use Heun’s method. More details can be found in [15, 18].
We perform all computations within our open-source software framework
OpenDiHu (“Digital Human Model”) [1], where modular time stepping schemes and
solvers for the individual parts of the multi-scale model can be combined. OpenDiHu
interfaces the numeric solver and preconditioner libraries PETSc [3–5], MUMPS
[2], and HYPRE [10]. Tailored optimizations are implemented for the 0D and 1D
models, e.g., to explicitly exploit vector instructions using the library ‘Vc’ [16]
or to communicate the data of every 1D model to a single core and then use an
efficient linear-complexity Thomas algorithm to solve the respective linear system of
equations.

3 Partitioning methods

As mentioned in Sect. 1.2, we first partition the meshes on all domains in an identical
way, such that each process owns a distinct part of the overall geometry. Figure 2
(a) and (b) illustrate this concept, showing the same domain decomposition into
16 subdomains for both the 3D mesh (fig. 2 (a)) and the 1D fiber meshes (fig. 2
(b)). While this partitioning approach prevents communication during data mapping
between the coupled model parts, it can hinder parallelization. In fact, in fig. 2 (a), the
3D mesh consists of 16 elements with quadratic ansatz functions, as used by the solid
mechanics model, and, consequently, a partitioning with more than 16 subdomains is
not possible. However, for the computationally intense fiber models we would like to
use a higher degree of parallelism, e.g., a partitioning with 1024 processes as shown
in fig. 2 (c).
To resolve this issue, we instead split our multi-scale solver implementation in
OpenDiHu into two separate programs. The first program solves the 0D and 1D
model parts and the 3D electric potential model part, which can be summarized as
the multi-scale electrophysiology model. The second program only solves the 3D
solid mechanics part. Both programs are coupled using the library preCICE and can,
thus, make use of individual domain decompositions.
Figure 3 presents the resulting coupled architecture. The yellow block on the
left-hand side represents the first simulation program, which solves the 3D model of
the electric potential and the EMG and the 0D/1D fiber models. The program uses
finely resolved 3D and 1D fiber meshes that are partitioned identically, here visualized
for 16 processes. The data mapping and numerical coupling between the 0D, 1D, and
3D models are performed entirely within OpenDiHu. The right yellow block in fig. 3
192 Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

Fig. 2: Partitioning schemes of muscle and fiber domains. (a) Domain decomposition
of a 3D mesh for 16 processes, given by the different colors. (b) The identical domain
decomposition applied to the 1D meshes of 81 fibers. (c) A different partitioning for
the same number of fibers, but for 1024 processes.

Fig. 3: Overview of the software coupling scheme between two different instances of
the solver OpenDiHu using the coupling library preCICE. The left program utilizes
more processes and uses finer meshes for the detailed electrophysiology simulations
than the program on the right-hand side, which uses a coarser 3D mesh for the
mechanics model. The figure shows the value exchanges between the solvers within
OpenDiHu and between the programs by blue arrows.
Scalable simulations of muscle contraction and EMG 193

represents the second OpenDiHu program that solely solves the mechanics model. It
uses a coarser mesh with different domain decomposition, in the shown example for
only two processes.
The bidirectional data mapping and data communication between the two programs
is handled by preCICE. The degrees of freedom (dofs) are mapped between the fine
3D mesh on the left-hand side and the coarse 3D mesh on the right-hand side. Note
that a direct mapping between the muscle fiber meshes on the left-hand side and the
coarse 3D mesh on the right-hand side would not be feasible because of the large
number of nodes in the muscle fiber meshes.
We use the serial-explicit coupling functionality of preCICE and employ a
consistent mapping with compact polynomial radial basis functions (cf. [8]).

4 Performance results

In this section, we present performance and simulation results for the described multi-
scale skeletal muscle model. Section 4.1 begins with a weak scaling investigation,
followed by results of our introduced coupling scheme in Sect. 4.2.

4.1 Weak scaling of the electrophysiology solver

In a first step, we study the weak scaling behavior of the multi-scale electrophysiology
model, i.e., the 0D, 1D, and 3D electric potential model parts. We increase the
problem size by adding more muscle fibers until the realistic number of 270 × 103
fibers for the biceps brachii muscle is reached. At the same time, we increase the
number of processes to 26,912. We run the simulation on the supercomputer Hawk at
the High Performance Computing Center Stuttgart. Each compute nodes consists of a
dual-socket AMD EPYC 7742 processor with 2.25 GHz base frequency and 256 GB
RAM per node. We use 64 processes per compute node.
The scenario contains the 0D model of Hodgkin and Huxley [14], which we solve
with a timestep width of 𝑑𝑡0D = 10−3 ms. We simulate a time span of 𝑡 end = 2 ms,
corresponding to two invocations of the 3D model solver in the described subcycling
scheme. To proportion the time for initialization, we multiply the measured runtimes
for all parts except the initialization by the factor 500 such that the results correspond
to a simulation time of 1 s.
Figure 4 shows the resulting runtimes of the different solvers and illustrates their
main characteristics. It can be seen that the computation of the 0D model contributes
most to the total runtime for the scenarios with up to 3600 muscle fibers, whereas the
1D model solver consistently exhibits relatively low runtimes. Further, the 0D and 1D
solvers show a perfect weak scaling behavior even for massively parallel simulations
194 Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

105
Total
3D model
104 0D model
1D model
103 Initialization
Communication 0D,1D
Runtime [s]

102

101

100

1
10

2
10
18 64 144 448 3600 18432
169 625 1369 4489 34969 182329
1152 7744 26912
11881 76729 273529
Number of processes
Number of fibers

Fig. 4: Weak scaling results for the electrophysiology solvers with runtimes for the
different model components (different colors, see legend) and standard deviation over
all processes (vertical bars) using between 18 and 26,912 cores and between 169 and
273,529 muscle fibers.

on up to 26,912 cores. The associated communication cost between the 0D and 1D

models is negligible. Also, the time for initialization is relatively low compared with
the solver runtimes.
We observe an increase in runtime for the 3D model solver, which is due to an
increasing number of iterations in the conjugate gradient scheme for larger numbers
of degrees of freedom, which we plan to address in future work by implementing a
multigrid preconditioner. However, the scaling behavior suffices to be able to run
simulations up to the maximum physiologically meaningful resolution of 270,000
muscle fibers of a biceps brachii muscle.

4.2 Performance evaluation of the full multi-scale solver

Next, we perform simulations of the entire multi-scale model of muscular EMG

and muscle contraction that was described in Sect. 2. We use the coupling scheme
presented in Sect. 2 where the total multi-scale solver is split into two parts, first, the
highly parallel multi-scale electrophysiology solver and, second, the solid mechanics
solver.
Scalable simulations of muscle contraction and EMG 195

In contrast to the weak scaling study in Sect. 4.1, the simulations here require
the more detailed and compute-intense sarcomere model of Shorten et al. [24] to
compute the muscle activation for the 3D mechanics model. In addition to the larger
number of dofs in this model, a smaller time step width of 𝑑𝑡0D = 1.25 × 10−5 ms has
to be used, which further increases the computational work.

Multi-scale Electrophysiology Solver Mechanics Solver

processes fibers fine 3D mesh total processes coarse 3D mesh total

nodes unknowns nodes unknowns

256 49 6321 4,070,185 1 441 2726

1024 169 21,801 14,037,985 1 129 3959
4096 625 80,625 51,915,625 1 225 1395
9216 1369 176,601 113,715,985 1 225 1395

Table 1: Simulation of muscle contraction and EMG, parameters for different scenarios
with varying problem sizes and numbers of processes.

We compute four differently sized scenarios for a time span of 𝑡 = 100 ms

and evaluate the runtimes of the individual solvers. Table 1 lists the discretization
parameters for the four scenarios. The electrophysiology solver is run with between
256 and 9216 processes on Hawk using 64 processes per compute node, similar
to the study in Sect. 4.1. Accordingly, we increase the problem size and consider
between 49 and 1369 1D muscle fiber meshes with a constant number of 1481 nodes
each. The mesh resolution of the 3D mesh used for the 3D electric potential model is
also increased, as listed in table 1. As a consequence, the total number of unknowns
increases to over 113 million for the largest scenario.
For all four scenarios, we execute the mechanics solver in serial using a coarser
3D mesh. This is justified, as we are only interested in the macroscopic, uniform
deformation of the contracting muscle, in contrast to the spatially higher resolved
electric potentials computed by the 3D EMG model. The preCICE library maps and
communicates data between the fine and coarse 3D meshes, whose node counts are
listed in the third and second last columns in table 1, respectively. The runtime of
the coupled mechanics solver without waiting time is below two minutes in every
scenario, which is negligible compared to the runtimes of several hours for the
electrophysiology solver.
Figure 5 presents the resulting runtimes in the electrophysiology solver. The 0D,
1D and 3D solver parts show characteristics similar to the weak scaling study in
Sect. 4.1. In addition, it can be seen that the preCICE mapping contributes negligible
additional runtime and exhibits good weak scaling properties. This shows us that our
software setup using preCICE can also be used for even larger runs in future work.
196 Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

total
2 0D model
10
1D model
Runtime [s]

3D EMG model
100 preCICE mapping

2
10

256 1024 4096 9216

Number of processes

Fig. 5: Weak scaling study of the full multi-scale model of muscular EMG and
contraction, including the preCICE volume coupling scheme. The runtimes of the
model parts (different colors, see legend) are given for the three scenarios specified
in table 1.

4.3 Simulation results of the full multi-scale model

Finally, we present simulation results of a large scenario corresponding to the second

last row in table 1. Solving this scenario for a simulation time span of 190 ms takes
around 8 h 27 min using 64 compute nodes on Hawk.
The left plot in Figure 6 shows the fine fiber meshes with the trans-membrane
potential 𝑉𝑚 . Action potentials can be seen spreading from the center of the fibers
towards to top and bottom ends of the muscle. The middle plot shows the resulting
EMG given by the extracellular potential 𝜙𝑒 . The right plot depicts the reference
configuration of the solid mechanics mesh by the yellow wireframe and the current
deformed mesh colored according to the active Piola–Kirchhoff stress. It can be seen
that the three different meshes are identically deformed, as a result of the numerical
coupling. Judging from experiences in experimental studies, the simulation results
look plausible. For example, the results exhibit a delay between the onset of the
muscle activation and the actual shortening of the muscle, which can also be observed
in real experiments.

5 Conclusion and outlook

In this work, an existing, detailed multi-scale model of the neuromuscular system

was described, which can be used to simulate EMG and muscle contraction. The
simulation software OpenDiHu is capable of solving different configurations of
this multi-scale model framework using High Performance Computing techniques.
However, the combination of the solid mechanics model on the macroscopic organ
scale with the electrophysiology models on a smaller spatial scale hindered highly
Scalable simulations of muscle contraction and EMG 197

Fig. 6: Simulation result at 𝑡 = 110 ms of the EMG and muscle contraction model
with 615 muscle fibers of the biceps brachii muscle. The simulation was computed
with 4096 processes on the supercomputer Hawk. The three images depicts the
trans-membrane voltage 𝑉𝑚 (left), the muscle activation 𝛾 (middle) and the active
Piola–Kirchhoff stress (right).

parallel executions. In the present paper, we described a scheme to separate the

partitionings of these models using the volume coupling functionality of the library
preCICE, ultimately resolving the issue.
We investigated the weak scaling behavior of the electrophysiology solver and
found good scaling properties for the 0D and 1D model parts, while the 3D part could
still be improved. Furthermore, we demonstrated the use of our new coupling scheme
with the full multi-scale model and, again, found good weak scaling behavior also for
the added coupling part.
Finally, we showed unprecedented simulation results of the combined muscular
EMG and muscle contraction model using a fine spatial resolution and a detailed
biophysical model with more than 100 million degrees of freedom in total. Such
simulations are only possible by exploiting High Performance Computing facilities
such as the supercomputer Hawk.
Future work needs to address the mediocre weak scaling behavior of the 3D action
potential model, which currently uses an unpreconditioned conjugate gradient solver.
Here, the use of a multigrid preconditioner seems promising. Further studies can also
extend the simulations to even higher numbers of muscle fibers.

Acknowledgements We thank the Deutsche Forschungsgemeinschaft (DFG, German Research

Foundation) for supporting this work by funding – EXC2075 – 390740016 under Germany’s
Excellence Strategy. Furthermore, we acknowledge the support by the Stuttgart Center for Simulation
Science (SimTech).
198 Benjamin Maier, David Schneider, Miriam Schulte and Benjamin Uekermann

References

1. OpenDiHu Documentation. https://opendihu.readthedocs.io/en/latest/, 2021. Accessed: 2021-

02-12.
2. Patrick Amestoy, Alfredo Buttari, Jean-Yves L’Excellent, and Théo Mary. On the complexity
of the block low-rank multifrontal factorization. SIAM Journal on Scientific Computing,
39(4):1710–1740, 2017.
3. Satish Balay, Shrirang Abhyankar, Mark F. Adams, Jed Brown, Peter Brune, Kris Buschelman,
Lisandro Dalcin, Alp Dener, Victor Eijkhout, William D. Gropp, Dmitry Karpeyev, Dinesh
Kaushik, Matthew G. Knepley, Dave A. May, Lois Curfman McInnes, Richard Tran Mills, Todd
Munson, Karl Rupp, Patrick Sanan, Barry F. Smith, Stefano Zampini, Hong Zhang, and Hong
Zhang. PETSc users manual. Technical Report ANL-95/11 - Revision 3.14, Argonne National
Laboratory, 2020.
4. Satish Balay, Shrirang Abhyankar, Mark F. Adams, Jed Brown, Peter Brune, Kris Buschelman,
Lisandro Dalcin, Alp Dener, Victor Eijkhout, William D. Gropp, Dinesh Kaushik, Matthew G.
Knepley, Dave A. May, Lois Curfman McInnes, Richard Tran Mills, Todd Munson, Karl Rupp,
Patrick Sanan, Barry F. Smith, Stefano Zampini, Hong Zhang, and Hong Zhang. PETSc Web
page. http://www.mcs.anl.gov/petsc, 2018.
5. Satish Balay, William D. Gropp, Lois Curfman McInnes, and Barry F. Smith. Efficient
management of parallelism in object oriented numerical software libraries. In E. Arge, A. M.
Bruaset, and H. P. Langtangen, editors, Modern Software Tools in Scientific Computing, pages
163–202. Birkhäuser Press, 1997.
6. Chris Bradley, Andy Bowery, Randall Britten, Vincent Budelmann, Oscar Camara, Richard
Christie, Andrew Cookson, Alejandro F. Frangi, Thiranja Babarenda Gamage, Thomas Heidlauf,
Sebastian Krittian, David Ladd, Caton Little, Kumar Mithraratne, Martyn Nash, David Nickerson,
Poul Nielsen, Øyvind Nordbø¸, Stig Omholt, Ali Pashaei, David Paterson, Vijayaraghavan
Rajagopal, Adam Reeve, Oliver Röhrle, Soroush Safaei, Rafael Sebastián, Martin Steghöfer,
Tim Wu, Ting Yu, Heye Zhang, and Peter Hunter. OpenCMISS: A multi-physics & multi-scale
computational infrastructure for the VPH/Physiome project. Progress in Biophysics and
Molecular Biology, 107(1):32–47, 2011. Experimental and Computational Model Interactions
in Bio-Research: State of the Art.
7. Chris P. Bradley, Nehzat Emamy, Thomas Ertl, Dominik Göddeke, Andreas Hessenthaler,
Thomas Klotz, Aaron Krämer, Michael Krone, Benjamin Maier, Miriam Mehl, Rau Tobias, and
Oliver Röhrle. Enabling detailed, biophysics-based skeletal muscle models on HPC systems.
Frontiers in Physiology, 9(816), 2018.
8. Hans-Joachim Bungartz, Florian Lindner, Bernhard Gatzhammer, Miriam Mehl, Klaudius
Scheufele, Alexander Shukaev, and Benjamin Uekermann. preCICE – a fully parallel library
for multi-physics surface coupling. Computers and Fluids, 141:250–258, 2016. Advances in
Fluid-Structure Interaction.
9. Autumn A. Cuellar, Catherine M. Lloyd, Poul F. Nielsen, David P. Bullivant, David P. Nickerson,
and Peter J. Hunter. An Overview of CellML 1.1, a Biological Model Description Language.
SIMULATION, 79(12):740–747, 2003.
10. Robert D. Falgout and Ulrike Meier Yang. Hypre: A library of high performance preconditioners.
In Proceedings of the International Conference on Computational Science-Part III, ICCS ’02,
pages 632–641, Berlin, Heidelberg, 2002. Springer-Verlag.
11. T. Heidlauf and O. Röhrle. Modeling the Chemoelectromechanical Behavior of Skeletal
Muscle Using the Parallel Open-Source Software Library OpenCMISS. Computational and
Mathematical Methods in Medicine, 2013:1–14, 2013.
12. T. Heidlauf and O. Röhrle. A multiscale chemo-electro-mechanical skeletal muscle model
to analyze muscle contraction and force generation for different muscle fiber arrangements.
Frontiers in Physiology, 5(498):1–14, 2014.
13. B. Hernández-Gascón, J. Grasa, B. Calvo, and J. F. Rodríguez. A 3D electro-mechanical
continuum model for simulating skeletal muscle contraction. Journal of Theoretical Biology,
335:108–118, 2013.
Scalable simulations of muscle contraction and EMG 199

14. A. L. Hodgkin and A. F. Huxley. A quantitative description of membrane current and its
application to conduction and excitation in nerve. The Journal of physiology, 117(4):500–544,
1952.
15. Aaron Krämer, Benjamin Maier, Tobias Rau, Felix Huber, Thomas Klotz, Thomas Ertl, Dominik
Göddeke, Miriam Mehl, Guido Reina, and Oliver Röhrle. Multi-physics multi-scale HPC
simulations of skeletal muscles. In Wolfgang E. Nagel, Dietmar H. Kröner, and Michael M.
Resch, editors, High Performance Computing in Science and Engineering ’ 20: Transactions of
the High Performance Computing Center, Stuttgart (HLRS) 2020, 2021.
16. Matthias Kretz and Volker Lindenstruth. Vc: A C++ library for explicit vectorization. Software:
Practice and Experience, 42(11):1409–1430, 2012.
17. Benjamin Maier, Nehzat Emamy, Aaron S. Krämer, and Miriam Mehl. Highly parallel multi-
physics simulation of muscular activation and EMG. In COUPLED PROBLEMS 2019, pages
610–621, 2019.
18. Benjamin Maier, Dominik Göddeke, Felix Huber, Thomas Klotz, Oliver Röhrle, and Miriam
Schulte. OpenDiHu - Efficient and Scalable Software for Biophysical Simulations of the
Neuromuscular System (forthcoming). Journal of Computational Physics, 2021.
19. Gary R. Mirams, Christopher J. Arthurs, Miguel O. Bernabeu, Rafel Bordas, Jonathan Cooper,
Alberto Corrias, Yohan Davit, Sara-Jane Dunn, Alexander G. Fletcher, Daniel G. Harvey,
Megan E. Marsh, James M. Osborne, Pras Pathmanathan, Joe Pitt-Francis, James Southern,
Nejib Zemzemi, and David J. Gavaghan. Chaste: An open source c++ library for computational
physiology and biology. PLOS Computational Biology, 9(3):1–8, 03 2013.
20. M. Mordhorst, T. Heidlauf, and O. Röhrle. Predicting electromyographic signals under realistic
conditions using a multiscale chemo-electro-mechanical finite element model. Interface Focus,
5(2):1–11, February 2015.
21. IUPS Physiome Project. Physiome Model Repository. https://models.physiomeproject.org/,
2020. [Online; accessed 8-December-2020].
22. O. Röhrle, J. B. Davidson, and A. J. Pullan. Bridging scales: a three-dimensional electrome-
chanical finite element model of skeletal muscle. SIAM Journal on Scientific Computing,
30(6):2882–2904, 2008.
23. O. Röhrle, J. B. Davidson, and A. J. Pullan. A physiologically based, multi-scale model of
skeletal muscle structure and function. Frontiers in Physiology, 3, 2012.
24. P. R. Shorten, P. O’Callaghan, J. B. Davidson, and T. K. Soboleva. A mathematical model of
fatigue in skeletal muscle force contraction. Journal of Muscle Research and Cell Motility,
28(6), 2007.
25. Gilbert Strang. On the construction and comparison of difference schemes. SIAM Journal on
Numerical Analysis, 5(3):506–517, 1968.
Computational Fluid Dynamics

The following presents a selection of research projects conducted in the field of

Computational Fluid Dynamics (CFD) during the reporting period. Numerical
simulations were carried out on the new HPE Apollo supercomputer “Hawk” of the
High Performance Computing Center Stuttgart (HLRS) as well as on the ForHLR II
of the Steinbuch Centre for Computing (SCC) in Karlsruhe. As in previous years,
CFD had the strongest demand for supercomputing resources of all disciplines.
A share of about 45% of the available computing power of Hawk as well as of
ForHLR II were made available to researchers in this field, helping them to expand
scientific knowledge in fundamental flow physics or to better understand complex
flow behaviour in technical applications. Many important results and new scientific
findings have been obtained and remarkable progress has been made. Research has
focused on the simulation of fluid dynamic phenomena that can only be captured
by extremely fine temporal and spatial discretization, often in combination with
high-order methods, as well as on the hardware-related optimization of the employed
numerical codes with respect to computational performance.
This year, 36 annual reports were submitted and subjected to a peer review process.
Research groups from various institutions across the country took advantage of
the supercomputer resources offered by the two centres. From the large number
of excellent reports, 11 contributions were selected for publication in this book.
The spectrum of projects presented in the following covers fundamental research,
investigations into application-oriented problems with industrial relevance as well as
further developments of codes and algorithms with respect to the HPC architecture.
Challenging fluid dynamic problems are addressed in the fields of bypass transition
of compressible boundary layers, gas turbine cooling, hydropower turbines, spray
painting processes, and aerodynamic propeller-wing interactions, to name but a
few. Different numerical methods, such as Finite Difference, Finite Volume, and
Discontinuous Galerkin, were employed using in-house, commercial or open-source
codes. The first two articles describe the results of so-called Gauss Large Scale
Projects, i.e. these projects were granted more than 100 million core hours per year
on Hawk from the Gauss Centre for Supercomputing (GCS).

201
202 Computational Fluid Dynamics

The contribution of Borgelt, Hösgen, Meinke and Schröder from the Institute of
Aerodynamics at RWTH Aachen University deals with rim seal flow in an axial
turbine stage. The rim seal shall prevent hot gas ingress into the wheel space between
the stator and the rotor, which could lead to machine failure. Experiments as well as
unsteady RANS simulation had shown that unsteady flow phenomena occur inside the
rim seal cavity. However, it had been observed that URANS could not reliably predict
the complex unsteady turbulent flow field for each operating condition. Therefore,
the authors studied the flow around the rim seal and in the cavity using large eddy
simulations (LES). They found that the instantaneous results strongly depend on the
cooling gas mass flow rate and are particularly pronounced for lower fluxes. In this
case, oscillations of the radial velocity in the rim seal gap occur which lead to an
ejection of the cooling gas out of the wheel space and an ingress of hot gas from the
main flow, resulting in an reduced cooling effectiveness. Since LES is quite expensive,
the authors also applied a zonal RANS/LES method to a generic turbine setup and
found good agreement with pure LES, qualifying the hybrid method for future work
on the prediction of hot gas ingress in axial turbines.
Ohno, Selent, Kloker and Rist from the Institute of Aerodynamics and Gas
Dynamics, University of Stuttgart, present results of their direct numerical simulations
(DNS) of bypass transition in a compressible boundary layer induced by isotropic free-
stream turbulence. To create the turbulent inlet condition, modes of the continuous
spectrum resulting from linear stability analysis were superposed. So far, this approach
by Jacobs and Durbin has been used by various researchers for incompressible flow
problems. Now, the authors have adapted it for compressible flows. Comparison with
DNS data from literature validates the numerical method for quasi-incompressible
flows as well as for higher Mach number (𝑀 = 0.7). In the latter case, the results
reveal a clear influence of the compressibility on the transition process. A performance
analysis of the applied in-house high-order finite-difference code NS3D shows a near
ideal parallel efficiency for both strong and weak scaling up to 1024 nodes on Hawk.
In industrial gas turbines, water is ingested to improve the thermal efficiency by
cooling the air before and during compression. However, the interaction between
the liquid droplets and the compressor components can lead to a degradation of
the structure. To study this interaction in detail, Schlottke, Ibach, Steigerwald and
Weigand from the Institute of Aerospace Thermodynamics at the University of
Stuttgart address the atomization process of a liquid rivulet at the trailing edge of
a compressor blade. Up to now, only experimental investigations are known from
literature. They carried out DNS with their in-house multiphase flow solver FS3D
and compared them with experimental data generated in their own test facility.
Additionally, a grid dependence study was performed revealing that the numerical
setup is very well suited to reproduce the different atomization processes qualitatively,
but even with a very high grid resolution (more than a billion grid cells), some
features still cannot be resolved. In a comprehensive performance study on Hawk
with different compilers installed on the new system and various optimization options,
significant efficiency gains could be achieved with the appropriate settings.
Computational Fluid Dynamics 203

Another working group at the same institute is concerned with the issue of turbine
blade cooling, which is mandatory if an increase of system efficiency is to be achieved
through higher turbine entry temperatures. Swirl cooling is a promising new technique
as it causes high heat transfer rates. On the other hand, high pressure losses occur due
to axial flow reversal resulting from vortex breakdown. Seibold and Weigand present a
numerical study in which they used delayed detached eddy simulations (DDES) with
OpenFOAM to analyse the impact of convergent tube geometries on the flow field
and the heat transfer in a swirl cooling system. They found that the converging tubes
enforce an axial and a circumferential flow acceleration, with the former counteracting
the flow reversal. By this, the vortex breakdown can be suppressed. Furthermore, the
flow becomes more insensitive to disturbances from the tube outlet. The heat transfer
in terms of Nusselt numbers increased significantly compared to a pipe flow without
swirl, but shows a strong dependency on the tube geometry. Overall, good agreement
between the numerical results and experimental data was achieved. The simulations
were performed on the ForHLR II.
Besides swirl cooling, other cooling methods are common to avoid thermal
damage to turbine blades, e.g. the use of pin fins or ribs inside the flow channels. In
combination with high Reynolds numbers, such complex geometries make the use
of LES very time consuming. Thus, for industrial applications, RANS methods are
usually chosen despite their lower accuracy. The project of Wellinger and Weigand
aims at providing very accurate results for various cooling features associated with
periodic pin-fin arrays and ribbed channels in order to evaluate the validity of the linear
Boussinesq hypothesis. The latter establishes a relationship between the unknown
Reynolds stresses and the mean strain rate tensor and is the basis of most existing
RANS models. To this end, LES is used to generate accurate data for different test
cases, which themselves are validated with experimental or DNS data. The plotted
results in terms of the misalignment of the eigenvectors of the Reynolds stress
tensor and the strain rate tensor give a very good impression of the areas where the
linear Boussinesq hypothesis can be considered valid and where it is violated. The
simulations were performed with the commercial STAR-CCM+ code on ForHLR II
using about 25,000 cells per core yielding a parallel efficiency of almost 80%.
The next two contributions are from the Numerical Methods group of the Institute
of Aerodynamics and Gas Dynamics at the University of Stuttgart and are based
on a high-order Discontinuous Galerkin (DG) method developed in the group over
the last years. While LES of flows along complex geometries are extremely time
consuming and RANS simulations suffer from well-known shortcomings with regard
to laminar-turbulent transition and flow separation, hybrid methods have proven to be
a valuable compromise for industrial application. This class of methods includes zonal
LES, which reduces the computational effort compared to pure LES significantly, but
requires the generation of time-accurate turbulent inflow conditions from the RANS
solution at the RANS-LES interface. Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz,
Schwarz and Munz present two turbulent inflow methods that they had implemented
in their open-source simulation framework FLEXI, a high-order DG Spectral Element
method. The inflow methods are a recycling-rescaling anisotropic linear forcing
and a synthetic eddy method. They are compared to numerical reference cases for
204 Computational Fluid Dynamics

turbulent boundary layers along a flat plate. A particular focus of the report is on the
HPC application on the new HAWK architecture, and current performance data are
provided.
For the modelling of compressible multi-phase flows, there are two main simulation
approaches, the diffuse-interface and the sharp-interface method. A variant of the
latter is the high-order level-set ghost-fluid method that has been developed in the
Numerical Methods group. The method used exhibits a dynamic variation in the
computational workload, and a dynamic load balancing strategy is required to ensure
an efficient resource utilization on large-scale supercomputers. Appel, Jöns, Keim,
Müller, Zeifang and Munz present such a strategy for their numerical framework,
which consists of the flow solver FLEXI and an interface-capturing scheme. The
strong scaling behaviour was investigated up to 16,384 cores for a generic setup
revealing near-ideal parallel efficiency and a significant performance gain compared
to the previous, unbalanced simulations. Actually, the authors demonstrate that
considerable runtime reductions are achievable even for more complex, realistic
scenarios.
The article of Ye and Dreher from the Fraunhofer Institute for Manufacturing
Engineering and Automation in Stuttgart and Shen from the Esslingen University of
Applied Science deals with industrial spray painting processes. In order to enhance
the appearance of the colour, as, for example, by metallic effect coatings in the
automotive industry, effect pigments (small flat flakes) are added to the paint. It is
known from comprehensive experimental studies that the orientation of the flakes
within the paint layer influences the final metallic effect decisively and strongly
depends on the processes during the impact of the paint droplets on solid surfaces.
However, numerical investigations have not been reported in literature so far. The
authors present for the first time a detailed numerical study of flake orientation
during droplet impact on dry and wetted solid walls. For that purpose, a 6-degree-
of-freedom model describing the rigid body motion of the flakes was implemented
in the commercial CFD-code ANSYS-FLUENT and validated with experimental
observations. Subsequently, a parameter study was performed at the HLRS for varying
grid resolutions, initial flake positions, Reynolds and Ohnesorge numbers using the
Volume-of-Fluid method. The study provides a valuable contribution to a better
understanding of the painting processes.
Denev, Zirwes, Zhang and Bockhorn from the Steinbuch Centre for Computing and
the Engler-Bunte Institute for Combustion Technology, respectively, at the Karlsruhe
Institute of Technology present a new three-dimensional, explicit low-pass Laplace
filter for linear forcing. The latter is often used in numerical codes for DNS or
LES to force turbulence and maintain it stable throughout the solution time and in
the complete computational domain. However, when applied in physical space, it
degrades the numerical efficiency, so low-pass filtering of the velocity field is used.
The authors provide a new filter that is numerically more efficient than existing ones,
has good scaling properties and resolves a larger scale range of turbulence. The new
filter and a second improvement, i.e., a particular form of the linear forcing term,
have been implemented and successfully tested in OpenFOAM. The report contains a
Computational Fluid Dynamics 205

mathematical description of the filter, DNS results, and a performance and efficiency
study performed during the installation phase of the new HoreKa cluster, which later
replaced ForHLR II at the SCC in Karlsruhe.
Hydropower turbines are expected to play a key role in compensating fluctuations
in the power grid. Wack, Zorn and Riedelbauch from the Institute of Fluid Mechanics
and Hydraulic Machinery at the University of Stuttgart investigated the vortex-induced
pressure oscillations in the runner of a model-scale Francis turbine at a far off-design
operating point. The focus was on determining the influence of mesh resolution
on torque and head at deep part load conditions. The mesh resolution considerably
impacts the prediction of the inner-blade vortices and further vertical structures
traveling upstream. Refinement in the main flow region as well as in the boundary
layer notably improved the results in terms of vortex location and induced pressure
fluctuations. For this study, the commercial CFD software ANSYS-CFX was applied
using a hybrid RANS/LES approach with an SBES turbulence model. Frequency
spectra of the pressure fluctuations are compared with model tests. The code showed
good scaling behaviour down to about 36,000 cells per core on Hawk.
The aim of reducing CO2 emission in aviation is the prominent goal of current
aeronautical research. One path is the use of full or hybrid electric propulsion systems.
The separation of energy supply and propulsive force generation via an electric motor
driving a propeller offers new design opportunities, two of which are wing-tip mounted
propellers and distributed electric propulsion with many propellers arranged along
the wingspan. If properly designed, the mutual interaction between the propeller(s)
and the wing can improve the aerodynamic performance of the aircraft. To find an
optimal solution, CFD simulations are very helpful, but also very expensive due to the
wide, multi-dimensional parameter space. One way to reduce the computational effort
is to model the propellers using an Actuator Disc (ACD) or an Actuator Line (ACL)
approach instead of fully resolving them. To ensure a physically correct prediction of
the actual interference effects when using these models, Schollenberger, Firnhaber
Beckers and Lutz from the Institute of Aerodynamics and Gas Dynamics at the
University of Stuttgart provide in their report a comparison and validation with fully
resolved simulations and experimental data known from literature. With the input
data for the ACD and the ACL methods derived from steady-state 3D simulations
from a single blade, the authors can demonstrate that both approaches can predict the
interactions between propeller and wing in both directions with sufficient accuracy
making them well suited for design studies. The TAU-code of the German Aerospace
Center was applied for this study showing an ideal scaling behaviour between 50,000
and 10,000 grid points per core on Hawk.
The selection of projects presented reflects the continuing progress of high perfor-
mance computing in the field of CFD. Numerical simulations on supercomputers
like those at the HLRS and the SCC are crucial to gain insight into often complex,
time-dependent flow physics. Some of these flow phenomena occur only on very
small temporal and spatial scales and can only be brought to light by an extremely
fine discretization of the temporal and/or physical domains. The sustained increase
in computational power at both HPC centres, recently realised by the replacement
206 Computational Fluid Dynamics

of the Hazel Hen with the Hawk supercomputer at the HLRS, enables the tackling
of increasingly complex fluid dynamic problems, including unsteady and transient
processes, within reasonable timeframes. This also facilitates industrial design pro-
cesses, where the use of high-fidelity numerical methods on modern supercomputers
helps to enhance the reliability and the efficiency of the simulations, mitigate de-
velopment time, risks and costs and, thus, increase industrial competitiveness. But,
the rise in hardware performance has to be accompanied by the development of new
numerical algorithms. Furthermore, to exploit the full potential of the new hardware,
code adaptations to the particular HPC architecture are indispensable. A very close
cooperation of the researchers with the experts from the HLRS and the SCC is key to
successfully accomplishing this demanding task. In this context, the staffs of both
computing centres deserve thanks for their individual and custom-tailored support.
Without their dedication, it would not have been possible and will not be possible in
the future to maintain the high scientific quality we see in the projects.

Institute of Aerodynamics and Gas Dynamics, Ewald Krämer

University of Stuttgart,
Pfaffenwaldring 21,
70550 Stuttgart,
Germany,
e-mail: [email protected]
Analysis of the hot gas ingress into the wheel
space of an axial turbine stage

Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

1 Introduction

The thermal efficiency of gas turbines strongly depends on the turbine inlet temperature.
Increasing the inlet gas temperature improves the engine performance, but results
in new challenges for the cooling of the turbine material. To avoid hot gas ingress
from the main annulus flow into the wheel space between the stator and the rotor
disks, which can lead to machine failure, cooling air from the turbine’s secondary
air system is used to seal the rim seal gap and to cool the wheel space. To increase
the thermodynamic efficiency of the turbine, it is necessary to minimize the cooling
mass flow rate.
The rim seal flow has been extensively studied over the past years. A recent review
of the progress made is given in [8]. Recent experimental studies showed the existence
of periodic flow phenomena inside the rim seal cavity [3, 20]. The unsteady flow
structures were unrelated to blade passing and expected to have significant impact
on the sealing effectiveness. These unsteady phenomena were only present at low
cooling gas mass flow rates and vanished abruptly when the cooling gas mass flux was
sufficiently increased. The fluctuations were later confirmed by numerical simulations
based on the unsteady RANS approach of the same 1.5-stage turbine test rig [13].
Further studies showed that the unsteady flow phenomena, apart from the amount of
injected cooling gas, depend on the rotor speed [1] and the rim seal geometry [9].
Although several authors observed unsteady flow phenomena using the unsteady
RANS approach, it does not reliably predict the complex unsteady turbulent flow
field for each flow condition and rim seal geometry [10, 11]. To further understand

Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

Institute of Aerodynamics, RWTH Aachen University, Wüllnerstraße 5a, Aachen, 52062, Germany,
e-mail: [email protected],[email protected]
Wolfgang Schröder
JARA Center for Simulation and Data Science, RWTH Aachen University, Seffenter Weg 23, Aachen,
52074, Germany

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 207
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_12
208 Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

the inherent unsteady nature of the rim seal flow, necessary for the optimization and
development of new rim seal designs, more profound methods with fewer modeling
assumptions need to be used.
The work presented in this paper is a continuation of previous investigations, where
the flow in the axial turbine was predicted by LES for the full 360° circumference [12,
17]. Two rim seal geometries were considered in [17]. Kelvin–Helmholtz type
instabilities were identified in a single lip rim seal gap, which interacted with the main
annulus flow. Adding a second sealing lip damped these fluctuations and reduced
the hot gas ingress. In [12], one important finding was the occurrence of standing
acoustic waves inside the wheel space, where the frequencies coincide with acoustic
modes of the wheel space. The computational costs of such large scale LES are,
however, extremely large.
To reduce the computational costs a computationally less expensive zonal RAN-
S/LES method is developed, with which the flow field in the rim seal gap, dominated
by the shear layers of the main annulus flow and the cavity flow, can be resolved by
LES, while the flow field radially above the seal gap, i.e. in the hot gas flow, and
radially below the seal gap, i.e. in the wheel space, is determined using the RANS
approach.
The paper is organized as follows. First, the governing equations and the numerical
approach are discussed. Second, the results of [12] are briefly summarized. Third,
the zonal RANS/LES approach is presented and validated based on a reduced
computational setup. Finally, some conclusions are drawn.

2 Numerical method

2.1 Governing equations

The motion of a viscous, compressible fluid is described by the Navier–Stokes

equations. For an arbitrarily moving control volume 𝑉 with surface 𝐴, the conservation
equations for mass, momentum, and energy read
∫ ∮
d
= 𝑄 d𝑉 + 𝐻¯ · 𝑛 d𝐴 = 0. (1)
d𝑡 𝑉 (𝑡) 𝐴(𝑡)

Here, 𝑄 = [𝜌, 𝜌𝑢, 𝜌𝐸, 𝜌𝑌 ] 𝑇 is the vector of the conservative variables with
the density 𝜌, the velocity vector 𝑢, the total energy 𝜌𝐸 = 𝜌𝑒 + 𝜌𝑢 2 /2, and the
concentration of a passive scalar 𝑌 . The variable 𝑒 denotes the specific internal energy.
The quantity 𝐻¯ is the flux tensor and 𝑛 the outward normal vector on the surface 𝐴.
The flux tensor in non-dimensional form is given by
Analysis of the hot gas ingress into the wheel space of an axial turbine stage 209

© 𝜌 (𝑢 − 𝑢 𝜕𝑉 ) ª © 0 ª
­ ® ­ ®
­ 𝜌𝑢 (𝑢 − 𝑢 𝜕𝑉 ) + 𝑝 𝐼¯ ®
­ ® ­ ®
¯
𝜏
®− 1 ­
­ ®
𝐻¯ = 𝐻¯ 𝑖𝑛𝑣 − 𝐻¯ 𝑣𝑖𝑠 = ­­ ® 𝑅𝑒 0 ­
®.
® (2)
­ 𝜌𝐸 (𝑢 − 𝑢 𝜕𝑉 ) + 𝜌𝑢 ® ¯
­ 𝜏𝑢 + 𝑞 ®
­ ® ­ ®
­ ® ­ ®
𝜌𝑌 (𝑢 − 𝑢 𝜕𝑉 ) −𝜌𝐷∇𝑌 /𝑆𝑐 0
« ¬ « ¬
where 𝑢 𝜕𝑉 is the velocity of the control volume’s surface, 𝑝 is the pressure, and
¯ Additionally, 𝐷 is the mass diffusion coefficient of the
the unit tensor is given by 𝐼.
passive scalar 𝑌 , and 𝑆𝑐 0 = 1 is the Schmidt number. The Reynolds number and the
speed of sound are expressed by

√︁
𝑅𝑒 0 = 𝜌0 𝑎 0 𝑙𝑟 𝑒 𝑓 /𝜂0 , and 𝑎 0 = 𝛾 𝑝 0 /𝜌0 (3)

with the ratio of specific heats 𝛾 = 1.4 and the characteristic length 𝑙𝑟 𝑒 𝑓 . For a
Newtonian fluid with zero bulk viscosity, the stress tensor is written

  1+𝑆
𝜏¯ = 2/3 𝜂 (∇ · 𝑢) 𝐼¯ − 𝜂 ∇𝑢 + (∇𝑢) 𝑇 , where 𝜂(𝑇) = 𝑇 3/2 , (4)
𝑇 +𝑆
is determined from Sutherland’s law with 𝑆 = 111𝐾/𝑇0 . For constant Prandtl number
𝑃𝑟 0 = 0.72, the non-dimensional vector of heat conduction according to Fourier’s
law is
−𝜂
𝑞= ∇𝑇 . (5)
𝑅𝑒 0 𝑃𝑟 0 (𝛾 − 1)
The equations are closed by the equation of state for an ideal gas, which is written
in non-dimensional form 𝛾 𝑝 = 𝜌𝑇. On the adiabatic considered walls the no-slip
condition is imposed. The pressure at the solid boundaries is determined via a
robin-type boundary condition derived from the momentum equation.
The Reynolds averaged Navier–Stokes equations (RANS) are closed using the
Boussinesq hypothesis
!
′ ′ 𝜕𝑢 𝑖 𝜕𝑢 𝑗 2
− 𝑢 𝑖 𝑢 𝑗 = 𝜈𝑡 + − 𝑘𝛿𝑖 𝑗 , (6)
𝜕𝑥 𝑗 𝜕𝑥𝑖 3

where the turbulent viscosity 𝜈𝑡 is determined by the density corrected Spalart–

Allmaras turbulence model [7].

2.2 Numerical method

A coupled finite-volume and a level-set solver is used to simulate the flow field.
The solvers own subsets of a shared hierarchical Cartesian mesh, which are adapted
independently based to the movement of the blade surfaces.
210 Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

The conservation equations (1) are discretized by a cell-centered finite-volume

method. In a LES, only the large scale turbulent structures are resolved. The sub-grid
contribution is modeled here by the monotone integrated LES (MILES) approach [6],
in which the dissipative part of the truncation error dissipates the energy at the
unresolved scales.
A low-dissipation variant of the advection upstream splitting method (AUSM)
proposed [15] is used to compute the inviscid flux tensor 𝐻¯ 𝑖𝑛𝑣 (2). The primitive
variables at the cell surfaces are obtained by a second-order accurate MUSCL
extrapolation. The viscous flux tensor 𝐻¯ 𝑣𝑖𝑠 (2) is computed using a central-difference
scheme. The cell-centered gradients are computed by a weighted least-square method
as described in [21]. The gradients at the cell-surface centroids are determined with
the recentering-approach proposed by [2].
The time-integration is performed by an explicit second-order accurate 5-stage
Runge–Kutta scheme [21], which is optimized for an efficient integration for cases with
moving domain boundaries. The Runge–Kutta coefficients are chosen for maximum
stability.
A strictly conservative cut-cell method is used to accurately represent the surfaces
of embedded bodies, which are prescribed by a zero level-set contour obtained from
the level-set solver. To prevent arbitrarily small time steps, a flux redistribution
method is applied in small cut-cells. A detailed description of the cut-cell method
and the flux redistribution method is given in [21].
A level-set method is applied to track the movement of embedded bodies. The
level-set is a signed-distance function that divides the domain into a solid 𝐺 𝑠 and fluid
𝐺 𝑓 part. The surface of the body is given by the zero level-set contour Γ. To preserve
the resolution of the initial level-set surface while moving the bodies, the kinematic
motion level-set approach described by [18] is used. In this method a reference
level-set field is generated from a high-resolution stl at the start of the simulation and
used to interpolated the level-set at later time steps. Each time step the grid cell from
which the level-set was interpolated is stored and serves to efficiently find the correct
grid cell for interpolation in the following time step. In a previously implemented
method, the reference level-set remains unchanged throughout the entire simulation.
Hence, for parallel operations, the reference grid cell can be located on an arbitrary
MPI rank, and after a certain rotor rotation a global MPI communication is required
until a full rotor rotation is completed and the level-set arrives again at the reference
location. Since this global communication strongly hampers the performance for a
large number of MPI ranks, the previous method was modified. In the new version,
the reference level-set field is reconstructed from the stl each time the reference grid
cell moves onto a neighboring MPI rank. Before the reconstruction the coordinated
of the stl are rotated according to the current rotor location. For a large number of
MPI ranks, the cost for the level-set reconstruction are significantly lower compared
to the idle time of individual MPI ranks generated by the global communication.
Analysis of the hot gas ingress into the wheel space of an axial turbine stage 211

2.3 Zonal RANS/LES method

The zonal RANS/LES or embedded LES method is a hybrid turbulence modeling

method, where the governing equations for LES or RANS are solved in segregated
predefined computational domains. The domains are linked via an overlapping region,
in which the transition from RANS to LES and vice versa occurs. The aim of the
zonal RANS/LES method is to reduce the domain size of the computationally more
expensive LES approach to the flow region around the rim seal of the axial turbine.

2.3.1 Synthetic turbulence generation (STG)

To ensure a smooth transition between the RANS domain located upstream of the
rim seal gap and the downstream LES domain, the time averaged RANS values are
used to determine the solution at the LES inflow. The time averaged pressure of the
LES domain is prescribed at the outflow of the RANS domain to ensure the correct
propagation of information from the LES domain the upstream RANS domain. The
velocity fluctuations at the LES inflow are generated by the reformulated synthetic
turbulence generation (STGR) method [19]. This method is a synthetic eddy method
(SEM), which composes a turbulent velocity field by a superposition of eddies in a
virtual volume.

2.3.2 Reconstruction of the eddy viscosity (RTV)

The coupling of the LES domain around the rim seal with the RANS domain located
further downstream, requires a reconstruction of the eddy viscosity for the Spalart–
Allmaras model from the LES solution. Here, the methodology of [14] is used. In
addition, the time averaged pressure of the downstream located RANS domain is
used as an outflow condition for the embedded LES domain. The temporal average
of the flow quantities in the LES domain is calculated by a moving average. The
turbulent kinetic energy 𝑘 and the specific rate of dissipation 𝜔 are computed based
on Bradshaw’s hypothesis using the time averaged and instantaneous flow quantities
of the LES domain.

3 Computing resources

The LES for the full circumference, i.e., 360◦ of the axial turbine is performed on 64
compute nodes of the HAWK high-performance computer installed at HLRS. Each
node consists of 2 AMD EPYC™ 7742 CPUs. Due to the low flow velocity in the
wheel space, it takes about 40 full rotations until the turbulent flow in the wheel space
becomes fully developed. Depending on the operating conditions, one degree of rotor
rotation can take between 8 and 15 minutes computing time on 8192 compute cores.
212 Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

The LES for the 360◦ axial turbine is conducted using approx. 450 million grid cells.
Therefore, a snapshot of the flow variables requires approximately 58.5 GBytes of
disk space.
By using a solution adaptive mesh refinement, the number grid cells on the
individual MPI ranks changes considerably due to the rotation of the turbine blades.
To prevent load imbalances between the compute cores and to guarantee a high
parallel efficiency a dynamic load balancing method is utilized. During the load
balancing the parallel subdomains are redistributed among the compute cores [16].
Necessary communication between solvers is minimized by partitioning the shared
hierarchical Cartesian grid based on a space-filling curve. The load balancing is based
on the individual compute loads of all MPI ranks, measured for 5 time steps after
each mesh adaptation. A detailed description of the load balancing method can be
found in [16].

4 One-stage axial turbine

4.1 Computational setup

In this section, the setup of the one-stage axial flow turbine also investigated in [12]
is described. The geometry is shown in Figure 1. Note that the distance of the stator
and rotor rows are increased in axial direction for a better visualization. The turbine
stage consists of 30 stator vanes and 62 rotor blades. A double lip rim seal is used to
seal the rotor-stator wheel space from the main annulus. More details of this setup
can be found in [4, 5].
62 stator
30 stator
blades
blades
rotor
stator lip
lip

secondary
inflow

main
inflow

Fig. 1: Cut through the one-stage axial Fig. 2: Dimensions and computational
turbine with an enlarged distance between mesh in an axial plane of the axial
rotor and stator row. turbine stage.
Analysis of the hot gas ingress into the wheel space of an axial turbine stage 213

The full 360◦ circumference of the main annulus flow and wheel space is resolved
by an adaptive Cartesian mesh with approximately 450 million cells. The subset used
by the finite-volume solver comprises approximately 400 million leaf cells of the
shared grid, while the level-set solver uses about 450 million leaf cells. Figure 2
shows an axial cut through the mesh. A zoom of the rim seal geometry for the region
marked in Figure 2 by the dashed line is displayed in Figure 3.

x
3mm 10mm

A
Fig. 4: Operating conditions.
r/R=0.98
d

sc

sc CW1K CW2K
sc

d=4.5mm 𝜌1 𝑐1 𝑅
𝑅𝑒𝑐1 = 0.8 · 106 0.8 · 106
d

sc=1.7mm 𝜇1
s=16.7mm
s
𝑀𝑐1 = √ 𝑐1 0.37 0.37
stator rotor
𝛾𝑅𝑇1
𝜌 Ω𝑅 2
𝑅𝑒𝑢 = 𝑐𝑔𝜇𝑐𝑔 0.8 · 106 0.8 · 106
x/sc=-8.7 x/sc=0.5
𝑚¤
𝑐𝑤 = 𝜇𝑐𝑔𝑐𝑔𝑅 1000 2000
Fig. 3: Schematic view of the rim seal
geometry.

The operating conditions are defined by the four non-dimensional parameters

listed in Table 4. The subscript 1 indicates the flow state 1.5 mm downstream of the
stator blades and cg denotes the conditions inside the cooling gas inlet. Two cases
with the same flow conditions in the main annulus, i.e., Reynolds and Mach number
𝑅𝑒 𝑐1 , and 𝑀𝑐1 ) and rotor speed 𝑅𝑒 𝑢 were conducted. The two cases differ in the
non-dimensional cooling gas mass flux 𝑐𝑤, which is reduced by 50% in case CW1K
compared to case CW2K.
At the main and secondary inflow boundaries the density and the three velocity
components are prescribed. At the outflow boundary, the static pressure is fixed.
To reduce numerical wave reflections sponge layers are used at all in- and outflow
boundaries. At the fixed and the moving walls, an adiabatic no-slip condition is
prescribed.

4.1.1 Reduced computational setup for zonal RANS/LES method

A simplified two-dimensional generic turbine setup (GT) without rotor blades based
on the geometry discussed in section 4.1 is used to test the zonal RANS/LES method.
Periodic boundary conditions in the 𝑧-direction are used. The investigations are
focused on the transition from RANS to LES domains especially for the shear layers.
The simplified turbine setup is shown Figure 5.
214 Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

Fig. 5: Schematic of the generic turbine geometry.

4.2 Results

First, the numerical results presented in [12] for the analysis of the flow field in a
one-stage axial flow turbine for the two cases with different cooling gas mass flow
rates are briefly summarized. The data shown for case CW2K is taken from [17]. For
both operating conditions, it takes approximately 40 full rotor rotations until the flow
field is fully developed. For the statistical analyses the data of the final 3 rotations are
used. The turbulent flow field in the stator hub boundary layer is shown in Figure 6 by
the instantaneous contours of the q-criterion. The contours a colored by the absolute
Mach number. The position of the rim seal gap is indicated by the black cylinder

Fig. 6: Instantaneous contours of the q-criterion colored by the absolute Mach number

between the vanes and the blades.

Figure 7 displays the radial distributions of the rms values of the velocity fluctua-
tions in the wheel space at the fixed azimuthal position of Θ = 0◦ . The distributions
in two axial positions are shown, i.e., near the stator wall at 𝑥/𝑠 𝑐 = −8.7 and near the
rotor wall at 𝑥/𝑠 𝑐 = 0.5, see Figure 3. While the rms values of the azimuthal and the
radial velocity components have somewhat distributions, the pressure fluctuations
Analysis of the hot gas ingress into the wheel space of an axial turbine stage 215

show large differences for the two cooling mass flow rates. For case CW2K, the
pressure rms values are almost constant along the radius. For case CW1K, fluctuations
of significantly higher amplitude are visible, which exceed the rms values of case
CW2K by more than a factor of two. To identify the reason for the increased pressure
fluctuations, the instantaneous flow field inside the wheel space is analyzed. The
pressure signals in the wheel space of the 3 full rotor rotations were sampled in 408
radially distributed bins along the stator wall. For these signals cross-correlations are
computed relative to the signal at the location 𝑟/𝑅 = 0.98.

x/sc=-8.7 x/sc=-8.7 x/sc=-8.7

1.0 1.0 1.0

0.8 0.8 0.8

r/R

r/R

r/R
0.6 0.6 0.6

0.4 0.4 0.4

0 0.05 0.1 0 0.05 0.1 0 2 4 6 8 -310

vθ,rms/ω*R vr,rms/ω*R prms / pgap x10
x/sc=0.5 x/sc=0.5 x/sc=0.5
1.0 1.0 1.0

0.8 0.8 0.8

r/R

r/R

r/R

0.6 0.6 0.6

0.4 0.4 0.4

0 0.05 0.1 0 0.05 0.1 0 2 4 6 8 -310

vθ,rms/ω*R vr,rms/ω*R prms / pgap x10

Fig. 7: Radial distributions of the azimuthal and radial velocity components 𝑣 𝜃 and
𝑣 𝑟 and the pressure fluctuations at two axial positions; – CW1K, ... CW2K.

The energy spectra computed from these cross-correlations, shown in Figure 8,

reveal a distinct peak at a frequency of approximately 94.33 times the rotor speed 𝑛
for case CW1K, which is seen along the entire stator wall [12]. The energy spectrum
for case CW2K only shows a minor peak at the blade passage frequency (BPF) of
𝑓 /𝑛 = 62.
This dominant frequency for case CW1K is also found in the instantaneous flow
field in the rim seal gap at 𝑟/𝑅 = 0.98. Since the radial velocity in the rim seal
gap determines the hot gas ingress, the signals of the effective radial velocity were
recorded along the entire circumference in 1440 equidistantly spaced positions. Using
the same procedure as for the energy spectra in Figure 8, the energy spectra of
the effective radial velocity fluctuations in the rim seal gap were obtained. The
energy spectrum for case CW1K in Figure 9 shows the same dominant frequency
of 𝑓 /𝑛 ≈ 94.33 as the pressure fluctuations for case CW1K in Figure 8 along with
additional peaks. Below the dominant frequency 𝑓 /𝑛 ≈ 94.33 only a broadband of
lower amplitudes is observed. This is also apparent for case CW2K, where no distinct
frequency can be observed for the effective radial velocity.
216 Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

CW1K; r/R=0.6
2
1.5

Ep’p’
1
0.5
0
0 50 100 150 200 250 300
f/n

CW2K; r/R=0.6
2
1.5 BPF
Ep’p’

1
0.5
0
0 50 100 150 200 250 300
f/n

Fig. 8: Energy spectra computed from cross-correlations of the pressure fluctuations

along the stator wall in the wheel space; CW1K (top), CW2K (bottom); x/sc =-8.7.

CW1K; θ=180°
0.4
0.3
Ev’v’

0.2
0.1
0
0 50 100 150 200 250 300
f/n

CW2K; θ=180°
0.4
0.3
Ev’v’

0.2
0.1
0
0 50 100 150 200 250 300
f/n

Fig. 9: Energy spectra computed from cross-correlations of the effective radial velocity
fluctuations in the rim seal gap; CW1K (top), CW2K (bottom); r/R=0.98.

To explain the origin of the peaks observed in Figure 9, the frequencies are
compared to the theoretical acoustic eigenfrequencies of a closed pipe. These
frequencies are defined by
𝑚
𝑓 = · 𝑎, (7)
2·𝐿
where 𝑎 is the speed of sound, 𝐿 the length of the resonator, and 𝑚 the order of
the harmonic. The speed of sound in the wheel space is almost constant and has
the value of approximately 𝑎/𝜔𝑅 = 2.853. Using the height of the wheel space
𝐿 = 𝑅 − (2 · 𝑑 + 𝑠 𝑐 ) − 𝑟 ℎ = 86.65 mm the harmonics in Table 1 can be determined.
Analysis of the hot gas ingress into the wheel space of an axial turbine stage 217

Table 1: Harmonics of a closed pipe normalized by rotor speed f/n.

m 7 8 11 12 13 16 17 18 21 24
Theory 84.17 96.20 132.27 144.29 156.32 192.39 204.42 216.44 252.51 288.59
LES 79.53 94.33 126.6 141.5 156.3 188.7 203.5 218.3 250.6 283.1
rel. Error -5.5 % -1.9 % 4.2 % -1.9 % -0.01 % -1.9 % -0.5 % 0.9 % -0.8 % -1.9 %

Despite the strong simplifications introduced by the 1D model, the comparison

of the LES data for CW1K from Figure 9 with the theoretical finding shows a good
match. Especially, the 8th harmonic is very close to 𝑓 /𝑛 ≈ 94.33. The strongest
deviations between the theoretical values and the LES results occur for the frequencies
close to a multiple of the blade passage frequency. For case CW1K several of the
wheel spaces’ harmonics are excited, which results in the development of a standing
wave inside the wheel space.
Filtering the flow field of the effective radial velocity inside the rim seal gap at the
frequency of the dominant 8th harmonic at 𝑓 /𝑛 = 94.33, shows how the fluctuations
can influence the hot gas ingress, see Figure 10. A square pattern consisting of

Fig. 10: Filtered effective radial velocity ṽr /𝜔R inside the rim seal gap at r/R = 0.98;
CW1K.

30 fields in the azimuthal direction and in time is evident for the filtered effective
radial velocity 𝑣˜ 𝑟 , which correspond to the 30 stator blades. The temporal fluctuation
repeats approximately every 3.82◦ rotor rotation and corresponds to the standing
wave inside the wheel space. This wave leads to alternating positive and negative
values of the effective radial velocity.
The radial distribution of the time and circumferentially averaged cooling effec-
tiveness for cases CW1K and CW2K is shown in Figure 11. The cooling effectiveness
is computed by
218 Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

100
80

η [%]
60
40
20
0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
r/R

Fig. 11: Radial distribution of the time and azimuthally averaged cooling effectiveness;
LES: CW1K (−), CW2K (− − −), exp. data [4]: CW1K (•), CW2K (■).

𝑌 (𝑟) − 𝑌ℎ𝑔
𝜂= , (8)
𝑌𝑐𝑔 − 𝑌ℎ𝑔
where 𝑌 is the concentration of a tracer gas mixed into the cooling gas. The
concentration is obtained from a passive scalar transport equation. The subscripts hg
and cg denote the concentrations at the main flow and the cooling gas inlet. In both
cases the LES results match the experimental data from [4] convincingly. The fact
that the cooling effectiveness is reduced for case CW1K compared to CW2K shows
the importance of accurately resolving the instantaneous flow field.

4.3 Reduced setup

In the following, the comparison of the flow field in the wheel space for the GT setup
is discussed for a pure LES and the zonal RANS/LES method. Figure 13 shows
the comparison of the instantaneous tracer gas concentration in the wheel space in
the axial cut at 𝑥/𝑠 = 0.875, a cross-section in the periodic direction and a radial
cross-section inside the boundary layer 𝑟/𝑅 = 1.0043. The position in periodic
direction is chosen at the trailing edge of the stator. The instantaneous cooling gas
concentration of the zonal RANS/LES setup shows a good agreement with the
instantaneous cooling gas concentration of the pure LES.
Figure 13 shows the radial distributions of the cooling effectiveness 𝜂 and the
components of the Reynolds stress tensor 𝑢 ′𝑢 ′ and 𝑣 ′ 𝑣 ′ at the axial position 𝑥/𝑠 = 0.98.
The zonal RANS/LES setup shows a good agreement in the radial distributions of the
cooling effectiveness 𝜂 compared to the pure LES setup validating the described STG
for shear layer flows. A good quantitative agreement of the relevant components of
the Reynolds stress tensor 𝑢 ′𝑢 ′ and 𝑣 ′ 𝑣 ′ is shown in Figure 13. The STG is therefore
capable of generating a physically meaningful turbulent energy spectrum from the
RANS domains reducing the extent of the scale resolving LES to a region of interest
whilst reducing the computing time for the zonal RANS/LES method by reducing the
number of grid cells.
Analysis of the hot gas ingress into the wheel space of an axial turbine stage 219

Fig. 12: Distribution of the instantaneous cooling gas concentration inside the wheel
space of the (left) pure LES setup and (right) zonal RANS/LES setup for the GT
geometry.

Fig. 13: Radial distributions of the cooling effectiveness 𝜂 and the components of the
Reynolds stress tensor 𝑢 ′𝑢 ′ and 𝑣 ′ 𝑣 ′ at the axial position 𝑥/𝑠 = 0.98

5 Conclusion

The flow field in an axial turbine stage was predicted by LES for two cooling gas
mass flow rates. After about 40 full rotor rotations a statistically converged flow field
in the wheel space was obtained. The analysis showed that the instantaneous results
strongly depend on the cooling gas mass flow rate. For the lower cooling gas mass
flux (case CW1K), several of the wheel space’s harmonics are excited, which generate
oscillations of the radial velocity component in the rim seal gap. Especially at the
lower frequent oscillations, cooling gas is ejected out of the wheel space, followed
by an injection of hotter gas from the main annulus. This leads to a reduced cooling
effectiveness compared to the case CW2K.
220 Jannik Borgelt, Thomas Hösgen, Matthias Meinke and Wolfgang Schröder

The presented zonal RANS/LES method employs a synthetic eddy method

generating the turbulent fluctuations at the inflow and a reconstruction of the turbulent
viscosity at the outflow of the embedded LES domain. Results were obtained for
a generic turbine setup and compared to a pure LES setup. The zonal RANS/LES
method showed good agreement with the pure LES setup in the instantaneous cooling
gas concentration, the time and spanwise averaged radial distributions of the cooling
gas concentration, and the components of the Reynolds-stress tensor. This zonal
RANS/LES method will be used in future work to predict the hot gas ingress in axial
turbines more efficiently.

References

1. P. F. Beard, F. Gao, K. S. Chana, and J. Chew. Unsteady flow phenomena in turbine rim seals.
Journal of Engineering for Gas Turbines and Power, 139(3), 2016.
2. M. Berger and M. Aftosmis. Progress towards a cartesian cut-cell method for viscous
compressible flow. In 50th AIAA Aerospace Sciences Meeting including the New Horizons
Forum and Aerospace Exposition, 2012.
3. D. Bohn, A. Decker, H. Ma, and M. Wolff. Influence of sealing air mass flow on the velocity
distribution in and inside the rim seal of the upstream cavity of a 1.5-stage turbine. In Volume 5:
Turbo Expo 2003, Parts A and B, Turbo Expo: Power for Land, Sea, and Air, pages 1033–1040,
2003.
4. D. Bohn and M. Wolff. Entwicklung von berechnungsansätzen zur optimierung von sperrgassys-
temen für rotor/stator-kavitäten in gasturbinen, 2001.
5. D. Bohn and M. Wolff. Improved formulation to determine minimum sealing flow – cw, min –
for different sealing configurations. In Volume 5: Turbo Expo 2003, Parts A and B, Turbo Expo:
Power for Land, Sea, and Air, pages 1041–1049, 2003.
6. J. P. Boris, F. F. Grinstein, E. S. Oran, and R. L. Kolbe. New insights into large eddy simulation.
Fluid Dynamics Research, 10(4):199–228, 1992.
7. S. Catris and B Aupoix. Embedded les-to-rans boundary in zonal simulations. Aerosp. Sc.
Technol, 4:1–11, 2000.
8. J. W. Chew, F. Gao, and D. M. Palermo. Flow mechanisms in axial turbine rim sealing. Journal
of Mechanical Engineering Science, 233(23-24):7637–7657, 2019.
9. M. Chilla, H. Hodson, and D. Newman. Unsteady interaction between annulus and turbine rim
seal flows. Journal of Turbomachinery, 135(5), 2013.
10. J. T. M. Horwood, F. P. Hualca, J. A. Scobie, M. Wilson, C. M. Sangan, and G. D. Lock.
Experimental and computational investigation of flow instabilities in turbine rim seals. Journal
of Engineering for Gas Turbines and Power, 141(1), 10 2018.
11. J. T. M. Horwood, F. P. Hualca, M. Wilson, J. A. Scobie, C. M. Sangan, G. D. Lock, J. Dahlqvist,
and J. Fridh. Flow instabilities in gas turbine chute seals. Journal of Engineering for Gas
Turbines and Power, 142(2), 01 2020.
12. T. Hösgen, M. Meinke, and W. Schröder. Large-eddy simulations of rim seal flow in a one-stage
axial turbine. Journal of the Global Power and Propulsion Society, 4:309–321, 2020.
13. R. Jakoby, T. Zierer, K. Lindblad, J. Larsson, L. deVito, Dieter E. Bohn, J. Funcke, and A. Decker.
Numerical simulation of the unsteady flow field in an axial gas turbine rim seal configuration.
In Volume 4: Turbo Expo 2004, Turbo Expo: Power for Land, Sea, and Air, pages 431–440,
2004.
14. D. König, M. Meinke, and W. Schröder. Embedded les-to-rans boundary in zonal simulations.
Journal of Turbulence, 67:1–25, 2010.
15. M. Meinke, W. Schröder, E. Krause, and Th. Rister. A comparison of second- and sixth-order
methods for large-eddy simulations. Computers & Fluids, 31(4):695–718, 2002.
Analysis of the hot gas ingress into the wheel space of an axial turbine stage 221

16. A. Niemöller, M. Schlottke-Lakemper, M. Meinke, and W. Schröder. Dynamic load balancing

for direct-coupled multiphysics simulations. Computers and Fluids, 199(104437), 2020.
17. A. Pogorelov, M. Meinke, and W. Schröder. Large-eddy simulation of the unsteady full 3d rim
seal flow in a one-stage axial-flow turbine. Flow, Turbulence and Combustion, 102(1):189–220,
2019.
18. A. Pogorelov, L. Schneiders, M. Meinke, and W. Schröder. An adaptive cartesian mesh
based method to simulate turbulent flows of multiple rotating surfaces. Flow, Turbulence and
Combustion, 100(1):19–38, 2018.
19. B. Roidl, M. Meinke, and W. Schröder. A zonal rans-les method for compressible flows.
Computers and Fluids, 67:1–15, 2015.
20. Bernd Rudzinski. Experimentelle Untersuchung des Heißgaseinzuges in die Rotor-Stator-
Zwischenräume einer eineinhalbstufigen Turbine für unterschiedliche Dichtkonfigurationen:
Zugl.: Aachen, Techn. Hochsch., Diss., 2009. Verlag Dr. Hut, München, 2009.
21. L. Schneiders, C. Günther, M. Meinke, and W. Schröder. An efficient conservative cut-cell
method for rigid bodies interacting with viscous compressible flows. Journal of Computational
Physics, 311:62–86, 2016.
Direct numerical simulation of bypass transition
under free-stream turbulence for compressible
flows

Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

Abstract Direct numerical simulations (DNS) of bypass transition of a flat-plate

boundary-layer subject to free-stream turbulence (FST) were performed. A synthetic
turbulent inflow is generated by superposition of modes of the continuous spectrum
resulting from linear stability theory (LST). The present article briefly explains the
adaption of the method for a fully compressible framework, including the generation
of turbulence as well as the treatment of the inflow boundary conditions. The results
of a flat-plate flow at three different integral turbulent length scales were compared
and validated with results from the literature. In addition, a performance analysis of
the codes and methods used was performed on the HLRS Hawk supercomputer.

1 Introduction

Fluid flows above a critical ratio of inertia to viscous forces Re = 𝑈˜ ∞ 𝑙˜char /𝜈˜ usually
undergo a transition from laminar to turbulent state when triggered by suitable
perturbations. Even though there is a wide range of low-order models to mimic
the effects of this transition it is still often necessary to perform highly resolved
numerical simulations of both the onset and the transition process itself. High-
fidelity simulations allow on the one hand to gain a deeper understanding of the
physical details of the transient and non-linear interactions and on the other hand they
enable the development of improved models to predict the impact on industrial-scale
applications. This is even more true for high-speed flow where not only increased
momentum transfer but also heat flux and a change in thermodynamical properties
result from the laminar-turbulent transition.

D. Ohno, B. Selent, M. Kloker and U. Rist

Institut für Aero- und Gasdynamik (IAG), Universität Stuttgart, Pfaffenwaldring 21, D-70550
Stuttgart, Germany, e-mail: {ohno,selent,kloker,rist}@iag.uni-stuttgart.de

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 223
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_13
224 Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

A common method for generating free-stream turbulence in DNS is the utilization

of multiple Fourier-modes scaled to a turbulent spectrum as inflow boundary condition.
However, this method is only applicable when the computational domain covers the
flow field well ahead of the aerodynamic body. Extending the DNS to the leading
edge of a flat-plate or an airfoil results in an increase in computational expenses and
may lead to numerical instabilities.
Jacobs & Durbin [1] described for the first time a method for generating artificial
free-stream turbulence, while taking the wall boundary condition at the inflow of the
simulation domain into account. In this approach, modes of the continuous spectrum
resulting from linear stability analysis are used instead of Fourier-modes. These
modes are a solution of the underlying linear stability problem of the laminar baseflow
including the boundary-layer. Introducing these modes with small amplitudes at the
inflow boundary satisfies the conservation equations, since the modes are a solution
of the linearized Navier-Stokes equations. In recent years, intensive studies were
carried out using this method, including investigations on 2D flows with pressure
gradient by Zaki & Durbin [2], 3D flows by Schrader et al. [3], corner flows by
Schmidt & Rist [4] or in an airfoil flow by Ohno et al. [5].
However, most of the previous studies were conducted for incompressible flow
problems. The present paper presents an adaption of this method for compressible
flows, using results from Brandt et al. [6] for validation.

2 Numerical methods

Direct numerical simulation

All simulations are conducted with a revised version of the high-order in-house
DNS code NS3D [7] which solves the three-dimensional compressible Navier-Stokes
equations in cartesian coordinates. Spatial derivatives are discretized with 8th-order
explicit finite differences, while an explicit 4th-order 4-step Runge–Kutta scheme is
used for time integration.
The solution vector consists of the numerical fluxes, i.e. Q = [𝜌, 𝜌𝑢, 𝜌𝑣, 𝜌𝑤, 𝐸] 𝑇 .
Here, 𝜌 is the density, 𝑢, 𝑣, 𝑤 are the velocity components in 𝑥-𝑦-𝑧 coordinates—
representing streamwise, wall-normal and spanwise directions, respectively—and 𝐸 is
the total energy per volume. Furthermore, the thermodynamic variables temperature
𝑇 and pressure 𝑝 are derived from the solution vector. Dimensional variables are
denoted with •˜ . The quantities in both, DNS and the linear stability solver (including
frequencies and wavenumbers), are nondimensionalized using free-stream values 𝑈˜ ∞ ,
𝜌˜ ∞ , 𝑇˜∞ and a characteristic length 𝑙˜char .
DNS of bypass transition under free-stream turbulence for compressible flows 225

Generation of free-stream turbulence

The continuous modes were calculated using an in-house solver for the compressible
Navier-Stokes equations linearized about a steady baseflow [8] at the inflow location
of the DNS domain. The resulting linear system is solved directly by LAPACK [9]
routines for eigenvalue problems.
The complex solution vector q̂ = ( 𝜌, ˆ 𝑣ˆ , 𝑤,
ˆ 𝑢, ˆ 𝑇 contains the 1D eigenfunctions
ˆ 𝑇)
of the flow field variables in 𝑦-direction of each mode. While 𝑢ˆ = 𝑣ˆ = 𝑤ˆ = 𝑇ˆ = 0 is
applied at the wall, a homogeneous Neumann boundary condition 𝜕 q̂/𝜕𝑦 = 0 is used
at the free-stream. Examples for the shape of these eigenfunctions are depicted in
Fig. 2(a). Evidently, the development of these modes is periodic in the free-stream,
but rapidly dampened in the boundary-layer due to shear sheltering [10].
In order to create a broad spectrum of disturbances, multiple modes need to be
selected from the wavenumber domain 𝛼-𝛾-𝛽, representing the wavenumbers in
the spatial directions 𝑥, 𝑦 and 𝑧, respectively. The free-stream modes move with
the free-stream velocity 𝑐 = 𝑈∞ = 𝜔/𝛼. Considering Taylor’s hypothesis of frozen
turbulence, the streamwise wavenumber is therefore connected to the frequency
𝛼𝑟 ≈ 𝜔𝑟 for both spatial and temporal analysis in the non-dimensional formulation
with 𝑈∞ = 1. The spanwise wavenumber 𝛽 is always a real number and can be
considered as a parameter. The corresponding (complex) wall-normal wavenumber 𝛾
can be determined using the dispersion relation

𝛾 2 = 𝑖Re 𝛿 ∗ · (𝛼𝑈∞ + 𝛽𝑊∞ − 𝜔) − 𝛼2 − 𝛽2 (1)

which is given in Mack [8] or Schrader et al. [3]. Using temporal theory—where 𝛼 is
a real number—and assuming 𝜔𝑟 = 𝛼 for free stream modes, the theoretical complex
eigenvalue
𝑖
𝜔=𝛼− (𝛼2 + 𝛽2 + 𝛾𝑟2 − 𝛾𝑖2 ) (2)
Re 𝛿 ∗
with
Re 𝛿 ∗
𝛾𝑖 = (𝛼(1 − 𝑈∞ ) − 𝛽𝑊∞ ) (3)
2𝛾𝑟
can be calculated. For 2D-baseflows with 𝑈∞ = 1 and 𝑊∞ = 0, 𝛾𝑖 = 0 applies,
simplifying Eq. (2). With this relation, the complex eigenvalue 𝜔 of the required
mode 𝑚 can be searched for in the spectrum of the stability solution at 𝛼𝑚 and 𝛽𝑚
with the desired wall-normal wavenumber 𝛾𝑚 .
The disturbances generated by superposition of individual modes can be described
with the modal ansatz
𝑀
∑︁
q′ (𝑥, 𝑦, 𝑧, 𝑡) = 𝐴𝑚 q̂𝑚 (𝑦) · e𝑖 ( 𝛼𝑚 𝑥+𝛽𝑚 𝑧−𝜔𝑚 𝑡) (4)
𝑚=1

where 𝐴𝑚 represents the amplitude coefficient. While the wall-normal wavenumber

𝛾𝑚 is determined by the eigenfunction q̂𝑚 (𝑦), a Fourier ansatz is chosen for the
disturbances in spanwise direction with the wavenumber 𝛽𝑚 thus satisfying period-
226 Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

icity. Aforementioned Taylor’s hypothesis of frozen turbulence allows streamwise

wavenumbers to be set via frequency 𝛼𝑚 ≈ 𝜔 𝑚 . In cases where the modes are
prescribed in more than one point of the flow direction 𝑥, it is necessary to trans-
form the temporal amplification rate 𝜔𝑖 to the spatial one 𝛼𝑖 by applying Gaster’s
transformation [11]
temporal
𝜔𝑖 𝜕𝜔𝑟
spatial
=− . (5)
𝛼 𝜕𝛼𝑟
𝑖
This transformation guarantees the correct growth of the modes within the forcing
region.

(a) (b)
3 10 2

2 k0
γ

1
-5/3
E(k)

10 1 ~k
0
3
2
1
0 3
β -1 2
-2 1 α 10 0
-3 0 1 2 3
k
Fig. 1: (a) Selection of free-stream modes in the 𝛼-𝛾-𝛽 wavenumber domain on 20
spheres (only 10 depicted) with identical wavenumber magnitudes. (b) Exemplary
von Kármán energy spectrum; the columns represent the energy of each sphere.

To ensure the isotropy of the free-stream turbulence, the wavenumber vector

k𝑚 = [𝛼𝑚 , 𝛾𝑚 , 𝛽𝑚 ] 𝑇 must fulfill certain constraints. Similar to the approach of
Brandt et al. [6], free-stream modes are distributed in the wavenumber√︁domain 𝛼-𝛾-𝛽
over concentric spheres with identical wavenumber magnitudes 𝑘 = 𝛼2 + 𝛾 2 + 𝛽2 .
This is achieved by random rotation of a dodecahedron in the wavenumber domain,
where the 20 vertices represent the wavenumber vectors k𝑚 . An example with 20
spheres, including the randomly distributed wavenumbers, is depicted in Fig. 1(a).
For the wall-normal and streamwise wavenumbers, the condition 𝛼 > 0 and 𝛾 > 0
applies, thus a limitation is applied.
Finally, the amplitudes of all modes 𝐴𝑚 in Eq. (4) need to be defined with the
spectral energy distribution 𝐸 (𝑘). In most investigations known from literature, the
von Kármán spectrum is applied, which is defined as

2 𝑎(𝑘 𝐿) 4
𝐸 (𝑘) = 𝐿𝑞. (6)
3 (𝑏 + (𝑘 𝐿) 2 ) 17/6
DNS of bypass transition under free-stream turbulence for compressible flows 227

Here, the constants 𝑎 = 1.606 and 𝑏 = 1.350 are used and the turbulent kinetic energy
is defined as 𝑞 = 23 Tu2 with the turbulence intensity Tu. The parameter 𝐿 denotes the
integral length scale. The amplitude coefficient is given by
Δ𝑘
𝐴2𝑚 (𝑘) = 𝐸 (𝑘) , (7)
𝑁shell
where Δ𝑘 denotes the difference in wavenumber between two contiguous shells and
𝑁shell the number of all spheres. An example of a von Kármán spectrum is depicted
in Fig. 1(b). The wavenumber of maximum energy 𝑘 0 , which can be calculated
according to Tennekes & Lumley [12] with the integral length scale 𝑘 0 = 1.8𝐿 , is
clearly visible.

Modified boundary conditions

In all simulations the wall boundary condition was chosen to be adiabatic, and periodic
boundary conditions were used for the spanwise direction. For compressible subsonic
flows, characteristic boundary conditions at the inflow, outflow, and free-stream are
advantageous, since acoustic reflections can be prevented, see Giles [14]. These
boundary conditions are effective in one spatial direction as well as in time and allow
harmonic perturbations to leave the domain while still prescribing the baseflow. In
the proximity of the boundary, the disturbance flow field

𝜙 𝑐′ = 𝜙 − 𝜙ref (8)

is transformed to characteristic variables based on a one-dimensional decomposition

of the Euler equations into upstream and downstream travelling waves. For an inflow
boundary on the left side of the domain, the characteristic variables can be calculated
with
©𝑐 1 ª ©−𝑎 2
­ ® ­ ref 0 0 0 1ª® ©­ 𝜌 𝑐′ ª®
­ ® ­ ® ­ ®
­𝑐 2 ® ­ 0
­ ® ­ 0 𝜌ref · 𝑎 ref 0 0®® ­­ 𝑢 𝑐′ ®®
­ ® ­ ® ­ ®
­𝑐 ® = ­ 0
­ 3® ­ 0 0 𝜌ref · 𝑎 ref 0®® · ­­ 𝑣 𝑐′ ®® (9)
­ ® ­ ® ­ ®
­𝑐 4 ® ­ 0 𝜌ref · 𝑎 ref 1® ­𝑤 𝑐′ ®
­ ® ­ ® ­ ®
0 0
­ ® ­ ® ­ ®
® ­ ′®
0 −𝜌ref · 𝑎 ref
­ ® ­
𝑐5 0 0 1 𝑝𝑐
« ¬ « ¬ « ¬
where 𝑎 represents the speed of sound. In case of steady-state boundary conditions, all
variables with the subscript ref are set to baseflow values 𝜙ref = 𝜙bf . The characteristic
variables represent entropy perturbations (𝑐 1 ), vorticity perturbations in the spanwise
and streamwise directions (𝑐 2 , 𝑐 3 ), as well as downstream and upstream travelling
sound waves (𝑐 4 , 𝑐 5 ). The boundary condition sets the incoming disturbances (𝑐 1 -𝑐 4 )
228 Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

to zero and extrapolates the characteristic variable 𝑐 5 of the upstream travelling

acoustic wave. In a final step, the primitive variables are computed by an inversion of
Eq. (9) and added to the current flow field.
For simulations of bypass transition due to free-stream turbulence, the characteristic
boundary conditions described above are suitable for the free-stream as well as for
the outflow. However, for the inflow condition, they must be manipulated in order
to allow the introduction of unsteady disturbances into the domain. In this case, the
calculated superposition of all modes from Eq. (4) is added to the baseflow

𝜙ref (𝑡) = 𝜙bf + 𝜙 ′ (𝑡) (10)

at every (sub)iteration and used for the characteristic boundary condition treatment.
This allows disturbances to be introduced into the domain, while at the same time
allowing upstream acoustic waves to exit the domain.
An additional approach to reduce reflections and improve numerical stability is to
apply sponge zones in front of the boundary conditions. Using equation

𝜕Q 𝜕Q
= − 𝐺 (x) · (Q − Qref ), (11)
𝜕𝑡 𝜕𝑡 NS

the time derivative of the conservative numerical fluxes Q of the unsteady solution is
forced to the reference state, with a gain function 𝐺 (x) specifying the magnitude and
spatial distribution. Typically, this function describes the fading from a maximum
value of the gain 𝐺 at the boundaries to the null parameter 𝐺 = 0 in the inner
evaluable computational domain. For simulations like presented in this paper it is
essential to use such damping zones in the free stream as well as at the outflow to
avoid distortions due to the boundary conditions. In those areas, the values of the
steady baseflow 𝜙bf are used to calculate the conservative reference field Qref in order
to reduce disturbances.
However, when using a sponge zone at the inflow, the introduced disturbances
must again be taken into account. Analogously to the treatment of the unsteady
characteristic boundary conditions, see equation Eq. (8), a primitive variable field can
be generated at each time step at the inflow region by adding the perturbation field
from Eq. (4) to the baseflow. This field is converted into a conservative field 𝑄 ref (𝑡)
at each time step and considered in Eq. (11). With this method, the sponge acts
simultaneously as an unsteady forcing region—which also contains the attenuation
rates of the modes via the Gaster transformation—as well as a dampening zone for
outgoing waves. This approach has also been used successfully for an inflow with a
turbulent boundary-layer flow, see Appelbaum et al. [15].
DNS of bypass transition under free-stream turbulence for compressible flows 229

3 Numerical simulations

Setup

For further validation and further investigations, a bypass transition scenario in a

Blasius boundary-layer flow—simulated by Brandt et al. [6]—has been employed.
The first three cases presented in this publication, ‘Case1’, ‘Case2’ and ‘Case3’, were
reproduced. The cases differ only in the spectrum of the free-stream turbulence (see
Tab. 1), but not in the domain dimensions.
The Blasius boundary-layer displacement thickness at the beginning of the
computational domain 𝛿˜0∗ determines the characteristic length 𝑙˜char for all presented
simulations. In the aforementioned paper, a domain size of 101.3𝛿0∗ ≤ 𝑥 ≤ 1101.3𝛿0∗ ,
0.0 ≤ 𝑦 ≤ 100.0𝛿0∗ and 0.0 ≤ 𝑧 ≤ 90.0𝛿0∗ was employed, with the Reynolds number
Re 𝛿0∗ = 300 at the inflow. Therefore, 𝑥 = 0 represents the theoretical location of the
flat-plate leading edge.

Table 1: Overview of performed simulations with parameters used to define FST

according to Brandt et al. [6]; FTT: flow-through time.

Free-stream turbulence Simulation properties

Ma Label 𝐿/ 𝛿0∗ Tu FTT total FTT stat. Iter. total CPU-h

0.3 Case1 5.0 4.7% 12.4 9.9 7270k 777.0k

0.3 Case2 2.5 4.7% 4.9 2.3 2840k 303.5k
0.3 Case3 7.5 4.7% 12.4 9.9 7270k 777.0k

0.7 Case1 5.0 4.7% 18.3 13.5 3060k 327.1k

0.7 Case2 2.5 4.7% 13.8 8.9 2300k 245.8k
0.7 Case3 7.5 4.7% 18.3 13.5 3060k 327.1k

However, in the present investigations a prolonged computational domain of

101.3𝛿0∗ ≤ 𝑥 ≤ 1450.6𝛿0∗ has been used in order to reduce the influence of the outflow
boundary and to allow an extended sponge zone. In contrast to the incompressible
simulations of the reference, a Mach number must be specified for the compressible
simulations with NS3D. In the investigations, two Mach numbers, Ma = 0.3 and
Ma = 0.7 have been selected, where the former can be considered as a quasi-
incompressible case. The baseflows for both cases were acquired with steady direct
numerical simulations using a 2D Blasius boundary-layer solution as boundary and
initial conditions.
230 Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

The free-stream turbulence intensity is set to Tu = 4.7% for all cases using 800
eigenmodes. At this relatively high turbulence intensity, instantaneous nonlinear
interaction occurs, although the low-amplitude single modes are still governed by the
linear theory when simulated individually. However, three different integral length
scales 𝐿 are selected to determine the free-stream turbulence: 𝐿 = 5.0𝛿0∗ (Case1),
𝐿 = 2.5𝛿0∗ (Case2) and 𝐿 = 7.5𝛿0∗ (Case3). The von Kármán energy spectrum for
𝐿 = 5.0𝛿0∗ is depicted in Fig. 1(b).
A grid resolution of 1800 × 180 × 256 points in (𝑥-𝑦-𝑧) direction was used.
The simulations were decomposed in all spatial directions to 60 × 6 × 6 = 1440
MPI processes. Furthermore, all domains were parallelized in 4 openMP threads
in spanwise direction, leading to a total of 5760 processes on 45 nodes. Regarding
the computing time, all important information is given in Tab. 1, including the flow-
through time (FTT, number of runs through domain) for the entire simulation and for
the recording/averaging of flow statistics. It is noted that compressible simulations
with lower Mach numbers require a smaller timestep for time integration. For the
given flow this means that the case with Ma = 0.3 needs about 3.5 times more steps
then for the calculation with Ma = 0.7 to simulate the same physical time span.

Single free-stream disturbances

To demonstrate the unsteady inlet boundary condition, results of 2D simulations

with single free-stream disturbances (FSD) in a shorter domain with 101.3𝛿0∗ ≤ 𝑥 ≤
460.8𝛿0∗ in streamwise direction are presented first. In three separate simulations, free-
stream modes—normalized with the maximum of 𝑢—with ˆ amplitude 𝐴𝑚 = 0.001
were introduced at the inflow boundary. The real part of the 𝑢ˆ components of the
modes with the same angular frequency of 𝜔 = 1.0 (therefore with streamwise
wavenumber 𝛼𝑟 = 1.0), but different wall-normal wavenumbers 𝛾, are shown in
Fig. 2(a). The selected wavenumbers lie within the range of the discretized free-stream
turbulence spectrum of the performed simulations of bypass transition. For larger
𝛾, the amplitudes of the modes at the boundary-layer edge increase. The modal
streamwise amplitude development of the corresponding modes in the simulations
is depicted in Fig. 2(b). In accordance with the temporal linear stability analysis
in conjunction with Gaster’s transformation (dashed lines) the decay rate 𝛼𝑖 of the
FSD increases with higher wall-normal wavenumbers 𝛾. The boundary conditions
described in the previous chapter can be considered validated since the introduced
disturbances enter the domain without discontinuity at the inlet and then behave
according to the damping rates predicted by LST in downstream direction.
DNS of bypass transition under free-stream turbulence for compressible flows 231

(a) (b)
20
1.4
1.2 γ=0.66, αi =0.0048
1.0 γ=1.00, αi =0.0067
γ=1.25, αi =0.0086
15 0.8
LST (Δn)

max{u'} × 10 3
0.6

10 0.4
y

5 0.2

0
-1 0 1 100 200 300 400
Re{û} x
Fig. 2: (a) Shape of eigenmodes from the continuous spectrum. (b) Streamwise
development of modal 𝑢 ′-amplitudes for different FSD with 𝜔 = 𝛼𝑟 = 1.0.

Bypass transition

Fig. 3(a) shows a very good agreement of the transition location indicated by the
skin friction coefficient 𝑐 𝑓 between the reference and the simulations performed with
NS3D for the low Mach number for Case1 and Case3. Furthermore, the overshoot is
resolved very well. Minor differences are caused by slight compressibility effects or
differences in the numerical schemes. As in the reference, transition does not occur
at all in Case2. However, the 𝑐 𝑓 curve lies somewhat closer to the analytical laminar
solution, which could possibly indicate that the presented adapted method involves
less numerical perturbations.
The skin friction coefficient 𝑐 𝑓 of the simulations with the higher Mach number
of Ma = 0.7 is depicted in Fig. 3(b). The transition behavior of Case1 and Case3 is
similar to the low Mach number case or the incompressible reference, but already
exhibits significant differences. The onset of the transition seems to start at similar
positions—however, the 𝑐 𝑓 value does not jump as rapidly to the analytic solution
of the turbulent flow. Furthermore, in both cases an overshoot is hardly visible. The
skin friction coefficient curve for Case2 is again even closer to the analytical solution.
Thus, it can be concluded that there is a clear Mach number effect, which can only be
achieved with a numerical setup of this type.
The wall normal distribution of the RMS values of the velocity components, 𝑢 rms ,
𝑣 rms , and 𝑤 rms versus the streamwise position (averaged in spanwise direction) can
be seen in Fig. 4(a)-(c) for Case3 (𝐿 = 7.5𝛿0∗ ) at Ma = 0.3. The distributions show
very good agreement with the results from the reference [6], where the same contour
lines were chosen. In addition, the distribution of the RMS values of the density 𝜌rms
can now be plotted analogously. As expected, 𝜌rms correlates most closely with the
232 Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

Fig. 3: Skin friction coefficient 𝑐 𝑓 for all three cases. Black lines: reference results by
Brandt et al. [6]. Red/blue lines: results with NS3D. (a) Ma = 0.3 (b) Ma = 0.7.

(a) 15
u rms
10 0.15
y

0.08
5 0.01

0
100,000 Rex 200,000
(b) 15
vrms
10
0.045
y

0.025
5 0.005

0
100,000 Rex 200,000
(c) 15
wrms
10 0.065
y

0.035
5 0.005

0
100,000 Rex 200,000
(d) 15
ρ rms
10
0.0025
y

0.0013
5 0.0001

0
100,000 Rex 200,000

Fig. 4: RMS-values for Case3 (𝐿 = 7.5𝛿0∗ ) with Ma = 0.3. (a)-(c) velocities 𝑢, 𝑣, 𝑤

(d) density 𝜌.
DNS of bypass transition under free-stream turbulence for compressible flows 233

RMS distribution of 𝑢. Fig. 5(a) and Fig. 5(b) are depicting visualizations of the

(a)
80
u
60
1
z

40 0.6
20 0.2
0
200 400 600 x 800 1000 1200
(b)
80
w
60
0.25
z

40 -0.05
20 -0.3
0
200 400 600 x 800 1000 1200
(c)
80
ρ
60
1
z

40 0.992
20 0.984
0
200 400 600 x 800 1000 1200

Fig. 5: Instantaneous flowfield at 𝑦/𝛿0∗ = 2 from simulation for Case3 (𝐿 = 7.5𝛿0∗ )

with Ma = 0.3, (a) Streamwise velocity 𝑢 (b) Spanwise velocity 𝑤 (c) Density 𝜌.

instantaneous streamwise velocity 𝑢 and spanwise velocity 𝑤 for Case3 (𝐿 = 7.5𝛿0∗ ),

respectively. The corresponding plots can also be found in [6], with which they display
a very good qualitative agreement. The clearly visible high and low speed streaks—
also known as Klebanoff modes—can be regarded as precursors to bypass transition.
Secondary instabilities of these modes are leading to patches of turbulence which are
finally causing a fully turbulent boundary-layer [13]. Due to the compressible setup,
the variation of the density 𝜌 can be shown again in this case, see Fig. 5(c).

4 Performance analysis

The NS3D code has been programmed from the start with a strong focus on parallel
efficiency and scalability. The finite difference method lends itself quite naturally to
techniques for high performance computing and allows for a combination of several
techniques such as vectorization of array operations, distributed computing and
symmetric multiprocessing.
The application of these methods makes NS3D to be equally well suited for
both vector and scalar CPU systems. The number of processes used in distributed
computing correlates with the number of block structured domains of equal size. The
234 Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

decomposition is realized by Message Passing Interface (MPI) [16] routines and is

arbitrary in all three spatial directions as long as the number of points in the respective
direction is equal. Within domain blocks a second layer of parallelization is achieved
by threading the spanwise coordinate direction using openMP [17] directives. An
exemplary decomposition of the computational domain is shown in Fig. 6. The setup
of all test cases, i.e. initial and boundary conditions as well as scheme settings, was
identical to the one described in §3 except for the domain size which was adapted
according to the tests run.

mpi_isend()
mpi_irecv() Proc 10

Proc 12
Proc 6

1
Proc 7 ead d 2 3
Proc 2 ThrThreahreadread e4ad 5 d 6
T Th Thr hrea
Proc 9 T
Proc 3

Proc 5
y Proc 0
x
Proc 1
z
!$omp parallel do
...
!$omp end parallel do

Fig. 6: Schematic of domain decomposition and parallel distribution within NS3D

Single node performance

At first a scaling analysis was done on a single node of the Hawk system. These tests
provided insight in how to optimally distribute processes and threads on the nodes
and thus make optimal use of the highly hierarchical structure of the AMD EPYC
Rome CPUs.
A mesh of 240 × 180 × 256 points in (𝑥-𝑦-𝑧) direction was used. The simulations
were run for 1000 time steps and used 1–128 cores. Three different modes, namely
pure distributed computing, pure symmetric multiprocessing and a combination of
processes and threads were compared. Fig. 7 shows the speed-up 𝑆 = 𝑡1 /𝑡NP and
parallel efficiency 𝐸 = 𝑆/NP, where 𝑡• is the CPU time and NP is the number of
processing elements. For pure MPI parallelization, a linear speed-up and almost
ideal efficiency up until 32 processes is achieved. For more than 64 processes no
performance gains are obtained. Pure openMP parallelization does not show any
significant performance gains for more than 16 cores. The hybrid mode demonstrates
a speed-up for up to 128 cores, albeit small beyond 64 cores. It can be concluded that
using 32 MPI processes and four openMP threads should generally be the preferred
DNS of bypass transition under free-stream turbulence for compressible flows 235

setting for simulations using NS3D. This combination ensures fast access on shared
L3 cache within a core complex (CCX) (cf. the linear speed-up up to four cores for
pure openMP).

(a) (b)
30 1.5
4 threads MPI
l
idea

openmp

efficiency [-]
hybrid
speed-up [-]

4 threads ideal
20 1
4 threads
2 threads
10 2 threads 0.5
2 threads

0 0 32 64 96 128 0 0 32 64 96 128
number of cores number of cores

Fig. 7: Single node performance on Hawk: (a) Strong scaling speed-up; (b) Strong
scaling efficiency

Multi node performance

Multi node scaling tests were subsequently using 32 processes and four threads on
each node. Both strong and weak scaling tests were run. Two different grids were used
for strong scaling tests in order to avoid simulations with an unfeasible small number
of grid points per core. The number of points and the respective distribution for the
computations are listed in Tab. 2. Furthermore in order to smooth out distortions due
to node errors tests have been run ten times each and averaged. The effectiveness

Table 2: Grid points and parallel distribution

Case / nb. nodes 1 4 16 64 256 1024

a: 480x180x256 4-2-4,4𝑎 8-4-4,4 16-4-8,4 x x x

b: 3720x240x256 4-2-4,4 x x 32-4- 64-8- 128-8-32,4
16,4 16,4
𝑎 MPIx-MPIy-MPIz,openMP

of the parallel programming techniques used in NS3D can be seen in Fig. 8. Strong
scaling computations show continuously super-linear speed-up for up to 256 nodes
and linear speed-up until 1024 nodes. The efficiency is always above one accordingly.
236 Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

(a) (b)
1.5
1100 30 b

time [s]
25
al

efficiency [-]
speed-up [-]

ide
20
151 2 3 4 5 6 7 8 910
1 ideal
600
b
b strong scaling
a weak scaling
100 a 0.5 weak scaling (main)
a,b
0 256 512 768 1024 0 256 512 768 1024
number of nodes number of nodes

Fig. 8: Multi node performance on Hawk: (a) Strong scaling speed-up; (b) Comparison
of strong and weak scaling efficiency

The scaling performance varied somewhat especially for the last test case on 1024
nodes because of individual node errors which led to poor performance as can be
seen by the inlay in Fig. 8(a). The probability to include a node with either faulty
hardware or poor inter-node connection obviously increases when using a number
of cores of order 105 . Weak scaling computations were done with the number of
grid points per computational element, i.e. a single node, set to 3686400. For 1–256
nodes the efficiency is almost ideal with a drop of only 6%. Even for 1024 nodes an
efficiency of 79% percent can be achieved. This performance drop can be attributed
to initialization routines. If the time to complete the main loop is compared only
(cf. solid line in Fig. 8(b)) the efficiency is almost ideal for up to 1024 nodes. As the
test cases had been run only for 1000 time steps this timer was deemed the more
appropriate indicator for performance measurements.
In conclusion it can be said that the historically proven strong parallel performance
of NS3D can also be uphold on the most recent HLRS system

5 Conclusions

This study presents results of high-fidelity direct numerical simulations of compress-

ible transitional boundary-layer flows. The laminar-turbulent transition is triggered
by introducing isotropic free-stream turbulence into the flow through appropriate
boundary conditions. The turbulent disturbances were computed from linear stability
analysis following an established method which was adapted to include compress-
ibility. It was shown that the disturbances are a realistic representation of isotropic
turbulence above walls; disturbance growth followed theoretical predictions and shear
sheltering is captured. Very good agreement with reference cases from the literature
was achieved for quasi-incompressible flow speeds.
DNS of bypass transition under free-stream turbulence for compressible flows 237

The results of the transition simulations show a clear Mach number dependency
of the flow. For identical length scales the onset of transition is delayed for higher
Mach number. The velocity fluctuations are in very good agreement with results from
literature. Additionally, the fluctuation density is quantified thus allowing for a future
estimation of not only wall friction but also wall heating.
In scaling tests the revised version of the NS3D solver proved to maintain its
strong parallel performance on a massively parallel system such as the Hawk. The
combination of techniques for distributed computing and symmetric multiprocessing
resulted in super-linear and linear speed-up when running on up to 256 and 1024
compute nodes res. Both strong and weak scaling consistently demonstrated near
ideal efficiency up to 1024 nodes.

Acknowledgements Duncan Ohno acknowledges funding from the Bundesministerium für Wirt-
schaft und Energie through the LuFo project “LAINA”. Direct numerical simulations were performed
on resources provided by the High Performance Computing Center Stuttgart (HLRS) under grant
GCS_Lamtur, ID44026.

References

1. Jacobs, R.G., Durbin, P.A.: Simulations of bypass transition. J. Fluid Mech. 428, 185–212
(2001)
2. Zaki, T.A., Durbin, P.A.: Continuous mode transition and the effects of pressure gradient. J.
Fluid Mech. 563, 357–388 (2006)
3. Schrader, L.U., Brandt, L., Henningson, D.S.: Receptivity mechanisms in three-dimensional
boundary-layer flows. J. Fluid Mech. 618, 209–241 (2009)
4. Schmidt, O., Rist, U.: Numerical investigation of classical and bypass transition in streamwise
corner-flow. Procedia IUTAM 14, 218–226 (2015)
5. Ohno, D., Romblad, J., Rist, U.: Laminar to turbulent transition at unsteady inflow conditions:
Direct numerical simulations with small scale free-stream turbulence. New Results in Numerical
and Experimental Fluid Mechanics XII (2020)
6. Brandt, L., Schlatter, P., Henningson, D.S.: Transition in boundary layers subject to free-stream
turbulence. J. Fluid Mech. 517, 167–198 (2004)
7. Babucke, A., Linn, J., Kloker, M., Rist, U.: Direct numerical simulation of shear flow phenomena
on parallel vector computers. In High Performance Computing on Vector Systems (pp. 229-247).
Springer, Berlin, Heidelberg. (2006)
8. Mack, L.M.: Boundary-layer linear stability theory. AGARD Report No. 709 (Special Course
on Stability and transition of Laminar Flow), pp. 3/1–81 (1984)
9. LAPACK: Linear Algebra PACKage. http://netlib.org/lapack
10. Hunt, J., Durbin, P.: Perturbed vortical layers and shear sheltering. Fluid Dyn. Res. 24(6),
375–404 (1999)
11. Gaster, M.: A note on the relation between temporally-increasing and spatially-increasing
disturbances in hydrodynamic stability. Journal of Fluid Mechanics. Vol. 14 (1962) 222–224.
12. Tennekes, H., Lumley, J.L.: A First Course in Turbulence. MIT Press, Cambridge (1972)
13. Durbin, P.A.: Perspectives on the phenomenology and modeling of boundary layer transition.
Flow Turbul. Combust. 99(1), 1–23 (2017)
14. Giles, M.B.: Nonreflecting boundary conditions for Euler equation calculations. AIAA journal
28.12 (1990): 2050-2058
15. Appelbaum, J., Ohno, D., Rist, U., Wenzel, C.: DNS of a Turbulent Boundary Layer Using
Inflow Conditions Derived from 4D-PTV Data. Experiments in Fluids. (2021) [In review]
238 Duncan Ohno, Björn Selent, Markus J. Kloker and Ulrich Rist

16. N.N.: MPI - A message-passing interface standard. Technical Report CS-94-230, University of
Tennessee, Knoxville (1994)
17. OpenMP: The OpenMP API Specification for Parallel Programming. http://www.openmp.org
Direct numerical simulation of a disintegrating
liquid rivulet at a trailing edge

Adrian Schlottke, Matthias Ibach, Jonas Steigerwald and Bernhard Weigand

Abstract Water ingestion in gas turbines is used in industrial applications to improve

the thermal efficiency by cooling the air before and throughout compression. However,
this also leads to interactions between the liquid droplets and the compressor parts,
which cause a faster degradation of the structure. The current work addresses the
numerical investigation of the atomization process at the trailing edge of a compressor
blade as there have only been experimental investigations in literature considering
the ambient conditions in a gas turbine compressor. Direct numerical simulations
(DNS) are carried out using the multiphase flow solver Free Surface 3D (FS3D).
Therefore, a numerical setup has been developed to model the trailing edge as a thin
plate corresponding to experiments performed at ITLR [22]. Four different cases
have been performed to take the experimentally observed atomization processes into
account. Additionally, the dependence of the simulation results on the grid resolution
and with it the limits to reproduce the experimental findings has been investigated in
a grid study. The results show that the developed numerical setup works well and
the different atomization processes are reproduced qualitatively, with best results for
very high grid resolutions.
Furthermore, an investigation of the available compilers on the new Hawk super-
computer platform reveals that a performance gain up to 36% for the amount of
completed calculation cycles per hour is possible compared to the standard compiler
setup used as standard practice over the previous years. Optimization options as well
as improvement during link time lead to a significant speed-up while simulation
results remain unaffected.

Adrian Schlottke
Institute of Aerospace Thermodynamics (ITLR), University of Stuttgart, Pfaffenwaldring 31,
70569 Stuttgart, Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 239
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_14
240 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand

1 Introduction

The ingestion of water droplets into gas turbine compressors, also called Fogging or
High-Fogging, is used to improve the thermal efficiency of the turbine by cooling
the air before and throughout the compression. However, many of the ingested
droplets interact with the compressor parts, e.g. liquid films accumulate on the
blade surface and disintegrate at the trailing edge, generating new droplets. The
droplets impacting onto the blades lead to erosion and a faster degradation of the
material. The focus of this work lies on the numerical investigation of the trailing edge
disintegration as an atomization process. The atomization process of different nozzles,
e.g. circular, flat, and also prefilming air-blast atomizers has been investigated in
literature, both experimentally and numerically. Atomization of liquid jets has been
investigated experimentally for a long time and many authors have contributed to
the understanding of the process, e.g. the work of Rayleigh [18] and Weber [25] are
well known. Dumouchel [2] gives a detailed overview of experimental investigations
on jet and sheet atomization including also more recent work as well. Simulations
and especially highly resolved Direct Numerical Simulations (DNS) of jet breakup
just became feasible with increasing computational power in recent years [4, 19].
However, these can only be performed at supercomputers and therefore Evrard [6]
proposed a hybrid Euler-Lagrange method, especially for jet breakup phenomena, to
reduce computational efforts.
Although the same morphological atomization behavior occurs for air-assisted jets
and for trailing edge disintegration, the limits and transition from one atomization
regime to another depend strongly on the application. The atomization process of
prefilming air-blast atomizers resembles qualitatively the trailing edge disintegration
at compressor blades. Experimental investigations on planar and swirl prefilming
air-blast atomizers were performed in [8] and [9] and have shown the dependence of
the atomization process and the resulting ejected droplets on the prefilmer geometry
and the ambient conditions. Numerically, Koch et al. [15] investigated planar prefilm
atomizer using a smoothed particle method and a 2D setup. The findings were
validated with experimental results from [9]. Nevertheless, the used 2D setup with
the smoothed particle method has some shortcomings, e.g. the small computational
domain in longitudinal direction, which disables the representation of the complete
liquid breakup process. Additionally, atomization at the trailing edge of a prefilmer
and of a compressor blade differ quantitatively due to the influence of the walls within
the prefilmer. The small geometrical distance between liquid film surface and top
wall influences the liquid film behavior, which is not the case for compressor blades.
Regarding the trailing edge disintegration at compressor blades, there exists a lack
of numerical data as only experimental investigations were reported in literature to
the best knowledge of the authors. The investigations mainly focus on the statistical
quantification of the atomization process of a liquid film at a trailing edge. Kim
[14] uses high-speed shadowgraphy imaging to evaluate the ejected droplet diameter,
acceleration and disintegration frequency. The experimental setup relates to conditions
of large steam turbine blades. Consecutively, Hammit et al. [10] investigate the change
in droplet diameter spectra as a function of distance from the trailing edge. A more
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 241

recent experimental test campaign of Javed et al. [12, 13] investigates the influence of
the trailing edge thickness and the blade’s angle of attack on the disintegration of a
liquid rivulet. It has been shown that at higher ambient velocities, i.e. when breakup
is dominated by aerodynamic forces, a larger trailing edge thickness leads to a smaller
distribution angle of the ejected droplets. At a constant Weber number, depending
on the trailing edge thickness, the ejected droplet diameters correlate with trailing
edge thickness. An increasing angle of attack leads to increased droplet sizes due
to decreased local air momentum in the region after the trailing edge. Additionally,
it is stated that the droplet size distribution is independent of the liquid mass flow
rate. So far, the process of trailing edge disintegration under gas turbine conditions
has only been investigated experimentally. The experiments focus on an integral
observation of the process. This is important for a first understanding of the process,
but for a complete prediction, it is important to understand the local process and
the underlying physics. For this reason, DNS performed with FS3D will enable the
insight into the liquid structure during breakup. As a first step, this work provides
the numerical setup and first results for comparison with the findings from detailed
experiments performed in [22].

2 Mathematical description and numerical approach

The in-house CFD code Free Surface 3D (FS3D) performs DNS of incompressible
multiphase flows and is continuously developed at ITLR for the last 25 years. Several
recent studies show the applicability of FS3D to simulate highly dynamic multiphase
processes like droplet deformation [20], droplet impacts onto dry and wetted surfaces
[1, 7], droplet collisions [17] and atomization of liquid jets [5]. Simulations of such
complex problems require high spatial and temporal resolution and thus a high
parallel efficiency of the used solver. For this reason, FS3D is parallelized with MPI
as well as OpenMP. FS3D solves the governing equations for mass and momentum
conservation on finite volumes


𝜕𝑡 𝜌 + ∇ · 𝜌u = 0, (1)

𝜕𝑡 𝜌u + ∇ · 𝜌u ⊗ u = ∇ · S − I 𝑝 + 𝜌g + f 𝛾 .




(2)
In equations (1-2), u denotes the velocity vector, 𝑝 the static pressure, g the gravita-
tional acceleration, S the shear stress tensor and I the identity matrix. The term f 𝛾
represents the body force which is used to model surface tension at the interface. The
governing equations are solved in a one-field formulation where the different phases
are regarded as a single fluid with variable physical properties across the interface.
FS3D uses the Volume-of-Fluid (VOF) method by Hirt and Nichols [11] to identify
these different phases. An additional variable 𝑓 is introduced, which is defined as
242 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand



0

 outside the liquid phase,



𝑓 (x, 𝑡) = (0, 1) at the interface, (3)




1
 inside the liquid phase.

The variable 𝑓 is then advected using the transport equation


𝜕𝑡 𝑓 + ∇ · 𝑓 u = 0. (4)

The corresponding 𝑓 -fluxes are calculated using the PLIC method by Rider and Kothe
[21] to maintain a sharp interface. By using the volume fraction 𝑓 local variables
such as the density can be calculated as


𝜌(x, 𝑡) = 𝜌𝑙 𝑓 (x, 𝑡) + 𝜌𝑔 1 − 𝑓 (x, 𝑡) . (5)

The surface tension forces are modeled by the continuous surface stress (CSS) model
by Lafaurie et al. [16]. Further details on numerical implementations and applications
of FS3D are given e.g. in Eisenschmidt et al. [3].

Fig. 1: Computational domain for simulation of a disintegrating liquid rivulet at a

trailing edge.
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 243

Table 1: Inflow conditions.

Inflow velocity (m/s) Circle segment (mm) 𝑅𝑒𝐺 (-) Disintegration regime
Air 𝑢𝑔 Liquid 𝑢𝑙 Height Width

11.5 0.103 1.35 2.70 15 × 103 Symmetrical Rayleigh breakup

19.2 0.172 0.81 2.70 25 × 103 Asymmetrical Rayleigh breakup
26.8 0.199 0.70 2.70 35 × 103 Bag breakup
34.5 0.199 0.70 2.70 45 × 103 Bag breakup

3 Computational setup

The simulation of the trailing edge disintegration at a thin plate is performed in a

rectangular computational domain as visualized in fig. 1. A rivulet flows on top and
over the edge of a thin horizontal plate of length 𝐿 in the middle of the domain.
The dimensions of the domain are 16𝐿 × 4𝐿 × 8𝐿 in streamwise, spanwise and
in plate-normal direction, respectively. In order to increase the spatial resolution,
only half of the rivulet is simulated. On the left side of the domain at 𝑥 = 0𝐿 an
inflow of both gaseous and liquid phase is applied. The gas flows with velocity 𝑢 𝑔
through the entire side into the domain whereas liquid flows with a lower velocity
𝑢 𝑙 on top of the thin plate into the domain with an idealized shape as a circular
segment. In the present study, four different settings are investigated. Both velocities
as well as the geometrical shape of the inflowing rivulet, defined by the maximum
height and width, correspond to experimental findings of Schlottke and Weigand [22].
An overview of the inflow conditions with the corresponding Reynolds numbers of
the air flow 𝑅𝑒 𝐺 = 𝜌𝑔 𝑢 𝑔 𝑑 ℎ /𝜇𝑔 is given in table 1. According to the experiments
in [22], the characteristic length is defined by the hydraulic diameter 𝑑 ℎ = 32𝐿 of the
experimental flow channel.
As can be seen in the last column of table 1, different disintegration scenarios were
investigated: Symmetrical and asymmetrical Rayleigh breakup and bag breakup. In
order to accurately reproduce these scenarios and to identify the necessary resolution
of the used equidistant Cartesian grid, all cases were simulated with three different
grid resolutions: A coarse grid with 512 × 256 × 128 cells (Δ𝑥 = 78 µm), a grid with
1024 × 512 × 256 cells (Δ𝑥 = 39 µm), and a fine grid with 2048 × 1024 × 512 cells
(Δ𝑥 = 20 µm).
The novelty of the present setup is the treatment of the thin plate. So far it has not
been feasible to represent a non-moving, rigid part with a finite contact angle inside
the computational domain with FS3D. To overcome this problem a new methodology
for a special treatment of artificial inner boundaries was implemented in FS3D
which is explained in the following. The whole computational domain is divided into
equally sized blocks. Between the sides of adjacent blocks all available boundary
conditions that are implemented in FS3D can be applied. In the present study the
244 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand

domain is divided into 16 × 2 × 1 blocks in 𝑥- ,𝑦-, and 𝑧-direction which are indicated
by the gray dashed lines in fig. 1. For reasons of clearer visualization only half of
the blocks in 𝑥-direction are shown. For the treatment of the thin plate a no-slip
boundary condition with a constant contact angle of 90◦ is applied between the first
two blocks in 𝑦-direction at the inlet side. In this way, an idealized thin plate with
infinitesimal thickness is reproduced. Besides the inflow and the symmetry conditions
at the domain boundary a continuous (homogeneous Neumann) boundary condition
is applied at the right side of the domain at 𝑥 = 16𝐿, representing the outlet. All other
sides use a free-slip boundary condition. Furthermore, the simulation accounts for
gravitational forces with 𝑔 = 9.81 m/s2 in 𝑦-direction. The physical properties of gas
and liquid are given in table 2.

Table 2: Physical properties of air and water.

Density 𝜌 (kg/m3 ) Dynamic viscosity 𝜇 (Pa s) Surface tension 𝜎 (N/m)

Air 1.190 1.824 × 10−5

73.0 × 10−3
Water 998.2 1.000 × 10−3

4 Results and discussion

Within this chapter, the results of the DNS will be evaluated and compared to
experimental findings from [22]. The discussion of the results is divided into three
parts. First, a qualitative, morphological comparison of the breakup process is
presented to highlight that the different breakup regimes and their individual features
can be reproduced. The second part focuses on the behavior of the ligament at
the trailing edge. Especially the geometrical dimensions, i.e. the width, height and
maximum length before breakup, are of interest. At last, the ejected droplet diameter
distribution will be discussed. The influence of grid resolution will be addressed in
each part separately. Regarding the last part it has to be mentioned that the evaluation
of droplets that are generated during the course of the simulation is not a trivial
task. In all previous studies with FS3D, the counting and characterization of these
droplets was always part of the post-processing. Only field data that were written
during the simulation could be used. Additionally, loading the data into the memory
again is very time-consuming. With the new generation of supercomputers like the
HPE Apollo system (Hawk), however, the spatial grid resolution of the computational
domain can be further refined. Due to that, the traditional way of droplet evaluation
via post-processing has become inefficient and almost unfeasible. For this reason, the
whole process of counting and full characterization of existing droplets/ligaments
inside the computational domain (volume, position, velocity) has been shifted from
the post-processing into the simulation cycle. A new fully parallelized output routine
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 245

that is based on a Connected-component labeling (CCL) algorithm for identifying

separate droplets was implemented in FS3D, which is described in the following. In a
first step, all liquid ligaments within a local processor domain are identified by means
of a CCL algorithm and they are marked incrementally and in a processor-dependent
manner. The properties of each ligament are calculated during the labeling as well.
The generated ligament marker field is shared afterwards by using an exchange routine
of the solver which is normally used to update the ghost cells of each local processor
domain. Due to that, a simple loop over all grid cells at the local domain boundary
can be used to check if two differently marked regions belong to the same ligament. In
case the markers of two neighboring grid cells differ, this pair of markers is added to
a global adjacency list, representing an undirected graph. The connected components
of the graph or the liquid ligaments, respectively, can then be identified by using the
complete adjacency list and a depth-first search. In a last step, associated ligament
characteristics are merged. This leads to a highly efficient way of extracting droplet
information from the simulation during runtime, independent from other data that is
written and almost independent from the chosen grid resolution.
The comparison of DNS and experimental results at specific points in time is
difficult due to the strong statistical behavior of the breakup process. This is mainly
affected by the changing initial conditions as each breakup relates on the previous
one rather than being independent. This dependence increases for higher ambient
gas velocities as the breakup frequency increases and, therefore, there is less time for
relaxation between two breakups. Nevertheless, it is important to evaluate if the DNS
can reproduce the morphological structures, which were observed experimentally
and how they are affected by the grid resolution. Figure 2 shows selected side view
snapshots of the trailing edge breakup for two different 𝑅𝑒 𝐺 and two different breakup
regimes, respectively. The experimental images are compared to DNS results for two
different grid resolutions. At 𝑅𝑒 𝐺 = 15 × 103 , symmetric Rayleigh breakup can be
observed. The experimental images show the typical behavior where one large droplet
is ejected straight in streamwise direction at breakup, often followed by few small
satellite droplets. This can also be seen in the DNS results for both grid resolutions. At
𝑅𝑒 𝐺 = 45 × 103 , the bag breakup regime can be observed. Here, the single breakup
event cannot be distinguished anymore, but rather the flapping motion at the trailing
edge is considered to distinguish between breakups. In the experiment, two states
of the bag breakup can be observed: One is the thinning of the lamella right at the
trailing edge where the bag is created. The other is the already fragmented liquid
bag at the front of the ligament, which creates many small droplets from the lamella
and few larger droplets from the rim. The DNS results also show the continuous
flapping motion of the ligament. However, the simulation could not reproduce the
creation of a bag. This may result from two reasons: One is the grid resolution. Even
for the highly resolved case the thickness of the lamella is in the order of the size
of a grid cell. Due to that, the interface cannot be accurately reconstructed anymore
since both sides of the lamella influence each other. The second reason may be due to
the symmetry plane at 𝑧 = 0. In the context of 2D simulations, it has been stated in
246 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand

Fig. 2: Instantaneous morphological comparison between DNS results and experi-

mental data for 𝑅𝑒 𝐺 = 15 × 103 and 𝑅𝑒 𝐺 = 45 × 103 . Side view.

literature that instabilities in spanwise direction trigger the formation of liquid bags
[9, 15]. Although spanwise instabilities are completely ignored in 2D simulations the
use of the symmetry plane still suppresses these instabilities drastically.
In the following, a quantitative analysis of the liquid rivulet will be presented. It
considers the geometrical dimensions, i.e. width and height of the rivulet close to
the trailing edge and the maximum length of the ligament directly before breakup.
This quantities are shown in fig. 3 in dependence of 𝑅𝑒 𝐺 for all simulated cases.
The results for the different grid resolutions are depicted side-by-side for better
visualization. However, it needs to be mentioned that the simulations are performed
at the same inflow conditions, i.e. the same 𝑅𝑒 𝐺 . The shown results contain the range
of evaluated values over time (up and down arrow), as well as the mean value (circle).
The mean values for width and height are time averaged whereas the maximum
length is averaged by the number of breakups. The dependence of the mean width
and height on the grid resolution is rather small and lies in a range of up to 200 µm.
While the mean width decreases from 𝑅𝑒 𝐺 = 15 × 103 to 𝑅𝑒 𝐺 = 25 × 103 and than
stays constant the mean height is decreasing monotonously for higher 𝑅𝑒 𝐺 .
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 247

This accords well with the experimental findings in [22], both qualitatively and
quantitatively. A closer look at the range of width and height reveals a small difference
of the results using different fine grids, especially for the cases with high ambient gas
flow at 𝑅𝑒 𝐺 ≥ 35 × 103 . It can be seen that with higher grid resolution the range is
increasing. This behavior can be explained by the wavy surface of the liquid rivulet at
these 𝑅𝑒 𝐺 . A sufficient grid resolution is needed to represent these surface waves and
their influence on the movement of the liquid rivulet. In contrast, the grid resolution
has a strong influence on the maximum length of the ligament before breakup. While
for 𝑅𝑒 𝐺 = 15 × 103 the mean values lie quite close together, the difference grows
up to 39% at 𝑅𝑒 𝐺 = 35 × 103 . It is apparent that the maximum length at breakup
diminishes for higher grid resolution. Looking at the range of maximum length, it can
be stated that for 𝑅𝑒 𝐺 > 25 × 103 the difference between minimum and maximum
value stays quite constant for the middle and high grid resolution with Δ ≈ 5.0 mm
and Δ ≈ 5.7 mm, respectively. The general behavior of the maximum length is the
same for all grid resolutions. It elongates from 𝑅𝑒 𝐺 = 15 × 103 to 𝑅𝑒 𝐺 = 25 × 103
Mean rivulet width, height, length at trailing edge in mm

18
Width
16 Height
Length
14 ∆x = 78 μm
∆x = 39 μm
12 ∆x = 20 μm

10

0
15 103 25 103 35 103 45 103
R eynolds number of air flow R eG

Fig. 3: Mean numerical results of rivulet width and height measured at the trailing
edge. Length is defined as the maximum distance between ligament tip and trailing
edge. Length is measured at the instance of breakup. Results are shown for the three
resolutions used.
248 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand

and decreases then monotonously. The same qualitative trend is observed in the
experiments. However, it needs to be mentioned that the quantitative values are
about 50% smaller than the experimental data. A possible reason for this is the inlet
condition of the liquid rivulet as the defined velocity strongly affects the liquid’s
momentum at the trailing edge. Higher velocities at the inlet cause higher momentum
at the trailing edge and lead to larger maximal length of the ligament before breakup.
As a conclusion, it can be assumed that in case of symmetrical Rayleigh breakup,
where only large droplets occur and individual breakups are nearly independent of
one another, all used grid resolutions produce comparable results. At other breakup
regimes, when the ejected droplets become smaller, more numerous and the influence
of small perturbations, e.g. surface waves, is growing, higher grid resolutions are
needed to resolve these phenomena adequately.
At last, the ejected droplet spectra will be evaluated and the influence of the grid
resolution is shown. Here, only the case at 𝑅𝑒 𝐺 = 45 × 103 is discussed in detail.
Figure 4 shows histogram plots for all three used grid resolutions.
The bars represent the probability of the corresponding equivalent spherical droplet
diameter where the normalization is performed with the total number of ejected
droplets for each case. The histogram bin size is 200 µm, starting from the droplet
diameter 𝑑 = 0 µm. The vertical dashed line indicates the minimal droplet diameter
which can still be resolved adequately. With a higher grid resolution this diameter is
decreasing, respectively. The first histogram bin is adapted in accordance with the
individual minimum resolvable diameter. The results show highest probabilities at
the smallest droplet diameters, excluding the first smaller bin at Δ𝑥 = 78 µm and
Δ𝑥 = 20 µm. Although the trend is the same, the grid resolution influences the mean
and the maximum ejected droplet diameters 𝑑 𝑚𝑒𝑎𝑛 and 𝑑 𝑚𝑎𝑥 , respectively. The mean
diameter 𝑑 𝑚𝑒𝑎𝑛 is defined as the arithmetic mean value of all identified droplets and
𝑑 𝑚𝑎𝑥 represents the largest droplet found throughout the investigation. Both values
are decreasing for a higher grid resolution, see table 3 for details and the comparison

μ μ μ

Fig. 4: Histogram of ejected droplets at the trailing edge for all three resolutions at
𝑅𝑒 𝐺 = 45 × 103 .
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 249

to the experimental values from [22]. The mean diameter in the experiments is even
smaller than the computed mean diameter with the highest grid resolution, although
it is capable to resolve droplets down to 120 µm. The trend towards smaller ejected
droplets, the limitation of the minimum resolvable droplet diameter, and the missing
appearance of liquid bags before breakup indicate that even the highest resolution
with Δ𝑥 = 20 µm is not sufficient to predict the ejected droplet diameter distribution.

Table 3: Mean and maximum ejected droplet diameter.

Grid resolution Δ𝑥 78 µm 39 µm 20 µm Experiment

[22]
Number of grid cells 16.8 Mio 134.2 Mio 1, 073.7 Mio

𝑑𝑚𝑒𝑎𝑛 0.83 mm 0.42 mm 0.36 mm 0.21 mm

𝑑𝑚𝑎𝑥 1.36 mm 1.11 mm 0.97 mm 1.13 mm

At this point, the reader is reminded of the already very high number of grid cells
used which was about 1 billion for the highest resolution and is close to the upper limit
of feasible computations using FS3D on HAWK. With the presented computational
domain further grid refinement is not possible. This leads to the final remarks of this
section where the shortcomings of the presented setup shall be addressed shortly.
Due to the use of the symmetry plane the spanwise disturbances, which support the
evolution of liquid bags, are limited. In comparison to the 2D simulations found
in literature, this is still an improvement, as 2D simulations completely suppress
interaction in spanwise direction. Furthermore, the trailing ligament is sensitive to
the inlet conditions of the liquid. As a larger computational domain is not feasible
in terms of computational power, further improvement of the inlet conditions is
necessary to match the experimental findings. At last, the results indicate that a grid
refinement is needed to enable the simulation of bag breakup and the subsequent
droplet diameter distribution. This will only be possible with a smaller computational
domain. The authors are aware of the limitations of the presented setup and future
studies will focus on the improvement of the simulations. Nevertheless, the results
show the successful representation of a thin plate and the atomization at the trailing
edge and good qualitative agreement with experimental results.

5 Computational performance

The transition from one generation of high performance computing systems to another,
such as from the Cray XC40 system “Hazel Hen” to the new AMD-based HPE Apollo
supercomputer “Hawk” at HLRS, always entails new challenges but also possibilities
when porting the users’ code and application. This step inevitably requires a good
250 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand

knowledge and thorough understanding of the interactions between the utilized code
and the provided hardware, e.g. at core-, socket- and node-level, where the code
gets executed and actual computational work is performed. One way of exploiting a
possible speedup of the ported application as a first step is to tweak environmental
settings and test available compilers and compiler options.
With the successful implementation of a tree structured communication in the
implemented multigrid solver in FS3D [23] simulations with more than 83 processors
are now practicable. This is of paramount importance especially for highly spatially
and temporally resolved direct numerical simulations such as for the presented case
of atomization of a liquid rivulet at a trailing edge. In the frame of this computational
performance study we investigated the available compilers and compiler options to
lay the foundation of an optimal setup for subsequent numerical investigations on the
HPE Apollo platform. In the following section we present an updated benchmark
case and provide a report on the performance analysis in terms of calculation cycles
per hour (CPH) measuring the parallel performance of both strong and weak scaling.

5.1 Benchmark case and performance analysis

For the setup of the representative benchmark case we simulate an isolated but
stretched droplet similar to fig. 2 resulting from the trailing edge disintegration at
𝑅𝑒 𝐺 = 25 × 103 . The reason for the choice of this setup is the symmetrical character
which distributes the computational load somewhat evenly making a performance
analysis more practicable and comparable. The computational domain consists of
an elongated and subsequently oscillating droplet initialized as an ellipsoid (semi-
principal axis of 𝑎 = 𝑏 = 1.357 mm and 𝑐 = 0.543 mm with an equivalent spherical
diameter of 𝑑 = 2 mm) in a cubic domain with an edge length of 8 mm. The fluid of
the droplet is water at standard ambient conditions 𝑇 = 293, 15 K and pressure of
one atmosphere similar to the cases presented in [23, 24].
For the analysis of strong and weak scaling we varied the number of processors
from 23 up to 163 which corresponds directly to the initiated MPI-processes as we did
not employ hyperthreading or parallelization on loop level with OpenMP. A baseline
case for single core performance is added for the weak scaling investigation. For all
simulations we tracked the exact time to initialize the simulation and set the walltime
to 20 minutes in order to estimate the amount of completed calculation cycles per
hour (CPH). To be independent of the interconnected file system and analyze plain
computational performance we omitted the output of restart files, integral and field
data. The simulations are performed with the latest revision of FS3D employing novel
implementations including an improved advection scheme [26], a revision of the
balanced CSF algorithm, the implicit viscosity calculation as well as the computation
of normal vectors. Note that these numerical settings have been adapted compared
to former FS3D benchmark cases and a direct comparison of calculation cycles per
hour to previous results is therefore pointless.
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 251

Several compilers and optimization options are available within the HPE Apollo
Hawk environment. The choices consist of an Intel Compiler (version 19.1.0), a
GNU Compiler (version 9.2.0) and an AOCC AMD Optimizing C/C++ Compiler
based on LLVM with a Fortran compiler flag (version 2.1.0) with the utilized version
denoted in brackets. The optimization options, which can reduce code size and
improve performance of the application, can be varied starting from a low level (O2)
to a higher optimization (O3) finishing with a more aggressive optimization option
(Ofast). Additionally, the technique of Link-Time Optimization (LTO for GNU and
AOCC) and Interprocedural Optimization (IPO for Intel)1 is available allowing the
compiler to optimize the code at link time. During this, further rearrangement of
the code from separate object files is performed. For the subsequent performance
analysis we use a combination of the available compilers and compiler options
with the possibility of link-time optimization to assess the optimal compiler setup
leading to maximal performance on the new supercomputer system. A version
without all optimization (optimization level O0, debug) was not investigated since the
computational performance is significantly lower compared to low level optimization
and the usage of this binary code is not feasible for the applied benchmark case.

5.1.1 Strong scaling

For the strong scaling performance we used a baseline case with a grid resolution of
5123 cells while increasing the number of cores from 23 up to 163 . The intermediate
steps of 2 · 83 and 4 · 83 MPI-processes are also taken into account although the
amount of cells per process is not cubic. As no calculation cycle could be performed
for a single core during the allotted walltime this case is disregarded. We pinned
processes and threads to CPU cores via omplace utilizing 64 of the available 128
cores to increase the available memory bandwidth per core. To have a representative
statistical evaluation we performed ten runs for each investigated case spanning a
period of several weeks for performance calculations. Simulations which included
obvious performance deviations of more than 25% due to instabilities in the system
were omitted. The presented results are therefore arithmetic means of the employed
walltime, initialization time and the performed calculation cycles.
It has been standard practice in the past, for reasons of compatibility of locally
available compilers, to compile with the Intel Compiler using the O2 optimization
option. Therefore, we chose this case as the reference case on which all results are
referred to. The strong scaling setup is summarized in table 4 along with the results
of the peak performance case with 2048 processors.
Figure 5 depicts the estimated cycles per hour for 23 , 83 and 4 · 83 MPI-processes
for all regarded compilers and compiler options.
For the cases of 23 and 83 processes a slight increase in performance from the
standard option O2 towards the highest optimization level Ofast can be observed
for the GNU and Intel Compiler whereas no effect can be discerned for the AOCC
1 In further context this option is denoted with a superscript 𝐿 for all compilers and compiler options
equivalently.
252 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand

Table 4: Setup for strong scaling and calculation cycles per hour (CPH) at peak
performance
(2048 processors) for all compilers and compiler options with Intel O2 as the reference
case.

Problem size 5123

MPI-processes 23 43 83 2 · 83 4 · 83 163
Cells per process 2563 1283 643 642 · 32 64 · 322 323
Nodes 1 1 8 16 32 64

Cycles per hour (CPH) at peak performance for 4 · 83 MPI-processes

Compiler options O2 O2 𝐿 O3 O3 𝐿 Ofast Ofast 𝐿

Intel (CPH) 3542 3741 3758 3815 3794 3862

− +5.6% +6.1% +7.7% +7.1% +9.0%

GNU (CPH) 3990 4281 4670 4701 4739 4831

+12.6% +20.8% +31.8% +32.7% +33.8% +36.4%

AOCC (CPH) 3528 3538 3542 3511 3420 3516

−0.4% −0.1% 0.0% −0.9% −3.4% −0.7%

Fig. 5: Comparison of CPH for different compiler options for 23 , 83 and 4 · 83 MPI-
processes (constant problem size of 5123 grid cells) and strong scaling performance
for the Ofast 𝐿 case.
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 253

Compiler. Generally, the best performance is achieved by the GNU Compiler whereas
the AOCC Compiler cannot reach the performance of the Intel reference case in any
configuration. For the peak performance case of 4 · 83 MPI-processes this trend is
more obvious: For both the GNU and the Intel Compiler the performance increases
with an increasing level of optimization whereas the AOCC Compiler cannot exploit
the potential of code optimization. The GNU Compiler shows a performance boost of
about +20% for the Ofast 𝐿 option compared to the GNU standard O2 case while the
Intel Compiler can still reach an approximate +9% rise. The best performance gain
comparing the compilers is achieved with the GNU Compiler which outperforms
the Intel reference case with more than +36% comparing the amount of calculated
cycles per hour. The link-time optimization can increase the performance for each
compiler and compiler optimization option up to 7% but mostly leads to a boost of
approximately 1 − 3% depending on compiler and optimization level. A drawback
of the linking with LTO/IPO is that it takes a considerable amount of time longer
than regular linking. However, this is quickly compensated with a steady use oft the
compiled binary in subsequent simulations saving computational resources.
For the analysis of strong scaling we chose the compiler option Ofast 𝐿 for the
sake of clarity as it revealed the maximum performance gain for all compilers.
The estimated cycles per hour show a similar trend as it is expected from previous
strong scaling analysis with FS3D. For the HPE Apollo Hawk system the peak
performance is achieved with 2048 MPI-processes following a decrease in CPH with
4096 processors. For even more processors (not depicted here) the completed cycles
per hour do not decrease drastically but stay constant, however, decreasing the strong
scaling efficiency continuously. The significant performance increase for the GNU
Compiler is most clearly visible at peak performance. It should be noted that the
Ofast optimization level enables all optimization options disregarding strict standards
compliance. This could lead to deviations in simulation results originating solely
from the compilation of the code. However, for this benchmark case we compared
integral and field data and confirmed that the results were binary identical.

5.1.2 Weak scaling

For the weak scaling analysis we kept the cells per processor constant at 643 while
progressively increasing the number of allotted MPI-processes from one to 163 . This
leads to problem sizes ranging from 643 up to 10243 grid cells. The setup is summed
up in table 5 along with the selected results of 163 processors.
Some general trends and findings from the strong scaling analysis are also
observed for the weak scaling measurements: Higher optimization levels also lead
to a performance gain (see figure 6) where an effective optimization of the code is
noticeable for the GNU and the Intel Compiler. Note that the case for the weak scaling
with 83 processors is already depicted in figure 5 as it is the same setup. The Intel and
AOCC Compiler are surpassed by the GNU Compiler for all options and investigated
cases for weak scaling by up to 45% for the single and the eight processor case and
still 7.5% with 4096 processors.
254 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand

Fig. 6: Comparison of CPH for different compiler options for 23 , 43 and 163 MPI-
processes (constant cells per process of 643 ) and weak scaling performance for the
Ofast 𝐿 case. In green: Results for the GNU Ofast 𝐿 case with 83 processors with a
variation of allotted processes per node (ppn).

For the weak scaling, however, the biggest influence of the compiler options can be
observed for one and eight MPI-processes and decreases with an increasing amount
of MPI-processes. The link-time optimization only leads to a maximum boost of

Table 5: Setup for weak scaling and calculation cycles per hour (CPH) for 4096
processors for all compilers and compiler options with Intel O2 as the reference case.

Cells per process 643

Problem size 643 1283 2563 5123 10243

MPI-processes 1 23 43 83 163
Nodes 1 1 1 8 64

Cycles per hour (CPH) for 163 MPI-processes

Compiler options O2 O2 𝐿 O3 O3 𝐿 Ofast Ofast 𝐿

Intel (CPH) 1061 1074 1094 1109 1110 1109

− +1.2% +3.1% +4.6% +4.6% +4.6%

GNU (CPH) 1140 1142 1140 1142 1141 1140

+7.5% +7.6% +7.4% +7.7% +7.5% +7.5%

AOCC (CPH) 1052 1051 1054 1057 1056 1061

−0.8% −1.0% −0.7% −0.3% −0.5% 0.0%
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 255

≈ 1% compared to the non-LTO option. Again, the AOCC compiler cannot compete
with the Intel reference case and does not represent a viable option for compiler
choices for the FS3D framework.
We chose the Ofast 𝐿 compiler option for the weak scaling representation again. A
decrease in calculated cycles per hour with a growing number of MPI-processes is
expected as this is inevitable for algorithms employing multigrid solvers. A sharp
drop of CPH going from eight to 64 processors on a single node can be observed.
This steep decrease for the weak scaling efficiency can be attributed to the reduced
available memory bandwidth for each processor on a node. An enhancement of the
weak scaling performance can be achieved by allotting only 32 processors per node
represented by the dashed lines in figure 6. This, however, not only increases the
amount of requested nodes but also the required computational resources by a factor
of two. The time when node-to-node-level communication becomes the limiting factor
is depicted by the green plot in figure 6 depicting the amount of assigned processors
per node for the GNU Ofast 𝐿 case with 512 processors2. The maximum performance
is achieved with four processors, however, leaving the remaining processors at idle.
Yet, this issue can be addressed in future endeavors by exploiting OpenMP
parallelization on loop level utilizing all the cores within a node. The update of all
critical routines in this respect within the FS3D framework is an ongoing process and
features considerable potential to further increase the performance on the new HPE
Apollo Hawk system. With the implementation of the tree structured communication
in the multigrid solver other bottlenecks were identified after the transfer to the new
platform. Major parts of the computational load can now be attributed to other parts of
the code where subsequent tracing and performance analysis can help to specifically
target the implementation of OpenMP in the identified routines.

6 Conclusions

Within this study the atomization process of a liquid rivulet at the trailing edge of a
thin plate has been investigated using DNS and the multiphase flow solver FS3D. To
do that, the computational domain has been structured into equally sized blocks and a
no-slip boundary condition between two blocks has been used to discretize the plate
with infinitesimal thickness. To validate the used computational setup simulations
were performed in analogy to experiments performed at ITLR. The inflow conditions
of the simulations were chosen to represent different occurring atomization regimes.
In addition, a grid study revealed the influence and the necessary grid resolution to
reproduce the experimental findings adequately. The results show that the setup is
generally capable of simulating the disintegration of a liquid rivulet at the trailing
edge. Also, the atomization regime changes with the inflow conditions and matches
the experimental results qualitatively well. However, the results show that even with
the highest grid resolution (with more than one billion cells) the very thin lamella at
2 For the plot depicting the processors per node (ppn) the appropriate scale for the 𝑥-axis is depicted
on top, the completed cycles per hour on the right hand side.
256 A. Schlottke, M. Ibach, J. Steigerwald and B. Weigand

bag breakup cannot be resolved. A quantitative comparison of the liquid rivulet’s

geometry at the trailing edge indicates a strong influence on the inlet condition of
the liquid. With the knowledge of the current limitations the authors will focus on
the enhancement of the resolution and a more precise specification of the liquid
inlet condition to match the experimental findings. The detailed insight into the
disintegration process will help to improve the existing empirical models for the
prediction of trailing edge disintegration.
Furthermore, several available compilers and compiler options have been tested
after the transition and adaption of the FS3D code to the new HPE Apollo Hawk
platform at HLRS. The usage of the optimization flags and link-time optimization
during compile time led to a significant performance increase compared to the
previously utilized O2 optimization. For the FS3D benchmark case the compiling
option Ofast with link-time optimization yields the best performance regarding
computational cycles per hour. For several years it has been standard practice to
use the Intel Compiler. A transfer to the usage of the GNU Compiler can achieve
performance gains of up to 36% for strong and 7.5% for weak scaling and should
become the standard for future numerical investigations at Hawk. The availability
of 128 cores per node engenders new challenges that can be addressed by means
of hybrid MPI (on node level) and OpenMP programming (inside each node on
loop level). Nearest neighbor communication (namely topology mapping) could also
ensure optimal system utilization especially for calculation involving multiple nodes.

Acknowledgements The authors kindly acknowledge the High Performance Computing Center
Stuttgart (HLRS) for support and supply of computational resources on the HPE Apollo (Hawk)
platform under the Grant No. FS3D/11142. In addition, the authors kindly acknowledge the financial
support of the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under
grant number WE2549/36-1, WE2549/35-1 and under Germany’s Excellence Strategy - EXC 2075
- 390740016. We also acknowledge the support by the Stuttgart Center for Simulation Science
(SimTech).

References

1. Baggio, M., Weigand, B.: Numerical simulation of a drop impact on a superhydrophobic surface
with a wire. Physics of Fluids 31(11), 112107 (2019)
2. Dumouchel, C.: On the experimental investigation on primary atomization of liquid streams.
Experiments In Fluids 45, 371–422 (2008)
3. Eisenschmidt, K., Ertl, M., Gomaa, H., Kieffer-Roth, C., Meister, C., Rauschenberger, P.,
Reitzle, M., Schlottke, K., Weigand, B.: Direct numerical simulations for multiphase flows: An
overview of the multiphase code FS3D. Journal of Applied Mathematics and Computation
272(2), 508–517 (2016)
4. Ertl, M., Reutzsch, J., Nägel, A., Wittum, G., Weigand, B.: Towards the Implementation of a
New Multigrid Solver in the DNS Code FS3D for Simulations of Shear-Thinning Jet Break-Up
at Higher Reynolds Numbers, p. 269–287. Springer (2017)
5. Ertl, M., Weigand, B.: Analysis methods for direct numerical simulations of primary breakup
of shear-thinning liquid jets. Atomization and Sprays 27(4), 303–317 (2017)
Direct numerical simulation of a disintegrating liquid rivulet at a trailing edge 257

6. Evrard, F., Denner, F., van Wachem, B.: A hybrid Eulerian-Lagrangian approach for simulating
liquid sprays. In: ILASS–Europe 2019, 29th Conference on Liquid Atomization and Spray
Systems. Paris, France (2019)
7. Fest-Santini, S., Steigerwald, J., Santini, M., Cossali, G., Weigand, B.: Multiple drops impact
onto a liquid film: Direct numerical simulation and experimental validation. Computers &
Fluids 214, 104761 (2021)
8. Gepperth, S., Bärow, E., Koch, R., Bauer, H.J.: Primary atomization of prefilming airblast
nozzles: Experimental studies using advanced image proprocess techniques. In: ILASS –
Europe 2014, 26th Annual Conference on Liquid Atomization and Spray Systems. Bremen,
Germany (2014)
9. Gepperth, S., Koch, R., Bauer, H.J.: Analysis and Comparison of Primary Droplet Characteristics
in the Near Field of a Prefilming Airblast Atomizer. In: Proceedings of ASME Turbo Expo
2013: Turbine Technical Conference and Exposition. San Antonio, Texas, USA (2013)
10. Hammit, F., Krzeczkowski, S., Krzyzanowksi, J.: Liquid film and droplet stability consideration
as applied to wet steam flow. Forschung im Ingenieurwesen 47(1) (1981)
11. Hirt, C.W., Nichols, B.D.: Volume of fluid (VOF) Method for the Dynamics of Free Boundaries.
Journal of Computational Physics 39(1), 201–225 (1981)
12. Javed, B., Watanabe, T., Himeno, T., Uzawa, S.: Effect of trailing edge size on the droplets size
distribution downstream of the blade. Journal of Thermal Science and Technology 12(2) (2017)
13. Javed, B., Watanabe, T., Himeno, T., Uzawa, S.: Experimental Investigation of Droplets
Characteristics after the Trailing Edge at Different Angle of Attack. International Journal of
Gas Turbine, Propulsion and Power Systems 9(3), 32–42 (2017)
14. Kim, W.: Study of liquid films, fingers, and droplet motion for steam turbine blading erosion
problem. Ph.D. thesis, University of Michigan, Michigan, USA (1978)
15. Koch, R., Braun, S., Wieth, L., Chaussonnet, G., Dauch, T., Bauer, H.J.: Prediction of primary
atomization using Smoothed Particle Hydrodynamics. European Journal of Mechanics B/Fluids
61(2), 271–278 (2017)
16. Lafaurie, B., Nardone, C., Scardovelli, R., Zaleski, S., Zanetti, G.: Modelling Merging and
Fragmentation in Multiphase Flows with SURFER. Journal of Computational Physics 113(1),
134–147 (1994)
17. Liu, M., Bothe, D.: Numerical study of head-on droplet collisions at high Weber numbers.
Journal of Fluid Mechanics 789, 785–805 (2016)
18. Rayleigh, L.: On The Instability Of Jets. Proc. London Math. Soc. 10, 4–13 (1878)
19. Reutzsch, J., Ertl, M., Baggio, M., Seck, A., Weigand, B.: Towards a Direct Numerical
Simulation of Primary Jet Breakup with Evaporation, chap. 16, p. 243–257. Springer, Cham
(2019)
20. Reutzsch, J., Kochanattu, G.V.R., Ibach, M., Kieffer-Roth, C., Tonini, S., Cossali, G., Weigand,
B.: Direct Numerical Simulations of Oscillating Liquid Droplets: a Method to Extract Shape
Characteristics. In: Proceedings ILASS–Europe 2019. 29th Conference on Liquid Atomization
and Spray Systems (2019)
21. Rider, W.J., Kothe, D.B.: Reconstructing Volume Tracking. Journal of Computational Physics
141(2), 112–152 (1998)
22. Schlottke, A., Weigand, B.: Two-Phase Flow Phenomena in Gas Turbine Compressors with
a Focus on Experimental Investigation of Trailing Edge Disintegration. Aerospace 8(4), 91
(2021). URL https://doi.org/10.3390/aerospace8040091
23. Steigerwald, J., Ibach, M., Reutzsch, J., Weigand, B.: Towards the Numerical Determination of
the Splashing Threshold of Two-component Drop Film Interactions. Springer (2022)
24. Steigerwald, J., Reutzsch, J., Ibach, M., Baggio, M., Seck, A., Haus, B., Weigand, B.: Direct
Numerical Simulation of a Wind-generated Water Wave. Springer (2021)
25. Weber, C.: Zum Zerfall eines Flüssigkeitsstrahles. Zeitschrift für Angewandte Mathematik und
Mechanik 11, 136–154 (1931)
26. Weymouth, G., Yue, D.K.P.: Conservative Volume-of-Fluid method for free-surface simulations
on Cartesian-grids. Journal of Computational Physics 229(8), 2853–2865 (2010)
Numerical Investigation of the Flow and Heat
Transfer in Convergent Swirl Chambers

Florian Seibold and Bernhard Weigand

Abstract Confined swirling flows are a promising technique for cooling applications
since they achieve high heat transfer rates. In such systems, however, an axial
flow reversal can occur, which corresponds to the axisymmetric vortex breakdown
phenomenon.
This report presents a numerical study using Delayed Detached Eddy Simulations
(DDES) in order to analyze the impact of convergent tube geometries on the flow field
and the heat transfer in cyclone cooling systems. For this purpose, a comparison is
drawn for a Reynolds number of 10, 000 and a swirl number of 5.3 between a constant-
diameter tube and four convergent tubes. The latter comprise three geometries with
linearly decreasing diameters yielding convergence angles of 0.42 deg, 0.61 deg and
0.72 deg, respectively. Additionally, a single tube with a hyperbolic diameter decrease
was analyzed.
The results demonstrate that converging tubes enforce an axial and circumferential
flow acceleration. The axial flow acceleration counteracts the flow reversal and thus
was proved capable of suppressing the vortex breakdown phenomenon. Further, the
heat transfer in terms of Nusselt numbers shows a strong dependency on the tube
geometry.

Florian Seibold
Institute of Aerospace Thermodynamics (ITLR), University of Stuttgart, Pfaffenwaldring 31, 70569
Stuttgart, Germany, e-mail: [email protected]
Bernhard Weigand
Institute of Aerospace Thermodynamics (ITLR), University of Stuttgart, Pfaffenwaldring 31, 70569
Stuttgart, Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 259
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_15
260 Florian Seibold and Bernhard Weigand

1 Introduction

In modern gas turbines, the maximum turbine entry temperature limits the overall
engine efficiency. As a consequence, this temperature is increased more and more
whereby the cooling system for turbine blades becomes one of the most critical parts
in the design process. The so called cyclone cooling or swirl cooling is an innovative
cooling technique that can be applied for instance in the leading edge of turbine
blades [2]. Cyclone cooling systems consist of a vortex chamber, which is an internal
flow passage, and one or more tangential inlets that induce the swirling motion. This
technique promises high heat transfer rates but is also accompanied by high pressure
losses [3]. Chang and Dhir [4] identified two major mechanisms that explain the heat
transfer enhancement in such swirl chambers: On the one hand, the high maximum
axial velocity close to the wall results in high heat fluxes at the wall. On the other
hand, a high turbulence level improves the fluid mixing.
An important phenomenon of swirling flows is the so called vortex breakdown,
which is defined as an abrupt change in the structure of the core flow [5]. The vortex
breakdown manifests itself in different types: spiral, double-helix and axisymmetric
form. The axisymmetric vortex breakdown is characterised by a stagnation point
followed by an axial flow reversal [6]. This flow reversal significantly influences the
flow field and, thus, also the heat transfer performance of cyclone cooling systems.
Although a great deal of research has been conducted on the vortex breakdown
phenomenon, it is not yet fully understood. Hence, several explanations exist among
which the most popular one is based on the propagation of waves. In this regard, Squire
[7] and Benjamin [8] hypothesized the existence of a critical flow state that separates
a supercritical state from a subcritical one. In a subcritical state, disturbances can
propagate upstream and downstream whereas in a supercritical state only downstream
propagation is possible. As a consequence of this categorization, Escudier et al. [9]
observed a strong dependency of subcritical flows on the tube outlet conditions.
In contrast, supercritical flows showed no influence at all. The authors related this
outcome to the axial flow reversal that enables disturbances to propagate upstream
and, hence, are able to affect the entire flow field within the swirl tube. Accordingly,
the supercritical state represents a more robust flow that is insensitive to disturbances
from downstream [10].
In the past, research on swirl chambers was mainly conducted on constant-diameter
tubes. Investigations on convergent tubes are rare and mainly focused on the heat
transfer enhancement as for instance in [11, 12] or the temperature separation in
Ranque–Hilsch vortex tubes [13, 14]. However, there is no detailed study of the flow
pattern in convergent vortex tubes and its impact on the heat transfer.
This report aims to summarize selected results from the conducted research on
convergent vortex tubes. For this purpose, the flow field and the heat transfer in a
constant-diameter tube is analyzed and systematically compared to four convergent
geometries with different diameter variations.
Flow and Heat Transfer in Convergent Swirl Chambers 261

2 Geometry

Figure 1 depicts the computational domain of the here presented simulations rep-
resenting an upscaled generic model of a cyclone cooling system. The fluid enters
the domain through two tangential inlets, which feature a height of ℎ = 5 mm and a
length of 15 times of their hydraulic diameter. After the inlet area with a diameter
of 𝐷 0 = 50 mm, a convergent section follows with an axial length of 20𝐷 0 . The
geometry is in accordance with the experimental setup of Biegger et al. [16]. The
subscript 0 denotes values at the axial location 𝑧 = 0.

Fig. 1: Sketch of the computational domain [15]

Based on these geometrical constraints, different non-dimensional numbers can

be defined. First of all, the local Reynolds number 𝑅𝑒 is based on the local axial bulk
∫𝑅
velocity 𝑢 𝑧 = 𝑅22 0 𝑟 𝑢 𝑧 (𝑟, 𝑧) d𝑟 and the local tube diameter 𝐷

𝑢𝑧 𝐷
𝑅𝑒 = . (1)
𝜈
Here, 𝜈 represents the kinematic viscosity. In addition, an inflow Reynolds number
𝑅𝑒 0 can be defined based on the tube inlet diameter 𝐷 0 and its corresponding axial
∫𝑅
bulk velocity 𝑢 𝑧0 = 𝑅22 0 𝑟 𝑢 𝑧 (𝑟, 𝑧=0) d𝑟 yielding

𝑢 𝑧0 𝐷 0 4𝑚¤
𝑅𝑒 = = . (2)
𝜈 𝜋𝐷 0 𝜇
Here, 𝑚¤ and 𝜇 denote the mass flow rate and the dynamic viscosity, respectively. The
inflow Reynolds number 𝑅𝑒 0 , defined by Eq. (2), can be evaluated prior to simulation
in order to determine the operating point of the device.
Moreover, a dimensionless swirl number 𝑆 can be defined as the ratio of axial flux
of circumferential momentum 𝐼¤𝜑 divided by the local tube radius 𝑅 and the axial flux
of axial momentum 𝐼¤𝑧 [17]
262 Florian Seibold and Bernhard Weigand

∫𝑅
𝜌𝑢 𝜑 𝑢 𝑧 2𝜋𝑟 2 d𝑟
𝐼¤𝜑 0
𝑆= = . (3)
𝑅 𝐼¤𝑧 ∫𝑅
𝑅 𝜌𝑢 2𝑧 2𝜋𝑟d𝑟
0

Here, 𝜌 denotes the fluid density. Equation (3) cannot be evaluated in advance since
the velocity profiles are unknown. However, when assuming a uniform velocity at the
inlets, a geometrical swirl number 𝑆 𝑔𝑒𝑜 can be determined prior to simulation
 
𝑅0 − ℎ2 𝜋𝑅02
𝑆 𝑔𝑒𝑜 = . (4)
𝑅0 2𝑤ℎ
All simulations were conducted for 𝑅𝑒 0 = 10, 000 and 𝑆 𝑔𝑒𝑜 = 5.3.
In the present report, five different geometries are investigated: One vortex tube
with a constant cross-section (𝛽 = 0 deg) is used for validation and as reference
for comparison. Moreover, three chambers with linearly decreasing diameters are
analyzed. these correspond to convergence angles 𝛽 of 0.42 deg, 0.61 deg and 0.72 deg
and area ratios 𝐴𝑜𝑢𝑡/𝐴0 of 1/2, 1/3 and 1/4, respectively. Finally, one geometry with
a hyperbolically decreasing diameter is investigated. The latter enforces a linear
increase in the local Reynolds number when assuming a constant bulk density 𝜌
yielding
𝐷0
𝐷= . (5)
1 + 𝐿𝑧
The geometrical features are summarized in Tab. 1.

Table 1: Investigated geometries

 
geometry 𝐷𝑜𝑢𝑡 𝐴𝑜𝑢𝑡 /𝐴0 𝛽 deg

1 50 mm 1 0
2 35.4 mm 1/2 0.42
3 28.9 mm 1/3 0.61
4 25 mm 1/4 0.72
5 25 mm 1/4 hyperbolic
Flow and Heat Transfer in Convergent Swirl Chambers 263

3 Numerical setup

The here presented numerical calculations are carried out as compressible Delayed
Detached Eddy Simulations (DDES) using the open source code OpenFOAM version
6. The DDES approach was introduced by Spalart et al. [18] as hybrid model that
applies a Reynolds Averaged Navier Stokes (RANS) model close to the wall and a
Large Eddy Simulation (LES) in the free stream.
Originally, Spalart et al. [19] designed this hybrid approach by modifying the
one-equation turbulence model from Spalart and Allmaras [20]
 2
D𝜈˜ 1 h
2
i 𝜈˜
= 𝑐 𝑏1 𝑆 𝜈˜ +
˜ ∇ · (𝜈 + 𝜈𝑡 ) ∇𝜈˜ + 𝑐 𝑏2 (∇𝜈)
˜ − 𝑐 𝑤1 𝑓𝑤 . (6)
D𝑡 𝜎 𝑑˜
𝜈˜ = 𝜈𝑡/ 𝑓𝑣1 represents an effective eddy viscosity and 𝑐 𝑏1 , 𝑐 𝑏2 , 𝑐 𝑤 and 𝜎 denote model
coefficients. Further, 𝑑˜ is the so-called DES-limiter

𝑑˜ = min{𝑑; 𝐶𝐷𝐸𝑆 Δ} . (7)

Here, 𝑑 represents the wall distance, Δ = max{Δ 𝑥 ; Δ 𝑦 ; Δ𝑧 } denotes the local grid
spacing and 𝐶𝐷𝐸𝑆 = 0.65 is a constant.
The definition of the DES-limiter 𝑑˜ in Eq. (7) allows to use the turbulence model
as hybrid model. In the wall distant region, where 𝑑 > 𝐶𝐷𝐸𝑆 Δ, the model operates
as Subgrid Scale (SGS) model for LES whereas it transforms into a RANS model
on approaching the wall, where 𝑑 < 𝐶𝐷𝐸𝑆 Δ [21]. This hybrid approach is termed
Detached Eddy Simulation (DES).
The original DES formulation in Eqs. (6, 7) comes along with a major drawback:
Within the RANS regime, a coarse grid resolution parallel to the wall is required
whereas a fine resolution is necessary in the LES regime. If this requirement is
violated in the RANS regime by a too small cell spacing, the model switches into the
LES mode too close to the wall. Concurrently, the wall-normal cell spacing is not yet
fine enough to resolve turbulent fluctuations. As a result, unphysical results can arise
such as modeled stress depletion and grid induced separation [21]. Therefore, Spalart
et al. [18] addressed this drawback by modifying the definition of the DES-limiter
using an additional blending function 𝑓 𝑑

𝑑˜ = 𝑑 − 𝑓 𝑑 max{0; 𝑑 − 𝐶𝐷𝐸𝑆 Δ} . (8)

The blending function 𝑓 𝑑 in Eq. (8) equals zero within the boundary layer and rapidly
approaches one in the free stream area. The exact equation is given in [18]. In contrast
to the original DES formulation in Eqs. (6,7), the modified DES-limiter 𝑑˜ in Eq. (8)
does not depend on the local grid size Δ but on the flow field solution [21]. This
modified model is termed Delayed DES (DDES). For a more detailed description,
the reader is referred e.g. to Fröhlich and von Terzi [21].
264 Florian Seibold and Bernhard Weigand

Finally, a constant turbulent Prandtl number 𝑃𝑟 𝑡 was prescribed in order to close

the equation system. In an experimental review, Kays [22] stated that 𝑃𝑟 𝑡 essentially
takes a constant value of 0.85 within a turbulent boundary layer. Based on this
conclusion, 𝑃𝑟 𝑡 = 0.85 was used in the here presented simulations.
All simulations were conducted using a finite volume approach. The Pressure-
Implicit-Split-Operator (PISO) was applied as pressure corrector with two outer
loops within every timestep. Convective and diffusive fluxes were discretized with
a second-order central scheme and the time dependency was approximated using a
second-order backward scheme. Besides, the numerical mesh consists of a structured
O-grid with 16.5 · 106 hexahedral cells. The boundary layer at the wall is fully resolved
with a maximum non-dimensional height of the first cell layer 𝑦 +1 = 𝑢 𝜏 𝑦 1 /𝜈 < 1.5.
√︃
𝜕𝑢
Here, 𝑦 1 and 𝑢 𝜏 = 𝜈 𝜕𝑦 𝑦=0
denote the height of the first cell layer at the wall and
the wall shear stress, respectively. Further mesh details are summarized in Tab. 2.
In the course of the simulations, the following boundary conditions were applied
for the fluid medium air: At the inlets, a fixed mean velocity was prescribed and
superimposed by random velocity fluctuations with a RMS-value of 15% of the
mean flow. The walls represent a no-slip boundary in combination with an effective
turbulent viscosity of 𝜈˜ = 0. At the outlet, a fixed mean pressure of 105 Pa was
prescribed. Isothermal boundary conditions were applied for the wall temperature
𝑇𝑊 = 293 K as well as for the inflow temperature 𝑇𝑖𝑛 = 333 K.
The velocity, pressure, and temperature fields were initialized with values from a
steady-state simulation. Then, the transient simulation time was defined based on the
flow through time (FFT) of the convergent tube segment 𝐿/𝑢 𝑧 . First, the calculations
run for 35 FFTs prior to averaging and then for another 10 FFTs in order to obtain the
mean quantities. The timestep Δ𝑡 was adapted automatically during run-time in order
to maintain a Courant number of 𝑢Δ𝑡/Δ = 0.9.

Table 2: Summary of mesh details including the total amount of cells, the height of
the first cell layer at the wall and the center cell size [15]
h i
𝑦1 10−5 m 10−4 m

 
Geometry Cells Δ 𝑥 , Δ 𝑦 , Δ𝑧 𝑐𝑒𝑛𝑡𝑒𝑟

0 deg 16.5 · 106 1.7 (6.6, 6.6, 11)

0.42 deg 16.5 · 106 1.7-3 (4.7-6.6, 4.7-6.6, 11)
0.61 deg 16.5 · 106 1.7-2.4 (3.8-6.6, 3.8-6.6, 11)
0.72 deg 16.5 · 106 1.7-2.1 (3.3-6.6, 3.3-6.6, 11)
hyperbolic 16.5 · 106 1.7-2.1 (3.3-6.6, 3.3-6.6, 11)

In the following, all results represent mean values. Figure 2 depicts a comparison of
circumferential and axial velocities between numerical and experimental outcomes for
a constant-diameter tube (𝛽 = 0 deg). The experimental data originates from Biegger
Flow and Heat Transfer in Convergent Swirl Chambers 265

[23], who measured flow velocities using the stereo particle image velocimetry (PIV)
technique. In Fig. 2, the radial coordinate 𝑟 and the velocity components 𝑢 𝜑 and 𝑢 𝑧
are normalized by the inlet tube radius 𝑅0 and the corresponding axial bulk velocity
𝑢 𝑧0 , respectively. Hereafter, this type of scaling is named global normalization since it
depends on values from the axial location 𝑧 = 0. In general, the results in Fig. 2 prove
overall good accuracy of the numerical prediction. Only some smaller deviations
occur for 𝑢 𝑧 in the tube center and for the maximum value of 𝑢 𝜑 . Furthermore,
Seibold et al. [24] and Seibold and Weigand [15] investigated the turbulence energy
spectrum and reported that the simulation reproduces the correct Kolmogorov slop
of −5/3 for the resolved scales. Therefore, the numerical setup achieves reasonable
turbulence behavior and the grid resolution proved to be fine enough to resolve
the large scales. Moreover, Biegger [23] carried out a more detailed validation for
the numerical DDES setup. For this purpose, the author conducted an extensive
mesh study for a channel flow without swirl and compared the outcomes to a Direct
Numerical Simulation (DNS) from literature. Based on this mesh study, Biegger
[23] simulated swirling flows with the same numerical setup and obtained overall
good agreement with experimental data in terms of mean velocity components and
turbulent kinetic energy.

PIV results [16]

DDES
uϕ /uz0 -6.5 0 6.5
1

0.5
r/R0

0
−0.5

−1
1.9 4.4 6.9 9.4 11.9 14.4 16.9
z/D0
(a) Circumferential velocity
uz /uz0 -2.5 0 2.5
1

0.5
r/R0

0
−0.5

−1
1.9 4.4 6.9 9.4 11.9 14.4 16.9
z/D0
(b) Axial velocity

Fig. 2: Comparison of experimental and numerical results for the geometry 𝛽 = 0 deg.
The numerical data was originally published in [15, 24]
266 Florian Seibold and Bernhard Weigand

4 Results

4.1 Flow field

β = 0 deg. β = 0.72 deg.

β = 0.42 deg. hyperbol.
β = 0.61 deg.
12

10

8
uz [m/s]

2
0 2.5 5 7.5 10 12.5 15 17.5 20
z/D0

Fig. 3: Local axial bulk velocity 𝑢 𝑧

Figure 3 depicts the values of the local axial bulk velocity 𝑢 𝑧 for the here
investigated swirl tubes. As might be expected in advance, the constant-diameter tube
(𝛽 = 0 deg.) shows almost constant values. Only the cooling of the gas at the cold
wall causes a slight decrease in density that leads to a minor deceleration. On the
contrary, all converging tubes feature a flow acceleration in axial direction in order to
satisfy mass conservation. Thus, the axial bulk velocity significantly increases and
leads to a favorable pressure gradient [15].
A more detailed analysis of the flow pattern is illustrated in Fig. 4 showing both
axial and circumferential velocity distributions. Here, the radial coordinate 𝑟 and
the velocity components 𝑢 𝜑 and 𝑢 𝑧 are normalized using the local tube radius 𝑅
and the local axial bulk velocity 𝑢 𝑧 from Fig. 3, respectively. In the following, this
type of normalization, which is based on local bulk values, is referred to as local
normalization.
The circumferential velocity component 𝑢 𝜑 is depicted in Fig. 4a with locally
normalized values. These results show that 𝑢 𝜑 is the largest velocity component
of the system for the here investigated swirl number. However, the corresponding
circumferential velocities of the constant diameter tube (𝛽 = 0 deg.) decline in axial
direction due to dissipation effects. In case of convergent geometries, an additional
effect counteracts this decrease that is an angular flow acceleration caused by the
conservation of circumferential momentum. This acceleration is included in the local
normalization using 𝑢 𝑧 in Fig. 4a.
Flow and Heat Transfer in Convergent Swirl Chambers 267

β = 0 deg. β = 0.72 deg

β = 0.42 deg hyperbol.
uϕ /uz -8.5 0 8.5 β = 0.61 deg
1

0.5
r/R

0
−0.5

−1
0 2.5 5 7.5 10 12.5 15 17.5 20
z/D0
(a) Circumferential velocity
uz /uz -3 0 3
1

0.5
r/R

0
−0.5

−1
0 2.5 5 7.5 10 12.5 15 17.5 20
z/D0
(b) Axial velocity

Fig. 4: Locally normalized velocity field

Further, the normalized velocity profiles of different geometries obviously coincide

in the middle section of the tube. In the upstream part, the flow pattern slightly varies
since the circumferential velocity has not yet fully evolved. Additionally, some smaller
deviations are apparent towards the tube outlet that are caused by an increased impact
of dissipation in convergent geometries. Nevertheless, the characteristic flow pattern
remains unaffected for the here investigated cases [15, 24].
Moreover, a change in shape is evident when comparing 𝑢 𝜑 of a single geometry
at different axial positions. In general, the circumferential velocity in a swirl tube
matches a Rankine vortex. Such a vortex consists of a solid body rotation in the
center, a potential vortex in the outer region and a boundary layer at the wall. In the
upstream part of the tube, the solid body rotation dominates the flow pattern, which
is indicated by a linear increase of 𝑢 𝜑 . Further downstream, the solid body vortex
shrinks in favor of the potential vortex.

Investigating the axial velocity 𝑢 𝑧 in Fig. 4b the constant-diameter tube shows

a pronounced backflow in the center that extends to the outlet. This phenomenon
corresponds to the axisymmetric vortex breakdown. The velocity magnitude of the
reversed flow increases in downstream direction while its area shrinks.
268 Florian Seibold and Bernhard Weigand

Convergent cross-sections impact considerably on the flow pattern of 𝑢 𝑧 in Fig. 4b.

The axial flow acceleration counteracts the flow reversal by imposing a favorable
pressure gradient [15]. In case of 𝛽 = 0.42 deg, a slight acceleration reaches stagnant
flow in the tube center. The vortex tubes with 𝛽 ≥ 0.61 deg as well as the hyperbolic
geometry cause a sufficient strong acceleration to reach purely positive axial flow.
Hence, convergent swirl chambers are able to suppress the axisymmetric vortex
breakdown if the flow acceleration is strong enough [15, 24]. Concurrently, purely
positive axial flow at the tube outlet indicates a supercritical flow state according
to Escudier et al. [9]. Thus, the upstream propagation of disturbances is prevented
and the flow within the convergent tubes is made insensitive to disturbances from
downstream. As a result, a more robust flow state is obtained, which is of crucial
importance for the design process of such systems.

4.2 Heat transfer

The cooling capability of cyclone cooling systems can be assessed by evaluating the
Nusselt number 𝑁𝑢
− 𝜕𝑇 | 𝑤 𝐷 ℎ𝐷
𝑁𝑢 = 𝜕𝑛 = . (9)
𝑇𝑤 − 𝑇𝑟 𝑒 𝑓 𝑘
Here, 𝑇, ℎ and 𝑘 denote the temperature, the heat transfer coefficient and the fluid’s
thermal conductivity, respectively. The indices 𝑤 and 𝑟𝑒 𝑓 indicate wall and reference
values.
The reference temperature 𝑇𝑟 𝑒 𝑓 in Eq. (9) is of crucial importance for the evaluation
of the Nusselt number. Nusselt numbers can only be compared to other investigations
with different thermal boundary conditions if the adiabatic wall temperature is used
for 𝑇𝑟 𝑒 𝑓 . In general, this value is difficult to obtain. However, it can be approximated in
swirling flows by assuming an adiabatic compression of the fluid from the centerline
(index 𝑐) to the wall [25]
  𝜅−1
𝑝𝑤 𝜅
𝑇𝑟 𝑒 𝑓 = 𝑇𝑐 . (10)
𝑝𝑐
Here, 𝜅 represents the isentropic exponent of the fluid medium air.

Figure 5 depicts the numerically obtained heat transfer results as circumferentially

averaged Nusselt numbers 𝑁𝑢. All of these values consider the local tube diameter 𝐷
in the definition of Eq. (9) as well as the local reference temperature 𝑇𝑟 𝑒 𝑓 according to
Eq. (10). In addition, experimental results are included in the diagram for comparison.
These were measured using the transient liquid crystal technique in combination with
a reference temperature that was also based on the adiabatic compression according
to Eq. (10). The complete experimental procedure was reported in [23, 27].
Flow and Heat Transfer in Convergent Swirl Chambers 269
num 0 deg. 0.61 deg. hyperbol
exp 0.42 deg. 0.72 deg.
250

200

150
Nu

100

50

0
0 2.5 5 7.5 10 12.5 15 17.5 20
z/D0 [-]

Fig. 5: Axial distribution of 𝑁𝑢 from numerical (solid lines) and experimental

investigations (dashed lines).

The results in Fig. 5 show overall good agreement between experimental and
numerical data. Close to the tangential inlets, all curves feature high values that decay
monotonically towards the tube outlet. This decay is caused by the decreasing swirl
intensity.
When comparing different geometries, obvious differences are apparent in the
downstream part of the tubes. There, the results show a distinct degression when
increasing the convergence angle 𝛽. However, caution is required when interpreting
these results. The aforementioned drop does not indicate a degression of the heat
transfer coefficient ℎ but is largely caused by the varying tube diameter 𝐷 in Eq. (9).
Consequently, the varying tube surface impacts considerably on the Nusselt numbers
in Fig. 5.

Furthermore, the circumferentially averaged Nusselt numbers can be scaled by a

Nusselt correlation from Gnielinski for a fully developed turbulent pipe flow without
swirl, which is presented in [26] as:

𝜉    23 
𝑅𝑒𝑃𝑟
1 + 𝐷 
8
 
𝑁𝑢 𝐺 = √︃   (11)
2  𝐿 
1 + 12.7 𝜉8 𝑃𝑟 3 − 1 
 

with
−2
𝜉 = 1.8 log10 𝑅𝑒 − 1.5 . (12)
The scaled results are depicted in Fig. 6 using a local normalization. In particular,
this means that all 𝑁𝑢-values are normalized by the Gnielinski-correlation from Eqs.
(11, 12), which are based on local values of the tube diameter 𝐷 and the Reynolds
270 Florian Seibold and Bernhard Weigand

number 𝑅𝑒. The normalized results then indicate a heat transfer enhancement over a
pure pipe flow without swirl at the same local flow conditions. The results in Fig. 6

num 0 deg. 0.61 deg. hyperbol

exp 0.42 deg. 0.72 deg.
7

5
N u/N uG

0
0 2.5 5 7.5 10 12.5 15 17.5 20
z/D0 [-]

Fig. 6: Axial distribution of locally normalized 𝑁𝑢 results from numerical (solid

lines) and experimental investigations (dashed lines)

show an even more pronounced drop of 𝑁𝑢 at the tube outlet in case of convergent
geometries. This effect is evoked by the axial flow acceleration that yields higher
values for 𝑁𝑢 𝐺 at the tube outlet in case of convergent geometries. As a result, the
heat transfer enhancement over a non-swirling pipe flow vanishes at the tube outlet
for 𝛽 ≥ 0.61 deg. as well as for the hyperbolic case.

In addition to the local heat transfer, global Nusselt numbers 𝑁𝑢 can also be
evaluated. For this purpose, globally averaged values of ℎ, 𝐷 and 𝑘 are introduced in
Eq. (9). These are calculated as integral means
∫ 𝐿
1
𝜙= 𝜙 d𝑧 . (13)
𝐿 0

Moreover, a proper mean value of the Gnielinski-correlation 𝑁𝑢 𝐺 is required for

normalization of the results. For this purpose, a local type of normalization is applied
by introducing integral mean values of 𝐷 and 𝑅𝑒 in Eqs. (11, 12). The corresponding
results are presented Tab. 3.
The outcomes in Tab. 3 again show overall good agreement between numerics and
experiments with a maximum deviation of less than 9%. These global values further
show the large potential of cyclone cooling systems by predicting high heat transfer
enhancements up to a factor of four compared to a pure pipe flow without swirl.
Flow and Heat Transfer in Convergent Swirl Chambers 271

Table 3: Globally averaged heat transfer

𝑁𝑢 𝑁 𝑢/𝑁 𝑢𝐺
𝛽
num exp num exp

0 deg. 140.8 135.0 3.81 3.65

0.42 deg. 122.6 134.5 2.98 3.27
0.61 deg. 103.8 111.6 2.38 2.55
0.72 deg. 102.7 104.5 2.25 2.29
hyperbolic 105.7 - 2.20 -

5 Usage of computational resources

This section contains an overview of the used computer resources, which are
summarized in Tab. 4. Furthermore, the speed-up of the conducted simulations is
depicted in Fig. 7 for a calculation with 16.5 · 106 cells. The diagram shows both an
ideal and a real speed-up. The latter was determined by comparing each simulation
to a calculation with 20 cores (1 node). With a small number of cores used in
parallel, the curve rises almost linearly and then flattens out. Based on these results, a
parallelization of 660 cores (33 nodes) is selected, which corresponds to 25, 000 cells
per core. This degree of parallelization takes advantage of massive parallelization but
does not waste computational resources.

ideal
real
50

40
speed-up

30

20

10

0
0 200 400 600 800 1000
cores

Fig. 7: Speed-up of numerical simulations

272 Florian Seibold and Bernhard Weigand

Table 4: Usage of resources

platform processors grid cells wall time software

ForHLR II 660 16.5 × 106 18-36 days OpenFOAM 6

6 Summary and conclusion

Convergent swirl tubes for cyclone cooling applications were investigated numerically
for a Reynolds number of 10, 000 and a geometrical swirl number of 5.3. The
numerical setup of the constant-diameter tube was validated with good accuracy. The
same geometry was further compared to three tubs that featured linearly decreasing
diameter distributions with convergence angles of 0.42 deg, 0.61 deg and 0.72 deg,
respectively. Additionally, one geometry with a hyperbolic diameter decrease was
analyzed. The results were investigated in terms of flow field and heat transfer.
The constant-diameter tube showed a pronounced flow reversal in the tube center
that corresponds to the axisymmetric vortex breakdown. Converging swirl chambers
enforced a circumferential and axial flow acceleration. Despite this influence, the
circumferential velocity retained its characteristic profile. The axial flow acceleration
counteracted the flow reversal in the center and reached unidirectional axial velocities
if the acceleration was strong enough. Consequently, convergent tube geometries
proved to be capable of suppressing the vortex breakdown phenomenon. Most
importantly, these convergent tubes reached a more robust flow that was insensitive to
disturbances from the tube outlet. Furthermore, the heat transfer showed overall good
agreement with experimental results and reached high heat transfer enhancements
up to a factor of four. A significant drop of Nusselt numbers was evident in case
of convergent tube geometries. The varying tube diameter, which was used as
characteristic length in the definition of the Nusselt number, turned out to be the main
factor for this decline.

Conflict of interest

The authors declare that they have no conflict of interest.

Acknowledgements The authors would like to acknowledge the funding of this project by the
German Research Foundation (DFG) under Grant No. WE 2549/38-1. The authors also thank the
Steinbuch Centre for Computing (SCC) for supply of computational time on the ForHLR II platform.
Flow and Heat Transfer in Convergent Swirl Chambers 273

References

1. Kreith, F., Margolis, D. (1959) Heat transfer and friction in turbulent vortex flow. Applied
Scientific Research 8:457–473
2. Glezer, B., Moon, H. K., O’Connell, T. (1996) A novel technique for the internal blade cooling.
In: Proceedings of the ASME 1996 International Gas Turbine and Aeroengine Congress and
Exhibition, Paper No. 96-GT-181, Birmingham, UK
3. Ligrani, P. M., Oliveira M. M., Blaskovich, T. (2003) Comparison of heat transfer augmentation
techniques. AIAA Journal 41(3):337–362
4. Chang, T., Dhir, V. K. (1995) Mechanisms of heat transfer enhancement and slow decay of
swirl in tubes using tangential injection. International Journal of Heat and Fluid Flow 16:78–87
5. Sarpkaya, T. (1971) Vortex breakdown in swirling conical flows. AIAA Journal 9(9):1792–1799
6. Escudier, M. P., Keller, J. J. (1985) Recirculation in swirling flows: A Manifestation of vortex
breakdown. AIAA Journal 23:111–116
7. Squire, H. B. (1960) Analysis of the vortex breakdown phenomenon, part 1. Department Report
No. 102, Imperial College of Science and Technology Aeronautics
8. Benjamin, T. B. (1962) Theory of vortex breakdown phenomenon. Journal of Fluid Mechanics
14:593–629
9. Escudier, M. P., Nickson , A. K., Poole, R. J. (2006) Influence of outlet geometry on strongly
swirling turbulent flow through a circular tube. Physics of Fluids 18:125103
10. Bruschewski, M., Grundmann, S., Schiffer, H.-P. (2020) Considerations for the design of swirl
chambers for the cyclone cooling of turbine blades and for other applications with high swirl
intensity. International Journal of Heat and Fluid Flow 86:108670
11. Ling, J. P. C. (2005) Development of heat transfer measurement techniques and cooling
strategies for gas turbines. PhD thesis, University of Oxford, UK
12. Yang, C. S., Jeng, D. Z., Yang, Y.-J., Chen, H.-R., Gau, C. (2011) Experimental study of
pre-swirl flow effect on the heat transfer process in the entry region of a convergent pipe.
Experimental Thermal and Fluid Science 35:73–81
13. Rafiee, S. E., Sadeghiazad, M. M., Mostafavinia, N. (2015) Experimental and numerical
investigation on effect of convergent angle and cold orifice diameter on thermal performance of
convergent vortex tube. Journal of Thermal Science and Engineering Applications 7(4):041006
14. Rafiee, S. E., Sadeghiazad, M. M. (2017) Efficiency evaluation of vortex tube cyclone separator.
Applied Thermal Engineering 114:300–327
15. Seibold, F., Weigand, B. (2021) Numerical analysis of the flow pattern in convergent vortex
tubes for cyclone cooling applications. International Journal of Heat and Fluid Flow 90:108806
16. Biegger, C., Sotgiu, C., Weigand, B. (2015) Numerical investigation of flow and heat transfer
in a swirl tube. International Journal of Thermal Sciences 96:319–330
17. Gupta, A. K., Lilley, D. G., Syred, N. (1984) Swirl Flows. Energy and Engineering Science
Series, Abacus Press
18. Spalart, P. R., Deck, S., Shur, M. L., Squires, K. D., Strelets, M. Kh., Travin, A. (2006) A
new version of detached-eddy simulation, resistant to ambigous grid densities. Theoretical and
Computational Fluid Dynamics 20:181–195
19. Spalart, P. R., Jou, W.-H., Strelets, M., Allmaras, S. R. (1997) Comments on the feasibility of
LES for wings, and on hybrid RANS/LES approach. In: Advances in DNS/LES
20. Spalart, P. R., Allmaras, S. R. (1994) A one-equation turbulence model for aerodynamic flows.
La Recherche Aerospatiale 1:5–21
21. Fröhlich, J., von Terzi, T. (2008) Hybrid LES/RANS methos for simulation of turbulent flows.
Progress in Aerospace Sciences 44:349–377
22. Kays, W. M. (1994) Turbulent Prandtl number – Where are we?. Journal of Heat Transfer
116(2):284–295
23. Biegger, C. (2017) Flow and heat transfer investigations in swirl tubes for gas turbine blade
cooling. PhD thesis, Institute of Aerospace Thermodynamics, University of Stuttgart, Germany
24. Seibold, F., Weigand, B., Marsik, F. and Novotny, P. (2017) Thermodynamic stability condition
of swirling flows in convergent vortex tubes. In: Proceedings of the 12th International Gas
Turbine Conference, Tokyo, Japan
274 Florian Seibold and Bernhard Weigand

25. Kobiela, B. (2014) Wärmeübertragung in einer Zyklonkülkammer einer Gasturbinenschaufel.

PhD thesis, Institute of Aerospace Thermodynamics, University of Stuttgart, Germany
26. VDI-Wärmeatlas (2013) 11th edn., Springer Berlin Heidelberg
27. Biegger, C., Weigand, B. (2015) Flow and heat transfer measurements in a swirl chamber with
different outlet geometries. Experiments in Fluids 56(4)
On the validity of the linear Boussinesq
hypothesis for selected internal cooling features of
gas turbine blades

Philipp Wellinger and Bernhard Weigand

Abstract In order to further decrease fuel consumption and pollutant emission of

modern, highly efficient gas turbine blades, the internal cooling air consumption
needs to be reduced. At the same time, the solid temperature must not exceed a
critical level to avoid thermal damage of the turbine blades. Different cooling methods
such as impingement cooling, pin fins or ribs are used to increase the internal heat
transfer and, thus, to decrease the material temperatures. In order to improve local
cooling air consumption the flow field and temperature distribution must be known.
High fidelity CFD like LES are extremely difficult or even not feasible for such
complex geometries and high Reynolds numbers. Therefore, RANS models must be
used accepting a higher degree of inaccuracy. In this study, LES of selected cooling
features are conducted for a better understanding of the validity of RANS models
which are based on the linear Boussinesq hypothesis. The results given below are part
of the research project “Improvement of the Numerical Computation of the Conjugate
Heat Transfer in Turbine Blades” supported by the Siemens Clean Energy Center and
the German Federal Ministry for Economic Affairs and Energy under grant number
03ET7073F.

Philipp Wellinger
Institute of Aerospace Thermodynamics (ITLR)
University of Stuttgart, Pfaffenwaldring 31 70569 Stuttgart,
e-mail: [email protected]
Bernhard Weigand
Institute of Aerospace Thermodynamics (ITLR)
University of Stuttgart, Pfaffenwaldring 31 70569 Stuttgart,
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 275
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_16
276 Philipp Wellinger and Bernhard Weigand

1 Introduction

Many of the RANS (Reynolds-averaged Navier–Stokes) turbulence models used

today are based on the linear Boussinesq hypothesis (LBH) assuming a linear relation
between the unknown Reynolds stresses and the mean strain rate tensor. The Reynolds
stresses are expressed by
2
− 𝜌 𝑢 𝑖′𝑢 ′𝑗 = 2𝜇𝑡 𝑆𝑖 𝑗 − 𝑘 𝜌𝛿𝑖 𝑗 (1)
3

where 𝜌 is the density, 𝑢 𝑖′ the velocity fluctuation, 𝜇𝑡 the eddy viscosity, 𝑆𝑖 𝑗 the mean
strain rate tensor, 𝑘 the turbulent kinetic energy and 𝛿𝑖 𝑗 denotes the Kronecker delta.
A main assumption of the LBH is that the eigenvectors of the Reynolds stress tensor
and the strain rate tensor are aligned. Schmitt [1] introduced a validity parameter 𝜌 𝑅𝑆
defined as
𝑅𝑖 𝑗 : 𝑆 𝑖 𝑗
𝜌 𝑅𝑆 = (2)
𝑅𝑖 𝑗 𝑆 𝑖 𝑗

to quantify the misalignment of the eigenvectors. Here, 𝑅𝑖 𝑗 = 𝑢 𝑖′𝑢 ′𝑗 − 2/3𝑘𝛿𝑖 𝑗 defines

the anisotropic stress tensor, 𝐴𝑖 𝑗 : 𝐵𝑖 𝑗 represents the Frobenius inner product and
√︃
𝐴𝑖 𝑗 = 𝐴𝑖 𝑗 : 𝐴𝑖 𝑗 . The validity parameter 𝜌 𝑅𝑆 ranges between 0 and 1. Values
of 1 indicate perfect alignment between the eigenvectors of the two tensors. If 𝜌 𝑅𝑆
tends to zero a large misalignment occurs and the LBH is violated. Schmitt [1]
suggested 𝜌 𝑅𝑆 > 0.86 as a soft limitation of the validity of the LBH and concluded
that the hypothesis fails for even very simple test cases like planar channels. In
order to better understand the validity of Eq. (1) for various flow scenarios, 𝜌 𝑅𝑆 is
evaluated for several more sophisticated test cases. Since an accurate prediction of
𝑅𝑖 𝑗 and 𝑆𝑖 𝑗 is needed, Large Eddy Simulations (LES) are conducted and validated
with experimental or DNS (Direct Numerical Simulations) data.
The complex flow of the internal cooling system of a gas turbine blade is divided
up into several test cases to investigate individual flow features separately. Test cases
for a planar channel, two different periodic pin fin arrays and a periodic ribbed channel
are chosen. It is important that suitable and accurate validation data for the Reynolds
stresses and the velocity components are available to validate the LES results. The
following test cases satisfy these requirements:
1. planar channel investigated by Abe et al. [2]1 at 𝑅𝑒 𝑏 = 24,400
2. periodic pin fin array investigated by Simonin and Barcouda [3] at 𝑅𝑒 𝑏 =
34,800
3. periodic ribbed channel investigated by Drain and Martin [4] at 𝑅𝑒 𝑏 = 24,000

1 https://www.rs.tus.ac.jp/t2lab/db/
On the validity of the linear Boussinesq hypothesis 277

4. Ames and Dvorak [5] pin fin array at 𝑅𝑒 𝑏 = 20,000 with experimental PIV
data kindly provided by Siemens Energy2
An overview of the four test cases is shown in Fig. 1. The first test case of a planar
channel implies two homogeneous directions. It is used as a reference case to validate
the methodology. Test cases number two and three are quasi-two-dimensional periodic
flows with one homogeneous direction. The fourth test case is a fully three-dimensional
periodic flow.

1 2 periodic
3.2 H 8D
2.0
6.4 H
H

2.2 D
fully
devel-
fully oped
devel-
oped periodic

3 7.2 e 4
7.2 5.0 D 2.5
e D

2.0 D
fully fully
devel- devel-
e
oped oped

e/H = 0.2 periodic

periodic
e/2

Fig. 1: Geometries and numerical domain of the four studied test cases [2–5].

2 Numerical setup

All simulations were run with the commercial CFD (Computational Fluid Dynamics)
solver Simcenter STAR-CCM+ [6]. First, each geometry was calculated using a
RANS simulation. The resulting wall normal grid spacing and the velocity field was
used to estimate a proper grid resolution and time step size for the LES. Grid size was
chosen to obtain wall normal grid spacing of 𝑦 +𝑛 < 1 for the first computational cell.
The streamwise and spanwise grid spacing in the entire domain was set to 𝑥 + ≈ 50
and 𝑧 + ≈ 30, respectively.

2 conducted at the Siemens Energy Center at the University of Central Florida

278 Philipp Wellinger and Bernhard Weigand

The time step size was chosen to ensure 𝐶𝐹 𝐿 < 1 in most regions of the flow
field. Simcenter STAR-CCM+ offers the opportunity to model convective fluxes with
a hybrid method blending between MUSCL 3rd-order upwind and 3rd-order central-
differencing reconstruction. A blending factor of 0.02 was chosen to reduce numerical
dissipation due to the upwind scheme. Time discretization was performed using a
2nd-order backward differentiation scheme with 5th-order correction. Gradients were
computed with a hybrid Gauss-Least Squares method limited with Venkatakrishnans
[7] method. All mentioned schemes are described in the Simcenter STAR-CCM+
user guide [6] in detail.
Hexahedral meshes were used for the LES since they were significantly faster
compared to the polyhedral meshes usually used in Simcenter STAR-CCM+. For the
test cases one and four, meshes were generated with Ansys ICEM CFD [8]. For the test
cases two and three, the trimmed mesher available in Simcenter STAR-CCM+ was
used to generate a two-dimensional mesh. Subsequently, the two-dimensional mesh
was extruded in the homogeneous direction. The overall number of computational
cells for the four test cases were 35, 45, 20 and 15 millions. These meshes were fine
enough to resolve ≈ 95% of the turbulent kinetic energy.
For this study, the WALE subgrid-scale model from Nicoud and Ducros [9] is
applied. This model uses an algebraic formulation for the small scales not resolved
by the computational grid. The coefficient used by the WALE model is usually less
sensitive to different test cases and it was kept constant at its default value for all
simulations.
In order to obtain mean values of the velocity components and turbulence statistics,
i.e. the Reynolds stresses, an averaging process was conducted. The mean velocity
field and the mean velocity gradients can be obtained from the time averaged flow
field. The resolved part of the Reynolds stresses were obtained using the empirical
variances and covariances of the instantaneous velocity field. The modeled part was
calculated from the subgrid-scale model. In addition, the time-averaged values were
spatially averaged in the homogeneous directions (test cases 1-3) to increase the
number of evaluation points significantly.

3 Results

3.1 Reference case: planar channel

As a first step, a comparison between the reference data and the LES for the planar
channel [2] is shown in Fig. 2. Instead of comparing the Reynolds stress as defined
in Eq. (1), only the product of the velocity fluctuations are compared. This term is
denoted as the specific Reynolds stresses and is defined by

𝜏𝑖 𝑗 = 𝑢 𝑖′𝑢 ′𝑗 . (3)
On the validity of the linear Boussinesq hypothesis 279

Fig. 2: Comparison of the normalized mean velocity component 𝑈/𝑈𝑚𝑎𝑥 and

the normalized specific Reynolds stress components 𝜏𝑖 𝑗 /𝑢 2𝜏 for the planar channel
obtained by LES with DNS data from [2].

The left diagram shows the mean velocity component normalized by the maximum
velocity in the center of the channel 𝑈/𝑈𝑚𝑎𝑥 for DNS and LES. The LES predicts
almost identical results in the entire domain. Since the DNS uses a much finer grid
resolution, 𝑦 + values below 0.3 are available. The diagram on the right compares the
specific Reynolds stresses normalized by the square of the friction velocity 𝜏𝑖 𝑗 /𝑢 2𝜏 . In
this case, a slight underprediction of 𝜏11 /𝑢 2𝜏 obtained by the LES can be observed.
However, a very good overall agreement can be determined. The LES data are applied
to compute 𝜌 𝑅𝑆 . The comparison with the data given by Schmitt [1] is show in Fig. 3.

Fig. 3 Comparison of the

validity parameter 𝜌𝑅𝑆 for the
planar channel computed from
LES data with data given by
[1], Fig. 2.

Although the Reynolds number of the reference case used by Schmitt [1] is
slightly lower, a good agreement between both data can be observed. This indicates a
proper implementation of the entire simulation and evaluation process. As already
mentioned, the alignment between both tensors is very poor close to the walls and
reaches larger values closer to the center of the channel. However, very low values
are again obtained in the symmetry plane since there the velocity gradient is zero and
the principal specific Reynolds stresses are not.

3.2 Validation of the flow field prediction

The validation of the flow field prediction of test cases 2 - 4 are presented below.
The evaluation lines used to compare the predicted flow field with the experimental
data are shown in Fig. 4. For the first pin fin array, experimental data of the specific
280 Philipp Wellinger and Bernhard Weigand

Reynolds stresses and the velocity components are available along four different lines.
The same accounts for the periodic ribbed channel. For the last test case, the pin fin
array originally investigated by Ames and Dvorak [5], PIV data are available between
the tail of the second row and the beginning of the third row. However, the results
were explicitly compared at seven different lines.

y0 y4
y3
x0 x3 y2
flow
x1 x2 y1
direction x0 x1 x2 x3
x1
x2

Fig. 4: Position of the experimental data for the periodic pin fin array (left) and the
periodic ribbed channel (middle). The evaluation lines of the 2nd pin fin array are
located in the symmetry plane (right).

Figures 5 and 6 compare the mean velocity field and the specific Reynolds stresses
of the LES with the experimental data of the pin fin array at selected locations. In this
case, the bulk velocity 𝑈𝑏 at the inlet is used for normalization. The prediction of
the velocity field (including the corresponding gradients) and the specific Reynolds
stresses are in very good agreement with the experimental data. Since all available
locations are predicted quite well, a well resolved flow field of the entire domain can
be assumed. Therefore, the validity parameter 𝜌 𝑅𝑆 can be accurately computed in
the entire flow field using the LES data. A detailed study of this test case including
a study of the turbulence structures can be found in the paper published during the
project [10].
Figures 7 and 8 show a selection of the numerical results of the normalized velocity
components and the specific Reynolds stress components for the ribbed channel in
comparison with experimental data. Again, the bulk velocity 𝑈𝑏 at the inlet was
used for normalization. The agreement between experimental results and numerical
data is very good for all positions and components. Based on this data, an accurate
prediction of 𝜌 𝑅𝑆 is expected in the entire domain.
The last test case was the most complex. Since it does not contain any homogeneous
direction the additional space averaging is not applicable. Therefore, longer averaging
times were needed. The comparison between the experimental data (PIV) and the
LES results normalized by the bulk velocity 𝑈 𝐵 at several locations is given in Figs. 9
and 10. In this case, larger deviations between both methods can be observed for
both velocities and the specific Reynolds stresses. Especially the specific Reynolds
stresses tend to be overpredicted. Although the agreement is not as good as for the
previous cases, the overall flow behavior is well captured. Therefore, the LES results
can be used for a rough prediction of the validity parameter 𝜌 𝑅𝑆 .
On the validity of the linear Boussinesq hypothesis 281

Fig. 5: Comparison of the normalized velocity components 𝑈/𝑈𝑏 and 𝑉/𝑈𝑏 obtained
by LES with experimental data from [3] at various locations of the quasi-two-
dimensional pin fin array.

Fig. 6: Comparison of the normalized specific Reynolds stress components 𝜏11 /𝑈𝑏2 ,
𝜏12 /𝑈𝑏2 and 𝜏22 /𝑈𝑏2 obtained by LES with experimental data from [3] at various
locations of the quasi-two-dimensional pin fin array.
282 Philipp Wellinger and Bernhard Weigand

Fig. 7: Comparison of the normalized velocity components 𝑈/𝑈𝑏 and 𝑉/𝑈𝑏 obtained
by LES with experimental data from [4] at various locations of the quasi-two-
dimensional ribbed channel.

Fig. 8: Comparison of the normalized specific Reynolds stress components 𝜏11 /𝑈𝑏2 ,
𝜏12 /𝑈𝑏2 and 𝜏22 /𝑈𝑏2 obtained by LES with experimental data from [4] at various
locations of the quasi-two-dimensional ribbed channel.
On the validity of the linear Boussinesq hypothesis 283

Fig. 9: Comparison of the normalized velocity components 𝑈/𝑈𝑏 and 𝑉/𝑈𝑏 obtained
by LES with experimental data (PIV) provided by Siemens Energy at various locations
of the three-dimensional pin fin array.

Fig. 10: Comparison of the normalized specific Reynolds stress components 𝜏11 /𝑈𝑏2
and 𝜏22 /𝑈𝑏2 obtained by LES with experimental data (PIV) provided by Siemens
Energy at various locations of the three-dimensional pin fin array.
284 Philipp Wellinger and Bernhard Weigand

3.3 Validity of the linear Boussinesq hypothesis for the selected test cases

The distribution of the validity parameter 𝜌 𝑅𝑆 based on the LES results are presented
in Figs. 11 and 12. Blue areas indicate regions where a large misalignment between
the eigenvectors of 𝑅𝑖 𝑗 and 𝑆𝑖 𝑗 can be observed. Hence, the LBH is not valid in
these regions. Red areas indicate good alignment between the eigenvectors. The
black line represent the soft limitation of 𝜌 𝑅𝑆 > 0.86 introduced by Schmitt [1]. The
validity parameter for the quasi-two-dimensional cases are shown in Fig. 11. For the
periodic pin fin array the LBH is valid in approximately one-third of the fluid domain.
Especially in the wake region (1) and in the narrow section between the pins (2) the
eigenvectors are aligned. However, the LBH is not applicable in the near wall region
(3) and close to the stagnation point (4). The results for the periodic ribbed channel
reveal a very interesting pattern. Overall, only very small areas with 𝜌 𝑅𝑆 > 0.86 can
be identified. The results in the predominantly undisturbed flow field in the upper
part between the ribs (1) are comparable to the planar channel. In the area between
the ribs at rib height (2), 𝜌 𝑅𝑆 increases but mainly remains below 0.86. However, in
the section above the ribs (3) a large misaligned occurs.

(2)
(1)
(1)
(3) (3)
flow
(4)
direction (2)

y y
x x

Fig. 11: Contour plot of 𝜌 𝑅𝑆 for the quasi-two-dimensional test cases: blue areas
represent regions in which the LBH is violated, red regions indicate good alignment.
The black lines highlight 𝜌 𝑅𝑆 > 0.86. left: periodic pin fin array, right: periodic
ribbed channel

In Fig. 12 the validity parameter on the symmetry planes of the three-dimensional

pin fin array is given. First of all, large red areas can be observed. However, in
most regions 𝜌 𝑅𝑆 is smaller than the weak limit of 0.86 indicating larger errors
introduced by the Boussinesq hypothesis. Similar to the two-dimensional pin fin
array, values of 𝜌 𝑅𝑆 close to 0.86 can be observed in the wake region behind the
pin (1). However, the behavior in the stagnation region (2) differ from each other.
For this test case, larger values of 𝜌 𝑅𝑆 are noticeable. This could be due the reduced
influence of the surrounding pin fins. In addition, blue areas close to solid walls of
the pins are less apparent. This might be an effect of the lower Reynolds number,
On the validity of the linear Boussinesq hypothesis 285

but needs further investigations. Interestingly, no larger, continuous blue areas exist
as they occur for the periodic ribbed channel. Only distinct narrow blue lines (3, 4)
separating several red regions are noticeable. The physical meaning of this sharp
separation in currently unclear and further investigations are needed. Although the
flow field is three-dimensional, the results only slightly differ along the pin height
in the 𝑧-direction (5). The largest gradients occur in front of the pin, where the blue
line forms a parabolic shape (4). Additionally, the LBH is not valid in corners in the
streamwise direction (6).

(3)

(1)
(2)

z y
y x
x

z z
x y
flow (5)

direction (6)
(4)

Fig. 12: Contour plot of 𝜌 𝑅𝑆 for the three-dimensional pin fin array: blue areas
represent regions in which the LBH is violated, red regions indicate good alignment.
The black lines highlight 𝜌 𝑅𝑆 > 0.86.

4 Computational performance

The parallel computing performance of Simcenter STAR-CCM+ on the ForHLR II

cluster was investigated. Therefore, the LES of the quasi-two-dimensional periodic
pin fin array (test case number two) containing ≈ 45 million cells was run on 40 up
to 3600 cores (2, 4, 8, 16, 32, 64, 128 and 180 nodes). Consequently, the number of
computational cells per cpu (cpc) vary between ≈ 1.1 million and 12,500. For the
286 Philipp Wellinger and Bernhard Weigand

speed-up test, only 70 iterations were simulated and the average computational time
per iteration was determined. For this purpose, the first 15 and the last 5 iterations
were excluded to eliminate potential errors due to loading or saving effects.
Figure 4 shows the parallel speedup against the ideal speedup and the efficiency of
Simcenter STAR-CCM+ on ForHLR II. The efficiency is defined as the respective
speedup divided by the ideal speedup. A very high efficiency of more than 80%
can be determined using up to 64 nodes (35,000 cpc). In addition, a high efficiency
just above 50% using 180 nodes (12,500 cpc) can be observed. This highlights the
excellent parallelization of Simcenter STAR-CCM+ on the ForHLR II cluster. For
this study, 25,000 cpc were used leading to an efficiency of ≈ 80%.

Fig. 13: Speedup vs. ideal speedup (left) and efficiency (right) of Simcenter STAR-
CCM+ on the ForHLR II cluster using ≈ 45 million cells.

Table 1 summarizes information concerning the usage of computing resources on

the ForHLR II. As shown before, Simcenter STAR-CCM+ shows good speedup up
to ≈ 25,000 cpc. Therefore, the number or cores were chosen in accordance to the
conducted speedup test. The wall time for each test case was adjusted to reach an
almost symmetrical mean flow field and turbulence statistics.

Table 1: Usage of resources on the ForHLR II

Simulation # of cells # of cores wall time core hours

planar channel 35 million 1400 8 days 270,000

pin fin Simonin 45 million 1800 12 days 520,000
ribbed channel 20 million 800 10 days 190,000
pin fin Ames 15 million 600 16 days 230,000

1,210,000
On the validity of the linear Boussinesq hypothesis 287

5 Conclusion

The validity of the linear Boussinesq hypothesis has been investigated for four
different test cases. For this purpose, Large Eddy Simulations have been performed
and the numerical results have been compared to experimental and DNS data. The
mean strain rate tensor and the anisotropic stress tensor have been computed from
the LES data. The validity parameter 𝜌 𝑅𝑆 introduced by Schmitt [1], representing the
misalignment between the eigenvectors of those two tensors, has been analyzed. The
first test case of a planar channel demonstrates that the applied numerical method
correctly reproduces the results given by Schmitt [1] for a similar planar channel.
Flow field and Reynolds stresses were well predicted for the second test case of a
quasi-two-dimensional pin fin array. Several distinct regions with good alignment
between the eigenvectors, e.g. the wake region, have been identified. The LES was
also able to predict the experimental results for the third test case of a periodic ribbed
channel very well. In the upper part of the channel 𝜌 𝑅𝑆 is quite similar to the planar
channel. Only small areas with good alignment have been observed. In contrast, a
high degree of misalignment has been identified in the region above the ribs. For the
three-dimensional pin fin array, larger deviations between the experimental data and
the LES have been observed. Nevertheless, the overall flow field is predicted with
sufficient accuracy to get a general idea of the distribution of 𝜌 𝑅𝑆 . Similar to the
quasi-two-dimensional test case, the LBH is valid in the wake region. However, a
different behavior close to the stagnation point has been observed.
Furthermore, speedup tests have been conducted. Good scaling up to 12,500 cpc
with an efficiency just above 50% has been observed. However, stronger scaling with
25,000 cpc have been used for this study corresponding to an efficiency of ≈ 80%.

Acknowledgements The investigations were conducted as part of the Siemens Clean Energy Centre
(CEC) joint research program. The authors acknowledge the financial support by Siemens Energy
and the German Federal Ministry for Economic Affairs and Energy in the project under grant number
03ET7073F. The authors also thank the Siemens Energy Center, University of Central Florida for
the permission to publish the experimental data shown in Figs. 9 and 10. The responsibility for
the content lies solely with its authors. This work was performed on the computational resource
ForHLR II funded by the Ministry of Science, Research and Arts Baden-Württemberg and DFG
(“Deutsche Forschungsgemeinschaft”).

References

1. F.G. Schmitt, About Boussinesq’s turbulent viscosity hypothesis: historical remarks and a
direct evaluation of its validity. Comptes Rendus Mécanique 335(9-10), 617–627 (2007) doi:
10.1016/j.crme.2007.08.004
2. H. Abe, H. Kawamura, Y. Matsuo, Direct numerical simulation of a fully developed turbulent
channel flow with respect to the Reynolds number dependence. J. Fluids Eng. 123(2), 382–393
(2001) doi: 10.1115/1.1366680
3. O. Simonin, M. Barcouda, Measurements of fully developed turbulent flow across tube bundle.
Proc. Third Int. Symp. Applications of Laser Anemometry to Fluid Mech., Lisbon (1986)
288 Philipp Wellinger and Bernhard Weigand

4. L.E. Drain, S. Martin, Two-component velocity measurements of turbulent flow in a ribbed-wall

flow channel. Int. Conf. on Laser Anemometry-Advances and Application, Manchester, 99–112
(1985)
5. F.E. Ames, L.A. Dvorak, Turbulent Transport in Pin Fin Arrays: Experimental Data and
Predictions. J. of Turbomach. 128(1), 71–81 (2005) doi: 10.1115/1.2098792
6. Siemens PLM Software, STAR-CCM+ Documentation Version 2020.1 (2020)
7. V. Venkatakrishnan, On the accuracy of limiters and convergence to steady state solutions. 31st
Aerosp. Sci. Meeting, Reno (1993) doi: 10.2514/6.1993-880
8. ANSYS, Inc., Ansys® ICEM CFD Documentation Release 17.1 (2016)
9. F. Nicoud, F. Ducros, Subgrid-scale stress modelling based on the square of the velocity gradient
tensor. Flow, Turbul. and Combust. 62(3), 183–200 (1999) doi: 10.1023/A:1009995426001
10. P. Wellinger, P. Uhl, B. Weigand, J. Rodriguez, Analysis of turbulence structures and the
validity of the linear Boussinesq hypothesis for an infinite tube bundle. Int. J. of Heat and Fluid
Flow 91, 108779 (2021) doi: 10.1016/j.ijheatfluidflow.2021.108779
Development of turbulent inflow methods for the
high order HPC framework FLEXI

Daniel Kempf, Min Gao, Andrea Beck, Marcel Blind, Patrick Kopper, Thomas Kuhn,
Marius Kurz, Anna Schwarz and Claus-Dieter Munz

Abstract Turbulent inflow methods offer new possibilities for an efficient simulation
by reducing the computational domain to the interesting parts. Typical examples
are turbulent flow over cavities, around obstacles or in the context of zonal large
eddy simulations. Within this work, we present the current state of two turbulent
inflow methods implemented in our high order discontinuous Galerkin code FLEXI
with special focus laid on HPC applications. We present the recycling-rescaling
anisotropic linear forcing (RRALF), a combination of a modified recycling-rescaling
approach and an anisotropic linear forcing, and a synthetic eddy method (SEM). For
both methods, the simulation of a turbulent boundary layer along a flat plate is used
as validation case. For the RRALF method, a zonal large eddy simulation of the
rear part of a tripped subsonic turbulent boundary layer over a flat plate is presented.
The SEM is validated in the case of a supersonic turbulent boundary layer using
data from literature at the inflow. In the course of the cluster upgrade to the HPE
Apollo system at HLRS, our framework was examined for performance on the new
hardware architecture. Optimizations and adaptations were carried out, for which we
will present current performance data.

Daniel Kempf, Min Gao, Andrea Beck, Marcel Blind, Thomas Kuhn, Marius Kurz, Anna Schwarz
and Claus-Dieter Munz
Institute of Aerodynamics and Gas Dynamics, University of Stuttgart,
e-mail: {kempf,mg,beck,blind,m.kurz,schwarz,munz}@iag.uni-stuttgart.de
Patrick Kopper
Institute of Aircraft Propulsion Systems, University of Stuttgart,
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 289
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_17
290 Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz, Schwarz and Munz

1 Introduction

Turbulent flows are relevant for a wide range of engineering applications. Therefore,
they are the subject of intensive research. In the field of computational fluid dynamics,
turbulent inflow conditions are of great relevance in order to simulate turbulent
flows as efficiently as possible by minimizing the computational domain. For the
computation of turbulent flows and their challenging multi-scale character, high order
methods are well suited due to their inherent low dissipation and dispersion errors.
Our open source simulation framework FLEXI1 has been developed and improved
during the last years in the Numeric Research Group at the IAG. A current overview
of the framework is given in [1]. The solver discretizes the compressible Navier–
Stokes–Fourier equations with an arbitrary high order discontinuous Galerkin spectral
element method (DGSEM) in space and uses low storage Runge–Kutta schemes for
explicit high order time integration. The focus of FLEXI is on scale-resolving large
eddy simulation (LES) of compressible flows, which includes aeroacoustics [2–4],
transitional and turbulent flows [5, 6], particle-laden flows [7] and many more.
LES poses significant computational requirements, still preventing its application
in industry. The less demanding Reynolds-averaged Navier–Stokes equations (RANS)
provides good results in attached flows but suffers from well-known shortcomings
with regards to intermittency, transition and separation. A zonal LES, sometimes
also called embedded LES, has emerged as a hybrid approach, restricting the
expensive application of LES to critical areas or regions of special interest, while
advancing the remaining domain with RANS. However, while this approach does
reduce the computation effort significantly, it poses the new challenge of requiring
a turbulent inflow method at the RANS-LES interface which reconstructs the time-
accurate LES inflow conditions from the RANS solution. An overview of existing
approaches is given in [8]. In this work, we present two turbulent inflow methods,
the recycling-rescaling anisotropic linear forcing (RRALF) by Kuhn et al. [9], a
combination of a modified recycling-rescaling approach and an anisotropic linear
forcing, and a synthetic eddy method (SEM) by Roidl et al. [10] in combination
with an anisotropic linear forcing (ALF) based on Laage de Meux et al. [11]. Both
methods are implemented in FLEXI and form the basis for RANS-LES coupling
approaches in future applications.
This work is structured as follows. In Sec. 2, we introduce both turbulent inflow
methods. In Sec. 3, we present a performance analysis of FLEXI on HAWK, the
new cluster at HLRS, and performance optimizations on this system as well as
implementation details of the turbulent inflow methods. In Sec. 3, both presented
turbulent inflow methods are validated with a simulation of a turbulent boundary
layer along a flat plate. In case of the RRALF method, a zonal large eddy simulation
of the rear part of a subsonic turbulent boundary layer is presented. The SEM is
validated with the case of a supersonic turbulent boundary layer. Sec. 5 concludes the
paper with a short outlook on further activities.

1 https://github.com/flexi-framework/flexi
Turbulent Inflow Methods for FLEXI 291

2 Methods

In this section, we introduce both turbulent inflow methods, which will be applied
throughout this work. First, in Sec. 2.1 we cover the RRALF method, an approach
for a zonal LES. Due to the recycling-rescaling used, the solution can suffer from
an artificial low frequency periodicity. In specific cases, like a shock-boundary
layer interaction with an oscillating shock, this frequency can influence the shock
movement. Therefore, a synthetic eddy method can be favorable in such cases, which
is described in Sec. 2.2.

2.1 Recycling-rescaling anisotropic linear forcing

In this work, we present an application of the combined recycling-rescaling an-

isotropic linear forcing technique by Kuhn et al. [9]. Recycling methods generally
produce the most accurate turbulent fields by employing the turbulence mechanisms
of the Navier–Stokes equations themselves. They are however restricted to fully
developed flows. An extension to zero pressure gradient (ZPG) boundary layer flows
was introduced by Lund et al. [12] by recycling and rescaling only the fluctuating part
of the solution at the interface. Anisotropic linear forcing introduced by de Laage de
Meux et al. [11] on the other hand is a volume forcing approach, introducing a source
term in the momentum equation to quickly stimulate the development of turbulence.
However, since the forcing is disconnected from the Navier–Stokes equations, it is
limited to weak source terms and thus large development regions are required to
form realistic turbulence. RRALF combines these two approaches, relying on the
Navier–Stokes equations for physical turbulence production but adding a forcing
zone in front of the recycling plane to break the limitation to ZPG boundary layers
and drive the solution towards the desired boundary layer profile. The rescaling
follows [12] by splitting the boundary layer into an inner and an outer region and
reconstructing the rescaled profile using
h i
+ (𝑢 𝑖′ )𝑖𝑛 + (𝑢 𝑖′ )𝑖𝑛
𝑖𝑛𝑛𝑒𝑟 𝑖𝑛𝑛𝑒𝑟
 𝑜𝑢𝑡𝑒𝑟 𝑜𝑢𝑡𝑒𝑟

𝑢 𝑖,𝑖𝑛 = ( 𝑢¯𝑖 )𝑖𝑛 [1 − 𝑊 (𝜂𝑖𝑛 )] + ( 𝑢¯𝑖 )𝑖𝑛 𝑊 (𝜂𝑖𝑛 ),

where 𝑢 𝑖 is the corresponding component of the instantaneous velocity field, [ ¯ ] and

[ ] ′ denoting mean and fluctuating quantities, and 𝑊 (𝜂𝑖𝑛 ) is a blending function
defined as "   #
1 𝑎(𝜂𝑖𝑛 − 𝑏)
𝑊 (𝜂𝑖𝑛 ) = 1 + tanh /tanh(𝑎) ,
2 (1 − 2𝑏)𝜂𝑖𝑛 + 𝑏
with the weights 𝑎 = 4 and 𝑏 = 0.2. The ALF volume forcing term is taken from [11],
which yields
𝜕 𝜌𝑢 𝑖
= 𝑅(𝑢 𝑖 ) + 𝜌 𝑓𝑖
𝜕𝑡
292 Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz, Schwarz and Munz

for each of the momentum vector components 𝑖 = 1, 2, 3, where 𝑅 represents the

spatial DG operator and 𝑓𝑖 the ALF vector given as 𝑓𝑖 = 𝐴𝑖 𝑗 𝑢 𝑗 + 𝐵𝑖 . 𝐴𝑖 𝑗 and 𝐵𝑖 are
the tensors for the Reynolds stresses and the mean velocity forcing, respectively. Due
to the recycling-rescaling, the second order turbulence statistics can overshoot the
target data. A modification of the original ALF introduces an explicit damping which
corrects the overshoot rapidly.
In the following paragraph, we discuss the algorithmic implementation of the
RRALF method in FLEXI. In a preprocessing step, the computational grid of
the recycling interface is built based on the properties of the target data. The
scaling is implemented as a static geometric scaling of the recycling plane, which is
computationally efficient. For every point on the recycling grid, we determine the
global element ID and the cell local coordinates of the scaled interface. Additionally,
we provide the necessary data for the MPI communication between the processors,
which handle the physical boundary and the interface plane. Besides the numerical
aspects of the preprocessing step, we also compute the wall-normal scaling function
from the target turbulent statistics of the initial flow field. This information is stored
to build the communication and the interface as well as to interpolate the turbulent
inflow fluctuations during runtime. The second preprocessing step creates an initial
solution based on turbulent target statistics. Time-averaged data can be provided
as 1D profiles and is interpolated to the flow domain to match the desired Re 𝜃 .
Additionally, field data of a precursor simulation can be used. Following Kuhn et
al. [9], we add superimposed wall-parallel streaks to improve the development of
turbulent structures inside the boundary layer. Further, the forcing zones where ALF
is used are defined. During runtime, the main overhead of the RRALF method is
created by the communication of data between the interface recycling plane and
the boundary as well as the update and application of the ALF forcing in every
Runge–Kutta step.

  






Fig. 1: Schematic of the SEM with an ALF zone and three virtual boxes containing
three different types of virtual eddies. The inflow boundary is highlighted red.
Turbulent Inflow Methods for FLEXI 293

2.2 Synthetic eddy method

Another method to produce the desired turbulent fluctuations at inflow boundaries

is the synthetic eddy method. SEM was proposed by Jarrin et al. [13], based on the
idea that turbulence can be seen as a superposition of coherent structures. Following
this idea, SEM creates turbulent fluctuations at the inflow by superimposing virtual
eddies in a virtual volume box at the inlet such that the prescribed turbulent statistics
are recovered. Since the virtual eddy cores at the inlet are distributed randomly,
SEM does not suffer from the well-known spurious periodicity of recycling-based
schemes [12]. Pamies et al. [14] extended the SEM by subdividing the single virtual
volume box containing one type of virtual vortices into multiple zones containing
different types of virtual vortices, as is shown in Fig. 1. This extension allowed to
account for the different turbulent scales in the wall-normal direction. Roidl et al. [10]
extended this method to compressible flows by enforcing the strong Reynolds analogy.
To reduce the required recovery length, we followed the idea of Kuhn et al. [9] that
ALF can be employed to improve the ability of the recycling-rescaling method to
recover the target turbulent statistics. To this end, we also adopted the ALF method
mentioned above to reduce the required recovery length of SEM.

3 Performance

The purpose of this section is twofold. In Section 3.1, the general performance
of FLEXI on the new HAWK system is evaluated and code optimizations for the
new system are discussed. In a second step, the parallelization strategy for SEM is
presented in Section 3.2.

3.1 FLEXI on HAWK

The system change at HLRS from the former Cray XC40 (Hazel Hen) to the current
HPE Apollo (HAWK) system included a shift from Intel® Xeon® CPUs to AMD
EPYCTM as well as a changed node-to-node interconnect, where an InfiniBand
architecture with a 9D-Hypercube topology replaced the Aries interconnect of Hazel
Hen. These drastic changes in hardware architecture and especially the increase
of available cores per node (from 24 to 128) motivated a detailed investigation of
the performance of our open source code FLEXI on the new system. To analyze
the performance, we use a free stream problem within a Cartesian box as test case,
for which we solve the Navier–Stokes–Fourier equations. To account for different
problem sizes, we further vary the number of elements in each direction for the
different cases. The polynomial degree was set to 𝑁 = 5, which results in a sixth-order
scheme and is a typical choice for production runs. We investigated the number of
nodes in the range from 1 to 512. The code was compiled with the GNU compiler
294 Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz, Schwarz and Munz

version 9.2.0 with the libraries mpt 2.23, hdf5 1.10.5 and aocl 2.1.0. The gradients of
the solution are computed with the BR1 lifting method [15], and the split formulation
by Pirozzoli [16] is employed to control aliasing errors. For each run, we computed
100 time steps. The influence of fluctuations on the performance of the system is taken
into account by running each configuration at least three times to collect statistics.
The performance data was obtained with the system’s configuration as of June 2021.

1.5 1.5
PID [𝜇s/DOF]

1 1

0.5 0.5

102 103 104 105 106 102 103 104 105 106
(a) Baseline, stride=1 (b) Baseline, stride=2
1.5 1.5 83 · 22 83 · 24
PID [𝜇s/DOF]

83 · 26 83 · 28
1 1

0.5 0.5

102 103 104 105 106 102 103 104 105 106
#DOF/Core #DOF/Core
(c) Performance optimized, stride=1 (d) Performance optimized, stride=2

Fig. 2: Results of the scaling analysis for different meshes and loads. Four different
cases are investigated: the baseline version of FLEXI as described in Krais et al. [1]
and the optimized version for HAWK with a stride of 1 and 2 for each version. The
legend lists the number of elements of the investigated meshes.

Fig. 2 presents the strong scaling behavior by plotting the performance index (PID)
over the number of degrees of freedom (DOF) per rank for all considered cases and
configurations. The PID is defined as

wall-clock-time · #cores
PID = , (1)
#DOF · #time steps · #RK-stages
and represents the time it takes to advance a single degree of freedom to the next
stage in the Runge–Kutta time-stepping algorithm.
Fig. 2a shows the results of the baseline version of FLEXI on the HAWK cluster.
The qualitative behavior is similar for all plots in Fig. 2, and also matches the behavior
observed on the former system as described in Krais et al. [1]. A small number of
nodes for a given problem size results in a high load. Here, an increasing amount
of data has to be regularly moved to and retrieved from the main memory, since
Turbulent Inflow Methods for FLEXI 295

the data does not fit into the fast CPU cache anymore. With increasing load, the
memory bandwidth per core of the AMD EPYCTM CPUs becomes the limiting factor
and the performance index increases dramatically. This is due to the specific CPU
architecture. Each socket consists of 8 CCDs (Core Chiplet Die), which comprise 2
CCXs (Core Complex) each. Both CCXs share a common interface to the I/O and
consist of 4 cores each. Thus, this hardware architecture leads to a comparably small
memory bandwidth per core, which can deteriorate the performance of memory
intensive operations. Therefore, the reduced code performance for high loads is more
pronounced on HAWK than on the former Hazel Hen system. For an increasing
number of nodes for a given problem size, i.e. lower load, the PID increases again. This
is due to decreasing local work on each CPU, which cannot hide the communication
latency effectively and the latency becomes dominant. For all cases, the optimal PID
is observed in between 103 and 104 DOF per core, where the communication latency
and caching effects are balanced optimally. A detailed discussion of the fundamental
scaling behavior of the DGSEM code FLEXI can be found in Krais et al. [1].
To analyze the impact of the architecture and especially the limited memory
bandwidth on the performance of FLEXI, the performance analysis was first carried
out on all available cores (stride=1, see Fig. 2a) and in a second step, while using
only every second available core (stride=2, see Fig. 2b). This artificially increases the
available memory bandwidth per active core. Further, the reduced usage of CPUs
within each node can have benefits for the internode communication and might
raise the power limits for the active cores. The results in Fig. 2 indicate that the
case with stride=2 gains about 30 % in performance compared to the stride=1 case.
It is important to stress that the overall improvement of the PID for the stride=2
configuration would have to exceed the factor of 2 to compensate for the unused cores,
which was not observed in any of the investigated cases. The qualitative behavior is
similar for both cases and the significant decrease in performance towards high loads
is still noticeable for the stride=2 case.
The Hypercube topology of the interconnect allows only for defined numbers
of nodes (64, 128, 256, 512, ...) for large jobs, which renders the aforementioned
scaling behavior unfavorable regarding flexibility and efficiency. Therefore, we strived
to decrease the memory footprint of FLEXI and improved the code optimizations
performed by the compiler. To this end, the lifting procedure was redesigned to
compute only the gradients of the variables which are actually required to compute the
parabolic fluxes of the Navier–Stokes–Fourier equations. While this causes the data
to be not contiguous in memory anymore, it proved to reduce the memory footprint
of the lifting procedure by about a fifth and improved the overall performance of
FLEXI. By profiling the code, we found two major performance-reducing issues.
Two frequently called functions, namely the Riemann solver as well as the solver for
the two-point volume split flux, were not inlined by the compiler. To this end, we
introduced a two step compilation process to employ profile-guided optimization
(PGO). By using PGO, the aforementioned functions get inlined by the compiler
and the overall cache usage is improved. In Fig. 2c and Fig. 2d the results with the
optimized code version using both stride=1 and stride=2 are depicted. A comparison
of Fig. 2a with Fig. 2c demonstrates a performance improvement by about 25 % in
296 Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz, Schwarz and Munz








 

 

Fig. 3: (left) Distribution of 8192 processing ranks (64 nodes) in the computational
domain. Only the ranks covering the inlet, i.e. the 𝑌 -𝑍-plane, up to 1.2𝛿99,𝑖𝑛 have
to be considered for the SEM and form the sub-communicator. (right) Schematic
view of the domain decomposition at the inflow boundary. For the first process (red),
only the vortices with their core inside the sampling area have to be evaluated. The
sampling area is larger than the occupied area at the inlet by the maximum length
scale of the virtual vortices.

case of an optimal load and even up to 40 % for high loads. This indicates the more
efficient usage of the CPU cache and the available memory bandwidth, especially
towards high loads, where the significant performance losses are mitigated. Since
the performance of the optimized code is thus less sensitive to the specific load per
core, FLEXI becomes more flexible and efficient, especially regarding the specific
Hypercube topology of the interconnect on HAWK. The comparison of Fig. 2c with
Fig. 2d again depicts a higher performance for the case with stride=2. However, the
observed improvement is not sufficient to compensate for the idling cores.

3.2 Parallel implementation of SEM

As discussed in Sec. 2.2, the SEM superposes virtual eddies in a virtual domain at
the inflow boundary, which are then used to derive the turbulent inflow state at each
point of the inflow plane. Since the inflow state is computed in a non-local fashion,
a consistent parallelization strategy is crucial for the method to be applicable for
large-scale HPC applications. Therefore, the following section focuses on the parallel
implementation of the reformulated SEM introduced in Sec. 2.2.
In a first step, a new MPI sub-communicator is introduced, which comprises
all processes with elements requiring velocity fluctuations at the turbulent inflow
boundary. All the information exchange within those processes computing the
turbulent fluctuations can thus be handled within this communicator. An exemplary
rank distribution at the inflow for practical applications is illustrated in Fig. 3 left.
According to the methodology of the SEM, the virtual eddies affect the velocity
only locally, while the range of influence is determined by the length scale of each
virtual eddy core. Therefore, the involved processors do not have to consider the
Turbulent Inflow Methods for FLEXI 297















Fig. 4: Flow chart of the MPI implementation of the reformulated SEM.

influence of all virtual vortices. Instead, only the effective vortices for each point
have to be evaluated. To this end, each process introduces a rectangular “sampling
region” containing a halo region. This region, which is depicted in Fig. 3 right,
is larger than the area occupied by the process by the maximum length scale of
the vortices and thus contains all vortices which influence the inflow state of this
processor. Clearly, the current restriction to rectangular sampling regions is not
optimal. Especially if a processor occupies a non-rectangular inflow area, vortices
might be evaluated unnecessarily, as shown in Fig. 3 right. In this example, the red
processor evaluates all the vortices within its sampling region. This might also include
vortices, which do not influence the red process but only the purple process. One way
to improve the sampling accuracy in the future could be to introduce an element-local
sampling procedure. However, the downside is the increase of the required memory
and communication latency. While such improvements could be easily integrated in
the future, the proposed approach already gives acceptable computational efficiency
at reasonable implementation effort.
In order to synchronize the distribution of the virtual eddy cores in each time
step, the root processor updates the locations of the virtual vortices. Afterwards,
the new distribution of vortices is broadcasted to all other processors inside the
sub-communicator. Finally, each process can evaluate its effective vortices. Figure 4
gives an overview of the implementation of the MPI strategy.

4 Results

In the following section, the introduced turbulent inflow methods are validated via a
turbulent flat plate simulation. First, we present a subsonic turbulent boundary layer
simulation at 𝑀𝑎 = 0.3 with a tripped simulation setup in Sec. 4.1. This is followed
by a zonal LES of the rear part of the identical tripped boundary layer case using
298 Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz, Schwarz and Munz

zonal region

sponge zone

   2536

RRALF

Fig. 5: Sketch of the mesh used for the tripped and zonal simulation of the turbulent
flat plate.

the RRALF method as turbulent inflow. The second configuration is a supersonic

boundary layer at 𝑀𝑎 = 2.0, which uses the SEM as turbulent inflow condition and
is presented in Sec. 4.2.

4.1 Subsonic turbulent flat plate

Two types of simulations of a weakly compressible turbulent flat plate were carried
out with the same numerics and resolution. First, a tripped turbulent boundary layer
simulation was conducted. These simulation results were used as target data for
a zonal LES of the rear half of the flat plate. The tripped boundary layer used
the incompressible, turbulent velocity profile provided by Eitel-Amor et al. [17] at
𝑅𝑒 𝜃 = 750 and a Mach number of 𝑀𝑎 = 0.3 as inflow data. The quadratic trip is placed
one boundary layer thickness behind the inflow and has the size of 𝑦 + = 50 in wall
units referring to the inflow profile. The domain size is 205 𝛿99,𝑖𝑛 ×12 𝛿99,𝑖𝑛 ×16 𝛿99,𝑖𝑛 ,
𝛿99,𝑖𝑛 being the boundary layer thickness at the inflow of the tripped case. The domain
resolves the boundary layer up to about 𝑅𝑒 𝜃 = 2, 800.
The mesh displayed in Fig. 5 comprises 459,150 elements, which results in about
235 million DOF for a polynomial degree of 𝑁 = 7. The mesh resolution is, improving
towards higher 𝑅𝑒 𝜃 , 𝑥 + = 25 → 20.5, 𝑦 + = 1 → 0.8 and 𝑧+ = 12 → 9.8. The
simulation was carried out for 531 convective time units 𝑇 ∗ = 𝛿99,𝑧 /𝑈∞ and was
averaged over the last 115 𝑇 ∗ to obtain the first and second order turbulence statistics.
Here, 𝑈∞ denotes the freestream velocity and 𝛿99,𝑧 the boundary layer thickness at
the inflow of the zonal domain at 𝑅𝑒 𝜃 = 1, 800 and is from now on the reference
for both cases. The simulation was carried out on HAWK with 256 nodes which
results in a load of 7,174 DOF/core. In this particular case, a PID of 0.95 𝜇𝑠 was
achieved with the optimized version of FLEXI, which corresponds to a reduction of
36% compared to a PID of 1.48 𝜇𝑠 for the non-optimized version. Due to additional
overhead, e.g. by the collection of turbulent statistics, the PID was slightly increased
compared to the results in Sec. 3. In total, the simulation required 445,000 CPUh.
The zonal LES had the same mesh resolution and resolved the boundary layer from
𝑅𝑒 𝜃 = 1, 800 − 2, 800. Hereby, the mesh was reduced to about 112 DOF. The region
where ALF was applied starts at the inflow and extends over 6 𝛿99,𝑧 in streamwise
Turbulent Inflow Methods for FLEXI 299

·10−3
Tripped LES
Zonal LES
5 Eitel-Amor et al. (2014)

cf 4

1,000 1,500 2,000 2,500 3,000

Reθ

Fig. 6: The friction coefficient 𝑐 𝑓 of the tripped and the zonal turbulent flat plate with
the numerical reference from Eitel-Amor et al. [17].

direction, whereby the first wall normal cell was skipped. The recycling plane was
placed at 8 𝛿99,𝑧 in streamwise direction. The simulation was carried out for 277 𝑇 ∗
and averaged over 69 𝑇 ∗ to obtain the turbulence statistics. The simulation used
227,000 CPUh with 128 nodes, which results in a load of 6,871 DOF/core. Due to
an additional overhead, introduced by the turbulent inflow method, the optimized
PID increased to 1.41 𝜇𝑠, which corresponds to a reduction of 25% referring to
the non-optimized version with a PID=1.88 𝜇𝑠. In Fig. 6, the friction coefficient
of the tripped and zonal LES as well as the reference from Eitel-Amor et al. [17]
are illustrated. The friction coefficient of the tripped simulation clearly indicates
the influence of the trip at low 𝑅𝑒 𝜃 . With increasing 𝑅𝑒 𝜃 , the friction coefficient
trends towards the reference but with a slightly steeper decline. For the zonal LES,
the turbulent inflow results in a significant drop in the friction coefficient at the
inflow. Due to the choice of the recycling method, the instantaneous fluctuations
were recycled without further amplitude scaling. However, in this flow regime the
changes of the amplitude of the Reynolds stresses over the recycling distance are
noticeable and lead to a relaxation of the flow field at the inflow. Further downstream,
the solution of the zonal LES tends towards the numerical reference.
Fig. 7 depicts the first and second order turbulence statistics at 𝑅𝑒 𝜃 = 2, 240 and
𝑅𝑒 𝜃 = 2, 536 within the zonal region, for which reference data from Eitel-Amor et al.
[17] are available. The time-averaged streamwise velocity in Fig. 7 (left) displays an
overall good agreement between the two simulations and the reference data for both
cases. At 𝑅𝑒 𝜃 = 2, 240 and 𝑅𝑒 𝜃 = 2, 536, the tripped LES matches the reference very
well in the viscous sublayer, the buffer layer and in the log-law region. In the outer
layer are minor deviations are noticeable due to a remaining influence of the trip. The
zonal LES slightly overestimates the reference from literature and the tripped LES
from the log-law region. At 𝑅𝑒 𝜃 = 2, 536, the zonal LES matches the two references
very well up to the log-law region. In the outer layer, the results of the zonal LES
lies between the two references. The second order turbulent statistics in form of
the normal stresses are illustrated in Fig. 7 (right) for both Reynolds numbers. For
300 Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz, Schwarz and Munz

both Reynolds numbers under investigation, there is an overall very good agreement
between the tripped and the zonal LES. However, both simulations tend to slightly
underestimate the Reynolds stresses.

3
Zonal LES

20 Tripped LES
Eitel-Amor et al. (2014)
U + = y+ 3 2

20 10 1 +
0.41 ln(y ) + 0.5
1

u′+ , v′+ , w′+

2
u+

Reθ = 2536
10 0 0
1 0 0.5 1 1.5
10−1 100 101 102 103 104
Reθ = 2240
0 0
10−1 100 101 102 103 104 0 0.5 1 1.5
y+ y/δ99

Fig. 7: Comparison of the mean velocity 𝑢 + = 𝑢/𝑢 𝜏 (left) and the Reynolds fluctuations
(right) at 𝑅𝑒 𝜃 = 2, 240 and 𝑅𝑒 𝜃 = 2, 536 for the tripped and the zonal turbulent
flat plate with a numerical
√︁ reference
√︁ from Eitel-Amor√︁et al. [17]. The Reynolds
fluctuations 𝑢 ′+ = 𝑢 ′𝑢 ′/𝑢 𝜏 , 𝑣 ′+ = 𝑣 ′ 𝑣 ′/𝑢 𝜏 and 𝑤 ′+ = 𝑤 ′ 𝑤 ′/𝑢 𝜏 are presented by
red lines ( ), green lines ( ) and blue lines ( ), respectively.

4.2 Supersonic turbulent flat plate

In the following, we present the LES results of a compressible turbulent boundary

layer at a Mach number of 𝑀𝑎 = 2.0, which is validated with the DNS data provided
by Wenzel et al. [18]. The SEM introduced in Sec. 2.2 is used as turbulent inlet
condition in conjunction with ALF. The computational domain covers a box with
the size of 34.78 𝛿99,𝑖𝑛 × 20.55 𝛿99,𝑖𝑛 × 3.95 𝛿99,𝑖𝑛 . The present simulation (LES02)
was conducted with a polynomial degree of 𝑁 = 6 with 102 × 104 × 27 elements in
streamwise, wall-normal and spanwise direction, respectively. This leads to a total
amount of about 98 million DOF. The grid resolution at the inlet was chosen in wall
units as Δ𝑦 +𝑚𝑖𝑛 = 1.6, Δ𝑥 + = 30 and Δ𝑧+ = 16. Periodic boundary conditions were
used in the spanwise direction and an adiabatic no-slip wall boundary condition was
used for the flat plate. At the upper boundary, the DNS data was set as Dirichlet
boundary condition to avoid reflections induced by the inconsistency between the
inflow and the upper boundaries. Finally, a sponge layer was placed before the outlet
and a supersonic outflow boundary condition was used to avoid spurious reflections.
The setup is displayed in Fig. 8. The simulation was performed with 8, 192 CPU
Turbulent Inflow Methods for FLEXI 301

Fig. 8: Q-criterion of the turbulent structures colored by the velocity magnitude. The
red region represents the zone where the ALF is enforced. The background mesh is
also presented up to 10.5 𝛿99 in the wall-normal direction.

cores, which leads to a load of about 11, 992 DOF/core. We obtained a PID of about
1.81 𝜇𝑠 for this LES02 case, which is about 33.3% larger compared to the results
in Sec. 3. In order to quantify the performance loss due to the reformulated SEM,
we calculated the same numerical case without using the additional ALF (LES01).
This resulted in a PID of around 1.68 𝜇𝑠, which is 24.4% larger than the baseline
code. In total, the computations used about 100,000 CPUh. In consideration of the
communication overhead due to the collection of the turbulent statistics for analysis,
it is assumed that the current MPI strategy works well but could be improved even
further.
The flow reached a quasi-steady state after about 77 convective time units
𝑇 ∗ = 𝛿99,𝑖𝑛 /𝑈∞ . In Fig. 8, an instantaneous snapshot of the turbulent structures in the
flow domain as well as the forcing zone, where the ALF was enforced, are illustrated.
In order to ensure that the forcing term remains weak, the forcing zone extends
from 5 𝛿99,𝑖𝑛 to 15 𝛿99,𝑖𝑛 in streamwise direction. After the quasi-steady state was
reached, the flow was advanced by another 350 𝑇 ∗ to obtain the turbulent mean flow
statistics. In Fig. 9, the mean velocity and the Reynolds stresses of the LES01 and the
LES02 are compared to the reference DNS. In order to account for the compressibility
∫ 𝑢+ √︁
effect, the Van Driest transformation of the mean velocity 𝑢 +𝑣𝐷 = 0 𝜌/𝜌 𝑤 𝑑𝑢 + as
√︁
well as Morkovin’s density scaling of the Reynolds fluctuations 𝑅 𝜙 𝜙 = 𝜌/𝜌 𝑤 𝜙 ′+ ,
𝜙 = 𝑢, 𝑣, 𝑤 are applied, where 𝜌 denotes the density and [ ] 𝑤 the statistics at the wall.
The results for the LES02 case show better agreement with the reference DNS data
than the LES01 case, which did not employ an ALF zone. This highlights the benefit
of the proposed method, which combines the SEM with ALF.
302 Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz, Schwarz and Munz

Fig. 9: Comparison of the averaged flow statistics at 15 𝛿99 between SEM (LES01),
SEM with additional ALF (LES02), and the reference DNS. Van Driest transformed
velocity profiles (left) and Morkovin’s density scaled Reynolds stress profiles (right).

5 Summary and outlook

In this work, we have presented the current capabilities of FLEXI with regards to
turbulent inflow methods as well as recent performance optimizations motivated by
the new hardware architecture of the HPE Apollo HAWK at HLRS. To adapt FLEXI
to the new CPU architecture, the optimizations were designed to reduce the memory
footprint and to achieve more efficient cache usage. Furthermore, the function call
overhead of two essential subroutines was reduced. We achieved a performance gain
of 25% at an optimal load and up to 40% at high loads, which was intended to give
FLEXI the necessary flexibility for an optimal use of the hypercube topology of
HAWK.
We presented two turbulent inflow methods, the recycling-rescaling anisotropic
linear forcing (RRALF), a combination of a modified recycling-rescaling approach
and an anisotropic linear forcing, and a synthetic eddy method (SEM). Both methods
were validated for a turbulent boundary layer along a flat plate. In case of the RRALF
method, a zonal large eddy simulation of the rear part of a tripped, subsonic turbulent
boundary layer was presented. By applying the RRALF method, we were able to
reproduce the target data from the tripped boundary layer within the zonal region. The
SEM was validated against a supersonic turbulent boundary layer using data from
literature at the inflow. Additionally, we showed that the SEM method benefits from
an additional anisotropic linear forcing behind the inflow to reach the target data faster.
In future work, the RRALF method will be applied for a zonal RANS-LES coupling
and the SEM method will be used to investigate shock-boundary layer interaction
with moving walls.

Acknowledgements We thank the Deutsche Forschungsgemeinschaft (DFG, German Research

Foundation) for supporting this work by funding - EXC2075 – 390740016 under Germany’s
Excellence Strategy and the DFG project Rebound - 420603919. The authors also gratefully
acknowledge the Deutsche Forschungsgemeinschaft DFG (German Research Foundation) for
Turbulent Inflow Methods for FLEXI 303

funding this work in the framework of the research unit FOR 2895. We acknowledge the support by
the Stuttgart Center for Simulation Science (SimTech) and the DFG International Research Training
Group GRK 2160. Min Gao recognizes the support of the China Scholarship Council (CSC). We all
truly appreciate the ongoing kind support by HLRS in Stuttgart.

References

1. N. Krais, A. Beck, T. Bolemann, H. Frank, D. Flad, G. Gassner, F. Hindenlang, M. Hoffmann,

T. Kuhn, M. Sonntag, and C.-D. Munz. Flexi: A high order discontinuous Galerkin framework
for hyperbolic–parabolic conservation laws. Computers and Mathematics with Applications,
2020.
2. T. Kuhn, J. Dürrwächter, F. Meyer, A. Beck, C. Rohde, and C.-D. Munz. Uncertainty quan-
tification for direct aeroacoustic simulations of cavity flows. Journal of Theoretical and Com-
putational Acoustics, 27(01):1850044, 2019.
3. H.M. Frank and C.-D. Munz. Direct aeroacoustic simulation of acoustic feedback phenomena
on a side-view mirror. Journal of Sound and Vibration, 371:132–149, 2016.
4. A. Beck, J. Dürrwächter, T. Kuhn, F. Meyer, C.-D. Munz, and C. Rohde. hp-multilevel Monte
Carlo methods for uncertainty quantification of compressible Navier–Stokes equations. SIAM
Journal on Scientific Computing, 42(4):B1067–B1091, 2020.
5. A.D. Beck, T. Bolemann, D. Flad, N. Krais, J. Zeifang, and C.-D. Munz. Application and
development of the high order discontinuous Galerkin spectral element method for compress-
ible multiscale flows. In High Performance Computing in Science and Engineering’18, pages
291–307. Springer, 2019.
6. A. Beck, T. Bolemann, D. Flad, H. Frank, G. Gassner, F. Hindenlang, and C.-D. Munz. High-
order discontinuous Galerkin spectral element methods for transitional and turbulent flow sim-
ulations. International Journal for Numerical Methods in Fluids, 76(8):522–548, 2014.
7. A. Beck, P. Ortwein, P. Kopper, N. Krais, D. Kempf, and C. Koch. Towards high-fidelity ero-
sion prediction: On time-accurate particle tracking in turbomachinery. International Journal of
Heat and Fluid Flow, 79:108457, 2019.
8. M.L. Shur, P.R. Spalart, M.K. Strelets, and A.K. Travin. Synthetic turbulence generators for
RANS-LES interfaces in zonal simulations of aerodynamic and aeroacoustic problems. Flow,
turbulence and combustion, 93(1):63–92, 2014.
9. T. Kuhn, D. Kempf, A. Beck, and C.-D. Munz. A novel turbulent inflow method for zonal large
eddy simulations with a discontinuous Galerkin solver. Submitted to Computers and Fluids,
2020.
10. B. Roidl, M. Meinke, and W. Schröder. A reformulated synthetic turbulence generation method
for a zonal RANS–LES method and its application to zero-pressure gradient boundary layers.
International Journal of Heat and Fluid Flow, 44:28–40, 2013.
11. B. De Laage de Meux, B. Audebert, R. Manceau, and R. Perrin. Anisotropic linear forcing
for synthetic turbulence generation in large eddy simulation and hybrid RANS/LES modeling.
Physics of Fluids, 27(3):035115, 2015.
12. T.S. Lund, X. Wu, and K.D. Squires. Generation of turbulent inflow data for spatially- developing
boundary layer simulations. Journal of Computational Physics, 140(2):233–258, 1998.
13. N. Jarrin, S. Benhamadouche, D. Laurence, and R. Prosser. A synthetic-eddy-method for
generating inflow conditions for large-eddy simulations. International Journal of Heat and
Fluid Flow, 27(4):585–593, 2006.
14. M. Pamiès, P.-E. Weiss, E. Garnier, S. Deck, and P. Sagaut. Generation of synthetic turbulent
inflow data for large eddy simulation of spatially evolving wall-bounded flows. Physics of
fluids, 21(4):045103, 2009.
15. F. Bassi and S. Rebay. A high-order accurate discontinuous finite element method for the
numerical solution of the compressible Navier–Stokes equations. J. Comput. Phys., 131:267–
279, 1997.
304 Kempf, Gao, Beck, Blind, Kopper, Kuhn, Kurz, Schwarz and Munz

16. S. Pirozzoli. Numerical methods for high-speed flows. Annual Review of Fluid Mechanics,
43:163–194, 2011.
17. G. Eitel-Amor, R. Örlü, and P. Schlatter. Simulation and validation of a spatially evolving
turbulent boundary layer up to Re 𝜃 = 8300. International Journal of Heat and Fluid Flow,
47:57–69, 2014.
18. C. Wenzel, B. Selent, M. Kloker, and U. Rist. DNS of compressible turbulent boundary layers
and assessment of data/scaling-law quality. Journal of Fluid Mechanics, 842:428–468, 2018.
A narrow band-based dynamic load balancing
scheme for the level-set ghost-fluid method

Daniel Appel, Steven Jöns, Jens Keim, Christoph Müller, Jonas Zeifang and
Claus-Dieter Munz

Abstract We present a dynamic load balancing scheme for compressible two-phase

flows simulations using a high-order level-set ghost-fluid method. The load imbalance
arises from introducing an element masking that applies the costly interface-tracking
algorithm only to the grid cells near the phase interface. The load balancing scheme is
based on a static domain decomposition by the Hilbert space-filling curve and employs
an efficient heuristic for the dynamic repartitioning. The current workload distribution
is determined through element-local wall time measurements, exploiting the masking
approach for an efficient code instrumentation. The dynamic repartitioning effectively
carries over the single-core performance gain through the element masking to
massively parallelized simulations. We investigate the strong scaling behavior for up
to 16384 cores, revealing near optimal parallel efficiency and a performance gain
of factor five on average compared to previous, unbalanced simulations without
element masking. The load balancing scheme is applied to a well-studied two- and
three-dimensional shock-drop interaction in the Rayleigh–Taylor piercing regime,
providing an overall runtime reduction of approximately 65%.

1 Introduction

Due to its inherent complexity, the modeling of compressible multi-phase flow remains
an active research topic. The two main simulation approaches for these types of flows
are sharp-interface and diffuse-interface methods. In the diffuse-interface method the
phase interface is resolved explicitly. In contrast, sharp-interface methods treat the
two phases as immiscible fluids separated by an interface. Thus, a physically sound

Daniel Appel, Steven Jöns, Jens Keim, Christoph Müller, Jonas Zeifang, and Claus-Dieter Munz
Institute of Aerodynamics and Gasdynamics, Pfaffenwaldring 21, 70569 Stuttgart, Germany,
e-mail: [email protected]
Jonas Zeifang
Faculty of Sciences, Universiteit Hasselt, Agoralaan Gebouw D, BE-3590 Diepenbeek, Belgium

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 305
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_18
306 D. Appel, S. Jöns, J. Keim, C. Müller, J. Zeifang and C.-D. Munz

coupling of the fluids has to be ensured and an additional interface-tracking algorithm

is required. The sharp-interface method was first introduced by Fedkiw et al. [11] and
has since been improved by many authors. Merkle and Rohde [24] presented a ghost-
fluid method based on the solution of a multi-phase Riemann problem for the evaluation
of the ghost states at the interface. The idea was extended by Fechter et al. [7–9]
to approximate Riemann solvers and applied in a level-set ghost-fluid framework.
Further improvements of the high-order level-set ghost-fluid method were presented
by Jöns et. al [19]. The simulation of multi-phase flows with complex topological
changes, present e.g. in shock-droplet and droplet-droplet interactions, is a common
high performance computing application due to its requirement for a high spatial
resolution. However, the used level-set ghost-fluid framework exhibits a dynamic
variation of the computational workload, which is caused by the locally applied
interface-tracking algorithm and the evaluation of the multi-phase Riemann problem at
the phase interface. To ensure an efficient resource utilization on large-scale computer
systems, dynamic load balancing strategies redistribute the workload evenly among
the processors at runtime. In this work, we present a load balancing approach for the
level-set ghost-fluid method. The approach is based on a static domain decomposition
through a Hilbert space-filling curve and employs the efficient H2-heuristic of Miguet
and Pierson [25] for the dynamic repartitioning. Similar approaches were successfully
applied to particle-based methods [29], direct-coupled multi-physics simulations of
acoustics [27], atmospheric modeling [22] and multiscale flow simulations using
Chimera grid techniques [3]. To evaluate the present workload distribution, we
perform element-local wall time measurements. The code instrumentation exploits
the masking concept to minimize the runtime overhead through these measurements
to a negligible extent. The remainder of the paper is organized as follows: First, the
governing equations of the level-set ghost-fluid framework are presented in section 2.
Section 3 briefly introduces the numerical framework to solve these equation systems.
This is followed by a description of the dynamic load balancing approach in section
4. Turning to simulation results, we first investigate the strong scaling behavior of the
load balancing scheme in section 5.1 and apply our framework to the simulation of a
two-dimensional and a three-dimensional shock-droplet interaction in section 5.2.
We conclude the report with a succinct summary.

2 Governing equations

In this work, we consider two pure, immiscible phases. Our domain Ω can generally
be divided into a liquid region Ω 𝐿 and a vapor region Ω𝑉 , which are separated by
an interface Γ. The fluid flow in Ω 𝐿 and Ω𝑉 are each described by the compressible
Euler equations:
Dynamic load balancing for the level-set ghost-fluid method 307

©𝜌ª © 𝜌v ª
­ ® ­ ®
­ 𝜌v® + ∇ · ­ 𝜌v ⊗ v + I𝑝 ® = 0,
­ ® ­ ®
(1)
­ ® ­ ®
­ ® ­ ®
𝜌𝑒 v(𝜌𝑒 + 𝑝)
« ¬𝑡 « ¬

where 𝜌 denotes the density, v = (𝑢, 𝑣, 𝑤) 𝑇 the velocity vector, 𝑝 the pressure and 𝑒
the specific total energy. The total energy per volume of the fluid 𝜌𝑒 is composed of
the internal energy per volume 𝜌𝜖 and the kinetic energy:
1
𝜌𝑒 = 𝜌𝜖 + 𝜌v · v. (2)
2
The pressure and specific internal energy of each phase are linked via an appropriate
equations of state (EOS). Within our framework, algebraic as well as multiparameter
EOS can be used. The tabulation technique of Föll et al. [12] ensures an efficient
evaluation.
Since we consider two-phase flows, the description of the phase interface is
essential. In our work we follow [33], in which the position of the interface is
implicitly given by the root of a level-set function 𝜙(x). From the level-set field,
geometrical properties such as the interface normal vector nLS and the interface
curvature 𝜅 can be calculated [19]. The level-set transport is governed by

𝜙𝑡 + vLS · ∇𝜙 = 0, (3)

where vLS is the level-set velocity-field. It is obtained at the interface and then
extrapolated into the remaining domain following [1], by solving the Hamilton–Jacobi
equations

𝜕𝑣 LS
𝑖
+ sign(𝜙)nLS · ∇𝑣 LS
𝑖 = 0, (4)
𝜕𝜏
with the direction-wise components 𝑣 LS LS and the pseudo time
𝑖 of the velocity field v
𝜏.
In general, the level-set function should fulfill the signed distance property.
However, Eq. (3) does not preserve this property. Hence, the level-set function has to
be reinitialized regularly. According to [33], we use the Hamilton–Jacobi equation


𝜙𝑡 + sign(𝜙) |∇𝜙| − 1 = 0, (5)

to retain the signed distance property.

308 D. Appel, S. Jöns, J. Keim, C. Müller, J. Zeifang and C.-D. Munz

3 Numerical method

Our numerical method can be separated into a fluid solver and an interface capturing
scheme. Both use the same computational mesh of non-overlapping hexahedral
elements Ω𝑒 . In the following, we will briefly describe both building blocks. A
detailed discussion can be found in [8, 9, 19, 26].

3.1 Fluid solver

The compressible Euler equations are discretized by the discontinuous Galerkin

spectral element method (DGSEM) [20]. In the high-order DGSEM, the solution is
a polynomial of degree 𝑁, defined at Gauss–Legendre points in each element. The
coupling between the neighboring elements is achieved via a classical numerical flux
function, e.g., HLLC or Rusanov. Since the DGSEM is a high-order method, discon-
tinuities, e.g., shocks and phase boundaries, will ultimately lead to the appearance of
the unwanted Gibbs phenomena. To stabilize the solution in these areas, we combine
the DGSEM with a finite-volume (FV) sub-cell method [2, 16, 21, 30, 32]. Areas in
which sub-cells are required are identified by a modal indicator [30] or the position
of the level-set root, for shocks or the phase-interface, respectively. In these elements,
the solution representation is switched conservatively from a polynomial of degree
𝑁 to 𝑁 + 1 equidistantly spaced FV sub-cells. To reduce the loss of accuracy in the
sub-cells, the underlying FV scheme is extended to a second-order TVD scheme in
combination with a minmod limiter.

3.2 Interface capturing

The level-set transport equation (3) is discretized with a DGSEM method for
hyperbolic equations with non-conservative products [6, 19], by the use of the
framework for path-conservative schemes of [5]. Herein, the solution is also described
by a polynomial of degree 𝑁. The coupling between the elements is handled via a so-
called path-conservative jump term, which we approximate with a path-conservative
Rusanov Riemann solver following [6].
Theoretically, the level-set function is a smooth signed-distance function. However,
in practical applications discontinuities may occur, e.g. in the case of topological
changes. To handle these, we use the FV sub-cell approach for the level-set transport
equation as well. The two additional operations required for the interface tracking,
reinitialization and velocity extrapolation, rely on Hamilton–Jacobi equations, which
are each solved with a fifth-order WENO scheme [18] in combination with a third-order
low-storage Runge–Kutta method with three stages.
Dynamic load balancing for the level-set ghost-fluid method 309

3.3 The level-set ghost-fluid method

The two building blocks described above are the fundamentals of our sharp interface
method. The remaining step is a physically sound coupling of the two fluid regions.
We follow the methodology presented in [9, 24], with a ghost-fluid method based on
the solution of two-phase Riemann problem. Given this solution, the numerical flux
at the interface as well as the velocity required for the extrapolation procedure can be
calculated for each phase.
Our numerical framework is able to predict complex two-phase flows. Phase
transition problems were considered in [10, 17]. Cases without phase transition,
e.g., colliding droplets, merging droplets and shock-drop interactions were shown in
[19, 26, 36].

4 Load balancing

4.1 Characterization of load imbalance

One of the major challenges for efficient high performance computing (HPC) is the
load balance [22, 34], i.e. the even distribution of the application workload across the
processor units. The need for dynamic load balancing arises if the workload changes
during runtime, as encountered in adaptive spatial grids [14] or local time stepping
techniques [4]. Aside from the numerical scheme, workload variations may also
originate from the considered physics. Examples range from particle-laden flows
[29], where the computational costs correlate with the local particle concentration,
to atmospheric modeling [22] and computational aeroacoustics [27], both of which
couple two physical models to capture the multiscale nature of problem.
Similarly, the outlined framework for multiphase flows introduces an uneven
workload distribution by applying the interface-tracking algorithm only in the vicinity
of the phase interface and solving the two-phase Riemann problem at the interface
itself. More specific, the solution of Eqs. (3)-(5) is only necessary in a narrow band
encompassing the interface [13]. Outside this narrow band, the level-set function is
set to the band’s fixed radius and the velocity field to zero.
The computational domain D can thus be decomposed into three intersecting sets
of grid cells: the bulk elements Dbulk ≡ D that discretize only the Euler equations,
the narrow band elements DNB ⊂ Dbulk in which additionally the aforementioned
level-set equations are solved, and the elements containing the phase interface itself,
DΓ ⊂ DNB , subject to the calculation of the interface curvature and to the solution of
the two-phase Riemann problem. The computational costs associated with these three
subsets consequently rise in the listed order. From an implementational perspective,
the distinction of Dbulk , DNB and DΓ is accomplished by introducing element masks
that filter the subset of elements relevant to the considered operations.
310 D. Appel, S. Jöns, J. Keim, C. Müller, J. Zeifang and C.-D. Munz

4.2 Domain decomposition and repartitioning

In general, the present framework operates on unstructured, curved grids. Conse-

quently, also the Cartesian grids in the investigated test cases below are internally
represented as unstructured meshes. For the given mesh representation, Hindenlang
[15] compared domain decomposition strategies for parallel simulations based on
space-filling curves (SFC) and graph partitioning. He demonstrated that graph parti-
tioning outperforms space-filling curves in terms of the ratio of inter-domain surfaces
to the total number of elements, which indicates a lower communication amount
per data operation. However, as the differences vanish with an increasing number of
subdomains (cores) and a space-filling curve greatly simplifies data structures and
parallel I/O, an SFC-based decomposition strategy was favored. In particular, the
Hilbert curve provided more compact subdomains than the Morton curve, and thus
led to a better domain composition for massively parallelized simulations.
With respect to dynamic load balancing, SFC-based strategies show further
advantages over graph partitioning approaches. Space filling curves generally provide
a fast mapping from the 𝑛-dimensional to the one-dimensional space while spatial
locality is preserved. This locality allows to reduce the distribution of 𝑁elem grid cells
among 𝑃 processors to the one-dimensional partitioning problem of decomposing
a task-chain of 𝑁elem elements with positive weights 𝑤 𝑖 , 𝑖 = 1, . . . , 𝑁elem into 𝐾
consecutive segments such that the bottleneck value, i.e. the maximum load among
all processors, is minimized. This problem is also referred to as Chains-on-Chains-
Partitioning (CCP) in the literature [31]. Herein, the weight 𝑤 𝑖 represents the
computational cost of the grid cell 𝑖 and the bottleneck value is defined as

𝐵 = max {𝐿 𝑘 } (6)
1≤𝑘 ≤𝐾
𝑠𝑘
∑︁
with the processor load 𝐿 𝑘 = 𝑤𝑗 , (7)
𝑗=𝑠𝑘−1 +1

where the symbol 𝑠 𝑘 denotes the index of the last task assigned to processor 𝑘.
Thus, SFC-based dynamic load balancing means to periodically recompute the
sequence of separator indices 𝑠0 = 0 ≤ 𝑠1 ≤ · · · ≤ 𝑠 𝐾 = 𝑁elem during runtime, by the
use of the updated chain of task weights 𝑤 𝑖 . The associated partitioning algorithms
typically supply medium-quality domain decompositions at low computational costs.
Graph partitioning, by contrast, often provides decompositions of higher quality, but
also at higher computational costs, which is not suitable for the applications with
frequent changes in the workload distribution [34].
Besides, SFC-based partitioning relies only on geometric mesh information. Graph-
based methods, on the other hand, abstract grid cells and shared cell faces as vertices
𝑉 and edges 𝐸 of an induced graph G = (𝑉, 𝐸). They thus require more detailed
information on the costs of the data and communication operations, respectively, to
accurately define G and deduce an optimal partitioning.
Dynamic load balancing for the level-set ghost-fluid method 311

Due to complex interface dynamics in compressible two-phase flows, the shape

and the position of the narrow-band region change significantly throughout the
simulation. This implies both pronounced spatial and temporal workload variations.
The necessitated frequent repartitioning at runtime, following the considerations
above, thus favors a SFC-based load balancing approach. Besides, this choice eases
the implementation over graph-based methods and allows to exploit the established
static domain decomposition framework for mesh parallelization of Hindenlang [15].
We employ the well-known, efficient H2-heuristic of Miguet and Pierson [25] with a
minor modification to minimize the deviation of the last processor’s load 𝐿 𝑃 from
the ideal bottleneck value
𝑁elem
1 ∑︁
𝐵∗ = 𝑤𝑖 . (8)
𝐾 𝑖

4.3 Code instrumentation and algorithmic details

To evaluate the processor load (7), the current computational cost of each grid cell
has to be determined. This can naturally be achieved through cell-local wall time
measurements. However, as they entail an additional runtime overhead and require
extensive implementational efforts, this approach is considered to be unsuited for
most scientific applications [27].
In contrast, Ortwein et al. [28] showed significant performance improvements for
a hybrid particle-mesh method with element-local wall time measurements compared
to a constant weighting factor which relates the workload of a grid cell and the particle
operations:
𝑤 𝑖 = 1 + 𝜈 𝑛particles,𝑖 . (9)
This ansatz assumes the total cost of one element to be composed of a constant part
through the grid cell operation itself, and a variable part, which scales linearly with
the number of particles 𝑛particles,𝑖 located on that element. As the proportionality
constant 𝜈 is in fact runtime-dependent and specific to the simulation scenario, the
authors of the cited work favored the element-local wall time measurements.
In the present work, we pursue a third, less intrusive approach, which exploits
the introduced distinction of the three element subsets outlined in the preceding
section. We only measure the total wall time of the level-set and ghost-fluid operations
listed in section 3.3 and distribute it evenly among the associated elements. This
averaging ansatz presumes that each of the masked elements contributes equally
to the total wall time of the measured subroutine. Nevertheless, it accounts for the
fact that the cost ratio of the three element subsets relative to each other is again
setup-specific. This ratio depends e.g. on the chosen numerical flux function and
two-phase Riemann solver, the polynomial degree used for the fluid and the level-set
equations, respectively [26], and the frequency of the level-set reinitialization and the
velocity extrapolation, which are not necessarily executed every iteration. Unlike in
312 D. Appel, S. Jöns, J. Keim, C. Müller, J. Zeifang and C.-D. Munz

Ortwein et al. [28], the runtime overhead of our approach for the time measurements
has proven to be negligible, as only very few calls to the code instrumentation
functions are necessary per time step.
Given the workload distribution {𝑤 𝑖 }, the processor loads 𝐿 𝑘 define the current
imbalance

𝐼max = max (𝐼 𝑘 ) , (10)

𝑘
𝐿𝑘
with the processor imbalance 𝐼 𝑘 = −1 (11)
𝐵∗
and 𝐼max = 0 indicates a perfect load distribution. The imbalance is evaluated every
𝑛th time step, where 𝑛 ≈ O (102 ) is adjusted to the temporal workload variations
in the considered simulation setup. Upon exceeding the threshold 𝐼threshold , the
imbalance triggers the actual repartitioning and the subsequent simulation restart. It
should be noted that the runtime overhead caused by the imbalance evaluation grows
significantly with increasing parallelism due to the involved collective communication
operation. The choice of evaluation period 𝑛 hence needs to account for the wall
time of the evaluation operation itself, on the one hand, and, on the other hand, the
performance gain through a repartitioning as well as the dynamics of the workload
distribution.
The dynamic load balancing algorithm can thus be summarized as follows:
𝑁elem
1. measure the element wall times {𝑤 𝑖 }𝑖=1 through outlined averaging procedure
2. compute the processor workloads {𝐿 𝑘 } 𝐾 𝑘=1
3. evaluate the current imbalance 𝐼max
4. if 𝐼max > 𝐼threshold , execute the repartitioning and restart the simulation
Fig. 1 depicts the domain decomposition before and after the repartitioning for a
two-dimensional droplet resting in a quiescent gas (v = 0 on Ω), by the use of a
representative grid resolution and solver setup.

5 Numerical results

5.1 Strong scaling behavior

In this section, we investigate the strong scaling behavior of the developed load
balancing algorithm. For reference, we include performance data of a previous code
version without the introduced element masking. All runs were performed on the
HPE Apollo System Hawk at HLRS, which is equipped with 5632 compute nodes of
two AMD EPYC 7742 CPUs (64 cores each) and 256GB memory. The internode
communication deploys the InfiniBand HDR200 interconnect with a bandwidth of
200GBit/s.
Dynamic load balancing for the level-set ghost-fluid method 313

𝑤𝑖
𝑤min 𝐼𝑘 𝐼𝑘

bulk region Dbulk

phase interface DΓ

narrow band DNB

(a) Dimensionless element wall (b) Processor imbalances before (c) Processor imbalances after
times repartitioning: 𝐼max = 0.921 repartitioning: 𝐼max = 0.023

Fig. 1: Domain decomposition for 𝐾 = 16 processors for a two-dimensional resting

droplet in a quiescent gas, using a representative solver setup and grid resolution (128
DOFs per diameter 𝐷).

The investigated test case considers a three-dimensional resting droplet in a

quiescent gas, centered in a cubic domain of edge length 5𝐷 and a resolution of 50
DOFs/𝐷.
The resolution and the solver setup are chosen to be representative of three-
dimensional compressible droplet dynamics simulations. Therefore, we restrict
ourselves to the strong scaling behavior, i.e. we keep the problem size constant
while we successively increase the number of cores. The stationary physics allow us
to isolate the performance gain and the scalability through the repartitioning. The
dynamic repartitioning is triggered after five time steps, followed by another 65 time
steps to evaluate the performance.
The main analyzed metric is the performance index PID, which expresses the wall
time required by a single core to advance one DOF for one stage of the RK time
integration scheme:
total runtime × #cores K
PID = . (12)
#DOFs × #time steps × #RK-stages
We also report the parallel efficiency to quantify the relative performance degradation
for an increasing number of cores. The parallel efficiency 𝜂 relates the obtained
performance index to the value obtained for a single compute node, that is in the
present case
PID𝐾=128
𝜂= . (13)
PID
Fig. 2a depicts the obtained PID over a range of 1 to 128 compute nodes
(𝐾 = 128, 256, . . . , 16384). The introduction of the element masking itself leads to a
performance gain in terms of the PID of approximately five for a single compute node.
314 D. Appel, S. Jöns, J. Keim, C. Müller, J. Zeifang and C.-D. Munz

#compute nodes #compute nodes

128 64 32 16 8 4 2 1 128 64 32 16 8 4 2 1
·10−5 1.2

1
4
0.8
PID

𝜂
0.6
2
0.4

0.2
103 104 105 103 104 105
#DOFs/processor #DOFs/processor

(a) Performance index (b) Parallel efficiency

Fig. 2: Strong scaling behavior for a three-dimensional resting droplet in quiescent

gas, with one compute node comprising 128 physical cores:
without element masking, unbalanced masking, balanced masking

As the domain granularity, i.e., the number of nodes, increases, this gain diminishes
and the PID approaches the reference value obtained without the masking. This
trend is to be expected, since a larger percentage of subdomains contains exclusively
narrow band elements, and thus does not benefit from the reduced workload due to
the element masking. In the present setup, for example, the narrow band region covers
only 1% of the elements. Through the dynamic repartitioning, the performance gain
is preserved over the entire range of cores.
Fig. 2b evaluates the parallel efficiency of the observed PID curves. In general,
the implementation benefits from the excellent scaling behavior of the FLEXI
framework [21] which it is built upon. The strong scaling behavior is governed by two
opposing effects: For decreasing processor workloads (#DOFs/processor), the memory
requirement per core is reduced such that a larger proportion of the used data can be
stored in the fast cache of the CPU, which increases the performance. Concurrently, the
communication effort raises relative to the core-local work and eventually outweighs
the caching effect. This gives rise to an optimal processor workload, which is reached
at 4048 DOFs/processor for the load balancing implementation. The significant
drop in the parallel efficiency to 𝜂 = 0.25 for the element masking alone underlines
the need for a dynamic load balancing. In total, a compute time of approximately
2900CPUh was spent on the depicted scaling results.
It should be noted that the runtime overhead through the wall time measurements
is negligible, in particular for three-dimensional setups. The additional runtime for
evaluating the imbalance, recomputing the partitioning and executing the simulation
restart is overcompensated by the reduced wall time, that is gained in return, after 40
time steps on average for the present test case.
Dynamic load balancing for the level-set ghost-fluid method 315

5.2 Shock-droplet interaction

In contrast to the rather generic setup with a static workload distribution in the
preceding section, we now apply the developed load balancing method to more
complex and realistic scenarios featuring temporal workload variations.
We first revisit the two-dimensional shock-droplet interaction in Jöns et al. [19], that
was previously studied by Winter et al. [35]. The authors of those works investigated
the deformation of an initially cylindrical water column after the interaction with a
shock in air. In the first stage of the deformation, the water column is flattened. This
flattening is independent of the Weber number

𝜌 𝑉 (𝑣 𝑉 ) 2 𝐷
We = , (14)
𝜎
which expresses the relative importance of the inertial to the surface tension forces.
The second stage starts with the onset of interfacial instabilities at the air-water
interface, where different droplet breakup modes can be distinguished depending on
the Weber number. At relatively small surface tension forces (We small), breakup
occurs in the shear-induced entrainment (SIE) regime, which is characterized by the
formation of a thin sheet at the top and the bottom droplet equator. Large surface
tension forces (We large), in contrast, suppress the growth of these sheets and have a
smoothing effect on the interface geometry. This so-called Rayleigh–Taylor piercing
(RTP) regime results in a bag-shaped droplet and is considered below due to the
negligible effect of the viscous forces.
We adopt the setup of Jöns et al. [19] and initialize a shock wave of Ma = 1.47
upstream of the water droplet with a Weber number of We = 12. The droplet is
discretized with 170 DOFs/𝐷 and we impose a symmetry boundary condition along
the symmetry axis. The numerical domain contains 106 DOFs and we deploy 4
compute nodes, which yields a nominal processor workload of 2048 DOFs and results
in a total compute time of 1200CPUh. For further details on the numerical setup,
the material parameters and the initial conditions as well as a more comprehensive
discussion of the involved physics, the reader is referred to the cited publications. In
Fig. 3, we provide the results at the dimensionless time 𝑡 ∗ = 1.58 to illustrate the impact
of the dynamic repartitioning. Initially, the domain is partitioned homogeneously
as no a priori knowledge of the workload distribution is available. The depicted
partitioning then reflects the current workload distribution, with the narrow band
elements being computationally two orders of magnitude more expensive than the
bulk elements. The performance gain through the dynamic load balancing for this
realistic simulation is expressed more adequately in terms of the overall runtime
saving; it has been evaluated for a coarser mesh with 42 DOFs/𝐷 and amounts to
64%.
Ultimately, we extend the given two-dimensional setup to three dimensions. The
resolution is reduced to 42 DOFs/𝐷, which results in a numerical domain of 8 × 106
DOFs. We deploy 32 compute nodes to attain the same nominal processor workload
as before. The simulation finished after a total compute time of 8200CPUh. Fig. 4
316 D. Appel, S. Jöns, J. Keim, C. Müller, J. Zeifang and C.-D. Munz

𝑤𝑖
𝑤min

𝑣1∗

Fig. 3: Two-dimensional shock-droplet interaction for Ma = 1.47, We = 12 at

𝑡 ∗ = 1.58: dimensionless element wall times (top) and dimensionless axial velocity
(bottom).

depicts the droplet contour together with the non-dimensional velocities in axial
and radial direction, 𝑣 ∗1 and 𝑣 ∗2 , respectively. The droplet exhibits a small concave
deformation on the upstream side and has a convex shape on the downstream side,
which indicates the onset of the characteristic bag growth [35]. An axially stretched
recirculation zone develops in the wake of the drop, however, with a less pronounced
detachment than in the cited work. The deviation may be attributed to the unmatched
and insufficient resolution as well as the neglected interface viscosity. The overall
runtime reduction by 69% emphasizes the effectiveness of the proposed dynamic
load balancing strategy.

6 Conclusion

We have presented a dynamic load balancing scheme for a high-order level-set ghost-
fluid method used for compressible two-phase flow simulations. The load imbalance
arises from introducing an element masking that applies the costly interface-tracking
routines only to the necessary grid cells. The load balancing scheme is based on a
static domain decomposition through the Hilbert space-filling curve and employs an
efficient heuristic for the dynamic repartitioning. The current workload distribution
is determined through element-local wall time measurements, which exploit the
masking approach for a minimally intrusive code instrumentation with negligible
runtime overhead. The developed load balancing scheme transfers the single-core
Dynamic load balancing for the level-set ghost-fluid method 317

𝑣1∗

𝑣2∗

Fig. 4: Three-dimensional shock-droplet interaction for Ma = 1.47, We = 7339

at 𝑡 ∗ = 1.58: dimensionless axial velocity (top) and dimensionless radial velocity
(bottom).

performance gain from the masking approach to massively parallelized simulations.

We provided strong scaling results, which reveal a parallel efficiency beyond 65 %
and superlinear scaling at the optimal processor workload. Compared to unbalanced
simulations without element masking, the performance gain amounts to approximately
five throughout the investigated range up to 16384 cores. We have demonstrated
the capability of the dynamic repartitioning scheme for two- and three-dimensional
shock-droplet interactions in the Rayleigh–Taylor piercing regime with an overall
runtime reduction by 64% and 69%, respectively.
The runtime overhead of the load balancing scheme is mainly attributed to the
parallel I/O associated with the simulation restart, which requires O (101 . . . 102 ) time
steps to be compensated through the induced performance gain. This amortization
rate could be reduced – i.e. the repartitioning could be called more frequently – by
recalculating only the static data like mesh metrics upon restart, while the dynamic
data (state variables) are migrated via point-to-point communication. Furthermore,
the current imbalance is evaluated periodically, which causes a growing runtime
overhead with increasing parallelism due to the involved collective communication.
Since the value of the evaluation period relies on an educated guess, future work may
adapt the idea of Menon et al. [23] to determine it automatically at runtime based on
application characteristics.

Acknowledgements We gratefully acknowledge the Deutsche Forschungsgemeinschaft (DFG)

with their support of D. Appel, C. Müller and J.Zeifang through the project GRK 2160/1 “Droplet
Interaction Technologies”, S. Jöns through the project SFB-TRR 75, Project number 84292822
-“Droplet Dynamics Under Extreme Ambient Conditions”, J. Keim and C.-D. Munz under Germany’s
Excellence Strategy - EXC 2075 - 390740016. The simulations were performed on the national
318 D. Appel, S. Jöns, J. Keim, C. Müller, J. Zeifang and C.-D. Munz

supercomputer HPE Apollo System Hawk at the High Performance Computing Center Stuttgart
(HLRS) under the grant number hpcmphas/44084, of which we used approximately 2.7M CPUh in
the current accounting period.

References

1. Tariq D. Aslam. A partial differential equation approach to multidimensional extrapolation.

Journal of Computational Physics, 193(1):349–355, 2004.
2. Andrea D. Beck, Thomas Bolemann, David Flad, Hannes Frank, Gregor J. Gassner, Florian
Hindenlang, and Claus-Dieter Munz. High-order discontinuous Galerkin spectral element
methods for transitional and turbulent flow simulations. International Journal for Numerical
Methods in Fluids, 76(8):522–548, 2014.
3. Andrea D. Beck, Thomas Bolemann, David Flad, Nico Krais, Jonas Zeifang, and Claus-Dieter
Munz. Application and development of the high order discontinuous galerkin spectral element
method for compressible multiscale flows. In High Performance Computing in Science and
Engineering ’18, pages 291–307. Springer International Publishing, 2019.
4. Carsten Burstedde, Omar Ghattas, Michael Gurnis, Tobin Isaac, Georg Stadler, Tim Warburton,
and Lucas Wilcox. Extreme-scale AMR. In SC’10: Proceedings of the 2010 ACM/IEEE
International Conference for High Performance Computing, Networking, Storage and Analysis,
pages 1–12. IEEE, 2010.
5. Manuel Castro, José Gallardo, and Carlos Parés. High order finite volume schemes based
on reconstruction of states for solving hyperbolic systems with nonconservative products.
applications to shallow-water systems. Mathematics of Computation, 75(255):1103–1134,
2006.
6. Michael Dumbser and Raphaël Loubère. A simple robust and accurate a posteriori sub-cell
finite volume limiter for the discontinuous Galerkin method on unstructured meshes. Journal
of Computational Physics, 319:163–199, 2016.
7. Stefan Fechter, Felix Jaegle, and Veronika Schleper. Exact and approximate Riemann solvers at
phase boundaries. Computers & Fluids, 75:112–126, 2013.
8. Stefan Fechter and Claus-Dieter Munz. A discontinuous Galerkin-based sharp-interface method
to simulate three-dimensional compressible two-phase flow. International Journal for Numerical
Methods in Fluids, 78(7):413–435, 2015.
9. Stefan Fechter, Claus-Dieter Munz, Christian Rohde, and Christoph Zeiler. A sharp interface
method for compressible liquid-vapor flow with phase transition and surface tension. Journal
of Computational Physics, 336:347–374, 2017.
10. Stefan Fechter, Claus-Dieter Munz, Christian Rohde, and Christoph Zeiler. Approximate
Riemann solver for compressible liquid vapor flow with phase transition and surface tension.
Computers & Fluids, 169:169–185, 2018.
11. Ronald P. Fedkiw, Tariq Aslam, Barry Merriman, and Stanley Osher. A non-oscillatory Eulerian
approach to interfaces in multimaterial flows (the ghost fluid method). Journal of Computational
Physics, 152(2):457–492, 1999.
12. Fabian Föll, Timon Hitz, Christoph Müller, Claus-Dieter Munz, and Michael Dumbser. On the
use of tabulated equations of state for multi-phase simulations in the homogeneous equilibrium
limit. Shock Waves, 2019.
13. Pablo Gómez, Julio Hernández, and Joaquín López. On the reinitialization procedure in a
narrow-band locally refined level set method for interfacial flows. International Journal for
Numerical Methods in Engineering, 63(10):1478–1512, 2005.
14. Daniel F. Harlacher, Harald Klimach, Sabine Roller, Christian Siebert, and Felix Wolf. Dynamic
load balancing for unstructured meshes on space-filling curves. In 2012 IEEE 26th International
Parallel and Distributed Processing Symposium Workshops & PhD Forum. IEEE, 2012.
15. Florian Hindenlang. Mesh Curving Techniques for High Order Parallel Simulations on
Unstructured Meshes. PhD thesis, Universität Stuttgart, 2014.
Dynamic load balancing for the level-set ghost-fluid method 319

16. Florian Hindenlang, Gregor J. Gassner, Christoph Altmann, Andrea D. Beck, Marc Staudenmaier,
and Claus-Dieter Munz. Explicit discontinuous Galerkin methods for unsteady problems.
Computers & Fluids, 61(0):86–93, 2012.
17. Timon Hitz, Steven Jöns, Matthias Heinen, Jadran Vrabec, and Claus-Dieter Munz. Comparison
of macro- and microscopic solutions of the Riemann problem II. Two-phase shock tube. Journal
of Computational Physics, 429:110027, 2021.
18. Guang-Shan Jiang and Danping Peng. Weighted ENO schemes for Hamilton–Jacobi equations.
SIAM Journal on Scientific Computing, 21(6):2126–2143, 2000.
19. Steven Jöns, Christoph Müller, Jonas Zeifang, and Claus-Dieter Munz. Recent advances
and complex applications of the compressible ghost-fluid method. In Muñoz-Ruiz, María
Luz, Carlos Parés, and Giovanni Russo, editors, Recent Advances in Numerical Methods for
Hyperbolic PDE Systems. SEMA SIMAI Springer Series, volume 28, pages 155–176. Springer,
Cham, 2021.
20. David A. Kopriva. Implementing spectral methods for partial differential equations: Algorithms
for scientists and engineers. Springer Publishing Company Incorporated, 1st edition, 2009.
21. Nico Krais, Andrea D. Beck, Thomas Bolemann, Hannes Frank, David Flad, Gregor Gassner,
Florian Hindenlang, Malte Hoffmann, Thomas Kuhn, Matthias Sonntag, and Claus-Dieter Munz.
FLEXI: A high order discontinuous Galerkin framework for hyperbolic-parabolic conservation
laws. Computers & Mathematics with Applications, 81:186–219, 2021.
22. Matthias Lieber and Wolfgang E. Nagel. Highly scalable SFC-based dynamic load balancing
and its application to atmospheric modeling. Future Generation Computer Systems, 82:575–590,
2018.
23. Harshitha Menon, Nikhil Jain, Gengbin Zheng, and Laxmikant Kalé. Automated load balancing
invocation based on application characteristics. In 2012 IEEE International Conference on
Cluster Computing, pages 373–381. IEEE, 2012.
24. Christian Merkle and Christian Rohde. The sharp-interface approach for fluids with phase
change: Riemann problems and ghost fluid techniques. ESAIM: Mathematical Modelling and
Numerical Analysis, 41(06):1089–1123, 2007.
25. Serge Miguet and Jean-Marc Pierson. Heuristics for 1d rectilinear partitioning as a low
cost and high quality answer to dynamic load balancing. In International Conference on
High-Performance Computing and Networking, pages 550–564. Springer, 1997.
26. Christoph Müller, Timon Hitz, Steven Jöns, Jonas Zeifang, Simone Chiocchetti, and Claus-
Dieter Munz. Improvement of the level-set ghost-fluid method for the compressible Euler
equations. In Grazia Lamanna, Simona Tonini, Gianpietro Elvio Cossali, and Bernhard Weigand,
editors, Droplet Interaction and Spray Processes. Springer, Heidelberg, Berlin, 2020.
27. Ansgar Niemöller, Michael Schlottke-Lakemper, Matthias Meinke, and Wolfgang Schröder.
Dynamic load balancing for direct-coupled multiphysics simulations. Computers & Fluids,
199:104437, 2020.
28. P Ortwein, T Binder, S Copplestone, A Mirza, P Nizenkov, M Pfeiffer, C.-D. Munz, and
S Fasoulas. A load balance strategy for hybrid particle-mesh methods. arXiv preprint
arXiv:1811.05152, 2018.
29. Philip Ortwein, Tilman Binder, Stephen Copplestone, Asim Mirza, Paul Nizenkov, Marcel
Pfeiffer, Torsten Stindl, Stefanos Fasoulas, and Claus-Dieter Munz. Parallel performance of
a discontinuous Galerkin spectral element method based PIC-DSMC solver. In Wolfgang E.
Nagel, Dietmar H. Kröner, and Michael M. Resch, editors, High Performance Computing in
Science and Engineering ‘14, pages 671–681, Cham, 2015. Springer International Publishing.
30. Per-Olof Persson and Jaime Peraire. Sub-cell shock capturing for discontinuous Galerkin
methods. In 44th AIAA Aerospace Sciences Meeting and Exhibit. American Institute of
Aeronautics and Astronautics, 2006.
31. Ali Pınar and Cevdet Aykanat. Fast optimal load balancing algorithms for 1d partitioning.
Journal of Parallel and Distributed Computing, 64(8):974–996, 2004.
32. Matthias Sonntag and Claus-Dieter Munz. Efficient parallelization of a shock capturing for
discontinuous Galerkin methods using finite volume sub-cells. Journal of Scientific Computing,
70(3):1262–1289, 2016.
320 D. Appel, S. Jöns, J. Keim, C. Müller, J. Zeifang and C.-D. Munz

33. Mark Sussman, Peter Smereka, and Stanley Osher. A level set approach for computing solutions
to incompressible two-phase flow. Journal of Computational Physics, 114(1):146–159, 1994.
34. James D Teresco, Karen D Devine, and Joseph E Flaherty. Partitioning and dynamic load
balancing for the numerical solution of partial differential equations. In Numerical Solution of
Partial Differential Equations on Parallel Computers, pages 55–88. Springer, 2006.
35. Josef M. Winter, Jakob W.J. Kaiser, Stefan Adami, and Nikolaus A. Adams. Numerical
investigation of 3d drop-breakup mechanisms using a sharp interface level-set method. In 11th
International Symposium on Turbulence and Shear Flow Phenomena, TSFP 2019, 2019.
36. Jonas Zeifang. A Discontinuous Galerkin Method for Droplet Dynamics in Weakly Compressible
Flows. PhD thesis, Universität Stuttgart, 2020.
Numerical simulation of flake orientation during
droplet impact on substrates in spray painting
processes

Qiaoyan Ye and Martin Dreher and Bo Shen

Abstract A numerical study of flake orientation by droplet impact on dry and wetted
solid surfaces for spray painting processes has been carried out. A dynamic contact
angle model was applied for the calculation of viscous droplet impact on the dry
surface after experimental validation. A user-defined 6-DOF model that concerns to
the rigid body motion was implemented in a commercial CFD program to calculate the
flake movement inside the droplet. The simulated flake orientations show interesting
results that are helpful to understand and to improve painting processes.

1 Introduction

Effect pigments (flat, 5 µm to 40 µm broad flakes) are widely used in industrial

spray painting processes for imparting colour and coating appearance, enabling e.g.
metallic effect coatings that are very popular in the automotive industry. From practical
observations, it is already well known that the initial pigment/flake orientation in the
paint layer, primarily at the early stage of the film formation and before the subsequent
solvent evaporation and baking, influences the final metallic effect significantly.
Therefore, it is strongly suspected that processes during viscous droplet impact have
a decisive influence on flake orientation. For a strong metallic effect, usually a flake
orientation parallel to the substrate is desired.
Regarding the droplet impact dynamics, there have been already many research
works [1–8]. In experimental studies, mainly large droplets (e.g. 𝐷 > 500 µm) with
low viscosity under hydrophilic/hydrophobic surface conditions were investigated.

Qiaoyan Ye and Martin Dreher

Fraunhofer Institute for Manufacturing Engineering and Automation, Stuttgart, Germany,
e-mail: [email protected]
Bo Shen
University of Applied Sciences Esslingen, Esslingen, Germany,
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 321
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_19
322 Q. Ye, M. Dreher and B. Shen

Different outcomes of droplet impact on dry/wet substrates were analysed [4]. The
air entrapment by droplet impact was studied [3]. The maximal droplet-spreading
diameters were correlated with non-dimensional numbers, i.e., Weber and Reynolds
numbers, using published experimental results [5]. The contact angle models that
have to be applied in the numerical simulation were studied and discussed in detail
[7, 8].
There are not so many studies that focus on the flake movement inside the droplet
and in the film. For small droplets (50 µm to 300 µm), especially for opaque liquids,
like in spray painting processes, it is very difficult to obtain high quality time-resolved
imaging of the flake orientation during droplet impingement experimentally. Although
a few predictions of flake orientation were carried out based on some mathematical
analysis [9] and the flow field of two-dimensional droplet impact calculation [10],
the corresponding results should be further verified.
The objective of the present paper is to carry out a detailed numerical study of
flake orientation, focusing on droplet impact processes on dry and wetted solid walls.
Thereby, a contact angle model that is necessary in the numerical simulation was
developed for the paint droplet impact process. In addition, a user-defined 6DOF
(6-degrees-of-freedom) solver was implemented in a CFD-program to perform the
rigid body (flake) motion calculation within the impacting droplet. The developed
models were applied in a parameter study, to further clarify the existing dependencies
on application and fluid parameters more quantitatively.

2 Basic numerical methods

The droplet impact and spreading on a surface is an example of an interfacial flow

problem that can be calculated using the Volume of Fluid (VOF) method. Hereby,
two or more immiscible fluids can be modelled by solving a single set of momentum
and mass equations and tracking the volume fraction of each of the fluids throughout
the domain. The numerical simulations in this work were carried out with the
commercial CFD code ANSYSFluent based on the finite-volume approach. Time
dependent VOF calculations were performed using an explicit scheme. A geometric
reconstruction scheme for the volume fraction discretization was used, ensuring a
sharp and low-diffusion interface discretization. To accommodate surface tension
effects, the CSF-model with wall adhesion modelling was used. The PRESTO scheme
was applied for the pressure discretization.
It is well known that the VOF-method is quite sensitive to grid resolution. A study
of grid independence was performed by comparing measured and calculated spread
and height factors of the droplet impact. A reasonable cell size was obtained at about
𝐷/Δ𝑥 = 150 and applied around the droplet and liquid film on the substrate. Here 𝐷
denotes the droplet diameter and Δ𝑥 the grid size. In the far-field, coarser hexahedral
meshes were used to reduce the total number of cells and therefore computational
cost. For the simulation of droplet impact on the wet solid surface, a relative large
computational domain had to be used. In order to reduce further computational cost,
Numerical simulation of flake orientation during droplet impact 323

a dynamic mesh adaption was applied, which was necessary especially for large
droplets. Consequently, the domains in the present study contain 20 to 150 million
cells. Further detailed important boundary conditions and models are described as
follows.

2.1 Simulation of viscous droplet impact on dry solid surfaces

A computational domain of 2𝐷 × 2𝐷 × 1.5𝐷 was created for the simulation of droplet

impact on the dry solid surface. Thereby, only a quarter of the spherical droplet was
calculated and symmetry boundary conditions were used to reduce the total amount of
cells. A flow field aware variable time step with the CFL (Courant–Friedrichs–Lewy)
number < 1 was set. The resulting time step sizes ranged from 5 × 10−9 s to 10−7 s.
Based on common spray-painting conditions, droplet impact velocities between
0.5 m·s−1 to 10 m·s−1 were applied. The initial droplet position was chosen to be
a few micrometers above the wall surface, so that the surrounding gas field can be
calculated, which is necessary in the study of droplet impact dynamics. The pressure
inside the droplet, induced by the surface tension of the liquid, was calculated and
properly initialized. The remaining domain is set to ambient pressure. A dry smooth
wall with no-slip boundary condition was used.
It is well known that the description of moving contact line or contact angle
in the numerical simulation of droplet impact on dry surfaces is not trivial and
quite challenging. Nevertheless, dynamic contact angles are widely used in VOF-
simulations to accommodate adhesion behaviour. However, the applicability of the
corresponding models is quite problem dependent, such as the variety of liquid and
substrate properties, as well as the numerous operating parameters. The hysteresis of
contact line and contact angle makes the model development more difficult. Further
information can be referred to [7, 8].
We have carried out the experimental observation of viscous droplet impact on dry
solid surfaces, focusing on the contour evolution during the spreading and receding
processes. The experimental results provided useful information to develop the model
for dynamic contact angle that was applied as boundary condition in the simulation.
Equation (1) describes the model in a mathematical manner,

𝜃A

 for 𝑣 cl > 0

 𝜃 int = 90°

 for 𝑣 cl < |𝛿| ∩ 𝑣 cl < 𝛿
𝜃𝐷 = (1)
 𝜃 ∇𝜙

 for 𝜃 ∇𝜙 < 𝜃 A

𝜃 for 𝜃 ∇𝜙 ≤ 𝜃 E
 E


where 𝛿 = −0.1. In the spreading phase, where the contact line velocity (𝑣 cl )
is significantly positive, a constant advancing contact angle (𝜃 A ) is deployed. At
maximum wetting spread, when the contact line movement stops and the droplet
usually starts to recede, an intermediate contact angle (𝜃 int ) is set. This helps to
324 Q. Ye, M. Dreher and B. Shen

suppress contact line movement in accordance to the experimental observation. If

the angle between the fluid interface and the wall, calculated by the gradient of the
VOF-field (∇𝜙), gets smaller than the advancing contact angle, the interphase-angle
is set as contact angle (𝜃 ∇𝜙 ). It is secured, that this angle decreases monotonically to
the static or equilibrium contact angle (𝜃 E ). On the wall cells and far away from the
contact line, the equilibrium contact angle (𝜃 E ) is used, which ensures the wetting
behaviour on the wall is more reasonable.

2.2 Simulation of viscous droplet impact on wet solid surfaces

A thin film with a height 𝐻 = 30 µm to 60 µm and an infinitive width was applied in

the present simulation. At first, droplet impact on the wet solid surface was calculated
in a two-dimensional axisymmetric coordinates, in order to track the gas-liquid
interface with a large amount of parameters and to define the impact events with some
dimensionless numbers, such as Re-, Oh- and We-number. Then, a computational
domain of 8𝐷 × 8𝐷 × 3𝐷 was created for the simulation of droplet impact with
flake on the wet solid surface. Clearly, the number of computational cells will be
tremendously increased by homogeneous grid because of the large domain required.
Therefore, local grid adaption and dynamic mesh adaption were carefully performed,
ensuring the necessary grid resolution. For large droplets, e.g., 300 µm, the region
with inhomogeneous grid could be enlarged because of the wave, which results in
also an increasing the computational cells. As shown in Fig. 1, for instance, the
mesh size around the interface between two phases was set to be 1 µm obtained by
using 𝐷/Δ𝑥 = 150 for a 150 µm droplet. Far away from the interface, a coarse mesh
with size of 16 µm was used, finally resulting a four-lever mesh adaption around the
interface.

Fig. 1: Mesh size distribution for the droplet impact with 𝐷=150 µm by using dynamic
mesh adaption model.
Numerical simulation of flake orientation during droplet impact 325

2.3 Simulation of flake orientation

For the simulation of flake movement, the effect pigment (flake) was incorporated
into the domain mesh utilizing the overset meshing option as a dynamic mesh method.
Hereby, a separate component mesh that may be adapted to its special geometry, is
necessary. As shown in Fig. 2, Cartesian grids with local mesh-refinement in the
relevant regions were created for the overall simulation domain. The component
mesh for the flake (Fig. 2b) and the background grid are overlaid as show in Fig. 2a
and coupled by an interpolated transition area, in which the flow solutions that are
separately calculated in both grids are exchanged. The resolution of the component
grid corresponds to that of the background grid. Comparing to other methods, such
as remeshing method, the overset meshing yields VOF-friendly meshing setup and
a good performance in the flake movement calculation because of the elimination
of distorted cells and the necessity of remeshing. In the present study, only one

(a) (b)

Fig. 2: a) 3D mesh setup with boundary conditions and initial droplet position,
b) flake/component mesh.

rectangular shaped flake in the computational domain is applied. The size of the
considered flake was chosen to be 1 × 16 × 16 µm3 that is a reasonable estimate
concerning typical effect pigment size distributions. Yet, this results in corresponding
moments of inertia of 1.6 × 10−23 kg·m2 , which are too marginal for the currently
available 6DOF motion solver by ANSYS. In consequence, a custom motion equation
solver has been implemented via a user-defined-function (UDF). This 6DOF solver
has been developed under the assumption of rigid body motion. Therefore, the flake
motion may be simplified to the movement of its centre of mass and the rotation
around it. Consequently, the momentum conservation (2) is calculated in the global
inertial coordinate system, and the angular momentum (3) is determined in body
coordinates:
1
v¤ = ΣF (2)
𝑚
𝜔¤ = I−1 ( ΣM − 𝝎 × I𝝎 ) (3)
326 Q. Ye, M. Dreher and B. Shen

Here, 𝑚 denotes the mass, I the inertia tensor and F, M the flow induced forces
and moments, respectively. The points over the velocity (v) and the angular velocity
(𝝎) represent time derivatives. In Equation (2), the sum includes pressure, viscosity
and gravitational forces. An Adams–Moulton algorithm of 4th order with a rather
complex variable time step formulation has been derived and implemented to time
integrate the above given ODEs. This enables simulations of flake movement or
orientation stable with a quite reasonable time step size, such as 𝑑𝑡 = 10−8 s to 10−7 s
that is solely adjusted by the flow solver.

3 Results and discussion

3.1 Evolution of interface of gas-liquid by droplet impact on wet solid

surface without flake

As stated in the introduction, there are many studies about outcomes of droplet
impact on wet substrates. Depending on the impact velocity and liquid viscosity,
basically, the outcome of droplet impact on a thin liquid film can be divided into
spreading/deposition, cratering/receding, crown formation and crown with splashing.
The latter will not be studied in the present investigation because it is very seldom
in spray coating applications. Typical mean diameter of spray droplet ranges from
30 µm to 50 µm. Considering the correlation between droplet impact velocity and
droplet diameter using different atomizers in spray painting processes [11], droplets
with diameter of less than 100 µm and the impact velocities up to 10 m·s−1 were used
in the simulation. The model paint was used and the rheological parameters that

Fig. 3: Phenomena of drop impact on a wet solid surface (red: liquid, blue: air). Film
𝐻 = 30 µm. a) 𝐷 = 30 µm, 𝑈 = 4 m·s−1 , b) 𝐷 = 45 µm, 𝑈 = 4 m·s−1 , c) 𝐷 = 80 µm,
𝑈 = 8 m·s−1 .
Numerical simulation of flake orientation during droplet impact 327

describe the shear thinning viscosity of the model paint were summarized in our
previous work [12]. Figure 3 shows typical outcomes of the paint droplet impact on
a thin liquid film. By small droplet and with a low impact speed the liquid of the
droplet deposits subsequently in the liquid layer (Fig. 3a). With increasing impact
energy of the droplet, namely either with a larger diameter or with a higher velocity,
a small crater can be formed in the target liquid layer (see Fig. 3b). If the impact
energy is high enough, at the circumference of the crater the so-called crater rim or
crown rises above the original surface of the target liquid layer (Fig. 3c).

Fig. 4: Droplet impact formations at different Reynolds and Ohnesorge numbers, the
ratio between film thickness and drop diameter, 𝐻/𝐷 = 0.375 to 2.16.

Fig. 5: Crater sizes as a function of the Reynolds number.

328 Q. Ye, M. Dreher and B. Shen

The dependence of √︁ the outcomes on the Reynolds and Ohnesorge numbers

(Re = 𝜌𝑢𝐷/𝜂, Oh = 𝜂/ (𝜌𝜎𝐷)) are depicted in Fig. 4. Based on the simulation
results, the dimensionless crater sizes, namely the ratio between the crater diameter
𝐷 𝑐 and the droplet diameter 𝐷 are plotted as a function of the Reynolds number in
Fig. 5, indicating a strong correlation with the Reynold number. The corresponding
quantitative correlations for curves in both Fig. 4 and 5 were obtained in [12]. The
simulation results in this section deliver useful information for further investigations,
such as flake orientation during droplet impact on wet solid surfaces, entrainment of
air bubbles in the impact process as well as film leveling.

3.2 Validation of simulation of droplet impact on dry solid surface

without flake

Validations of the numerical models were carried out for the impact process of viscous
droplet on dry solid surfaces. Thereby, using a droplet generator, glycerol/water
droplets (𝐷 = 400 µm, 𝜂 = 20 mPa·s, 𝜎 = 0.063 N·m−1 , 𝜃 E = 55◦ ) and paint droplets
(𝐷 = 300 µm, 𝜂 = 17 mPa·s, 𝜎 = 0.025 N·m−1 , 𝜃 E = 53◦ ) were created. Subsequently,
the droplet spread-factors 𝑑/𝐷 0 and height ratios ℎ/𝐷 0 induced by the impact process
that was recorded using a high-speed camera were measured and calculated. The
advancing angles in the simulation were set for glycerol/water droplet 𝜃 𝐴 = 120◦ and
paint droplet 𝜃 𝐴 = 95◦ , respectively. Figure 6 shows a comparison of these values
between the experimental data and the simulation results. A quite good agreement
can be observed, which indicates that the proposed dynamic contact angle model
delivers good performance and is suitable for the following further application.

Fig. 6: Comparison of droplet spread-factors between experiment and simulation

during the impact process.
Numerical simulation of flake orientation during droplet impact 329

Table 1: Parameters used in the simulation for Newtonian drops.

Parameter Range Parameter Range

Droplet diameter D (µm) 50, 100 and 300 Flake size (µm3 ) 1 × 16 × 16
Droplet velocity U (m·s−1 ) 0.5 to 10 Flake density (kg·m−3 ) 3200
Liquid viscosity 𝜂 (Pa·s) 0.01, 0.02 and Ratio of flake to droplet 𝑙/𝐷 0.053 and 0.32
0.04
Surface tension 𝜎 (N·m−1 ) 0.025 and 0.063 Reynolds-number, 5 to 100
𝑅𝑒 = 𝜎𝑈𝐷/𝜂
Liquid density 𝜌 (kg·m−3 ) 1020 and 1200 Weber-number, 3 to 1200
𝑊𝑒 = 𝜎𝑈 2 𝐷/𝜎
Static contact angle (◦ ) 50 to 60 Ohnesorge-number,
√ 0.1 to 2.2
𝑂ℎ = 𝜂/ 𝜎𝐷𝜌
Height of film (µm) 60 and 30

3.3 Flake orientation by droplet impact on dry solid surfaces

Considering the shear thinning and thixotropic non-Newtonian behaviour of most

paint liquids, fluids with a constant viscosity between 10 mPa·s to 40 mPa·s were
chosen that represent a good approximation of the viscosity of typical paint materials
at high shear rate by the atomization as well as the impact processes. Moreover, a
surface tension of 𝜎 = 0.025 N·m−1 was used in the simulations for paint liquids.
Depending on different atomizer and droplet size, the droplet impact velocity could
be from 0.5 m·s−1 to 50 m·s−1 [11]. Nevertheless, velocity < 10 m·s−1 was used, since
a representative mean diameter of paint droplet is usually smaller than 50 µm. Table 1
summarizes some parameters used in this study. The position and orientation of effect
pigments within a droplet may be randomly distributed. For the present simulation,
the flake has initially been located at the vertical centre at one half of the droplet
radius and oriented vertically to the solid surface. This initial orientation angle is
considered as the worst-case whilst a horizontal orientation as optimal in the resulting
film. In order words, the ideal flake is aligned parallel to the solid surface after the
curing process. Figure 7 shows the contour evolution and the flake orientation during
a droplet impact process. The colour bar gives also the development of the contact
angle based on Eq. (1) during the droplet impact process. Figure 7a) depicts the initial
state. The inertia driven spreading phase, as shown in Fig 7b) and c), turns the flake
quite substantially. In Fig. 7d) the maximum spreading is reached and surface tension
effects become dominant. During this stage, despite droplet contour movement, the
mentioned contact line hysteresis is observed, which is identical to our experimental
findings. Here, the flake is found to increase its orientation angle again. This is partly
reversed during recoiling and whilst the final equilibrium state is reached, as shown
in Fig 7 e), f), g) and h). Air entrapment by droplet impact and air bubble (blue blobs)
330 Q. Ye, M. Dreher and B. Shen

development on the wall, namely at the beginning a thin air layer/small bubbles on
the wall and late large bubbles with time, can also be observed in Fig. 7, which is
another topic and will not be discussed in the present report.
The effects of the droplet liquid properties and impact parameters on the flake
orientation are obtained and shown in Fig. 8. Glycerol/water and paint drops with
𝜂 = 20 mPa·s were applied. For small droplets, such as 𝐷 = 50 µm, the quasi-static
state is reached at 𝑡 = 0.3 ms. Since the size ratio between flake and droplet diameter
is relative large, 𝑙/𝐷 = 0.32 for 50 µm droplet, the movement of the flake is quite
limited. For 300 µm droplets, a larger change in orientation angle can be observed.
It takes a relative long time to reach the static state for large We- or Re-number.
The final flake angle in the quasi-static state depends, presumably, mainly on the
Re-number. It seems that the flake orientation improves with increasing Re-number,
which should be further studied by increasing the parameter variation.

Fig. 7: Evolution of the droplet contour and flake orientation during dry wall droplet
impact (𝐷 = 100 µm, 𝜂 = 20 mPa·s, 𝜎 = 0.063 N·m−1 , 𝜃 E = 55◦ , 𝑈 = 6 m·s−1 ,
𝑊 𝑒 = 68.6, 𝑅𝑒 = 36).

The interaction between flake and flow field may be observed more detailed
in Figure 9. At 𝑡 = 0.17 ms (Fig. 9c), there is no macroscopic contour movement
left, which is referred to as quasi-static state. Although the velocity magnitude is
quite small, there are eddies arising around the pigment (Fig. 9d, c), resulting in
some disorientation of the flake. This effect requires further investigation. Also, the
entrapment of air bubbles on the solid surface, which are later noticed also in the
paint film, can be seen in Fig. 7 and Fig. 9.
Numerical simulation of flake orientation during droplet impact 331

Fig. 8: Comparison of flake orientation with different viscous droplets.

(a) (b)

(c) (d)

Fig. 9: Flake within coat droplet (𝐷 = 100 µm, 𝜃 E = 53◦ , 𝑈 = 6 m·s−1 , 𝑊 𝑒 = 153.3,
𝑅𝑒 = 30.7); overlaid with velocity vector (m·s−1 ) and streamline: a) advancing
process b) receding process, c) and d) quasi-static state.

3.4 Flake orientation by droplet impact on wet solid surfaces

Flake orientation by droplet impact on the wet surface is more concerned than on the
dry surface, since the optical properties is mainly influenced by flakes located close
to the film surface. Since the simulation is quite time consuming for large droplet
because of relative large computational domain in this case, detailed parameter studies
were carried out with droplet with diameter of 50 µm.
332 Q. Ye, M. Dreher and B. Shen

Typical spreading and cratering processes are shown in Fig. 10 and Fig. 11,
respectively. The crater occurs by increasing the impact velocity. With deepening the
crater, as shown in Fig. 11, the flake angle decrease and even tends to parallel to the
solid surface. However, by crater receding the flake angle increases again.

Fig. 10: Evolution of droplet-, film-contour, and flake orientation (black ruler
corresponds to 50 µm), (𝐷 = 50 µm, 𝜂 = 20 mPa·s, 𝜎 = 0.063 N·m−1 , 𝑈 = 2 m·s−1 ,
𝑊 𝑒 = 3.8, 𝑅𝑒 = 6, 𝑂ℎ = 0.33).

Fig. 11: Evolution of droplet-, film-contour, and flake orientation (black ruler
corresponds to 50 µm), (𝐷 = 50 µm, 𝜂 = 20 mPa·s, 𝜎 = 0.063 N·m−1 , 𝑈 = 10 m·s−1 ,
𝑊 𝑒 = 95, 𝑅𝑒 = 30, 𝑂ℎ = 0.33).

Some comparison of flake orientations using different viscous droplets (glycerine

and paint) is depicted in Fig. 12. At quasi-static state, we can observe that for droplet
with lower viscosity (20 mPa·s) the flake angle with lower impact velocity is smaller
(32◦ ) than that (50◦ ) with high impact velocity. The paint droplet with constant
viscosity (𝜂 = 40 mPa·s) that undergoes only the spreading process shows the largest
flake angle (58◦ ). The surface tension shows weak effect on the final flake orientation
at the quasi-static state, however, the time to reach the static state is longer for the
paint droplet.
Numerical simulation of flake orientation during droplet impact 333

Fig. 12: Comparison of flake orientations by droplet impact on the wet solid surface
using different viscous droplets, 𝐻 = 60 µm, 𝐷 = 50 µm, We, Re and Oh are
dimensionless numbers and are defined in Table 1.

Fig. 13: Effect of initial position of the flake on the final flake orientation, 𝐻 = 60 µm,
𝐷 = 50 µm and 300 µm, 𝜎 = 0.025 N·m−1 , 𝜌 = 1020 kg·m−3 , 𝜂 = 20 mPa·s,
𝑈 = 0.5 m·s−1 to 10 m·s−1 .

The effect of the initial position of the flake inside droplet on the final flake
orientation was also studied. Simulations with four representative initial flake positions
shown in Fig. 13, were carried out. Initial position 2 and 3 show the lower angle to
horizontal. In general, the final angles are quite scatted with increasing Reynolds
number, indicating a worse flake orientation by a higher droplet impact velocity for a
droplet with diameter around 𝐷 = 50 µm. For a large droplet, the flake distribution in
the droplet may be more complicated. Noticeable gas-liquid interface wave occurs
because of crown formation / splashing by large Reynolds and Weber numbers, which
results in a complicated flake movement and an extreme time consuming for the
corresponding numerical simulation.
334 Q. Ye, M. Dreher and B. Shen

The aforementioned parameter study was performed at a constant droplet viscosity.

The effect of shear thinning viscosity of paint liquid on the droplet impact [12] and
consequently on the flake movement are also investigated. Simulation results are
summarized in Fig. 14 a - d). In the main region of droplet impact, which is also the
cratering region, the viscosity is about 10 mPa·s to 20 mPa·s. The maximum liquid
viscosity is 80 mPa·s where is far away from the impact centre. For convenience, a
characteristic shear rate defined as 𝛾¤ ∗ = 2𝑈/𝐷 is introduced [12] here, where 𝑈 is
the impact velocity and 𝐷 the droplet diameter. The 𝛾¤ ∗ is 2 × 105 (s−1 ) for the case
shown in Fig. 14. The corresponding viscosity is approx. 14 mPa·s. The comparison
of simulation results (Fig. 14 d and e)) using a shear thinning viscosity and a constant
viscosity of 10 mPa·s shows a quite similar flake orientation, which indicates that the
flake orientation in a real droplet could be estimated by the simulation results using
the constant viscosity corresponding to the paint viscosity at the characteristic shear
rate.

Fig. 14: Simulation results using paint droplet with shear thinning viscosity (𝑈 =
5 m·s−1 , 𝐷 = 50 µm): a) VOF contours (red: air, blue: liquid); b) Strain rate (s−1 )
distribution (values that are large than 7×105 are blanked out); c) viscosity distribution
(values that are large than 50 mPa·s are blanked out); d) Flake orientation at static
state; e) Flake orientation at static state using constant droplet viscosity of 10 mPa·s.
Numerical simulation of flake orientation during droplet impact 335

4 Computational performance

The numerical flow simulations were performed using the commercial flow solver
ANSYSFluent v19.5 on the HPE APOLLO (HAWK) of the High Performance
Computing Center (HLRS) in Stuttgart. This supercomputer is equipped with 5632
compute nodes (2021), where each node has 256 GB memory and maximal 128
cores/node can be suggested to use. Figure 15 shows the performance study for
simulations of flake orientation by droplet impact on the dry solid surface with a grid
of 24 million cells for the large case and 12 million cells for the small case. Although
dynamic load-balancing is used during simulation with AnsysFluent, clearly, by the
large case, the speed-up decreases tremendously, when the number of nodes is larger
than 16. A reasonable number of 8 to 10 nodes were therefore applied.

Fig. 15: Scaling with number of nodes.

For the simulation of flake orientation by droplet impact on the wet surface, more
memory per core is required because of the overset method coupled with dynamic
mesh adaption. We reduced the number of the used cores per node extensively. The
corresponding performance is depicted in Fig. 16. Obviously, the calculation with 32
cores/node and 256 cores used in total shows a better performance. It seems that a
higher number of cores per node will not benefit our current application.
A simulation case with 130 million cells was calculated up to 0.1 m·s−1 to 0.5 m·s−1
of physical time with time step size < 10−7 s. Using 8 nodes and 256 cores, it consumed
about 150 to 760 hours. The latter corresponds to the case with a large droplet.
336 Q. Ye, M. Dreher and B. Shen

Fig. 16: Mean wall-clock-time of one time step for the droplet impact on the wet solid
surface with 130 Mio. computational cells.

5 Conclusions

It is the first time that a detailed numerical study of the flake orientation by the viscous
droplet impact on dry/wet solid surfaces has been carried out. A dynamic contact
angle model that is suitable to the spray painting process was proposed and validated
by the experimental observations. We have succeeded to implement a rigid body
motion solver for calculating pigment movement in ANSYSFluent. Some numerical
treatments have been applied, ensuring accurate calculation of derivatives and stable
solution by solving the flake motion equations, which makes the simulation of the
flake movement possible in a practical sense. A first parameter study was performed.
In the case of droplet impact on a dry surface, a relatively large flake angle with
respect to the horizontal surface could be still expected after the receding process
at quasi-static state, although the flake angle changes tremendously with increasing
Reynolds number. By droplet impact on wet surfaces, the simulation results of crater
size in dependence on the Reynolds number deliver useful information for further
analysis of the flake orientation. In the case of droplet with a flake impact on the wet
surface, detailed parameter study for medium droplet (𝐷 = 50 µm) was performed.
Effects of initial flake position, droplet viscosity and impact velocity on the final flake
orientation were investigated. For a given low viscous droplet and at the quasi-static
state, it seems that the flake angle increases with increasing impact velocity because
of the effect of crater receding. It is clearly observed that higher viscosities of droplet
and film reduce the flake movement, resulting in smaller differences between the
initial and the final flake angles. It is found that the height of film has weak effect on
the flake orientation.
Further detailed parameter study should be performed with non-Newtonian liquids,
especially for large droplets, such as 𝐷 = 80 µm to 200 µm, for which time consumed
numerical simulations have to be applied.
Numerical simulation of flake orientation during droplet impact 337

Acknowledgements The present investigations were supported by the German Federal Economical
Affairs and Energy through the Arbeitsgemeinschaft industrieller Forschungsvereinigungen (AIF).
The simulations were performed in the High Performance Calculation Centre Stuttgart (German
federal project: DropImp). These supports are gratefully acknowledged by the authors.

References

1. Chandra, S. and Avedisian, C. T.: On the collision of a droplet with a solid surface. Proc. R.
Soc. London, Ser. A 432, 13 (1991)
2. Thoroddsen, S. T. and Sakakibara, J.: Evolution of the fingering pattern of an impacting drop.
Phys. Fluids 10(6):1359—1374 (1998)
3. Thoroddsen,S. T., Takehara, K. and Etoh, T. G.: Bubble entrapment through topological change.
Phys. Fluids 22(051701):1—4 (2010)
4. Weiss Daniel A., Yarin Alexander L.: Single drop impact onto liquid films: neck distortion,
jetting, tiny bubble entrainment, and crown formation. Journal of Fluid Mechanics, vol.385:
229–254 (1999)
5. Arogeti, M., Sher, E. and Bar-Kohany, T.: A single spherical drop impact on a flat, dry surface
– a unified correlation. Atomization and Sprays, 27(9):759–770 (2017)
6. Kim, E. and Baek, J.: Numerical study of the parameters governing the impact dynamics
of yield-stress fluid droplets on a solid surface. Journal of Non-Newtonian Fluid Mechnics
173-174,62–71(2012)
7. Šikalo, Š., Wilhelm, H.-D., Roisman, I. V., Jakirlić, S. and Tropea, C.: Dynamic contact angle
of spreading droplets: Experiments and simulations. Physics of Fluids 17, 062103 (2005);
https://doi.org/10.1063/1.1928828
8. Linder, N., Criscione, A., Roisman, Ilia V., Marschall, H., Tropea, C.: 3D computation of
an incipient motion of a sessile drop on a rigid surface with contact angle hysteresis. Theor.
Comput. Fluid Dyn. 2015, DOI 10.1007/s00162-015-0362-9
9. Kirchner, E.: Flow-induced orientation of flakes in metallic coatings – II. The orientation
mechanism, Progress in Organic Coatings, 124, 104–109 (2018)
10. Schlüsener, T.: Untersuchungen zum Einfluss der thermo- und hydrodynamischen Vorgänge bei
der Lackapplikation und -trocknung auf die Farbtonausbildung wasserbasierter Metallic-Lacke,
PhD thesis in German, University Darmstadt, (2000)
11. Ye, Q. and Domnick, J.: Analysis of droplet impingement of different atomizers used in spray
coating processes, J. Coat. Technol. Res., 14 (2) 467—476 (2017)
12. Shen, B., Ye, Q., Tiedje, O., Domnick, J.: On the impact of viscous droplets on wet solid
surfaces for spray painting processes, accepted by ICLASS 2021, 15th Triennial International
Conference on Liquid Atomization and Spray Systems, Edinburgh, UK, 29 Aug. – 2 Sept.
(2021)
A low-pass filter for linear forcing in the
open-source code OpenFOAM – Implementation
and numerical performance

Jordan A. Denev, Thorsten Zirwes, Feichi Zhang and Henning Bockhorn

Abstract Direct Numerical Simulations or Large-Eddy Simulations often require

that turbulence is forced and maintained throughout the solution time in the complete
computational domain. In numerical codes written in physical space—like the open
source library OpenFOAM—a modeling technique called linear forcing is often
used for this purpose. It consists of adding a body-force term to the Navier–Stokes
equations which is linearly proportional to the velocity. When compared to codes
written in spectral space, in physical space this technique can only capture integral
length scales half the size of those in spectral codes and is therefore inferior in terms
of numerical efficiency. As shown by Palmore and Desjardins 2018 ([1], Physical
Review Fluids 3, 034605) this drawback can be overcome through low-pass filtering
of the velocity field used for forcing. However, the filter proposed in [1], although
one-dimensional, is implicit and increases considerably the CPU-time. In order to
overcome this, the present work proposes a new three-dimensional, explicit, low-pass
Laplace-filter which is numerically efficient, shows very good scaling features and is
easy to implement and parallelize. With this, the integral length-scale of turbulence
increases more than two times thus resolving successfully a larger scale range of
turbulence. A second improvement proposed in this work concerns the particular
form of the linear forcing term: it applies the forcing individually to each velocity
component. It is found that the new forcing term prevents unphysical, numerically
triggered growth of only one velocity component and hence stabilizes the numerical
process in OpenFOAM.

Jordan A. Denev and Thorsten Zirwes

Steinbuch Centre for Computing, Karlsruhe Institute of Technology, Hermann-von-Helmholtz-Platz
1, 76344 Eggenstein-Leopoldshafen, Germany, e-mail: [email protected]
Thorsten Zirwes, Feichi Zhang, and Henning Bockhorn
Engler-Bunte-Institute/Combustion Technology, Karlsruhe Institute of Technology, Engler-Bunte-
Ring.7, 76131 Karlsruhe, Germany

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 339
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_20
340 J.A. Denev, T. Zirwes, F. Zhang and H. Bockhorn

1 Introduction

Turbulence is the most common fluid flow type in engineering. Its ability to strongly
enhance mixing of species, to intensify heat transfer and chemical reactions allows
decreasing the size of gas turbines, of car engines as well as of other equipment in
the process industry. Through properly controlling the features of turbulence—like
its intensity and length scales—one can control the burning regimes and hence the
burning intensity of premixed flames as it has been summarized in the Borghi regime
diagram and its modifications [2–4]. Another application which strongly depends
on the turbulence properties is connected to the distribution and clustering of spray
particles and their interaction with eddies of the continuous phase as described in [5].
Appreciating the importance of turbulence for engineering applications, more
accurate numerical methods for resolving the time and spatial scales of turbulence on
very fine numerical grids—like Direct Numerical Simulation (DNS) or Large-Eddy
Simulation (LES)—have been developed. These methods require that corresponding
initial and boundary conditions which reflect the development of turbulent eddies in
space and time are imposed. The generation of proper boundary conditions for the
simulation of these methods is not a trivial task and more details can be found in [6–8].
Although with such boundary conditions the desired properties of turbulence are well
defined at the inflow boundaries, the energy cascade and the action of viscous forces
could still change the properties of turbulence on the way downstream until it reaches
the close vicinity of the flame front.
Therefore, a second class of numerical methods for DNS has been developed
which generate and maintain the proper turbulence features throughout the complete
numerical domain. This is achieved through imposing an additional body force on
the gas mixture thus maintaining the desired turbulence properties everywhere in the
domain, including areas of special interest like the very vicinity of a flame front. The
method is known as “forcing of turbulence” [9–12] and the present work introduces
features that lead to improvements of its numerical performance.
Historically, forcing of turbulence has been first introduced for codes operating
in wave-number space [13, 14]. Here, the forcing is applied to low-wave number
modes [10] which resembles the energy input of large-scale turbulent vortices.
However, as recognized by Rosales and Meneveau [10], the forcing of low-wave
numbers is not easy to implement in codes which are written only in terms of physical
space. But extending the forcing to such codes is desirable as then simulation setups
are no more limited to periodic boundary conditions thus enabling the study of, e.g.,
freely propagating premixed turbulent flames.
An important step allowing the forcing to be extended to codes in physical space
has been given by the work of Lundgren [9] who introduced a method which nowadays
is known as “linear forcing”. Although he still used a pseudospectral computation,
his idea was to apply a forcing function which is directly proportional to the velocity.
The usability of the idea of Lundgren [9] for codes in physical space has been
recognized and studied in detail by [10]. In their work [10], they introduce the linear
forcing for codes in physical space by adding a local force that is proportional to the
velocity in each computational point. In their work the authors showed that “linear
A low-pass filter for linear forcing in openFOAM 341

forcing gives the same results as in spectral implementations” and that “the extent of
Kolmogorov −5/3 range is similar to that achieved using the standard band-limited
forcing schemes”.
One general problem of physical forcing observed in [10] is that the integral length
3 /𝜀 with 𝑢
scale (defined as ℓ = 𝑢 rms rms being the rms of velocity and 𝜀 being the
dissipation rate of the kinetic energy of turbulence) is smaller than in spectral space.
As the smallest scales resolved in numerical simulations are determined from the
resolution of the numerical grid, this means that the scale range which can be resolved
in physical space is smaller than that resolved by codes in spectral space on the same
numerical grid.
Carroll and Blanquart [15] found that the implementation of [10] exhibits an
oscillatory behavior at the beginning of the computations. As the oscillations would
require relative long computations until steady-state is reached, [15] suggest a slight
modification of the forcing term which they show to reduce this oscillatory nature
without altering the physics of turbulence. The code of Carroll and Blanquart uses
a staggered numerical grid and we found that for the OpenFOAM implementation
which uses a collocated grid, one of the components could raise non-proportionally
high (on the expense of the other two), if the velocity components have a mean value
which is different from zero and the forcing is following the implementation suggested
by [15]. Therefore, in the present work we extend the idea of [15] and use forcing
that is applied to each velocity component individually. This way the collocated
grid arrangement is found to work stable and the forcing can be controlled for each
component separately thus allowing, at least in theory, to obtain even non-isotropic
turbulence. However, application of this technique to obtain non-isotropic turbulence
would require a further investigation which goes beyond the scope of the present
work.
Palmore and Desjardins [1] investigated the problem with the reduced integral
scale of the turbulent eddies in physical forcing. They suggested a remedy to this
problem by applying a low-pass filter to the velocity field before calculating the
forcing term. It has been shown that the range of resolved scales can be doubled by
applying a low-pass filter. In [1], they apply successively a one-dimensional (1D) filter
in each spatial direction, 𝑥, 𝑦 and 𝑧, which requires solving three times a complex,
tridiagonal matrix. The application of this filter leads to a considerable increase in
the CPU-time of the computations: between 48 and 652 percent, depending on the
order of the filter which in turn determines its sharpness.
The present work extends the ideas of Palmore and Desjardin [1] by introducing a
new, explicit, three-dimensional (3D) Laplacian filter in physical space. Similar to
[1], this new filter also increases by a factor of two the integral length-scale of the
turbulence, but, as shown in chapter 4, it is computationally considerably cheaper
than the filter proposed in [1].
The paper starts with a mathematical description of the physical forcing (chapter 2)
where the new low-pass Laplace-filter is introduced to the reader. The code for this
filter—in the framework of the open source software OpenFOAM—is presented in
the text. The idea to use a separate (individual) filtering for each velocity component
is also introduced and described mathematically in chapter 2. In chapter 3, first
342 J.A. Denev, T. Zirwes, F. Zhang and H. Bockhorn

a DNS-simulation of [10] is repeated and used for validation of the case without
filtering. Then, on the same setup the effect of filtering on the integral length-scale is
simulated and assessed. The numerical performance of the introduced Laplace-filter
and of the idea for individual filtering of each velocity component is thoroughly
examined in chapter 4. A summary of the work completed and an outlook to future
work is given in chapter 5.

2 Mathematical description

In the present study, the incompressible Navier–Stokes equations are solved in the
framework of Direct Numerical Simulations (DNS) of the turbulent flow:
𝜕𝑢 𝑖
= 0, (1)
𝜕𝑥𝑖
𝜕𝑢 𝑖 𝜕𝑢 𝑖 𝑢 𝑗 1 𝜕 𝑝 𝜕 (𝜈2𝑆𝑖 𝑗 )
+ =− + + 𝑓𝑖 , (2)
𝜕𝑡 𝜕𝑥 𝑗 𝜌0 𝜕𝑥𝑖 𝜕𝑥 𝑗

where 𝑢 is the fluid velocity, 𝜈 the kinematic viscosity, 𝑝 the pressure and 𝑓 the
forcing term. The density 𝜌0 is constant in case of incompressible flows. The strain
rate tensor 𝑆 is given by: !
1 𝜕𝑢 𝑖 𝜕𝑢 𝑗
𝑆𝑖 𝑗 = + . (3)
2 𝜕𝑥 𝑗 𝜕𝑥𝑖
Following the idea of Lundgren [9], the forcing term 𝑓 is presented by:

𝑓𝑖 = 𝐶𝑢 𝑖 , (4)

where 𝐶 is a constant. For a stationary isotropic turbulence, Lundgren [9] considered

the balance of the turbulent energy equation and could show that the value of the
constant 𝐶 corresponds to:
𝐶 = 𝜀/(3𝑈 2 ) (5)
where 𝑈 2 is the mean square of one component of the velocity: 𝑈 2 = ⟨𝑢 𝑖 𝑢 𝑖 ⟩/3 = 𝑢 2rms
and 𝜀 is the mean (in spatial sense) dissipation rate of turbulence. Here, 𝑢 𝑖 are the
velocity fluctuations and summation over repeating indices is assumed. The angular
brackets denote spatial averaging.
Rosales and Meneveau [10] realized that the forcing-term from equation 5 is
suitable for forcing in physical space and showed the applicability of this idea in their
work. They noticed that the constant 𝐶 from equation 5 has a meaning of a turnover
time scale and kept the value of this constant unchanged during the computations.
Carroll and Blanquart [15] improved the convergence towards steady-state by
rearranging the forcing term into the form:
A low-pass filter for linear forcing in openFOAM 343
 
𝑘0
𝑓𝑖 = 𝐶 𝑢𝑖 , (6)
𝑘

where 𝑘 0 is the desired kinetic energy and 𝑘 is the instantaneously calculated kinetic
energy. Equation 6 reduced the amplitude of the oscillations compared to equation 5
and the computations with the former reached quicker steady-state conditions. After
reaching these conditions, the two equations deliver identical (in statistical sense)
results.
Palmore and Desjardins [1] made a physical interpretation of equation 6. Con-
sidering the equation for the energy production, they could show that the pre-factor
𝑘 0 /𝑘 dynamically adjusts the forcing in such a way that a constant dissipation rate is
achieved. In the same sense, by modifying the constant 𝐶 to depend on the current
𝑘 or 𝜀 values, Basenne et al. [16] introduced physical forcing that maintains either
constant turbulent kinetic energy or constant turbulent dissipation (also constant
enstrophy for incompressible flows), or also a combination of the two showing further
shortening of the initial transient period.
Palmore and Desjardins [1] applied a low-pass filter for the velocity field before
using it as a source-term in the Navier–Stokes equations. Through this, they could
double the integral length-scale ℓ of the resulting turbulent velocity field. For the
forcing type represented by equation 6, the use of a filtered velocity field is written as:
 
𝑘0
𝑓𝑖 = 𝐶 𝑢˜ 𝑖 . (7)
𝑘

Here, .̃ denotes the low-pass filtering operation. As already mentioned, in the work
of [1] the filtering was applied to each spatial direction individually.
In the present work we introduce two new features to linear forcing in physical
space which improve the numerical performance of this method.
The first new feature proposed in this work consists of introducing of a real
three-dimensional filter for the filtering operation in equation 7. This filter is an
explicit Laplace-filter. Since this filter is applied to the velocity field and smoothes
the velocities in analogy to physical diffusion, its form is determined from the
implementation of the incompressible flow solvers of OpenFOAM:
 !
1
ũ = u + ∇ · (𝐵∇u) + ∇ · 𝐵 (∇u) 𝑇 − ∇ · u , (8)
3

with u being the vector of the instantaneous velocity components. We use this form
of the filter in the simulations in chapter 3. Before applying the filter, the velocity
components have been centered with respect to their space-average value.
For an incompressible fluid, in analogy to the viscous term from equation 2, the
above equation can be simplified to:
  
ũ = u + ∇ · 𝐵 ∇u + (∇u) 𝑇 . (9)
344 J.A. Denev, T. Zirwes, F. Zhang and H. Bockhorn

The filter constant 𝐵 is not bound to the physical viscosity. Its value depends on the
computational grid:
2
(𝑉 𝑜𝑙) 3
𝐵= , (10)
𝑊coeff
where 𝑉 𝑜𝑙 is the volume of the numerical cell and 𝑊coeff is a user-specified constant
with the meaning of a width (diffusion) coefficient. Through variation of the value
of 𝑊coeff , respectively of 𝐵, the desired length-scale of turbulence can be achieved.
Details about the particular value of the constant 𝐵 and the way it has been obtained
in the present work are given in section 3.
A Laplace-filter is already available in OpenFOAM for LES models, but the
diffusive term there is formulated only for scalar quantities, not vector quantities
like the velocity here. So, the filter has been modified accordingly and the coding
is provided to the reader in the following; one major difference to the available
OpenFOAM-filter is that now we made the filter entirely explicit thus allowing
its value to be immediately obtained and directly used in the forcing term. In
OpenFOAM’s syntax, the implementation of the Laplace-filter for the velocity field
is written as:

1 Foam : : tmp <Foam : : v o l V e c t o r F i e l d > l a p l a c e F i l t e r

2 (
3 c o n s t Foam : : v o l V e c t o r F i e l d& U,
4 s c a l a r widthCoeff ,
5 c o n s t fvMesh& mesh
6 )
7 {
8 s c a l a r a v e r a g e C e l l V o l u m e = a v e r a g e ( mesh . V ( ) ) . v a l u e ( ) ;
9

10 d i m e n s i o n e d S c a l a r c o e f f ( dimArea ,
11 Foam : : pow ( a v e r a g e C e l l V o l u m e , 2 . 0 / 3 . 0 ) / w i d t h C o e f f ) ;
12

13 r e t u r n U + f v c : : l a p l a c i a n ( c o e f f , U) +
14 f v c : : d i v ( c o e f f ∗ dev ( T ( f v c : : g r a d (U ) ) ) ) ;
15 }

In the next sections it will be shown that the explicit and localized nature of the
Laplace-filter leads to considerable savings of the CPU-time usage when compared
to the filters originally used by [1]. Similarly to the original work of [1], the present
Laplace-filter also leads to doubling of the integral velocity length scale of the
resulting turbulent field and hence to an increase of resolved length scales similar to
the one observed by spectral codes.
The second new feature introduced in the present work consists of modifying
equations 6 or 7 so that each velocity component can be forced individually. For
velocity in direction of axis 𝑖 this results in:

2 2
© 𝑢 𝑖,𝑟 𝑚𝑠,0 ª © 𝑢 𝑖,𝑟 𝑚𝑠,0 ª
𝑓𝑖 = 𝐶 ­ 2 ® 𝑢𝑖 (𝑎) or 𝑓𝑖 = 𝐶 ­ 2 ® 𝑢˜𝑖 (𝑏). (11)
𝑢 𝑖,𝑟 𝑚𝑠 𝑢 𝑖,𝑟 𝑚𝑠
« ¬ « ¬
A low-pass filter for linear forcing in openFOAM 345

Here, 𝑢 2𝑖,𝑟 𝑚𝑠,0 is the targeted value of the squared velocity fluctuations of component
𝑢 𝑖 and 𝑢 2𝑖,𝑟 𝑚𝑠 is the instantaneously calculated value of these fluctuations; both values
are averaged in space. In case of application of a filter, the filtered velocity field 𝑢˜𝑖
according to equation 11b is applied instead.
We found that the above form of the forcing equation 11a, unlike equation 6,
prevents a divergence from a non-proportional growth of one velocity component on
the expense of the other two in case when the average value of this component ⟨𝑢 𝑖 ⟩ is
not zero. Such a non-zero average value is typical for setups with a prescribed flow
direction as in the case of premixed flames with inflow velocity equal to the flame
velocity. Therefore, the proposed modification enables also the handling of cases
with non-periodic, i.e. inflow-outflow boundary conditions. Although not explicitly
demonstrated in the present work, but by simply setting 𝑢 21,𝑟 𝑚𝑠,0 ≠ 𝑢 22,𝑟 𝑚𝑠,0 ≠ 𝑢 23,𝑟 𝑚𝑠,0 ,
equation 11 enables a forcing which, at least in theory, would produce also a non-
isotropic turbulence.

3 DNS results with the proposed filter for linear forcing

In order to assess the modeling and the numerical performances of the suggested
filter (equation 8), a Direct Numerical Simulation following case 3 from the work
of Rosales and Meneveau [10] is carried out. This case has been chosen because it
has the largest forcing constant and therefore produces the highest kinetic energy of
turbulence. In the following, some details of the DNS are presented. The domain
dimensions are 2𝜋 m in all directions and periodic boundary conditions apply on
all boundaries. The kinematic viscosity is equal to 𝜈 = 0.004491,m2 /s. The value
of the forcing constant for both cases (with and without filtering) is set to 𝐶 = 0.2
and the forcing scheme is according to equation 11 with a value of 𝑢 2𝑖,𝑟 𝑚𝑠,0 equal to
0.44. While the DNS of [10] was simulated on a numerical grid of 1283 cells, here
the calculations have been made also on a two times finer grid with 2563 cells. One
reason for this is the need to assess the numerical performance of the introduced
filter on finer grids which are typical for DNS. A second reason is to compensate the
numerical error of the second-order scheme applied in the present work as compared
to the sixth-order numerical scheme used originally by [10]. The results presented in
this chapter are obtained on the finest grid with 2563 numerical cells.
The initial turbulent velocity field was generated using a recently implemented
OpenFOAM utility called createBoxTurb (see description in [17] for more details)
according to the spectrum of [18]. This is different from the initial spectra used in
[10], however, according to [10] the spectra of the initial velocity field has no effect
on the final value of the integral length scale. To make a verification of the current
results before the application of the filter, the energy spectrum from the current work
is compared with that from [10] (Figure 2, case 3c). The comparison is performed
after reaching a final stationary state and is presented in Figure 1. The figure shows
that the spectrum from the present work, when reaching a stationary state, practically
overlaps with that of [10].
346 J.A. Denev, T. Zirwes, F. Zhang and H. Bockhorn

The value of the filter constant 𝐵 for all grids was set equal to 0.0277 which
corresponds to 𝑊coeff = 0.087 for the 1283 grid and to 𝑊coeff = 0.02175 for the 2563
grid (see equation 10). The above value of 𝐵 was first iteratively adjusted (by trial
and error) on a grid with 643 grid points (𝑊coeff = 0.348). As shown further in this
section, the target for the adjustment was to reach—at the statistically stationary state
(times larger than 100𝑠)—a doubling of the integral length-scale of the turbulent field
ℓ in comparison with the original study of [10]. Each of the above test runs had a
duration of 100 minutes by using only a single core.
Figure 2 allows to get a first impression of the influence of the filter on the velocity
field presented by the magnitude of the velocity vector. It can be clearly seen that the
structures appearing on the left-hand side of the Figure are on average visually larger
than those without the filter; also the amount of structures at the left-hand side is less
than in the case without the filter.

Fig. 1: Energy spectrum of the velocity field without filtering. Comparison of present
stationary results at 𝑡 = 300 s with stationary results of [10].

In the following, the effect of the filter on the statistical quantities of the obtained
turbulence is examined. First of all, it has been observed that the filter leads to a
decrease of the value of the energy dissipation 𝜀. This can be seen in Figure 4 where
the spatially averaged value of 𝜀 is plotted against the number of time-steps; for
convenience, the first 5000 time-steps with wide-spread values for 𝜀 are omitted. It
can be seen that the resulting value for the energy dissipation rate is approximately
two times smaller with the filter (0.115) than without the filter (0.254).
A low-pass filter for linear forcing in openFOAM 347

Fig. 2: Turbulent structures of the velocity magnitude |u| at 𝑥 = 2.1𝑚 and 𝑡 = 300 s
appearing in the case without the filter (left) and with it (right).

As it is shown in Figure 3, the forcing scheme according to equations 11a and 11b
leads statistically to the same amount of kinetic energy of turbulence. Bearing in
mind the two times higher values for the energy dissipation when the filter is applied,
the relation ℓ = 𝑢 3rms /𝜀 reveals a major outcome from the Laplace-filtering, namely
that the integral length scale ℓ also increases by a factor of two. Therefore, through
the application of the filter, the main drawback of forcing in physical space, i.e. the
smaller size of the integral length scale, - can be successfully overcome and thus the
range of resolved scales is considerably increased. Following the ideas of [1], the
smallest eddies are always fixed by the grid size and therefore the two times larger
integral length scale means that the range of scales captured by the same numerical
grid is also doubled. The corresponding increase of the integral length scale is shown
in Figure 5.

Fig. 3: Time evolution of the turbulence kinetic energy 𝑇 𝐾 𝐸 (m2 /s2 ) for the case
with the filter and when no filtering is applied.

Showing the advantage which filtering brings for the range of scales resolved,
the question about the numerical performance of the computations with the filter
becomes important. This performance is evaluated in section 4.
348 J.A. Denev, T. Zirwes, F. Zhang and H. Bockhorn

Fig. 4: Time evolution of the energy dissipation rate 𝜀 (m2 /s3 ) for the case with the
filter and when no filtering is applied.

Fig. 5: Integral length scale ℓ (m) compared for the case with the filter and when no
filtering is applied.

4 Parallel performance

The method for parallel computing used by OpenFOAM is known as domain

decomposition, where the entire grid with a large number of computational cells are
partitioned into a number of sub-domains and assigned to one CPU core. Ideally
each sub-domain or CPU shares the same number of grid elements. directly. The
implementation of the filter described in this work requires the evaluation of spatial
derivatives which utilizes OpenFOAM’s implementation based on the finite volume
method. For this, pair-to-pair communication in MPI is required between the tasks
that share internal domain boundaries. In addition, all-to-all communication has to
be performed to obtain global quantities like the total turbulent kinetic energy, global
dissipation rate and velocity fluctuation.
The scale-up performance depends mainly upon the number of allocated cells to
each single CPU core 𝑛cell . A good parallel scalability is expected for sufficiently
large 𝑛cell . In case of small 𝑛cell , the time occupied by communications or data
exchanges between the processors cores may constitute a large portion of the total
computing time and reduce parallel efficiency. Table 1 lists the measured computing
performance by running the proposed solver in parallel for a test case with 𝑁cell = 2563
A low-pass filter for linear forcing in openFOAM 349

grid cells on the HoreKa cluster at SCC/KIT. The number of processor cores 𝑁core
in the x-axis is selected to be multiples of 76 (𝑛core = 76, 152, 304, 608, 1216 and
2432), because each computing node from Horeka contains 76 cores. In this case, the
consumed wall clock time for running the solver for 2000 iteration time steps (Δ𝑡)
has been recorded and the measurement has been repeated for 3 times in each case
by using different 𝑛core . The averaged consumed clock times per time step 𝑡 Δ𝑡 from
these 3 simulations, shown in the second column of Tab. 1, are used to evaluate the
speed-up factor and the efficiency factor with respect to parallel scalability
𝑡Δ𝑡 ,ref 𝑛core,ref 𝑡Δ𝑡 ,ref
𝑆𝑛 = , 𝐸𝑛 = · . (12)
𝑡Δ𝑡 ,𝑛 𝑛core 𝑡Δ𝑡 ,𝑛

𝑆 𝑛 and 𝐸 𝑛 are normalized to the measured computing time with 𝑛core,ref = 76

(corresponding to use of one single node from Horeka). 𝑃 given in the 5th column of
Tab.1 represents an estimation of consumed clock time per time step and per cell on
one single CPU core:
𝑡 Δ𝑡 · 𝑛core
𝑃= . (13)
𝑁cell

Table 1: Scale-up performance of running the implemented forced-turbulence code

with OpenFOAM-v2006 on the HoreKa cluster at SCC/KIT on a computational mesh
with 2563 cells.

𝑛core [-] 𝑡Δ𝑡,𝑛 [s] 𝑆𝑛 [-] 𝐸𝑛 [-] 𝑃 [𝜇s/Δt/cell] 𝑛cell [-]

76 4.420 1 1 20.0 220.8×103
152 1.932 2.29 1.14 17.5 110.4×103
304 0.904 4.88 1.22 16.4 55.2×103
608 0.457 9.67 1.21 16.6 27.6×103
1216 0.292 15.12 0.95 21.1 13.8×103
2432 0.276 16.0 0.50 40.0 6.9×103

The measured parallel scalability is further illustrated in Fig. 6 with the speed-up
factor on the left and the efficiency factor on the right. A superlinear scaling can
be detected by increasing 𝑛core from 76 to 608 for the current setup, which may be
explained by a beneficial usage of the cache memory instead of assessing the RAM
by using more processor cores. In addition, the solution procedures of the governing
equations occupy a major breakdown of the total computing time compared with the
consumed time for data-exchanges, communication and synchronization between
the processors. Further increase of 𝑛core from 1216 to 2432 results however in a
considerable decrease of 𝐸 below unity (see the 4th column in Tab. 1 and Fig. 6 on
the right), which may be attributed to a strongly increased share of communication
compared with the computational effort for solving the mathematical equations.
Therefore, the optimal number of cells per CPU core, shown in the last column of
Tab. 1, should be selected as large as more than 10,000 for good parallel scalability.
350 J.A. Denev, T. Zirwes, F. Zhang and H. Bockhorn

Fig. 6: Strong scaling performance of running DNS of forced turbulence on a

computational grid with 2563 cells on the HoreKa cluster at KIT/SCC Karlsruhe.

5 Conclusions

Numerical methods like DNS and LES are established tools for the detailed inves-
tigation of turbulent flows. These methods are known to be very computationally
intensive and therefore the increase of their numerical efficiency is a subject of active
research. In order to sustain a stable isotropic turbulence that does not decay in time,
a class of methods exists that is called linear forcing. The present work introduces two
numerical improvements of linear forcing applicable for codes operating in physical
space.
The first one introduces a new explicit Laplace-filter for linear forcing in physical
space by equation 8. The new filter is a low-pass filter that is three-dimensional
in space. It is implemented in OpenFOAM and is listed in chapter 2. On a given
computational mesh and with a fixed domain size, the application of the filter is shown
to increase two times the integral length scale ℓ resolved by the numerical simulation.
This is done with only 9 % increase of the CPU-time. Compared to a recent work
of [1], where the filtering operation requires at least 48% longer CPU-time, there is a
considerable improvement in the numerical efficiency with the present filter.
The second improvement is introduced through equation 11. The novelty is that
the forcing is applied to each velocity component separately thus allowing its level
to be adjusted individually in the course of the computations. This prevents the
appearing of non-physical solutions that lead to much stronger forcing of one velocity
component on the expense of the other two when forcing is applied according to
equation 6 suggested by [15]. The overhead for this improvement is below 1% of the
overall CPU-time of the simulations.
In chapter 4 it has been shown that the strong scaling with the new filtering
method performs as expected for the incompressible solvers in OpenFOAM. A
recommendation is given to use more than 10,000 cells per core to achieve good
parallel scalability.
A low-pass filter for linear forcing in openFOAM 351

Acknowledgements The authors gratefully acknowledge the financial support by the Helmholtz
Association of German Research Centers (HGF), within the research field MTET (Materials and
Technologies for the Energy Transition), subtopic “Anthropogenic Carbon Cycle” (38.05.01). This
work utilized computing resources provided by the High Performance Computing Center Stuttgart
(HLRS) at the University of Stuttgart and on the ForHLR II and HoreKa Supercomputers at the
Steinbuch Centre for Computing (SCC) at the Karlsruhe Institute of Technology.

References

1. J. A. Palmore and O. Desjardins, “Linear forcing in numerical simulations of isotropic turbulence:

Physical space implementations and convergence properties,” Physical Review Fluids, vol. 3,
pp. 034 605(1)–034 605(18), 2018.
2. R. Borghi, “On the Structure and Morphology of Turbulent Premixed Flames,” in Recent
Advances in Aeronautical Science. Plenum Press, 1985.
3. J. Warnatz, U. Maas and R.W. Dibble, Combustion., 4th ed. 378 p., Springer, Berlin, 2006.
4. T. Poinsot and D. Veynante, Theoretical and numerical combustion., 2nd ed. 522 p., Edwards,
2005.
5. N. S. H. Lian and Y. Hardalupas, “Evaluation of the topological characteristics of the turbulent
flow in a ‘box of turbulence’ through 2D time-resolved particle image velocimetry,” Experiments
in Fluids, vol. 58:118, p. 17, 2017.
6. F. Galeazzo, F. Zhang, T. Zirwes, P. Habisreuther, H. Bockhorn, N. Zarzalis, and D. Trimis,
“Implementation of an Efficient Synthetic Inflow Turbulence-Generator in the Open-Source
Code OpenFOAM for 3D LES/DNS Applications,” in High Performance Computing in Science
and Engineering ’20. Springer, 2020, (accepted).
7. M. Klein, A. Sadiki, and J. Janicka, “A digital filter based generation of inflow data for spatially
developing direct numerical or large eddy simulations,” Journal of computational Physics, vol.
186, no. 2, pp. 652–665, 2003.
8. N. Kornev and E. Hassel, “Synthesis of homogeneous anisotropic divergence-free turbulent
fields with prescribed second-order statistics by vortex dipoles,” Physics of Fluids, vol. 19(6),
pp. 068 101(1)–068 101(4), 2007.
9. T. Lundgren, “Linearly forced isotropic turbulence,” in Annual Research Briefs. Center for
Turbulence Research, Stanford University, Stanford, CA, 2003, pp. 461–473.
10. C. Rosales and C. Meneveau, “Linear forcing in numerical simulations of isotropic turbulence:
Physical space implementations and convergence properties,” Physics of Fluids, vol. 17, p.
095106, 2005.
11. J. A. Denev and H. Bockhorn, “Dns of lean premixed flames,” in High Performance Computing
in Science and Engineering ‘13, W. E. Nagel, D. H. Kröner, and M. M. Resch, Eds. Cham:
Springer International Publishing, 2013, pp. 245–257.
12. J. A. Denev, I. Naydenova, and H. Bockhorn, “Lean premixed flames: A direct numerical
simulation study of the effect of lewis number at large scale turbulence,” in High Performance
Computing in Science and Engineering ‘14, W. E. Nagel, D. H. Kröner, and M. M. Resch, Eds.
Cham: Springer International Publishing, 2015, pp. 237–250.
13. V. Eswaran and S. B. Pope, “An examination of forcing in direct numerical simulations of
turbulence,” Comput. Fluids, vol. 16, p. 257, 1988.
14. S. M. Sullivan and R. Kerr, “Deterministic forcing of homogeneous, isotropic turbulence,” Phys.
Fluids, vol. 6, p. 1612, 1994.
15. L. Carroll and G. Blanquart, “A proposed modification to Lundgren’s physical space velocity
forcing method for isotropic turbulence,” Physics of Fluids, vol. 25, pp. 105 114(1)–105 114(9),
2013.
16. M. Bassenne, J. Urzay, G. I. Park, and P. Moin, “Constant-energetics physical-space forcing
methods for improved convergence to homogeneous-isotropic turbulence with application to
particle-laden flows,” Physics of Fluids, vol. 28, pp. 035 114(1)–035 114(16), 2016.
352 J.A. Denev, T. Zirwes, F. Zhang and H. Bockhorn

17. T. Saad, D. Cline, R. Stoll, and J.C. Sutherland, “Scalable tools for generating synthetic isotropic
turbulence with arbitrary spectra,” AIAA Journal, vol. 55(1), pp. 327–331, 2017.
18. G. Comte-Bellot and S. Corrsin, “Simple Eulerian Time Correlation of Full- and Narrow-Band
Velocity Signals in Grid-Generated, Isotropic Turbulence,” Journal of Fluid Mechanics, vol.
48(2), pp. 273–337, 1971.
Numerical simulation of vortex induced pressure
fluctuations in the runner of a Francis turbine at
deep part load conditions

Jonas Wack, Marco Zorn and Stefan Riedelbauch

Abstract For hydropower applications far off-design operating points like deep part
load are more and more investigated, as these turbines can play a key role for the
compensation of fluctuations in the electrical grid. In this study the single-phase
simulation results of a Francis turbine at model scale are investigated for three mesh
resolutions with the commercial CFD software ANSYS CFX. For the investigated
deep part load operating point the typical inter-blade vortices can be observed. Further
vortex structures are on the one hand traveling upstream close to the suction side
and are on the other hand a result of a flow detachment at the runner trailing edge.
The evaluation of the mesh resolution shows that a mesh refinement especially in
the region of the inter-blade vortices results in a better prediction of the pressure
minimum of these vortices.
The strong scaling test indicates an acceptable scaling up to 1536 cores for the mesh
with 56 million cells. For the mesh with 82 million cells the scaling is acceptable
even up to 2048 cores. A comparison of the MPI methods Open MPI and HMPT
MPI showed that the latter is 16.5% slower.

1 Introduction

Due to the higher share of volatile renewable energies like solar and wind, other
technologies are necessary that can compensate fluctuations in the electrical grid. For
this purpose hydropower can play a key role. However, this requires more start-stop
sequences and operation outside of the designed operating range [1]. These modified
operating conditions result in higher structural loads and thus can have a negative
impact on the lifetime of the turbine. Consequently, for future turbine designs reliable
load evaluations at off-design conditions are required in the design process.

J. Wack, M. Zorn and S. Riedelbauch

Institute of Fluid Mechanics and Hydraulic Machinery, Pfaffenwaldring 10, D-70569 Stuttgart,
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 353
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_21
354 J. Wack, M. Zorn and S. Riedelbauch

To integrate off-design operating points into the design process is challenging.

For example at deep part load conditions stochastic pressure fluctuations occur [2]
that cannot be captured by a standard RANS approach [3]. For that reason, either
algebraic stress models [4] or hybrid RANS-LES models [3, 5] are typically applied
for these operating conditions. However, especially the latter are computationally
expensive as they need fine meshes and small time steps. In the design process a
low computational effort is desired, as a lot of different designs are simulated and
compared. Contrary to this, it is the goal of this project to provide highly resolved
CFD results. Even though these simulations are not feasible in the design process,
they can serve as a reference to derive more simplified simulation setups (e.g. smaller
mesh and reduction of the simulation domain) that give the best compromise between
computational effort and accuracy.

2 Numerical setup

CFD simulations of a Francis turbine model (24 guide vanes and 13 runner blades) of
mid-range specific speed (𝑛𝑞 = 40−60) were performed with the commercial software
ANSYS CFX. The investigations were carried out for a deep part load operating
point with the following characteristics: discharge 𝑄/𝑄 𝐵𝐸 𝑃 ≈ 0.33, discharge factor
𝑄 𝑒𝑑 = 0.07, speed factor 𝑛𝑒𝑑 = 0.37 and 8° guide vane opening. Within this study,
single-phase simulations are carried out and consequently cavitation effects are
neglected. Taking cavitation into account is an interesting research objective and will
be published in the future.
The simulation domain consists of the following components: spiral case (SC),
stay and guide vanes (SVGV), runner (RU), draft tube (DT) and downstream tank.
In this study the focus is on the impact of mesh refinement on the resolution of
vortex structures in the runner channels, as different studies already highlighted the
necessity of fine grids to properly resolve the inter-blade vortices at deep part load
conditions [6, 7]. For these analyses, three different meshes are investigated that are
listed in table 1. All meshes have in common that the majority of cells are used for
resolving the vortex structures in the runner. The coarsest mesh has around 25 million
cells (25M) and uses wall functions in the boundary layer as the averaged 𝑦 + -value is
around 40. This mesh is already finer compared to standard industrial applications in
the field of hydraulic machinery. Mesh 56M is refined in the main flow, with the focus
on a better resolution of the vortex structures in the runner, while the boundary layer
resolution and consequently the 𝑦 + -value remains constant. This allows to investigate
the impact of a better mesh resolution of the vortices. Contrary to that, mesh 82M
has the same mesh resolution in the main flow as mesh 56M but the mesh in the
boundary layer is refined to have an averaged 𝑦 + -value that is slightly below one.
With this methodology it is possible to distinguish between mesh effects that result
from a refinement of the main flow and those that originate from a refinement of the
boundary layer.
Numerical simulation of vortex induced pressure fluctuations in a Francis turbine 355

Table 1: Mesh size of the subdomains for the used grids in million elements.

Mesh SC SVGV RU DT Tank

25M 0.5 7.9 10.8 5.2 0.6

56M 1.6 11.2 31.1 10.1 1.7
82M 1.9 16.9 48.0 13.8 1.7

A comparison of simulations with the SBES [8] and the SAS [9] turbulence
model (both hybrid RANS-LES models) did not show significant differences [10] and
consequently only the more recently developed SBES model has been selected for
more thorough investigations. This model is blending between the SST model in the
RANS region and the WALE model in the LES region. The idea of hybrid RANS-LES
models is to resolve large eddies in the main flow, while the computational effort can
be reduced by using a RANS approach in the wall boundary layers [11]. Due to the
selected turbulence model it is necessary to perform the simulations with a small time
step that fulfills LES criteria. For mesh 25M the time step is approximately 0.4° of
runner revolution while for the meshes 56M and 82M it is around 0.2°. With these
settings the RMS averaged Courant number is below one for all simulations. Due to
the small time step sizes the unsteady simulations converge quite fast within one time
step and consequently only four coefficient loops are performed.
For spatial discretization a bounded central differencing scheme is selected and
for temporal discretization a second order backward Euler scheme is applied. At the
spiral case inlet a constant mass flow is set and the static pressure is set at the outlet
of the tank. It can be expected that runner seal leakage has a higher impact on the
velocity distribution for low discharges as the leakage flow is almost independent of
the operating point [12]. For that reason seal leakage is considered in the numerical
model by defining a sink at the runner inlet (outlet boundary condition) and a source
at the runner outlet (inlet boundary condition). The seal leakage flow is assumed to be
0.33 % of the discharge for the investigated deep part load operating point. Between
the rotating runner and stationary components a transient rotor stator interface is
applied. To ensure that the interface between runner and draft tube is far away from the
inter-blade vortices, it is located at the end of the draft tube cone. Consequently, walls
of the draft tube that are within the runner domain are prescribed as counter-rotating
walls.
Within this study all averaged results comprise 50 runner revolutions. Before the
averaging has started, additional 20 runner revolutions have been simulated within
the initialization process.
356 J. Wack, M. Zorn and S. Riedelbauch

3 Experimental setup

A model test has been performed for the Francis turbine with multiple sensors in
the stationary and rotating frame. Theses sensors include both, pressure sensors as
well as strain gauge sensors on the runner blades. To investigate cavitation effects
measurements have been carried out for three different pressure levels: ℎ 𝑠 = −10 m,
5.6 m and 6.9 m. Even for the highest pressure level (ℎ 𝑠 = −10 m) some cavitation
regions could be observed, which is a result of the strong off-design operating point.
More information on the experimental setup can be found in [10, 13].

4 Results

First, a comparison of the integral quantities head and torque is carried out. The
deviations compared to the experiment with ℎ 𝑠 = −10 m are listed in table 2. For
this comparison the experimental results for the highest pressure level are used as it
has the smallest cavitation volume and for that reason is closest to the single-phase
treatment of the simulations.

Table 2: Head and torque deviation compared to experiment with ℎ 𝑠 = −10 m.

Mesh 25M 56M 82M

Head -4.5% -4.9% -4.4%

Torque +7.1% +3.2% +2.6%

For all meshes the head is underestimated by around 4.5%. A more significant
difference between the meshes can be found for the torque. All simulation results
overestimate torque, but the coarsest mesh has the highest deviation. Especially
the refinement in the main flow from mesh 25M to 56M results in a significant
improvement of almost 4%. Refining the mesh of the boundary layer (mesh 82M)
leads to a further improvement of 0.6% of torque deviation.
The remaining deviations in head and torque can result from several effects. First,
the neglection of cavitation causes some uncertainty. Furthermore, in the measurement
the angle of the guide vane opening has some uncertainty. While a small deviation
in guide vane opening has only a relatively small effect for operating points with
high power, it is much more significant for operating points with small guide vane
opening like deep part load conditions. Additionally, the complex flow structures in
the runner channels might even need a further mesh refinement.
Numerical simulation of vortex induced pressure fluctuations in a Francis turbine 357

4.1 Vortex movement

For the investigated deep part load operating point the highest pressure fluctuations are
caused by vortices. Three main vortex structures can be observed that are visualized
in Fig. 1 for mesh 82M with an isosurface of the velocity invariant (𝑄 = 5 · 105 s−2 ).
The inter-blade vortex ranges from the hub, at a location close to the leading edge, to
the runner outlet close to the shroud, where it decays into small vortices. Over a wide
range of the runner channel, the inter-blade vortex is located closer to the suction
side (not visible in Fig. 1). Only close to the runner outlet it is located in the middle
between pressure and suction side.

Fig. 1: Vortex structures in one runner channel.

Furthermore, the investigated operating point is characterized by a huge backflow

region in the center of the draft tube cone. Due to this reversed flow, it comes to a
detached flow at the trailing edge on the pressure side close to the hub. Vortices are
generated from the flow detachment that are moving upstream close to the suction
side in areas close to the hub.
To evaluate the effect of mesh refinement, it is investigated if the pressure minimum
in the vortex core of the inter-blade vortices is well resolved. For that reason an
isosurface of the pressure (threshold: vapor pressure for ℎ 𝑠 = 5.6 m) is visualized
in Fig. 2. While for the coarsest mesh the instantaneous pressure (top) is not falling
below vapor pressure in the region of the inter-blade vortex, the results for mesh 56M
and 82M show a low pressure region at the inter-blade vortex. This can be explained
by the fact that these meshes are refined in the main flow compared to mesh 25M and
consequently can better resolve the pressure minimum in the vortex core.
Similar results can be found for the time-averaged pressure (Fig. 2 bottom).
However, for that quantity it can be observed that the isosurface is a connected region
from hub to the runner outlet for mesh 82M, while the isosurface is only present in
the rear part of the runner channel for mesh 56M. This shows that with the better wall
resolution the inter-blade vortex develops slightly stronger. Nevertheless, it can be
stated that the mesh refinement in the main flow has a bigger impact on the pressure
minimum in the inter-blade vortices compared to the wall resolution.
358 J. Wack, M. Zorn and S. Riedelbauch

82M
25M 56M

Fig. 2: Visualization of low pressure regions in one runner channel by an isosurface of

pressure with threshold vapor pressure for ℎ 𝑠 = 5.6 m. Top: Instantaneous snapshot.
Bottom: Time-averaged values for 50 runner revolutions.

In addition to the inter-blade vortex, a low pressure region can be found in Fig. 2
at the trailing edge near the hub. This region develops due to the backflow in the draft
tube cone that causes the flow detachment at the trailing edge.
The main fatigue damage potential of vortices results from the vortex movement.
While a vortex that remains at the same location causes a constant load, it is the
movement that results in permanent changes of the pressure field on the runner blades.
Consequently, it is of interest to have a closer look on the vortex movement in the
runner channel. This movement is visualized in Fig. 3 by the golden isosurface, which
represents the envelope of vortex movement. The idea of this envelope is to visualize
the region that falls below vapor pressure (for pressure level ℎ 𝑠 = 5.6 m) in at least
one time step using the minimum pressure of the simulated 50 runner revolutions.
For better clarity this envelope is clipped at a surface that is located close to the
runner trailing edge. The part of this surface, where the pressure falls below vapor
pressure, is shown in Fig. 3 and colored with the minimum pressure within the 50
runner revolutions. In addition, an isosurface of the instantaneous pressure (threshold
vapor pressure, isosurface is not clipped) is visualized in purple. This isosurface
allows to visualize the vortex size and compare it to the envelope, which enables to
get an impression about the vortex movement.
A comparison of the simulation results shows that for mesh 25M the envelope is
much smaller in the inter-blade vortex region compared to the finer meshes. This is
in agreement to the previous results and can be explained by the poorer resolution
Numerical simulation of vortex induced pressure fluctuations in a Francis turbine 359

of the pressure minimum in the vortex core. The significantly higher pressure in the
inter-blade vortex core for mesh 25M can also be seen in Fig. 3 in the minimum
pressure on the surface close to the runner outlet. There, very intense pressure minima
occur for meshes 56M and 82M, while this turns out to be significantly weaker for
mesh 25M.

𝑝𝑣

82M
25M 56M

Fig. 3: Vortex movement in a runner channel. The movement is visualized by an

isosurface (gold) that captures the region where the pressure is falling below vapor
pressure 𝑝 𝑣 for at least one time step within the simulated 50 runner revolutions.
Furthermore an isosurface of the pressure (threshold set to vapor pressure) for a
randomly chosen time step is displayed in purple. The envelope of vortex movement is
clipped at a plane in spanwise direction. On the part of this plane where the minimum
pressure is falling below vapor pressure the minimum pressure is visualized and the
rest of the plane is clipped.

Generally, it seems that the vortex movement in the hub region is smaller compared
to the shroud region. The second envelope region results from the other vortices and
the detachment region at the runner trailing edge. It gives an impression how far
vortices with significant intensity travel upstream. The results indicate that close to
the hub this region extends further upstream for meshes 56M and 82M, which again
is related to a better vortex resolution.

4.2 Pressure fluctuations on the runner blade

The impact of the vortices on the pressure fluctuations on one runner blade can
be found in Fig. 4. It shows the standard deviation on suction (top) and pressure
side (bottom). Additionally, an isosurface of the time-averaged velocity invariant
360 J. Wack, M. Zorn and S. Riedelbauch

(𝑄 = 5 · 105 s−2 ) is displayed to get an impression about the location of the vortices.
The vortices that are moving upstream close to the suction side are not visible by the
time-averaged velocity invariant because of their strong movement.
𝑠𝑡 𝑑 ( 𝑝)
0.0 𝜌𝑔𝐻 0.05

P11

P13

P7

P12

P6

P8

82M
25M 56M

Fig. 4: Standard deviation of pressure on the suction side (top) and pressure side
(bottom) of a runner blade. The standard deviation is calculated from 50 runner
revolutions.

Over a wide range of the runner blade the standard deviation and consequently
the pressure fluctuations are significantly higher on the suction side compared to
the pressure side. This can be explained by the fact that the vortices are located
closer to the suction side. While the inter-blade vortex causes for all meshes the
maximum standard deviation close to the shroud in the middle of the blade suction
side, over a wide range of the pressure side no impact of the inter-blade vortex on
the pressure fluctuations can be detected. Only close to the trailing edge, where
this vortex is located in the middle between pressure and suction side, an increased
standard deviation can also be observed on the pressure side.
A comparison of the different simulations shows qualitatively similar results.
Nevertheless, differences can be found. While mesh 25M and 56M have a region
of high pressure fluctuations close to the leading edge near the hub on the suction
side, for mesh 82M significantly smaller pressure fluctuations occur in that region.
This might result from a slightly different location of the inter-blade vortex in this
Numerical simulation of vortex induced pressure fluctuations in a Francis turbine 361

area that is shifted slightly downstream for mesh 82M compared to the other meshes.
Furthermore, a region of high standard deviation can be found on the suction side
in the area around measurement location P7 for mesh 25M. This region is located
further away from the shroud for the other meshes and might result from a different
interaction between the inter-blade vortex with the other vortices that is caused by the
mesh refinement in the main flow. On the pressure side all simulations show increased
pressure fluctuations at the leading edge close to the shroud that are very likely a
result of the rotor stator interaction. Further regions of higher standard deviation can
only be found close to the trailing edge for all simulation results. For mesh 82M a
region of significantly higher pressure fluctuations compared to the other meshes
occurs close to the shroud at the trailing edge.
While different simulation results can be compared all over the runner blade (see
Fig. 4), a comparison to experimental results is only possible at specific measurement
locations. In Fig. 5 the results of an FFT are presented for six sensor locations. Sensors
with an even number (top) are located on the pressure side and sensors with an odd
number (bottom) are on the suction side. The respective sensor location is marked in
Fig. 4. To get an impression about the impact of cavitation, experimental results are
displayed for ℎ 𝑠 = −10 m and 5.6 m.

Fig. 5: FFT results for different locations in on the pressure (top) and suction side
(bottom) of one runner blade. The positions are marked in Fig. 4.

Sensor P6 is located around midspan close to the trailing edge on the pressure
side. For 𝑓 / 𝑓 𝑅𝑈 < 5 and especially 𝑓 / 𝑓 𝑅𝑈 < 3 high amplitudes are present at this
sensor location in the simulation and measurement results. However, all simulations
show a significant overestimation of the amplitude. A comparison of the experimental
362 J. Wack, M. Zorn and S. Riedelbauch

results at different pressure level shows that the amplitude of the pressure fluctuations
decreases with a reduction of the pressure level. For that reason the overestimation of
the simulation results is probably a result of the neglection of cavitation.
At sensor P7 noticeable amplitudes of pressure fluctuations occur over a wide
spectrum of frequencies. This broadband spectrum is a result of vortices that are
located close to the sensor position. The amplitude of the simulation results decreases
with increasing mesh size. For mesh 82M the amplitude is in good agreement to
the experimental results at both pressure levels. The difference in the amplitude of
pressure fluctuations is also visible in Fig. 4 and is very likely a result of shifted
vortex locations.
A variety of different frequencies with noticeable amplitude can also be found
for sensor P8 ( 𝑓 / 𝑓 𝑅𝑈 < 10). At this location this broadband spectrum is caused by
the inter-blade vortex that is located close to that sensor (see Fig. 4). The simulation
results show a trend of overestimating the amplitude. Again, mesh 82M is closest
to the experimental results. However, the differences are not as big as for sensor P7.
For rotor stator interaction ( 𝑓 / 𝑓 𝑅𝑈 = 24) the measurement with ℎ 𝑠 = −10 m shows a
huge peak that is neither present in the simulation results nor at the measurement at
ℎ 𝑠 = 5.6 m. It is not clear at this point why this peak is so pronounced but it might be
a result of measurement issues and should be taken with caution.
Sensor P11 is located just upstream of the low pressure region at the trailing edge
close to the hub (see Fig. 2 and 4). It can be observed that the simulation results behave
totally different compared to the measurement - especially with ℎ 𝑠 = 5.6 m. While the
simulation results show a broadband frequency spectrum up to around 𝑓 / 𝑓 𝑅𝑈 < 20,
the experiment with ℎ 𝑠 = 5.6 m has only a noticeable peak at 𝑓 / 𝑓 𝑅𝑈 = 1. The
measurement with ℎ 𝑠 = −10 m also has this peak but shows also some broadband
frequency spectrum that has a lower amplitude compared to the simulation results. The
difference between simulation an experiment is most likely a result of the neglection
of cavitation in the simulation. Due to the absence of the broadband frequencies for
ℎ 𝑠 = 5.6 m, it seems that the occurrence of cavitation changes the flow and results in
a suppression of vortices in this region.
Sensor P12 is located at a similar location like P11 but on the pressure side
of the blade. There, the deviation between simulation and experiment is smaller.
Nevertheless, all simulations show relevant amplitudes up to around 𝑓 / 𝑓 𝑅𝑈 = 5,
while the measurement at low pressure level shows only noticeable peaks up to
𝑓 / 𝑓 𝑅𝑈 = 1.5. Again, this deviation is probably caused by the neglection of cavitation,
which is already indicated by preliminary two-phase simulation results.
Finally, the FFT results for sensor P13, which is located around midspan, also show
a broadband frequency spectrum with overestimated amplitude by the simulations.
Due to the location close to the low pressure region at the trailing edge near the hub
the neglection of cavitation is the probable cause of the deviation. A comparison of
the different pressure levels of the experimental results indicates that the size of the
cavitation region has an impact on the pressure fluctuations. While for ℎ 𝑠 = 5.6 m
the highest amplitude occurs around 𝑓 / 𝑓 𝑅𝑈 = 5, the maximum amplitude is shifted
to lower frequencies for ℎ 𝑠 = −10 m.
Numerical simulation of vortex induced pressure fluctuations in a Francis turbine 363

Preliminary results of two-phase simulations support that the occurrence of cavi-

tation results in a damping of the amplitude of pressure fluctuations. Furthermore,
frequency shifts, as observed at sensor location P13, can be captured by consider-
ing cavitation in the simulations. Detailed comparisons of single- and two-phase
simulation results will be performed in the future.

5 Computational resources

The results of the numerical simulations have been performed on the HPE Apollo
(Hawk) at the High Performance Computing Center Stuttgart. The architecture of the
HPE Apollo consists of 5632 compute nodes with AMD EPYC Rome processors
with 128 cores and 256 GB memory. On Hawk an infiniband HDR based interconnect
with a 9-dimensional enhanced hypercube topology is used.
A strong scaling test is carried out for the meshes 56M and 82M on the file system
ws10. For mesh 56M the speedup curve that is normalized to the simulation with 256
cores and time per time step are shown in Fig. 6 for the ANSYS CFX versions 19.5
and 21.1. A normalization of the speedup to 128 cores is not possible as this case
needs more than 256 GB of memory.

8 56M-v195
56M-v211 250
time per time step in s

7 ideal
6 200
speedup

5
150
4
3 100
2
1 50
256 512 768 1024 1280 1536 1792 2048 256 512 768 1024 1280 1536 1792 2048
number of cores number of cores

Fig. 6: Speedup and time per time step for mesh 56M.

For the two versions of ANSYS CFX the speedup is almost identical. As the
time per time step is also not affected by the version of the solver, it can be stated
that parallel performance has not been improved from version 19.5 to 21.1. The
code shows an acceptable scaling up to approximately 1536 cores, where parallel
performance is above 87%. This corresponds to around 36000 cells per core.
As ANSYS CFX version 21.1 did not show any improvement in parallel perfor-
mance for mesh 56M, the speedup test for mesh 82M is only performed for version
19.5. The results are presented in Fig. 7. For the speedup the results are normalized to
364 J. Wack, M. Zorn and S. Riedelbauch

the simulation with 512 cores, which again is selected due to memory requirements.
The results indicate a not ideal but reasonable scaling up to 1536-2048 cores, where
parallel performance is 87% or 84%, respectively.

5.0 82M-v195 200 82M-v195-HMPT-MPI

ideal runs on job with 64 nodes

time per time step in s

4.5 180 run on 16 nodes with 64 cores
4.0 160
3.5
speedup

140
3.0
120
2.5
2.0 100
1.5 80
1.0 60
512 1024 1536 2048 2560 512 1024 1536 2048 2560
number of cores number of cores

Fig. 7: Speedup and time per time step for mesh 82M.

As all jobs that use less than 64 nodes are performed in a special partition, for one
case it is tested how the ideal use of the hypercube topology affects the simulation
time. This is achieved by running a job on 64 nodes that performs eight simulations
on 1024 cores each. Analyzing the time per time step shows that all simulations of the
64-node-job are slightly but not significantly faster compared to the 8-node-job that
runs on the special partition. The fastest and slowest simulation of the 64-node-job
are 0.6% faster or slower, respectively, compared to the averaged time per time step
of the eight simulations. A comparison to the 8-node-job shows that it is 0.8% slower
compared to the averaged time per time step of the simulations with the 64-node-job.
To investigate the effect of only using 64 cores per node, one 1024-core run is
performed on 16 nodes instead of only using 8 nodes when all 128 cores per node
are used. This gives a reduction of time per time step of 2.3%. Consequently, it is
not advisable to use this approach as the gain in simulation time does not justify the
usage of significantly higher computational resources.
Finally, the impact of the MPI method is investigated. Currently it is possible to
use Open MPI or HMPT MPI on Hawk with ANSYS CFX. The comparison of these
methods for a job on 1536 cores shows that the simulation with HMPT MPI is 16.5%
slower. Consequently, Open MPI is used for all simulations on Hawk.
Within the reporting period around 18 million core-h have been used. From that
approximately 11 million core-h were needed for the results of the single-phase
simulations presented in this study. The other core-h were used for two-phase
simulations that are not finished yet. The two-phase results will be published in the
future.
Numerical simulation of vortex induced pressure fluctuations in a Francis turbine 365

6 Conclusion and outlook

In this study a deep part load operating point of a Francis turbine at model scale
has been investigated. Three different meshes have been used for the single-phase
simulations. The results show that head is almost not affected by the mesh resolution,
while for the torque the mesh refinement results in a reduction of the deviation to the
experiment. Especially the refinement of the main flow has an impact on torque as
well as the vortex movement. A finer resolution of the boundary layer also has some
effect that, however, is lower compared to the refinement of the main flow.
For the different meshes a slightly changed location of the vortices is possible.
This is accompanied by shifted locations of high pressure fluctuations as they are
mainly caused by the vortices. A comparison to experimental results indicates that at
some pressure sensors the amplitude of the pressure fluctuations are overestimated
due to the neglection of cavitation. Furthermore, cavitation can result in a shifted
frequency spectrum that cannot be captured by single-phase simulations.
The strong scaling test did not show a difference in the parallel performance of the
ANSYS CFX versions 19.5 and 21.1. For simulations with the mesh 56M the parallel
performance is acceptable up to 1536 cores. With the finest mesh (82M) even 2048
cores are feasible. In terms of the MPI method the Open MPI is preferable as it is
16.5% faster compared to the HMPT MPI method.
The pressure field in the runner of the presented single-phase simulations will now
be used as input for structural mechanical investigations. With the results a reduced
simulation setup will be developed that can be used in the design process of future
turbines. Furthermore, two-phase simulations are carried out to investigate the impact
of cavitation on the pressure fluctuations more detailed.

Acknowledgements The simulations were performed on the national supercomputer HPE Apollo
Hawk at the High Performance Computing Center Stuttgart (HLRS) under the grant number 44047.
Furthermore, the authors acknowledge the financial support by the Federal Ministry for Economic
Affairs and Energy of Germany in the project FrancisPLUS (project number 03EE4004A).
366 J. Wack, M. Zorn and S. Riedelbauch

References

1. Seidel, U., Mende, C., Hübner, B., Weber, W., Otto, A.: Dynamic loads in Francis runners
and their impact on fatigue life. In: IOP Conference Series: Earth and Environmental Science,
vol. 22 p 032054 (2014)
2. Dörfler, P., Sick, M., Coutu, A.: Flow-induced pulsation and vibration in hydroelectric
machinery: Engineer’s guidebook for planning, design and troubleshooting. Springer (2012)
3. Wack, J., Riedelbauch, S., Yamamoto, K., Avellan, F.: Two-phase flow simulations of the
inter-blade vortices in a Francis turbine. In: Proceedings of the 9th International Conference on
Multiphase Flow, Florence, Italy (2016)
4. Conrad, P., Weber, W., Jung, A.: Deep Part Load Flow Analysis in a Francis Model Turbine by
means of two-phase unsteady flow simulations. In: Proceedings of the Hyperbole Conference,
Porto, Portugal (2017)
5. Yamamoto, K.: Hydrodynamics of Francis turbine operation at deep part load condition. Ph.D.
thesis École Polytechnique Fédérale de Lausanne (2017)
6. Stein, P., Sick, M., Dörfler, P., White, P., Braune, A.: Numerical simulation of the cavitating
draft tube vortex in a Francis turbine. In: Proceedings of the 23rd IAHR Symposium on
Hydraulic Machinery and Systems, Yokohama, Japan p 228 (2006)
7. Wack., J., Riedelbauch, S.: Numerical simulations of the cavitation phenomena in a Francis
turbine at deep part load conditions. In: Journal of Physics: Conference Series, vol. 656 p
012074 (2015)
8. Menter, F.R.: Stress-Blended Eddy Simulation (SBES) - A new Paradigm in hybrid RANS-LES
Modeling. In: Proceedings of the 6th HRLM Symposium, Strasbourg, France (2016)
9. Menter, F.R., Egorov, Y.: The Scale-Adaptive Simulation Method for Unsteady Turbulent Flow
Predictions. Part 1: Theory and Model Description. In: Flow, Turbulence and Combustion,
vol. 85, pp. 113-138 (2010)
10. Wack, J., Beck, J., Conrad, P., von Locquenghien, F., Jester-Zürker, R., Riedelbauch, S.:
A Turbulence Model Assessment for Deep Part Load Conditions of a Francis Turbine. In:
Proceedings of the 30th IAHR Symposium on Hydraulic Machinery and Systems, Lausanne,
Switzerland (2021)
11. Menter, F.R.: Best practice: Scale-resolving simulations in ANSYS CFD - Version 2.0. ANSYS
Germany GmbH (2015)
12. Čelič, D., Ondráčka, H.: The influence of disc friction losses and labyrinth losses on efficiency
of high head Francis turbine. In: Journal of Physics: Conference Series, vol. 579 p 012007
(2015).
13. von Locquenghien, F., Faigle, P., Aschenbrenner, T.: Model test with sensor equipped Francis
runner for Part Load Operation. In: Proceedings of the 30th IAHR Symposium on Hydraulic
Machinery and Systems, Lausanne, Switzerland (2021)
Validation of ACD and ACL propeller simulation
using blade element method based on airfoil
characteristics

Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

Abstract Investigations concerning the aerodynamic interactions of propellers at

new configurations such as distributed propulsion, wingtip mounted propellers, etc.,
are characterized by a large, multidimensional parameter space. Simplified models
like the steady-state Actuator Disk (ACD) or the unsteady Actuator Line (ACL) are
therefore common to reduce the calculation time. Since the quality of the models
depends on the accuracy of the input data, it is essential that these are represented
in a physically correct manner. In this study a validation and comparison of blade
element method (BEM) based ACD/ACL models with fully resolved simulations and
experimental data known from literature is performed. It is demonstrated that both
the effects of the propeller on the wing and vice versa are accurately captured with
BEM-based propeller models.

1 Introduction

The CFD simulation of propeller flows is of increasing relevance due to research

into new configurations such as distributed propulsion, wingtip mounted propellers,
etc., which benefit from positive aerodynamic interactions between the propellers
and the other components of the aircraft. Investigations concerning the aerodynamic
interactions involving propellers are characterized by a large, multidimensional
parameter space. In order to reduce the calculation time simplified models like the
steady-state Actuator Disk (ACD) or the unsteady Actuator Line (ACL) are thereby
necessary. Stokkermans has impressively demonstrated the capability of ACD and
ACL to simulate the complex effects of the propeller slipstream on the wing when
using prescribed propeller loadings as input [1]. However, for the design of novel

Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

University of Stuttgart, Faculty of Aerospace Engineering and Geodesy, Institute of Aerodynamics
and Gas Dynamics (IAG), Pfaffenwaldring 21, 70569 Stuttgart,
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 367
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_22
368 Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

configurations it is crucial to also capture the effect of the wing on the propeller
loads, both on the averaged propeller coefficients and on the phase dependence over
one propeller revolution. In addition to the use of prescribed loadings, in the DLR
TAU-code [2] the blade forces for ACD and ACL can be calculated via blade element
method (BEM) [3, 4]. In the present study the airfoil characteristics of the propeller
blades in the terms of lift and drag polars (𝑐 𝑙 , 𝑐 𝑑 ) are first extracted from a steady-state
single blade simulation and are thereafter used for ACD, ACL calculations and
validated by comparison with fully resolved simulations and experimental data. Three
configurations were simulated in this study, see Fig. 1:
• A single propeller blade: to generate the airfoil characteristics for the ACD, ACL
models
• An isolated propeller with an axisymmetric nacelle: to compare and validate the
slipstream data
• An installed tractor propeller mounted at the wingtip of an semi-span wing: to
compare and validate the effect of the slipstream on the wing as well as the wing
effect on the propeller

blade propeller flap

R
c
nacelle
spinner half-wing
spinner

(a) single blade (b) isolated propeller (c) installed propeller

Fig. 1: Geometries of the simulated configurations.

The geometries are identical to those of Stokkermans et al. [1] and Sinnige et
al. [5]. The single propeller blade, with a diameter of 0.237 𝑚 and a blade pitch angle
of 𝛽75 = 23.9◦ , was simulated for a range of advance ratios between 𝐽 = 0.6 − 1.2.
The four-bladed propeller at the isolated and installed case uses the same blade
geometry and was simulated with an advance ratio of 𝐽 = 0.8 at a freestream velocity
of 40 𝑚/𝑠 (𝑀𝑎 = 0.12) and a freestream angle of attack of 𝛼 = 0◦ . The wing of the
installed case has a semi-span of 𝑏 = 0, 327 𝑚 with a symmetrical NACA 642 𝐴015
airfoil, a chord length of 0.240 𝑚 and a plain flap with a deflection of 𝜑 = +10◦ to
generate lift.
Validation of ACD and ACL coupled with BEM 369

2 Simulation methods and numerical setup

2.1 Propeller Simulation

Three different approaches to model the propeller impact within the CFD simulation
are used: Actuator Disk (ACD), Actuator Line (ACL) and fully resolved propeller
blades (full).

2.1.1 Fully resolved propeller (full)

The fully resolved propeller simulation in TAU is enabled by the Chimera technique,
see Stürmer [7]. Thereby, a propeller grid rotates in front of the fixed background
grid with a cylindrical hole at the position of the propeller. A sufficient overlap of
the two grids ensures a smooth transition of the flow values. In this study a viscid
modeling of the propeller blades is achieved by resolving the blade boundary layer to
capture the propeller friction losses.

2.1.2 Actuator Disk (ACD)

In the standard TAU version, an Actuator Disk method is available, where the blade
forces averaged over one propeller revolution and introduced stationary into the flow
field on a circular boundary condition (BC), see Raichle et al. [3]. The local effective
angle of attack 𝛼 and the local inflow velocity are directly determined from the
flow at each point of the disk. The steady-state calculation enormously reduces the
computation time compared to the fully resolved simulation. However, the ACD has
limitations due to the simplification made, the slipstream is stationary and blade tip
vortices are not taken into account.

2.1.3 Actuator Line (ACL)

An unsteady Actuator Line (ACL) method was implemented into the TAU code by
the presenting authors [4]. Here the blade forces are also introduced into the flow field
on the propeller disk BC, but along discrete lines rotating in time. The calculation
process of the ACL consists of four steps: 1. the calculation of the time-dependent
ACL positions and the inflow conditions, 2. the calculation of the sectional blade
loads via BEM, 3. the distribution of the forces and 4. the insertion into the flow field.
The determination of the local effective angle of attack is more difficult than with the
ACD, since it is not possible to take the flow values directly at the line. The circulation
bound to the blade element causes an upwash upstream and a downwash downstream
from the ACL, both of similar magnitude. By averaging the flow values over an
angular range up- and downstream from each ACL point, this effect can therefore be
eliminated and the effective angle of attack can be determined. The angular ranges are
defined by two angles. In order to prevent discontinuities by reducing the force applied
to one cell, the forces are distributed two-dimensionally in the radial and chordwise
directions of the blade. In the TAU ACL three different probability density functions
370 Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

(PDF) were implemented for the distribution, see [4]. Besides the usual isotropic
Gauss distribution, an anisotropic Gauss distribution following Churchfield et al. [8]
is available, where the forces are distributed in chordwise direction as a function
of the local blade chord length. In addition, an anisotropic Gumbel distribution
can be used, where the force distribution is more similar to that of a typical airfoil.
Compared to the ACD, the unsteady ACL simulation increases the computation time,
but compared to the fully resolved computation, significantly fewer cells are required
for the propeller meshing, which reduces the computational effort.

2.2 Single blade simulation

For BEM calculation with the ACD and ACL models the airfoil characteristics must
be available as input data in the form of lift and drag polars. To obtain these, different
methods are available [9]. Here, 3D CFD simulations are carried out with a single
propeller blade via a quarter propeller model, in order to reduce the number of cells.
Periodic BC enable a steady-state rotational in- and outflow through the computational
domain. The tangential velocity is generated by the rotation of the grid and the axial
velocity by the farfield BC. To determine a propeller polar the inflow conditions have
to be varied, either by variation of the propeller rotation Ω at constant inflow velocity
𝑣 ∞ or by variation of the inflow velocity at constant rotation. After the simulations,
the coefficients 𝑐 𝑙 and 𝑐 𝑑 have to be determined at different sections of the propeller
blade for each inflow case. They are obtained from the effective angle of attack and
by integration of the difference in pressure coefficients 𝑐 𝑝 between the upper and
the lower side of the respective blade section and the skin friction coefficients 𝑐 𝑓 .
The local induced angle of attack 𝛼𝑖 can be extracted from the CFD data using the
reduced axial velocity technique (RAV), see Johansen et al. [10]. For this purpose,
two planes are used, shaped as a circular ring sector in front of and behind the blade
section, on which the flow data are averaged, see Fig. 4b. The effective angle of attack
is than estimated by linear interpolation between the planes.

2.3 Grids and numerical setup

The hybrid grids used in this study are shown in figure 2. Structured cells are utilized
in the cylindrical slipstream area up to the end the nacelle with a radius of 1.1 𝑅. The
boundary layer around the spinner, the nacelle and the wing is resolved with structured
cells and 𝑦 + < 1. The structured domain is embedded in a refined unstructured domain
with a radius of 10 𝑅 and a length of 30 𝑅. The latter domain is itself embedded
in a coarser farfield area with a distance of up to 50 𝑅. By using a farfield BC all
influences by wind tunnel walls are neglected. In order to take advantage of the
symmetry of the installed propeller configuration a symmetry plane BC is applied
here. For the single blade simulation, a one-quarter section of the propeller grid is
used with rotationally periodic BC on the sides.
Validation of ACD and ACL coupled with BEM 371

unstructured cells
structured cells 10 R

50 R single blade 1.1 R

grid

periodic plane BC

farfield BC 2R
20 R

50 R
2R

10 R
chimera grid or ACD/ACL

symmetry plane BC installed propeller grid

Fig. 2: Computational Grids.

The research was preceded by two grid convergence studies: one for discretization
of the single propeller blade and one for discretization of the propeller with the
ACD/ACL methods. The results are shown in Fig. 3. The number of cells in chordwise
and radial direction as well as the number of cells inside the boundary layer were
varied for the single propeller blade. The selected grid has a deviation from the
Richardson-extrapolated value of Δ𝑐 𝑡 = 0.26% and Δ𝑐 𝑝 = 0.19%. The airfoil is
discretized by 116 points, the blade by 62 points in radial direction, excluding the hub,
and the boundary layer by 24 structured cells and 𝑦 + < 1. The same propeller blade
grid was also used for the fully resolved simulation, copied to obtain four blades. For
the ACD/ACL study, the discretization of the propeller disk BC was varied in radial
𝑛𝑟 and tangential 𝑛𝑡 direction, the selected grid has a deviation of Δ𝑐 𝑡 , 𝑝 < 0, 5%. The
single blade grid has about ≈ 1.7 millions cells. With ACD/ACL, the isolated propeller
grid has ≈ 8.7 million cells and the installed propeller grid ≈ 12 million. With the
fully resolved propeller ≈ 5 million more cells are necessary. For all calculations
the second order central scheme was used for spatial discretization, the implicit
Backward Euler for time discretization and the one-equation Spalart–Allmaras model
to generate the turbulence. No laminar-turbulent transition was considered. The
calculation time of the time-resolved simulations covered at least five propeller
revolutions. A physical timestep corresponding to 4◦ rotation of the propeller was
chosen. For the last revolution, over which the results were later averaged, the physical
timestep was reduced to a value corresponding to 1◦ rotation.
372 Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

2 2
blade, full cT ACL cT
1.5 blade, full cP 1.5 ACL cP
[%]

[%]
1 1

0.5 0.5

00 00
0.5 1 1.5 2 2.5 3 3.5 0.5 1 1.5 2
3/2
(1/n) 10 [-] 4 1/(nrnt)104[-]

(a) single blade, fully resolved (b) ACL

Fig. 3: Spatial discretization.

3 Results

3.1 Airfoil characteristics

For the determination of the propeller blade polars, seven advance ratios were
simulated with the single blade. Two methods were thereby applied: first, the
rotational speed was varied at a constant inflow velocity, and second, the inflow
velocity was varied at a fixed rotational speed, both producing almost the same thrust
and power coefficients. For the following polar calculation the varying velocities are
used, because it was assumed that this corresponds more to the velocity change due to
the wing influence. Simulation data from Stokkermans [6] is shown as a comparison,
which slightly differs with increasing distance from the advance ratio 𝐽 = 0.8, which
was used as the (Reynolds number) design point of the blade grid. However, the
general characteristic is matched.

averaging plane
0.15
2 dr~0.1R
0.1
c
cT, cP [-]

v var. ct 1
0.05 v var. cp dx
var. ct
var. cp
0
Stokkermans [1] ct
dx
Stokkermans [1] cp
-0.05
0.55 0.65 0.75 0.85 0.95 1.05 1.15 1.25
J [-]
(a) Global propeller coefficients (b) RAV method

Fig. 4: Determination of the propeller airfoil characteristics. Data from Stokkermans

[6].
Validation of ACD and ACL coupled with BEM 373

6
c0.5
c2
4 c5
ACD
2

-2
0.2 0.4 0.6 0.8 1
r/R [-]
(a) induced axial velocity (b) effective angle of attack

Fig. 5: Propeller blade angle of attack.

Figure 5a shows the axial velocity induced by the propeller blade and different
positions of the RAV averaging planes (0.5𝑐, 0.75𝑐, 1𝑐, 2𝑐, 3𝑐, 4𝑐, 5𝑐). As can be
seen in figure Figure 5b, the distance between the plane and the propeller disk has an
influence on the angle of attack determined with the RAV, which varies by approx.
𝛼 = 0.5◦ . For comparison, the effective angle of attack determined with the ACD is
also shown. Figure 6 shows the forces along the radius obtained with polars based
on the different RAV distances as well as the single blade results for 𝐽 = 0.8. The
thrust and tangential force distributions present the influence of the polars that are
used as input. For the polar with RAV distance 5𝑐, with the largest angle of attack
difference between ACD and RAV, the forces are recognizably underestimated. The
ACD forces with RAV distance 0.5𝑐 capture the blade data most accurately, although
an underestimation remains visible in the medium blade area. The results demonstrate
that it is crucial for the method of extracting the polars to use an effective angle of
attack as closely as possible to that of the propeller model.

90 40
ACD 0.5c ACD 0.5c
70 ACD 1c ACD 1c
ACD 3c 30 ACD 3c
ACD 0.75c ACD 0.75c
50 blade blade
20
30

10
10

-10 0
0.1 0.25 0.4 0.55 0.7 0.85 1 0.1 0.25 0.4 0.55 0.7 0.85 1
r/R [-] r/R [-]
(a) normal force distribution (b) tangential force distribution

Fig. 6: Radial force distributions at the propeller disk.

 374 Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

3.2 Isolated propeller

3.2.1 Propeller slipstream

Figure 7 shows the radial distribution of the total pressure coefficient and the axial
and tangential velocity components from the ACD, ACL, fully resolved simulations
as well as experimental data from Stokkermans [1] at different positions. The time
averaged axial and tangential velocity distributions for ACD and ACL are a little
lower than the fully resolved results. The blade tips vortex at 𝑟/𝑅 = 1 is predicted by
both unsteady methods, but is less pronounced than in the experimental data. The less
pronounced blade tip vortices, caused by numerical diffusion, were also described by
Stokkermans [1]. In the root region, the tangential velocity is slightly underestimated
by the polar data, which is also reflected in lower tangential forces. However, in
general, the slipstream values are captured well by the BEM-based models.

1.2 exp [1]

full
ACD
1
ACL

0.8
r/R [-]

0.6

0.4

0.2
0.6 0.8 1 1.2 1.4 1.6 1.8 0.6 0.8 1 1.2 1.4 1.6 1.8 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.05 0.1 0.15
cpt [-] va/V [-] va/V [-] vt/V [-]

(a) total pressure (b) axial velocity (c) axial velocity (d) tang. velocity

Fig. 7: Radial total pressure and velocity distributions in the slipstream. Experimental
data from Stokkermans [1].

3.2.2 Propeller blade flow

Figure 8 shows the ACL forces on the propeller disk BC with the three PDFs (see
Sec. 2.1.3) compared to the fully resolved blades. In contrast to the isotropic force
distribution, which is concentrated on a narrow line, the anisotropic Gauss and
Gumbel distributions follow the blade geometry. The closest match is obtained with
the Gumbel distribution, where the asymmetric distribution in chordwise direction
is visible. Figure 9 shows the pressure distribution around a propeller blade in an
orthogonal section at 𝑟/𝑅 = 0.75 for all distributions as well as the fully resolved
simulation. For better comparability, the blade section is also plotted in the ACL
results. Due to the ACL force insertion on the propeller disk BC, the pressure
distribution around the blade is in a way projected onto the disk. The too dense
concentration of the isotropic PDF is also manifested in the pressure distribution.
Validation of ACD and ACL coupled with BEM 375

The asymmetric pressure distribution in the chordwise direction is only reflected

by the Gumbel distribution. However, all ACL approaches show a shifted pressure
distribution. An improvement could probably be achieved by a distribution around
the pressure point instead of the geometrical center of the blade.

(a) Iso. Gauss (b) Aniso. Gauss (c) Gumbel (d) full

Fig. 8: Force distribution of ACL and fully resolved propeller.

(a) Isotropic Gauss (b) Anisotropic Gauss

(c) Gumbel (d) full

Fig. 9: Pressure distribution around the propeller blade.

376 Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

3.3 Installed propeller

3.3.1 Propeller wing interaction
Figure 10a shows the influence of the propeller slipstream on the averaged pressure
distribution of the wing near the outboard flap edge. The fully resolved simulation
captures the experimental data [1] accurately. On the upper side, ACD and ACL
predict a slightly lower pressure. Figure 10b shows the pressure distribution in the
propeller blade tip area for the ACL and the fully resolved methods (single timestep,
time-accurate) as well as for comparison reason the distribution for the ACD method
(steady-state). Compared to the ACD distribution, the time-accurate distributions
present pressure fluctuations caused by the blade tip vortices. The position and the
amplitude of these fluctuations are similar between the ACL and the fully resolved
solutions.
-0.8 -0.8
-0.6 -0.6
-0.4 -0.4
-0.2 -0.2
cp [-]
cp [-]

0 0
0.2 ACD 0.2
0.4 ACL 0.4 ACD
full ACL
0.6 0.6
exp [1] full
0.8 0.8
10 0.2 0.4 0.6 0.8 1 10 0.2 0.4 0.6 0.8 1
x/c [-] x/c [-]
(a) pressure in slipstream (b) pressure at tip
0.01 0.45
ACL
cL, w

ACD
0 full
0.35 off
cl [-]

-0.010 90 180 270 360

0.25
cD, w

0 90 180 270 360 0.15 0.2 0.4 0.6 0.8 1

[°] y/b [-]
(c) phase dependent lift and drag (d) spanwise lift
Fig. 10: Influence of the propeller slipstream on the wing. Experimental data from
Stokkermans [1].

Figure 10d shows the spanwise lift distribution of the propeller simulation compared
to the isolated wing. The increase in lift is captured by all methods and can be divided
into regions inside and outside the area covered by the propeller slipstream. The ACD
generally overestimates the lift increase, which is also indicated in the simulations by
Stokkermans [1] using prescribed forces. The ACL overestimates the lift in the region
covered by the propeller slipstream, but the distribution outside of it agrees well
with the fully resolved one. The effect of the slipstream on the wing is in principle
captured correctly by the BEM-based models. However, compared to results with
prescribed forces [1] a slightly larger deviation occurs.
Validation of ACD and ACL coupled with BEM 377

3.3.2 Wing propeller interaction

The advantage of BEM-based ACD and ACL simulations over prescribed forces
is that, in addition to the influence of the propeller on the wing, the change of the
propeller forces due to the wing’s presence are also captured, which is demonstrated
in Figure 11. Figure 11a shows the variation of the blade lift coefficient over one
revolution at a blade position of 𝑟/𝑅 = 0.75 for all three methods.

0.045 0.001 ACL

ACD
full

cT
ACL
0.03 0
full

-0.0010 90 180 270 360

0.015
cl, p [-]

0.001

cP
0
0

-0.0150 -0.0010 90 180 270 360

90 180 270 360
[°] [°]

(a) phase dependent blade lift (b) phase dependent thrust and power

Fig. 11: Influence of the wing slipstream on the propeller.

The influence of the wing leads to non-constant blade loads. When passing
through the reduced velocity field in front of the wing, an increase in lift occurs on
the propeller blade. With the ACD, for which a continuous evaluation is shown, a
maximum appears at about 250◦ . The ACL and fully resolved methods are evaluated
only at discrete positions of the blades. In principle, the lift variations correspond
well across all methods. However, the ACL shows a shift of the maximum to an earlier
propeller position in comparison. While the ACD captures the position dependence
of the propeller forces, they are introduced into the flow field in a purely steady-state
manner. With the ACL and fully resolved methods the individual blade forces vary
accurately in time. Figure 11b shows the unsteady propeller thrust and power variation
for both methods. The amplitudes match for both methods, whereby the periodicity
of the ACL is superimposed by higher-frequency fluctuations. The reason for this is
assumed to be that the distribution of the discrete ACL line loads on the neighboring
cells varies slightly depending on the position of the line relative to the grid. The
small shift in the phase dependence by less than an eighth of a rotation can again
be seen. This shift can be attributed to the ACL-internal calculation of the local
effective angle of attack (see Sec. 2.1.3). The phase shift depends on the distance
of the averaging ranges and the blade and decreases with a smaller distance. The
strongest influence of the wing is thus not captured when the blade passes in front
of the wing, but instead when the upstream averaging range passes in front of it. In
the results shown, the averaging range is from 25 − 40◦ . At smaller distances, the
influence of the blade dominates and the effective angles of attack are not calculated
correctly. A reduction of the phase shift could be achieved with a modified effective
angle of attack determination method, which is planned for the future: instead of
378 Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

averaging over constant lines, the ranges could be coupled to the chord length for
each ACL point, whereby a larger distance at the root would then be possible along
with a smaller distance at the tip.

4 Scaling test
To investigate the efficiency of the TAU code, a preliminary scaling test on the HPC
Hawk was conducted using the isolated propeller case with the fully resolved propeller
rotated by the chimera-technique. This propeller simulation method was chosen for
the scaling test, because it is considered the most complex. Unsteady simulations
were conducted for various number of domains with 3000 inner iterations, which
corresponds to a one third rotation of the propeller with a physical timestep of 4◦ and
100 inner iterations per time step. The grid was partitioned with the TAU internal
private partitioner with a varying numbers of subgrids, ranging from 64 to 1280
domains. The results of the scaling test are shown in Figure 12.

nodes cores points/domain

1 64 100961
1 128 50481
104
2 256 25240
3 384 16827
time [s]

4 512 12620 103

5 640 10096 ideal

total time
6 768 8413 solver time
102
7 896 7212 104 105
points/domain [-]
8 1024 6310
9 1152 5609
10 1280 5048

Fig. 12: Test cases and scaling behavior of the DLR TAU-code on Hawk.

The monitored values were the pure solver time without initialization and output
routines and the total computation time. The TAU-code, using the private partitioner
and the chimera-technique, shows a reasonably efficient scaling behavior on the
HPC Hawk with decreasing ratio of points/domain. A performance optimum exists
in the range of 10, 000 to 50, 000 grid points per domain, where an almost ideal
scaling behavior is reached. For a lower ratio the total computation time seems to
be limited by the input/output time even though the pure solver time continues to
decrease. A further increase of the number of cores was therefore rejected, since the
Validation of ACD and ACL coupled with BEM 379

total computation time would not decrease any further. The simulations of the study
with the different grids (described in Sec. 2.3) were performed accordingly with a
points/domain ratio in the ideal scaling range.

5 Conclusion

In the study, simulations using Blade Element method (BEM) based Actuator Disk
(ACD) and Actuator Line (ACL) propeller modeling approaches are compared with
fully resolved simulations and experimental data known from literature, for an isolated
and an installed propeller. The characteristics of the propeller blades as input data for
the BEM are determined by steady-state 3D simulations of a single blade with different
inflow velocities. The effective angles of attack are determined with the Reduced
Axial Velocity method (RAV). It is shown that the effective angle of attack distribution
determined by the RAV method varies with the distance of the averaging planes from
the propeller disk. Therefore, the effective angles of attack distribution should be
chosen as close as possible to that of the method used for propeller simulation (ACD,
ACL). With appropriate polars, force, pressure and velocity distributions obtained
from ACD and ACL calculations agree well with fully resolved and experimental
data. Using the ACL method, the force and pressure distribution of a line with a
Gumbel distribution agree better with the fully resolved blade than with an isotropic
or anisotropic Gauss distribution. The slipstream calculated with BEM-based ACD
and ACL achieves a similar effect on the wing as the fully resolved one. Compared
with predefined forces, a slightly larger deviation must be accepted, however. Due
to the BEM the phase-dependent influence of the wing on the propeller flow is also
captured with ACD and ACL. The study demonstrates that BEM-based propeller
models are suitable for design studies, to consider the interactions between propeller
and wing in both directions.

Acknowledgements This study was carried out within the LuFo V 3 project ELFLEAN, which is
funded by the Federal Ministry for Economic Affairs and Energy. The authors gratefully thank Leo
Veldhuis from TU-Delft for providing the propeller geometry for the single blade and fully resolved
simulations to validate the BEM-based models.
380 Michael Schollenberger, Mário Firnhaber Beckers and Thorsten Lutz

References

1. Stokkermans, T.C., v. Arnhem, N., Sinnige, T., Veldhuis, L.L.: Validation and Comparison of
RANS Propeller Modeling Methods for Tip-Mounted Applications. AIAA SciTech Forum,
(2018). doi:10.2514/6.2018-0542
2. Schwamborn, D. and Gerhold, T. and Heinrich, R.: The DLR TAU-code: Recent applications
in research and industry, ECCOMAS CFD 2006 CONFERENCE,(2006), https://elib.d
lr.de/22421/
3. Raichle A.: Flusskonservative Diskretisierung des Wirkscheibenmodells als Unstetigkeitsfläche.
PhD thesis, DLR, (2017).
4. Schollenberger M. , Lutz T. and Krämer E.: Boundary Condition Based Actuator Line Model
to Simulate the Aerodynamic Interactions at Wingtip Mounted Propellers. New Results in
Numerical and Experimental Fluid Mechanics XII, (2020)
5. Sinnige, T., de Vries, R., Della Corte, B., Avallone, F., Ragni, D., Eitelberg, G., Veldhuis,
L.L: Unsteady Pylon Loading Caused by Propeller-Slipstream Impingement for Tip-Mounted
Propellers, Journal of Aircraft 55:4, (2018)
6. Stokkermans, T.C.: Aerodynamics of Propellers in Interaction Dominated Flowfields: An
Application to Novel Aerospace Vehicles, (2020). https://doi.org/10.4233/uuid:46178824-bb80-
4247-83f1-dc8a9ca7d8e3
7. Stuermer, A.: Unsteady CFD Simulations of Propeller Installation Effects. 42nd AIAA/AS-
ME/SAE/ASEE Joint Propulsion Conference & Exhibit, (2006). doi:10.2514/6.2006-4969
8. Churchfield, M.J., Schreck, S.J., Martinez, L.A., Meneveau, C., Spalart, P.R.: An Advanced Ac-
tuator Line Method for Wind Energy Applications and Beyond. 35th Wind Energy Symposium,
AIAA SciTech Forum, (2017). doi:10.2514/6.2017-1998
9. Schollenberger M. and Lutz T.: Comparison of Different Methods for the Extraction of Airfoil
Characteristics of Propeller Blades as Input for Propeller Models in CFD. New Results in
Numerical and Experimental Fluid Mechanics XIII, (2021).
10. Johansen J. and Sørensen N.N.: Aerofoil Characteristics from 3D CFD Rotor Computations,
Wind Energy, 7,4, (2004). https://doi.org/10.1002/we.127
 Transport and Climate

The simulations in the category “Transport and Climate” have consumed a total of
approximately 50 million core-hours over the past granting period. Four projects have
utilized the system Hawk (HLRS), and nine were hosted on the now retired system
ForHLR II (SCC). The majority (67%) of computational resources in this category
was spent on ForHLR II.
In the project “GLOMIR+” by Schneider et al. the platform ForHLR II was
leveraged for the analysis of spectra measured via satellite-bound infrared atmospheric
sounding interferometry. The particular focus of the post-processing campaign is to
identify atmospheric trace gases such as H2O, HDO, CH4, N2O and HNO3. The
results of these retrievals constitute highly sought-after data for a variety of other
studies in earth system science.
The authors of the second project “MIPAS” (Kiefer et al.) have performed similar
analysis of Michelson interferometer data taken aboard the satellite Envisat over one
decade. Here the distributions of more than 30 species of trace gases are analyzed,
constituting one of the largest available databases for the composition of the middle
atmosphere. The post-processing operations carried out successfully on Hawk are
not only computationally intensive, but they also feature an exceptional load on I/O.
The third project by Bauer et al. (“WRFSCALE”) uses the large-eddy technique for
the description of atmospheric turbulence in the context of the Weather and Research
Forecasting (WRF) model. The project is aimed at understanding the interaction
between land-surface processes and the dynamics of the atmosphere, which is studied
for a mid-western US region covered by the Land Atmosphere Feedback Experiment
(LAFE). Their setup exhibits good parallel scaling on Hawk making use of the full
number of available compute cores. Overall, the project “WRFSCALE” convincingly
showcases how the nested large-eddy approach down to a grid resolution of 20 meters
can contribute to further our understanding of atmospheric processes.

Institute for Hydromechanics, Markus Uhlmann

Karlsruhe Institute of Technology,
Kaiserstr. 12,
76131 Karlsruhe,
Germany,
e-mail: [email protected]

381
The HPC project GLOMIR+ (GLObal MUSICA
IASI Retrievals - plus)

Matthias Schneider, Benjamin Ertl, Christopher Diekmann and Farahnaz Khosrawi

1 Introduction

The HPC project GLOMIR+ (GLObal MUSICA IASI Retrievals - plus) has the
objective to generate a global, multi-year dataset for various atmospheric trace gases
based on thermal nadir spectra measured by the sensor IASI (Infrared Atmospheric
Sounding Interferometer) aboard the EUMETSAT’s polar orbiting satellites Metop-A,
-B, and -C. The Metop/IASI mission combines high spectral resolution and low
measurement noise with high horizontal and temporal coverage (12 km ground
pixel, global, twice daily). Figure 1 shows the measurement geometry of IASI. The
polar orbit in combination with a wide across track scanning angle ensures a global
coverage twice per day. The IASI measurements started in 2007 and in the context of
an already approved successor mission (launch of IASI-NG on Metop-SG satellites is
currently scheduled for early 2024) observations are guaranteed until the late 2030s.
This unprecedented climate and weather research potential is due to the strong support
by the European meteorological organizations. However, it comes along with the
challenge of the large amount of spectra that have to be processed. The currently three
orbiting IASI instruments provide about 3.8 million measured spectra each 24 h. For
their analysis, the application of high performance computing is indispensable. For
the global retrievals the MUSICA PROFFIT-nadir retrieval code is used. It has been
developed recently within the ERC (European Research Council) project MUSICA
(MUlti-platform remote Sensing of Isotopologues for investigating the Cycle of

Matthias Schneider, Benjamin Ertl, Christopher Diekmann and Farahnaz Khosrawi

Institute of Meteorology and Climate Research – Atmospheric Trace Gases and Remote Sensing
(IMK-ASF), Karlsruhe Institute of Technology (KIT), Hermann-von-Helmholtz Platz 1, Eggenstein-
Leopoldshafen, 76344, Germany,
e-mail: [email protected],[email protected],
[email protected], [email protected]
Benjamin Ertl
Steinbuch Centre for Computing (SCC), Karlsruhe Institute of Technology (KIT), Hermann-von-
Helmholtz Platz 1, Eggenstein-Leopoldshafen, 76344, Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 383
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_23
384 M. Schneider, B. Ertl, C. Diekmann and F. Khosrawi

Atmospheric water, [13]. The target of the MUSICA processing is the atmospheric
composition of H2 O, HDO/H2 O ratio, CH4 , N2 O and HNO3 (the HDO/H2 O ratio
retrieval is truly innovative). The processing has been continuously improved during
MUSICA follow-up projects and the current MUSICA IASI data product and the
corresponding processing chain is presented in [15] and shown as a flowchart in
Figure 2. The processing relies on six different components:
1. Preprocessing stage: the merging of the EUMETSAT IASI Level 1c data (L1c,
the calibrated IASI spectra) and Level 2 products (L2, the first guess profiles of
atmospheric humidity and temperature).
2. PROFFIT-nadir retrieval: this algorithm inverts the spectral measurements and
generates the atmospheric composition profiles [11–13].
3. Output generation: tool for generating storage efficient MUSICA IASI netcdf
output files [23] in agreement to the CF (Climate and Forecast, http://cfco
nventions.org/) metadata conventions. The high quality of the products is
demonstrated by several validation studies, e.g. [1, 10, 24].
4. A posteriori data reusage: algorithms ensuring tailored data reusage, e.g. for the
generation of {H2O,𝛿D} pairs [3] or for the synergetic use of MUSICA IASI
and TROPOMI methane products [16].
5. Retrieval simulator: the retrieval simulator is needed for high quality observation-
model comparison studies, e.g. [14].
6. Re-gridded L3 products: the level 3 data generation tool is still not available,
but highly asked for by the scientific community. It will enable the generation
and distributing of the data products on a regular horizontal and time step grid
tailored to the needs of the individual data user.

Fig. 1: Depiction of a polar orbiting satellite (left, source: http://tornado.

sfsu.edu/) and indication of the IASI observation geometry (right, source:
https://www.eumetsat.int).
The HPC project GLOMIR+ 385

In Sec. 2 we give details on the executed computations, the parallelism and scaling
of our computations, and the used HPC resources. The second part of this paper
presents the scientific results we achieved based on the HPC calculations (Sec. 3).

Fig. 2: Flow chart with the different components (blue symbols) of the MUSICA
IASI processing chain.

2 Data processing

2.1 Computations, parallelism, and scaling

Using a single processing unit the retrieval of one observation takes about 30 seconds.
We make retrievals for all spectra measured under cloud-free conditions, which are
typically 25000 observations per orbit. Currently three different IASI instruments
are in operation and provide spectra for 42 orbits per 24 h. This results in about
600000 observations per day meaning a processing time of 1.8 million seconds (208
days) when using a single processing unit. High performance computing is thus
indispensable for processing the satellite data. In the interest of optimum utilization of
the HPC resources, our calculations are distributed across the different processors, so
that they can operate in parallel (data parallelism). For this purpose, we use the Python
Multiprocessing package and the Python Luigi package. The Multiprocessing package
supports spawning processes using an API similar to the threading module. It allows
the full leverage of processors on a machine (https://docs.python.org/3.6/li
brary/multiprocessing.html). The Python Luigi package (2.7, 3.6, 3.7 tested)
helps to build complex pipelines of batch jobs. It handles dependency resolution,
386 M. Schneider, B. Ertl, C. Diekmann and F. Khosrawi

workflow management, visualization, handling failures, command line integration, and

much more (https://luigi.readthedocs.io/en/stable/index.html). In
the following, we give details on the parallelism and scaling of our three computational
most intensive applications: (1) the PROFFIT-nadir retrieval algorithm, (2) the CF
data generation and output compression, and (3) the processing in context of the a
posteriori data reusage.

Fig. 3: Charts demonstrating parallelism and scaling of the MUSICA IASI retrieval
processing (left panel) and the processing for a posteriori data reusage (right panel).

1. The main MUSICA IASI retrieval Python program uses a thread pool to distribute
the input data (individual spectra observations and corresponding auxiliary data
like temperature profiles) to the available number of worker threads, which can
execute the Fortran retrieval code in parallel (see left chart on Fig. 3). The thread
pool schedules as many worker threads as possible (40 threads on a single HoreKa
computing node, i.e. 2 threads per processor core). The available computing
performance is fully explored until the retrievals of all the observations of a single
orbit are completed. The retrieval of one observation takes about 30 seconds.
One orbit contains typically 25 observations. Because 40 observations can be
processed in parallel, we can process a whole orbit in typically less than 15 h on
a single node.
2. The compression and generation of CF conform netcdf MUSICA IASI data files
are also scaling with the number of processors that can be used on a given compute
node [23]. For this purpose, our Python compression code utilizes the Python
Luigi package. The centralized Luigi scheduler coordinates the operations that
can be executed in parallel. The tasks are distributed to the available workers and
it is ensured that two instances of the same task are not running simultaneously.
This procedure ensures that all available processors are utilized to capacity.
3. The processing in context of the a posteriori isotopologue data reusage distributes
the input data on the available computing resources as depicted on the right side
of Fig. 3. For each observation, the post-processing (Python program) reads
in a subset of the retrieval output variables and takes a fraction of seconds for
The HPC project GLOMIR+ 387

processing. Therefore, the memory of one computing node offers the capacity
for post-processing all orbit files for one day simultaneously by distributing them
to the available processors. Using this parallelization method, the processing of
one day takes about 40 min. Thus, one single computing node can be used to
do the a posteriori processing of one month of data (≈ 30 Mio. observations) in
about 20 h.

2.2 Summary of used computation resources

The computationally most intensive tasks were as follows:

1. Execution of the retrievals using the algorithm PROFFIT-nadir: we processed
IASI (Infrared Atmospheric Sounding Interferometer) spectra measured aboard
the satellites Metop-A, -B, and -C between June 2019 and December 2020. The
retrievals for 2019, the generation of the first half of 2020 have been performed
on ForHLR II. The second half of 2020 has been processed on HoreKa.
2. Generation of the data output files (including data compression): we made an
output data compression and generated the CF (Climate and Forecast metadata
standard) conform netcdf files for the whole period of MUSICA IASI retrieval
data (for more than six years between October 2014 and December 2020). For
the October 2014 to June 2019 period, this processing has been made on ForHLR
II. These data have been published with DOIs [15, 17, 18]. The output data
generation and compression for the period between July 2019 and December
2020 has been made on HoreKa.
3. Isotopologue data post-processing: we made the MUSICA IASI isotopologue
post-processing for the October 2014 to June 2019 period on ForHLR II. These
data have been published with DOIs [3, 5]. The post-processing for the July 2019
to December 2020 period has been realized on HoreKa.
4. Forward operator and LAGRANTO backward trajectory calculations: for
the West African Monsoon period in 2016 and 2018 we have made forward
operator calculation on ForHLR II. These calculations allow simulating the
MUSICA IASI observations of the atmospheric state as modelled by COSMO-
iso. For the same period and area we performed LAGRANTO back trajectory
calculations using the fields as modeled by COSMO-iso. These calculations have
also been made on ForHLR II. The COSMO-iso model data were provided by
the project partner (ETH Zürich) of the MOTIV (MOisture Transport pathways
and Isotopologues in water Vapour) project. In Table 1 we give an overview of
the computing resources used for the different tasks and of the period when the
processing has been made.
388 M. Schneider, B. Ertl, C. Diekmann and F. Khosrawi

Table 1: Summary of consumed computing resources. We calculate here with 20

cores on one node and two CPUs per core (i.e. 40 CPUs per node)

Task Period Consumed CPUh

(1) Execution of the retrievals 03/2020–02/2021 10 millions

(2) Generation of the data output 10/2020 and 02/2021 0.75 million
(3) Isotopologue data post-processing 10/2020–12/2020 0.75 million
(4) Forward operator and trajectory 06/2020–12/2020 0.5 million

Total: 12 million

3 Scientific results

The GLOMIR+ HPC activities contributed to different international research projects.

This results in a large number of scientific publications with the written acknowl-
edgment “Retrieval calculation for this work were performed on the supercomputer
ForHLR funded by the Ministry of Science, Research and the Arts Baden-Württemberg
and by the German Federal Ministry of Education and Research”. In the following,
we give some brief examples of these contributions (order by scientific field), and list
the related scientific publications.

3.1 Data set dissemination and presentation

During the GLOMIR+ project and the related precursor HPC projects we have been
able to perform MUSICA IASI retrievals for the whole period between October 2014
and December 2020. In the beginning of 2021 we published the data in the research
data repository RADAR4KIT (https://radar.kit.edu/, [4, 17, 18] and describe
the data products in detail in two publications in the Copernicus journal “Earth System
Science Data” (https://www.earth-system-science-data.net/, [3, 15]. The
full standard data product [17] contains the trace gas profiles of H2 O, 𝛿D (the stan-
dardised HDO/H2 O ratio), CH4 , N2 O and HNO3 as well as atmospheric and surface
temperatures. The full standard data files also contain auxiliary data like uncertainty
covariances, information of constraint and representativeness (e.g. averaging kernels),
which is essential for a comprehensive re-use of the data. The file with the extended
data product [18] contains additional information about the response of the data on
different error sources (e.g. unrecognized clouds, uncertainties in surface emissivity),
but only considers 74 exemplary observations for a polar mid-latitudinal and tropical
site. The MUSICA IASI full retrieval products (standard and extended output files)
are described in detail in the Earth System Science Data publication of [15]. The
The HPC project GLOMIR+ 389

{H2O,𝛿D} pair product [5] contains H2 O and 𝛿D data generated by a posteriori

processing method. Such a posteriori processing is essential for a successful scientific
usage of the data. Details on the data product and the a posteriori processing are
given in the Earth System Science Data publication of [3]. Table 4 gives an overview
on the related publications and Fig. 4 gives an example of the MUSICA IASI data
showing a continuous time series (Oct. 2014 – Dec. 2019) of the different trace gas
products over Karlsruhe (49.0◦ N, 8.5◦ E).

Table 2: Publications in the field of MUSICA IASI data set dissemination and
presentation

Description of publication Reference DOI

Presentation of full data product Schneider et al. (2021a) 10.5194/essd-2021-75

Presentation of {H2 O, 𝛿D} 10.5194/essd-2021-87
pair data product
Dissemination of full standard Schneider et al. (2021c) 10.35097/408
output data product
Dissemination of full extended Schneider et al. (2021d) 10.35097/412
output data product
Dissemination of {H2 O, 𝛿D} pair Diekmann et al. (2021c) 10.35097/415
data product

3.2 Research of the atmospheric water cycle

The data generated during GLOMIR+ have been essential for achieving the scientific
goals of the two international projects MOTIV (MOisture Transport and Isotopologues
in water Vapour) and TEDDY (TEsting isotopologues as Diabatic heating proxy for
atmospheric Data analYses). Table 3 gives an overview on the related publications.
In [2] the MUSICA IASI {H2O,𝛿D} pair data are used together with COSMOiso
model simulations to disentangle the different moisture transport pathways to the
subtropical North Atlantic middle troposphere. We have been able to demonstrate
that the {H2O,𝛿D} pair data allow identifying the different transport pathways.
This gives new possibilities for investigating the related processes and its possible
response to global warming. Similarly, [4] has used the measured and simulated
{H2O, 𝛿D} pair data together with Lagrangian trajectory calculations for research
on the West African monsoon. We have found distinct {H2O,𝛿D} pair distributions
390 M. Schneider, B. Ertl, C. Diekmann and F. Khosrawi

Fig. 4: October 2014 to December 2019 time series of MUSICA IASI data products
retrieved over Karlsruhe (49.0◦ N, 8.5◦ E): H2 O, 𝛿D (the standardised HDO/H2 O
ratio), CH4 , N2 O and HNO3 .

for air masses that experienced a convective process. These results indicate that the
isotopologue data can be used for quantifying the importance of convective transport
for free tropospheric humidity. In [7] we present a new method for clustering pair
distributions that allows for constraining with respect to the correlation of the pairs
(CoExDBSCAN: Correlation Extended DBSCAN). We apply this method for pattern
recognitions with the {H2O,𝛿D} pair data. [8] shows that CoExDBSCAN can very
efficiently identify temporal sequences that are characterised by a specific correlation
of data pairs. In [9] we use CoExDBSCAN for clustering the temporal sequences
of {H2O,𝛿D} pairs along an atmospheric Lagrangian trajectory. This allows for an
automated recognition of moisture processes along a Lagrangian trajectory.
In [20] we show that assimilating the MUSICA IASI water vapour isotopologue
data in addition to traditional atmospheric observations (humidity, temperature, wind,
etc.) can significantly improve the knowledge about the atmospheric state, which in
turn allow for improved weather predictions (see example of Fig. 5). The MUSICA
IASI water vapour isotopologue data have also been of central importance for the
PhD thesis of Christopher J. Diekmann [6]. He further developed the MUSICA
IASI water vapour isotopologue processing and interpreted the MUSICA IASI
{H2O, 𝛿D} pair data with the help of COSMOiso model simulations and also by
using complementary observations from other satellite sensors. Fig. 6 gives an
example of the latter. There the IMERGE (Integrated Multi satellitE Retrievals for
GPM, https://gpm.nasa.gov/data/imerg) satellite rain data are used for
The HPC project GLOMIR+ 391

Table 3: Publications in the field of atmospheric water cycle research.

Description of publication Reference DOI (or URL)

Investigating the North Atlantic Dahinden et al. (2021) 10.5194/acp-2021-269

moisture transport pathways
Water isotope signals and Diekmann et al. (2021b) 10.1029/2021JD034895
the West African Monsoon
Method for clustering {H2 O, 𝛿D} Ertl et al. (2020) 10.5220/00101312010540115
data (CoExDBSCAN)
Time point clustering with Ertl et al. (2021b) caiac.pubpub.org/pub/a3py333z
CoExDBSCAN
Time point clustering of Ertl et al. (2021a) 10.1007/978-3-030-77961-0_23
{H2 O, 𝛿D} pair data
Assessment of isotopologue Toride et al. (2021) 10.1029/2020GL091698
data assimilation
PhD thesis Diekmann (2021) doi:10.5445/IR/1000134744

Fig. 5: Improvements of forecast skills for 𝛿D, humidity, wind, temperature, geopo-
tential height, and precipitation (a-f) by assimilating MUSICA IASI data in addition
to the traditional atmospheric observations (green: additional assimilation of H2 O;
yellow: of 𝛿D; violet: of {H2 O, 𝛿D} pairs). Figure has been taken from [20].
392 M. Schneider, B. Ertl, C. Diekmann and F. Khosrawi

classifying the IASI observations in affected by rain and not affected by rain. For air
that has been affected by rain a distinct {H2O,𝛿D} pair distribution can be clearly
observed. Currently, we also discuss a MUSICA IASI water isotopologue data usage
in support of the project EUREC4A (http://eurec4a.eu/). EUREC4A aims at an
advanced understanding of the interplay between cloud processes and the large-scale
atmospheric state and circulation important for weather and climate prediction.

Fig. 6: Left: H2O, 𝛿D pair distributions as observed over the Sahel zone by the
MUSICA IASI data for locations with a previous rain event (post-rain, red contours)
and no rain during 48 h before the observation (non-rain, blue contours). Right:
Location and frequency of occurrence of non-rain (a) and rain events (b). For the rain
event identification the IMERG satellite data products are used. Figures have been
from [6].

3.3 Research on atmospheric methane emissions

The MUSICA IASI methane (CH4 ) data product is very useful for investigating local
methane sources. Table 4 gives an overview of the related publications. With this
product we have also supported the ESA projects FRM4GHG (https://frm4ghg.
aeronomie.be/) and COCCON-PROCEEDS (https://www.imk-asf.kit.edu
/english/3225.php). In [16] we present a method for a synergetic combination of
the MUSICA IASI CH4 profile data products with the CH4 total column averaged
(XCH4) data product of TROPOMI (TROPOspheric Monitoring Instrument). We show
that the combination allows for detecting lower tropospheric column averaged CH4
amounts (TXCH4) independently from the upper tropospheric/lower stratospheric
amounts (this is not possible by one of the data sets alone). In [22] we use this
synergetic combination method for investigating surface near CH4 anomalies due to
waste disposal site emissions nearby Madrid (see Fig. 7).
The HPC project GLOMIR+ 393

Table 4: Publications in the field of atmospheric methane emissions.

Description of publication Reference DOI

Synergetic combination of Schneider et al. (2021b) 10.5194/amt-2021-31

MUSICA IASI and TROPOMI CH4
Quantification of local CH4 Tu et al (2021b) 10.5194/acp-2021-437
emissions from ground and space

Fig. 7: Methane anomalies as observed in the tropospheric column averaged product

(generated by the synergetic combination of TROPOMI and MUSICA IASI CH4
data) ordered by predominant wind directions: all winds (left); south-west winds
(centre); north-east winds (right). Similar to Fig. 8 of [22].

3.4 Validation of satellite data

Comparison of different remote sensing data products is essential for understanding

the data quality. We have participated with our MUSICA IASI H2 O and 𝛿D data
products in two inter-comparison studies (see Table 5). The first study resulted in the
publication of [21]. There water vapour total column data (XH2O) obtained from
IASI by our MUSICA processor and data generated from TROPOMI measurements
have been successfully compared with different ground-based XH2O remote sensing
data sets. Fig. 8 gives an example of the good agreement between the MUSICA IASI
XH2O data and the XH2O reference measured within the COCCON (Collaborating
Carbon Column Observation Network). Furthermore, in the context of the ESA
project S5P+I H2O-Iso our MUSICA IASI 𝛿D data product has been used to estimate
the precision of a newly developed TROPOMI column integrated 𝛿D data product
[19].
394 M. Schneider, B. Ertl, C. Diekmann and F. Khosrawi

Table 5: Publications in the field of satellite data validation

Description of publication Reference DOI (or URL)

Validation of total column Tu et al. (2021a) 10.5194/amt-14-1993-2021

water vapour
Validation of column Schneider et al. (2021e) s5pinnovationh2o-iso.le.ac_uk/
integrated water vapour wp-content/uploads/2021/02/
isotopologue ratios S5p-I-VR-Version 1.1.pdf

Fig. 8: Scatter plot of the COCCON XH2O compared with coincident MUSICA IASI
retrievals at the Kiruna site (a) and the Sodankylä site (b). Figure taken from [21].

References

1. Borger, C., Schneider, M., Ertl, B., Hase, F., García, O. E., Sommer, M., Höpfner, M., Tjemkes,
S. A., and Calbet, X.: Evaluation of MUSICA IASI tropospheric water vapour profiles using
theoretical error assessments and comparisons to GRUAN Vaisala RS92 measurements, Atmos.
Meas. Tech., 11, 4981–5006, doi:10.5194/amt-11-4981-2018, 2018.
2. Dahinden, F., Aemisegger, F., Wernli, H., Schneider, M., Diekmann, C. J., Ertl, B., Knippertz, P.,
Werner, M., and Pfahl, S.: Disentangling different moisture transport pathways over the eastern
subtropical North Atlantic using multi-platform isotope observations and high-resolution numer-
ical modelling, Atmos. Chem. Phys. Discuss., 21, 16319–16347, https://doi.org/10.5194/acp-
21-16319-2021, 2021.
3. Diekmann, C. J., Schneider, M., Ertl, B., Hase, F., García, O., Khosrawi, F., Sepúlveda, E.,
Knippertz, P., and Braesicke, P.: The global and multi-annual MUSICA IASI {H2O, 𝛿D}
pair dataset, Earth Syst. Sci. Data, 13, 5273–5292, https://doi.org/10.5194/essd-13-5273-
2021,2021a.
The HPC project GLOMIR+ 395

4. Diekmann, C. J., Schneider, M., Knippertz, P., de Vries, A. J., Pfahl, S., Aemisegger, F.,
Dahinden, F., Ertl, B., Khosrawi, F., Wernli, H., Braesicke, P.: A Lagrangian perspective on
stable water isotopes during the West African Monsoon, J. Geophys. Res.: Atmospheres, 126,
e2021JD034895. https://doi.org/10.1029/2021JD034895, 2021b.
5. Diekmann, C. J., Schneider, M., Ertl, B.: MUSICA IASI water isotopologue pair product (a
posteriori processing version 2), Institute of Meteorology and Climate Research, Atmospheric
Trace Gases and Remote Sensing (IMK-ASF), Karlsruhe Institute of Technology (KIT). DOI:
10.35097/415, 2021c.
6. Diekmann, C. J., Analysis of stable water isotopes in tropospheric moisture during the West
African Monsoon, PhD thesis, successfully completed at the Faculty of Physics, Karlsruhe
Institute of Technology (KIT), doi:10.5445/IR/1000134744, 2021.
7. Ertl, B., Meyer, J., Schneider, M., and Streit, A.: CoExDBSCAN: Density-based Clustering with
Constrained Expansion, Proceedings of the 12th International Joint Conference on Knowledge
Discovery, Knowledge Engineering and Knowledge Management - Volume 1: KDIR, pages
104–115, INSTICC, SciTePress, ISBN 978-989-758-474-9, doi:10.5220/0010131201040115,
2020.
8. Ertl, B., Meyer, J., Schneider, M., Diekmann, C. and Streit, A.: A Semi-Supervised Approach
for Trajectory Segmentation to Identify Different Moisture Processes in the Atmosphere,
Computational Science – ICCS 2021, Springer International Publishing, https://doi.org/
10.1007/978-3-030-77961-0_23, 2021a.
9. Ertl, B., Meyer, J., Schneider, M., and Streit, A.: Semi-Supervised Time Point Clustering
for Multivariate Time Series, Advances in Artificial Intelligence, accepted for publication,
https://caiac.pubpub.org/pub/a3py333z, 2021b.
10. García, O. E., Schneider, M., Ertl, B., Sepúlveda, E., Borger, C., Diekmann, C., Wiegele,
A., Hase, F., Barthlott, S., Blumenstock, T., Raffalski, U., Gómez-Peláez, A., Steinbacher,
M., Ries, L. and de Frutos, A. M.: The MUSICA IASI CH4 and N2 O products and their
comparison to HIPPO, GAW and NDACC FTIR references, Atmos. Meas. Tech., 11, 4171-4215,
doi:10.5194/amt-11-4171-2018, 2018.
11. Hase, F., Hannigan, J.W., Coffey, M. T., Goldman, A., Höpfner, M., Jones, N. B., Rinsland, C.
P., and Wood, S.: Intercomparison of retrieval codes used for the analysis of high-resolution, J.
Quant. Spectrosc. Ra., 87, 25–52, 2004.
12. Schneider, M. and Hase, F.: Optimal estimation of tropospheric H2 O and 𝛿D with IASI/METOP,
Atmos. Chem. Phys., 11, 11 207–11 220, https://doi.org/10.5194/acp-11-11207-2011, 2011.
13. Schneider, M., Wiegele, A., Barthlott, S., González, Y., Christner, E., Dyroff, C., García, O. E.,
Hase, F., Blumenstock, T., Sepúlveda, E., Mengistu Tsidu, G., Takele Kenea, S., Rodríguez,
S., and Andrey, J.: Accomplishments of the MUSICA project to provide accurate, long-term,
global and high-resolution observations of tropospheric {H2O, 𝛿D} pairs – a review, Atmos.
Meas. Tech., 9, 2845-2875, doi:10.5194/amt-9-2845-2016, 2016.
14. Schneider, M., Borger, C., Wiegele, A., Hase, F., García, O. E., Sepúlveda, E., and Werner, M.:
MUSICA MetOp/IASI H2O, 𝛿D pair retrieval simulations for validating tropospheric moisture
pathways in atmospheric models, Atmos. Meas. Tech., 10, 507-525, doi:10.5194/amt-10-507-
2017, 2017.
15. Schneider, M., Ertl, B., Diekmann, C. J., Khosrawi, F., Weber, A., Hase, F., Höpfner, M.,
García, O. E., Sepúlveda, E., and Kinnison, D.: Design and description of the MUSICA IASI
full retrieval product, Earth Syst. Sci. Data Discuss. [preprint], https://doi.org/10.5194/essd-
2021-75, in review, 2021a.
16. Schneider, M., Ertl, B., Diekmann, C. J., Khosrawi, F., Röhling, A. N., Hase, F., Dubravica,
D., García, O. E., Sepúlveda, E., Borsdorff, T., Landgraf, J., Lorente, A., Chen, H., Kivi, R.,
Laemmel, T., Ramonet, M., Crevoisier, C., Pernin, J., Steinbacher, M., Meinhardt, F., Deutscher,
N. M., Griffith, D. W. T., Velazco, V. A., and Pollard, D. F.: Synergetic use of IASI and
TROPOMI space borne sensors for generating a tropospheric methane profile product, Atmos.
Meas. Tech. Discuss. [preprint], https://doi.org/10.5194/amt-2021-31, in review, 2021b.
396 M. Schneider, B. Ertl, C. Diekmann and F. Khosrawi

17. Schneider, M., Ertl, B., Diekmann, C. J.,: MUSICA IASI full retrieval product standard output
(processing version 3.2.1), Institute of Meteorology and Climate Reasearch, Atmospheric
Trace Gases and Remote Sensing (IMK-ASF), Karlsruhe Institute of Technology (KIT), DOI:
10.35097/408, 2021c.
18. Schneider, M., Ertl, B., Diekmann, C. J.: MUSICA IASI full retrieval product extended
output (processing version 3.2.1), Institute of Meteorology and Climate Research, Atmospheric
Trace Gases and Remote Sensing (IMK-ASF), Karlsruhe Institute of Technology (KIT), DOI:
10.35097/412, 2021d.
19. Schneider, M., Röhling, A. N., Khosrawi, F., Diekmann, C. J., Trent, T., Bösch, H.,
and Sodemann, H.: Validation Report (VR), Version 1.1, 2020-11-11, ESA project:
S5p+I H2O-ISO, https://s5pinnovationh2o-iso.le.ac.uk/wp-content/uploads/2021/02/S5p-I-VR-
Version1.1.pdf, 2021e.
20. Toride, K., Yoshimura, K., Tada, M., Diekmann, C., Ertl., B., Khosrawi, F., and
Schneider, M.: Potential of mid-tropospheric water vapor isotopes to improve large-
scale circulation and weather predictability, Geophys. Res. Lett., 48, e2020GL091 698,
https://doi.org/10.1029/2020GL091698, 2021.
21. Tu, Q., Hase, F., Blumenstock, T., Schneider, M., Schneider, A., Kivi, R., Heikkinen, P.,
Ertl, B., Diekmann, C., Khosrawi, F., Sommer, M., Borsdorff, T., and Raffalski, U.: In-
tercomparison of arctic XH2O observations from three ground-based Fourier transform
infrared networks and application for satellite validation, Atmos. Meas. Tech., 14, 1993–2011,
https://doi.org/10.5194/amt-14-1993-2021, 2021a.
22. Tu, Q., Hase, F., Schneider, M., García, O., Blumenstock, T., Borsdorff, T., Frey, M., Khosrawi,
F., Lorente, A., Alberti, C., Bustos, J. J., Butz, A., Carreño, V., Cuevas, E., Curcoll, R.,
Diekmann, C. J., Dubravica, D., Ertl, B., Estruch, C., León-Luis, S. F., Marrero, C., Morgui,
J.-A., Ramos, R., Scharun, C., Schneider, C., Sepúlveda, E., Toledano, C., and Torres, C.:
Quantification of CH4 emissions from waste disposal sites near the city of Madrid using
ground- and space-based observations of COCCON, TROPOMI and IASI, Atmos. Chem. Phys.
Discuss. [preprint], https://doi.org/10.5194/acp-2021-437, in review, 2021b.
23. Weber, A.: Storage-efficient analysis of spatio-temporal data with application to climate
research, Master Thesis, doi:10.5281/zenodo.3360021, https://zenodo.org/recor
d/3360021#.Xl4pZEoo-bg, 2019.
24. Wiegele, A., Schneider, M., Hase, F., Barthlott, S., García, O. E., Sepúlveda, E., González, Y.,
Blumenstock, T., Raffalski, U., Gisi, M., and Kohlhepp, R.: The MUSICA MetOp/IASI H2 O
and 𝛿D products: characterisation and long-term comparison to NDACC/FTIR data, Atmos.
Meas. Tech., 7, 2719-2732, doi:10.5194/amt-7-2719-2014, 2014.
Global long-term MIPAS data processing

Michael Kiefer, Bernd Funke, Maya García-Comas, Udo Grabowski, Andrea Linden,
Axel Murk and Gerald E. Nedoluha

Abstract The aim of this project is to perform a level 2 (L2) processing of the
version 8 global infrared spectra data set (V8 L1b data) of the Earth’s atmosphere,
measured in limb-viewing geometry by the space-borne instrument MIPAS (Michelson
Interferometer for Passive Atmospheric Sounding) operated by the European Space
Agency (ESA) from 2002–2012. MIPAS was a Fourier transform mid-infrared limb
scanning high resolution spectrometer which allowed for simultaneous measurements
of more than 30 atmospheric trace species related to atmospheric chemistry and
global change. At the Institute for Meteorology and Climate Research (IMK) MIPAS
spectra are used for retrieval of vertically resolved profiles of abundances of trace
species of the atmosphere. The trace gas distributions are used for the assessment of
e.g. stratospheric ozone chemistry, stratospheric cloud physics and heterogeneous
chemistry, stratospheric exchange processed with troposphere and mesosphere,
intercontinental transport of pollutants in the upper troposphere, effects of solar proton
events on stratospheric chemistry, mesospheric dynamics, atmospheric coupling,
thermospheric temperature, and validation of climate-chemistry models. In the
reporting period 2020/2021 MIPAS data processing was performed on the XC40
(Hazel Hen) and on the HPE Apollo (Hawk) supercomputers. The latter machine
was mainly used to perform computationally expensive non-local thermodynamic
equilibrium (NLTE) calculations. In the test phase/configuration of the HPE Apollo
computer small obstacles had to be overcome, however, our approach which has

M. Kiefer, U. Grabowski and A. Linden

Institut für Meteorologie und Klimaforschung, Karlsruhe Institute of Technology, Karlsruhe,
Germany, e-mail: [email protected]
M. García-Comas and B. Funke
Instituto de Astrofísica de Andalucía, Granada, Spain
A. Murk
Institute of Applied Physics, University of Bern, Bern, Switzerland
G. Nedoluha
Remote Sensing Division, Naval Research Laboratory, Washington DC, USA

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 397
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_24
398 M. Kiefer et al.

been proven to successfully and efficiently work on the XC 40, also initially worked
for the HPE Apollo and showed good performance. A configuration change with
respect to the /tmp-filesystem of the compute nodes, which was heavily used by
our processing software for efficiency reasons, unfortunately altered this state. We
were forced to use the temporary workspace filesystem instead, which implies a
performance degradation. Most of the processing work presented here was done
in close collaboration between the Instituto de Astrofísica de Andalucía (IAA) in
Granada, Spain, and KIT-IMK. The focus of the current work is the processing of
atmospheric trace gases which require computationally expensive NLTE calculations,
i.e., H2 O, CO, NO, NO2 . The progress in the quality of the V8 water vapour data
product is demonstrated. First presentations of V8 L2 data of CO, NO, and NO2 are
shown. The vertical profiles of version V8 data processed at IMK are available to
external data users on http://www.imk-asf.kit.edu/english/308.php.

1 The MIPAS/Envisat mission

The Michelson Interferometer for Passive Atmospheric Sounding (MIPAS) was part
of the core-payload of the sun-synchronous polar orbiting satellite Envisat, operated
by the European Space Agency (ESA). Envisat orbits the Earth 14.4 times per day.
MIPAS measured in the mid-infrared spectral region 4.15–14.6 𝜇m with a design
spectral resolution of 0.035 cm−1 . It measured thermal emission interferograms of
the Earth’s limb, centered around a latitude/longitude point (called limb scan or
geolocation), whereby variation of the limb tangent altitude provided altitude-resolved
information [1, 2]. Mid-infrared limb observations are essentially independent of the
illumination status of the atmosphere. Hence, e.g., the atmosphere at polar winter
conditions can be observed without problems.
The MIPAS mission improved, and still is improving, the understanding of
the composition and dynamics of the Earth’s atmosphere by measurement of 4D
distributions of more than 30 trace species relevant to atmospheric chemistry and
climate change. Operation of satellite and instrument and level-1b (L1b) data
processing have been performed by ESA.
MIPAS was operational in its original, full spectral resolution (FR) specification
from June 2002 to March 2004. Due to an instrument problem the measurements
could not be resumed before the beginning of 2005. Since then a reduced spectral
resolution (RR) of 40% of the FR value was used, but at the same time a finer vertical
scan grid was implemented.
Just weeks after celebrating its tenth year in orbit, communication with the Envisat
satellite was suddenly lost on 8th April 2012, which led to the end of the mission
being declared by ESA.
Global long-term MIPAS data processing 399

2 L1b data delivery and L2 data generation

2.1 L1b data processing and delivery by ESA

The L1b data processing, i.e. essentially Fourier transformation of the measured
interferograms, phase correction, and calibration, were performed by the European
Space Agency. A description of this processing step can be found in [3]. During
the MIPAS mission the acquired knowledge about deficiencies in the L1b data led
to the release of several data versions with respective improvements. Among these
improvements are better values for the line of sight (LOS), i.e. the knowledge on the
exact altitude at which the instrument points.
A major problem of older data versions was caused by an instrumental drift in
time due to the ageing of the MIPAS detectors, in conjunction with their non-linear
response curve. This instrumental drift in turn leads to a drift in the atmospheric
temperature and constituents calculated from the spectra. Therefore for the V7
L1b data product a new calibration scheme was devised to compensate for this
effect of the detector ageing (see [4, 5]). However, it soon turned out that there was
an inconsistency between the FR and RR measurement periods. Hence a further
correction was implemented for V8 L1b data. This V8 L1b data is the final release,
since the financial and administrative efforts for the MIPAS mission were suspended
by ESA in 2019.
It should be noted that initially our HPC-project was proposed to last from April
2017 to December 2019 and it was planned to do the L2 processing with the V7 data
set. However, as soon as it became clear that there would be one further L1b data
version with clear improvements over V7, the decision was to wait for the new V8
L1b data set. The delivery of the V8 L1b data was expected for mid 2017. However,
although the processing of the V8 L1b data had been finished by end of January 2018,
there were some investigations and corrective actions necessary by ESA. The data
delivery to the end users did not start before January 2019.

2.2 L2 Data generation

Retrieval of abundances of atmospheric constituents from infrared spectra requires the

inverse solution of the radiative transfer equation. Measured radiances are compared
to radiative transfer calculations and residuals are minimized in a least squares sense
by adjustment of the abundances of constituents which are fed into a forward model.
The Retrieval Control Program (RCP), together with the forward model, is the core
of the processing chain.
The forward model, the Karlsruhe Optimized and Precise Radiative Transfer
Algorithm (KOPRA) [6] is a computationally optimized line-by-line model which
simulates radiative transfer through the Earth’s atmosphere under consideration of
all relevant physics: the spectral transitions of all involved molecules, atmospheric
400 M. Kiefer et al.

refraction, line-coupling and non-local thermodynamic equilibrium (NLTE). Along

with the spectral radiances, KOPRA also provides the Jacobian matrices, i.e. the
sensitivities of the spectral radiances to changes in the atmospheric state parameters.
Atmospheric state parameters are retrieved by constrained multi-parameter non-
linear least-squares fitting of simulated to measured spectral radiances. The inversion
is regularized with a Tikhonov-type constraint which minimizes the first order finite
differences of adjacent profile values [7, 8].
Instead of performing a simultaneous retrieval of all target parameters from a limb
sequence, the retrieval is decomposed in a sequence of analysis steps. First, spectral
shift is corrected, then the instrument pointing is corrected along with a retrieval
of temperature, using CO2 transitions. Then the dominant emitters in the infrared
spectrum are analyzed one after the other (H2 O, O3 , HNO3 ...), each in a dedicated
spectral region where the spectrum contains maximum information on the target
species but least interference by non-target species. Finally, the minor contributors
are analyzed sequentially, whereby pre-determined information on major and minor
contributors is used. The analysis is done limb sequence by limb sequence, i.e. limb
scans are processed independently from each other by separate calls of the RCP.

3 Computational considerations

The original V8 L1b data consists of binary data sets with one data file containing all
the spectra of one orbit (14.4 orbits per day, 70–100 geolocations per orbit, 17–35 IR
spectra per geolocation, approx. 40000 radiance values per spectrum). The size of
one L1b file is approx. 300 MB. At IMK facilities each file was converted once into a
less complex binary file to keep the reading and conversion costs low at the actual
data processing, since every single spectrum will be read multiple times. Converted
data files are put into archives representing 35 days each (the repeat cycle of the
Envisat orbits). The resulting 110 cycle archives, together with small meta data files,
are stored at the HPSS. The latter files contain information for the efficient access of
the archives. Since the actual processing of the infrared spectra includes a lot of I/O,
reading of spectral and auxiliary data and writing of intermediate results, our setup
included the use of the /tmp-filesystem of the compute nodes (essentially a RAM-disk).
This approach had been chosen to reduce the I/O-load on the temporary workspace
($WORK) filesystem of the XC40 “Hazel Hen” computer. Initially, during the testing
phase, this approach also worked for the “Hawk”. However, the configuration of the
/tmp-filesystem was changed (due to memory leakage problems, it seems) for the
production phase and the approach did no longer work.
As outlined in the last report, 4800 cores were selected as the “sweet spot” and used
throughout the processing of the L2 data on the XC40 “Hazel Hen”, and increasing
the number of used cores would lead to an increased number of corrupted result
data, mainly due to I/O-problems. However, in the course of the adaptation of the
processing software to the “Hawk” machine it was found that the I/O-problems might
have been caused by an error in a script. Testing of this issue was stopped short due
Global long-term MIPAS data processing 401

to the shutdown of “Hawk” in May 2020 caused by the infamous security incident.
After “Hawk” was available again it turned out that indeed most probably the faulty
script was responsible for some of the I/O-error problems. However, the configuration
change with respect to the /tmp-filesystem of the compute nodes, which was heavily
used by our processing software for efficiency reasons, unfortunately forced us to
use the $WORK filesystem for all processing steps instead, which implies a severe
performance degradation. The combined impact of the corrected script (improvement)
and the necessity to switch from the compute node’s /tmp to $WORK (deterioration)
leads to a number of 6000 parallel cores, which can be used safely, without having to
accept I/O-error problems any more. A maximum number of 8000 cores has been
tried and this seems to be the limit for the current configuration. We are eagerly
hoping for the node configuration with respect to /tmp to switch back to the state as it
was during the “Hawk” testing phase.
The processing steps consist of:
• preprocessing, always performed at IMK
• collection of preprocessed data into archives, according to the chosen size of
jobs on the “Hawk”
• transfer of archives containing preprocessed data from IMK to the $WORK
directory
• in parallel: extraction of the necessary V8 L1b data from the HPSS to the $WORK
directory
• submission of compute jobs, i.e. retrieval (core processing step) with the Retrieval
Control Program (RCP) on “Hawk” with approx. 6000 cores used in parallel
• postprocessing to generate some elementary result diagnostics and corresponding
diagnostics plots, again approx. 6000 cores are used in parallel
• collection of results and diagnostics into archives, and transfer of these back to
IMK
At the time of this writing (Nov. 2021) 62% of the granted core hours have been
used.

4 First results

Until the final shutdown of the Cray XC40 spectral shift, temperature and line-of-sight
([8]), H2 O, and O3 have been processed from V8 L1b data. This has been done for
the entire mission, i.e. for 2002–2012 for all major measurement modes: nominal
(NOM), middle and upper atmosphere (MA and UA), noctilucent cloud (NLC) and
upper troposphere lower stratosphere (UTLS) modes. 20% of the allocated compute
time to the MIPAS_V7 project on the Cray XC40 has been used for this task.
The “Hawk” computer has been used for the processing of H2 O (a modified version,
see below), CO, NO, and NO2 , i.e. those species which due to NLTE treatment are
most expensive to calculate.
402 M. Kiefer et al.

The following list gives a short overview over the status of the data processing:
Water Vapour H2 O data for all major measurement modes (NOM, MA, UA, and
UTLS) has been processed for the entire mission already on the XC40. However,
there is a modification to the setup, which is currently in the processing. The
modification essentially shows better validation results in the upper stratosphere
and above.
Nitric Oxide NO data for all major measurement modes (NOM, MA, UA, and
UTLS) has been processed for the entire mission.
Nitrogen Dioxide NO2 data for all major measurement modes (NOM, MA, UA, and
UTLS) has been processed for the entire mission.
Carbon Monoxide CO data for all major measurement modes (NOM, MA, UA, and
UTLS) has been processed for the entire mission.

4.1 Water vapour

As mentioned in our previous report, water vapor retrievals from NOM and
MA/UA/NLC measurement modes need to consider NLTE effects. Non-LTE means
that the populations of the emitting vibrational states, from which atmospheric
variables are derived, do not follow the Boltzmann distribution, which depends
only on the local kinetic temperatures. Since the atmospheric radiation emission
is a function of the populations of the vibrational levels, these populations must
therefore be calculated in detail, considering all possible processes that populate and
de-populate them. The correct abundances can only be retrieved with a complete
model in which all levels, bands and mechanisms affecting the emitting levels are
included. Indeed, the achieved accuracy of the retrieved abundances strongly depends
on the non-LTE model.
For this task, we use our non-LTE model GRANADA (Generic RAdiative traNsfer
AnD non-LTE population algorithm, [9]), which is a sophisticated algorithm that
calculates one by one the impact of the processes at each atmospheric layer, considering
thermal and non-thermal collisions with other atmospheric molecules, chemical
productions and losses, absorption and emission of external radiation and exchange of
energy among atmospheric layers. In the case of water vapor, the effect of non-LTE on
retrievals is noticeable above 50 km [10]. For V8 data we thus include full non-LTE
calculations not only in MA/UA/NLC mode retrievals but also in NOM retrievals.
For the H2 O retrievals, GRANADA calculates the state populations from the
surface up to 120 km in 1 km-steps not only for the H2 O 6.3-micron emitting
levels, but also for a number of other H2 O and O2 states, that also affect the former.
In short, the setup for the calculation of H2 O populations accounts for 23 water
vapor vibrational levels, including three water vapor isotopologues, and 44 molecular
oxygen electronic-vibrational levels. The H2 O levels are connected by 185 radiative
transitions, 10 of them considering full radiative transfer (calculated line-by-line
with Kopra by means of the Curtis matrix formalism). Fifteen radiative transitions
Global long-term MIPAS data processing 403

Fig. 1: Comparison of MIPAS H2 O-retrieval results to ground-based microwave

instrument data. Green: V5 products, red: current V8 products, black: V8 products
with horizontal H2 O-gradient. Symbol for full resolution data is the asterisk; triangle
indicates reduced resolution data. The names of the microwave instrument stations
are given in the plot title.
404 M. Kiefer et al.

connect the electronic O2 states, two of them considering full radiative transfer.
Finally, consideration of 46 processes describing thermal and non-thermal collisions
and chemical productions for H2 O and O2 molecules, not only among each of them
but also with other molecules (e.g., CO2 , N2 or O3 ), is required [9].
The use of these calculations for the H2 O-O2 system with GRANADA coupled
to RCP is expensive in terms of computer time, which increases by a factor of four
in comparison with LTE retrievals. Nevertheless, that seems to be justified by the
benefit of the quality of the MIPAS NOM water vapor, leading to H2 O abundance
improvements of up to 5% at 45-50 km and 10% at 60-70 km. Above that altitude,
i.e., for MA/UA/NLC retrievals, non-LTE are a must in order to retrieve reasonable
H2 O (better than 50% below 90 km).
The delay in V8 L1b data delivery was used to improve the setup for the processing
of water vapour. In addition to the setup for V5 and V7 processing, more spectral
data points and computational expensive radiative transfer calculations in full NLTE
are used to achieve an improvement especially in the altitude range of the lower
Mesosphere (approx. 50–70 km or 1–0.05 hPa).
The gain of these modifications for the water vapour profiles calculated from
NOM observation mode is depicted in Fig. 1, where mean relative differences are
shown between coincident measurements of MIPAS and ground-based microwave
instruments (GBMW) at several sites (Bern [11], Mauna Loa, Lauder and Table
Mountain [12], and Seoul [13]). The criterion for data pairs MIPAS/GBMW to be
considered as coincident are a maximum time difference of 24 hours and a spatial
difference of at most 1000 km. The method of this comparison is described in [14]
In the lower Mesosphere the V8 based water vapour data (red) differs almost
everywhere less than 10% from the GBMW data. This is a clear improvement over
V5 data products (green, details see figure caption).
One further modification in the setup for the retrieval of H2 O profiles was devised
after the first reprocessing was completed (on the XC40). It turned out that inclusion of
one further retrieval parameter, namely the horizontal gradient of H2 O, significantly
improves the results in the upper stratosphere. This is demonstrated in Fig. 1, where
this new V8 data set is shown with black curves/symbols. In the altitude range
0.1–2 hPa the difference compared to the GBMW H2 O profiles are mostly smaller
than for the preceding data version (red, also V8).
In summary the new processing setup for NOM data, with inclusion of com-
putational expensive NLTE treatment and inclusion of the horizontal gradient of
H2 O, clearly leads to very satisfactory agreement of the V8 water vapour data with
independent comparison data in the upper stratosphere and above.
Global long-term MIPAS data processing 405

NO, orbit 40537, 20091201, 00:08-01:40 UTC ppbv NO2, orbit 40537, 20091201, 00:08-01:40 UTC ppbv
80 20 80 20
-30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0 -30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0

70 70
15 15
60 60
Altitude / km

Altitude / km
50 50
10 10
40 40

30 5 30 5

20 20
10 0 10 0
-30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0 -30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0
Latitude / deg Latitude / deg

NO, orbit 37216, 20090412, 23:53-01:32 UTC ppbv NO2, orbit 37216, 20090412, 23:53-01:32 UTC ppbv
80 20 80 20
-30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0 -30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0

70 70
15 15
60 60
Altitude / km

Altitude / km
50 50
10 10
40 40

30 5 30 5

20 20
10 0 10 0
-30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0 -30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0
Latitude / deg Latitude / deg

NO, orbit 38646, 20090721, 21:29-23:08 UTC ppbv NO2, orbit 38646, 20090721, 21:29-23:08 UTC ppbv
80 20 80 20
-30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0 -30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0

70 70
15 15
60 60
Altitude / km

Altitude / km

50 50
10 10
40 40

30 5 30 5

20 20
10 0 10 0
-30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0 -30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0
Latitude / deg Latitude / deg

NO, orbit 39648, 20090929, 21:29-23:08 UTC ppbv NO2, orbit 39648, 20090929, 21:29-23:08 UTC ppbv
80 20 80 20
-30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0 -30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0

70 70
15 15
60 60
Altitude / km

Altitude / km

50 50
10 10
40 40

30 5 30 5

20 20
10 0 10 0
-30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0 -30 0 30 60 90 60 30 0 -30 -60 -90 -60 -30 0
Latitude / deg Latitude / deg

Fig. 2: Result for NO (left column of plots) and NO2 (right) for four orbits representing
northern winter, northern spring, northern summer, and northern fall (from top to
bottom).
406 M. Kiefer et al.

4.2 Nitrogen oxide and nitrogen dioxide

Since the V8 data products have just been processed there are no thorough analyses
available yet. Hence, by now we only show a case of interest.
Nitrogen compounds play a role in the ozone chemistry, e.g. NO contributes
as a free radical catalyst to ozone depletion. Normally NO is only available in the
atmosphere under sunlight conditions. In the darkness it quickly reacts to become
NO2 . This is shown in Fig. 2, where the left column shows NO, while the right column
shows NO2 volume mixing ratios for four orbits representing different atmospheric
conditions (northern winter, spring, summer, and fall). Clearly NO is well constrained
to the dayside of the orbits (marked by red plus signs near the bottom of the plots),
while NO2 mainly shows up during the night (marked by crosses). However, there
are some occurrences of high NO (and of NO2 as well) in the polar mesosphere
(i.e., between 50 and 80 km), especially for hemispheric winter conditions. These
enhancements are produced by ionization of nitrogen and oxygen molecules at higher
altitudes due to precipitating energetic particles. Such precipitation events are caused
by the interaction of the solar wind with the Earth’s magnetic field and affect the
atmosphere in the polar regions, where these particles are guided by the magnetic
field lines towards the Earth. Under polar winter conditions, the produced NO is then
transported downwards with the meridional circulation without being photolyzed
by sunlight and partly converted into NO2 . [15, 16]. Enhanced NO and NO2 is also
visible at 45–50 km during the 2009 Northern hemispheric polar spring (see second
row of Fig. 2) as a consequence of the extraordinarily efficient and confined downward
transport of NO in the following weeks after the strong sudden stratospheric warming
event in January 2009 [17].
The MIPAS V8 NO product from the upper atmosphere mode measurements will en-
ter the successor of the empirical global reference atmospheric model NRLMSISE-00
[18].

4.3 Carbon monoxide

Since the V8 data products have just been processed there are no thorough analyses
available yet. Hence, by now we only present monthly means for a selected year
(2009). Figure 3 shows latitude/altitude distributions of monthly averages of CO for
all twelve months of 2009 (note that the winter month December is put on top to have
the four seasons continuously displayed). The most notable feature is the downward
transport of CO over the hemispheric winter poles. With some caution, CO (plus
additional chemistry modelling) can be used as tracer for atmospheric dynamics. This
is done (together with H2 O and several other species) to retrieve the long term large
scale meridional circulation of the middle atmosphere [19].
Global long-term MIPAS data processing 407
CO in December 2009 log(vmr) CO in January 2009 log(vmr) CO in February 2009 log(vmr)
80 0.5 0.5

80 0.5 0.5

80 0.5 0.5

0.0 0.0

0.0 0.0

0.0 0.0

60 60 60
Altitude / km

Altitude / km

Altitude / km
-0.5 -0.5

-0.5 -0.5

-0.5 -0.5

40 40 40
-1.0 -1.0

-1.0 -1.0

-1.0 -1.0

20 20 20
-1.5 -1.5

-1.5 -1.5

-1.5 -1.5

0 -2.0 -2.0

0 -2.0 -2.0

0 -2.0 -2.0

-50 0 50 -50 0 50 -50 0 50

Latitude / deg Latitude / deg Latitude / deg

CO in March 2009 log(vmr) CO in April 2009 log(vmr) CO in May 2009 log(vmr)

80 0.5 0.5

80 0.5 0.5

80 0.5 0.5

0.0 0.0

0.0 0.0

0.0 0.0

60 60 60
Altitude / km

Altitude / km

Altitude / km
-0.5 -0.5

-0.5 -0.5

-0.5 -0.5

40 40 40
-1.0 -1.0

-1.0 -1.0

-1.0 -1.0

20 20 20
-1.5 -1.5

-1.5 -1.5

-1.5 -1.5

0 -2.0 -2.0

0 -2.0 -2.0

0 -2.0 -2.0

-50 0 50 -50 0 50 -50 0 50

Latitude / deg Latitude / deg Latitude / deg

CO in June 2009 log(vmr) CO in July 2009 log(vmr) CO in August 2009 log(vmr)

80 0.5 0.5

80 0.5 0.5

80 0.5 0.5

0.0 0.0

0.0 0.0

0.0 0.0

60 60 60
Altitude / km

Altitude / km

Altitude / km
-0.5 -0.5

-0.5 -0.5

-0.5 -0.5

40 40 40
-1.0 -1.0

-1.0 -1.0

-1.0 -1.0

20 20 20
-1.5 -1.5

-1.5 -1.5

-1.5 -1.5

0 -2.0 -2.0

0 -2.0 -2.0

0 -2.0 -2.0

-50 0 50 -50 0 50 -50 0 50

Latitude / deg Latitude / deg Latitude / deg

CO in September 2009 log(vmr) CO in October 2009 log(vmr) CO in November 2009 log(vmr)

80 0.5 0.5

80 0.5 0.5

80 0.5 0.5

0.0 0.0

0.0 0.0

0.0 0.0

60 60 60
Altitude / km

Altitude / km

Altitude / km

-0.5 -0.5

-0.5 -0.5

-0.5 -0.5

40 40 40
-1.0 -1.0

-1.0 -1.0

-1.0 -1.0

20 20 20
-1.5 -1.5

-1.5 -1.5

-1.5 -1.5

0 -2.0 -2.0

0 -2.0 -2.0

0 -2.0 -2.0

-50 0 50 -50 0 50 -50 0 50

Latitude / deg Latitude / deg Latitude / deg

Fig. 3: Monthly means of CO in the middle atmosphere. The rows show, from top to
bottom: northern winter (DJF), northern spring (MAM), northern summer (JJA), and
northern fall (SON). The depicted values are common logarithms of the CO volume
mixing ratio (ppmv).

5 Summary and conclusions

The HPE Apollo “Hawk” has been successfully used to run the IMK/IAA L2-processor
with the V8 L1b data of the MIPAS instrument on board the European Environmental
Satellite Envisat. Three species (NO, NO2 , CO) have been processed for the entire
mission duration (2002–2012)., i.e. for approximately 2.5 million geolocations each.
Additionally, a further improved version of H2 O has just entered processing.
The transfer of our processing environment and programs from the XC40 to
the HPE Apollo “Hawk” has been performed successfully. After the testing phase,
however, there was a change with respect to the /tmp-filesystem of the compute nodes.
This change prevents the use of the /tmp-filesystem of the nodes in a way which is
408 M. Kiefer et al.

essential for the efficiency of our retrieval system. We designed our system in a way
to use /tmp (essentially a RAM-disk) to store temporary data and intermediate result
files to minimize the traffic for the $WORK filesystem. As a consequence the current
setup is far from being efficient with respect to I/O-load on the $WORK. Parallel
calculations on 6000 cores has been found to be safe with respect to I/O-problems.
Using 8000 cores occasionally leads to I/O-problems. The HPSS, used to hold the
complete pool of the V8 L1b data necessary for the processing, has proved to be a
reliable and sufficiently fast data repository.
It could already be shown that the new V8 L1b data set is superior to the preceding
versions and especially that the effects of an instrumental drift could largely be reduced
and that now there is much better consistency between FR and RR temperatures
[8]. Additionally, it has been demonstrated that the most recent retrieval setup for
water vapour gives significantly better results, especially in the region of the upper
stratosphere and in the mesosphere.
A part of the data already is available on a data server where already the preceding
data versions can be accessed by interested scientists (http://www.imk-asf.kit
.edu/english/308.php).
The V8 L2 data set, with description of improvements, appropriate characterisa-
tions, error discussion, and validation, will be presented in a cross-journal special issue
of Atmospheric Measurement Techniques and Atmospheric Chemistry and Physics,
see https://www.atmos-meas-tech.net/special_issue1094.html.

Acknowledgements Sylvia Kellmann (KIT-IMK) provided a part of the figures.

References

1. Fischer, H., Blom, C., Oelhaf, H., Carli, B., Carlotti, M., Delbouille, L., Ehhalt, D., Flaud,
J.M., Isaksen, I., López-Puertas, M., McElroy, C.T., Zander, R.: Envisat-MIPAS, an instrument
for atmospheric chemistry and climate research. European Space Agency-Report SP-1229, C.
Readings and R. A. Harris (eds.), ESA Publications Division, ESTEC, P. O. Box 299, 2200 AG
Noordwijk, The Netherlands (2000)
2. Fischer, H., Birk, M., Blom, C., Carli, B., Carlotti, M., von Clarmann, T., Delbouille, L., Dudhia,
A., Ehhalt, D., Endemann, M., Flaud, J.M., Gessner, R., Kleinert, A., Koopmann, R., Langen,
J., López-Puertas, M., Mosner, P., Nett, H., Oelhaf, H., Perron, G., Remedios, J., Ridolfi, M.,
Stiller, G., Zander, R.: MIPAS: an instrument for atmospheric and climate research. Atmos.
Chem. Phys. 8(8), 2151–2188 (2008). DOI 10.5194/acp-8-2151-2008
3. Kleinert, A., Aubertin, G., Perron, G., Birk, M., Wagner, G., Hase, F., Nett, H., Poulin, R.:
MIPAS Level 1B algorithms overview: operational processing and characterization. Atmos.
Chem. Phys. 7, 1395–1406 (2007)
4. Birk, M., Wagner, G.: Complete in-flight detector non-linearity characterisation of MI-
PAS/Envisat (2010). Available at: https://earth.esa.int/documents/700255/707720/Technical+
note+DLR+on+MIPAS+non_linearity_0810.pdf (last access: 27 May 2021)
5. Kleinert, A., Birk, M., Wagner, G.: Technical note on MIPAS non-linearity correction (2015).
Available at: https://earth.esa.int/documents/700255/707720/Kleinert_20151030___TN_KIT
_DLR_nonlin_20151030.pdf, (last access: 29 October 2020)
6. Stiller, G.P. (ed.): The Karlsruhe Optimized and Precise Radiative Transfer Algorithm (KOPRA),
Wissenschaftliche Berichte, vol. FZKA 6487. Forschungszentrum Karlsruhe, Karlsruhe (2000)
Global long-term MIPAS data processing 409

7. von Clarmann, T., Glatthor, N., Grabowski, U., Höpfner, M., Kellmann, S., Kiefer, M., Linden,
A., Mengistu Tsidu, G., Milz, M., Steck, T., Stiller, G.P., Wang, D.Y., Fischer, H., Funke, B., Gil-
López, S., López-Puertas, M.: Retrieval of temperature and tangent altitude pointing from limb
emission spectra recorded from space by the Michelson Interferometer for Passive Atmospheric
Sounding (MIPAS). J. Geophys. Res. 108(D23), 4736 (2003). DOI 10.1029/2003JD003602
8. Kiefer, M., von Clarmann, T., Funke, B., García-Comas, M., Glatthor, N., Grabowski, U.,
Kellmann, S., Kleinert, A., Laeng, A., Linden, A., López-Puertas, M., Marsh, D., Stiller, G.P.:
IMK/IAA MIPAS temperature retrieval version 8: nominal measurements. Atmos. Meas. Tech.
14(6), 4111–4138 (2021). DOI 10.5194/amt-14-4111-2021
9. Funke, B., López-Puertas, M., García-Comas, M., Kaufmann, M., Höpfner, M., Stiller, G.P.:
GRANADA: A Generic RAdiative traNsfer AnD non-LTE population algorithm. J. Quant.
Spectrosc. Radiat. Transfer 113(14), 1771–1817 (2012). DOI 10.1016/j.jqsrt.2012.05.001
10. Stiller, G.P., Kiefer, M., Eckert, E., von Clarmann, T., Kellmann, S., García-Comas, M., Funke, B.,
Leblanc, T., Fetzer, E., Froidevaux, L., Gomez, M., Hall, E., Hurst, D., Jordan, A., Kämpfer, N.,
Lambert, A., McDermid, I.S., McGee, T., Miloshevich, L., Nedoluha, G., Read, W., Schneider,
M., Schwartz, M., Straub, C., Toon, G., Twigg, L.W., Walker, K., Whiteman, D.N.: Validation of
MIPAS IMK/IAA temperature, water vapor, and ozone profiles with MOHAVE-2009 campaign
measurements. Atmos. Meas. Tech. 5(2), 289–320 (2012). DOI 10.5194/amt-5-289-2012
11. Deuber, B., Kämpfer, N., Feist, D.G.: A new 22-GHz radiometer for middle atmospheric water
vapour profile measurements. IEEE Transactions on Geoscience and Remote Sensing 42(5),
974 – 984 (2004). DOI 10.1109/TGRS.2004.825581
12. Nedoluha, G.E., Gomez, R.M., Neal, H., Lambert, A., Hurst, D., Boone, C.D., Stiller, G.P.:
Validation of long term measurements of water vapor from the midstratosphere to the mesosphere
at two Network for the Detection of Atmospheric Composition Change sites. J. Geophys. Res.
118(2), 934–942 (2013). DOI 10.1029/2012JD018900
13. De Wachter, E., Haefele, A., Kämpfer, N., Ka, S., Lee, J.E., Oh, J.J.: The Seoul water vapor
radiometer for the middle atmosphere: Calibration, retrieval, and validation. IEEE Trans. Geosci.
Remote Sens. 49(3), 1052–1062 (2011). DOI 10.1109/TGRS.2010.2072932
14. Nedoluha, G.E., Kiefer, M., Lossow, S., Gomez, R.M., Kämpfer, N., Lainer, M., Forkman,
P., Christensen, O.M., Oh, J.J., Hartogh, P., Anderson, J., Bramstedt, K., Dinelli, B.M.,
Garcia-Comas, M., Hervig, M., Murtagh, D., Raspollini, P., Read, W.G., Rosenlof, K., Stiller,
G.P., Walker, K.A.: The SPARC water vapor assessment II: intercomparison of satellite and
ground-based microwave measurements. Atmos. Chem. Phys. 17(23), 14543–14558 (2017).
DOI 10.5194/acp-17-14543-2017
15. Funke, B., López-Puertas, M., Holt, L., Randall, C.E., Stiller, G.P., von Clarmann, T.: Hemi-
spheric distributions and interannual variability of NOy produced by energetic particle pre-
cipitation in 2002-2012. J. Geophys. Res. Atmos. 119(23), 13,565–13,582 (2014). DOI
10.1002/2014JD022423
16. Funke, B., López-Puertas, M., Stiller, G.P., von Clarmann, T.: Mesospheric and stratospheric
NOy produced by energetic particle precipitation during 2002-2012. J. Geophys. Res. Atmos.
119(7), 4429–4446 (2014). DOI 10.1002/2013JD021404
17. Funke, B., Ball, W., Bender, S., Gardini, A., Harvey, V.L., Lambert, A., López-Puertas,
M., Marsh, D.R., Meraner, K., Nieder, H., Päivärinta, S.M., Pérot, K., Randall, C.E., Redd-
mann, T., Rozanov, E., Schmidt, H., Seppälä, A., Sinnhuber, M., Sukhodolov, T., Stiller,
G.P., Tsvetkova, N.D., Verronen, P.T., Versick, S., von Clarmann, T., Walker, K.A., Yushkov,
V.: HEPPA-II model-measurement intercomparison project: EPP indirect effects during the
dynamically perturbed NH winter 2008-2009. Atmos. Chem. Phys. 17(5), 3573–3604 (2017).
DOI 10.5194/acp-17-3573-2017
18. Picone, J.M., Hedin, A.E., Drob, D.P., Aikin, A.C.: NRLMSISE-00 empirical model of the
atmosphere: Statistical comparisons and scientific issues. J. Geophys. Res. 107(A12), 1468–1484
(2002). DOI 10.1029/2002JA009430
19. von Clarmann, T., Grabowski, U., Stiller, G.P., Monge-Sanz, B.M., Glatthor, N., Kellmann, S.:
The middle atmospheric meridional circulation for 2002-2012 derived from MIPAS observations.
Atmos. Chem. Phys. 21(11), 8823–8843 (2021). DOI 10.5194/acp-21-8823-2021
WRF simulations to investigate processes across
scales (WRFSCALE)

Hans-Stefan Bauer, Thomas Schwitalla, Oliver Branch and Rohith Thundathil

Abstract Several scientific aspects ranging from boundary layer research and land
modification experiments to data assimilation applications were addressed with the
Weather Research and Forecasting (WRF) model from the km-scale down to the
turbulence-permitting scale.
Due to the transition to the new Hawk system, most of the work done in the sub-
projects during the reporting period was related to cleaning up, configuration testing,
transfer of data, and publication of the results of earlier simulations. Investigations
were extended to a second case of the Land Atmosphere Feedback Experiment
(LAFE) in the central United States. The model grid increment was refined from
100 m with 100 vertical levels to 20 m with 200 vertical levels, resulting in more
detailed turbulence structures and stronger variability in the boundary layer. These
are promising results and match well with lidar observations.

1 Introduction and motivation

Numerical models provide the full 4D representation of the atmosphere and a

consistent state with respect to the 3D thermodynamic atmospheric fields, cloud
water, cloud ice, and diagnostic variables such as precipitation. Therefore, they are
excellent tools to improve our understanding of atmospheric processes across scales
and complement to observation-based approaches.
Commonly-used mesoscale models with resolutions on the km scale do not permit
the explicit simulation of smaller-scale processes such as large turbulent eddies (100–
2000 m), and the whole turbulence spectrum is parameterized. Further improvements

Hans-Stefan Bauer, Thomas Schwitalla, Oliver Branch and Rohith Thundathil

Institute of Physics and Meteorology, Garbenstrasse 30, 70599 Stuttgart, Germany,
e-mail: [email protected],
[email protected], [email protected],
[email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 411
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_25
412 Hans-Stefan Bauer, Thomas Schwitalla, Oliver Branch and Rohith Thundathil

are expected if a chain of model runs is performed down to the turbulence-permitting

(TP) scale (100 m resolution and below), as further details of land-surface atmosphere
(LA) interaction are resolved. Nevertheless, processes between the atmosphere and
the land surface are still to date not completely understood and subject of ongoing
research.
The WRF model system [4], described with more detail in earlier reports, provides
the opportunity to perform LES simulations under realistic conditions because of the
wide range of scales it can be applied over. Using a nesting strategy that covers the
synoptic weather situation and the mesoscale circulations, in combination with data
assimilation, will ensure a forcing of the LES that is as close as possible to the real
weather situation.

2 Work done since March 2020

During the report period, apart from some testing, only sub-project one “LES
simulations to better understand boundary layer evolution and high-impact weather
(LES-PROC)” performed simulations.
The sub-projects “Seasonal land surface modification simulations over the United
Arab Emirates (UAE-1)” and “Assimilation of Lidar water vapor measurements
(VAP-DA)” only performed some domain tests during the preparation of the follow-up
proposal for the project. Causes were the delays in the introduction of the new super
computer Hawk and its file system and the necessity to publish the results of the
simulations performed in the previous report period [2, 5].
The sub-project “Turbulence-permitting particulate matter forecast system (Open
Forecast)” performed the last simulation in February 2020 in the previous report
period. The corresponding report was submitted in 2020. The related EU project
ended and the results were published in [3]. Therefore, this annual report focuses on
sub-project LES-PROC and the needed scaling tests done to optimize the operation
of WRF on the new Hawk system.

2.1 Cases and model setup

The Land Atmosphere Feedback Experiment (LAFE) took place in August 2017 at
the Southern Great Plains site of the Atmospheric Radiation Measurement Program
(ARM) in Oklahoma. Many different instruments were brought to the site and the
measurement strategy was optimized in a way to derive as much information as
possible. More details about the campaign and the applied measurement strategy can
be found in [7].
WRF simulations to investigate processes across scales (WRFSCALE) 413

We so far focused our simulations on two LAFE cases on the 23rd of August 2017
and 08th of August 2017. Both were clear sky days with operation of the lidar systems.
On the 23rd of August, the operation was temporally extended to include the evening
transition of the convective boundary layer. On August 8th , although synoptically
similar, a clearly deeper boundary layer developed during day compared to the 23rd .
During the transition to the new Hawk system, we changed to the more recent
version 4.2.1 of WRF. The simulations were started at 06 UTC (01 a.m. local time)
and run for 24 hours within the 24 hour wall time limit on the HAWK system. During
the last report period, we tested the influence of the grid ratio during the nesting to
the target resolution of 100 m. It was found that the results in the innermost domain
are almost identical and the transition to the finer scales performs well in both setups.
Therefore, we focus on the grid ratio of five in the new simulations described in this
report.
WRF was set-up using three domains with 2500 m, 500 m and 100 m resolution.
The outer domain was driven by the operational analysis of the European Centre
of Medium Range Weather Forecasting (ECMWF). The size of the domains is
1000 × 1000 grid cells in the outer domains and 1201 × 1001 grid cells in the
innermost 100 m domain. The vertical resolution is the same in all domains with 100
vertical levels up to a height of 50 hPa with 30 levels in the lowest 1500 m of the
troposphere. Figure 1 shows the third domain with 100 m and the fourth domain with
20 m horizontal resolution.

Fig. 1: Domain configuration of the first LAFE simulation. From left to right the
domains with 2500 m, 500 m and 100 m resolution. The outer and middle domain
have sizes of 1000 × 1000 grid cells and the inner domain consists of 1201 × 1001
points.

In addition to this standard configuration, a fourth nest with a horizontal resolution

of 20 m was added. So far it is not nested “online” by performing all 4 domains in
one simulation. Rather, the NDOWN capability of WRF is applied to calculate the
lateral and initial forcing for this 20 m domain from the 100 m nest. Advantage of
this procedure is that the vertical resolution can be increased for this nest during
NDOWN. In our case we double the vertical resolution to 200 levels ending with
414 Hans-Stefan Bauer, Thomas Schwitalla, Oliver Branch and Rohith Thundathil

a representation of the lowest 1500 m of the troposphere by a approx. 60 levels.

Furthermore, the expensive simulation of the 20 m domain runs separately and can
be reduced to a part of the full simulation period. In the presented first test, the 20
m domain is only simulated for the 3-hour period 18 to 21 UTC (13 to 16 LT in
Oklahoma). Figure 2 shows the third domain with 100 m and the 20 m nest created
with NDOWN.

Fig. 2: Surface orography of the 100 m domain (left) and additionally added 20 m
domain (right). The location of the 20 m domain in the 100 m domain is shown by
the red box.

2.2 Optimization of the model setup on Hawk

Before repeating the simulations for the 23rd of August with the newer model version
and new case studies, tests were performed to optimize the operation of the simulations
on the new Hawk system.
To estimate the performance of the WRF model on the new HAWK system, strong
scaling tests for the configuration applied in LES_PROC subproject were performed.
The scaling tests were done for the three-domain configuration 200 m – 500 m –
100 m described above. Table 1 summarizes the results. The times mentioned are
averaged values, since the numbers vary slightly from timestep to timestep.
The smallest possible configuration was six nodes using 32 MPI tasks per node
with four OpenMP threads each. According to our experience with the system, this is
the best combination of MPI tasks and OpenMP threads to fully use the number of
cores available per node with the applied WRF model. With less nodes, the simulation
crashes because more memory than available is needed. More than 128 nodes with
WRF simulations to investigate processes across scales (WRFSCALE) 415

32 MPI tasks per node is not possible with the selected configuration because then
the sub-domain per MPI task is getting too small. Figure 3 shows that WRF scales
nicely on the HAWK system.

Table 1: Tested configurations for the strong scaling test.

Nodes MPI OpenMP Number Factor Needed Speedup Needed

tasks per threads of cores cores time per timestep time per
Node per MPI timestep radiation
task (s) timestep
(s)

6 32 4 768 1 1.58 1 9.8

8 32 4 1024 1.33 1.17 1.35 7.57
10 32 4 1280 1.67 0.88 1.8 5.45
12 32 4 1536 2 0.7 2.26 4.45
16 32 4 2048 2.67 0.5 3.16 3.3
24 32 4 3072 4 0.32 4.94 2.1
32 32 4 4096 5.33 0.26 6.08 1.7
64 32 4 8192 10.67 0.14 11.29 0.9
128 32 4 16384 21.33 0.08 19.75 0.56

To minimize the time needed for I/O, the WRF model is compiled with parallel
NetCDF support and in addition I/O quilting is applied. The latter reserves some
cores for I/O and reduces the time, the computation is interrupted for I/O to at most a
few seconds even if several files larger than 15 GB are written at the same time.

2.3 Results of the simulations

First, the well-developed boundary layer for both cases is illustrated with the horizontal
distribution of vertical velocity and water vapor mixing ratio interpolated to 1000
m above sea level. Figure 4 shows the two fields for 2 p.m. LT and the two case
studies. For a better illustration of the evolution of turbulence, a subregion of the 100
m domain is presented.
Both cases show a predominant northeasterly to easterly circulation. On the 23rd of
August, clearly drier air is advected into the region. The well-developed turbulence is
clearly seen. The size of the updraft plumes (blue) and the compensating downdrafts
(red) are nicely captured and correspond in size to observed eddy sizes in the region
(e.g. [1, 6, 7]). With this updraft plumes moister air is transported from near surface
416 Hans-Stefan Bauer, Thomas Schwitalla, Oliver Branch and Rohith Thundathil

Fig. 3: Scaling of the different configurations presented in Table 1. T768 is the time
needed with 768 cores, the smallest possible configuration, and TXX is the time
needed by the larger configurations.

to higher levels in the boundary layers seen as green, orange and red plumes on the
right panels of the plot. The compensating downdrafts, on the other hand, transport
drier air downwards. The stronger turbulence developing on August 8 is clearly seen,
transporting larger amounts of moisture upwards.
Figure 5 shows time-height cross sections of water vapor mixing ratio and
horizontal wind velocity to illustrate the temporal evolution of the boundary layer.
The forecast length of 24 hour allows the analysis of a full diurnal cycle of the
temporal evolution of the daytime boundary layer. To increase the temporal resolution,
time series output at selected model grid points were written in addition to provide
data in 10 s resolution.
The different stages of the development of the boundary layer are clearly seen.
During night, a shallow nighttime boundary layer is seen, overlaid by the residual
layer, namely the convective boundary layer of the previous day. In the morning local
time, with the onset of turbulence, the new convective boundary layer is growing. The
turbulent fluctuations are typical for the development of an undisturbed convective
boundary layer. In the evening after sunset, turbulence diminishes and finally a new
nighttime stable boundary layer develops.
Another interesting feature, commonly found in this region is the nighttime
low-level jet, prominently shown by the lower panel, transporting moist air from the
Gulf of Mexico to the region. During daytime, the jet is replaced by the developing
turbulent eddies transporting heat and moisture vertically.
WRF simulations to investigate processes across scales (WRFSCALE) 417

Fig. 4: Representation of the convective boundary layer zoomed into a small region
around the ARM SGP site (black dot) at 2 p.m. LT on 23 August 2017 (top) and for 8
August 2017 (bottom). Vertical velocity 𝑚/𝑠 1000 m above sea level is shown in the
left column and water vapor mixing ratio 𝑔/𝑘𝑔 1000 m above sea level in the right
column.

Another new result is the higher 20 m resolution inner domain. Figure 6 compares
the same fields vertical velocity and water vapor mixing ratio interpolated to 1000 m
above sea level for the 100 m and the 20 m resolutions over the domain of the 20 m
simulation at 19 UTC (15 LT in Oklahoma) during the well-established turbulent
boundary layer.
The comparison demonstrates that the finer resolution reveals more details of the
evolution of turbulence. Especially much more movement is seen along the edges of
the turbulent updrafts and outflow boundaries are seen in the downdraft regions.
418 Hans-Stefan Bauer, Thomas Schwitalla, Oliver Branch and Rohith Thundathil

Fig. 5: Time-height cross sections of water vapor mixing ratio [𝑔𝑘𝑔−1 ] (top) and
horizontal wind velocity [𝑚/𝑠] (bottom) on August 23 rd for the grid cell the
Hohenheim lidar systems were located during the campaign. The X-Axis marks the
time in hours since the beginning of the forecast (00 corresponds to 06 UTC or 01 LT
in Oklahoma).

Finally, the vertical representation of turbulence in the two resolutions is compared.

This is done in Figure 7 showing the time-height cross sections of vertical velocity
for the three-hour time window 18 to 21 UTC (13 to 16 LT) for the grid point the
ARM Doppler lidar was located. The two simulated representations are compared
with the observation of the ARM lidar.
Comparing the two model simulations reveals the finer details simulated by the
20 m resolution. This is more in accordance with the lidar observation leading to
stronger variability of vertical velocity in the boundary layer (see e.g. Bauer et al.,
2020). The PBL height, on the other hand is represented by both resolutions. At the
moment, it is investigated in detail of whether the higher horizontal or the doubled
vertical resolution is the major cause for the more realistically represented boundary
layer.
WRF simulations to investigate processes across scales (WRFSCALE) 419

Fig. 6: Comparison of vertical wind velocity [𝑚/𝑠] (top) and water vapor mixing
ratio [𝑔𝑘𝑔−1 ] 1000 m above sea level. Shown are results with 100 m resolution (left)
and 20 m resolution (right) for the region covered by the 20 m domain.

3 Used resources

Table 2 lists the resources used during the report period March 2020 to June 2021.
The term “clean up” also contains the move of all needed data from the old work
space “ws9” to the new “ws10” and “testing” refers to the optimization of WRF on
the new HAWK system.
420 Hans-Stefan Bauer, Thomas Schwitalla, Oliver Branch and Rohith Thundathil

Fig. 7: Time-height cross section of vertical velocity [𝑚/𝑠] of the model simulations
with 100 m resolution (upper left), 20 m resolution (upper right) and the observation
of the Doppler lidar (lower left) for the three-hour time window 18 to 21 UTC (13 to
16 LT).

Table 2: Resources used by the WRFSCALE project between March 2020 and June
2021.

Subproject User Computing time (Rth.) Purpose

LES_PROC Hans-Stefan Bauer 10899.2848 Clean up, testing and

case study simulations
mesoscale to LES
UAE-1 Oliver Branch 1373.6066 Clean up + testing
VAP-DA Rohith Thundathil 46.7555 Clean up
Open Forecast Thomas Schwitalla 574.9204 Clean up

Total 12894.5673
WRF simulations to investigate processes across scales (WRFSCALE) 421

References

1. H.-S. Bauer, S.K. Muppa, V. Wulfmeyer, A. Behrendt, K. Warrach-Sagi, and F. Späth.

Multi-nested WRF simulations for studying planetary boundary layer processes on the
turbulence-permitting scale in a realistic mesoscale environment. Tellus A, 72:1–28, 2020.
doi:10.1080/16000870.2020.1761740.
2. O. Branch, T. Schwitalla, M. Temimi, R. Fonseca, N. Nelli, M. Weston, J. Milovac, and
V. Wulfmeyer. Seasonal and diurnal performance of daily forecasts with WRF v3.8.1 over the
united arab emirates. Geosci. Model Dev., 14:1615–1637, 2021. doi:10.5194/gmd-14-1615-2021.
3. T. Schwitalla, H.-S. Bauer, K. Warrach-Sagi, T. Bönisch, and Wulfmeyer. V. Turbulence-
permitting air pollution simulation for the stuttgart metropolitan area. Atmos. Chem. Phys.,
21:4575–4597, 2021. doi:10.5194/acp-21-4575-2021.
4. W.C. Skamarock, J.B. Klemp, J. Dudhia, O. Gill, Z. Liu, J. Berner, W. Wang, J.G. Powers, M.G.
Duda, D.M. Barker, and X.-Y. Huang. A description of then advanced research WRF version 4.
NCAR Tech. Note, NCAR/TN-556+STR:145pp, 2019. doi:10.5056/1dfh-6p97.
5. R. Thundathil, T. Schwitalla, A. Behrendt, and V. Wulfmeyer. Impact of assimilating lidar
water vapour and temperature profiles with a hybrid ensemble transform kalman filter – three-
dimensional variational analysis on the convection permitting scale. Quart. J. Roy. Meteorol.
Soc., 147:1–23, 2021. doi:10.1002/qj.4173.
6. D.D. Turner, V. Wulfmeyer, L.K. Berg, and J.H. Schween. Water vapor turbulence profiles in
stationary continental convective mixed layers. J. Geophys. Res. Atmos., 119:151–165, 2014.
doi:10.1002/2014JD022202.
7. V. Wulfmeyer, D.D. Turner, B. Baker, A. Behrendt, T. Bonin, A. Brewer, M. Burban,
A. Choukulkar, E. Dumas, R.M. Hardesty, T. Heus, J. Ingwersen, D. Lange, T.R. Lee, S. Met-
zendorf, S.M. Muppa, T. Meyers, R. Newsom, M. Osman, and S. Raasch. A new research
approach for obswerving and characterizing land-atmosphere feedback. Bull. Amer. Meteorol.
Soc., 99:1639–1667, 2018.
 Computer Science

The following part of this volume typically is a smaller one, and it deals with research
labelled as “Computer Science”, mostly due to the fact that the respective groups
are affiliated with Computer Science departments. As in previous years, however,
if we look at the topics addressed in the five annual reports submitted this year,
the impression is a bit different: we find both classical informatics topics (such as
load balancing, machine learning, or discrete algorithms) and classical application
domains (such as molecular dynamics or multi-physics problems). Nevertheless,
the common theme in the projects and the reports is their focus on state-of-the-art
informatics topics related to HPC.
Out of those five submissions undergoing the usual reviewing process, two project
reports were selected for publication in this volume: GCS-MDDC and SDA.
The contribution Dynamic Molecular Dynamics Ensembles for Multiscale Simula-
tion Coupling by Neumann, Wittmer, Jafari, Seckler and Heinen reports on recent
progress in the GCS-MDDC project. The project’s basis are, first, the software ls1
mardyn, a molecular dynamics (MD) framework for multi-phase and multi-component
studies at small scales with applications in process engineering, and, second, the
coupling software MaMiCo that addresses the challenge of efficiently sampling
hydrodynamic quantities from MD. The overall goal is to couple CFD solvers with
ls1 mardyn, thus allowing for making larger scales accessible. The report presented
in this volume deals with a first step in that direction, an extension of MaMiCo to
handle dynamic ensembles. The system used for the computations of this project was
HAWK at HLRS.
The second report, Scalable Discrete Algorithms for Big Data Applications by
Hespe, Hübschle-Schneider, Sander, Schreiber and Hübner is definitely not on the
usual suspects side. In contrast to classical HPC applications based on continuous
(numerical) algorithms, the SDA project focuses on discrete algorithms, to be precise
SAT solving, malleable job scheduling, load balancing, and fault tolerance. In
the period to be reflected in this volume, emphasis was put on SAT solving and
fault-tolerant algorithms. The system used for that was ForHLR II at KIT.

423
424 Computer Science

Together, both papers nicely reflect the breadth of HPC use in science: continuous
and discrete, simulation as well as data engineering and analytics.

Technical University of Munich, Hans-Joachim Bungartz

Department of Informatics,
Chair of Scientific Computing,
Boltzmannstrasse 3,
D-85748 Garching,
Germany
Dynamic molecular dynamics ensembles for
multiscale simulation coupling

Philipp Neumann, Niklas Wittmer, Vahid Jafari, Steffen Seckler and Matthias Heinen

Abstract Molecular dynamics (MD) simulation has become a valuable tool in process
engineering. Despite our efforts in software developments for large-scale molecular
simulations over several years, which have amongst others enabled record-breaking
trillion-atom runs, MD simulations are—as stand-alone simulations—limited to
rather small time and length scales. To make bigger scales accessible, we propose to
work towards a coupling of CFD solvers with our efficient MD software ls1 mardyn.
As a first step, we discuss extensions to our coupling software MaMiCo, addressing
the challenge of efficiently sampling hydrodynamic quantities from MD: due to
the high level of thermal fluctuations, MD ensemble considerations are required
for sampling. We propose an extension of MaMiCo that allows to handle dynamic
ensembles, i.e. to launch and remove MD simulations on-the-fly over the course of
a coupled simulation. We explain the underlying implementation and provide first
scalability results.

1 Introduction

Molecular dynamics (MD) simulation has become a valuable tool for various
application fields, in particular also for process engineering. Over several years,
we have developed the software ls1 mardyn [13] for this purpose, i.e. ls1 mardyn
specializes on the handling of large numbers of rather small molecules. Recently,
ls1 mardyn was extended by (1) auto-tuning at node-level, always choosing the

Philipp Neumann, Niklas Wittmer and Vahid Jafari

Helmut Schmidt University, Chair for High Performance Computing, Holstenhofweg 85, 22043
Hamburg, e-mail: [email protected]
Steffen Seckler
Technical University of Munich, Chair of Scientific Computing, Boltzmannstr. 3, 85748 Garching
Matthias Heinen
Technical University of Berlin, Chair of Thermodynamics, Ernst-Reuter-Platz 1, 10587 Berlin

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 425
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_26
426 P. Neumann, N. Wittmer, V. Jafari, S. Seckler and M. Heinen

optimal combination of particle data container, traversal scheme and corresponding

shared-memory parallelization scheme on-the-fly during a simulation [2, 14] and (2)
an improved dynamic load balancing approach to tackle highly heterogeneous loads
such as liquid-gas systems at inter-node level [14].
This software development, in conjunction with the provision of computational
resources in the scope of a GCS Large Scale Project, allowed us to simulate nanoscale
droplet coalescence at unprecedented size, cf. Figure 1. Our biggest study with two
200nm-sized (in terms of diameter; twice as big as the example shown in Fig. 1)
droplets using the dynamic load balancing of ls1 mardyn consumed approx. 11
million core hours. It also enabled direct comparison with coarse-grained modelling

Fig. 1: Snapshot, rendered by MegaMol [3], of two coalescing argon droplets with an
initial diameter of 100 nm (bottom). The molecular system consists of ≈ 25 million
particles in total. Particles constituting the initial left and right droplet were colored
green and red, respectively. The magnified view (top) shows the upper part of the
growing bridge between the droplets.
Dynamic molecular dynamics 427

approaches such as phase-field theory which showed very good agreement between
both modelling approaches, see amongst others [1].
These findings on consistency between coarse- and fine-scale models suggest that
a coupling of coarse-scale, continuum models and fine-scale, molecular models could
be a promising option to resolve molecular behavior only when and where necessary.
Molecular-continuum methods have been established for this purpose in the past.
We believe that this will make even bigger scenarios accessible through available
supercomputing resources, such as the current machine HAWK hosted at HLRS. One
major challenge for these methods lies, however, in efficient sampling of the highly
fluctuating thermodynamics at the molecular scale.
In this regard, we present preparatory results that we have worked on to push our
software technology towards multiscale molecular-continuum simulations. Over the
past years, much development effort also went into our macro-micro-coupling tool
(MaMiCo) [10, 11] which allows the coupling of arbitrary continuum flow solvers and
MD packages. In the mid term, we strive for coupling ls1 mardyn with MaMiCo and
thus bring together both highly efficient molecular simulation software and the
multiscale approach. To address the challenge of sampling the highly fluctuating
quantities on the molecular scale, we present work that extends MaMiCo towards
dynamic MD ensembles in the following.
We describe the underlying idea of molecular-continuum simulations in Sect. 2,
including a short recap of related work and the involved software. The implementation
of dynamic MD ensembles is discussed in Sect. 3. In Section 3, we validate this
ensemble handling in parallel simulations, considering a Couette flow scenario.
We also present scalability results in the same section. Findings and next steps are
summarized in Sect. 5. Parts of this report have been accepted for publication [5].

2 Molecular-continuum coupling

2.1 Short-range Molecular Dynamics

In the field of molecular dynamics (MD) we investigate fluids, suspensions, and

materials at the nanoscale. The system under consideration is modelled by particles
which are characterized by their positions x𝑖 , velocities v𝑖 and forces F𝑖 𝑗 .
Forces typically arise from pairwise inter-molecular potentials. In the following,
we will restrict our considerations to the Lennard–Jones potential from which the
following inter-molecular force arises between two particles 𝑖 and 𝑗:
 ! 14 ! 8
48  𝜎 1 𝜎 
F𝑖 𝑗 = 2  − (xi − xj ), (1)
𝜎  𝑟𝑖 𝑗
 2 𝑟 𝑖 𝑗 
 
428 P. Neumann, N. Wittmer, V. Jafari, S. Seckler and M. Heinen

where 𝑟 𝑖 𝑗 := ||x𝑖 − x 𝑗 ||. This distance is additionally limited to a cut-off radius

𝑟 𝑐 , such that F𝑖 𝑗 = 0 for 𝑟 𝑖 𝑗 > 𝑟 𝑐 (short-range molecular dynamics). 𝜖, 𝜎 are
material-dependent parameters. Furthermore, we restrict our considerations to NVT
ensembles, this means that beside a fixed density, the temperature of the investigated
systems is maintained to be constant via thermostats.

2.2 Computational fluid dynamics

In the following, we make use of a simple Lattice Boltzmann solver with the standard
BGK collision model, see e.g. [9], to compute the fluid flow on the continuum scale:
1
𝑓𝑖 (x + c𝑖 Δ𝑡, 𝑡 + Δ𝑡) = 𝑓𝑖 (x, 𝑡) − ( 𝑓𝑖 (x, 𝑡) − 𝑓𝑖𝑒𝑞 (x, 𝑡)), 𝑖 = 1, ..., 𝑄. (2)
𝜏
Here 𝑓𝑖 (x, 𝑡) denotes the probability to find particles in a cell with midpoint x and
time 𝑡 moving with lattice velocity c𝑖 . The set of lattice velocities is fixed, we rely
on the well-known D3Q19 discretization with 19 lattice velocities in 3D. The fluid
density 𝜌(x, 𝑡), pressure 𝑝(x, 𝑡) and velocity u(x, 𝑡) evolve locally from:
𝑄
Í
𝜌(x, 𝑡) = 𝑓𝑖 (x, 𝑡),
𝑖=1

𝑝(x, 𝑡) = 31 𝜌(x, 𝑡), (3)

𝑄
Í
𝜌(x, 𝑡)u(x, 𝑡) = 𝑓𝑖 (x, 𝑡)c𝑖 .
𝑖=1

As boundary conditions, we rely on periodic conditions as well as half-way bounce

back modeling of solid stationary and moving walls.
LB methods are very efficient and are known to scale very well up to full
supercomputer machine size due to their local nature in computations and since they
operate typically on standard Cartesian grids (or adaptive variants thereof). Note that
the choice of the fluid flow solver is however more or less arbitrary from a technical
perspective, as long as density and velocity values are available and can be provided
to our coupling tool MaMiCo, cf. Sect. 2.4, on a mesh. From a physical perspective,
the choice of the CFD solver may impact accuracy and stability of the arising coupled
scheme.
Dynamic molecular dynamics 429

2.3 Molecular-continuum algorithm in a nutshell

In the following, we will restrict views to concurrent transient coupling of MD and

CFD. The method sketched in this section follows the algorithm described in [12]
and slightly adopted in [10]; cf. the latter reference for further details.
Our coupling method makes use of an overlapping domain decomposition, i.e.
both MD and CFD domain overlap to some extent in space and this overlapping
region is used to couple both simulations. In our case, the MD domain is basically
embedded in the CFD domain, and overlapping is assumed in a three Cartesian grid
cell-wide section within the MD domain. Mass and momentum are transferred from
the CFD solver to the MD solver inside this overlap region in separate sub-regions:
from the MD perspective, the outermost layer of overlapping Cartesian grid cells is
used to transfer mass using particle insertion and removal mechanisms. The following
two cell layers are used to exchange momentum by imposing the average acceleration
from the CFD solution onto the MD system. Boundaries of the MD region are either
modelled with periodic boundaries where appropriate or with reflecting boundaries
to hinder molecules from escaping, together with additional boundary forcing to
impose the correct average hydrodynamic pressure on the MD system.
This approach is sufficient for one-way coupling, i.e. to couple the CFD solver to
the MD solver, which is subject to the current work. Note that, to enable two-way
coupling, MD data such as average flow velocities are sampled cellwise in the inner
region of the MD simulation and are imposed on CFD side as boundary condition.

2.4 MaMiCo

MaMiCo stands for macro-micro coupling tool and shall facilitate the coupling of
particle- and mesh-based solvers for multiscale fluid flow simulations, with focus on
molecular-continuum coupling [6,7,10,11]. It provides both steady-state and transient
coupling schemes, with the latter given in the scope of the algorithm presented in
Sect. 2.3.
MaMiCo especially provides the ability to couple CFD solvers to ensembles
of MD simulations: let a MD domain be embedded somewhere in a big CFD
domain. Then, the MD simulation is computed multiple times starting from random
initial configurations. This yields an ensemble of MD simulations whose average
hydrodynamic data such as flow velocities can be computed locally (i.e. per CFD-
related grid cell). An averaging procedure is essential, since MD exhibits strong
thermal fluctuations in typical scenarios of interest. These fluctuations may be critical
for the stability of the CFD solver, if this solver is not compressible and incorporating
thermal fluctuations in a comparable way. Ensemble averaging (1) removes potential
biases from time-dependent averaging, (2) allows to investigate scenarios on shorter
time scales, yet relying on a coarse-scale CFD view and (3) is preferable in terms of
leveraging supercomputing capacities due to the embarrassingly parallel nature of
running the different ensemble members.
430 P. Neumann, N. Wittmer, V. Jafari, S. Seckler and M. Heinen

MaMiCo is parallelized with MPI, supporting arbitrary domain decompositions

on CFD side, blockwise domain decomposition on MD side and ensemble-wise
parallelism for the MD simulations. The tool further provides coupling interfaces
to a multitude of MD codes such as LAMMPS and ls1 mardyn [11]. For the current
work, we focus on the use of SimpleMD, a built-in MD simulation that is shipped
with MaMiCo, and a built-in LB solver.

2.5 Related work

Molecular dynamics-continuum coupled simulations have been the subject of numer-

ous studies; we will only review selected approaches which make use of ensembles
or well-chosen sets of MD simulations.
In [15], a molecular-continuum method is introduced in which MD simulations
are dynamically launched to investigate unsteady channel flow. MD simulations are
executed at different cross-sections of the channel, their data is gathered in a database
and interpreted via Gaussian process regression to steer accuracy of the flow over
long time scales. This is different to our approach, in which we currently only resolve
one particular sub-region of the considered computational domain with MD.
Besides, various sampling methods for MD systems have been developed in the
past to assess long time scale behavior or large molecular systems. Examples are
given by equation of state sampling [8] or the investigation of rare events [4]. In our
work, we consider quasi-identical MD simulations to explore a single thermodynamic
non-equilibrium state.
To the best of our knowledge, no HPC-aware, massively parallel molecular-
continuum simulation approach with dynamic ensemble handling has been developed
until now.

3 Implementation of dynamic MD ensemble handling

The MultiMDCellService of MaMiCo is the central class to incorporate the coupling

of an MD ensemble and a single instance of a CFD solver [10]. This service
manages an array of MacroscopicCellService instances, each being responsible
for a single MD simulation in the coupling [11]. Each MPI rank of a running
simulation holds one instance of the MultiMDCellService, which in turn manages a
MacroscopicCellService for every available MD simulation. Two implementations
of the MacroscopicCellService are used: MacroscopicCellServiceImpl manages the
interactions between the CFD solver and MD solvers. The other implementation,
MacroscopicCellServiceMacroOnly, is necessary if the MPI parallelization of the
respective CFD solver requires communication with an MD simulation that does not
have a sub-domain to be computed on the CFD solver ranks.
Dynamic molecular dynamics 431

During the setup of MaMiCo, the MPI ranks are being grouped into equally sized
blocks. The size of these groups corresponds to the number of processes per MD
instance, which is the same across all instances. These groups of MPI ranks are also
ordered according to their respective ranks; e.g., assuming a 2 × 2 × 2 block-wise
decomposition of the MD domain and a total of 24 processes being available, we
obtain 3 groups with ranks 0-7, 8-15, 16-23.
The initialization of MaMiCo works such that the MD simulations are homo-
geneously distributed across the process groups. Oversubscription is possible, i.e.
multiple MD simulations can execute within a process group. For example, in the
constellation from above, we could run 6 MD simulations, with each process group
handling 2 MD simulations.
Inserting or removing arbitrary numbers of MD simulations would break this
homogeneity in MD simulation load. We therefore add a layer of abstraction by
introducing slots in MaMiCo. The number of slots in a process group corresponds
to the size of the MacroscopicCellService array of the MultiMDCellService. It can
furthermore be active or inactive. An active slot’s MacroscopicCellService takes
part in the running simulation. If it is inactive, no work or communication will be
performed.
For management of the slots we introduce two new classes in MaMiCo. First,
InstanceHandling can be used to centralize all tasks regarding the use of MD
instances. The InstanceHandling holds an STL vector of MD simulations and one for
MicroscopicSolverInterfaces. Initialization, execution of MD time steps and shutdown
are abstracted into this class. Second, the class MultiMDMediator manages the slots;
adding new slots as well as activation/ deactivation is handled. The MultiMDMediator
is closely coupled with the MultiMDCellService and the InstanceHandling.
The process of launching a new MD simulation works as follows. A slot is chosen
either manually by declaring the exact slot or the MPI process group from which a
slot should be chosen, or according to the MultiMDMediator, which tries to keep the
number of MD simulations per rank balanced by applying a round-robin scheme.
After a slot has been selected, the MultiMDMediator delegates to the Multi-
MDCellService which initializes a new MacroscopicCellService which in turn
corresponds to the selected slot. In the activated slot, a new MD simulation is
launched, together with a corresponding communicator (related to the process group).
Each process group regularly saves checkpoints. Upon launch, the new MD instance
is initiated from the last available checkpoint. As this checkpoint will basically always
reflect a deprecated state of the flow behavior, the new MD instance first needs to
be equilibrated and pushed towards the current CFD state over a defined number
of coupling cycles. Currently, we use 10 coupling cycles, with each coupling cycle
comprising 50 MD time steps, in the concurrently coupled simulation. During this
equilibration phase, the rest of the coupling simulation continues. The new simulation
receives coupling information from the CFD solver, but does not take part in the
ensemble sampling or sending of information to the CFD solver; note that this
sampling is also relevant in the one-way coupling to extract flow information from
432 P. Neumann, N. Wittmer, V. Jafari, S. Seckler and M. Heinen

the MD system for post-processing. After the equilibration phase, the new simulation
is activated and starts behaving like all other MD simulations, i.e. it is fully integrated
into the MD ensemble.
On the one hand, to achieve high performance, the frequency at which the required
checkpoints are written should be low to avoid frequent I/O. On the other hand, for the
sake of physical consistency and rapid equilibration, checkpoints should be written
as often as possible. This opposition will require further research and evaluations to
find a good compromise in the future.
Furthermore, the checkpoint used for initialization of the simulation resembles the
state of another running MD simulation from an unknown number of coupling cycles
before; the checkpoint might even originate from the very last finished cycle—which
is good in terms of physical consistency with regard to the CFD state, but which
is very bad in terms of adding an entirely independent ensemble member to the
ensemble of MD simulations. In addition to the equilibration phase, we therefore
vary the particles’ velocities of the checkpoint. For particles residing in one CFD
grid cell, we set their new velocities to the mean flow velocity in this grid cell and
add Gaussian noise, resembling the corresponding temperature. This will introduce
enough chaos into the system so that the state of the new MD simulation will rapidly
diverge from the original MD simulation which produced the checkpoint.
Removal of simulations is simpler. The MD simulation and its corresponding
MDSolverInterface are shut down. Then, the respective instance of the Macroscopic-
CellService is removed. Finally, the selected slot will be set to inactive. This slot is
now available again for the launch of a new MD instance in the future.

4 Results

In the following subsections, the name MD30 refers to a specific single-site Lennard–
Jones (LJ) MD domain and simulation configuration shown in Tab. 1. This MD
configuration is coupled with the LB solver in order to simulate Couette flow: in this
scenario, we consider the 3D flow between two plates, with the lower plate moving at
fixed velocity. Although the evolving flow profile is typically one-dimensional, we run
the full 3D simulation for the sake of preparing our entire simulation methodology for
more complex, potentially fully 3D flow scenarios in the future. From the configuration
MD30, we derive MD60 by doubling all MD and CFD domain sizes, while keeping
the other parameters fixed, cf. also [10] for more details.
We first validate the correctness of our implementation. For this purpose we prepare
an MD30 scenario using MaMiCo’s integrated MD simulation code SimpleMD in
Sect. 4.1. We then investigate the scalability of the implementation running two
distinctive scenarios on the HAWK system in Sect. 4.2.
Dynamic molecular dynamics 433

Table 1: Configuration parameters for MD30 setup.

Parameter Value Parameter Value

molecular mass (𝑚) 1 density 𝜌 0.81

𝜎 1 kinematic viscosity 2.64/𝜌
𝜖 1 channel size (50,50,50)
Boltzmann constant 1 MD domain size (30,30,30)
temperature 1.1 moving wall velocity 0.5
cut-off radius 𝑟𝑐 2.2 MD domain center (25,25,17.5)
MD time step 0.005 CFD time step 0.25

4.1 Validation

We validate our implementation in a MD30 Couette flow scenario, cf. Tab. 1 for
the parametrization. The coupling is initialized using 128 SimpleMD simulations
that are coupled to the LB solver. The simulation is executed over 1000 coupling
cycles, i.e. CFD time steps. After the first 100 cycles, dynamic launch and removal is
activated. For this purpose, every 20 coupling cycles, we choose a random number
𝑟 ∈ (−50, 50) which corresponds to the number of MD simulations that shall be
removed from (𝑟 < 0) or added to (𝑟 > 0) the ensemble.
This study was performed using one full node on the Hawk system. We applied a
domain decomposition scheme with 2 × 2 × 2 processes on LB side and 2 × 2 × 2
processes for every MD simulation. We use the notation LB2 × 2 × 2-MD2 × 2 × 2
for these parallel configurations in the following. The entire simulation consumed
approx. 105 core-h.
Figure 2 shows the Couette flow profile between the plates after different coupling
cycles. The velocity is measured across a 1D cross-section between the plates. It
can be observed, that the MD ensemble solution follows the analytical solution very
well. It is further demonstrated that varying the size of the MD ensemble affects the
degree of deviation from the analytical profile: before the start of the dynamic MD
simulation handling, the measured state fits the expected state well. At cycle 130, we
can see some slightly higher fluctuations around the expected state. This fluctuation
is even larger at cycle 410, whereas at cycle 640 it has decreased again and is thus
closer to the expected state.
Comparing these observations to Fig. 3, which shows the number of MD simu-
lations 𝑀 over the course of the simulation, we see that the number of active MD
instances first decreases and reaches a number of 32 at cycle 410. After this cycle,
the number of MD instances within the ensemble starts to increase again, reaching its
434 P. Neumann, N. Wittmer, V. Jafari, S. Seckler and M. Heinen

0.5
cycle 30
cycle 50
0.4 cycle 80
Velocity (x-direction)
cycle 130
0.3 cycle 250
cycle 410
cycle 640
0.2

0.1

0.0
0 10 20 30 40 50
z (channel cross-section)

Fig. 2: Couette flow profile of the simulation after different coupling cycles. The
profile is sampled across a 1D cross-section between the plates. Lines depict the
analytical flow profile, diamonds depict the state sampled from the MD ensemble.

300
M instances
250

200

150

100

50

00 100 200 300 400 500 600 700 800 900 1000
Cycle

Fig. 3: Dynamic change of the number of MD instances over the course of the entire
coupled simulation.

maximum of 298 after cycle 640. We can thus conclude that our dynamic coupling
scheme qualitatively works as expected. A more rigorous error estimator is under
current development and has already confirmed our findings.
Dynamic molecular dynamics 435

4.2 Scalability

We investigate the strong scalability on Hawk in two Couette scenarios, considering

MD30 and MD60. For MD30, we run each simulation over 100 coupling cycles and
perform random launch/removal of MD instances every 10 cycles.
In the second scenario based on MD60, we run a coupling simulation over a total
of 20 coupling cycles. Random launch or removal happens every 5 cycles.
To retain comparability among all runs, we impose a lower limit of 32 active
MD instances. This is necessary due to the fact that the actual reasonable minimum
number of MD simulations is defined by the number of MPI communicators:
once a process group becomes entirely empty, we waste these idle resources and
can—without significant loss in performance—execute at least one MD simulation
in the corresponding slots, since there will be at least one other process group
that executes at least one MD simulation. As we investigate strong scalability, the
number of communicators varies across setups. The number of communicators
effectively sets the minimal number of MD simulations running at any time. To
circumvent this behavior and thus ensure comparability between the runs, we enforce
a minimal number of 32 instances on setups that run with less communicators. This
furthermore limits the range of possible setups to those consisting of no more than
32 communicators.

Table 2: Time-to-solution (in seconds) on the MD30 scenario. A configuration of

(LB𝑀 × 𝑀 × 𝑀-MD𝑁 × 𝑁 × 𝑁) corresponds to a decomposition of the LB solver
into 𝑀 × 𝑀 × 𝑀 equally sized sub-domains and of the MD domain into 𝑁 × 𝑁 × 𝑁
sub-domains.

Topology # cores

1 2 4 8 16 32 64 128 256 512

LB 1 × 1 × 1-MD1 × 1 × 1 19608 10188 5174 3098 2044 1037 461

LB 1 × 1 × 1-MD2 × 2 × 2 3894 2059 1034 680 573 286 120
LB 2 × 2 × 2-MD2 × 2 × 2 3920 1982 1030 675 562 285 117

Table 2 shows the runtime results of the MD30 configuration. The 1-core topology
was run in sequential mode, i.e. without MPI parallelization. This study consumed
about 280 core hours.
The run time results of MD60 are shown in Tab. 3. This study consumed about
920 core hours.
The speedups for both MD30 and MD60 are displayed in Fig. 4. We observe
that both scenarios scale reasonably well. In the MD30 scenario, there is a drop in
efficiency visible for the larger setups of the LB2 × 2 × 2 decomposition. This is due
to the relatively small-sized MD domain and the actual communication pattern: with
436 P. Neumann, N. Wittmer, V. Jafari, S. Seckler and M. Heinen

Table 3: Time-to-solution (in seconds) on different parallel topologies of MD60.

Nomenclature follows that of MD30 (cf. Table 2).

Topology # cores

1 2 4 8 16 32 64 128 256 512 1024 2048 4096

LB1 × 1 × 1-
48933 26897 14008 7589 4574 2320 1290
MD 1 × 1 × 1
LB1 × 1 × 1-
7908 3972 2084 1202 889 453 225
MD2 × 2 × 2
LB2 × 2 × 2-
7877 3966 2085 1191 901 466 231
MD2 × 2 × 2
LB2 × 2 × 2-
1688 1053 567 301 177 91 50
MD4 × 4 × 4
LB4 × 4 × 4-
1677 937 526 299 178 90 46
MD2 × 2 × 2

a domain decomposition of MD2 × 2 × 2, each sub-domain contains around 2740

particles. Considering the initial number of 2 MD simulations per communicator,
the load per process is relatively small. Additionally, each process has to execute a
relatively high number of small MPI send operations. In contrast, in the scenario
LB1 × 1 × 1-MD2 × 2 × 2, the same amount of data needs to be sent on average via
MPI, but each process needs to execute less send operations per MD instance.
For MD60, we do not observe this drop in performance. As discussed above, it
was not possible with this setup to scale MD60 any further without increasing the
minimum number of MD instances. Larger simulations are to be executed in the
future.
We also investigated the time spent in the actual launch/removal phases of MD30.
We achieve a mean of 1.7% of the overall runtime. None of the runs exceeded a total
launch/removal time fraction of 4.5%.

5 Summary and outlook

We have introduced a mechanism for dynamic launch and removal of coupled MD

ensembles. Our results show that our implementation works in reasonable time frames,
even when applying it in drastic scenarios of dynamically launching/ removing large
numbers of MD simulations at short time intervals. We have yet to perform further
scalability studies at larger scale, especially with regard to weak scaling, and also in
combination with larger setups. This is subject to a HLRS/GCS project extension
that we recently applied for.
Dynamic molecular dynamics 437

1024
Ideal 4096 Ideal
512 lb1x1x1-md1x1x1 lb1x1x1-md1x1x1
256 lb1x1x1-md2x2x2 lb1x1x1-md2x2x2
lb2x2x2-md2x2x2 1024 lb2x2x2-md2x2x2
128 lb4x4x4-md2x2x2
64 256
Speedup

Speedup
128
32
64
16 32
8 16
4 8
4
2 2
11 2 4 8 16 32 64 128 256 512 1024 11 2 4 8 16 32 64 128256 1024 4096
Cores Cores

(a) MD30 (b) MD60

Fig. 4: Scalability of dynamic launch/removal for scenarios (a) MD30 and (b) MD60.

This work forms the ground for the ongoing research at Helmut-Schmidt-University
Hamburg to enable error estimation and fault tolerance in molecular-continuum
simulations, exploiting the MD ensemble approach.

Acknowledgements We thank HLRS and GCS for providing computational resources in the scope
of the project GCS-MDDC, ID 44130. We further thank HSU for supporting our project through the
HSU-internal research funding program (IFF), project “Resilience and Dynamic Noise Reduction at
Exascale for Multiscale Simulation Coupling”.

References

1. F. Diewald, M.P. Lautenschlaeger, S. Stephan, K. Langenbach, C. Kuhn, S. Seckler, H.-J.

Bungartz, H. Hasse, and R. Müller. Molecular dynamics and phase field simulations of
droplets on surfaces with wettability gradient. Computer Methods in Applied Mechanics and
Engineering, 361:112773, 2020.
2. F.A. Gratl, S. Seckler, H.-J. Bungartz, and P. Neumann. N Ways to Simulate Short-Range Particle
Systems: Automated Algorithm Selection with the Node-Level Library AutoPas. Submitted,
2020.
3. S. Grottel, M. Krone, C. Muller, G. Reina, and T. Ertl. Megamol – a prototyping framework
for particle-based visualization. Visualization and Computer Graphics, IEEE Transactions on,
21(2):201–214, 2015.
4. S. Hussain and A. Haji-Akbaria. Studying rare events using forward-flux sampling: Recent
breakthroughs and future outlook. Journal of Chemical Physics, 152:060901, 2020.
5. V. Jafari, N. Wittmer, and P. Neumann. Massively parallel molecular-continuum flow simulation
with error control and dynamic ensemble handling. International Conference on High
Performance Computing in Asia-Pacific Region, page 52–60, 2022.
438 P. Neumann, N. Wittmer, V. Jafari, S. Seckler and M. Heinen

6. P. Jarmatz, F. Maurer, and P. Neumann. MaMiCo: Non-Local Means Filtering with Flexible
Data-Flow for Coupling MD and CFD. Lecture Notes in Computer Science (ICCS 2021
proceedings), pages 576–589, 2021.
7. P. Jarmatz and P. Neumann. MaMiCo: Parallel Noise Reduction for Multi-Instance Molecular-
Continuum Flow Simulation. Lecture Notes in Computer Science (ICCS 2019 proceedings),
pages 451–464, 2019.
8. A. Köster, T. Jian, G. Rutkai, C. Glass, and J. Vrabec. Automatized determination of fundamental
equations of state based on molecular simulations in the cloud. Fluid Phase Equilibria, 425:84–
92, 2016.
9. T. Krüger, H. Kusumaatmaja, A. Kuzmin, O. Shardt, G. Silva, and E.M. Viggen. The Lattice
Boltzmann Method. Principles and Practice. Springer, 2016.
10. P. Neumann and X. Bian. MaMiCo: Transient Multi-Instance Molecular-Continuum Flow
Simulation on Supercomputers. Comput. Phys. Commun., 220:390–402, 2017.
11. P. Neumann, H. Flohr, R. Arora, P. Jarmatz, N. Tchipev, and H.-J. Bungartz. MaMiCo:
Software design for parallel molecular-continuum flow simulations. Comput. Phys. Commun.,
200:324–335, 2016.
12. X. Nie, S. Chen, W. E, and M. Robbins. A continuum and molecular dynamics hybrid method
for micro-and nano-fluid flow. J. Fluid Mech., 500:55–64, 2004.
13. C. Niethammer, S. Becker, M. Bernreuther, M. Buchholz, W. Eckhardt, A. Heinecke, S. Werth, H.-
J. Bungartz, C.W. Glass, H. Hasse, J. Vrabec, and M. Horsch. ls1 mardyn: The massively parallel
molecular dynamics code for large systems. Journal of Chemical Theory and Computation,
10(10):4455–4464, 2014.
14. S. Seckler, F. Gratl, M. Heinen, J. Vrabec, H.-J. Bungartz, and P. Neumann. AutoPas in ls1
mardyn: Massively Parallel Particle Simulations with Node-Level Auto-Tuning. Journal of
Computational Science, 50:101296, 2021.
15. D. Stephenson, J.R. Kermode, and D.A Lockerby. Accelerating multiscale modelling of fluids
with on-the-fly Gaussian process regression. Microfluidics and Nanofluidics, 22:139, 2018.
Scalable discrete algorithms for big data
applications

Demian Hespe, Lukas Hübner, Lorenz Hübschle-Schneider, Peter Sanders and

Dominik Schreiber†

Abstract In the past year, the project “Scalable Discrete Algorithms for Big Data
Applications” dealt with High-Performance SAT Solving, Malleable Job Scheduling
and Load Balancing, and Fault-Tolerant Algorithms. We used the massively parallel
nature of ForHLR II to obtain novel results in the areas of SAT solving and fault-
tolerant algorithms.

1 Introduction

Developing scalable algorithms is a challenging task that requires careful analysis

and extensive experimental evaluation. CPU technology shifts to deliver increasing
amounts of cores with relatively low clock rates, since these are cheaper to produce
and operate. Parallelizing computationally intensive algorithms at the core of all
applications is therefore an important research topic. Developing distributed algo-
rithms on cluster computers such as the ForHLR II is an integral part of this scalability
challenge. Our focus is especially on discrete algorithms, such as graph partitioning,
text search, and propositional satisfiability (SAT) solving.
In the past year, we used the ForHLR II cluster in the scope of three subprojects.
First, we designed and implemented a massively-parallel and distributed SAT solving
system [22, 24] (Section 2.1). The central novelty of our solver is a succinct and
communication-efficient exchange of information among the core solvers which helps
† All authors contributed equally and are sorted alphabetically.

Demian Hespe, Lukas Hübner, Lorenz Hübschle-Schneider, Peter Sanders and Dominik Schreiber
Institute for Theoretical Informatics: Algorithms II, Karlsruhe Institute of Technology (KIT), Am
Fasanengarten 5, Karlsruhe, 76131, Germany, e-mail: [email protected], [email protected],
[email protected], [email protected], [email protected]
Lukas Hübner
Computational Molecular Evolution Group, Heidelberg Institute for Theoretical Studies, Schloss-
Wolfsbrunnenweg 35, Heidelberg, 69118, Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 439
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_27
440 D. Hespe, L. Hübner, L. Hübschle-Schneider, P. Sanders and D. Schreiber

to speed up the solving process. Secondly, in order to improve resource efficiency and
to reduce scheduling times for the resolution of difficult SAT problems in cloud-like
HPC environments, we explored novel dynamic scheduling and load balancing
approaches (Section 2.2). For this means, we exploit malleability of tasks; that is, the
task’s capability of handling a fluctuating number of processing elements during its
execution. Thirdly, we explored parallel fault-tolerance mechanisms and algorithms,
the software modifications required, and the performance penalties induced via
enabling parallel fault-tolerance by example of RAxML-NG, the successor of the
widely used RAxML tool for maximum likelihood based phylogenetic tree inference
(Section 2.3).
In previous years, we studied distributed online sorting and string sorting in our Big
Data toolkit Thrill, developed a scalable approach to edge partitioning, developed and
evaluated algorithms for maintaining uniform and weighted samples over distributed
data streams (reservoir sampling), and designed new approaches to massively parallel
malleable job scheduling applied to propositional satisfiability (SAT) solving. To
conclude our project on scalable discrete algorithms in the scope of ForHLR II, we
provide a compact project retrospective in Section 6.

2 Scientific work accomplished and results obtained

2.1 High-Performance SAT Solving

Satisfiability (SAT) solving deals with one of the most famous N P-complete
problems and has many interesting applications in software verification, theorem
proving and automated planning. We designed and implemented a massively-parallel
and distributed SAT solving system [22, 24] which is also able to gracefully handle
fluctuating computational resources (see Section 2.2). The central novelty of our
solver is a succinct and communication-efficient exchange of information among the
core solvers which helps to speed up the solving process.
Experiments on up to 128 compute nodes of the ForHLR II showed that our
solving system named Mallob significantly outperforms its precursor HordeSat [2]
and shows much better scaling properties [24]. Fig. 1 shows a direct comparison
between HordeSat and our new solver Mallob whereas both make use of exactly the
same backend solvers. As HordeSat fails to scale beyond 32 nodes, Mallob on 32
nodes outperforms HordeSat on 128 nodes. Furthermore, Mallob can make effective
use of up to 128 nodes. We provide more detailed scaling results in Table 1. To
the best of our knowledge, the speedups achieved by Mallob are the best speedups
reported by any SAT solver in an HPC environment so far.
Our solver scored the first place in the first Cloud Track of the International SAT
Competition 2020 [1] where Mallob-mono was executed on 100 8-core nodes of an
AWS infrastructure and solved more formulae than any other solver in the competition.
Scalable discrete algorithms 441

HordeSat Mallob
70 70

60 60
# instances solved in ≤ t s

# instances solved in ≤ t s
128 × 5 × 4
50 50
32 × 5 × 4
40 40 8×5×4
2×5×4
30 30 1×3×4
Kissat
20 20
Lingeling
10 10

0 0
0 100 200 300 0 100 200 300
Run time t / s Run time t / s

Fig. 1: Scaling behaviour of HordeSat (with updated solvers) and untuned Mallob
compared to two sequential solvers [24].

Table 1: Parallel speedups for HordeSat (H) and Mallob (M). In the left half, “#”
denotes the number of instances solved by the parallel approach and 𝑆 𝑚𝑒𝑑 (𝑆𝑡𝑜𝑡 )
denotes the median (total) speedup for these instances compared to Lingeling / Kissat.
In the right half, only instances are considered for which the sequential solver took at
least (num. cores of parallel solver) seconds to solve. Here, “#” denotes the number
of considered instances for each combination [24].

All instances Hard instances

Lingeling Kissat Lingeling Kissat
Config. # 𝑆𝑚𝑒𝑑 𝑆𝑡𝑜𝑡 𝑆𝑚𝑒𝑑 𝑆𝑡𝑜𝑡 # 𝑆𝑚𝑒𝑑 𝑆𝑡𝑜𝑡 # 𝑆𝑚𝑒𝑑 𝑆𝑡𝑜𝑡

H1×3×4 36 3.84 51.90 2.22 29.55 32 4.39 52.01 31 4.03 32.49

H2×5×4 40 12.00 95.80 5.06 64.44 35 12.27 96.83 33 9.11 69.63
H8×5×4 49 22.83 135.55 9.76 90.08 38 32.00 142.76 32 24.88 105.94
H32×5×4 56 42.12 203.66 15.25 112.14 34 97.61 231.77 19 114.86 208.68
H128×5×4 59 50.35 204.10 17.38 111.46 21 356.33 444.12 10 243.42 375.04
M1×3×4 35 4.83 58.15 3.62 64.66 31 5.37 58.24 30 5.29 66.08
M2×5×4 44 12.98 94.44 10.52 67.71 39 14.37 95.28 37 11.54 69.25
M8×5×4 52 28.38 154.62 12.06 89.61 41 34.29 162.23 34 23.43 106.85
M32×5×4 60 53.75 220.92 23.41 148.57 37 152.19 245.54 23 134.07 262.04
M128×5×4 65 81.60 308.48 25.97 175.58 25 363.32 447.97 12 363.32 483.11

We submitted a new version of Mallob to the upcoming International SAT Competition

2021 as well [23] and are excited to see how our system performs this year. Mallob is
open-source and can be obtained from https://github.com/domschrei/mallob.
442 D. Hespe, L. Hübner, L. Hübschle-Schneider, P. Sanders and D. Schreiber

2.1.1 Usage statistics

The following measures also subsume Section 2.2, as the two subprojects are integrated
in a single software system and have been evaluated together.
In total, 675 219 TRES hours have been spent on the ForHLR II for experiments
and evaluations involving Mallob, which amounts to 71% of our project’s overall
usage in the reported time frame. We used twelve to 2560 cores (1 to 128 compute
nodes). As we are committed to responsible and resource-efficient experiments, we
identified a statistically significant selection of SAT instances on which we performed
most experiments as opposed to running all experiments on a much larger set of
benchmarks [24]. Furthermore, we limited the run time of parallel solvers to 300 s
per formula in most cases. For the experiments with sequential solvers, we scheduled
multiple solvers to run in parallel on a single compute node in order to make efficient
use of resources.

2.2 Malleable job scheduling and load balancing

In order to improve resource efficiency and to reduce scheduling times for the
resolution of difficult problems in cloud-like HPC environments, we explored novel
dynamic scheduling and load balancing approaches. For this means, we exploit
malleability of tasks: A malleable task is capable of handling a fluctuating number of
processing elements during its execution. We developed a novel system named Mallob
for the scheduling and load balancing of malleable tasks [24]. New jobs entering
the system are scheduled virtually immediately (mostly within tens of milliseconds)
and balanced through a fully decentralized load balancing protocol: Lightweight
asynchronous and event-driven message passing ensures a fair distribution of resources
according to the priorities and demands of active jobs. As mentioned in Section 2.1,
we developed a scalable distributed SAT solver which can handle malleability and
integrated it into our system as an exemplary application.
Experiments on up to 128 nodes on the ForHLR II [24] showed that our job
scheduling and load balancing framework imposes minimal computational and
communication overhead and dynamically assigns active processing elements to
jobs in a fair manner. Most jobs arriving in the system are initiated within tens of
milliseconds. In the context of our application, we showed that our system is able to
find an appealing trade-off between the “trivial” resource efficiency of solving many
formulae at once and the speedups of flexible parallel SAT solving. For instance, we
experimentally compared Mallob’s malleable processing of 400 SAT jobs with an
“embarrassingly parallel” processing using 400 sequential SAT solvers in parallel and
with a hypothetical optimal sequential scheduling (HOSS) where 400 corresponding
runs of 128-node Mallob-mono are sorted ascendingly by their run time. As Fig. 2
shows, Mallob with malleable scheduling outperforms any of these extremes and
achieves low response times and a high number of solved instances.
Scalable discrete algorithms 443

300

250
# solved instances

𝑅 𝑎𝑙𝑙 𝑅𝑠𝑙𝑣
200
Configuration avg. med. avg. med.
150 Mallob 𝐽 = ∞ 2422.4 679.8 808.6 260.6
Mallob, 400 jobs @ start
100 400×Kissat 400×Kissat 2998.4 1362.5 975.5 355.5

Sorted runs of Mallob-mono HOSS 2774.7 2024.5 1396.4 937.3

50
400×Lingeling 400×Lingeling 4436.0 7200.0 1559.2 819.9
0
0 2000 4000 6000
Total run time (s)

Fig. 2: Cumulative solved instances by different scheduling approaches on 128

compute nodes within two hours. The table shows average and median response
times, calculated for all 400 instances (𝑅 𝑎𝑙𝑙 ) and for the solved instances per approach
(𝑅𝑠𝑙𝑣 ). Each unsolved instance leads to a response time of 7200 s [24].

2.3 Fault-tolerance for massively parallel phylogenetic inference

Failing hardware is projected to be one of the main challenges in future exascale

systems [25]. In fact, one may expect that a hardware failure will occur in exascale-
systems every 30 to 60 min [5, 8, 26]. The reasons for failures are manifold: HPC
systems can fail due to core hangs, kernel panics, file system errors, file server failures,
corrupted memories or interconnects, network outages, air conditioning failures, or
power halts [12, 18].
We apply fault-tolerance to RAxML-NG [17], a phylogenetic tree inference tool.
Phylogeny is one of the central applications in bioinformatics. Given a set of aligned
input sequences, the task is to reconstruct the phylogenetic tree which maximizes the
(negative) log-likelihood of this tree showing the evolutionary relationships among
the input entities. Phylogenetic trees are now routinely inferred on large scale HPC
systems with thousands of cores as the parallel scalability of phylogenetic inference
tools has improved over the past years to cope with the molecular data avalanche [17].
Thus, the parallel fault-tolerance of phylogenetic inference tools has become a relevant
challenge. To this end, we explore parallel fault-tolerance mechanisms and algorithms,
the software modifications required, and the performance penalties induced via
enabling parallel fault-tolerance by example of the RAxML-NG software [17], the
successor of the widely used RAxML tool for phylogenetic inference [27]. We find
that the slowdown induced by the necessary additional recovery mechanisms in
RAxML-NG is on average 1.00 ± 0.04. The overall slowdown by using these recovery
mechanisms in conjunction with a fault-tolerant MPI implementation amounts to
on average 1.7 ± 0.6 for large empirical datasets [14]. Via failure simulations, we
show that RAxML-NG can successfully recover from multiple simultaneous failures,
subsequent failures, failures during recovery, and failures during checkpointing.
Recoveries are automatic and transparent to the user [14].
444 D. Hespe, L. Hübner, L. Hübschle-Schneider, P. Sanders and D. Schreiber

We developed, tested, and benchmarked this fault-tolerant version of RAxML-NG

(FT-RAxML-NG) on the ForHLR II. We used the results of these experiments in our
first submission of this work to the Oxford University Press Bioinformatics journal.
As the ForHLR II was then discontinued we had to use another HPC system for
further experiments as part of the revision.
In the scope of this project, we deployed and tested ULFM, a failure-mitigating
MPI implementation on the ForHLR II. This also included experiments to determine
the most reliable failure detection method and optimizing recovery times after a
failure. Various methods to simulate failures on the cluster without interfering with
other jobs or needing administrative intervention were evaluated [14].

2.3.1 Usage statistics

For the experiments described in Section 2.3 we used 20 to 400 cores. We did some
preliminary tests with large jobs using 512 nodes. As we saw runtime fluctuations of
over 300 % in these tests and each job took around two weeks to be scheduled, we
decided against using the ForHLR II for such large jobs.
RAxML-NG supports parallelization at three levels. At the single thread level, it
uses parallelism as provided by the x86 vector intrinsics (SSE3, AVX, AVX2). At
the single node level, RAxML-NG leverages the available cores by parallelization
using PThreads. If we run RAxML-NG on a distributed memory HPC system, it uses
parallelization via message passing (using MPI) [19, 28]. We can enable all three
levels of parallelism at the same time.
As this was not the focus of this work, we did not perform any scaling experiments
using FT-RAxML-NG.

3 Publications with project results

The technical report describing our winning submission of our SAT solving system
(Section 2.1) to the Cloud Track of the SAT Competition 2020 was published
in the Proceedings of SAT Competition 2020 [22]. A full paper introducing the
Mallob system for job scheduling and load balancing (Section 2.2) and analyzing the
scalability of our SAT solving system has been accepted at the 24th International
Conference on Theory and Applications of Satisfiability Testing [24]. We are currently
in the process of editing another publication which will describe our scheduling and
load balancing approaches in greater detail. Our new version of Mallob (Section 2.2)
competes in the upcoming International SAT Competition 2021 [23]. Our work on
fault-tolerant phylogenetic inference (Section 2.3) has been accepted for publication
at the Bioinformatics Journal [14].
Scalable discrete algorithms 445

4 Ongoing theses within the project

We are supervising a promising ongoing master thesis project that used the ForHLR
II cluster:

4.1 Low overhead fault-tolerant MapReduce for HPC clusters

The MapReduce framework [7] is a popular solution for executing tasks in parallel
among multiple machines by formulating the task using map and reduce operations.
The framework then takes care of parallelization, load balancing and fault-tolerance, i.e.
continuing work when one of the machines used stops working. In the past years there
have been several implementations of MapReduce both for cloud computing [7, 9, 13]
and for HPC clusters using MPI [10, 11, 20]. However, fault-tolerance is usually
implemented by storing checkpoints to a distributed file system [7, 9, 11, 13] or is
omitted entirely [10, 20]. The master thesis of Charel Mercatoris aims at developing a
MapReduce implementation in MPI that achieves fault-tolerance by storing all relevant
information redundantly in memory: During normal computation, all messages sent
remain stored on the sending machine. Additionally, all information that would be
lost in case of a failure is sent to a different machine as backup. After a failure,
only the most recent map and reduce functions have to be re-executed only on the
data that resided on the machine that stopped working. Due to the type of data flow
defined through the map and reduce functions, we only have to store redundant
data for one round of map and reduce functions. As soon as the next backup-cycle
is complete, the previous backup data can be discarded. Preliminary experiments
have been performed on the ForHLR-II cluster and are now continued on a different
system since ForHLR-II is no longer available. Figure 3 shows preliminary scaling
experiments executed on ForHLR-II. We can see that the performance and scaling
behavior virtually does not change when activating fault-tolerance mechanisms (like
storing backup data redundantly). Even when we simulate a failure of 10% of all MPI
ranks, we only observe a small slowdown.

5 Theses completed within the project

The master’s thesis of Lukas Hübner looked into load-balance and fault-tolerance for
massively parallel phylogenetic inference. After completion of this Master’s Thesis
we continued to look into this topic; our work is described in Section 2.3.
The dissertation of Lorenz Hübschle-Schneider [15] considers communication-
efficient probabilistic algorithms for three fundamental Big Data problems: selection,
sampling, and checking. ForHLR II was used to evaluate the weighted reservoir
sampling algorithm, a batched distributed streaming algorithm for maintaining a
 Page Rank: grid graph x=4096, y=4096 p=50.0%
446 D. Hespe, L. Hübner, L. Hübschle-Schneider, P. Sanders and D. Schreiber

1000

500

300
200
time (s)

100

50

30
20

10
20 40 80 160 320
# CPUs

Fig. 3: Preliminary scaling experiments of our MapReduce implementation calculating

PageRanks on a random grid graph with 224 vertices. no FT: No fault-tolerant
mechanisms enabled FT: Fault-tolerance mechanisms enabled but no faults occurred.
FT + 10% failure: Fault-tolerance mechanisms enabled and the failure of 10% of the
MPI ranks is simulated.

weighted random sample over an input that arrives over time. The results and which
we previously described in last year’s report, show good scalability on up to 256
nodes [16].

6 Project retrospective

In the following, we conclude our report by briefly reiterating the most notable project
results which we achieved with the help of ForHLR II over the past years.

• Thrill – A High-Performance Big Data Framework in C++. We improved

the sorting algorithm of Thrill [6], our next-generation C++ framework for
distributed Big Data batch processing on a cluster of homogeneous machines
which enables writing distributed applications conveniently using “dataflow”
graph-like computations [3].
• Graph Partitioning. We developed a fast distributed graph construction algo-
rithm and showed that combining it with advanced parallel node partitioning
algorithms yields high-quality edge partitions scaling to networks with billions
of edges on thousands of processors on large real-world networks [21].
Scalable discrete algorithms 447

• Communication-Efficient Distributed String Sorting. We proposed two new

distributed algorithms for string sorting, Distributed String Merge Sort (MS) and
Distributed Prefix-Doubling String Merge Sort (PDMS); experiments on up to
1280 cores revealed that the new algorithms are often more than five times faster
than previous algorithms [4].
• Reservoir Sampling. We developed algorithms for uniform and weighted
reservoir sampling in a distributed mini-batch model which scaled up to 256 nodes
(5120 cores) and handily outperformed a centralized comparison algorithm [16].
• Malleable Job Scheduling and Load Balancing. We developed a novel system
named Mallob for the scheduling and load balancing of malleable tasks [24],
imposing minimal computational and communication overhead, dynamically
assigning active processing elements to jobs fairly and scheduling most jobs
within tens of milliseconds on 128 nodes.
• High-Performance SAT Solving. We designed and implemented a massively
parallel and distributed SAT solving system [22, 24] which won the Cloud Track
of the International SAT Competition 2020 [1] and which enables our Mallob
system to resolve SAT jobs in a scalable manner.
• Fault-Tolerance for Massively Parallel Phylogenetic Inference. We applied
fault-tolerance to RAxML-NG[17], a phylogenetic tree inference tool, and
showed that the resulting software can successfully recover from multiple
simultaneous failures, subsequent failures, failures during recovery, and failures
during checkpointing [14].
For the mentioned subprojects, ForHLR II proved to be an invaluable resource for
testing and scientifically evaluating our software. As such, we close with thanking
the maintainers and the steering committee of the ForHLR II cluster for their kind
support.

References

1. Tomáš Balyo, Nils Froleyks, Marijn JH Heule, Markus Iser, Matti Järvisalo, and Martin Suda.
SAT competition, 2020. Accessed: 2021-03-19.
2. Tomáš Balyo, Peter Sanders, and Carsten Sinz. Hordesat: A massively parallel portfolio SAT
solver. In International Conference on Theory and Applications of Satisfiability Testing, pages
156–172. Springer, 2015. Preprint arXiv:1505.03340 [cs.LO].
3. Timo Bingmann, Michael Axtmann, Emanuel Jöbstl, Sebastian Lamm, Huyen Chau Nguyen,
Alexander Noe, Sebastian Schlag, Matthias Stumpp, Tobias Sturm, and Peter Sanders. Thrill:
High-performance algorithmic distributed batch data processing with C++. In 2016 IEEE
International Conference on Big Data, pages 172–183. IEEE, 2016. Preprint arXiv:1608.05634
[cs.DC].
4. Timo Bingmann, Peter Sanders, and Matthias Schimek. Communication-Efficient String
Sorting. In 34th IEEE International Parallel and Distributed Processing Symposium (IPDPS).
IEEE, May 2020. to appear, preprint arXiv:2001.08516 [cs.DC].
5. Franck Cappello, Al Geist, William Gropp, Sanjay Kale, Bill Kramer, and Marc Snir. toward
exascale resilience: 2014 update. Supercomputing Frontiers and Innovations, 1(1), September
2014.
448 D. Hespe, L. Hübner, L. Hübschle-Schneider, P. Sanders and D. Schreiber

6. Jonas Dann. improving distributed external sorting for big data in Thrill. Master Thesis.
Karlsruhe Institute of Technology, Germany, September 2019.
7. Jeffrey Dean and Sanjay Ghemawat. MapReduce: simplified data processing on large clusters.
In Eric A. Brewer and Peter Chen, editors, 6th Symposium on Operating System Design and
Implementation, pages 137–150. OSDI, 2004.
8. Jack Dongarra, Thomas Herault, and Yves Robert. Fault tolerance techniques for high-
performance computing. https://www.netlib.org/lapack/lawnspdf/lawn289.pdf,
2015.
9. Jaliya Ekanayake, Hui Li, Bingjing Zhang, Thilina Gunarathne, Seung-Hee Bae, Judy Qiu, and
Geoffrey Fox. Twister: a runtime for iterative mapreduce. In Proceedings of the 19th ACM
international symposium on high performance distributed computing, pages 810–818, 2010.
10. Tao Gao, Yanfei Guo, Boyu Zhang, Pietro Cicotti, Yutong Lu, Pavan Balaji, and Michela Taufer.
Mimir: memory-efficient and scalable MapReduce for large supercomputing systems. In 2017
IEEE international parallel and distributed processing symposium (IPDPS), pages 1098–1108.
IEEE, 2017.
11. Yanfei Guo, Wesley Bland, Pavan Balaji, and Xiaobo Zhou. Fault-tolerant MapReduce-MPI for
HPC clusters. In Proceedings of the International Conference for High Performance Computing,
Networking, Storage and Analysis, pages 1–12, 2015.
12. Saurabh Gupta, Tirthak Patel, Christian Engelmann, and Devesh Tiwari. Failures in large scale
systems. In Proceedings of the International Conference for High Performance Computing,
Networking, Storage and Analysis, pages 1–12, November 2017.
13. Apache Hadoop website. https://hadoop.apache.org/. Accessed: 2021-06-24.
14. Lukas Hübner, Alexey M. Kozlov, Demian Hespe, Peter Sanders, and Alexandros Stamatakis.
Exploring Parallel MPI Fault-Tolerance Mechanisms for Phylogenetic Inference with RAxML-
NG.
15. Lorenz Hübschle-Schneider. communication-efficient probabilistic algorithms: selection,
sampling, and shecking. PhD thesis, Karlsruher Institut für Technologie (KIT), 2020.
16. Lorenz Hübschle-Schneider and Peter Sanders. brief announcement: communication-efficient
weighted reservoir sampling from fully distributed data streams. In 32nd ACM Symposium on
Parallelism in Algorithms and Architectures (SPAA), pages 543–545, 2020. extended preprint:
arXiv:1910.11069 [cs.DS].
17. Alexey M. Kozlov, Diego Darriba, Tomáš Flouri, Benoit Morel, and Alexandros Stamatakis.
RAxML-NG: a fast, scalable and user-friendly tool for maximum likelihood phylogenetic
inference. Bioinformatics, 35(21):4453–4455, May 2019.
18. Charng-Da Lu. failure data analysis of HPC systems. Computer Science, February 2013.
19. Wayne Pfeiffer and Alexandros Stamatakis. hybrid MPI/Pthreads parallelization of the RAxML
phylogenetics code. In 2010 IEEE International Symposium on Parallel & Distributed
Processing, Workshops and PhD Forum (IPDPSW), pages 1312–1313, 2010.
20. Steven J. Plimpton and Karen D. Devine. MapReduce in MPI for large-scale graph algorithms.
Parallel Computing, 37(9):610–632, 2011.
21. Sebastian Schlag, Christian Schulz, Daniel Seemaier, and Darren Strash. Scalable edge
partitioning. In 2019 Proceedings of the Twenty-First Workshop on Algorithm Engineering and
Experiments (ALENEX), pages 211–225. SIAM, 2019.
22. Dominik Schreiber. Engineering HordeSat towards malleability: mallob-mono in the SAT 2020
cloud track. SAT COMPETITION 2020, page 45, 2020.
23. Dominik Schreiber. Mallob in the SAT competition 2021. SAT COMPETITION 2021, 2021.
38–39.
24. Dominik Schreiber and Peter Sanders. Scalable SAT solving in the cloud. In International
Conference on Theory and Applications of Satisfiability Testing, pages 518–534. Springer, 2021.
25. John Shalf, Sudip Dosanjh, and John Morrison. Exascale computing technology challenges. In
Lecture Notes in Computer Science, pages 1–25. 2011.
26. Marc Snir, Robert Wisniewski, Jacob Abraham, Sarita Adve, Saurabh Bagchi, Pavan Balaji,
Jim Belak, Pradip Bose, Franck Cappello, Bill Carlson, Andrew Chien, Paul Coteus, Nathan
DeBardeleben, Pedro Diniz, Christian Engelmann, Mattan Erez, Saverio Fazzari, Al Geist,
Scalable discrete algorithms 449

Rinku Gupta, Fred Johnson, Sriram Krishnamoorthy, Sven Leyffer, Dean Liberty, Subhasish
Mitra, Todd Munson, Rob Schreiber, Jon Stearley, and Eric Van Hensbergen. Addressing
failures in exascale computing. The International Journal of High Performance Computing
Applications, 28(2):129–173, March 2014.
27. Alexandros Stamatakis. RAxML version 8: a tool for phylogenetic analysis and post-analysis of
large phylogenies. Bioinformatics, 30(9):1312–1313, January 2014.
28. Alexandros Stamatakis, T. Ludwig, and H. Meier. Computing large phylogenies with sta-
tistical methods: Problems & solutions. In Proceedings of 4th International Conference on
Bioinformatics and Genome Regulation and Structure (BGRS2004), 2014.
 Miscellaneous Topics

The Miscellaneous Topics section documents, besides the beauty, the breadth of
numerical simulations. It supports research in many fields other than just the topics
like fluid dynamics, aerodynamics, structure mechanics, and so forth. The following
five articles show that today’s computational approaches are by far not complete from
a numerical or from a modeling point of view. However, the physically correct focus
will lead to reliable prediction methods derived from data-driven or physics-driven
approaches to substantiate new theories.
The report of the Goethe Universität Frankfurt is a first performance analysis of
the software framework UG4 on the Hawk Apollo supercomputer. In Computational
Science and Engineering (CSE) the access to high-end computational systems
naturally sparks the desire to match the computational resources and to solve larger
problems. Overall increasing complexity requires scalable and efficient computational
methods for very large problems. Solving intricate applications requires a flexible
and robust software infrastructure. In a classic setting the numerical solution of
transient PDEs is obtained in a pipeline including the following steps. The governing
equations are discretized in time and space. In a next step, the problem is linearized,
before a resulting linear system can be solved. This strategy is incorporated in many
open source projects for general purpose simulations. In the article, the focus is on
the software UG4. A scaling study is provided for benchmark problems on Hawk.
Moreover, results for a thermohaline flow problem are discussed.
The contribution from the Technische Universität Berlin addresses scaling issues
related to the field of molecular dynamics simulations focusing on computational
details. In the beginning, finite size effects in the context of multicomponent diffusion
are investigated. Different methods to correct the influence of the finite simulation
domain are compared. Next, the structure of a fluid near its critical point is discussed
in the context of the strong scaling behavior. Subsequently, droplet coalescence
dynamics determined by large molecular dynamics simulations and a macroscopic
phase field model is analyzed. The performance of the new supercomputer Hawk is
compared to that of Hazel Hen. Finally, the influence of the direct sampling of the
energy flux on the performance of the code ls1 mardyn is described. Various speed
tests are carried out and analyzed in detail.

451
452 Miscellaneous Topics

The third article is a joint contribution of the Ruhr Universität Bochum and
the Universität Hamburg. The topic is scalable multigrid for fluid dynamics shape
optimization. To be more precise, a parallel approach for the shape optimization of
an obstacle in an incompressible Navier–Stokes flow is investigated. A self-adapting
nonlinear extension equation within the method of mappings is used, which links
a boundary control to a mesh deformation. It is demonstrated how the approach
preserves mesh quality and allows for large deformations by controlling nonlinearity
in critical regions of the domain. Special focus is on reference configurations, where
the transformation has to remove and create obstacle boundary singularities. This
is particularly relevant for the employed multigrid discretizations. Aerodynamic
drag optimizations for 2d and 3d configurations define benchmark problems. The
efficiency of the algorithm is demonstrated in weak scalability studies performed on
the supercomputer HPE Hawk for up to 5,120 cores.
The fourth article “Numerical calculation of Lean-Blow-Out (LBO) of a premixed
and turbulent burner consisting of an array of jet flames” comes from the Engler-
Bunte-Institute of Combustion Technology, Karlsruhe Institute of Technology. Swirl-
stabilized flames have been the preferred choice in gas turbine combustors due to
their aerodynamic means of stabilization. However, swirl flames can be susceptible
to the issue of combustion instability in the lean combustion regime leading to
Lean-Blow-Out and therefore damaging the hardware. The main task of this project
is improving the low load capability of stationary gas turbines.
The focus is on new numerical calculations of the Lean-Blow-Out limit of
a premixed and turbulent burner consisting of an ensemble of jet flames using
large eddy simulations and two different approaches of modeling the turbulent
flame interaction. The solver used for calculating the reactive flow is based on an
OpenFOAM solver. The turbulent-chemistry interaction are treated by two different
combustion models. In the first model the calculation of the source term is done via a
density function of the reaction progress variable. The second model calculates the
source term on the basis of the turbulent flame speed.
Simulations are validated through corresponding experiments. The numerically
calculated Lean-Blow-Out limit is in good agreement with the experimental values
In the fifth contribution “Data-driven multiscale modeling of self-assembly and
hierarchical structural formation in biological macro-molecular systems” the authors
present and evaluate a multiscale modelling framework called the “Molecular Discrete
Element Method” where interaction potentials are calculated via “Universal Kriging”
from fine scale molecular dynamics simulations. The approach is tested for the
hepatitis B viral protein HBcAg.

Institute of Aerodynamics, Wolfgang Schröder

RWTH Aachen University,
Wüllnerstr. 5a,
52062 Aachen,
Germany,
e-mail: [email protected]
Large scale simulations of partial differential
equation models on the Hawk supercomputer

Ruben Buijse, Martin Parnet and Arne Nägel

Abstract This work presents a first performance analysis of the software framework
UG4 on the Hawk Apollo supercomputer. The software demonstrated excellent
scaling properties before. It has now been demonstrated that this also holds true for
HLRS’s new flagship and its architecture. Three aspects are emphasized: (i) classic
weak scaling of a multi-grid solver for the Poisson equation, (ii) strong scaling for the
heat equation using multi-grid-in-time, and (iii) application to a thermo-haline-flow
problem in a fully-coupled fashion.

1 Introduction

Today, access to high-end computational systems is widely spread and available

in ever increasing amounts. In Computational Science and Engineering (CSE) this
naturally sparks the desire to match the computational resources and to solve larger
and larger problems in the applications. Overall complexity increases not only with
the mesh size, but also with the level of interactions and couplings between the
physical unknowns. This requires scalable and efficient computational methods for
very large problems.
Solving complex applications requires a flexible and robust software infrastructure.
In a classic setting the numerical solution of transient PDE is obtained in a pipeline
including the following steps: starting from the governing equations, (i) the problem is
discretized in time and space first. In a next step (ii) the problem is linearized, before
(iii) a resulting linear system can be solved. This strategy is incorporated in many open
source projects for general purpose simulations such as deal.II [1, 18], Dune [3, 4],
DuMu 𝑥 [9], FEFLOW [6], FEniCS[8, 20], IPARS [16], the Matlab Reservoir Toolbox
(MRST) [22], or OpenGeoSys [19]. In this work, we focus on the software UG4

Arne Nägel
Goethe-Universität Frankfurt, G-CSC, Kettenhofweg 139, 60325 Frankfurt,
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 453
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_28
454 Ruben Buijse, Martin Parnet and Arne Nägel

[25, 30]. We provide a scaling study for benchmark problems on Hawk. Moreover,
we provide results for a thermohaline flow problem. This PDE couples fluid flow,
substance transport and heat transport and is an example for a complex PDE problem
solved using parallel computing.

2 Methods

2.1 Multilevel solvers

For a large-scale simulations, the solution of the linear systems, turns out to be
the limiting factor. In order to benefit from high performance computing, this step
is performed using multigrid methods, as introduced in greater detail, e.g. in the
monographs [14, 29]. These methods have proven to be highly scalable to 100,000
processes and beyond, e.g. [2, 11, 12, 17, 25, 25, 27]. Details of the MPI based
parallelization of UG4’s multigrid solver, highly scalable on hundreds of thousands
of processes, have been also been described in previous reports, e.g. [15, 26]. Results
on Hawk are provided in Section 3.1.

2.2 Time integration

Classic time stepping by one-step or multi-step-methods is a serial procedure. On

modern and future architectures, clock speeds are stagnant and speedups will be
likely only be available through greater concurrency. Hence researchers face a serial
time-integration bottleneck. To that end, parallel in time methods, which have been
developed for more than 50 years, recently have regained popularity [10] provides
an overview of existing methods. An attractive approach is the multigrid in time
method (MGRIT) [7]. This has been implemented in form of the XBraid library [21],
which has been coupled with UG4 successfully [24]. This approach is investigated in
Section 3.2.

2.3 Hawk Apollo supercomputer

The Hawk Apollo is a 26 Petaflop supercomputer at HLRS Stuttgart 1. It consists of

5,632 compute nodes which are connected by an InfiniBand HDR200 interconnect
network in a 9D-Hypercube. Each node consists of two AMD EPYC™ 7742 processors.
This provides 2 × 64 = 128 cores per node, and a total of 720, 896 cores. The structure
of the nodes is hierarchical: 4 cores form a core complex (CCX) sharing a 16 MB L3

1 https://kb.hlrs.de/platforms/index.php/HPE_Hawk
Large scale simulations on the Hawk supercomputer 455

cache 2. 2 CCXs are combined to a cluster complex die (CCD), which forms a group
of cores with a common interface to the I/O die. Each node is equipped with 256
GB DRAM memory. This structure will be investigated in the performance results
later. All results are obtained using the aocc/clang compiler (version 2.1.0/9.0.0)
in combination with the mpt-toolkit (version 2.23) and mkl-library (version 19.1.0).

3 Results

In this section we present results for three different simulation scenarios in 3D space:
• Weak scaling study for Poisson’s equation,
• Strong scaling study for the heat equation for MGRIT,
• Results for a thermo-haline transport problem.

3.1 Weak scaling for Poisson’s equation

We use a classic setup that has been used before[25]: Poisson’s equation has been is
solved on the unit cube with Dirichlet boundary conditions. The equation is discretized
using finite elements on a structured hexagonal grid. Afterwards, the system is solved
with a geometric multigrid solver. Damped Jacobi is used as a smoother, with 1
pre-/post smoothing step in a V-cycle respectively. Times for three different phases
have been measured:
• The assemble-phase computes the finite element stiffness matrix. This is
primarily dominated by computation, but memory access is required for reading
element information and writing element stiffness matrix into a sparse-matrix
format. No communication via MPI is taking place in this phase.
• The init-phase refers to the setup of the multilevel solver. Here the commu-
nication interfaces are constructed. Moreover, data for a proper application of
matrix-vector operations is exchanged.
• The apply-phase involves primarily computations on vectors and sparse-matrix
data structures. MPI communications occurs along vertical and horizontal process
interfaces. For this phase, timings for a fixed number of five BiCGStab-sweeps
are reported. This roughly corresponds to a reduction of the residual by eight
orders of magnitude.
In a classic weak scaling, the number of cores and the number of degrees of
freedom is increased simultaneously. Starting from a 3D grid with 2 × 2 × 2 cubes, the
grid has been refined uniformly. This results in an eightfold increase of the number

2 https://kb.hlrs.de/platforms/upload/Processor.pdf
456 Ruben Buijse, Martin Parnet and Arne Nägel

of degrees of freedom, which is matched by an increase of number of cores by a

factor of eight. This strategy allows for an optimal distribution among processes. (cf.
Tab. 1).
The scaling study starts with 8 cores on refinement level 6, which corresponds
to roughly 250, 000 degrees of freedom per core3. On the levels ≤ 5, the number of
processes is limited. The first two levels are stored on a single process. Afterwards,
the number of increased processes is increased by a factor of 64 after every second
refinement. This strategy limits the number of processes a single process needs to
communicate with to a maximum of 64.

Table 1: Number of degrees of freedom.

Level Processes used Degrees of freedom Process limit

0 n.a. 27 1
1 n.a. 125 1

2 n.a. 729 ≤ 64
3 n.a. 4,913 ≤ 64

4 n.a. 35,937 ≤ 4, 096

5 n.a. 274,625 ≤ 4, 096

6 8 2,146,689 none
7 64 16,974,593
8 512 135,005,697
9 4,096 1,076,890,625
10 32,768 8,602,523,649
11 262,144 68,769,820,673

3.1.1 Standard process assignment

In the first test, we consider a fixed number 𝑚 of cores per node. If 𝑛 denotes the
number of nodes, the number of processes is given by 𝑛 ∗ 𝑚. This corresponds to the
setting select=n:node_type=rome:mpiprocs=m .
As shown in Tab. 2, the wallclock times in the range 64–32,768 processes are
almost constant. This is valid for all three code regions (assemble, init, apply).
When only eight cores are used, the times reduce slightly. and almost identical in all

3 One core corresponds to one MPI process.

Large scale simulations on the Hawk supercomputer 457

Table 2: Standard process assignment (w/o stride): Weak scaling with 16, 32, 64, and
128 MPI processes per node (ppn).

16-ppn 8 64 512 4,096 32,768 262,144

assemble 2.403 2.463 2.548 2.572 3.514 𝑎

init 0.699 0.975 0.986 1.091 1.606

apply 1.783 2.932 2.987 2.982 3.072

32-ppn 8 64 512 4,096 32,768 262,144

assemble 2.417 2,757 2,705 2,675 3,397 𝑎

init 0.702 0,931 1,021 1,081 1,251

apply 1.789 2.931 2.989 3.013 3.085

64-ppn 8 64 512 4,096 32,768 262,144

assemble 2.418 3.311 2.963 3.293 3.307 𝑏

init 0.686 0.956 1.063 1.044 1.100

apply 1.780 2.945 3.010 3.003 3.093

128-ppn 8 64 512 4,096 32,768 262,144

assemble 2.394 3.388 3.402 3.436 3.482 4.803

init 0.715 0.979 1.056 1.004 2.186 2.631
apply 1.796 2.944 3.006 2.997 3.224 3.281
𝑎 A job with this configuration exceeds the maximum number of 5,632 compute nodes.
𝑏 Calls of mpirun terminated In this configuration before the custom executable could be launched.

four cases, which can be expected. One exception is the case with fully populated
nodes (128 cores per node): Here the run time for 32,768 processes is slightly
increased. When 262,144 processes are used, time increases significantly again.
These results must be taken with a grain of salt and require further analysis and
detailed profiling.

3.1.2 Topology aware process assignment

As mentioned earlier, the Rome nodes feature an internal hierarchical structure. As

four cores share a common L3 cache, collisions in the previous test are very likely. In
an additional test, we tried to address the topology of the nodes by enforcing a limited
458 Ruben Buijse, Martin Parnet and Arne Nägel

number of MPI processes per node. Thus, we repeat the previous experiment using
only every eighth, fourth and second core, respectively. The distribution is achieved by
the commands omplace -c 0-127:st=s with a stride 𝑠 = 8, 4, 2. This corresponds
to using 2 (𝑠 = 2) or 1 (𝑠 = 4) cores per CCX, and 1 core per CCD (𝑠 = 8), respectively.

Table 3: Topology aware process assignment: Weak scaling with 16, 32, and 64 MPI
processes per node (ppn). The stride was 𝑠 = 8, 4, and 𝑠 = 2 respectively.

16-ppn
(𝑠 = 8) 8 64 512 4096 32,768

assemble 2.254 2.276 2.292 2.480 4.000

init 0.521 0.527 0.533 0.533 0.621
apply 0.605 0.618 0.627 0.642 0.705

32-ppn
(𝑠 = 4) 8 64 512 4096 32,768

assemble 2.330 2.287 2.290 2.375 2.946

init 0.559 0.555 0.622 0.677 0.887
apply 0.796 0.796 0.808 0.881 0.903

64-ppn
(𝑠 = 2) 8 64 512 4096 32,768

assemble 2.387 2.585 2.601 2.755 2.882

init 0.660 0.717 0.718 0.769 0.824
apply 1.418 1.437 1.463 1.478 1.558

As shown in Tab. 3, the impact of the topology aware process assignment w.r.t the
code region:
• For the assemble-phase, a small acceleration of about 10% is observed. This
phase includes a few read/write memory accesses, but is primarily computation
dominated. In addition, no communication is involved. Hence, this result can be
expected.
• In the init-phase communication becomes more important. In this case a slightly
more pronounced acceleration is observed, and run times reduce by 20-30%.
Large scale simulations on the Hawk supercomputer 459

• The biggest gains are achieved in the apply-phase. This phase features many
memory accesses to vector and sparse-matrix data structures. Moreover, the
network is used heavily for all-reduce operations ( e.g., for stopping criteria)
as well as nearest neighbor communication (e.g., master-slave-exchange along
the process boundaries). The acceleration approaches the optimal factors, i.e., 2
when using 2 out of 4 cores per CCX (𝑠 = 2), and 4 when using only a single
core per CCX (𝑠 = 4). Restricting the resources to a single core per CCD (𝑠 = 8)
only, yields additional acceleration.

3.1.3 Summary

For the benchmark problem, weak scaling capabilities have been demonstrated for a
wide range of MPI processes. The results show, that using all available 128 cores
per node is reasonable and economic. In our experiments, using a reduced number
of processes per node, only provides extra benefit in the apply-phase. At the same
time, the number of allocated nodes grows and resources by a factor of four. Hence,
this is not a very economic alternative. From a practical perspective, this should be
considered only when small wall-clock timings are crucial.

3.2 Strong scaling for the heat equation

The previous section demonstrated the excellent weak scaling properties of the
multilevel method for a steady state problem. In this case, the grid resolution in space
tends to zero. However, in particular for some transient problems, it is sometimes
desirable to work with a spatial grid with a given resolution, which corresponds to a
fixed number of degrees of freedom in space.
The following test, which was first suggested by [7], is an extension of the tests
conducted in [24] on a small cluster in 2D. We consider the heat equation

𝜕𝑡 𝑢(x, 𝑡) − 𝛼△𝑢(x, 𝑡) = 𝑓 (x, 𝑡)

for (x, 𝑡) ∈ (0, 1) 3 × (0, 4𝜋) and 𝛼 = 0.1. Initial value, Dirichlet boundary conditions
and right hand side 𝑓 are chosen such that the solution

𝑢(x, 𝑡) = sin(𝜋𝑥1 ) · sin(𝜋𝑥2 ) · sin(𝜋𝑥3 ) · cos(𝑡)

is obtained. The equation is discretized using an implicit Euler method in time and
𝑄 1 finite elements in space. Assuming that 𝑝 MPI processes are available, two setups
are compared:
• In serial time integration, all processes are dedicated to the spatial domain. We
consider the sequence 𝑝 ∈ {8, 64, 512}. The time domain is split into 16384
equidistant intervals that are treated sequentially.
460 Ruben Buijse, Martin Parnet and Arne Nägel

• The parallel-in-time integration uses 𝑝 = 𝑝 𝑠 × 𝑝 𝑡 ∈ {512, . . . , 16384} processes.

A fixed number of 𝑝 𝑠 = 64 processes is dedicated to spatial domain. For the time
domain a variable number 𝑝 𝑡 ∈ {8, . . . , 256} is used. Each process then owns an
interval with 16, 384 : 𝑝 𝑡 equidistant time points.
On the finest level, the unit cube Ω = (0, 1) 3 is split into 128 × 128 × 128 elements,
with 2, 146, 689 degrees of freedom. Splitting this onto eight processes, this means
that workload is in a similar order of magnitude as in the previous example for
the steady state. Note, however, that due to the fixed time step size, assembling the
operator is only required once. Instead of a BiCGStab, a linear iteration is used. On
each node, every second core has been used (cf. previous section with a stride 𝑠 = 2).
Results are provided in Fig. 1: We observe that for serial time stepping the strong
scaling w.r.t the spatial variables is limited: From eight to 64 processes, a speedup of
≈ 5.5 is observed, which is already smaller than the optimal value of eight. For 512
processes the run time even deteriorates. In this case, there is only roughly 4, 000
degrees of freedom per process left, which means that computation is only of little
importance, but communication becomes the dominant factor. Since the number of
process increases, more communication is required.
The MGRIT method initially suffers from the computational overhead of being an
iterative method. Several iterations are required to achieve the discretization error of
the serial method. To that end, observe that using 512 processes for 𝑝 𝑡 = 8 = 512/64
equidistant time intervals roughly takes the same time as the serial computation using
𝑝 = 8 processes. However, from that point on, adding more processes in the time
dimension always lead to a speedup. Doubling the number of processes in each step,
the observed speedup is 1.88 ≈ 2 initially, but still 1.52 in the last step. With respect
to wall-clock time, the break even point is roughly at 4, 096 cores. From that point
on, MGRIT is faster than the fastest run with serial time integration.
The results are in agreement with results [7, 24]. However, problems of the size
investigated here were not computed, but only predicted. It is important to emphasize
that the MGRIT is not attractive as long as a scaling from Section 3.1 is to be
sought. However, when all solutions w.r.t. the temporal dimension, are to be computed
simultaneously, as is, e.g. important for optimization, multigrid-in-time can provide
an attractive alternative which is not intrusive w.r.t. software design.

3.3 Thermohaline transport

The previous two sections focused primarily on scaling properties of the algorithms
for select models problems. In addition to the parallel capabilities, one major benefit of
UG4 is, however, the versatile applicability for a wide range of problems. in particular
for coupled PDE problems. This includes, e.g., density-driven flow, computational
neuro-science, poroelasticity.
To that end, this report is concluded by results for a thermohaline flow problem.
In this special instance of density-driven flow, where the fluid density is modeled
as a variable depending on both temperature and fluid composition. Hot fluids, e.g.,
Large scale simulations on the Hawk supercomputer 461

100000
SERIAL
MGRIT-FACTOR2

10000
wall clock
time [s]

1000

100
1 8 64 512 4096 32768
#processes

Fig. 1: Strong scaling for heat equation. Comparison of a classic serial time integration
(SERIAL) and multigrid in time (MGRIT-FACTOR2) with factor of 2 coarsening.

typically have a lower density than cold fluids. Similarly in solutions, an increase of
solute concentration yields a higher density. Systems of this type are important, e.g.,
for modeling transport of CO2 or NaCl in repositories and deep geological layers.
In this work, we concentrate on a parcel benchmark problem suggested in [13]:
In this test, a hot parcel with saline solution is placed in a rock matrix of lower
temperature. The behavior of the system then depends on the configuration of the
parcel, e.g., its temperature 𝑇 𝑝 and its salt mass fraction 𝜔 𝑝 . As outlined above, for
very hot parcels, the density of the parcel is lower than in the surrounding rock. This
yields a configuration with positive initial buoyancy, i.e, the parcel initially starts
moving upward, before the fluid is cooled down by the surrounding rock. If, on the
contrary, the salt mass fraction is very high, this increases density and yields a case
with negative initial buoyancy, that is the parcel immediately starts sinking down.
As observed in [13], one important feature of the negative buoyancy case is, that
this configuration leads to a fingering effect. However the structure of the convection
cells, or, question whether the number of fingers is finite, have not been resolved
yet. Computers like HLRS’s Hawk now allow to address these research topics. In
addition to the massive computing resources, algorithmically a proper control of the
discretisation error is crucial. To that end, novel time integration schemes have been
introduced [5, 23].
The following images provides preliminary results for the parcel benchmark with
negative buoyancy. We employ the implementation provided by the 𝑑 3 𝑓 -framework
[28]. Figs. 2 and 3 show the evolution of the parcel at an early and late stage for
different grid resolutions respectively. The simulation were performed using 128,
462 Ruben Buijse, Martin Parnet and Arne Nägel

1024, 8192 cores. Since the problem is symmetric w.r.t the center axis, all fields can
be shown in a single plot. In the background, the temperature field is shown from blue
to red. A selection of ten isosurfaces of the salt mass fraction is shown in greyscale on
the left. Streamlines resulting from the velocity field are shown on the right. Rainbow
colors indicate the corresponding magnitude.
Fig. 2 shows that the center of mass of the parcel has moved downwards from its
original position at the center already. Depending on the spatial resolution, different
fingers evolve. As expected, an increased spatial resolution yields a better separation
of the fingers. The structure is quite similar, however. In the late stage, as depicted in
Fig. 3, several branches have developed. At low spatial resolution (top image), one
big finger has evolved. However, a splitting becomes visible at the tip. When the
spatial resolution is increased, additional smaller fingers branch from the center finger,
which persists with a smaller diameter (center image). At even higher resolution, an
additional branching occurs, and three layers of fingers occur.

Conclusion

This report provided first results for the performance of the software UG4 on the
Hawk Supercomputer. The results indicate that the highly integrated CPUs yield an
excellent weak scaling. It is demonstrated that a topology aware scheduling of the
MPI processes leads to an increased performance. This seem to be related to the
hierarchical organization of the nodes an the internal NUMA architecture. Results at
this stage use about 1/3 of the full machine. One goal for the next reporting period is to
extend this to the full machine. For transient problems, strong scaling can be improved
using the multigrid in time approach. The presented scaling analysis focuses on
scalar linear problems. With respect to real-world applications with complex physics,
it is equally important to address transient non-linear coupled problems requiring
high spatio-temporal resolution. To that end, preliminary results for a thermohaline
flow have been presented. This should be extended to adaptive grid refinement and
time-stepping. Earlier studies, e.g. for poroelasticity or density-driven flow without
transfer of heat [23], showed that the linearly-implicit extrapolation method can
provide a useful tool. This will be investigated in greater detail in the next steps of
this project.

Acknowledgements The authors would like to thank Björn Dick, Bernd Krischok and all members
the HLRS staff for all technical support, guidance, and advice regarding the Hawk system.
Large scale simulations on the Hawk supercomputer 463

Fig. 2: Early stage (three different levels of spatial refinement from top to bottom):
The parcel sinks down. Depending on the spatial resolution, a branching of the central
finger becomes visible.
464 Ruben Buijse, Martin Parnet and Arne Nägel

Fig. 3: Late stage (three different levels of refinement): A layered fingering evolves.
In the highest resolution, three layers of fingers are visible.
Large scale simulations on the Hawk supercomputer 465

References

1. G. Alzetta et al. The deal.II Library, Version 9.0. Journal of Numerical Mathematics 26(4),
pp. 173–183 (2018). https://doi.org/10.1515/jnma-2018-0054.
2. Allison H. Baker et al. Scaling Hypre’s Multigrid Solvers to 100,000 Cores. In: High-
Performance Scientific Computing: Algorithms and Applications. Ed. by Michael W. Berry et
al., pp. 261–279, Springer, London, (2012). isbn: 978-1-4471-2437-5. https://doi.org/10
.1007/978-1-4471-2437-5_13.
3. Peter Bastian. DUNE – the Distributed and Unified Numerics Environment. https://www.du
ne-project.org/.
4. Markus Blatt et al. The Distributed and Unified Numerics Environment, Version 2.4. eng.
In: Archive of Numerical Software Vol 4 (2016), Starting Point and Frequency: Year: 2013.
https://doi.org/10.11588/ans.2016.100.26526.
5. Ruben Buijse. Numerische Verfahren zur Simulation thermohaliner Strömungen in der Software
d3f++. B.Sc. thesis; Goethe Universität Frankfurt (2018).
6. Hans-Jörg G. Diersch. FEFLOW: Finite Element Modeling of Flow, Mass and Heat Transport
in Porous and Fractured Media. Springer-Verlag Berlin Heidelberg (2014).
7. R.D. Falgout et al. Parallel time integration with multigrid. SIAM Journal on Scientific
Computing 36(6), pp. C635–C661 (2014). issn: 1064-8275. https://doi.org/10.1137/
130944230.
8. FEniCS Project. The FEniCS computing platform. https://fenicsproject.org/.
9. B. Flemisch et al. DuMux: DUNE for multi-phase, component, scale, physics ,... flow and
transport in porous media. Advances in Water Resources 34(9), 1102–1112 (2011). issn:
0309-1708. https://doi.org/10.1016/j.advwatres.2011.03.007.
10. Martin J. Gander. 50 years of Time Parallel Time Integration. In: Multiple Shooting and Time
Domain Decomposition. Springer (2015).
11. Björn Gmeiner et al. Towards textbook efficiency for parallel multigrid. Numerical Mathematics:
Theory, Methods and Applications 8(1), 22–46 (2015).
12. Björn Gmeiner et al. A quantitative performance study for Stokes solvers at the extreme scale.
Journal of Computational Science 17, pp. 509–521 (2016).
13. Alfio Grillo, Michael Lampe, and Gabriel Wittum. Three-dimensional simulation of the
thermohaline-driven buoyancy of a brine parcel. Comput Visual Sci 13, pp. 287–297 (2010).
14. W. Hackbusch. Multi-Grid Methods and Applications. Springer, Berlin (1985). isbn: 3-540-
12761-5.
15. Myra Huymayer. First steps towards a scaling analysis of a fully resolved electrical neuron
model. In: High Performance Computing in Science and Engineering ’19 (2021).
16. Mary F. Wheeler. IPARS: A New Generation Framework for Petroleum Reservoir Simulation.
http://csm.ices.utexas.edu/ipars/.
17. O. Ippisch and M. Blatt. Scalability Test of 𝜇 𝜑 and the Parallel Algebraic Multigrid solver
of DUNE-ISTL. In: Jülich Blue Gene/P Extreme Scaling Workshop 2011, Technical Report
FZJ-JSC-IB-2011-02, April 2011. Ed. by B. Mohr and Wolfgang Frings (2011).
18. Guido Kanschat. Web site: deal.II – an open source finite element library. http://www.deal
ii.org/.
19. O. Kolditz, S. Bauer, L. Bilke et al. OpenGeoSys: an open-source initiative for numerical
simulation of thermo/hydro/mechanical/chemical (THM/C) processes in porous media. Environ
Earth Sci 67(2), 589–599 (2012).
20. Hans Petter Langtangen and Anders Logg. Solving PDEs in Python – The FEniCS Tutorial I.
Springer International Publishing (2016). https://doi.org/10.1007/978-3-319-5246
2-7.
21. Lawrence Livermore National Laboratory. Web site: XBraid: Parallel multigrid in time.
http://llnl.gov/casc/xbraid.
22. Knut-Andreas Lie et al. Open-source MATLAB implementation of consistent discretisations
on complex grids. Computational Geosciences 16(2), 297–322 (Mar. 2012). issn: 1573-1499.
https://doi.org/10.1007/s10596-011-9244-4.
466 Ruben Buijse, Martin Parnet and Arne Nägel

23. Arne Nägel and Gabriel Wittum. Scalability of a Parallel Monolithic Multilevel Solver for
Poroelasticity. In: High Performance Computing in Science and Engineering ’18. Springer
Nature Switzerland AG (2019).
24. Martin Parnet. Zeitparallele Lösungsverfahren für die Wärmeleitungsgleichung (2020).
25. Sebastian Reiter et al. A massively parallel geometric multigrid solver on hierarchically
distributed grids. Computing and Visualization in Science 16(4), 151–164 (Aug. 2013).
https://doi.org/10.1007/s00791-014-0231-x.
26. S. Reiter et al. A massively parallel multigrid method with level dependent smoothers for
problems with high anisotropies. High Performance Computing in Science and Engineering
’16, pp. 667–675, Springer (2017).
27. Johann Rudi et al. An extreme-scale implicit solver for complex PDEs. In: Proceedings of the
International Conference for High Performance Computing, Networking, Storage and Analysis
– SC ’15. ACM Press (2015). https://doi.org/10.1145/2807591.2807675.
28. Anke Schneider, ed. Modeling of Data Uncertainties on Hybrid Computers. GRS Bericht 392
(2016). isbn: 978-3-944161-73-0.
29. U. Trottenberg, C.W. Oosterlee, and A. Schüller. Multigrid. Contributions by A. Brandt,
P. Oswald and K. Stüben. Academic Press, San Diego, CA (2001).
30. Andreas Vogel et al. UG 4: A novel flexible software system for simulating PDE based models
on high performance computers. Computing and Visualization in Science 16(4), 165–179 (Aug.
2013). https://doi.org/10.1007/s00791-014-0232-9.
Scaling in the context of molecular dynamics
simulations with ms2 and ls1 mardyn

Simon Homes, Robin Fingerhut, Gabriela Guevara-Carrion, Matthias Heinen and

Jadran Vrabec

Abstract This chapter covers scaling issues related to our recent work in the field
of molecular dynamics simulations with a focus on computational details. The
first section deals with finite size effects in the context of multicomponent diffusion.
Different methods to correct the influence of the finite simulation domain are compared.
In the second section, the structure of a fluid near its critical point is discussed in the
context of the strong scaling behavior of the code ms2. The third section discusses
droplet coalescence dynamics investigated by large molecular dynamics simulations
and a macroscopic phase field model, respectively. The performance of the new
supercomputer Hawk is compared to that of Hazel Hen. The last section describes
the influence of the direct sampling of the energy flux on the performance of the code
ls1 mardyn. Various speed tests are carried out and analyzed in detail.

1 Finite size effects of multicomponent diffusion coefficients

Diffusion processes in multicomponent mixtures are highly complex and still not
well understood due to the presence of coupling effects. Moreover, experiments to
measure transport diffusion coefficients are difficult and time consuming. Molecular
dynamics simulation has become a powerful alternative to accurately predict diffusion
coefficients and thus to improve data availability.
Molecular dynamics simulations under periodic boundary conditions are typically
performed employing a small number of molecules which is far away from the
thermodynamic limit. In this context, systematic errors associated with the system
size are present when diffusion coefficients are calculated. It is thus necessary to correct
the simulation data to account for such effects. The most widely employed correction
method is based on the shear viscosity 𝜂 and the edge length of the simulation volume

Simon Homes, Robin Fingerhut, Gabriela Guevara-Carrion, Matthias Heinen and Jadran Vrabec
Lehrstuhl für Thermodynamik und Thermische Verfahrenstechnik, Technische Universität Berlin,
Ernst-Reuter-Platz 1, 10587 Berlin, Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 467
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_29
468 S. Homes, R. Fingerhut, G. Guevara-Carrion, M. Heinen and J. Vrabec

𝐿, i.e. 2.837297 · 𝑘 𝐵𝑇/(6𝜋𝜂𝐿), and does not require additional simulation runs. This
method was proposed originally by Yeh and Hummer [21] to correct self-diffusion
coefficients has been demonstrated to not always be adequate [12].
In case of multicomponent mixtures, finite size effects are observed not only for self-
diffusion coefficients but also for mutual diffusion coefficients according to Maxwell-
Stefan or Fick. Typically, the Fick diffusion coefficient matrix of a multicomponent
mixture with 𝑛 components is not obtained directly by equilibrium molecular
dynamics simulation, but it is calculated from the sampled phenomenological
diffusion coefficients 𝐿 𝑖 𝑗 and the thermodynamic factor matrix 𝚪, employing the
relation [D] = [B] −1 [𝚪], in which all three symbols represent (𝑛 − 1) × (𝑛 − 1)
matrices and [𝑩] = [𝚫] −1 , where
! 𝑛
!
𝐿 𝑖 𝑗 𝐿 𝑖𝑛 ∑︁ 𝐿 𝑘 𝑗 𝐿 𝑘𝑛
Δ𝑖 𝑗 = (1 − 𝑥𝑖 ) − − 𝑥𝑖 − . (1)
𝑥𝑗 𝑥𝑛 𝑘=1≠𝑖
𝑥𝑗 𝑥𝑛
In fact, the observed system size dependence of the Fick diffusion coefficient
matrix elements is directly associated with the corresponding effects of the underlying
phenomenological coefficients [8]. Therefore, a finite size correction approach based
on the correction of the phenomenological coefficients is more adequate than a direct
correction of Fick diffusion coefficients [14].
Here, this system size dependence was studied for the ternary mixture methanol
+ ethanol + isopropanol, performing series of simulations with varying system size
containing 512 to 8000 molecules. The infinite size values were obtained from the
extrapolation of the respective diffusion coefficients to the thermodynamic limit
𝐿 −1 → 0. Figure 1 shows the predicted intra-diffusion coefficients and their values
corrected with the Yeh and Hummer term [21] as a function of the system size. It can
be seen that this correction yields an overestimation between 4% and 10% from the
extrapolated values for systems containing 1000 and 8000 molecules, respectively.
To study the finite size effects of the mutual diffusion coefficients, the values for
an infinite system size were calculated for all main 𝐿 𝑖𝑖 and cross phenomenological
coefficients 𝐿 𝑖 𝑗 of the studied ternary mixture as depicted in Figure 2. The proposed
fast correction procedure based on normalized coefficient values [8, 9] leads to
corrections of the main and cross phenomenological coefficients for simulations with
6000 molecules of approximately 5% and 4%, respectively. Infinite size extrapolated
and corrected diffusion values exhibit a good agreement, with relative deviations
below 4%.
Because the phenomenological coefficients are underlying to the Fick diffusion
coefficient coefficients, a propagation of finite size effects is expected. Figure 3 shows
the observed system size dependence for all elements of the Fick diffusion coefficient
matrix of the studied mixture. The extrapolated values in the thermodynamic limit
are compared with values that were obtained with the proposed correction procedure
based on the phenomenological coefficients [8,9] and those corrected with the method
by Jamali et al. [14]. Note that this method only considers corrections for the main
elements of the Fick diffusion coefficient matrix. The proposed correction method
based on the phenomenological coefficients agrees on average within 2% with the
Scaling in the context of molecular dynamics simulations with ms2 and ls1 mardyn 469


  


 






            

  

Fig. 1: Intra-diffusion coefficients of methanol (a), ethanol (b) and isopropanol (c) in
their ternary mixture (𝑥CH4O = 0.125, 𝑥 C2H6O = 0.625 and 𝑥C3H8O = 0.25 mol·mol−1 )
as a function of the inverse edge length of the simulation volume 𝐿 −1 at 298.15 K and
0.1 MPa. The uncorrected simulation results (empty circles) are shown together with
the corrected values using the Yeh and Hummer approach [21] (green crosses). The
gray dashed line is a linear fit to the uncorrected simulation results (empty circles)
and the blue line represents the extrapolated value in the thermodynamic limit.

  

  






  

  

  

  




 


 



 


  

              
  

Fig. 2: Phenomenological coefficients of the mixture methanol (1) + ethanol (2) +

isopropanol (3) (𝑥H2O = 0.125, 𝑥 C2H6O = 0.625 and 𝑥C3H8O = 0.25 mol·mol−1 ) as a
function of the inverse edge length of the simulation volume 𝐿 −1 at 298.15 K and 0.1
MPa. The uncorrected simulation results (empty circles) are shown together with
the corrected values using the fast correction procedure [8, 9] for 𝑁 = 8000 (red
squares). The gray dashed line is a linear fit to the uncorrected simulation results
(empty circles).
470 S. Homes, R. Fingerhut, G. Guevara-Carrion, M. Heinen and J. Vrabec

extrapolated data. For the cross elements of the Fick diffusion coefficient matrix, this
method [8, 9] is also in good agreement with the extrapolated values, cf. Figure 3.
For the studied ternary mixture, the proposed Fick diffusion correction method based
on the phenomenological coefficients yields better results than the method by Jamali
et al. [14], which neglects the correction of the cross elements of the Fick diffusion
coefficient matrix.

 



 
 









 

 
 

 

 

 

 

 

 
         
 

Fig. 3: Elements of the Fick diffusion coefficient matrix of the mixture methanol (1)
+ ethanol (2) + 2-propanol (3) (𝑥 CH4O = 0.125, 𝑥 C2H6O = 0.625 and 𝑥 C3H8O = 0.25
mol·mol−1 ) as a function of the inverse edge length of the simulation volume 𝐿 −1 at
298.15 K and 0.1 MPa. The coefficients calculated with the corrected values using the
fast correction procedure [8, 9] for 𝑁 = 8000 (green crosses) are compared with those
according to the procedure by Jamali et al. [14] (blue triangles) and Fick diffusion
coefficients based on the individually extrapolated phenomenological coefficients
(red squares). The gray dashed line is a linear fit to the uncorrected simulation results
(empty circles).
Scaling in the context of molecular dynamics simulations with ms2 and ls1 mardyn 471

2 Strong scaling of ms2 when sampling the structure

It is well known that sampling thermodynamic equilibrium properties with standard

system sizes of a few thousand molecules may lead to challenges in regions close to
the critical point. Because structure and dynamics of such states exhibit pronounced
fluctuations, small systems may not cover the relevant physics.
A recent study [16] has shown that molecular dynamics (MD) simulations combined
with sampling the radial distribution function (RDF) 𝑔(𝑟) are well suited to analyze
structural phenomena. Here, the Lennard-Jones (LJ) potential was applied in MD
simulations utilizing the program ms2 [1, 5, 6, 18]. Three isotherms at 𝑇 = 1.4, 1.5
and 1.6 were studied somewhat above the critical temperature 𝑇𝑐 = 1.32. Each
isotherm was sampled by varying the density from ±5 % to ±20 % around the
critical density 𝜌 𝑐 = 0.31 in steps of 5 %. MD simulations were carried out with
𝑁 = 120,000 particles and 4 · 105 equilibration as well as 106 production steps in
the canonical ensemble with a time step of Δ𝑡 = 0.004. RDF were sampled over the
entire production period up to the maximum radius 𝑟 given by half of the simulation
volume edge length 𝐿/2, which corresponds to 𝑟/𝜎 = 39.25 for the smallest and
𝑟/𝜎 = 34.29 for the largest density. A discretization of 8,000 shells over the radius
was specified. Fig. 4 shows a magnified view on the RDF.

Fig. 4: Magnified view on the RDF sampled by MD simulations with ms2; left:
𝑇 = 1.4, center: 𝑇 = 1.5, right: 𝑇 = 1.6.

It can be clearly seen that the first peak of the RDF decreases with rising temperature,
whereas its position shifts only slightly towards smaller values of 𝑟. Moreover, the
area below 𝑔(𝑟) decreases with rising density or temperature.
The parallel performance of ms2 was assessed with respect to strong scaling,
where the problem size is fixed while the number of processing elements is increased.
The strong scaling efficiency of ms2 was analyzed for its hybrid MPI + OpenMP
parallelization. This scheme allows simulations with a larger number of particles
𝑁 because of more efficient data handling. The vertical axis of Fig. 5 shows the
472 S. Homes, R. Fingerhut, G. Guevara-Carrion, M. Heinen and J. Vrabec

computing power (nodes) times computing time per computing intensity (problem
size). A horizontal line would indicate a perfect strong scaling efficiency. Here,
concurrent sampling of several state points in one program execution was allocated
with ms2. In each ensemble, MD simulations with 𝑁 = 120,000 LJ particles with a
cutoff radius of 𝑟 𝑐 /𝜎 = 6 and 𝑟 𝑐 /𝜎 = 35.3 were chosen. From Fig. 5, it becomes clear
that ms2 is close to optimal strong scaling, however, 100 % efficiency is not reached.
In ms2, the computing intensity for traversing the particle matrix is proportional to
𝑁 2 , but the intermolecular interactions are calculated for particles that are in the cutoff
sphere only. Thus, the computational cost of ms2 should be in-between 𝑁 and 𝑁 2 . The
effective proportionality depends on the computing intensity of the intermolecular
interactions. Because single LJ particles were considered here, the cost of calculating
these interactions is not much higher than that of traversing the particle matrix. Fig. 5
compares simulations with a cutoff radius of 𝑟 𝑐 /𝜎 = 6 and simulations with the
maximum cutoff radius of 𝑟 𝑐 /𝜎 = 35.3, which have a ratio of around 5.883. The
number of particles in the cutoff sphere is proportional to its volume. Thus, a radius
ratio of 5.883 indicates a volume ratio of 5.8833 = 203.64, which is proportional
to the computational load for the intermolecular interactions. However, increasing
𝑟 𝑐 /𝜎 from 6 to 35.3 leads to an increase of execution time of only 4.38 (comparison
for 300 nodes). Thus, traversing the particle matrix is dominating in the present LJ
particle scenarios. Because one of the parallelization schemes in ms2 is implemented
on the particle matrix traversing level, an execution time ratio of around 2.52 was
achieved for larger parallelization (comparison of the data for 2,400 nodes).

30
10 molecules * steps

20
time * nodes

10
6

0
0 500 1000 1500 2000 2500
nodes (with 128 cores)

Fig. 5: Strong scaling efficiency of MD simulations with ms2 for a LJ fluid with
𝑁 = 120,000 particles measured on HPE Apollo (Hawk) for the hybrid MPI +
OpenMP parallelization (each MPI process with four OpenMP threads); red circles:
𝑟 𝑐 /𝜎 = 35.3; blue squares: 𝑟 𝑐 /𝜎 = 6.
Scaling in the context of molecular dynamics simulations with ms2 and ls1 mardyn 473

3 Droplet coalescence on Hawk and Hazel Hen

Large scale MD simulations were conducted to investigate the coalescence process of

two initially resting argon droplets, considering three sizes, i.e. with diameter 𝑑 = 50,
100 and 200 nm. The droplets were arranged on the central axis of a cuboid system
with a distance of only a single particle diameter, cf. Fig. 6. They were surrounded
by saturated vapor at a temperature of 110 K (≈ 74 % of the critical temperature).
The systems consisted of approximately 3, 25 and 200 million particles, respectively,
denoted with 3M 25M and 200M in the following. To cope with such a large number
of particles, the MD code ls1 mardyn [17] was employed for all simulations, which is
continuously developed and optimized for massively parallel execution [20].

𝑦 𝑥

𝑡 = 0.0 ns 𝑡 = 0.5 ns 𝑡 = 3.0 ns

0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30
𝜌 [mol/l] 𝜌 [mol/l] 𝜌 [mol/l]

Fig. 6: Snapshots, rendered by MegaMol [7], from the MD simulation with two argon
droplets of diameter 𝑑 = 50 nm (top) and two-dimensional density fields sampled
during simulation, depicted with a color map identifying the vapor phase (light blue),
the liquid phase (yellow) and the interface in-between (bottom color map).

During simulation, two-dimensional density fields where sampled by employing a

rotationally symmetric binning scheme. For this purpose, the system was divided by
equidistant cylinder shells around and equidistant slices along the 𝑧 axis. The sampled
density field 𝜌(𝑧, 𝑟), with 𝑟 being the distance to the 𝑧 axis, represents a centered cut
trough the droplets and can be visualized by a color map to identify vapor, liquid and
the interface in-between, cf. Fig. 6.
These data sets were used as a benchmark for a macroscopic solution based on a
phase-field model, i.e. results obtained from computational fluid dynamics (CFD)
simulations, considering exactly the same scenarios. This model is described in detail
in Ref. [2] and was already successfully used to simulate the motion of a droplet on a
surface controlled by a wettability gradient [3].
For a comparison of the results of MD and CFD simulations, the contour of the
interface of the coalescing droplets was extracted from the sampled (MD) or calculated
(CFD) density fields and superimposed in the same plot, cf. Fig. 7. This representation
474 S. Homes, R. Fingerhut, G. Guevara-Carrion, M. Heinen and J. Vrabec

shows a very good agreement of the results. A more elaborate comparison, e.g. based
on the growth rate of the bridge that forms between the coalescing droplets, was
recently published in [11].

MD
CFD
t = 0.0 ns t = 0.5 ns t = 3.0 ns
Radius r / nm

25

−25

−50 0 50 −50 0 50 −50 0 50

t = 0.0 ns t = 2.0 ns t = 20.0 ns
Radius r / nm

100

−100

−200 0 200 −200 0 200 −200 0 200

Coordinate z / nm

Fig. 7: Interface contour of coalescing argon droplets with a diameter of 𝑑 = 50 nm

(top) and 𝑑 = 200 nm (bottom) at three instances of time 𝑡 = 0, 0.5 and 3 ns and 𝑡 = 0,
2 and 20 ns, respectively. Results of MD (solid) and CFD (dash-dotted) simulations
are superimposed, showing an almost perfect agreement.

For the present MD simulation series, comprehensive scaling experiments were

already conducted in the framework of the project TaLPas, funded by the Federal
Ministry of Education and Research (BMBF), at that time on the machine Hazel Hen,
the previous supercomputer of the High Performance Computing Center Stuttgart
(HLRS). In this work, after a revision of the initial configurations, composed from
carefully equilibrated droplets, the simulations were conducted on the new machine
Hawk. For these production runs, the results of the scaling experiments on Hazel Hen
were used to estimate a good choice for the allocation of computational resources.
Since the entire coalescence process up to the complete fusion of the droplets had to
be considered, the time-to-solution should be as short as possible. However, in order
to use the resources efficiently, the number of cores to be allocated was set in a range
before the course of the scaling curves starts to diverge substantially from perfect
strong scaling.
Results of the present time-to-solution measurements (Hawk) compared to the
previous ones (Hazel Hen) are depicted in Fig. 8. They show a speed-up of up to a
factor of 2.7 for the 200M scenario (factors 2.2 and 2.1 for the 3M and 25M scenarios).
These results underline the excellent performance of Hawk, pushing forward the MD
simulations’ accessible scales towards micrometers and microseconds.
Scaling in the context of molecular dynamics simulations with ms2 and ls1 mardyn 475

26

24
Time steps / s
22

20 Hazel Hen Hawk

3M 3M
2−2 25M 25M
200M 200M

21 23 25 27 29 211
Number of cores / 24

Fig. 8: Performance of ls1 mardyn for the droplet coalescence scenario measured on
Hawk compared to previous results of scaling experiments conducted on Hazel Hen.

4 Scaling of energy flux implementation in ls1 mardyn

The energy flux plays a crucial role in many scientific and engineering problems.
Until now, our open-source MD code ls1 mardyn [17] was not capable to sample the
energy flux directly. Instead, it had to be calculated by post-processing as it was done
in Ref. [13]. The utilized equation results from the first law of thermodynamics

𝑗e = (ℎ + 𝑒 kin ) 𝑗 p + 𝑞,
¤ (2)
with 𝑗 e being the total energy flux, ℎ the enthalpy, 𝑒 kin the kinetic energy, 𝑗p the
particle flux and 𝑞¤ the heat flux. Until now, only the kinetic energy and the particle flux
could be sampled directly. For the calculation of the remaining quantities, external
data like correlations for the thermal conductivity [15] or equations of state [10] had
to be used. In order to overcome this, the direct sampling of the energy flux was
implemented into our code. The governing equation reads [4]
𝑁 𝑁 ∑︁𝑁
" #
1 ∑︁ ∑︁ 𝜕𝑢 𝑖 𝑗
𝑱𝑒 = 𝑚 𝑖 𝑣 2𝑖 𝒗 𝑖 − 𝒓𝑖 𝑗 ⊗ − 𝑰 · 𝑢𝑖 𝑗 · 𝒗 𝑖 . (3)
2 𝑖=1 𝑖=1 𝑗>𝑖
𝜕 𝒓𝑖 𝑗

Several quantities, like the particle mass 𝑚 and the particle velocity 𝑣, are summed
up over 𝑁 particles, while 𝒓 𝑖 𝑗 , 𝑢 𝑖 𝑗 and 𝜕𝑢 𝑖 𝑗 /𝜕 𝒓 𝑖 𝑗 are the distance vector, potential
energy and virial between two particles 𝑖 and 𝑗, respectively, and 𝑰 stands for the
identity matrix.
Since the off-diagonal elements of the virial tensor are needed, additional code had
to be added to the core of ls1 mardyn, which adds significant additional computational
load. In order to get an idea about the impact, multiple tests were conducted. First,
476 S. Homes, R. Fingerhut, G. Guevara-Carrion, M. Heinen and J. Vrabec

the fastest compiler/MPI module combination was identified and the impact of
vectorization analyzed. In a second step, the performance of the new implementation
was compared to the legacy code to determine its influence. The last step was a short
evaluation of the strong scaling behavior of the new implementation.
There are several different compiler and MPI modules available on Hawk. In
this work, the compilers aocc (AMD), gcc (GNU) and icc (Intel) were combined
with the MPI implementations mpt, openmpi and impi (Intel). Some of the possi-
ble combinations are not operable as they lead to building errors. The following
five combinations were tested: aocc+aocc, aocc+aocc, aocc+aocc, aocc+aocc and
aocc+aocc. All of these were used to run the same test case which stands exemplarily
for a study in which stationary evaporation was investigated. This typical test case
consisted of about 2 · 106 LJ particles, constituting one liquid and one vapor phase.
The simulations were conducted on 8 · 128 = 1024 cores for 25,000 time steps and
the execution time was measured.
The results of the compiler/MPI study are shown in Fig. 9. It can be seen that
vectorization leads to a speed-up of about 30%, regardless of the chosen compiler/MPI
combination. Furthermore, the tests show that the gcc compiler produces the fastest
code. The choice of the MPI implementation has little effect on the performance,
since the two fastest simulations were both performed with gcc-compiled code.
Nevertheless, choosing openmpi for MPI communication can speed-up the test
simulation for another 2%. The poorest performance was achieved when utilizing the
icc (Intel) compiler. One explanation for this finding may be that Hawk consists of
AMD processors for which the icc compiler may be not well optimized.

35 without vectorization with vectorization

30
Time steps per second

25
20
15
10
5
0
aocc+mpt gcc+mpt gcc+openmpi icc+mpt icc+impi

Fig. 9: Performance of different compiler/MPI combinations with and without

vectorization. Performance was measured in time steps per second as an average over
25,000 time steps.

In a second study, the impact of the implementation of the direct sampling of the
energy flux was investigated. The most performant compiler/MPI combination was
used to conduct the simulation runs. Again, the simulations were run for 25,000 time
steps, respectively, and the average number of time steps per second was calculated.
Even though significantly more calculations have to be conducted in order to sample
Scaling in the context of molecular dynamics simulations with ms2 and ls1 mardyn 477

the energy flux directly, the new code takes just about 25% longer to execute the
simulation of the same test scenario. Furthermore, with enabled vectorization, the
new code is still faster compared to the old one without vectorization.

40
new code legacy code
35
Time sptes per second

30
25
20
15
10
5
0
without vectorization with vectorization

Fig. 10: Performance comparison of ls1 mardyn with and without the implementation
of direct energy flux sampling. Performance is measured in time steps per second as
an average over 25,000 time steps.

In a last test, the strong scaling behavior of ls1 mardyn was investigated for the
aforementioned test case. Three scaled-up simulations were run in total. The smallest
run utilized 2 · 128 = 256 cores, the mid-sized one 8 · 128 = 1024 cores and the biggest
one 16 · 128 = 2048 cores. The speed-up of all three simulations was compared to a
run on a single node, i.e. 128 cores. Results are shown in Fig. 11. Ideal scaling as
well as the speed-up of the test case simulations are plotted over the number of cores.

16
Ideal
12
Test case
Speed-up

0
0 512 1024 1536 2048
Number of cores

Fig. 11: Strong scaling of a test case with 𝑁 = 2 · 106 particles executed with ls1
mardyn on a varying number of cores.
478 S. Homes, R. Fingerhut, G. Guevara-Carrion, M. Heinen and J. Vrabec

It became apparent that the test case scales almost perfectly up to a total number of
1024 cores. Utilizing even more cores speeds up the simulation, although the scaling
is not close to ideal anymore. This is a consequence of the investigated special test
case which consists of a very elongated domain in combination with simple domain
decomposition. For other test cases and more sophisticated domain decompositions,
the scaling behavior of our code ls1 mardyn was studied in detail as well [19].

Acknowledgements The co-authors R.F., G.G.-C., M.H. and J.V. acknowledge funding by Deutsche
Forschungsgemeinschaft (DFG) through the Project SFB-TRR 75, Project number 84292822 -
“Droplet Dynamics under Extreme Ambient Conditions”. This work was carried out under the
auspices of the Boltzmann-Zuse Society of Computational Molecular Engineering (BZS), and it
was facilitated by activities of the Innovation Centre for Process Data Technology (Inprodat e.V.),
Kaiserslautern. The simulations were performed on the HPE Apollo (Hawk) at the High Performance
Computing Center Stuttgart (HLRS).

References

1. Deublein, S., Eckl, B., Stoll, J., Lishchuk, S.V., Guevara-Carrion, G., Glass, C.W., Merker, T.,
Bernreuther, M., Hasse, H., Vrabec, J.: ms2: A molecular simulation tool for thermodynamic
properties. Computer Physics Communications 182, 2350–2367 (2011)
2. Diewald, F.: Phase field modeling of static and dynamic wetting, Forschungsbericht / Technische
Universität Kaiserslautern, Lehrstuhl für Technische Mechanik, vol. 19 (2020)
3. Diewald, F., Lautenschlaeger, M.P., Stephan, S., Langenbach, K., Kuhn, C., Seckler, S., Bungartz,
H.J., Hasse, H., Müller, R.: Molecular dynamics and phase field simulations of droplets on
surfaces with wettability gradient. Computer Methods in Applied Mechanics and Engineering
361, 112773 (2020)
4. Fernández, G., Vrabec, J., Hasse, H.: A molecular simulation study of shear and bulk viscosity
and thermal conductivity of simple real fluids. Fluid Phase Equilibria 221, 157–163 (2004)
5. Fingerhut, R., Guevara-Carrion, G., Nitzke, I., Saric, D., Marx, J., Langenbach, K., Prokopev,
S., Celný, D., Bernreuther, M., Stephan, S., Kohns, M., Hasse, H., Vrabec, J.: ms2: A molecular
simulation tool for thermodynamic properties, release 4.0. Computer Physics Communications
262, 107860 (2021)
6. Glass, C.W., Reiser, S., Rutkai, G., Deublein, S., Köster, A., Guevara-Carrion, G., Wafai, A.,
Horsch, M., Bernreuther, M., Windmann, T., Hasse, H., Vrabec, J.: ms2: A molecular simulation
tool for thermodynamic properties, new version release. Computer Physics Communications
185, 3302–3306 (2014)
7. Grottel, S., Krone, M., Müller, C., Reina, G., Ertl, T.: Megamol – a prototyping framework for
particle-based visualization. IEEE Transactions on Visualization and Computer Graphics 21,
201–214 (2015)
8. Guevara-Carrion, G., Fingerhut, R., Vrabec, J.: Fick diffusion coefficient matrix of a quaternary
liquid mixture by molecular dynamics. Journal of Physical Chemistry B 124, 4527–4535 (2020)
9. Guevara-Carrion, G., Fingerhut, R., Vrabec, J.: Diffusion in multicomponent aqueous alcoholic
mixtures. Scientific Reports 11, 12319 (2021)
10. Heier, M., Stephan, S., Liu, J., Chapman, W.G., Hasse, H., Langenbach, K.: Equation of state
for the Lennard-Jones truncated and shifted fluid with a cut-off radius of 2.5 sigma based on
perturbation theory and its applications to interfacial thermodynamics. Molecular Physics 116,
2083–2094 (2018)
11. Heinen, M., Hoffman, M., Diewald, F., Seckler, S., Langenbach, K., Vrabec, J.: Droplet
coalescence by molecular dynamics and phase-field modeling. Physics of Fluids 34, 042006
(2022)
Scaling in the context of molecular dynamics simulations with ms2 and ls1 mardyn 479

12. Heyes, D.M., Cass, M.J., Powles, J., Evans, W.A.B.: Self-diffusion coefficient of the hard-sphere
fluid: System size dependence and empirical correlations. Journal of Physical Chemistry B 111,
1455–1464 (2007)
13. Homes, S., Heinen, M., Vrabec, J., Fischer, J.: Evaporation driven by conductive heat transport.
Molecular Physics, in press (2021). DOI 10.1080/00268976.2020.1836410
14. Jamali, S.H., Bardow, A., Vlugt, T.J.H., Moultos, O.A.: Generalized form for finite-size correc-
tions in mutual diffusion coefficients of multicomponent mixtures obtained from equilibrium
molecular dynamics simulation. Journal of Chemical Theory and Computation 16, 3799–3806
(2020)
15. Lautenschlaeger, M.P., Hasse, H.: Transport properties of the Lennard-Jones truncated and
shifted fluid from non-equilibrium molecular dynamics simulations. Fluid Phase Equilib. 482,
38–47 (2019)
16. Mausbach, P., Fingerhut, R., Vrabec, J.: Structure and dynamics of the Lennard-Jones fcc-solid
focusing on melting precursors. Journal of Chemical Physics 153, 104506 (2020)
17. Niethammer, C., Becker, S., Bernreuther, M., Buchholz, M., Eckhardt, W., Heinecke, A.,
Werth, S., Bungartz, H.J., Glass, C.W., Hasse, H., Vrabec, J., Horsch, M.: ls1 mardyn: The
massively parallel molecular dynamics code for large systems. Journal of Chemical Theory and
Computation 10, 4455–4464 (2014)
18. Rutkai, G., Köster, A., Guevara-Carrion, G., Janzen, T., Schappals, M., Glass, C.W., Bernreuther,
M., Wafai, A., Stephan, S., Kohns, M., Reiser, S., Deublein, S., Horsch, M., Hasse, H., Vrabec,
J.: ms2: A molecular simulation tool for thermodynamic properties, release 3.0. Computer
Physics Communications 221, 343–351 (2017)
19. Seckler, S., Gratl, F., Heinen, M., Vrabec, J., Bungartz, H.J., Neumann, P.: AutoPas in ls1 mardyn:
Massively parallel particle simulations with node-level auto-tuning. Journal of Computational
Science 50, 101296 (2021)
20. Tchipev, N., Seckler, S., Heinen, M., Vrabec, J., Gratl, F., Horsch, M., Bernreuther, M., Glass,
C.W., Niethammer, C., Hammer, N., Krischok, B., Resch, M., Kranzlmüller, D., Hasse, H.,
Bungartz, H.J., Neumann, P.: TweTriS: Twenty trillion-atom simulation. International Journal
of High Performance Computing Applications 33, 838–854 (2019)
21. Yeh, I.C., Hummer, G.: System-size dependence of diffusion coefficients and viscosities from
molecular dynamics simulations with periodic boundary conditions. Journal of Physical
Chemistry B 108, 15873–15879 (2004)
Scalable multigrid algorithm for fluid dynamic
shape optimization

Jose Pinzon, Martin Siebenborn and Andreas Vogel

Abstract We investigate a parallel approach for the shape optimization of an obstacle

in an incompressible Navier–Stokes flow. For this purpose, we used a self-adapting
nonlinear extension equation within the method of mappings, which links a boundary
control to a mesh deformation. It is demonstrated how the approach preserves mesh
quality and allows for large deformations by controlling nonlinearity in critical regions
of the domain. Special focus is given to reference configurations, where the transfor-
mation has to remove and create obstacle boundary singularities, which is particularly
relevant for the employed multigrid discretizations. As benchmark problems, we
demonstrate the aerodynamic drag optimization in 2d and 3d configurations. The
efficiency of the algorithm is demonstrated in weak scalability studies performed on
the supercomputer HPE Hawk for up to 5,120 cores.

1 Introduction

Our project aims at furthering our understanding of optimization schemes for domains
that experience large deformations. With this purpose, we employ shape optimization
to obtain the optimal shape of an obstacle in terms of a physical quantity, e.g. the
aerodynamic drag or lift.
Our main focus is on extension operators within the method of mappings [2, 8, 13],
with which a scalar variable defined on the surface of an obstacle is related to a
deformation field defined on the surrounding domain, c.f. fig. 1. This approach
has been implemented with linear elastic extension equations in [7], however it

A. Vogel
High Performance Computing, Ruhr University Bochum, Universitätsstraße 150, 44801 Bochum,
Germany, e-mail: [email protected]
J. Pinzon and M. Siebenborn
Department of Mathematics, University Hamburg, Bundesstraße 55, 20146 Hamburg, Germany,
e-mail: {jose.pinzon,martin.siebenborn}@uni-hamburg.de

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 481
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_30
482 J. Pinzon, M. Siebenborn and A. Vogel

Γwall

Γin Γobs Ωobs Γout

Γwall

Fig. 1: Sketch of the domain used in the 2d simulations. An obstacle Γobs is located
inside a flow tunnel.

suffers under large deformations. To allow for large deformations, it turns out to be
advantageous to add nonlinear terms to the extension model. For instance, in [14] a
nonlinear advective term leads the obtained displacements on the major directions
of deformations, thus preserving mesh quality even under large displacements with
respect to the reference configuration.
The domain of the underlying optimization experiment is sketched in fig. 1. It
represents an obstacle Ωobs and its surface Γobs inside a holdall domain Ω, which
represents a flow tunnel. The obstacle is subject to an incompressible flow with
inflow at Γin . The objective functional is the drag, as expressed in [12]. We follow the
continuous adjoint approach (see for instance [6, 9, 12]) where the determination of
the fixed dimensional multipliers is realized via an augmented Lagrangian approach.
For an overview of the ongoing research on various approaches in the area of shape
optimization in fluid dynamics, the reader is referred to [2, 4, 11, 17]. For a complete
monograph on discretization schemes, solvers, and numerical stabilization techniques
for the Navier–Stokes equations, please refer to [3] and the citations therein.
The core of the method of mappings is to rephrase the shape optimization problem
into an optimal control problem over the set of admissible transformations acting on a
reference domain. This is in contrast to defining a set of admissible shapes itself and
to optimize within this. The advantage is that the appropriate choice of the extension
equation allows to require properties of the optimal configuration such as a certain
degree of mesh quality in the deformed domain or preventing self intersections in the
discretization. This means that there are no mesh deformations performed throughout
the algorithm, but all operators are traced back to the reference domain. In order to
maintain mesh quality and prevent element overlappings, a threshold is enforced on
the deformation gradient. We express this as an inequality constraint in the problem
formulation, as in [7].
Scalable Multigrid Algorithm for Shape Optimization 483

2 Mathematical background

The following strategy and optimization problem are outlined in detail in [15]. We
start by introducing the shape optimization problem in a general form. Via the method
of mappings it is formulated by means of an optimal control problem for a generic
objective 𝑗 and a PDE constraint 𝑒. Let further 𝑋 = 𝐿 2 (Γobs ) × 𝐿 2 (Ω), 0 < 𝑢 lb < 𝑢 ub ,
𝜂det > 0, 𝛼 > 0 and consider the problem
2 𝜒
min 𝑗 (𝑦, 𝐹 (Ω)) + 𝛼
2 ∥𝑢∥ 𝐿 2 (Γobs ) + 2 ∥𝜂 − 12 (𝜂ub + 𝜂lb ) ∥ 2𝐿 2 (Ω) (1)
(𝑢, 𝜂) ∈𝑋

s.t. 𝑒(𝑦, 𝐹 (Ω)) = 0 (2)

𝐹 = id + w (3)
w = 𝑆(𝜂, 𝑢, Ω) (4)
det(𝐷𝐹) ≥ 𝜂det in Ω (5)
𝜂lb ≤ 𝜂 ≤ 𝜂ub in Ω (6)
𝑔(w) = 0. (7)

In the objective function (1), 𝑗 refers to the quantity to be optimized, for instance the
drag of the obstacle. 𝑦 denotes the state variable of the PDE. The boundary control
variable is denoted with 𝑢, w is its extension to a displacement field via the extension
equation 𝑆 (cf. (4)), which then defines the domain mapping function 𝐹. In (7),
𝑔(w) represent geometric constraints, see [7, 14] for a complete explanation of how
to treat these. The domain transformation 𝐹 (Ω) in (3) is based on the perturbation
of identity. Furthermore, (5) models a threshold on the deformation gradient towards
local injectivity of 𝐹 as investigated in [7]. Finally, (6) defines box constraints on the
factor 𝜂, which limits the nonlinearity in 𝑆, taking into account that this affects the
convergence of the iterative solver.
In the optimization problem (1) to (7), we consider the PDE constraint 𝑒(𝑦, 𝐹) to be
the stationary, incompressible Navier–Stokes equations in terms of a velocity v and a
pressure 𝑝 given by

−𝜈Δv + (v · ∇)v + ∇𝑝 = 0 in 𝐹 (Ω) (8)

div v = 0 in 𝐹 (Ω) (9)
v = 𝑣∞ on Γin (10)
v=0 on Γobs ∪ Γwall (11)
𝜕v
𝑝𝑛 − 𝜈 =0 on Γout . (12)
𝜕𝑛
∫
For compatibility, it is assumed that Γ v∞ ·𝑛 d𝑠 = 0 holds at the inflow. The extension
in
𝑆(𝜂, 𝑢, Ω) is given as the solution operator of the PDE
484 J. Pinzon, M. Siebenborn and A. Vogel

div (𝐷w + 𝐷w⊤ ) + 𝜂ext (w · ∇)w = 0 in Ω (13)

(𝐷w + 𝐷w⊤ )𝑛 = 𝑢n on Ωobs (14)
w=0 on Γin ∪ Γout ∪ Γwall . (15)

The condition (5) is approximated via a non-smooth penalty term. This results in the
regularized objective function

2 𝜒
𝐽 (𝑢, 𝜂) = 𝑗 (𝑦, 𝐹 (Ω)) + 𝛼
2 ∥𝑢∥ 𝐿 2 (Γobs ) + 2 ∥𝜂 − 12 (𝜂ub + 𝜂lb ) ∥ 2𝐿 2 (Ω)
+ 𝛽2 ∥(𝜂det − det(𝐷𝐹)) + ∥ 2𝐿 2 (Ω) . (16)

In contrast to the PDE constraints (3) to (5), the geometric constraints (7) are fixed
dimensional (here it is 𝑑 + 1 where 𝑑 ∈ {2, 3}). Thus, the multiplier associated to
these conditions is not a variable in the finite element space but an 𝑑 + 1-dimensional
vector. We incorporate this into our optimization algorithm in form of an augmented
Lagrange approach. This leads to the augmented objective function

𝐽aug (𝑢, 𝜂) = 𝐽 (𝑢, 𝜂) + ∥𝑔(w) ∥ 22 (17)

and the corresponding Lagrangian

L (w, v, 𝑝, 𝑢, 𝜂ext , 𝜓w , 𝜓v , 𝜓 𝑝 , 𝜓bc 𝜓vol ) =

∫ ∫
𝜈 −1 −1 𝛼
(𝐷v(𝐷𝐹) ) : (𝐷v(𝐷𝐹) ) det(𝐷𝐹)d𝑥 + 𝑢 2 d𝑠
2 Ω 2 Γobs
∫ ∫
𝛽 + 2
+ (𝜂det − det(𝐷𝐹)) d𝑥 − [𝜈(𝐷v(𝐷𝐹) −1 ) : (𝐷𝜓v (𝐷𝐹) −1 )
2 Ω Ω
+ (𝐷v(𝐷𝐹) −1 ) · 𝜓v − 𝑝Tr (𝐷𝜓v (𝐷𝐹) −1 )] det(𝐷𝐹)d𝑥
∫ ∫
+ 𝜓 𝑝 Tr (𝐷v(𝐷𝐹) −1 ) det(𝐷𝐹)d𝑥 − [(𝐷w + 𝐷w⊺ ) : 𝐷𝜓w + 𝜂ext (𝐷w · w)]d𝑥
Ω Ω
∫ ∫ ∫
+ 𝑢n · 𝜓w d𝑠 − 𝜓bc · 𝐹 (𝑥) det(𝐷𝐹)d𝑥 − 𝜓vol (det(𝐷𝐹) − 1)d𝑥 (18)
Γobs Ω Ω

under the assumption that the barycenter of Ω is 0 ∈ R𝑑 .

The underlying numerical scheme is shown in algorithm 1. Conceptually, the
algorithm consists of two nested loops. The outer one is the augmented Lagrange
algorithm, which iteratively determines the fixed dimensional multipliers for the
geometric constraints 𝑔(w) (cf. (7)). For each approximation of 𝜆 𝑔 and penalty factor
𝜈𝑔 the actual shape optimization problem is then solved in the inner loop where the
descent direction is chosen according to a quasi-Newton update. Line 7 shows the
termination condition for the iterated shape optimization problems, which is based on
the [𝜂lb , 𝜂ub ]-box-projection of the gradient (∇𝑢, ∇𝜂). The termination of the entire
algorithm is based on the change of 𝑢 between two successive augmented Lagrange
iterations, as shown in line 14.
Scalable Multigrid Algorithm for Shape Optimization 485

Algorithm 1 Augmented Lagrange Method

Require: 𝜆𝑔 = 0, . . .
1: repeat
2: repeat
3: (w, ∇𝑢 𝑘,ℓ , ∇𝜂 𝑘,ℓ ) ← Solve for(𝑔def , 𝑢 𝑘,ℓ , 𝜂 𝑘,ℓ )
4: (𝑢 𝑘,ℓ+1 , 𝜂 𝑘,ℓ+1 ) ← Apply Quasi-Newton(𝑚, 𝑔def , 𝜈𝑔 , ∇𝑢 𝑘,ℓ , ∇𝜂 𝑘,ℓ )
5: 𝑔def ← 𝑔 (w) − 𝑔 (0)
6: ℓ←  ℓ +1 
 
7: until ∇𝑢 𝑘,ℓ−1 , 𝑃( 𝜂 𝑙 , 𝜂 𝑢 ) 𝜂 𝑘,ℓ−1 − ∇𝜂 𝑘,ℓ−1 − 𝜂 𝑘,ℓ−1 < 𝜖inner
𝑋
8: if ∥𝑔def ∥ 2 < 𝜖𝑔 then
9: 𝜈𝑔 ← 𝜈inc 𝜈𝑔
10: else
11: 𝜆𝑔 ← 𝜆𝑔 + 𝜆inc 𝑔def
12: end if
13: 𝑘 ← 𝑘+1
14: until 𝑘 ≥ 1 and ∥𝑢 𝑘,ℓ − 𝑢 𝑘−1,ℓ ∥ 𝐿 2 (Γobs ) < 𝜖outer

The Lagrange multipliers 𝜆 𝑔 are updated based on the norm of the geometrical defects,
i.e. how close the current deformation field and its corresponding transformation
fulfill the barycenter and volume constraints.

3 Numerical results

The simulations are performed with UG4 [19], a framework to solve PDE systems on
massive parallel systems. The general parallel scalability of the software is reported in
[16, 18]. The uncoupled equation systems are solved in a block-like fashion, making
it necessary to pass the solution of a PDE system to another as integration point
data. UG4 provides this functionality in the form of data imports. The computational
meshes are created using GMSH [5] and the visualization of vector and scalar
quantities, as well as of the deformed grids, employs ParaView [1].
Case studies for 2d and 3d domains are presented. The former uses a square obstacle
as reference domain to illustrate the effects of algorithm 1 on grids with pronounced
singularities. The latter demonstrates the successful generation of singularities using
surface elements. In all cases, the simulations utilize 1 MPI process per core.

3.1 2d results

The 2d simulations are carried out using a discretization of the domain shown in fig. 1.
A 𝑃1 discretization is used for all PDEs, except for the solution of the Navier–Stokes
equations, where 𝑃2 − 𝑃1 stable triangular elements are used. The grid has a total
of 421,888 triangular elements, with 5 refinement levels. The simulations runs each
486 J. Pinzon, M. Siebenborn and A. Vogel

Fig. 2: At the top, the reference configuration is shown with the optimal 𝜂 (left) and
w = 𝑆(𝑢). At the bottom, the transformed grid 𝐹 (Ω) with resulting singularities (left)
is shown, altogether with a zoom on the singularity where mesh quality is preserved
due to the choice of 𝑆

on 320 cores, divided among 4 nodes. As a rule of thumb, we choose to use the
number of refinements that better represent the used mathematical spaces. This can
be empirically measured by the smoothing of grid singularities, e.g. edges, and by
their creation. Nevertheless, it is shown in Sect. 3.2 that the obtained results are
grid independent, with only minor variations proper of the iterative solution of PDE
systems.
In fig. 2, the final results for a 2d simulation are shown. The deformation field
obtained after 1000 steps is applied to transform the reference domain. In order to
prove the efficiency of the optimization scheme shown in algorithm 1, a reference
domain with pronounced edges is chosen. The algorithm is able to detect the edges,
which are not part of the optimal configuration, to promote a concentration of both
the extension factor 𝜂ext and boundary control variable 𝑢. Likewise, it can promote
domain transformations, which result in the creation of new boundary edges. As
can be seen on the bottom of fig. 2, the front and back tips appear as a result of the
optimization process, whereas the previously existing corners are smoothed down by
the transformation.
The effects of the transformation on the reference domain, throughout several
optimization steps, is shown in fig. 4. The transformation, i.e. the deformation field,
is applied to the reference domain. In the initial steps, the deformation field incurs
in notorious violations of the geometric constraints, i.e. there is shrinkage of the
obstacle’s area. This is related to the poor initial values for the deformation field, 𝑢
Scalable Multigrid Algorithm for Shape Optimization 487

1.7 0.4
𝐽 ( 𝑢, 𝜂 )
∥𝜆𝑔 ∥ 0.35
1.6
0.3

Lagr. Mult. Norm

Objective value

1.5 0.25
0.2
1.4 0.15
0.1
1.3
0.05
1.2 0
1 10 100 1000
Optimization step

Fig. 3: The objective function per optimization step is shown in relation to the norm
of the Lagrange multipliers. The vertical blue lines indicate major changes of the
multipliers

scalar field, and 𝜂 extension factor. As the simulation progresses, the front and back
tips are formed, and the optimization scheme starts to smooth out the corners of the
reference domain.
The extension factor 𝜂ext is shown with respect to the reference domain. In (13),
𝜂ext enriches the extension equation in order to allow for nonlinear deformations.
While the extension factor could be chosen as a constant, refer to [14], the advantages
of implementing it as a scalar grid function become evident from fig. 4. In the
first steps of the simulation, the extension factor is already promoting the advective
movement of the nodes which require large deformations. This can be appreciated
both on the corners as well as the faces of the square obstacle.
The value of the objective function is shown for a 1000 steps in fig. 3. The dashed
blue lines show the optimization steps where major Lagrange multiplier updated
occurred. This is directly related to line 11 in algorithm 1, where the update is subject
to the value of the small number 𝜈𝑔 . The first steps usually imply large and expected
violations of the geometric constraints, due to the poor selection of an initial guess
for 𝜆 𝑔 . Upon each update, we see a peak in the value of the objective function. This
is expected, since the optimization process can be considered global within the given
values of the multipliers. Once the latter have converged, the value of the objective
function reaches a minimum.
488 J. Pinzon, M. Siebenborn and A. Vogel

Step Transformed domain Extension factor

10

20

40

Fig. 4: Transformation of the reference domain by the application of the deformation

field compared to the accumulation of the extension factor, given for steps 2,5,10,20,40
Scalable Multigrid Algorithm for Shape Optimization 489

Fig. 5: The convergence of several refinement levels (indicated by the colored nodes)
is shown with the superimposed grids

Fig. 6: Side-by-side comparison of the generated tips for several refinement levels.
From left to right 2,3,4,5 refinements of the discretized domain shown in fig. 1

3.2 2d grid study

In this section, we provide results which demonstrate that the obstacle shape obtained
by applying the transformation after a given number of steps is independent of
the number of refinements used, and that the detection and generation of domain
singularities is grid-independent. The coarse grid is refined from 6,592 to 421,888
elements. All simulations had the same settings, including viscosity 𝜈 = 0.02. Figure 5
shows the final grid after 400 optimization steps. The number of MPI processes used
went from 80 for the lowest refinement level to 320 for the highest, running an equal
490 J. Pinzon, M. Siebenborn and A. Vogel

number of optimization steps for all settings to provide a meaningful comparison. It

can be seen that the different surfaces across all refinement levels are superimposed
to each other or have small differences at the magnified tip in fig. 5.
In addition to this, the side-to-side comparison in fig. 6 shows that, after applying
the optimal transformation to the reference domain, the obtained geometry is the
same in terms of shape and position. Therefore, the results are grid-independent
which is a necessary fact to move towards more demanding, industrial applications.

3.3 3d results

In fig. 7, we present results for a 3d grid that consists of 12,296,192 tetrahedral

elements with 4 levels of refinement. The obstacle is composed of 54,784 triangular
faces. A 𝑃1 − 𝑃1 discretization is used across all PDE systems, with a stabilization
term for the Navier–Stokes equations, which is necessary for numerical stability. The
viscosity is set to 1, while the box constraints imposed on the extension factor 𝜂ext
are 0 ≤ 𝜂ext ≤ 60.
The reference configuration is chosen as a spherical obstacle inside a flow tunnel
and the creation of singularities is demonstrated in fig. 7. Given the high levels of
computation for these simulations, it was necessary to use an average of 4 nodes
with at least 80 cores each. Moreover, the success of the simulations, in terms of the
generated shape, could only be determined after a high number of optimization steps,
usually more than 100.

4 Scalability results

Weak scaling results for the 2d problem shown in Fig. 2 were performed on the HPE
Hawk supercomputer at HLRS. The machine features 5,632 nodes, each with two
AMD EPYC 7742 CPUs consisting of 64 cores per CPU. In fig. 8, we present the
accumulated timings and iteration counts for the first 3 optimization steps for up to
5,120 cores and more than 6 million elements. ParMETIS [10] is employed for load
balancing. Following the recommendation from the Hawk online documentation, we
chose to base our results on an 80 core count per node and increase the cores fourfold
upon each mesh refinement.
We present results for the solution of the extension equation (13) to (15), for the
shape derivative of (16), and the gradient of the extension factor system of equations,
shown below:
1
𝑀 + 𝜉 (𝜂 − (𝜂lb + 𝜂ub )) − (𝐷w · w) · 𝜓w = 0.
2
Scalable Multigrid Algorithm for Shape Optimization 491

Fig. 7: High viscosity 3d simulation

The nonlinear extension equation is solved using a Newton method, while the linear
problems are solved with a BiCGStab method preconditioned with a geometric
multigrid method. A Jacobi smoother is used within the multigrid, with 3 pre- and
post-smoothing steps in a V-cycle. The base level is solved on a single process by a
serial LU solver in all cases. The convergence criteria for the linearization within
the Newton solver are a maximum of 2000 iterations or an error reduction of 10−14 .
Whereas, for the Newton method they are a maximum of 20 steps or an error reduction
of 10−12 . The linear solvers must fulfill a maximum of 2000 steps or an error reduction
of 10−16 .
Figure 8 presents timings and speedup for the solve phases of the Newton
(newton) and linear solvers (solve), the multigrid initialization (init), and the fine
matrix assembly (ass). A good weak scaling is found in general, but sub-optimal
degradations are observed for the Newton solver of the extension equation, the
assembly phases of the shape derivative and the extension gradient. The iteration
counts remain constant for the Newton method steps and the linearization problem
even for more than 6 million elements. The slight variation in the shape derivative
492 J. Pinzon, M. Siebenborn and A. Vogel

could be related to the numerical differences caused by the PDE system’s dependency
on integration point data from the solution of other grid functions. At the same time,
the extension gradient performance loss could be related to the selection of the cores
per node within the same hypercube topology. The timings increase for 16 and 64
nodes, this might suggest that the sweet spot in terms of cores per node is yet to be
achieved and must be further investigated.

500
𝑇lin,init 512 𝑆lin,init
𝑇lin,solve 𝑆lin,solve
400 𝑇newton 𝑆newton
𝑆ideal

Speedup (weak)
64
Time [ms]

300

200 8

100
1
0
80 320 1280 5120 80 320 1280 5120
Number of processes Number of processes

(a) Extension Equation Non-Linear Solver

500
512
𝑇init 𝑆init
𝑇ass 𝑆ass
400 𝑇solve 𝑆solve
𝑆ideal
Speedup (weak)

64
Time [ms]

300

200 8

100
1

0
80 320 1280 5120 80 320 1280 5260

Number of processes Number of processes

(b) Shape Derivative Linear Solver

Fig. 8: (continued on next page).

5 Conclusion and outlook

We presented an optimization result for a 2d square obstacle in a Navier–Stokes flow

channel. It has been demonstrated that the employed numerical scheme is able to
remove the edges present in the reference configuration, as well as introducing new
ones through the accumulation of the boundary control variable and extension factor.
Scalable Multigrid Algorithm for Shape Optimization 493

120
𝑇init 512 𝑆init
100 𝑇ass 𝑆ass
𝑇solve 𝑆solve
𝑆ideal

Speedup (weak)
80 64
Time [ms]

60
8
40

20
1
0
80 320 1280 5120 80 320 1280 5120
Number of processes Number of processes

(c) Extension Gradient Linear Solver

Procs Refs NumElems Linear solver Newton solver Linear solver

(shape derivative) (deformation field) (extension gradient)

80 4 105,472 56 12 9
320 5 421,888 70 12 9
1,280 6 1,687,552 77 12 9
5,120 7 6,750,208 77 12 9

(d) Accumulated iteration counts for the weak scaling study

Fig. 8: Weak Scaling: For the first three optimization steps, the accumulated wallclock
time is shown for: (a) the nonlinear extension equation, (b) the derivative of the
objective function with respect to the deformation field, and (c) the extension factor
gradient, equation given in Sec. 4. In (d), the accumulated iteration counts are
presented for the geometric multigrid preconditioned linear solver of the shape
derivative, the number of Newton steps and linear iterations necessary to solve the
extension equation and its linearization

The results for the objective function reflect the fact that, after the Lagrange multipliers
have iteratively calculated within a certain tolerance, the functional converges to a
minimum. Corresponding 3d results demonstrate that the algorithm is not restricted
to creating singularities on obstacle contours with edges, but can also achieve similar
patterns with surface elements.
A grid study emphasizes that the computed optimal shape is independent of the
number of refinements. This allows the usage of the geometric multigrid method with
grid-independent convergence. Although good weak scaling speedup results have
been achieved for up to 5,120 cores and more than 6 million elements, an optimal
494 J. Pinzon, M. Siebenborn and A. Vogel

parallel setup for Newton’s method and assembly phases has to be further investigated.
To this end, better core counts per node have to be identified, taking into account the
hypercube topology of the HPE Hawk supercomputer.
As a next step, we plan to investigate more detailed 3d configurations with higher
levels of refinement to extend the weak scaling studies and to demonstrate the
applicability of the method for large-scale, real-world applications.

Acknowledgements Computing time on the supercomputer Hawk at HLRS under the grant
ShapeOptCompMat (ACID 44171, Shape Optimization for 3d Composite Material Models) is
gratefully acknowledged.
The current work is part of the research training group ‘Simulation-Based Design Optimization of
Dynamic Systems Under Uncertainties’ (SENSUS) funded by the state of Hamburg under the aegis
of the Landesforschungsförderungs-Project LFF-GK11.

References

1. URL www.paraview.org
2. Brandenburg, C., Lindemann, F., Ulbrich, M., Ulbrich, S.: A Continuous Adjoint Approach
to Shape Optimization for Navier Stokes Flow. In: K. Kunisch, G. Leugering, J. Sprekels,
F. Tröltzsch (eds.) Optimal Control of Coupled Systems of Partial Differential Equations,
Internat. Ser. Numer. Math., vol. 160, pp. 35–56. Birkhäuser, Basel (2009)
3. Elman, H., Silvester, D., Wathen, A.: Finite Elements and Fast Itertative Solvers With Applica-
tions in Incompressible Fluid Dynamics, vol. 1. Oxford Science Publications (2014)
4. Garcke, H., Hinze, M., Kahle, C.: A stable and linear time discretization for a thermodynamically
consistent model for two-phase incompressible flow. Applied Numerical Mathematics 99,
151–171 (2016)
5. Geuzaine, C., Remacle, J.F.: Gmsh: A 3-D finite element mesh generator with built-in pre- and
post-processing facilities. International Journal for Numerical Methods in Engineering 79(11),
1309–1331 (2009). DOI 10.1002/nme.2579
6. Giles, M., Pierce, N.: An introduction to the adjoint approach to design. Flow, turbulence and
combustion 65(3-4), 393–415 (2000)
7. Haubner, J., Siebenborn, M., Ulbrich, M.: A continuous perspective on shape optimization via
domain transformations. SIAM Journal on Scientific Computing 43(3), A1997–A2018 (2021).
DOI 10.1137/20m1332050
8. Iglesias, J.A., Sturm, K., Wechsung, F.: Two-dimensional shape optimization with nearly
conformal transformations. SIAM Journal on Scientific Computing 40(6), A3807–A3830
(2018)
9. Jameson, A.: Aerodynamic shape optimization using the adjoint method. Lectures at the Von
Karman Institute, Brussels (2003)
10. Karypis, G., Schloegel, K., Kumar, V.: Parmetis, parallel graph partitioning and sparse matrix
ordering library (2013). URL http://glaros.dtc.umn.edu/gkhome/metis/parmetis
/overview
11. Müller, P.M., Kühl, N., Siebenborn, M., Deckelnick, K., Hinze, M., Rung, T.: A novel 𝑝-
harmonic descent approach applied to fluid dynamic shape optimization (2021)
12. Mohammadi, B., Pironneau, O.: Applied shape optimization for fluids. Oxford university press
(2010)
13. Murat, F., Simon, J.: Etude de problèmes d’optimal design. In: J. Cea (ed.) Optimization
Techniques Modeling and Optimization in the Service of Man Part 2: Proceedings, 7th IFIP
Conference Nice, September 8–12, 1975, pp. 54–62. Springer-Verlag, Berlin, Heidelberg (1976)
Scalable Multigrid Algorithm for Shape Optimization 495

14. Onyshkevych, S., Siebenborn, M.: Mesh quality preserving shape optimization using nonlinear
extension operators. Journal of Optimization Theory and Applications 16(5), 291—-316 (2021).
DOI 10.1007/s10957-021-01837-8
15. Pinzon, J., Siebenborn, M.: Fluid dynamic shape optimization using self-adapting nonlinear
extension operators with multigrid preconditioners (in preparation)
16. Reiter, S., Vogel, A., Heppner, I., Rupp, M., Wittum, G.: A massively parallel geometric
multigrid solver on hierarchically distributed grids. Comp. Vis. Sci. 16(4), 151–164 (2013)
17. Schmidt, S., Ilic, C., Schulz, V., Gauger, N.R.: Three-dimensional large-scale aerodynamic
shape optimization based on shape calculus. AIAA journal 51(11), 2615–2627 (2013)
18. Vogel, A., Calotoiu, A., Strube, A., Reiter, S., Nägel, A., Wolf, F., Wittum, G.: 10,000
performance models per minute – scalability of the UG4 simulation framework. in: Euro-Par
2015: Parallel Processing, J. L. Träff et al. (eds), Springer pp. 519–531 (2015)
19. Vogel, A., Reiter, S., Rupp, M., Nägel, A., Wittum, G.: UG 4: A novel flexible software system
for simulating PDE based models on high performance computers. Comp. Vis. Sci. 16(4),
165–179 (2013)
Numerical calculation of the lean-blow-out in a
multi-jet burner

Alexander Schwagerus, Peter Habisreuther and Nikolaos Zarzalis

Abstract The report presents the results of the project TurboRe, which focuses on
the numerical investigation of lean-blow-out phenomenon in a multi jet flame array.
Successful calculations have been conducted for a model combustor and the developed
numerical setup has been applied for the calculation of blow-out in a complex burner.

1 Introduction

The project TurboRe focuses on the numerical calculation of Lean-Blow-Out (LBO)

of a premixed and turbulent burner consisting of an array of jet flames (matrix
burner) using large eddy simulation and two different approaches of modelling
the turbulence-flame-interaction. The main objective is to simulate turbulent lean
premixed jet-flames and determine their LBO-limits as a function of basic geometrical
and thermo-dynamical parameters. Simulations are validated through corresponding
experiments and correlations are to be determined. The results of these investigations
show the ability to numerically predict LBO and lead to a better understanding of
the flame stability in matrix burner arrangements. The investigation provides crucial
information for the design and scaling of the matrix burner. In the last several decades,
a swirl-stabilized flame has been the preferred choice in gas turbine combustors due
to their aerodynamic means of stabilization. However, swirl flames can be susceptible

Alexander Schwagerus
DVGW Research Center, Engler-Bunte-Institute of the Karlsruhe Institute of Technology, Karlsruhe,
76131, Germany, e-mail: [email protected]
Peter Habisreuther
Karlsruhe Institute of Technology, Institute for Technical Chemistry, Eggenstein-Lepoldshafen,
76344, Germany, e-mail: [email protected]
Nikolaos Zarzalis
Karlsruhe Institute of Technology, Engler-Bunte-Institute, Combustion Technology, Karlsruhe,
76131, Germany, e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 497
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_31
498 Alexander Schwagerus, Peter Habisreuther and Nikolaos Zarzalis

to the issue of combustion instability in the lean combustion regime leading to

lean-blow-out (LBO) and therefore damaging the hardware. The corresponding task
of the EU-project Turbo-Reflex focuses on improving the low load capability of
stationary gas turbines to reduce the number of hot starts. By having a higher load
capability, plants can be run more flexible without being forced to shut-down/start-up
completely, resulting in less cycling, and therefore leading to a longer lifespan of the
turbine in the increasingly volatile energy demand environment. A matrix burner
(arrangement of jet flames) is regarded as an alternative to the swirl burners mentioned
earlier. However, since such matrix burners have only scarcely been investigated,
further understanding of such systems is needed to correctly describe and scale them.
This was the motivation to carry out a test campaign involving velocity, exhaust gas
and LBO measurements at the combustion technology department of the Engler-
Bunte-Institute of the KIT. A sketch of the used nozzles can be seen in Figure 1. For
detailed information regarding the experiments and the investigated combustor, the
reader is referred to previous publications [1, 6]. To complement the experimental
investigation, detailed numerical simulations have been performed to further deepen
the understanding of the dominating flame stability effects in this highly complex
combustion system.

000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00

TĚ,ŽůĞ

Fig. 1: Sketch of the investigated nozzles (from left: 𝐷 1 , 𝐷 2 and 𝐷 3 ) [1]

For the simulation of the flame, a model for the calculation of the reactive source
terms is required to calculate the distribution of temperature and species concentrations.
Since detailed kinetic mechanisms usually contain hundreds of elementary reactions,
this approach nowadays still surpasses the computation capacities generally available
for complex, realistic geometries and therefore is limited to one-dimensional flames
or simple laminar systems. Two methods to reduce the computational effort have been
used and compared, that enable to calculate the reactive system by the introduction a
quantity for the reaction progress.
Numerical calculation of the lean-blow-out in a multi-jet burner 499

2 Numerical setup

The main task is the numerical calculation of the lean-blow-off (LBO) limits.
For that purpose, the open-source C++ libraries of OpenFOAM have been used.
The solver used for calculating the reactive flow is based on the already existing
OpenFOAM solver rhoPimpleFoam. It solves the compressible formulation of the
Navier-Stokes equations. The pressure-velocity coupling is implemented through the
PIMPLE method. Among the different approaches for the description of turbulence
by numerical models, the LES (Large-Eddy-Simulation) technique was chosen, which
enables the investigation of transient flows. This is considered useful, since LBO is
also a transient phenomenon. Computational grids for four different nozzle have been
created. The numerical grids consists of each around 1.3 mio. cells. The turbulence
of the inlet has, as already shown in literature [5] and in preliminary simulations, a
huge influence on the flow in the wake, but the specification of a suitable velocity
field at the inlet is a non-trivial problem in LES-simulations. For this work, the
turbulence generator proposed by Klein et al. [4] has been applied and implemented
in OpenFoam. This method is based on digital filtering of a series of uncorrelated
random data to generate correlated velocity fields according to user-defined turbulence
properties. In order to reduce the computational effort and enable to the computation
within a reasonable timeframe, a FGM (flame-generated manifold) approach is used,
which reduces the reaction mechanism to a pre-tabulated reaction progress variable Θ.
The turbulent-chemistry interaction (TCI) were captured by two different combustion
models. In both models, a transport equation of a reaction progress variable is solved.
The difference between the combustion models is in the source term modelling. The
first model (JPDF model) involves the calculation of the source term via a presumed
density function (PDF) of the reaction progress variable. In order to link the reaction
state to the reaction progress variable a model reactor, like the one-dimensional
premixed flame, is needed. This allows a tabulation of the source term as a function
of the progress variable itself. The calculation of the necessary mean source term can
then be conducted by integrating it with the PDF. The PDF is exactly defined by its
mean and variance. Therefore two transport equation for the mean and the variance
of the reaction progress variable need to be solved. This is already implemented in
a solver developed at the Engler-Bunte-Institute and have been already successful
used in different numerical investigations [2, 3]. For a more detailed description, the
reader is referred to literature, e.g. the dissertation of Kern [3]. The second model
(TFC model) calculates the source term through the KPP-theorem, which states that
the source term can be calculated on basis of the turbulent flame speed, which is a
measure of the turbulent volumetric conversion rate. The turbulent flame speed is
calculated using the H.P. Schmid model [3] as a function of the Damkoehler number
and the laminar flame speed.
500 Alexander Schwagerus, Peter Habisreuther and Nikolaos Zarzalis

3 General description of the flame

For comparison, the calculated distribution of the reaction progression variable of

the flames under stable near stoichiometric conditions is shown in Fig. 2 for a median
cut. Here, the unburned mixture (marked in blue) flows from left into the combustion
chamber and is reacting then to fully burnt exhaust gases (light red color). Simulations
of the JPDF model (top) and the TFC model (bottom) are compared at the same
operating conditions. Snapshots are shown on the left and time averaged fields on
the right side. In these simulations, only about half of the combustor length was
discretized. The results of both combustion models show that the jets are very straight,
similar to what was observed in the experiments. The simulations of the JPDF model
shows separated reaction zones for each jet, while the TFC model predicts more
overlaps of the individual jets. In addition, there are considerable differences in the
width of the reaction zone: While the JPDF model shows very thin flame fronts, the
flame zone of the TFC model extends up to the end of the combustion chamber. This
is mainly because the second half of the reaction progress (the change from Θ = 0.5
to 1) is predicted to be slower. The TFC model does not achieve complete burnout in
the shown part of the combustion chamber, whereas the burnout is complete using
the JPDF. Although the flames differ strongly in the main and post-flame zone, the
flame root is predicted to be similar. This can be seen in the averaged images, in
which both combustion models show the turquoise and green marked flame zones
in almost identical axial positions. This fact is even more visible in Fig. 3 showing
the iso-surface Θ = 0.25. It can be seen that the predicted flame shape for the jets
at their root is almost identical in both models. The flame root is in most cases an
anchor point of the flame and is therefore decisive for flame stability. The beginning
of the flame is predicted to look very similar in both models, while the flames differ
particularly in the main and afterburning zone. The impact of these effects on the
determination of LBO are investigated in the following section.

Fig. 2: Comparison of the simulated reaction progress distribution between the JPDF-
and TFC-model for the nozzle 𝐷 2 at 𝑇0 = 100 °C, 𝑢 𝐻𝑜𝑙𝑒 = 35 m/s and 𝜆 = 1.15
Numerical calculation of the lean-blow-out in a multi-jet burner 501

Fig. 3: Flame root indicated by the iso-surface Θ = 0.25 simulated by the two
combustion models for the nozzle 𝐷 2 at 𝑇0 = 100 °C, 𝑢 𝐻𝑜𝑙𝑒 = 35 m/s and 𝜆 = 1.15

4 LBO results

Above, only flames at constant operating conditions have been investigated. In order
to determine the explicit blow-off points, transient inlet conditions are set, which
enables the determination of a blow-off point using only one simulation. There are
two basic methods to induce LBO: On the one hand, the mass flow of fuel can be
reduced, which leads to an increase of the air equivalence ratio 𝜆 and, consequently,
a reduction of the flame speed until the flame blows off. LBO has been induced by
increasing the total mass flow, characterized by the average inlet velocity 𝑢 𝐻𝑜𝑙𝑒 . A
stepwise gradual increase in velocity was implemented. A schematic course of the
input velocity of a simulation is shown in Fig. 4. In this example, the velocity is
kept constant for Δ𝑡 = 40 ms and is then increased by Δ𝑢 𝐻𝑜𝑙𝑒 = 4 m/s. In order to
observe whether the time of constant velocity is long enough, global quantities are
observed and checked whether a convergence is found within this time interval. The
schematic plot also shows that the velocity increase does not occur instantaneously
but in a steep rise. This prevents the creation of numerical pressure waves due to
abrupt velocity changes, which could erroneously cause LBO too early. The overall
LBO simulation always start at stable conditions (low velocity) and then are increased
in the aforementioned stepwise way until LBO is observed.
As has already been shown, there are differences in the representation of the
reaction zone and the jet flame interactions between the two combustion models. It
is expected, that the flame shape is decisive for the behavior approaching unstable
conditions and therefore also for the choice of the LBO criterion. In this section, it is
examined how these two models exhibit LBO and how accurately they are able to
reproduce the experimental data. For this purpose LBO simulations were conducted
for the nozzle 𝐷 2 at a preheating temperatures of 𝑇0 = 100°C. The corresponding
experimental LBO data show a blow-off at an air equivalence ratio of 𝜆 = 1,661 at
an inlet speed of 𝑢 𝐻𝑜𝑙𝑒 = 53 m/s. First the LBO simulations with the JPDF model
are discussed. The simulations were initialized with an inlet velocity of 𝑢 𝐻𝑜𝑙𝑒 = 35
m/s and were increased by Δ𝑢 𝐻𝑜𝑙𝑒 = 4 m/s every Δ𝑡 = 30 ms. Using these settings,
the experimentally measured LBO condition is reached after 0.15 s simulated time.
502 Alexander Schwagerus, Peter Habisreuther and Nikolaos Zarzalis

Fig. 4: Schematic change of the inlet velocity for the LBO procedure

Figure 5 shows snapshots of the reaction progress variable from the resulting LBO
simulation at different inlet velocities. At these conditions, the flame forms a conical
flame as shown in the upper picture. In the middle picture, where the inlet velocity
is increased to 𝑢 𝐻𝑜𝑙𝑒 = 43 m/s, a qualitatively similar conical flame can be seen,
which still burns stable but already extends considerably further into the combustion
chamber. The bottom picture shows the effect of a further inlet velocity increase to
𝑢 𝐻𝑜𝑙𝑒 = 47 m/s: LBO occurs and the flame almost completely disappears. There is
still ongoing reaction in the recirculation zones and at the edges of the outlet. The
fact that reaction is still occurring at the edge of the combustion chamber outlet can
be explained by the use of the outer domain (not shown in the pictures): This outflow
zone causes a slow-down of the gas mixture due to the cross-sectional expansion. In
reality, a dilution by sucked in air would take place, which would lead to a massive
reduction of the flame speed and would prevent flame stabilization. Since the solver
assumes a perfectly premixed composition, the flame always stabilizes in these low
velocity regions and therefore is able to occasionally retaliate back into the combustion
chamber. In spite of this fact, LBO can be still clearly detected, indicating that the
flame stabilization in the outer domain is no problem for the simulation of LBO.
In summary, the JPDF model is able to reproduce a blow-off behavior, marked
by a sudden extinction of the flame after reaching a critical velocity. The blow-off
occurs between 𝑢 𝐻𝑜𝑙𝑒 = 43 m/s and 47 m/s, which is about 15 % lower than the
measured value 𝑢 𝐻𝑜𝑙𝑒 = 53 m/s. This good agreement with the measured value is
particularly surprising as no heat loss model is included in this simulation. It is to
be expected that the calculated values will decrease and thus deviate even further
from the measured values if heat losses are included. This is the reason no more
LBO simulation were conducted with the inclusion of heat losses and is a strong
indication that heat losses are not dominant for flame stability in the current multi jet
burner system. The following section discusses the LBO simulation using the TFC
model. The TFC-model predicts long reaction zones where complete burnout is not
Numerical calculation of the lean-blow-out in a multi-jet burner 503

Fig. 5: Snapshots of the reaction progress variable inside the combustion chamber
from the LBO simulation with the JPDF-model at different inlet velocities (from top
𝑢 𝐻 𝑜𝑙𝑒 = 35 m/s, 43 m/s and 47 m/s)

reached until the end of the combustion chamber, even for velocity conditions far from
experimentally measured LBO conditions. The effect of this flame pattern predicted
by the TFC model, that is being different from the one observed with the JPDF model,
on the approach of unstable operating conditions will be discussed in this section. For
the LBO simulations with the TFC-model, the same conditions were chosen as for
the JPDF-LBO simulation (nozzle 𝐷 2 at 𝑇0 = 100 °C and 𝜆 = 1.661). The inlet speed
𝑢 𝐻𝑜𝑙𝑒 begins at 30 m/s and is increased incrementally by Δ𝑢 𝐻𝑜𝑙𝑒 = 5 m/s every Δ𝑡 =
40 ms. The experimentally determined blow-off velocity is again 𝑢 𝐻𝑜𝑙𝑒 = 53 m/s, an
inlet velocity which is reached after 0.2 s of simulated time. Snapshots of the reaction
progress of the flame at different inlet velocities are shown in Figure 6. The top
picture shows the starting flame at 𝑢 𝐻𝑜𝑙𝑒 = 30 m/s, characterized by long single jets,
where the outer jets are significantly shorter than the inner ones. In the course of the
velocity increase, two different stages are observed: when comparing the snapshots
up to 𝑢 𝐻𝑜𝑙𝑒 = 50 m/s, it is noticeable that the flame root (illustrated by the end of
the blue zone) remains almost at the same location. While the root of the flame is
stationary at this stage, the reaction zone expands downstream, leading to a reduction
of burnout at the exit. Afterwards, especially between the third and fourth image,
which shows the increase in velocity at which LBO was measured experimentally, it
can be observed that the flame root of the outer jets shifts significantly downstream
(marked by black ellipses). From this point on, a displacement of the entire flame
504 Alexander Schwagerus, Peter Habisreuther and Nikolaos Zarzalis

takes place, which can be seen in particular in comparison to the last image, in which
the unburned area (marked by blue color) extends almost to the end of the combustion
chamber. However, even under these conditions a stable flame root is still predicted
at the end of the combustion chamber. A critical velocity that suddenly leads to a
disappearance of the flame is, in contrast to the JPDF-model, not observed here. For
this reason, it is difficult to find a suitable blow-off criterion in order to determine
a clear blow-off velocity. If possible, a simple and global criterion should be used.
Different globally averaged variables are considered for this purpose.

Fig. 6: Snapshots of the reaction progress variable inside the combustion chamber
from the LBO simulation with the TFC-model of the nozzle 𝐷 2 at increasing inlet
velocities

To give an overview, various possible globally averaged quantities are shown

in Figure 7. In the upper two diagrams, the temperature and the reaction progress
variable are shown followed by the mass fractions of CO2 and OH. As can be seen,
all these quantities decrease continuously with an increase in velocity. The presence
of OH is often used as an indicator for reaction and, as the graph shows, the mean OH
Numerical calculation of the lean-blow-out in a multi-jet burner 505

concentration converges with increasing velocity towards zero. It is an obvious idea

to define a limit value, where the flame is considered to be “blown off”. However, it
is difficult to define a general limit here and so a different LBO-criterion is required.

Fig. 7: Progression of globally volume averaged quantities over time in the LBO
simulation with the TFC-model for the nozzle 𝐷 2 at 𝑇0 = 100 °C and 𝜆 = 1,661

In the last diagram of Figure 7 the reaction rate is shown. Due to the increase
of the incoming mass flow of fuel, the reaction rate initially increases. This is the
previously described first stage in which the reaction zone expands, while maintaining
a stationary flame root. By slowly discharging the flame in the second phase, the
increased fuel flow can no longer be converted to the same extent in the combustion
chamber, which results in a reduction of the reaction rate. It is reasonable to define the
transition point between increase due to increasing fuel mass flow and the discharging
of the flame as LBO. This characteristic point of consumption behavior can be
506 Alexander Schwagerus, Peter Habisreuther and Nikolaos Zarzalis

determined very precisely by the beginning of the decline of the reaction rate and
takes place at the velocity increase from 50 m/s to 55 m/s, which corresponds exactly
to the experimentally determined value of 53 m/s. The resulting LBO points of all
calculated conditions and used nozzles are plotted in a Peclet diagram in Figure 8.
The diagram presents, in addition to experimentally determined LBO values (stars),
the LBO points calculated with the JPDF model (circles) and those calculated with
the TFC model (diamonds). It can be seen that for the nozzles at a DR = 2.8 there
is very good agreement between both TCI models and the experimental data over
the entire measured Peclet range. The JPDF model predicts the LBO values with an
almost perfect agreement with the Peclet curve for the nozzles 𝐷 2 and 𝐷 3 , while
smaller deviations occurred at the highest 𝑃𝑒 𝑢 values of the nozzle 𝐷 1 . Here, the
highest velocities are present and the observed discrepancy may possibly be due to
the cell resolution, which has been kept constant and may not be able to sufficiently
resolve the turbulence at these large velocities. While the JPDF model in a small
range both underestimates and overestimates the LBO values, the TFC model predicts
the LBO values at generally higher blowout velocities compared to the experiments.
The distance of the TFC points for the DR = 2.8 nozzles to the Peclet curve is
almost constant. The deviation was quantified in more detail in the following: As
measured values show some scatter due to the probabilistic character of the underlying
turbulent flow, it is useful to compare the numerically calculated LBO limits not
with the individual measured data but with the experimentally determined Peclet
correlation, which provides an average of the measured values. The relative deviations
between the numerical LBO points and the experimental Peclet curve were calculated
and are shown in Table 1: Comparison of the numerically calculated LBO values
to the experimental Peclet curve for the nozzles at DR = 2.8. It can be seen that
the JPDF model on average predicts the LBO limits more accurately with a mean
deviation of about 11% compared to the TFC model with a mean deviation of about
15%. Both values are comparable to the 6% mean deviation of the experimental
values themselves. However, the TFC model scatters much less here, as evidenced by
the low variance. The higher variance of the JPDF model is mainly caused by the
overestimation of the LBO limits at high Peclet numbers.

Table 1: Comparison of the numerically calculated LBO values to the experimental

Peclet curve for the nozzles at DR = 2.8

JPDF-Model TFC-Model

Mean relative deviation 11.3 % 14.8 %

Variance relative deviation 1.0 % 0.3 %
Numerical calculation of the lean-blow-out in a multi-jet burner 507

Two additional calculations to check the prediction of the dump ratio influence
were performed with both models for nozzle 𝐷 5 , which has a higher DR of 6 (red
dots). Here, the JPDF model can predict the drop to lower 𝑃𝑒 𝑢 values, while the TFC
model calculates the blowout limits at significantly higher values. Thus, the influence
of the DR on the LBO limits cannot be predicted by the TFC model. It could be
shown that a determination of the blowout limits and thus of the Peclet parameters is
possible by CFD simulations with only moderate grid resolution. Both TCI models
allow to predict the LBO limits for the nozzles at DR = 2.8 and this despite the fact
that both models represent the flame shape as well as the blowout process itself very
differently. The mean deviation between the resulting LBO is quite small for both
models. However, the influence of a variation of the dump ratio on the LBO limits
can only be described using the JPDF model.

Fig. 8: Comparison of the numerically calculated LBO limits to the experiments in a

Peclet diagram

5 Calculation of an industrial burner

The calculation of the model burner already showed that the investigated models are
able to calculate the blowout of the flame with good accuracy. Anyhow, the calculation
of industrial gas turbine combustors imposes a much bigger challenge, since the
geometries are larger and considerably more complex in their overall geometry. As
an additional difficulty, in these cases, additional air is usually introduced into the
508 Alexander Schwagerus, Peter Habisreuther and Nikolaos Zarzalis

combustion chamber to cool down the walls and regulate the flame temperature
and the combustors are operated with strongly preheated gases and at elevated
pressure conditions. The goal was therefore to investigate whether the developed
numerical setup can also be used to calculate LBO for a complex multi-jet flame
burner under industrially more realistic operating conditions. To this end, the blowout
point of an industrial prototype burner developed and experimentally investigated
by Siemens was calculated numerically using the developed models. Since this
burner has additional air inlets into the combustion chamber, the additional impact of
the spatially and temporally varying mixture field on the reaction is required. The
JPDF model already inherently includes the effect of mixing by the aid of the first
and second statistical moment (the mean and the variance) of the mixture fraction.
Because of this feature, and as, additionally, it was shown that this model was the
only one capable of correctly predicting the effect of DR variation, this model was
used to simulate the LBO point for the industrial set-up. The burner consists of a
multitude of non-swirling jet flames and includes a swirled pilot burner in the center.
The combustor was operated with natural gas at significantly elevate pressure levels
and high preheating temperature. The complete numerical domain consists of 33 mio.
cells. The experimental LBO point has been recalculated using the numerical methods
developed to show its applicability to industrial burners. In contrast to the former
determination of the LBO point of the model matrix burner, for determination of the
LBO in the industrial set-up the air mass flow rate was set identical to the experiment,
while the flame stabilized at this part-load condition. To induce LBO as the second
step the fuel mass flow rate was then stepwise decreased until LBO of the flame
could be observed. In Figure 9, meridian slices through the combustion chamber,
showing the temperature distributions are shown for three different equivalence
ratios inside the mixing tubes and is set leaner from top to bottom. In all pictures
only the two inner fuel inlets are activated, which can be seen by the slightly lower
temperature (dark blue) compared to the other inlets (dark cyan), due to the lower
temperature of the natural gas compared to the preheated air. At Φ = 0.77, a wide
flame is formed, occupying almost the entire combustion chamber diameter. When
the equivalence ratio is reduced to Φ = 0.57, the diameter of the flame decreases,
while still maintaining a stable flame, which is mostly located in the central inner
recirculation zone. At these conditions, LBO was observed in the experiment, but
as can be clearly seen, the simulation still predicts a stable flame. Only when the
equivalence ratio is further reduced to the next equivalence ratio Φ = 0.48, as can be
seen in the lowest picture, the flame can no longer stabilize, and the reaction zone
disappears. At these conditions, there is still a slightly elevated temperature in the
recirculation zone (around 100 °C above the inlet temperature), but can be explained
by the long residence time in this zone and the associated slow response. It can
be assumed that if for this operation point the calculation is continued for a longer
time a uniform temperature field would form. Nevertheless, the blowout is clearly
noticeable visually. Summarizing the observations, it could be shown that even for
high pressure level and complex geometry, which are important characteristics for
Numerical calculation of the lean-blow-out in a multi-jet burner 509

industrial applications, the developed numerical setup using the JPDF model can be
used to predict LBO. Flame blowout was calculated exactly in the experimentally
observed interval.

Fig. 9: Temperature distribution inside the combustion chamber at different equiva-

lence ratios

6 Performance and required resources

Figure 10 presents the results of a performance and scalability test carried out on the
ForHLR II. For this test, a configuration with 28 million cells was used. It can be seen
that the code scales well up to 1000 cores and reasonably up to 2000 cores which
corresponds to a minimum required number of cells per core of 14 000. Because
the grid size of the model combustor (the matrix burner) with a total cell count of
1.3 mio. was rather small typically only 200 processors have been used. Due to the
510 Alexander Schwagerus, Peter Habisreuther and Nikolaos Zarzalis

number of LBO points calculated and the needed long physical time to reach LBO
around 1 mio core-h have been used to achieve the LBO calculation of the matrix
burner. Lastly, the simulation of the industrial burner with a cell count of 33 mio.
cells has been conducted with a total physical time of 0.7s that has been simulated to
reach LBO. The case was calculated with typically 1000 processors and resulted in
around 4.2 mio. core-h used.

100
90
Speed-up normalized to 40 procs

80
28 mio. cells Ideal
70
60
50
40
30
20
10
0
0 1000 2000 3000
Number of cores

Fig. 10: Scaling of the used reactive Solver in OpenFOAM on the supercomputer
FORHLR II using different number of cores

7 Conclusion

The current report demonstrates the applicability of two different combustion models
for the calculation of the reacting cases under investigation. It describes the results of
the reactive flow simulations that were applied for LBO calculations of a wide range
of operating conditions. Lastly, an LBO simulation for a highly complex industrial
burner has been successfully conducted. The numerically calculated LBO limit is in
good agreement to the experimental value.

References

1. Robbin Bhagwan, Alexander Schwagerus, Christof Weis, Peter Habisreuther, Nikolaos Zarzalis,
Michael Huth, Berthold Koestlin, and Stefan Dederichs. Combustion characteristics of natural
gas fueled, premixed turbulent jet flame arrays confined in a hexagonal combustor. In Turbo
Expo: Power for Land, Sea, and Air, volume 58615, page V04AT04A018. American Society of
Mechanical Engineers, 2019.
2. Fabian Eiberger. Wechselwirkung zwischen turbulenz und wärmestrahlung, 2018.
3. Matthias Kern. Modellierung kinetisch kontrollierter, turbulenter Flammen für Mager-
brennkonzepte. KIT Scientific Publishing, 2013.
Numerical calculation of the lean-blow-out in a multi-jet burner 511

4. Markus Klein, Amsini Sadiki, and Johannes Janicka. A digital filter based generation of inflow
data for spatially developing direct numerical or large eddy simulations. Journal of computational
Physics, 186(2):652–665, 2003.
5. Andrea Montorfano, Federico Piscaglia, and Giancarlo Ferrari. Inlet boundary conditions for
incompressible les: A comparative study. Mathematical and Computer Modelling, 57(7-8):1640–
1647, 2013.
6. C. Weis, A. Schwagerus, S. Faller, R. Bhagwan, P Habisreuther, and N. Zarzalis. Determination
of a correlation for predicting lean blow off limits of gaseous fueled, premixed turbulent jet flame
arrays enclosed in a hexagonal dump combustor. page S5_AIII_48, 2019.
Data-driven multiscale modeling of self-assembly
and hierarchical structural formation in
biological macro-molecular systems

P.N. Depta, M. Dosta and S. Heinrich

Abstract Self-assembly and hierarchical structural formation are essential in many

systems of both nature and technology. Examples are the dynamic self-assembly
of multi-protein clusters and the interdependency with catalytic activity, as well as
the self-assembly of virus capsids critical for virus function. In an attempt to better
understand and model these systems, we develop a multiscale modeling methodology
to capture macro-molecular self-assembly on the micro-meter and milli-second scale.
Focus of this report is the derivation and implementation of data-driven interaction
potentials based on molecular dynamics simulations using Universal Kriging on the
example of the hepatitis B core antigen (HBcAg).

1 Introduction

The dynamic self-assembly of multi-protein structures has received increasing interest

in recent years [1]. This is especially true for the inter-dependency of structure and
function. Understanding these processes remains a challenge and is critical in enabling
the targeted modification and optimization of e.g. efficient bioreaction cascades [2].
Additionally, such hierarchical structural formation processes are at the basis of
most macromolecular systems [3]. Namely, this includes material science and also
structures of physiological systems in the human body, e.g. viruses. In the context of
this work, viruses are investigated in the scope of virus-like-particles (VLP), which
are formed out of monomers (primary proteins) of their structural capsid proteins
providing critical properties for their function [4]. Examples are the hepatitis B viral
protein (HBcAg) [5], which we are working on as an application for this framework,
but potentially also, in light of current events, the SARS-CoV-2 virus.

Philipp Nicolas Depta, Maksym Dosta and Stefan Heinrich

Institute of Solids Process Engineering and Particle Technology, Hamburg University of Technology,
e-mail: [email protected], [email protected], [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 513
W. E. Nagel et al. (eds.), High Performance Computing in Science and Engineering ’21,
https://doi.org/10.1007/978-3-031-17937-2_32
514 P.N. Depta, M. Dosta and S. Heinrich

The main challenge of investigating macro-molecular structural formation both

experimentally and by means of simulation is that the physical phenomena involved
therein typically spread over vast scales in length and time [6–8]. These scales are
typically too fast and small for experimental methods to capture the dynamic structural
formation / self-assembly in detail and are therefore limited to higher scales and
rather steady-state investigations when structures have already formed.
On the simulation side, the main challenge is that methods that provide the
appropriate molecular resolution (typically in the framework of molecular dynamics
(MD) both atomistic and coarse-grained) are limited to system sizes below 100
nm due to high computational complexity. While many larger-scale methods have
been proposed in the context of e.g. coarse-grained molecular dynamics (CG MD),
Brownian dynamics (BD), dissipative particle dynamics (DPD), ballistic cluster-
cluster aggregation (BLCA), diffusion-limited cluster(-cluster) aggregation (DLCA),
reaction-limited cluster aggregation (RLCA), and others - methods capturing the 𝜇m
and ms scale, while maintaining sufficient detail and being parameterized primarily
from lower scales and first principles, are rather scarce. In this work, we investigate a
modeling approach called by us the “Molecular Discrete Element Method” (MDEM)
for abstracting entire macro-molecules as an object with a position and orientation
and deriving their interaction in a data-driven fashion from molecular dynamics.
In this context, we have already successfully established the means to describe
and parameterize anisotropic diffusion and enforce the correct thermodynamics
(i.e. canonical ensemble) [9]. This project and contribution focuses on the derivation
of interaction potentials from MD for the hepatitis B core antigen (HBcAg).

2 Methods

2.1 Modeling framework

In the proposed modeling approach, each macro-molecule is abstracted as an ob-

ject with a position and orientation. Newton’s equation of motion is solved using
explicit time-stepping and interaction between objects as well as the environment
(thermodynamics) is modeled. An overview of the framework is shown in Fig. 1.
At the basis of the framework is the protein / molecular reference structure, which
is assumed stable / stiff for the model, but is flexible during parameterization. For
the purpose of this work, the reference structure and MD model are assumed given
(see acknowledgments). Diffusion of the objects is modeled for both translation and
rotation anisotropically using the previously published diffusion model [9] based on
Langevin dynamics and implicit water modeling. In order to describe the interaction
of objects (i.e. macro-molecules), we propose a data-driven potential field approach in
which the interaction potential for relative positions and orientations (6-D) of pairwise
contacts is saved in a field. Hierarchical fields (grids) are possible and implemented
including automatic refinement. A numerical gradient operation can then be used to
derive forces and torques acting between molecules during the simulation.
  
 
 
  
 
 

 

 
  


  
 


 




         
Multiscale modeling of macro-molecular structural formation



     

  

Fig. 1: Overview of the MDEM framework.

 
   

515
516 P.N. Depta, M. Dosta and S. Heinrich

In order to estimate the interaction potential at each relative position and orientation
from MD data, a Universal Kriging approach was implemented and a sampling
strategy derived. The main advantage of such an approach over more traditional
surrogate modeling approaches using functional descriptions or neural networks /
machine-learning descriptions is that the potential field approach allows for arbitrary
potential shapes and is only limited by memory size. Additionally, the Kriging
approach provides the ‘best linear unbiased estimate’ based on certain mathematical
requirements and presents a ’glass box’ model providing not only an estimate, but
also an error estimate. This error estimate is used for iterative resampling. As the MD
based interaction potential cannot account for repulsion due to overlapping molecules,
in a last step a repulsive potential has to be added as a function of the molecular
overlap. In the following, these components will be further described.

2.2 Molecular dynamics setup

All MD simulations were performed using the open-source software package Gromacs
[10,11] version 2020.1 using the Martini force-field version 2.2P [12,13]. Polarizable
water (PW) [13] and the particle mesh Ewald (PME) technique [14] for electrostatics
were employed in order to improve accuracy over the standard Martini water. Since
atomistic MD simulations are slower by 1-2 orders of magnitude and consequently
not possible for the iterative potential sampling approach, we rely on this previously
employed and atomistically validated coarse-grained model for modeling the Pyruvate
Dehydrogenase Complex (PDC) [15–17]. Credit for setting up the MD model and
deriving the reference structure for the HBcAg system is given to Dr. U. Jandt (see
acknowledgments) and will be summarized in the following.
The HBcAg virus capsid is composed of either 90 or 120 dimer units, which
are considered the smallest unit structure in this work. Typically, two dimers (AB
and CD) build from the HBcAg monomer are distinguished in literature with slight
conformational differences. While the majority of the conformation is very similar,
differences in regions for inter-dimer interactions are larger [18]. In the context of
this work, only one dimer kind is modeled based on a reference structure derived
from representative clustering. As done for the entire molecule, the regions for
inter-dimer interactions are also modeled flexibly when determining the interaction
potential using MD sampling and consequently the differences in conformation
during the inter-dimer interaction are implicitly captured in the MD data and derived
interaction potential. The atomistic structure for the HBcAg dimer was provided by
Dr. M. Kozlowska based on a modified version of PDB 6HTX [19] and PDB 1QGT
[18]. Representative clustering was then performed by Dr. U. Jandt based on the
martinized [12] coarse-grained structure of the AB dimer using the linkage method
as implemented in Gromacs [20] on the conformations of a 10 ns MD run with a
10 ps savings interval at 293 K and 150 mM sodium chloride ions. The determined
reference structure differs by a root-mean-square deviation (RMSD) of 0.39 nm from
the original conformation.
Multiscale modeling of macro-molecular structural formation 517

The MD setup is based on the ‘new’ parameter set for the Martini force-field
with PW and PME unless otherwise stated. A time step of 20 fs was employed
for all simulations unless otherwise stated, temperature maintained at 293 K, the
Parrinello-Rahman barostat with a compressibility of 3 × 10−4 bar−1 and coupling
constant of 12 ps used, and all systems charge neutralized with an additional 150 mM
of sodium chloride ions added. Systems contained two dimers A and B at a specific
relative position and orientation centered in a triclinic box with a minimum of 5.5
nm to any periodic boundary condition (PBC). A convergence study with a large
distance of 8 nm to the PBC showed no notable differences.
The simulation procedure consisted of two energy minimizations using the steepest
descent algorithm with tolerance of 10’000 kJ/mol/nm (first with normal Martini
water and no PME, second with PW and PME); an equilibration for 50 ps using
a time step of 5 fs with position restraints on the back-bone atoms and Berendsen
barostat with a coupling constant of 4 ps to avoid oscillations; and lastly a production
MD run for 0.6 ns. Energies between all groups (A, B, PW, ions) were calculated
every 20 steps and saved along with trajectories every 500 steps.
Postprocessing was performed using Gromacs utilities. All energy components
were extracted and reference structures fitted to determine relative positions and
orientation. Energies, positions, and orientations were then averaged between 0.5 and
0.6 ns. Overall, the following potentials between respective groups were investigated
for determining the overall interaction potential and Lennard-Jones and Coulomb
potentials added when applicable: A-B, A-A + B-B, A-PW + B-PW, PW-PW, A-
ions + B-ions, PW-ions, ions-ions, bonds, G96-angles, improper dihedral angles,
Coulomb reciprocal. As it can be seen, not only the interaction between the molecules
themselves, but also effects on the water, ions, bonds, and long range electrostatics in
the reciprocal Coulomb term were evaluated.

2.3 Iterative multivariant interpolation using Universal Kriging

In order to estimate the interaction potential at each relative position and orientation on
the grid, a Universal Kriging approach was implemented. Kriging is most frequently
applied in the field of geostatistics. Due to the scope of this report, only a brief
overview and not all details can be given. For more detail on the mathematical
background the interested reader is referred to literature [21–23].
The goal of Kriging is the determination of optimal weights for the estimation of a
spatially distributed (random) variable based on a linear combination of observations.
Optimality refers to the minimum estimation variance. In the context of this work,
the interaction potential 𝑈 𝑘𝑟𝑖𝑔, 𝑝 (x, q) of a potential component 𝑝 as a function of
relative position x and orientation q has to be estimated based on 𝑁 𝑘𝑟𝑖𝑔 observations
as
𝑁∑︁
𝑘𝑟𝑖𝑔

𝑈 𝑘𝑟𝑖𝑔, 𝑝 (x, q) = 𝑤 𝑖 𝑈 𝑝,𝑖 (x𝑖 , q𝑖 ) . (1)

𝑖=1
518 P.N. Depta, M. Dosta and S. Heinrich

Typical Kriging is performed over a subset of observations (𝑁 𝑘𝑟𝑖𝑔 ⊆ 𝑁𝑡𝑜𝑡 ) in the local
neighborhood. Universal Kriging assumes that the underlying process, in this case
potential 𝑈 (for simplicity index 𝑝 is dropped), can be decomposed into a systematic
trend 𝜇(x, q) and random component 𝑌 (x, q) as

𝑈 (x, q) = 𝜇(x, q) + 𝑌 (x, q) , (2)

where the systematic trend can be described by the linear combination of deterministic
basic functions and is in this work determined in the lower-dimensional space of the
minimum distance 𝛿 𝑚𝑖𝑛 between back-bone atoms of molecule A and B referenced
to zero by their collision distance 0.3 nm. Universal Kriging requires the remaining
random component 𝑌 to be intrinsically stationary with zero mean. While the zero
mean requirement is directly fulfilled in minimum distance space for all tested
systems, intrinsic stationarity is strictly speaking not fulfilled in the case of molecular
interaction, as the typical Gaussian distribution at small minimum distances tends
towards a (degenerate) delta distribution of zero at large minimum distances. In
order to resolve this, spatial continuity is investigated in sections for which intrinsic
stationarity is approximately fulfilled. Spatial continuity of 𝑌 is described by the
(residual) variogram 𝛾𝑌 using the root-mean-square distance 𝛿𝑟 of the back-bone
atoms as a distance measure. The optimal weights providing the unbiased and
minimum prediction variance can then be determined by solving a linear system
of equations. This is done separately for all potential components and the overall
interaction potential can then be calculated as a superposition of all components.
The overall procedure for iterative multivariant interpolation and resampling using
Universal Kriging then becomes
1. Trend fitting using weighted-least-squares for all potential components;
2. Sectional Variogram determination and fitting of trend-compensated residual 𝑌
for all potential components using weighted-least-squares;
3. Universal Kriging for qualifying potential components. Convergence studies
showed that a minimum of 𝑁 𝑘𝑟𝑖𝑔 = 500 is necessary for potential estimation,
while 𝑁 𝑘𝑟𝑖𝑔 = 100 is sufficient for estimation of the variance;
4. Summation of all potential components (trend only for those not qualifying for
Kriging) for determination of overall potential estimate 𝑈 𝑘𝑟𝑖𝑔 ;
5. Accounting of molecular collisions by increasing potential as a function of
atom collisions and proximity to MD data, as these conformations cannot be
sampled using MD;
6. Resampling based on variance reduction and extrema (potential minima/maxima,
gradient maxima) localization and specification. For variance reduction, virtual
points are iteratively placed at the maximum variance location and the variance
of the field recalculated (actual value at location not needed).
The algorithm was implemented in a custom C++ code with hybrid MPI+OpenMP
parallelization using the Eigen library for solving the linear system of equations and
employing Matplotlib in Python for fitting.
2-D Example A 2-D example of the algorithm can be found in the appendix.
Multiscale modeling of macro-molecular structural formation 519

Insertion of Experimental Findings and Expectations In order to incorporate

experimental knowledge on e.g. binding locations, virtual data points can be inserted
into the data set. Similar approaches are common in the generation of force-fields in
MD, as a precise estimation of parameters based on lower-scale models is often not
sufficient. These points do not influence the trend and variogram derived from the
MD data set, but only the neighborhood where they are located. Multiple approaches
were investigated and the final approach will be presented in the results section.

2.4 MDEM: Molecular Discrete Element Method

The modeling of structural formation based on Langevin dynamics was implemented in

the open-source DEM simulation framework MUSEN [24]. The code is implemented
in C++ and CUDA for the simulation on graphics processing units (GPU) using
the CUDA Toolkit v11.2 by NVIDIA [25]. The leap-frog algorithm was used for
time integration and contact detection performed using a Verlet list implementation
[24]. The previously published diffusion model [9] was employed and the numerical
gradient operation for deriving forces and torques of contacts from interaction
potentials implemented. A temperature of 293 K, a dynamic viscosity of 0.001 Pa s,
a time step of 10−13 s, and periodic boundary conditions were used unless otherwise
indicated. As the assembly process is diffusion limited, a two step procedure was
employed: In a first step, an accelerated simulation was performed at a lower dynamic
viscosity of 0.0001 Pa s and lower temperature of 100 K. This enabled a time step of
10−12 s and reduced simulation time for 1 ms to approximately 3 days on an Nvidia
Titan RTX. In a second step, equilibration was performed using the previously stated
settings for 0.01 ms.

2.5 Parallelism and scaling

The framework for iterative resampling and generation of the interaction potential
consists of five steps with varying degree of parallelism, which will be presented in
the following. At multiple points within the jobs consistency and error checks were
implemented to catch rare issues.
1. Pre-processing (1 node): First, the next resampling points are determined based
on either (normalized) variance minimization, potential minima/maxima, or gra-
dient maxima resampling. Variance minimization employs a hybrid parallelized
(MPI+OpenMP) and extrema identification is largely single-threaded with partial
OpenMP parallelized sections. Second, MD input files are generated in parallel.
2. MD (16 - 64 nodes): A custom implemented MPI-based scheduler runs the
individual MD simulations within an overall job. The sub-jobs are run in
descending system volume order to improve synchronization at the end (typically
<5 min run-time differences at the end). For the HBcAg system typically either
520 P.N. Depta, M. Dosta and S. Heinrich

16 or 32 cores per MD simulation were used. Each sub-job reads a 60 KB input

file (generated in last step) from the parallel file system, performs all temporary
operations in a temporary RAM disk to reduce file-server load, and saves one
compressed results file of approximately 9-10 MB.
3. Post-processing (1 node): Post-processing was performed on one node using
the Gromacs utilities in parallel for multiple jobs.
4. Post-processing - Variogram (1 node): Analysis of interaction data was per-
formed to determine potential trends and spatial correlation to derive Variogram
models (custom OpenMP parallelized code). For small data-sets, which did
not require the large SMP nodes, this step was carried out within the previous
post-processing.
5. Kriging (10 - 16 nodes): For interaction potential field generation, the custom
Universal Kriging code was used. The code employs a hybrid MPI+OpenMP
parallelization and MPI-IO for reading/saving of potential fields.
In order to efficiently use the resources granted to us, in a first step a parallel scaling
analysis of all codes employed on more than one node was performed, specifically the
MD code using Gromacs and the custom Kriging code. For the MD scaling analysis,
15 random HBcAg pairwise MD simulations at minimum atom distances of 0.5 -
2.5 nm between interaction partners using Gromacs 2020.1 on AMC EPYC 7742
CPUs at HLRS were performed. As individual MD runs are only performed on the
node-level, scaling with an increasing number of cores was investigated. Results of the
parallel efficiency can be seen in Fig. 2. Both full MD points (including single-core
pre-processing), as well production MD only are compared. As it can be seen, parallel
efficiency is good until 32 cores after which, the single-core pre-processing severely
degrades performed. Consequently, core counts of 16 or 32 were employed to ensure
sufficient efficiency.

Fig. 2 Scaling efficiency of

15 random pairwise HBcAg 1.1
MD simulations at minimum
Scaling Efficiency [-]

atom distances of 0.5 - 2.5 nm 1.0

between interaction partners
using Gromacs 2020.1 on 0.9
AMC EPYC 7742 CPUs.
Note that error bars indicate 0.8
standard deviations and all
cores were fully utilized during 0.7 MD (full point)
test. Efficiencies over 1 (super- MD (production only)
0.6
linear speedup) are attributed
0 10 20 30 40 50 60
to cache increase. Number of Cores [-]

Furthermore, the scaling of the Kriging code was investigated based on the random
VLP dataset and 0.63 nm grid. One MPI process was used per socket with 64 threads.
Primarily, Kriging sizes of 100 (used during variance minimization) were investigated
and differentiated between the full job (overall Kriging code) and only the Kriging
Multiscale modeling of macro-molecular structural formation 521

portion (without parsing of input data). Reason for this is that parsing of input
data takes place currently from a text-file format, as during method development
changes are still frequent. Once the format is final, a shift towards a binary format
will improve this efficiency. Additionally, Kriging sizes of 500 (convergent size for
potential estimate, used for extrema resampling) were estimated. As these runs are
computationally expensive (quadratic scaling leads to an increase by a factor of
25, leading to approx. 14’000 core-h per run), one test on 4 nodes was performed
to determine the cost-increase of the Kriging portion to be 23.5 and then used to
estimate scaling.

Fig. 3 Scaling efficiency of the

Kriging code for two Kriging
1.0
sizes (100 and 500) based
Scaling Efficiency [-]

on the random VLP dataset

before iterative resampling 0.9
and 0.63 nm grid. Note that
only Kriging (without reading
of input data and writing of 0.8
results), as well as overall Only Kriging (100)
Kriging are shown. Overall Kriging Code (100)
0.7 Overall Kriging Code (500), est.
0 5 10 15 20 25 30
Number of Nodes [-]

As the results in Fig. 3 show, the Kriging portion (’only’) of the code scales with a
parallel efficiency of at least 98 % at 32 nodes (4096 cores). However, the overall code
scales with a parallel efficiency of 96 % at 32 nodes for a Kriging size of 500 and
only 69 % at 32 nodes for a Kriging size of 100. As discussed earlier, this is caused
by the inefficient parsing of the input data from a text-file and is to be improved
by binary parsing once the format is final. Consequently, in order to use resources
efficiently, during variance minimization with a Kriging size of 100 only 4-10 nodes
(512-1280 cores) were used and for Kriging sizes of 500 up to 16 nodes (2048 cores).
Additionally, note that an analysis into using less than four cores per CCX showed no
improvements in overall performance.

3 Results

In the following, the results of the interaction potential for HBcAg and virus-like-
particle assembly will be presented. Beforehand, validation and convergence studies
of components were carried out. This included a 2-D validation test (see appendix Fig.
7), a MD box size convergence analysis, and a convergence analysis of the number of
Kriging points in a neighborhood based on the random HBcAg data set, which will
be discussed in the following.
522 P.N. Depta, M. Dosta and S. Heinrich

3.1 HBcAg interaction potential

In a first step, a random interaction data set was sampled from MD. For this,
MD simulations were performed at random relative positions and orientations for
molecules A and B at difference distance classes (minimum distance between atom
centers). 20’000 simulations were performed between 0.4 - 0.5 nm with a focus
of sampling binding locations, 5’000 simulations were performed in each 0.2 nm
interval between 0.5 - 2.5 nm, and 5’000 simulations were performed in each 0.5
nm interval between 2.5 - 5.0 nm. This lead to a total of 95’000 random data points,
which is doubled due to symmetry.
Upon analysis of the (random) interaction data it was found that potentials A-B
(attractive), A-PB + B-PW (repulsive), PW-PW (attractive), and A-ion + B-ion
(repulsive) posses a significant trend in 𝛿 𝑚𝑖𝑛 space, while potentials A-A + B-B,
bond/G96-angle/improper dihedral, PW-ion, and ion-ion posses no significant trend.
This shows that no trends in intra-dimer conformation could be detected (including
bonded) and of ion related potentials only the interaction with the molecules, but not
the water and between ions is significant. Consequently, the dominating components
to the potential are the molecular interaction, interplay with the water, as well as
ion mediation effects. During resampling, the decision on significance of trends
and incorporation into the overall potential was left flexibly to the algorithm. Of all
residuals 𝑌 , only the A-B potential contained a significant spatial correlation and was
further evaluated using Universal Kriging.
In the following, the interaction potential was iteratively refined and results can be
seen in Fig. 4 and Fig. 5. Resampling was performed as following: ten iterations of
variance minimization of each 5’000 samples, ten iterations of normalized variance
minimization (focus at larger distances) of each 5’000 samples, followed by potential
minima, potential maxima, and gradient extrema in consecutive order for three times
with each 20’000 samples (15’000 at main extrema locations, 5’000 at random
neighboring grid points). For variance related Kriging, 𝑁 𝑘𝑟𝑖𝑔 = 100 was used and for
all other 𝑁 𝑘𝑟𝑖𝑔 = 500. A total of 375’000 data points were sampled.
As it can be seen in Fig. 4, during variance resampling the average changes in
potential between iterations decrease continuously while the estimation variance
remains essentially unchanged. Furthermore, the maximum changes in potential
between iterations remained essentially unchanged for all iterations between 100 -
200 kJ/mol. This is attributed to the high dimensionality of the interaction space
and large residual noise. Additionally, extrema resampling (iteration 21-29) lead
to an increase in average potential change as well as estimation variance. This is
attributed to the inclusion of comparatively larger potential difference from the trend,
which also increases the overall variance of the variogram, leading to an increase
in estimation variance. This attribution is also consistent as no further increase in
maximum potential change is observed, indicating that these samples merely posses
a larger variance due to their proximity to extrema locations (e.g. binding locations).
The overall potential can be seen in Fig. 5 as a function of 𝛿 𝑚𝑖𝑛 (average and
standard-deviation over all grid locations) as well as the minimum cross-section in
X-Y. As it can be seen, the interaction potential possesses a slight potential barrier
Multiscale modeling of macro-molecular structural formation 523

Fig. 4: Convergence of the iterative resampling procedure for potential changes (left,
𝑈𝑖 − 𝑈𝑖−1 ) and variance development (right). Note that iterations 1-10 are variance
resampling (5’000 samples each), iterations 11-20 are normalized variance resampling
(5’000 samples each), and iterations 21-29 are extrema resampling (20’000 samples
each).

at 𝛿 𝑚𝑖𝑛 ≈ 1.5 nm, an intermediate potential well around 𝛿 𝑚𝑖𝑛 ≈ 0.5 nm, and three
regions of potential minima at the top left/right (aside the dimer spike) and underneath
the dimer. As it can be seen in Fig. 5 (right) in the visualization of the binding
locations, they are notably different and were found to be not sufficient for a stable
capsid.

Fig. 5: HBcAg interaction potential after resampling over the minimum distance
(left) and in X-Y cross-section (right, color scale in kJ/mol, minimum of all remaining
degrees of freedom) with overlaid visualization of the binding locations on a trimer.
Note that interaction potential and binding locations vary depending on molecular
collision model.

Based on inspection of the data and binding modes, we attribute this to the
main challenge that specific conformations are necessary for inter-molecular binding
during capsid formation. The reference structure based on the Martini force-field and
524 P.N. Depta, M. Dosta and S. Heinrich

representative clustering shows some deviations, especially at the flexible residues

involved in inter-molecular interaction. It appears that the differences might either be
too large or not sufficiently captured by the Martini force-field to sufficiently sample
the binding locations. In order to evaluate this further, extended MD simulations of 10
ns with 1’016 samples at each of the four main binding conformations were performed.
While trimers exhibited improved stability and slight changes in interaction potential
occurred including a decrease at the binding locations, no significant improvements
in capsid stability could be achieved at this point.
Consequently, in a next step we investigated the impact of introducing additional
artificial points at the binding locations into the data set. This approach is similar
to the development of force-fields in MD, where experimental data is incorporated.
Various approaches were tested and the following found to provide good results. Two
sets of virtual points for potential AB were introduced into the data set (no influence
on trends and variogram): A first set of constant potential 𝑈bind,center = −1400 kJ/mol
is located on a regular grid with -0.1, 0.0, and 0.1 nm distance in all directions
(including rotation equivalent). A second set with Gaussian shape and a range of 1
nm and asymptotic value of -1000 kJ/mol is added on a regular grid of -0.4, -0.2, 0.0,
0.2, 0.4 nm up to a distance of 1 nm (𝛿𝑟 ) from the binding location and an increase of
10 back-bone atom collisions. VLPs based on this potential were found to be stable
and assembly results will be discussed in the following.

3.2 VLP assembly

In order to study assembly, a variety of systems have been investigated thus far and a
selected system will be presented to show some of the results and challenges, which
are currently being addressed. The selected system can be seen in Fig. 6 and consists
of a 1 µm3 box with a concentration of 10 µM after 1.01 ms simulation time (see
Sec. 2.4 for simulation procedure, simulation time approx. 3.5 days on an Nvidia
Titan RTX). Note that both system size and simulation time are well beyond anything
possible with traditional, also coarse-grained, MD. As it can be seen, the formation of
capsid components as well as some close to fully formed capsids can be observed after
the simulation time of approx. 1 ms. Over 20% of dimers are involved in structures
of more than 85 dimers and can hence be considered as pre-stages of VLPs, while
the remainder forms capsid components which can be clearly identified as fractions
of spheres. Less than 0.3 % of dimers participate in structures of less than 5 dimers
indicating along with the time-dependent data that structural formation to capsid
components occurs very quickly. The largest fraction of structures after 1 ms consists
of approx. 70 dimers.
During stability analysis of a fully formed capsid, it was found that such a capsid
is very stable. During assembly, two main challenges have been identified: Firstly,
the process is largely diffusion limited. Initial capsid components form at very small
time scales and with growing size their diffusion kinetics slow down, leading to a
reduction in assembly kinetics. In order to address this, the employed simulation
Multiscale modeling of macro-molecular structural formation 525



Fig. 6: Assembly of virus like particles from random state (left) after 1.01 ms (right).

approach with reduced water viscosity was developed. Furthermore, it was found that
overgrowing of capsids occurs up to sizes of approx. 160 dimers. These capsids are
not stable and change shape and size over time. Such phenomena are experimentally
known, especially for larger concentrations, and are considered pre-stages of fully
formed stable capids. In order to address this, we are working on further increasing
the simulation times for improved equilibration, investigating lower concentrations to
avoid kinetic traps during capsid formation, increasing system sizes for improved
statistics, and exploring additional avenues (e.g. Monte-Carlo related).

3.3 Computational resources

In order to generate the presented results and to perform method development, a total
of 5.85 million core-h on the Hawk system at HLRS were used, which were graciously
provided in the context of federal project Acid 44178. GPU computations for the
agglomeration studies were performed in-house. An overview of all computational
resources can be found in Tab. 1 and will be discussed in the following. Overall, more
than 383,000 molecular dynamics simulations of the HBcAg virus protein system
in pairwise interaction were carried out providing more than 0.385 ms of overall
simulation time and a wealth of information on the interaction of HBcAg necessary
for virus capsid formation.
As it can be seen in Tab. 1, random sampling and iterative refinement accounted
for approximately 2.0 million core-h and is consequently slightly over the anticipated
1.75 million core-h. However, three additional aspects were not accurately estimated
during the planning of the project, which is attributed to the method development
nature of the project: First, in the beginning, additional testing and validation including
a MD box size and Kriging size convergence study was performed, leading to an
526 P.N. Depta, M. Dosta and S. Heinrich

additional 0.714 million core-h. Second, significantly more sampling was necessary
and the increased data size made it necessary to utilize the large SMP nodes for
statistical analysis. The authors are thankful for the availability of such resources at
HLRS, as this made the analysis of such data sets feasible in the first place. However,
the cost factor of 100 lead to an unanticipated 1.06 million core-h (0.01 million core-h
on SMP nodes). Lastly, limitations of molecular dynamics had to be investigated
in a limited set of longer simulations at the binding locations and it was needed to
incorporate experimental information into the potential fields. This lead to additional
2.37 million core-h.

Table 1: Computational resources used on Hawk at HLRS. All times are in million
core-h.

Context MD Kriging Pre/Post

Testing, MD box size convergence, number of 0.573 0.128 0.013

Kriging points
Random sampling 0.371 0.001 0.001
Iteration 1–20 0.384 0.017 0.020
Iteration 20-29 0.659 0.252 0.296
Full potential component spatial statistical 0.438 1.931 0.768
analysis, biased sampling, Kriging with ad-
ditional binding data

Sum 2.425 2.329 1.097

Total 5.85

4 Conclusion

In conclusion, we have gained valuable insight into the proposed data-driven method-
ology for deriving macro-molecular interaction potentials from MD using Universal
Kriging on the example of HBcAg for VLP assembly. The main challenges iden-
tified were found to be MD related, as force-fields are (currently) not specifically
parameterized for potential sampling and capturing inter-molecular binding remains
challenging, especially with CG-MD, which is required for sufficient sampling. We
have proposed ways to overcome these limitations by biased MD sampling and
inclusion of additional data similarly to traditional force-fields and achieved good
Multiscale modeling of macro-molecular structural formation 527

results concerning VLP assembly. Overall, the proposed method showed merit in
capturing self-assembly and post-processing MD data. Currently, we are exploring
further approaches for improvement and are testing the method on the PDC system.

Appendix

In order to perform validation and visualize the algorithm, a random scalar 2-D
example field (no units) between two spherical objects of radius 0.15 was generated
using sequential Gaussian simulation and can be seen in Fig. 7. For this, a random
truth field with similar statistical properties as typical MD data was generated (see
Fig. 7 for details), overlaid with a Gaussian trend of -400 at contact and zero at range
one, and a scaling to zero performed between a minimum distance of 0.4 and 1.2
using a Gaussian function. As it can be seen, the overall trends and binding locations
(minima) are identified and the estimation error consists of small-scale discontinuities
and noise.

Acknowledgements The authors would like to acknowledge the German Research Foundation
(DFG) for funding within the focus program SPP 1934 (HE 4526/19-2), as well as the High-
Performance Computing Center Stuttgart (HLRS, Acid 44178) for providing the computational
resources. Furthermore, the authors would like to acknowledge Dr. Uwe Jandt for setting up the
molecular dynamics system, and various MD related scripts in the context of the framework, as well
as Dr. Mariana Kozlowska for providing the atomistic reference structures for the HBcAg system
and many discussions in understanding the VLP assembly. Lastly, the authors thank the Institute
of Bioprocess and Biosystems Engineering at TUHH for collaboration in the context of further
application of the developed framework to the Pyruvate Dehydrogenase Complex, as well as critical
reading of manuscripts.

References

1. M. Castellana, M.Z. Wilson, Y. Xu, P. Joshi, I.M. Cristea, J.D. Rabinowitz, Z. Gitai, N.S.
Wingreen, Nat. Biotechnol. 32(10), 1011 (2014). DOI 10.1038/nbt.3018
2. Y.H.P. Zhang, Biotechnol. Adv. 29(6), 715 (2011). DOI 10.1016/j.biotechadv.2011.05.020
3. L.J. Sweetlove, A.R. Fernie, Nat. Commun. 9(1), 2136 (2018). DOI 10.1038/s41467-018-045
43-8
4. E.V. Grgacic, D.A. Anderson, Methods 40(1), 60 (2006). DOI 10.1016/j.ymeth.2006.07.018
5. W. Fiers, M. De Filette, K.E. Bakkouri, B. Schepens, K. Roose, M. Schotsaert, A. Birkett,
X. Saelens, Vaccine 27(45), 6280 (2009). DOI 10.1016/j.vaccine.2009.07.007
6. K.A. Henzler-Wildman, M. Lei, V. Thai, S.J. Kerns, M. Karplus, D. Kern, Nature 450(7171),
913 (2007). DOI 10.1038/nature06407
7. J.Z. Ruscio, J.E. Kohn, K.A. Ball, T. Head-Gordon, J. Am. Chem. Soc. 131(39), 14111 (2009).
DOI 10.1021/ja905396s
8. K. Steiner, H. Schwab, Comput. Struct. Biotechnol. J. 2(3), e201209010 (2012). DOI
10.5936/csbj.201209010
9. P.N. Depta, U. Jandt, M. Dosta, A.P. Zeng, S. Heinrich, J. Chem. Inf. Model. 59(1), 386 (2019).
DOI 10.1021/acs.jcim.8b00613
528 P.N. Depta, M. Dosta and S. Heinrich

Fig. 7: 2-D Universal Kriging example after 17 iterations with 10 samples per iteration
and 20 initial samples. For variogram determination the entire truth field was provided
to ensure sufficient statistics.
Multiscale modeling of macro-molecular structural formation 529

10. M.J. Abraham, T. Murtola, R. Schulz, S. Páll, J.C. Smith, B. Hess, E. Lindahl, SoftwareX 1–2,
19 (2015). DOI 10.1016/j.softx.2015.06.001
11. H. Berendsen, D. van der Spoel, R. van Drunen, Comput. Phys. Commun. 91(1), 43 (1995).
DOI 10.1016/0010-4655(95)00042-E
12. D.H. de Jong, G. Singh, W.F.D. Bennett, C. Arnarez, T.A. Wassenaar, L.V. Schäfer, X. Periole,
D.P. Tieleman, S.J. Marrink, J. Chem. Theory Comput. 9(1), 687 (2013). DOI 10.1021/ct3006
46g
13. S.O. Yesylevskyy, L.V. Schäfer, D. Sengupta, S.J. Marrink, PLoS Comput. Biol. 6(6), e1000810
(2010). DOI 10.1371/journal.pcbi.1000810
14. T. Darden, D. York, L. Pedersen, J. Chem. Phys. 98(12), 10089 (1993). DOI 10.1063/1.464397
15. S. Hezaveh, A.P. Zeng, U. Jandt, J. Phys. Chem. B 120(19), 4399 (2016). DOI 10.1021/acs.jp
cb.6b02698
16. S. Hezaveh, A.P. Zeng, U. Jandt, ACS Omega 2(3), 1134 (2017). DOI 10.1021/acsomega.6b0
0386
17. S. Hezaveh, A.P. Zeng, U. Jandt, J. Chem. Inf. Model. 58(2), 362 (2018). DOI 10.1021/acs.jc
im.7b00557
18. S. Wynne, R. Crowther, A. Leslie, Molec. Cell 3(6), 771 (1999). DOI 10.1016/S1097-2765(01
)80009-5
19. B. Böttcher, M. Nassal, J. Mol. Biol. 430(24), 4941 (2018). DOI 10.1016/j.jmb.2018.10.018
20. E. Lindahl, M.J. Abraham, B. Hess, D. Van Der Spoel, Zendo (2020). DOI 10.5281/ZENODO
.3685920
21. N.A.C. Cressie, Statistics for Spatial Data, revised edition edn. (John Wiley & Sons, Inc,
Hoboken, NJ, 2015)
22. R. Webster, M.A. Oliver, Geostatistics for Environmental Scientists (Wiley, 2007)
23. A. Lichtenstern, Kriging Methods in Spatial Statistics. Bachelorarbeit, Technische Universität
München (2013)
24. M. Dosta, V. Skorych, SoftwareX 12, 100618 (2020). DOI 10.1016/j.softx.2020.100618
25. NVIDIA Corporation, CUDA Toolkit V11.2 Programming Guide (NVIDIA Corporation, 2021)