Smart Innovation, Systems and Technologies 173

Lakhmi C. Jain
Margarita N. Favorskaya
Ilia S. Nikitin
Dmitry L. Reviznikov Editors

Advances in Theory and

Practice of Computational
Mechanics
Proceedings of the 21st International
Conference on Computational Mechanics
and Modern Applied Software Systems

123
[email protected]
Smart Innovation, Systems and Technologies

Volume 173

Series Editors
Robert J. Howlett, Bournemouth University and KES International,
Shoreham-by-sea, UK
Lakhmi C. Jain, Faculty of Engineering and Information Technology,
Centre for Artificial Intelligence, University of Technology Sydney,
Sydney, NSW, Australia

[email protected]
The Smart Innovation, Systems and Technologies book series encompasses the
topics of knowledge, intelligence, innovation and sustainability. The aim of the
series is to make available a platform for the publication of books on all aspects of
single and multi-disciplinary research on these themes in order to make the latest
results available in a readily-accessible form. Volumes on interdisciplinary research
combining two or more of these areas is particularly sought.
The series covers systems and paradigms that employ knowledge and intelligence
in a broad sense. Its scope is systems having embedded knowledge and intelligence,
which may be applied to the solution of world problems in industry, the environment
and the community. It also focusses on the knowledge-transfer methodologies and
innovation strategies employed to make this happen effectively. The combination of
intelligent systems tools and a broad range of applications introduces a need for a
synergy of disciplines from science, technology, business and the humanities. The
series will include conference proceedings, edited collections, monographs, hand-
books, reference books, and other relevant types of book in areas of science and
technology where smart systems and technologies can offer innovative solutions.
High quality content is an essential feature for all book proposals accepted for the
series. It is expected that editors of all accepted volumes will ensure that
contributions are subjected to an appropriate level of reviewing process and adhere
to KES quality principles.
** Indexing: The books of this series are submitted to ISI Proceedings,
EI-Compendex, SCOPUS, Google Scholar and Springerlink **

More information about this series at http://www.springer.com/series/8767

[email protected]
Lakhmi C. Jain Margarita N. Favorskaya
• •

Ilia S. Nikitin Dmitry L. Reviznikov

•

Editors

Advances in Theory
and Practice
of Computational Mechanics
Proceedings of the 21st International
Conference on Computational Mechanics
and Modern Applied Software Systems

123
[email protected]
Editors
Lakhmi C. Jain Margarita N. Favorskaya
University of Technology Sydney Department of Informatics and Computer
Sydney, Australia Techniques, Institute of Informatics
and Telecommunications
Liverpool Hope University
Reshetnev Siberian State University
Liverpool, UK
of Science and Technology
KES International Krasnoyarsk, Russia
UK
Dmitry L. Reviznikov
Ilia S. Nikitin Department of Numerical Mathematics
Institute of Computer Aided Design and Computer Programming
of the Russian Academy of Sciences Moscow Aviation Institute (National
(ICAD RAS) Research University)
Moscow, Russia Moscow, Russia
Federal Research Centre “Information
and Control” of the Russian Academy
of Sciences
Moscow, Russia

ISSN 2190-3018 ISSN 2190-3026 (electronic)

Smart Innovation, Systems and Technologies
ISBN 978-981-15-2599-5 ISBN 978-981-15-2600-8 (eBook)
https://doi.org/10.1007/978-981-15-2600-8
© Springer Nature Singapore Pte Ltd. 2020
This work is subject to copyright. All rights are reserved 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.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore

[email protected]
Preface

The research book is a continuation of our previous books which are focused on the
recent advances in computer vision methodologies and technical solutions using
conventional and intelligent paradigms.
This book presents the selected papers reported at the Twenty-first International
Conference on Computational Mechanics and Modern Applied Software Systems,
CMMASS’2019, which was held during 24–31 May, 2019. The contributions include
the modern numerical methods for solving problems of the continuum mechanics and
numerical simulation of physical processes occurring in gas, fluid, and solid media.
Also, methods of mathematical modeling of dynamic systems, optimization methods,
and optimal control algorithms are considered. Part I “Computational Fluid
Dynamics” involves Chaps. 2–6, Part II “Numerical Simulation of Multiphase Flows,
Combustion and Detonation” contains Chaps. 7–12, Part III “Computational Solid
Mechanics” includes Chaps. 13–19, and Part IV “Numerical Study of Dynamic
Systems” consists of Chaps. 20–25.
We wish to express our gratitude to the authors and reviewers for their contri-
butions. The assistance provided by Springer-Verlag is acknowledged.

Sydney, Australia Lakhmi C. Jain

Krasnoyarsk, Russia Margarita N. Favorskaya
Moscow, Russia Ilia S. Nikitin
Moscow, Russia Dmitry L. Reviznikov

[email protected]
Contents

1 Advances in Computational Mechanics and Numerical

Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ......... 1
Lakhmi C. Jain, Margarita N. Favorskaya, Ilia S. Nikitin
and Dmitry L. Reviznikov
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Chapters in the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Part I Computational Fluid Dynamics

2 The Splitting Scheme for Mathematical Modeling of the Mixed
Region Dynamics in a Stratified Fluid . . . . . . . . . . . . . . . . . . . ... 11
Valentin A. Gushchin and Irina A. Smirnova
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Mathematical Model. Foundation of the Problem . . . . . . . . . . . 13
2.3 Splitting Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Finite-Difference Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Modeling of Some Astrophysical Problems on Supercomputers
Using Gas-Dynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 23
Alexander V. Babakov, Alexey Y. Lugovsky
and Valery M. Chechetkin
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 23
3.2 Mathematical Models of the Evolution of the Stellar
Accretion Disk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 24
3.3 Numerical Simulation of the Evolution of the Stellar
Accretion Disk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 26

vii

[email protected]
viii Contents

3.3.1 Simulation of the Evolution of the External Area

of Stellar Accretion Disk . . . . . . . . . . . . . . . . . . .... 26
3.3.2 Numerical Results of the Formation of Vortex
Structures in Stellar Accretion Disks . . . . . . . . . . .... 28
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 31
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 34
4 Numerical Modeling of the Kolmogorov Flow
in a Viscous Media . . . . . . . . . . . . . . . . . . . . . . ............... 37
Svetlana V. Fortova and Alexey N. Doludenko
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Description of the Experiment . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Problem Statement and Numerical Method . . . . . . . . . . . . . . . . 39
4.3.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3.2 Numerical Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5 On Structures of Supersonic Flow Around Plane System
of Cylindrical Rods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 49
Sergey V. Guvernyuk and Fedor A. Maksimov
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2 Formulation of Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.3 Multiblock Computing Technology . . . . . . . . . . . . . . . . . . . . . 51
5.4 Calculation Results for Lattice of 10 Elements . . . . . . . . . . . . . 54
5.5 Calculation Results for Infinite Lattice with Periodical
Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 59
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 61
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 61
6 Limiting Functions Affecting the Accuracy of Numerical Solution
Obtained by Discontinuous Galerkin Method . . . . . . . . . . . . . . . .. 63
Marina E. Ladonkina, Olga A. Nekliudova and Vladimir F. Tishkin
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.2 Discontinuous Galerkin Method for Euler Equations . . . . . . . . . 65
6.2.1 Cockburn Limiter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.2.2 Moment Limiter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.2.3 Limiter Based on WENO Reconstruction . . . . . . . . . . . 68
6.2.4 Limiter Based on Averaging the Solution . . . . . . . . . . 68
6.2.5 Slope Limiter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.3 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

[email protected]
Contents ix

Part II Numerical Simulation of Multiphase Flows, Combustion,

and Detonation
7 Numerical Simulation of Detonation Initiation: The Quest
of Grid Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 79
Alexander I. Lopato, Artem G. Eremenko, Pavel S. Utkin
and Dmitry A. Gavrilov
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.3 Mathematical Model and Numerical Method . . . . . . . . . . . . . . 83
7.4 Technical Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7.5 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
8 On the Stability of a Detonation Wave in a Channel of Variable
Cross Section with Supersonic Input and Output Flows . . . . . . . .. 91
Vladimir Yu. Gidaspov and Dmitry S. Kononov
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
8.2 Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.3 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.4 The Results of Mathematical Modeling . . . . . . . . . . . . . . . . . . 100
8.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
9 Physical and Kinematic Processes Associated with Meteoroid
When Falling in the Earth’s Atmosphere . . . . . . . . . . . . . . . . . . . . 107
Viktor A. Andrushchenko, Vasiliy A. Goloveshkin
and Nina G. Syzranova
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
9.2 Heat Transfer to the Surface of Meteoroids . . . . . . . . . . . . . . . 108
9.3 Fragmentation of Meteoroids . . . . . . . . . . . . . . . . . . . . . . . . . . 110
9.4 Thermal Stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
9.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
10 Computational Modeling of Rarefied Plasma and Neutral Gas
Effusion into Vacuum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Vadim A. Kotelnikov and Mikhail V. Kotelnikov
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
10.2 Physical, Mathematical, and Computational Models
of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
10.3 Methodical Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
10.4 Results of Computational Experiments . . . . . . . . . . . . . . . . . . . 125

[email protected]
x Contents

10.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
11 Numerical Simulation of the Process of Phase Transitions
in Gas-Dynamic Flows in Nozzles and Jets . . . . . . . . . . . . . . . . . . . 133
Igor E. Ivanov, Vladislav S. Nazarov, Vladimir Yu. Gidaspov
and Igor A. Kryukov
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
11.2 Mathematical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
11.2.1 Method of Moments . . . . . . . . . . . . . . . . . . . . . . . . . . 135
11.2.2 Quasi-Chemical Model of Homogeneous
Condensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
11.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
11.3.1 Test: Ideal Constant Volume Adiabatic Reactor . . . . . . 146
11.3.2 Wet Steam Flow with Spontaneous Condensation
in Laval Nozzle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
11.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
12 Numerical Study of the Injection Parameters Impact
on the Efficiency of a Liquid Rocket Engine . . . . . . . . . . . . . . . . . . 153
Yulia S. Chudina, Evgenij A. Strokach, Igor N. Borovik
and Vladimir Yu. Gidaspov
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
12.2 Features of the Working Process . . . . . . . . . . . . . . . . . . . . . . . 154
12.3 The Studied Object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
12.4 Numerical Modeling of the Processes
in Combustion Chamber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
12.5 Study of the Fuel Injection . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
12.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

Part III Computational Solid Mechanics

13 Methods for Calculating the Dynamics of Layered and Block
Media with Nonlinear Contact Conditions . . . . . . . . . . . . . . . . . . . 171
Ilia S. Nikitin, Nikolay G. Burago, Vasily I. Golubev
and Alexander D. Nikitin
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
13.2 Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
13.3 Layered Model System of Equations . . . . . . . . . . . . . . . . . . . . 174
13.4 Block Model System of Equations . . . . . . . . . . . . . . . . . . . . . . 175
13.5 Numerical Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
13.6 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

[email protected]
Contents xi

13.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
14 Algorithms for Calculating Contact Problems in the Solid
Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Nikolay G. Burago, Ilia S. Nikitin and Alexander D. Nikitin
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
14.2 Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
14.3 Numerical Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
14.4 Use of Lagrange Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . 192
14.5 Using Penalty Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
14.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
14.6.1 The Impact of Two Bodies at an Angle . . . . . . . . . . . . 195
14.6.2 Explosion Welding Problem . . . . . . . . . . . . . . . . . . . . 196
14.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
15 Different Approaches for Solving Inverse Seismic Problems
in Fractured Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Vasily I. Golubev, Maxim V. Muratov and Igor B. Petrov
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
15.2 Direct Seismic Problem Solution . . . . . . . . . . . . . . . . . . . . . . . 201
15.3 Inverse Problem Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
15.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
16 Elastic Wave Scattering on a Gas-Filled Fracture Perpendicular
to Plane P-Wave Front . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Alena V. Favorskaya
16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
16.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
16.3 Elastic Wave Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
16.4 Amplitudes of Private Solutions Components . . . . . . . . . . . . . . 216
16.5 Scattering Amplitudes and Angles . . . . . . . . . . . . . . . . . . . . . . 220
16.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
17 Discrete Element Method Adopting Microstructure
Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Andrew A. Zhuravlev, Karine K. Abgaryan
and Dmitry L. Reviznikov
17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
17.2 Multiscale Discrete Element Model . . . . . . . . . . . . . . . . . . . . . 227
17.3 Computational Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

[email protected]
xii Contents

17.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
18 Durability Evaluation of Bonded Repairs for the Damaged
Metallic Structures Subjected to Mechanical
and Thermal Loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Alexey A. Fedotov and Anton V. Tsipenko
18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
18.2 Analytical Model of the Bonded Repair . . . . . . . . . . . . . . . . . . 242
18.2.1 Stage 1. Stress Computation for Bonded Joint
Without Damage . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
18.2.2 Stage 2. Computation of Stress Intensity Factors
in the Skin with Damage . . . . . . . . . . . . . . . . . . . . . . 243
18.3 Elastic Properties Degradation of the Repair
Patch Material . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
18.3.1 Testing Procedure and Materials . . . . . . . . . . . . . . . . . 246
18.3.2 Results of Fatigue Tests . . . . . . . . . . . . . . . . . . . . . . . 247
18.3.3 Elastic Modulus Change Evaluation Based
on Fatigue Test Data . . . . . . . . . . . . . . . . . . . . . . . . . 248
18.4 Bonded Repair Effectiveness Calculation Results . . . . . . . . . . . 249
18.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
19 Parametric Identification of Tersoff Potential
for Two-Component Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Karine K. Abgaryan and Alexander V. Grevtsev
19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
19.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
19.3 Comparison of Optimization Methods . . . . . . . . . . . . . . . . . . . 261
19.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
19.4.1 Silicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
19.4.2 Germanium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
19.4.3 Aluminum Nitride . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
19.4.4 Boron Nitride . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
19.5 Description of Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
19.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268

Part IV Numerical Study of Dynamic Systems

20 Multi-agent Optimization Algorithms for a Single Class
of Optimal Deterministic Control Systems . . . . . . . . . . . . . . . . . . . 271
Andrei V. Panteleev and Maria Magdalina S. Karane
20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271

[email protected]
Contents xiii

20.2 Description of Multi-agent Methods . . . . . . . . . . . . . . . . . . . . . 272

20.2.1 Optimization Problem . . . . . . . . . . . . . . . . . . . . . . . . . 272
20.2.2 Fish School Search Algorithm . . . . . . . . . . . . . . . . . . . 272
20.2.3 Krill Herd Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 275
20.2.4 Imperialist Competitive Algorithm . . . . . . . . . . . . . . . . 280
20.2.5 Hybrid Multi-agent Algorithm . . . . . . . . . . . . . . . . . . . 283
20.3 Application of Multi-agent Methods for Optimal Open-Loop
Control Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
20.3.1 Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . 284
20.3.2 Search Algorithm of Optimal Open-Loop Control . . . . 285
20.3.3 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
20.3.4 Solving the Problem of Finding Optimal Open-Loop
Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
20.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
21 Spectral Method for Analysis of Diffusions
and Jump Diffusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Gevorg Y. Baghdasaryan, Marine A. Mikilyan, Andrei V. Panteleev
and Konstantin A. Rybakov
21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
21.2 Spectral Method Formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
21.2.1 Multidimensional Matrices . . . . . . . . . . . . . . . . . . . . . 294
21.2.2 Spectral Characteristics of Functions and Linear
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
21.3 Spectral Method for Analysis of Diffusions . . . . . . . . . . . . . . . 298
21.3.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
21.3.2 Spectral Method for Solving
Fokker–Planck–Kolmogorov Equation . . . . . . . . . . . . . 299
21.3.3 Dryden Wind Turbulence Model . . . . . . . . . . . . . . . . . 303
21.4 Spectral Method for Analysis of Jump Diffusions . . . . . . . . . . . 306
21.4.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
21.4.2 Spectral Method for Solving Kolmogorov–Feller
Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
21.4.3 Dryden Wind Turbulence Model with Jumps . . . . . . . . 310
21.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory
Process: Theory and Observations . . . . . . . . . . . . . . . . . . . . . . . . . 315
Sergej S. Krylov, Vadim V. Perepelkin and Alexandra S. Filippova
22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
22.2 Mathematical Description of the Earth Pole Oscillatory
Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316

[email protected]
xiv Contents

22.2.1 Celestial-Mechanical Model of the Earth

Pole Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
22.2.2 Lunar–Solar Perturbations in the Model of Earth Pole
Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
22.2.3 Variations of the Geopotential Coefficients . . . . . . . . . . 321
22.3 Gravitational and Tidal Perturbations in the Model of the Earth
Pole Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
22.3.1 Gravitational-Tidal Lunar–Solar Moment of Forces . . . 322
22.3.2 The Oscillatory Process of the Earth Pole
at the Frequency of the Moon’s Orbit Precession . . . . . 326
22.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
23 Application of Modified Fireworks Algorithm for Multiobjective
Optimization of Satellite Control Law . . . . . . . . . . . . . . . . . . . . . . . 333
Andrei V. Panteleev and Alexander Yu. Kryuchkov
23.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
23.2 Dynamic System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
23.3 Sketch of Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
23.4 Solution of Multiobjective Problem . . . . . . . . . . . . . . . . . . . . . 337
23.4.1 Multiobjective Optimization Problem . . . . . . . . . . . . . . 337
23.4.2 Modification of Multiobjective Fireworks
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
23.4.3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
23.5 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
23.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
24 Approximate Filtering Methods in Continuous-Time Stochastic
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Konstantin N. Chugai, Ivan M. Kosachev and Konstantin A. Rybakov
24.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
24.2 Optimal Filtering Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
24.3 Equations for Conditional Probability Density . . . . . . . . . . . . . 354
24.4 Particle Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
24.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
25 Essentials of Fractal Programming . . . . . . . . . . . . . . . . . . . . . . . . . 373
Alexander S. Semenov
25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
25.2 The Elastic Object Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374

[email protected]
Contents xv

25.3 Fractal Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378

25.3.1 The Model of Container–Component
Elastic Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
25.3.2 The Model of Architectural Elastic Objects . . . . . . . . . 381
25.3.3 The Model of Fractal Petri Nets as Elastic Objects . . . . 384
25.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386

[email protected]
About the Editors

Lakhmi C. Jain, Ph.D., M.E., B.E.(Hons) Fellow (Engineers Australia), is with

the Faculty of Education, Science, Technology & Mathematics at the University of
Canberra, Australia, and Bournemouth University, UK. Professor Jain founded the
KES International for providing a professional community the opportunities for
publications, knowledge exchange, cooperation, and teaming. Involving around
5,000 researchers drawn from universities and companies world-wide, KES facili-
tates international cooperation and generates synergy in teaching and research. KES
regularly provides networking opportunities for professional community through
one of the largest conferences of its kind in the area of KES. His interests focus on
the artificial intelligence paradigms and their applications in complex systems,
security, e-education, e-healthcare, unmanned air vehicles, and intelligent agents.

Dr. Margarita N. Favorskaya is a Professor and Head of the Department of

Informatics and Computer Techniques at Reshetnev Siberian State University of
Science and Technology, Russian Federation. Professor Favorskaya is a member
of KES organization since 2010, the IPC member, and the Chair of invited sessions
of over 30 international conferences. She serves as a reviewer in international
journals (neurocomputing, knowledge engineering and soft data paradigms, pattern
recognition letters, engineering applications of artificial intelligence), an Associate
Editor of Intelligent Decision Technologies Journal, International Journal of
Knowledge-Based and Intelligent Engineering Systems, International Journal of
Reasoning-based Intelligent Systems, a Honorary Editor of the International Journal
of Knowledge Engineering and Soft Data Paradigms, the Reviewer, Guest Editor,
and Book Editor (Springer). She is the author/co-author of 200 publications and 20
educational manuals in computer science/engineering. She co-authored/co-edited
seven books for Springer recently. She supervised nine Ph.D. candidates and pre-
sently supervising four Ph.D. students. Her main research interests are digital image
and video processing, remote sensing, pattern recognition, fractal image processing,
artificial intelligence, smart systems design, and information technologies.

xvii

[email protected]
xviii About the Editors

Dr. Ilia S. Nikitin is a Professor and Director at Institute of Computer Aided

Design RAS (ICAD RAS), a Professor at Moscow Aviation Institute (MAI), a
member of the Russian National Committee on Theoretical and Applied Mechanics,
expert RAS, expert RSF, expert Minobrnauki RF. He graduated from Moscow
Institute of Physics and Technology. His scientific interests are mathematical
modeling, numerical methods in continuum mechanics, moving adaptive meshes,
dynamics of elastoplastic media, fatigue fracture, durability of operation, and
high-frequency loading. The main scientific results are the numerical methods for
solving non-stationary problems of continuum mechanics on moving and adaptive
grids, methods for calculating the stress state of elements of aircraft structures and
assessing the durability for various fatigue failure modes, refined models of layered
and block media with different sliding conditions at the contact boundaries, the
problems of propagation, transformation, and reflection of waves in such media,
and models of sintering powder materials under thermomechanical and pulsed
high-energy effects.

Dr. Dmitry L. Reviznikov is a Professor of the Department of Numerical

Mathematics and Programming at Moscow Aviation Institute (National Research
University), Russian Federation. Professor Dmitry L Reviznikov is a member of the
Russian National Committee on Heat and Mass Transfer, a member of the Scientific
Council of International Centre for Heat and Mass Transfer (ICHMT). He is a
reviewer in international journals (International Journal of Heat and Mass Transfer,
Computational Thermal Sciences, International Journal of Fluid Mechanics
Research). He supervised eight Ph.D. candidates and presently supervising three
Ph.D. students.
Scientific interests: mathematical modeling, computational physics, heat and
mass transfer, multiphase flows, nonlinear dynamics, and data analysis. Scientific
results: author of more than 100 scientific papers in Russian and international
journals, 4 monographs. Fundamental results in the fields of modeling of conju-
gated heat and mass transfer, supersonic heterogeneous flows, thermal erosion
destruction of heat-shielding coatings, anomalous diffusion, numerical methods for
fractional differential equations, nonlinear wave dynamics, and interval analysis.

[email protected]
Chapter 1
Advances in Computational Mechanics
and Numerical Simulation

Lakhmi C. Jain, Margarita N. Favorskaya, Ilia S. Nikitin

and Dmitry L. Reviznikov

Abstract The chapter contains a brief description of chapters that contribute to

the development and applications of computational methods and parallel algo-
rithms in different areas of gas dynamics, aerodynamics, hydrodynamics, turbulence,
solids dynamic, dynamic systems, optimal control. The first part presents the recent
advances in computational fluid dynamics and aerodynamics. The second part intro-
duces a numerical simulation of multiphase flows, combustion, and detonation. The
third part is devoted to computational solid mechanics, and the fourth part provides
a numerical study of dynamic systems.

L. C. Jain (B)
University of Technology Sydney, Ultimo, Australia
e-mail: [email protected]; [email protected]
Liverpool Hope University, Liverpool, UK
KES International, North Yorkshire, UK
M. N. Favorskaya
Institute of Informatics and Telecommunications, Reshetnev Siberian State University of Science
and Technology, 31, Krasnoyarsky Rabochy ave., Krasnoyarsk 660037, Russian Federation
e-mail: [email protected]
I. S. Nikitin
Institute of Computer Aided Design of the RAS, 19/18, Vtoraya Brestskaya ul., Moscow 123056,
Russian Federation
e-mail: [email protected]
D. L. Reviznikov
Moscow Aviation Institute (National Research University), 4, Volokolamskoe shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
Federal Research Centre “Information and Control” of the RAS, 44, Vavilova ul., Moscow
119333, Russian Federation

© Springer Nature Singapore Pte Ltd. 2020 1

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_1

[email protected]
2 L. C. Jain et al.

1.1 Introduction

This book presents the selected papers reported at the Twenty-first International
Conference on Computational Mechanics and Modern Applied Software Systems
(CMMASS ’2019), which was held during 24–31 May, 2019. The book includes
the modern numerical methods for solving problems of the continuum mechanics
and numerical simulation of physical processes occurring in gas, fluid, and solid
media. Also, methods of mathematical modeling of dynamic systems, optimization
methods, and optimal control algorithms are considered. Part I “Computational Fluid
Dynamics” involves Chaps. 2, 3, 4, 5, and 6, Part II “Numerical Simulation of Mul-
tiphase Flows, Combustion, and Detonation” contains Chaps. 7, 8, 9, 10, 11, and 12,
Part III “Computational Solid Mechanics” includes Chaps. 13, 14, 15, 16, 17, 18,
and 19, and Part IV “Numerical Study of Dynamic Systems” consists of Chaps. 20,
21, 22, 23, 24, and 25.

1.2 Chapters in the Book

Part I presents the recent advances in computational fluid dynamics and includes five
chapters. Chapter 2 reports the development of a splitting method for incompressible
fluid flows aiming at modeling of various types of phenomena and processes occur-
ring in such heterogeneous media as the atmosphere and the ocean [1, 2]. Research
has led to the assumption that turbulence existing only in thin layers registered in
the ocean represents the intrusion of collapsing spots mixed liquids. Four stages of
splitting scheme are considered. The obtained results are well within a range with
theoretical estimates, experimental data, and calculations by other authors. Chapter 3
studies the structure of accretion disks around compact astrophysical objects on the
basis of numerical simulation using gas-dynamic models of the environment as a
continuation of previous research findings [3, 4]. It considers two approaches for
modeling the structure of the accretion disk using various mathematical models of
the disk and various numerical methods for solving the modeling problem. Chapter 4
introduces a direct numerical simulation of the vortex flow formation regime in a layer
of weakly compressible medium based on Navier–Stokes equations [5]. Turbulence
in the two-dimensional case is studied using the plane flow of an incompressible fluid
under the action of an external force, periodic in the transverse direction. Chapter 5
reports the results of numerical simulation of two-dimensional laminar flows near a
regular system of cylinders. The supersonic flows around a geometrically unchanged
lattice of cylinders with a variation in Mach number in the direction of increasing
and decreasing in the range from 2.0 to 4.5 are visualized as the flow patterns [6,
7]. During experiments, three ranges of flow ambiguity and the corresponding hys-
teresis of Mach number characteristics were revealed. Chapter 6 considers a model
approach for discontinuous Galerkin method implementation. Numerical results
show that using discontinuous Galerkin method and applying moment limiter, slope

[email protected]
1 Advances in Computational Mechanics and Numerical Simulation 3

limiter, Weighted Essentially Non-Oscillatory (WENO) limiter, or limiter based on

averaging allows one to obtain a high order of accuracy on smooth solutions [8, 9]. In
addition, slope limiter, WENO limiter, and averaging limiter are easy to implement
and provide the generalized solutions on multidimensional unstructured grids.
Part II introduces a numerical simulation of multiphase flows, combustion, and
detonation and involves six chapters. Chapter 7 conducts the numerical investigation
of the grid resolution influence on the detonation initiation process in the multifo-
cused system with the profiled end-wall [10, 11]. The study of detonation issues in
multi-focused systems and clarification of the basic mechanisms accompanying the
detonation process using unstructured computational grids, as well as, the study of
the manifested features using the unstructured grids approach is the main goal of
this chapter. Chapter 8 is dedicated to the analysis of possible flow variants with the
stationary shock and detonation waves in a variable cross section channel, consisting
of two consecutive Laval nozzles, with hydrogen–air and hydrogen–oxygen mixtures
in a quasi-one-dimensional nonstationary formulation [12, 13]. It was obtained that
a stationary detonation wave is stable in the first expanding part of the channel and
unstable in narrowing parts. The authors clarify that for a hydrogen–air mixture in
the investigated channel, the range of flow rates, at which a stationary detonation
wave exists, can be predicted with a high degree of accuracy by the equilibrium
stationary theory. Chapter 9 considers the motion of known meteor bodies in the
Earth’s atmosphere and their fall out on the Earth’s surface [14, 15]. The mecha-
nisms of destruction of the bodies due to thermal stresses are under consideration.
The obtained results qualitatively correctly reflect the observed processes of destruc-
tion of bodies in the atmosphere. Chapter 10 presents the results of research of the
neutral gas and plasma effusion into vacuum space found by computational mod-
eling using the kinetic theory [16, 17]. The proposed physical, mathematical, and
computational models are based on the computational solutions referred to Vlasov
kinetic equation. Distribution functions of charged and neutral particles at various
points of the region of interest, as well as, the momentums of those functions (the
concentrations and velocities fields) were found through computational experiments.
Chapter 11 aims to develop the condensation and evaporation in flows of two-phase
gas–droplet mixture in the nozzles, jets, and external area in front of the nozzle. This
study develops two ways for condensation modeling. The first one is a continual
approach based on the method of moments. The second one is a kinetic approach
based on a quasi-kinetic model. Using a quasi-chemical model for water vapor in
nitrogen, the saturation curves in the pressure–temperature phase plane depending
on the mass fraction of water vapor were obtained. Also, a qualitative agreement
was obtained between the numerical and experimental pressure distributions on the
plane of symmetry of the nozzle. Chapter 12 investigates the effect of various param-
eters of fuel injection in an oxygen–kerosene rocket thruster on the efficiency of the
workflow, particularly, the droplet injection velocity components by a centrifugal
nozzle in a cylindrical coordinate system and droplet size distribution parameters.
The working process was modeled without and with film cooling. The main features
of the numerical experiment are highlighted, as well as, the recommendations based
on the obtained results are formulated.

[email protected]
4 L. C. Jain et al.

Part III is devoted to computational solid mechanics and contains seven chapters.
Chapter 13 contains the continual models of solid media with a discrete set of slip
planes (layered, block media) and with nonlinear type slip conditions at the con-
tact boundaries of structural elements [18, 19]. The developed model can be used
for the numerical simulation of the seismic survey process in complex geological
fractured media. Numerical simulations of the dynamic scattering process for sub-
surface layered and block objects in elastic 2D and 3D media were carried out using
high-performance computing systems. Chapter 14 explores the explicit and implicit
non-matrix finite element algorithms for calculating contact interactions between
elastic–plastic bodies [20, 21]. The algorithms of Lagrange multiplier methods for
explicit schemes and algorithms of penalty functions for implicit schemes are con-
sidered in detail. The effectiveness of the algorithms is illustrated by two nontrivial
examples: the impact of two plates at an angle and axisymmetric welding of two dis-
similar tube samples under the action of a detonation wave. Chapter 15 examines the
inverse seismic problem for oil and gas exploration using different approaches [22,
23]. First, the functional of minimization based on synthetic responses from layered
and fractured media is constructed, and all parameters of the model are estimated
by the appropriate minimization procedure. Second, machine learning techniques
are used to reconstruct the fractured structure of the geological medium. Third, the
classic migration problem using adjoint operators and the grid-characteristic method
on structured meshes is solved. Chapter 16 discusses the features of the scattering
of plane P-waves on gas-filled fractures located along the motion of the incident
wave front [24, 25]. This problem has practical mining in the areas of nondestruc-
tive testing and seismic exploration, primarily in the area of railway nondestructive
testing. The analytical formulas for the scattering of a plane P-wave on a gas-filled
fracture located along the motion of the front of this wave were derived. Scattered
S-waves have been identified and studied. Chapter 17 focuses on the elastic behav-
ior of modeled structures using a novel multiscale method for materials modeling,
which requires the information only from the atomic level (atomic structure and
potential of atomic interaction). Modeled structures are virtually divided into tetra-
hedral elements. Each element contains a small but representative sample of atomic
structure [26, 27]. The whole system evolution is governed by equations of motion
for every element vertex. Computational experiments show a good correspondence
of the results obtained from the proposed model with classical molecular dynam-
ics results, which can be considered as the exact solution. Chapter 18 develops the
analytical calculation model to determine the repair joint parameters of the aircraft
structural elements. The model utilizes the inclusion methodology for heterogeneous
materials in the joint and Hart-Smith model of adhesive layer. This study covers the
variation of the longitudinal and transversal elastic moduli of the carbon fiber plastic
specimens at different temperatures and variations of Poisson ratio under cyclic load
at the same values of temperature. The proposed analytical technique can be tuned for
specific structural and repair materials and solutions. Chapter 19 covers a parametric
identification of Tersoff potential for one-component and two-component materi-
als based on the molecular-dynamic modeling [28]. Each potential has a certain set
of parameters, the values of which are unique for each material. The problem of

[email protected]
1 Advances in Computational Mechanics and Numerical Simulation 5

parametric identification is multi-extremal. Therefore, a comparison of Monte Carlo

and simulated annealing methods for global minimization and Hooke–Jeeves and
Radial Granular Search methods for local minimization was implemented. Also, a
software tool for parametric identification of certain materials that used the parallel
calculations is discussed.
Part IV provides a numerical study of dynamic systems and includes six chapters.
Chapter 20 develops the algorithms and software of three metaheuristic multi-agent
methods: fish school search, krill herd, and imperialist competitive algorithm. Using
multi-agent approach, one can optimize not only multi-extremal functions of many
variables, but also find a solution for optimal open-loop control problems in aviation
and space technology [29, 30]. On the basis of krill herd and imperialist competitive
algorithm, a hybrid extremum search algorithm is formulated. Also, an algorithm
for finding open-loop control for a single class of dynamic systems based on the
use of the described multi-agent algorithms is suggested. Software that allows to
find the optimal open-loop control, criterion value, and coordinates of switching
points of the control law on the basis of the suggested algorithms was formed. It is
shown that the numerical solution is closed to the optimal one. Chapter 21 discusses
the use of the spectral form of mathematical description for the statistical analysis
of stochastic dynamical systems [31, 32]: diffusions and jump diffusions, i.e., for
solving Fokker–Planck–Kolmogorov equation and Kolmogorov–Feller equation for
the probability density of the state vector for these dynamical systems. A detailed
description of the proposed methods was supplemented by step-by-step algorithms
for solving analysis problems and numerical experiments. Dryden wind turbulence
model and its jump diffusion modification are used for testing the spectral method.
Chapter 22 analyzes the dynamic effects of the Earth’s pole motion in the celestial
mechanical problem statement as the “deformable Earth–Moon problem in the grav-
itational field of the Sun” [33, 34]. The aim of this research is to study the effect of
lunar–solar long-period disturbances on the Earth’s pole motion, in other words, to
find the influence of the perturbations from the Earth–Moon system spatial motion
on the Earth’s pole oscillatory process. A mathematical description of the Earth’s
pole motion model and gravitational–tidal lunar–solar disturbances, as well as, the
gravitational–tidal mechanism of the Earth pole motion with a frequency close to
the frequency of the lunar orbit precession are discussed. Chapter 23 discusses the
application of modified metaheuristic global optimization algorithm “fireworks” in
order to solve the problems of multiobjective optimization [35, 36]. A solution is a
set of Pareto optimal possible solutions. Searching of control includes two stages.
At the first stage, the optimization problem is solved for each of the objectives with
penalties. The values of the penalties are selected to satisfy the terminal constraints.
At the second stage, the penalties found are used to solve a multiobjective optimiza-
tion problem. Modification of the one-objective optimization “fireworks” algorithm
and its application to find programmed control which stabilizes a satellite is pro-
posed. Chapter 24 contributes in solving the optimal filtering problem for nonlinear
continuous-time stochastic observation systems [37]. Particle filters are proposed
on the basis of Duncan–Mortensen–Zakai equation, as well as, on the basis of the

[email protected]
6 L. C. Jain et al.

robust Duncan–Mortensen–Zakai equation. To find the mode of the conditional dis-

tribution approximately, Edgeworth series is used for the conditional probability
density expansion that allows to reduce significantly a computation time in contrast
to find the mode by estimating the conditional probability density, for example, by the
histogram or kernel estimations. Chapter 25 describes fractal programming as a pro-
gramming paradigm based on the concept of “elastic objects” [38, 39]. The concept
supposes that elastic objects can be transformed (unfolded and folded) dynamically
at runtime using strategy planning model and production rules. These rules are keep-
ing the object structure self-similar that defines a fractal property. This provides a
new type of abstraction, encapsulation, inheritance, modularity, and concurrency of
objects in fractal-oriented programming.

1.3 Conclusions

The book presents the research work of major experts in the field of numerical meth-
ods and mathematical modeling the dynamics of continuous media: gases, liquids,
deformable solids, as well as, dynamic systems and optimal control. Using com-
putational methods of continuum mechanics, such diverse gas and hydrodynamics
processes have been studied as inhomogeneous flows in the ocean and atmosphere,
the formation of accretion disks near astrophysical objects, laminar, and turbulent
flows near bodies systems, shock, and detonation waves in channels of variable cross
section. In the dynamics of solids, numerical methods have been developed to study
the processes of wave propagation and scattering in structured media (seismic sur-
vey), the motion and destruction of meteoroids in the atmosphere, and the contact
interaction of inelastic bodies. To solve most of the problems, researchers reported the
use of parallel algorithms for multiprocessor high-performance computing systems
(supercomputers). The book will be useful to scientists, researchers, undergraduate,
and postgraduate students specializing in the field of computational methods, paral-
lel algorithms, gas dynamics, aerodynamics, hydrodynamics, turbulence, multiphase
flows, combustion and detonation, solids dynamic, dynamic systems, and optimal
control.

References

1. Gushchin, V.A.: Family of quasi-monotonic finite-difference schemes of the second order of

approximation. Math. Model. Comput. Simul. 8(5), 487–496 (2016)
2. Gushchin, V.A., Matyushin, P.V.: Simulation and study of stratified flows around finite bodies.
Comput. Math. Math. Phys. 56(6), 1034–1047 (2016)
3. Babakov, A.V., Popov, M.V., Chechetkin, V.M.: Mathematical simulation of a massive star
evolution based on a gasdynamical model. Math. Model. Comput. Simul. 10(3), 357–362
(2018)

[email protected]
1 Advances in Computational Mechanics and Numerical Simulation 7

4. Babakov, A.V., Lugovsky, A.Y., Chechetkin, V.M.: Mathematical modeling of the evolution
of compact astrophysical gas objects. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N.,
Simakov, S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development
of Grid-characteristic Method, SIST, vol. 133, pp. 210–227. Springer (2019)
5. Fortova, S.V., Oparina, E.I., Belotserkovskaya, M.S.: Numerical simulation of the Kolmogorov
flow under the influence of the periodic field of the external force. J. Phys. Conf. Ser. 1128,
012089.1–012089.5 (2018)
6. Maksimov, F.A., Churakov, D.A., Shevelev, Y.D.: Development of mathematical models and
numerical methods for aerodynamic design on multiprocessor computers. Comput. Math. Math.
Phys. 51(2), 284–307 (2011)
7. Guvernyuk, S.V., Maksimov, F.A.: Supersonic flow past a flat lattice of cylindrical rods.
Comput. Math. Math. Phys. 56(6), 1012–1019 (2016)
8. Ladonkina, M., Nekliudova, O., Ostapenko, V., Tishkin, V.: On the accuracy of the discontin-
uous Galerkin method in the calculation of shock waves. Comput. Math. Math. Phys. 58(8),
1344–1353 (2018)
9. Ladonkina, M., Nekliudova, O., Tishkin, V.: Construction of the limiter based on averaging of
solutions for discontinuous Galerkin method. Math. Model. 30(5), 99–116 (2018)
10. Lopato, A.I., Utkin, P.S.: Numerical study of detonation wave propagation in the variable
cross-section channel using unstructured computational grids. J. Combustion 3635797, 1–8
(2018)
11. Lopato, A.I., Utkin, P.S.: The usage of grid-characteristic method for the simulation of flows
with detonation waves. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov, S.S.,
Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development of Grid-
characteristic Method, SIST, vol. 133, pp. 281–290. Springer (2019)
12. Gidaspov, V.Y., Severina, N.S.: Numerical simulation of the detonation of a propane-air
mixture, taking irreversible chemical reactions into account. High Temp. 55(5), 777–781 (2017)
13. Gidaspov, V.Y.: Numerical simulation of chemically non-equilibrium flow in the nozzle of the
liquid-propellant rocket engine. Aerospace MAI J. 20(2), 90–97 (2013)
14. Syzranova, N.G., Andrushchenko, V.A.: Simulation of the motion and destruction of bolides
in the Earth’s atmosphere. High Temp. 54(3), 308–315 (2016)
15. Andrushchenko, V.A., Maksimov, F.A., Syzranova, N.G.: Simulation of flight and destruction
of the Benešov bolid. Comput. Res. Model. 10(5), 605–618 (2018)
16. Kotelnikov, M.V.: The distribution functions of charged particles in the vicinity of a cylindrical
body in a flow of collisionless plasma in magnetic field. High Temp. 46(6), 757–762 (2008)
17. Kotelnikov, V.A., Kotelnikov, M.V.: Current–voltage characteristics of a flat probe in a rarified
plasma flow. High Temp. 54(1), 20–25 (2016)
18. Nikitin, I.S., Burago, N.G., Nikitin, A.D.: Continuum model of the layered medium with
slippage and nonlinear conditions at the interlayer boundaries. Solid State Phenom. 258, 137–
140 (2017)
19. Golubev, V.I.: The usage of grid-characteristic method in seismic migration problems. In:
Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov, S.S., Jain, L.C. (eds) Proceedings
of the Conference on 50 Years of the Development of Grid-characteristic Method, SIST, vol.
133, pp. 143–155. Springer (2019)
20. Burago, N.G., Nikitin, I.S.: Matrix-free conjugate gradient implementation of implicit schemes.
Comput. Math. Math. Phys. 58(8), 1247–1258 (2018)
21. Burago, N.G., Nikitin, I.S., Nikitin, A.D., Stratula, B.A.: Algorithms for calculation damage
processes. Frattura ed Integrità Strutturale 13(49), 212–224 (2019)
22. Golubev, V.I., Voinov, O.Y., Petrov, I.B.: Seismic imaging of fractured elastic media on the
basis of the grid-characteristic method. Comput. Math. Math. Phys. 58(8), 1309–1315 (2018)
23. Muratov, M.V., Petrov, I.B.: Application of fractures mathematical models in exploration seis-
mology problems modeling. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov,
S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development of
Grid-characteristic Method, SIST, vol. 133, pp. 120–131. Springer (2019)

[email protected]
8 L. C. Jain et al.

24. Favorskaya, A.V., Zhdanov, M.S., Khokhlov, N.I., Petrov, I.B.: Modeling the wave phe-
nomena in acoustic and elastic media with sharp variations of physical properties using the
grid-characteristic method. Geophys. Prospect. 66(8), 1485–1502 (2018)
25. Favorskaya, A.V.: The use of multiple waves to obtain information on an underlying geological
structure. Procedia Comput. Sci. 126, 1110–1119 (2018)
26. Abgaryan, K.K., Zhuravlev, A.A., Zagordan, N.L., Reviznikov, D.L.: Discrete-element simu-
lation of a spherical projectile penetration into a massive obstacle. Comput. Res. Model. 7(1),
71–79 (2015)
27. Abgaryan, K.K., Eliseev, S.V., Zhuravlev, A.A., Reviznikov, D.L.: High-speed penetration.
Discrete-element simulation and experiments. Comput. Res. Model. 9(6), 937–944 (2017)
28. Abgaryan, K.K., Posypkin, M.A.: Optimization methods as applied to parametric identification
of interatomic potentials. Comput. Math. Math. Phys. 54(12), 1929–1935 (2014)
29. Panteleev, A.V., Metlitskaya, D.V.: Using the method of artificial immune systems to seek
the suboptimal program control of deterministic systems. Autom. Remote Control 75(11),
1922–1935 (2014)
30. Karane, M.M.C.: Comparative analysis of multi-agent methods for constrained global optimiza-
tion. In: IV International Conference on Information Technologies in Engineering Education,
pp. 128–133 (2018)
31. Rybakov, K.A., Sotskova, I.L.: Spectral method for analysis of switching diffusions. IEEE
Trans. Autom. Control 52(7), 1320–1325 (2007)
32. Baghdasaryan, G., Mikilyan, M.: Effects of magnetoelastic interactions in conductive plates
and shells. Springer, Cham (2016)
33. Markov, YuG, Perepelkin, V.V., Filippova, A.S.: Analysis of the perturbed Chandler wobble of
the Earth pole. Dokl. Phys. 62(6), 318–322 (2017)
34. Perepelkin, V.V., Rykhlova, L.V., Filippova, A.S.: Long-period variations in oscillations of the
Earth’s pole due to lunar perturbations. Astron. Rep. 63(3), 238–247 (2019)
35. Panteleev, A., Kryuchkov, A.: Metaheuristic optimization methods for parameters estimation
of dynamic systems. Civ. Aviat. High Technol. 20(2), 37–45 (2017)
36. Panteleev, A., Pis’mennaya, V.: Application of a memetic algorithm for the optimal control
of bunches of trajectories of nonlinear deterministic systems with incomplete feedback. J.
Comput. Syst. Sci. Int. 57(1), 25–36 (2018)
37. Rybakov, K.A.: Solving the nonlinear problems of estimation for navigation data processing
using continuous particle filter. Gyroscopy Navig. 10(1), 27–34 (2019)
38. Semenov, A.S.: Pattern-type reachability analysis of distributed systems based on fractal Petri
nets. In: 5th International Conference on Control, Decision and Information Technologies.
Thessaloniki, Greece, pp. 346–351 (2018)
39. Semenov, A.S.: Pattern recognition technique for synthesis fractal Petri nets. In: 6th Inter-
national Conference on Control, Decision and Information Technologies. Paris, France,
pp. 1798–1803 (2019)

[email protected]
 Part I
Computational Fluid Dynamics

[email protected]
Chapter 2
The Splitting Scheme for Mathematical
Modeling of the Mixed Region Dynamics
in a Stratified Fluid

Valentin A. Gushchin and Irina A. Smirnova

Abstract Study of wave motions’ fluid is one of the most important and complex
problems of modern hydrodynamics. A mathematical model for dynamics of incom-
pressible uniform viscous liquid spots in the stratified medium is considered. This
model is described by Navier–Stokes equations in Boussinesq approximation. Strat-
ification component of the medium is saltiness. Bearing in mind that in such flows
there are areas with large gradients of hydrodynamic parameters, required meth-
ods should possess such properties as a high order of accuracy, minimum scheme
dissipation and dispersion, as well as monotony. To solve the task, the authors are
developing a method of splitting by physical factors called as Splitting on physical
factors Method for Incompressible fluid Flows (SMIF) possessing by the above-
mentioned properties. Four stages of splitting scheme are considered. This chapter
provides a brief description of SMIF method. The test calculations and compari-
son with some theoretical and experimental data respect to the calculations of other
authors are demonstrated.

2.1 Introduction

Study of various types of phenomena and processes occurring in such heterogeneous

medium as the atmosphere and the ocean is both an academic and practical interest.
The heterogeneity of these media is linked to the effects of buoyancy that is the
presence of gravity. It is known that density of sea and ocean medium depends on
temperature, pressure, and salinity. A number of mathematical models describing
the dynamics of stratified fluids have been suggested [1–4]. Fine structure of hydro-
physical fields observed in the ocean is an alternation of deep sites with low and

V. A. Gushchin (B) · I. A. Smirnova

Institute for Computer Aided Design of the RAS, 19/18, Vtoraya Brestskaya ul., Moscow 123056,
Russian Federation
e-mail: [email protected]
I. A. Smirnova
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 11

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_2

[email protected]
12 V. A. Gushchin and I. A. Smirnova

high vertical gradients of varying characteristics. Almost uniform in its properties,

layers, vertical dimensions that vary from tens of centimeters to tens of meters, are
separated by streaks. In these streaks, vertical gradients of physical properties can
significantly exceed their averages. Time scales of such irregularities ranges from
several hours to several days and more. The first systematic analysis of the physical
mechanisms, forming fine structure, was taken in [3].
The causes of the initial formation of the mixed areas in stratified fluids are the
following: shear instability benthic stratified large-scale currents and tidal waves,
overturn of surface and internal waves, and convection in layers with unstable den-
sity stratification. Next, a shear velocity, under the action of which the initially
mixed areas of turbulent energy balance can be positive and begins to increase the
turbulence, starts playing a role [5]. In the ocean, turbulence exists only in thin lay-
ers—“turbulent pancakes.” Field studies in the ocean have shown that turbulence has
a pronounced “island,” intermittent nature [4]. Research has led to the assumption
that “pancake” structures registered in the ocean represent the intrusion of collapsing
spots mixed liquids, mainly in viscous phase of its evolution. Collapse process of
mixed areas, where turbulence lives or degenerates, is one of the fundamental pro-
cesses responsible for the formation of the fine structure of the ocean waters [3–6].
It is known [5] that the arising and development of turbulence in fluid density strat-
ified steadily is inseparable from the dynamics of internal waves and is as follows.
Under the influence of outside forces, the large-scale internal waves originate in the
stratified fluids. As a result, their nonlinear interactions and subsequent overturning
or buckling occur field mixed liquid-stains (sometimes referred to as mixing zones).
These spots of mixed turbulent fluid are evolving, gradually flattening (collapse of
turbulent spots) that, in turn, leads to the formation of new spots, etc.
It is naturally consider as the three main stages of its development in the process
of evolution [7]:
• Stage I. Initial stage: a force is acting on a particle of the fluid inside the spot,
vastly superior force of resistance; an intensive generation of internal waves by
spot takes place.
• Stage II. Intermediate fixed stage: a driving force is balanced by mainly resis-
tance forms and wave resistance caused by radiation of internal waves; the size
of the spots is increasing depending on the time of happening almost linear, i.e.,
acceleration curve is negligibly small.
• Stage III. Final viscous stage: a driving force is balanced by mostly viscous
resistance; the horizontal size of the spots varies slightly.
Stages I and II are short and were estimated in [7–9] completed through a period
of time, approximately equal to 4T b , where T b = 2π/N is the period and N is Brent–
Viasel frequency. The duration of Stage I does not exceed T b /2. Basically, the same
observed spots of turbulence are at the final Stage III of evolution. Further, as a result
of the diffusion the stain is mixed with the ambient fluid and disappears.
In this chapter, we adapt a mathematical model [10] previously used in tasks of
stratified fluid flows around a sphere and circular cylinder [11, 12] to the task on
collapse spots that we solved earlier excluding diffusion of stratification component

[email protected]
2 The Splitting Scheme for Mathematical Modeling of the Mixed … 13

[13, 14]. Bearing in mind that in such flows there are areas with large gradients of
hydrodynamic parameters, the required methods should possess such properties as
a high order of accuracy, minimum scheme dissipation and dispersion, as well as
monotony. This is essentially impotent for flows with internal and surface waves,
where arising of “numerical” waves is unacceptable. To solve the task, we develop
SMIF method possessing by the abovementioned properties [15]. Four stages of
splitting scheme are considered.
The chapter is structured as follows. Section 2.2 provides a foundation of the
problem, mathematical model, and initial and boundary conditions. Four stages of
splitting scheme are considered in Sect. 2.3. Short review of finite-difference scheme
is described in Sect. 2.4. The results and comparison with theoretical estimates, exper-
imental data, and calculations of other authors are presented in Sect. 2.5. Section 2.6
gives the conclusion.

2.2 Mathematical Model. Foundation of the Problem

Consider 2D unsteady flow that occurs, when collapse area of homogeneous fluid A
is surrounded by sustainably and continuously stratified by density (for clarity, use
a linear low) incompressible fluid (Fig. 2.1).
The course develops in a uniform gravity field with the acceleration of free fall g.
Undisturbed linear density distribution [10] is defined as:
 y 
ρ(x, y) = ρ0 1 − + s(x, y) ,

  −1
 
where  is the stratification scale,  =  ρ10 ∂ρ∂y 
, N is the buoyancy frequency,
 
N = g , T b is the buoyancy period, T b = 2π /N, C = /R0 is the scale ratio,
R0 is the radius of spot, s is the perturbation of salinity (stratification component),
includes salt ratio of compression.

Fig. 2.1 Initial and

boundary conditions

[email protected]
14 V. A. Gushchin and I. A. Smirnova

Navier–Stokes equations in Boussinesq approximation describing the flow of this

type can be written as

∂v 1 1 g
+ (v · ∇)v = −∇ p + v + s ,
∂t Re Fr g
∇ · v = 0,
∂s 1 v
+ (v · ∇)s = s + ,
∂t Sc · R e C
where v is the velocity vector with components u, v along the x- and y- axes of
Cartesian coordinate system selected as indicated in Fig. 2.1, respectively, ρ is the
density, p is the pressure minus hydrostatic one, s is the perturbation of salinity, Re is
the Reynolds number, Re = ρ 0 R20 N/μ, Fr is the Froude number, Fr = R0 N 2 /g, Sc is
the Schmidt number = μ/ρ 0 k s , k s is the diffusion coefficient of salts, μ is dynamic
viscosity coefficient, g = (0, −g), g is acceleration of free fall, ρ 0 is the density at
the level y = 0, C = /R0 is the scale ratio.
We assume that the initial time t = 0 the system on the plane R2 is at rest, i.e.,

u = v = 0, (x, y) ∈ R2 ,

the density of fluid at the spot A is

ρ = 1, (x, y) ∈ A,

and outside of spot, i.e., in the area of R2 \A, is

y
ρ =1− + s, (x, y) ∈ R2 \A,
C
the peturbation of salinty is

y/C (x, y) ∈ A,
s=
0 (x, y) ∈ R 2 \A.

As an initial approximation necessary to solve the equation for pressure distribu-

tion, the following system is selected:

−y/Fr, (x, y) ∈ A,
p=
− y − y 2 /2C /Fr, (x, y) ∈ R 2 \A.

As the pressure in the case of an incompressible fluid shall be determined with an

accuracy of up to an arbitrary constant (without limiting a generality), we can select
it to zero, at level y = 0.

[email protected]
2 The Splitting Scheme for Mathematical Modeling of the Mixed … 15

Effect of symmetry tasks concerning the plane x = 0, naturally seeks a solution

in only one half-plane, for example, if x ≥ 0. Solution will search in the rectangular
area {x, y: 0 ≤ x ≤ X, −Y ≤ y ≤ Y}. In the left boundary (line 1 in Fig. 2.1), this
area is the conditions of symmetry:

∂v ∂p ∂ρ ∂s
u = 0, = = = = 0.
∂x ∂x ∂x ∂x
The top (line 2), bottom (line 4), and right (line 3) borders should be chosen far
enough away from the source of disturbance (from spots) so that setting any boundary
conditions at these borders, which are necessary for the solution of the problem, is
not providing a significant influence on the flow.
To solve the task, we use one of the latest versions of SMIF method. Finite-
difference scheme of this method possesses by properties such as a second-order
approximation for the spatial variable, minimum scheme viscosity and dispersion,
functioning in a wide range of Reynolds and Froude numbers, and more importantly,
when solving such problems, the monotony [15].

2.3 Splitting Scheme

Let some time be t n = n · τ , where τ is the time step, n is the number of steps
with known velocity v, pressure p, and perturbation of salinity s. Then the unknown
scheme functions at time t n+1 = (n + 1) · τ can be represented as follows:

ṽ − v 1 1 ng
= −(vn · ∇)vn + vn + s ,
τ Re Fr g
τ p = ∇ · ṽ,
vn+1 − ṽ
= −∇ p,
τ
s n+1 − s n 1 vn+1
= −(vn+1 · ∇)s n + s n + .
τ Sc · R e C
At Stage I, it is expected that the transfer of momentum (the momentum of a
unit of mass) is performed only by the convention, diffusion, and buoyancy forces.
At Stage II, using the found interim velocity ṽ, a pressure field is calculated using
Poisson equation. Here, we take into account that ∇ · vn+1 = 0. At Stage III, it is
anticipated that the transfer is carried out only at the expense of the pressure gradient
(convection and diffusion are not available). At Stage IV, using the found velocity
field vn+1 , a perturbation of salinity is calculated.

[email protected]
16 V. A. Gushchin and I. A. Smirnova

Fig. 2.2 Grid stencil

2.4 Finite-Difference Scheme

The study area is covered by a uniform x and y grid cells:


xi+1/2 = i · δx, δx > 0, i = 0, 1, . . . , L; L · δx = X,
=
y j+1/2 = j · δy, δy > 0, j = 0, 1, . . . , M; M · δy = 2Y,

where δx, δy are the dimensions grid steps, L and M are the numbers of grid cells
in the directions of x and y, respectively (the point with coordinates (i, j) is at the
center of the cell). Here, as in the original method of splitting, we use the “Chess”
grid, i.e., the grid coordinates functions are separated in space, as shown in Fig. 2.2.
This enables to interpret visually each cell, such as volume, which is characterized
by calculations in its central pressure pi,j , density ρ i,j (possibly, temperature, energy,
etc.), as well as, divergence of Di,j (D in according to the sign determines whether
source or drain in this volume). Knowledge of the normal components of the velocity
vector on the sides of the cell is able to directly calculate the flow of momentum
through this side. The finite-difference scheme for 1D case is presented in [15].
For the solution of Poisson equation for pressure the successive over the relaxation
method is used.

2.5 Results

Calculations were carried out in the field with dimensions X = 10, Y = 5, R0 = 1

with the following coefficients and parameters: μ/ρ 0 = 0.01 cm2 /s−1 , k s = 1.41 10−5
cm2 /s−1 , N = 1 s−1 , T b = 2π s,  = 10 cm, C = 10, Re = 100, Fr = 0.1, Sc = 709.2
that is close to the laboratory experimental conditions. As boundary conditions at the
top, the resting states of bottom and right borders of the computational domain are
chosen, i.e., u = v = s = 0. The computational domain is covered with a uniform grid
with the steps in both directions δx = δy = 0.1. With a view to program verification,
the calculations in the absence of stain and on different grids were performed. The
results confirmed the execution of conservation laws with the required accuracy.

[email protected]
2 The Splitting Scheme for Mathematical Modeling of the Mixed … 17

Fig. 2.3 Time dependence of horizontal spot size: a received by our method, b comparison with
other authors

Time dependence of horizontal spot size l at the level of y = 0 is shown in Fig. 2.3a.
The comparison with analytical estimations [7]—curve 2, experimental data [5]—
curve 1, and calculation of other authors [4, 6]—curves 3, 4, is shown in Fig. 2.3b.
Here, curve 5 is our numerical results. Time dependence of vertical spot size h at the
level of x = 0 is shown in Fig. 2.4. Here, thick lines are our results obtained with
physical model without diffusion [13] and thin lines are our present results, where
diffusion of stratification component (perturbation of salinity) is taken into account.
It should be noted that, as in [13], changes in both horizontal and vertical sizes of
spots have non-monotonous character.
The isoclines of stream function for t = 2, 4, 6, 8 are shown in Fig. 2.5. Figure 2.5b
shows the emergence of a second vortex pair, Fig. 2.5c depicts a clear visibility of two
developed vortex pairs, and by the time t = 8 a third vortex pair arises (Fig. 2.5d).

Fig. 2.4 Vertical spot size.

Thick lines are taken from
[13], thin lines are our
present results

[email protected]
18 V. A. Gushchin and I. A. Smirnova

Fig. 2.5 The isoclines of stream function: a t = 2, b t = 4, c t = 6, d t = 8

It should be noted that the vortices located in the upper half-plane have a greater
intensity than those vortices from the lower half-plane.
The isoclines of perturbation of salinity for t = 2, 4, 6, 8 are shown in Fig. 2.6,
while the isoclines of perturbation of salinity for t = 20 and t = 30 are shown in
Fig. 2.7.
The isoclines of stream function for t = 20 and 30 are shown in Fig. 2.8.
From Figs. 2.6, 2.7, and 2.8, it is possible to estimate a length of internal waves.
Simultaneously, Figs. 2.7, 2.8 show us that for correct calculations, outer boundaries
X and Y should be taken larger for time moments more than t = 20.

2.6 Conclusions

The mathematical model of the dynamics of the spots in the stratified fluid is sug-
gested. The model is based on Navier–Stokes equations in Boussinesq approximation

[email protected]
2 The Splitting Scheme for Mathematical Modeling of the Mixed … 19

Fig. 2.6 The isoclines of perturbation of salinity: a t = 2, b t = 4, c t = 6, d t = 8

Fig. 2.7 The isoclines of perturbation of salinity: a t = 20, b t = 30

[email protected]
20 V. A. Gushchin and I. A. Smirnova

Fig. 2.8 The isoclines of stream function: a t = 20, b t = 30

taking into account the diffusion of stratification component (salt). To solve the task,
four-stage scheme of splitting was considered. The code is tested on the example of
calculation of a problem in the absence of spot and on different grids. The results
confirmed the execution of conservation laws with the required accuracy.
As the results of the calculations, the horizontal and vertical sizes of spots are
changing depending on the time. The obtained results are well coinciding with the-
oretical estimates, experimental data, and calculations of other authors. Previously
detected not monotonous changes of the linear sizes of the spots are confirmed. The
isoclines of stream function and perturbation of salinity are presented for different
moments of time. It is anticipated that the proposed model will receive better results,
when dealing with similar tasks.

References

1. Turner, J.S.: Buoyancy Effects in Fluids. Cambridge University Press (1973)

2. Scorer, R.S.: Environmental Aerodynamics. Ellis Horwood Publishers (1978)
3. Fedorov, K.N.: Fine Structure of the Thermohaline Ocean Waters. Gidrometeoizdat, Leningrad
(1976) (in Russian)
4. Monin, A.S., Ozmidov, R.B.: Ocean Turbulence. Gidrometeoizdat, Leningrad (1981) (in
Russian)
5. Barenblatt, G.I.: Similarity, Self-Similarity and Intermediate Asymptotics. Gidrometeoizdat,
Leningrad (1982). (in Russian)
6. Zatsepin, A.G., Fedorov, K.N., Voropaev, S.I., Pavlov, A.M.: Experimental research of collaps
mixed spot in stratified fluid. Izv. of the Academy of Sciences of the USSR. Phys. Atmos.
Ocean 14(2), 234–237 (1978) (in Russian)
7. Wu, J.: Mixed region collapse with internal wave generation in a density-stratified medium. J.
Fluid Mech. 35(3), 531–544 (1969)
8. Wessel, W.R.: Numerical study of the collapse of a perturbation in an infinite density stratified
fluid. Phys. Fluids 12(12), 170–176 (1969)
9. Kao, T.W.: Principal stage of wake collapse in a stratified fluid: two-dimensional theory. Phys.
Fluids 19(8), 1071–1074 (1976)
10. Chashechkin, Y.D., Zagumennyi, Y.V., Dimitrieva, N.F.: Dynamics of formation and fine struc-
ture of flow pattern around obstacles in laboratory and computational experiment. In: Voevodin,

[email protected]
2 The Splitting Scheme for Mathematical Modeling of the Mixed … 21

V., Sobolev, S. (eds.) RuSCDays 2016, CCIS, vol. 687, pp. 41–56. Springer International
Publishing AG (2016)
11. Gushchin, V.A., Mitkin, V.V., Rozhdestvenskaya, T.I., Chashechkin, Y.D.: Numerical and
experimental study of the fine structure of a stratified fluid flow over a circular cylinder. J.
Appl. Mech. Tech. Phys. 48(1), 34–43 (2007)
12. Gushchin, V.A., Matyushin, P.V.: Simulation and study of stratified flows around finite bodies.
Comput. Math. Math. Phys. 56(6), 1034–1047 (2016)
13. Gushchin, V.A.: The splitting method for problems of the dynamics of an inhomogeneous
viscous incompressible fluid. U.S.S.R. Comput. Math. Math. Phys. 21(4), 190–204 (1981)
14. Gushchin, V.A., Kopysov, A.N.: The dynamics of a spherical mixing zone in a stratified fluid
and its acoustic radiation. U.S.S.R. Comput. Math. Math. Phys. 31(6), 51–60 (1991)
15. Gushchin, V.A.: Family of quasi-monotonic finite-difference schemes of the second order of
approximation. Math. Models Comput Simul 8(5), 487–496 (2016)

[email protected]
Chapter 3
Modeling of Some Astrophysical
Problems on Supercomputers Using
Gas-Dynamic Model

Alexander V. Babakov , Alexey Y. Lugovsky

and Valery M. Chechetkin

Abstract In the current study, the vortex structures that occur in accretion disks are
investigated using mathematical modeling methods. The simulation of the processes
of formation of large-scale vortex structures in stellar accretion disks is carried out
by two methods with different numerical schemes. The first numerical technique is
based on conservative difference scheme with “upwind” approximation for fluxes.
The second numerical technique is based on an explicit, conservative, monotone in
the linear approximation Godunov-type Roe–Einfeldt–Osher scheme, which approx-
imates, with order no higher than the third, the conservation laws in the form of Euler
equations. Visualized pictures of the vortex structure are given by both methods for
accretion disks. The qualitative similarity of the obtained results is discussed. Evolu-
tionary calculations are carried out on the basis of parallel algorithms implemented
on the supercomputing complex of the cluster architecture.

3.1 Introduction

This chapter of the studying the structure of accretion disks around compact astro-
physical objects on the basis of numerical simulation using gas-dynamic models of
the environment is a continuation of researches [1–4]. In these researches, the calcu-
lations of the evolution of a massive third-generation star, which is the predecessor of

A. V. Babakov (B)
Institute for Computer Aided Design of the RAS, 19/18, Vtoraya Brestskaya ul., Moscow 123056,
Russian Federation
e-mail: [email protected]
A. Y. Lugovsky · V. M. Chechetkin
Keldysh Institute of Applied Mathematics of the RAS, 4, Miusskaya pl., Moscow 125047,
Russian Federation
e-mail: [email protected]
V. M. Chechetkin
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 23

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_3

[email protected]
24 A. V. Babakov et al.

a supernova, are carried out and the simulation of the development of hydrodynamic
instability in accretion disks with the formation of large vortexes is carried out also.
One of the main issues in the study of the structure of accretion disks is the
physics of redistribution and transfer of angular momentum [5]. Currently, within
the framework of the hydrodynamic approach, there are at least two mechanisms
that influence this process. The first mechanism is associated with the viscosity and
development of turbulence [6], while the second mechanism is associated with the
development of shear instability and with the formation of a large-scale structure of
turbulence [7–9]. Since the characteristic scales of turbulence in them are different,
there are also different times of transition of kinetic energy into thermal energy,
and different temperatures of accretion disks are corresponded to them. This leads
to different flow structures and times of evolution of accretion disks, as well as, to
the different rates of angular momentum transfer from the accretion disk [8, 9]. This
research is devoted to the second mechanism and explores the processes of formation
of large vortexes in the flow of a disk.
It should be noted that the formation of a large-scale vortex flow structure in
accretion disks is observed not only in the framework of the hydrodynamic approach.
For example, it is shown in [10] that in a gas-dust cloud rotating around a protostar in
the presence of a magnetic field, the large vortex structures leading to the transfer of
angular momentum are appeared. The mechanism of the occurrence of such structures
in the accretion gas-dust disk is the magneto-rotational instability. The occurrence
of a large-scale flow structure in the disks of spiral galaxies is demonstrated in [11].
In this chapter, two approaches with different formulations and difference schemes
are used to study the formation of large vortices in the flow of the disk, allowing us
to study this issue from different angles.
This chapter is organized as follows. Section 3.2 shortly presents the mathematical
models of studied problems. The results of the numerical modeling of the vortex
structures in accretion disks are introduced in Sect. 3.3. Lastly, Sect. 3.4 presents the
conclusions.

3.2 Mathematical Models of the Evolution of the Stellar

Accretion Disk

The chapter considers two approaches to modeling the structure of the accretion disk
using various mathematical models of the disk and various numerical methods for
solving the modeling problem.
The first approach to simulate a nonstationary motion of matter in fast-rotating
accretion disks uses a conservative numerical flux method [12, 13]. It is based on
a finite-difference representation of the conservation laws of density, momentum
components, and total energy for each finite volume that occurs when the integra-
tion region is discretized. In this case, for the approximation of the vectors of the
flux densities of the indicated characteristics, “upwind” approximations are used.

[email protected]
3 Modeling of Some Astrophysical Problems … 25

Previously, based on the algorithms of the method, a complex of parallel programs

for computing systems with cluster architecture was developed [14]. Parallel algo-
rithms of the complex are based on the standardized message transfer system called
as Message Passing Interface (MPI). During modeling spatial-nonstationary prob-
lems, paralleling in space in two or three directions is implemented, depending on
the specifics of the task, the size of the integration domain, and the parameters of the
computational mesh. During modeling fast-rotating gas objects, for which a rotation
is a substantially dominant motion, the modified finite-difference technique of the
flux method is used. It allows to preserve the components of the angular momentum,
for which conservation laws are not included in the basic system of equations based
on the laws of conservation of mass, components of momentum, and total energy
(when using the finite-difference analogue of conservation laws and time integra-
tion for large times, the conservation of components of the angular moment in the
area of integration may be disturbed). To preserve these components, the modified
finite-difference approximations of fluxes on the surface of a finite volume are used.
This make it possible, when integrating even for long times, to achieve the conser-
vation of the component of the angular momentum associated with the rotation with
an accuracy of 0.5% during the entire integration time. Moreover, for the other two
components, the angular momentum errors do not exceed 10−12 .
The second approach also uses a conservative numerical method based on the
finite-difference representation of the density, components of the momentum, and
total energy conservation laws for each finite volume in a flux form using the integro-
interpolation method. To obtain the fluxes at the cell boundaries, Godunov-type
Roe–Einfeldt–Osher method [15–18] based on an approximate solution of Riemann
problem is used. As is known, a linear monotone difference scheme cannot have an
order higher than the first. In this scheme, to circumvent this limitation, the nonlinear
limiters of anti-diffusion fluxes are used to increase the order of approximation to the
third, while preserving the monotony of the scheme in the linear approximation. The
system of Euler equations for gas dynamics is a nonlinear hyperbolic system, one of
the important features of which is a possibility of the appearance of discontinuous
solutions from smooth initial data. For problems of gas dynamics and especially
for astrophysical problems, flows are typical, in which shock waves and contact
discontinuities arise [11]. The accuracy of the numerical solution depends strongly
on the ability of the difference scheme to resolve gas-dynamic features. The used
Godunov-type Roe–Einfeldt–Osher scheme is conservative by recording the scheme
in the flux form and has a low numerical viscosity, but at the same time retains
a monotony of the solution. The efficiency of the schemes was confirmed in [8,
9, 19], in the same works it was numerically shown that the scheme retains the
angular momentum including the rotating component, which is especially important
for fast-rotating accretion disks. The built software package, when implementing a
parallel algorithm, uses a method of decomposition of the domain of calculation into
subdomains and the communication library MPI [4].
Below, on the basis of these program complexes, methods of mathematical mod-
eling are used to simulate the nonstationary motions of matter in accretion disks with
fast rotation. A gas-dynamic model of a perfect, non-viscous gas is used.

[email protected]
26 A. V. Babakov et al.

3.3 Numerical Simulation of the Evolution of the Stellar

Accretion Disk

Below we consider two model problems on the structure of an accretion disk that
is rapidly rotating around the gravitating center. Simulation of the evolution of the
external area of stellar accretion disk is represented in Sect. 3.3.1, while numerical
results of the formation of vortex structures in stellar accretion disks are discussed
in Sect. 3.3.2. The tasks differ in the geometric and mass parameters of the accretion
disk and the gravitating center.

3.3.1 Simulation of the Evolution of the External Area

of Stellar Accretion Disk

The results of modeling the structure of the outer region of the accretion disk, rapidly
rotating around a neutron star of mass M = 2.7846 * 1033 g (which corresponds to
1.4M ) and radius r 0 = 1.0 * 106 cm, are considered. Hereinafter, the subscript 
indicates the value of the solar parameter. The results of modeling the nonstationary
behavior of the outer part of the accretion disk of size R0 = 10r 0 and mass 1.088 * 1031
g are presented. The ratio of specific heats of gas γ in the disk is taken to be 5/3.
Simulation is carried out on the basis of parallel flux method algorithms [10, 11].
The mass of the central gravitating region (neutron star) exceeds the mass of
matter in the accretion disk by 257 times. Given this mass ratio, self-gravity inside
the accretion disk is neglected.
To divide the region of integration into finite volumes, a cylindrical coordinate
system (r, ϕ, z) is introduced. Finite volumes m are formed by splitting in constant
steps along the coordinate z, the radial r, and angular ϕ coordinates. The integration
is carried out in the cylindrical region  bounded by the inner surface of radius r 0
(the surface of a neutron star), the outer surface of radius R1 , and the lateral planes
z = –z0 and z = z0 :  = (r 0 ≤ r ≤ R1 ) * (0 ≤ ϕ ≤ 2π ) * (–z0 ≤ z ≤ z0 ).
The boundary conditions at the outer boundary of the integration region r = R1 are
determined by the zero derivative of the normal to the external integration region with
a positive value of the normal velocity component and the prohibition of convective
transfer into the integration region and zero gas dynamic variables with a negative
normal velocity component. On the inner boundary of the integration region r = r 0 ,
the conditions on a solid surface (impermeability condition) are specified. Along with
a cylindrical coordinate system intended for building a computational grid, Carte-
sian coordinate system is introduced with the center coinciding with the center of
the cylindrical coordinate system. The equations of motion are written in Cartesian
coordinate system for Cartesian components of the velocities, and the approximation
of Cartesian vectors of flux densities is carried out by reference points of finite vol-
umes m . The following geometrical parameters of the integration domain are used
in the calculation: R1 = 1.5R0 , z0 = 0.5R0 . The calculations apply the computational

[email protected]
3 Modeling of Some Astrophysical Problems … 27

grids containing up to 30 million finite volumes. The simulation is carried out on

computing complexes of cluster architecture using up to 2000 multicore processors.
The initial fields of pressure p and density ρ are set in such a way that at zero
speeds of motion equilibrium in the region of the accretion disk located in the central
gravitational field (neutron star) is realized. The polytropic equation of state of the gas
environment p = Kρ γ is used. The entropy in the initial field is constant throughout
the integration domain, which means hydrodynamic equilibrium, which, however,
does not prevent the formation of flows.
Further, dimensionless variables will be used: linear dimensions will be assigned
to R0 , pressure p̄ = p/ p0 , density ρ̄ = ρ/ρ0 , Cartesian velocity components v̄k =
vk /a0 , temperature T̄ = T /T0 , and time t¯ = t/t0 , where p0 , ρ 0 , and T 0 are the
pressure, density, and temperature on the surface of the accretion disk (r = r 0 )
at the initial time (t = 0), respectively, a0 is the characteristic speed, t0 = r0 /a0 ,
a02 = p0 /ρ0 .
Figure 3.1 shows the fields of the density (Fig. 3.1a) and angular velocity
(Fig. 3.1b) in the ϕ = 0 plane, as well as, their profiles in the radial direction (Fig. 3.1c,
d) at the initial time t = 0 (the density is in the logarithmic scale).
In the evolutionary calculation in the region, in which the density takes a value
less than the specified minimum ρ min , the convective transfer velocities, pressure,
and temperature are taken equal to 0. In the calculation ρ min = 10−8 .
The differences between the discrete assignment of gas-dynamic variables of the
initial field in the accretion disk and the gravitational field from the analytical repre-
sentation are a type of small perturbations. In the evolutionary calculation (especially
under conditions of a large gravitational mass of the star), these perturbations can
lead to a nonstationary numerical solution.
In the numerical integration over a sufficiently large time interval (t < 30), the
main parameters of the environment do not undergo noticeable changes. However,
with further integration over time, the numerical solution at the outer boundary of
the accretion disk in the region of rarefaction acquires a nonstationary character
with the formation of characteristic structures. We note here that a similar behavior
of a numerical solution was observed previously [20] when simulating subsonic
flow around a circular cylinder, where also initially implemented a stationary, but
unstable in a hydrodynamic sense, numerical solution is realized. This solution later
is integrated without introducing external disturbances and passed in the near wake
into the nonstationary, but stable mode of the vortex track—an analogue of Karman
vortex street. Thus, in Fig. 3.2 for the different points in time, the isosurfaces of the
density field of level 1 * 10−4 are shown in plane z = 0.
The accretion disk loses its axial symmetry. Proof of this in the form of density
fields is presented in Fig. 3.3 for different times (in a logarithmic scale). The pictures
are shown in the plane ϕ = 0 and ϕ = π, passing through the axis of rotation of the
accretion disk.
These pictures give an idea of the spatially nonstationary structure of the outer
boundary of the accretion disk.

[email protected]
28 A. V. Babakov et al.

Fig. 3.1 Visualization of initial fields and profiles in the equatorial plane at the initial moment of
time (the density is in the logarithmic scale): a initial field of the density, b initial field of angular
velocity, c profile of the density, and d profile of the angular velocity

3.3.2 Numerical Results of the Formation of Vortex

Structures in Stellar Accretion Disks

The task of modeling the structure of an accretion disk fast rotating around a central
gravitating compact object with mass M = 2 * 1033 g (which corresponds to M )
is considered. Specific radius of the accretion disk is R0 = 1011 cm. The ratio of
specific heats γ is equal to 5/3. Self-gravitation of the matter of a disk is not taken
into consideration, since it is consumed that the mass of the central gravitating body
is almost two orders of magnitude greater than the mass of the matter of the disk.
The integration domain is divided into finite volumes m by entering a uniform
grid in cylindrical coordinates (r, ϕ, z), where r is the cylindrical radius, ϕ is the polar
angle, and z is the height. Integration is carried out in a cylindrical area  bounded
by an inner radius r 0 , by an outer radius R1 and by side planes z = –z0 and z = z0 :
 = (r 0 ≤ r ≤ R1 ) * (0 ≤ ϕ ≤ 2π ) * (–z0 ≤ z ≤ z0 ).

[email protected]
3 Modeling of Some Astrophysical Problems … 29

Fig. 3.2 Isosurfaces of the density in plane z = 0 of the accretion disk at different points in time:
a t = 50, b t = 70, c t = 100, d t = 150, e t = 170, and f t = 200

On the boundaries r = r 0 , r = R1, z = –z0 , z = z0 of the computational domain,

we set “free” boundary conditions, which are determined by the zero of the nor-
mal derivative. In the calculations below geometrical parameters r 0 = 0.15R0 , R1
= 1.8R0 , z0 = 0.2R0 and the region  = (0.15 ≤ r ≤ 1.8) × (0 ≤ ϕ ≤ 2π ) ×
(−0.2 ≤ z ≤ 0.2) are used. Computational meshes containing up to several mil-
lion finite volumes are used in calculations. The modeling is carried out by cluster
architecture computer systems with up to 256 processors.

[email protected]
30 A. V. Babakov et al.

Fig. 3.3 The density fields in the plane ϕ = 0 and ϕ = π of the accretion disk at different points
in time: a t = 0, b t = 50, c t = 200, and d t = 300 (in the logarithmic scale)

The initial fields of the pressure p, the density ρ, and the velocity ν̄ are selected
by the equilibrium state obtained in work [21] as an analytical solution for the
equilibrium gas configuration near the gravitating center.
Further dimensionless variables are used. As scale factors, we choose R0 , M, and
G, where G is the gravitational constant. The dimensionless variables, which are
marked with a prime, are introduced according to the formulas:

r = R0 r  , z = R0 z  , n̄ = v0 n̄ , t = t0 t  , p = p0 p  , ρ = ρ0 ρ  , e = e0 e .

The multipliers v0 , t0 , p0 , ρ0 , e0 are given by

GM GM R03 M G M2
v20 = , e0 = , t02 = , ρ0 = , p0 = .
R0 R0 GM R03 R04

In Fig. 3.4, the fields of the density (Fig. 3.4a) and the angular velocity (Fig. 3.4b)
in the plane ϕ = 0 are shown, as well as their profiles in the radial direction (Fig. 3.4c,
d) at the initial time t = 0.
In the evolutionary calculation in the area, in which the density takes a value less
than the specified minimum ρ min convective transfer velocities, the pressure and the
temperature are taken equal to 0. In the calculation, the minimum density is ρ min =
10−6 . Using the well-known technique of introducing small perturbations in a small
region into the initial state of the accretion disk [4], let us follow the evolution of the
flow in the accretion disk. Note that in contrast to [4], perturbations are injected to
density and closer to the outer edge of the disk.

[email protected]
3 Modeling of Some Astrophysical Problems … 31

Fig. 3.4 The initial fields and profiles in the equatorial plane at the initial moment of time: a initial
fields of the density, b initial fields of the angular velocity, c profiles of the density, and d profiles
of the angular velocity

In numerical integration, the main parameters of the environment undergo fairly

rapid changes, and the numerical solution in the accretion disk acquires a nonstation-
ary character with the formation of characteristic vortex structures. Thus, in Fig. 3.5
for different points in time, the density field patterns in the plane z = 0 are shown.
The accretion disk loses its axial symmetry. Proof of this is the density fields
presented in Fig. 3.6 for different times. The pictures are shown in plane ϕ = 0
passing through the axis of rotation of the accretion disk.
These pictures reflect the vortex structure of the flow of the accretion disk.

3.4 Conclusions

The important result of the completed numerical studies is the occurrence of vortex
motion in stellar accretion disks obtained by different numerical methods. In this

[email protected]
32 A. V. Babakov et al.

Fig. 3.5 Density patterns in the plane z = 0 of the accretion disk at various points in time: a t =
1, b t = 2, c t = 3, d t = 4, e t = 5, and f t = 6

[email protected]
3 Modeling of Some Astrophysical Problems … 33

Fig. 3.6 Density fields in the plane ϕ = 0 of the accretion disk at different points in time: a t = 1,
b t = 2, c t = 3, d t = 4, e t = 5, and f t = 6

[email protected]
34 A. V. Babakov et al.

case, the flows obtained in different models have a qualitative similarity in the sense
that large spiral vortex structures leading to a loss of flow symmetry are formed
in the flow, while the flow remains vortex. It is already known that the vortex flows
have a significant influence on the evolution of accretion disks. Therefore, the authors
consider their studying necessary not only to understand the properties and evolution
of the vortex flows but also for research of the evolution of accretion disks, which
are significantly influenced by the vortexes.
Another important result is a demonstration of the possibility to model the real
astrophysical objects on the basis of supercomputers in real physical conditions
including huge space-time areas. Calculations are carried out on the computational
resources of the JSCC of the RAS as well as using the equipment of the Research
Computing Center of the Lomonosov Moscow State University.

Acknowledgements The authors are grateful to A.G. Aksenov for helpful discussions of the setting
of the initial field for the modeling of fast-rotating stellar accretion disks.
The work is performed in the framework of state assignments of the ICAD RAS and KIAM
RAS.

References

1. Babakov, A.V., Popov, M.V., Chechetkin, V.M.: Mathematical simulation of a massive star
evolution based on a gasdynamical model. Math. Model. Comput. Simul. 10(3), 357–362
(2018)
2. Aksenov, A.G., Babakov, A.V., Chechetkin, V.M.: Mathematical simulation of the vortex struc-
tures in the fast rotation astrophysical objects. Comput. Math. Math. Phys. 58(8), 1287–1293
(2018)
3. Babakov, A.V., Lugovsky, A.Y., Chechetkin, V.M.: Mathematical modeling of the evolution
of compact astrophysical gas objects. In: Petrov I.B., Favorskaya, A.V., Favorskaya, M.N.,
Simakov, S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development
of Grid-Characteristic Method, SIST, vol. 133, pp. 210–227. Springer (2019)
4. Lugovsky, A.Y., Popov, Y.P.: Roe–Einfeldt–Osher scheme as applied to the mathematical sim-
ulation of accretion disks on parallel computers. Comput. Math. Math. Phys. 55(8), 1407–1418
(2015)
5. Boyarchuk, A.A., Bisikalo, D.V., Kuznetsov, O.A., Chechetkin, V.M.: Mass Transfer in Close
Binary Stars. Taylor and Francis, London and New York (2002)
6. Shakura, N.I., Sunyaev, R.A.: Black holes in binary systems. Observational appearance. Astron.
Astrophys. 24, 337–355 (1973)
7. Belotserkovskii, O.M., Oparin, A.M., Chechetkin, V.M.: Turbulence: New approaches Nauka.
Moscow (2003)
8. Velikhov, Y.P., Lugovsky, A.Y., Mukhin, S.I., Popov, Y.P., Chechetkin, V.M.: The impact of
large-scale turbulence on the redistribution of angular momentum in stellar accretion disks.
Astron. Rep. 51(2), 154–160 (2007)
9. Lugovsky, A.Y., Mukhin, S.I., Popov, Y.P., Chechetkin, V.M.: The development of large-scale
instability in stellar accretion disks and its influence on the redistribution of angular momentum.
Astron. Rep. 52(10), 811–814 (2008)
10. Velikhov, Y.P., Sychugov, K.R., Chechetkin, V.M., Lugovskii, A.Y., Koldoba, A.V.: Magneto-
rotational instability in the accreting envelope of a protostar and the formation of the large-scale
magnetic field. Astron. Rep. 56(2), 84–95 (2012)

[email protected]
3 Modeling of Some Astrophysical Problems … 35

11. Lugovskii, A.Y., Filistov, E.A.: Numerical modeling of transient structures in the disks of spiral
galaxies. Astron. Rep. 58(2), 48–62 (2014)
12. Belotserkovskii, O.M., Severinov, L.I.: The conservative “flow” method and the calculation of
the flow of a viscous heat-conducting gas past a body of finite size. Comput. Math. Math. Phys.
13(2), 141–156 (1973)
13. Belotserkovskii, O.M., Babakov, A.V.: The simulation of the coherent vortex structures in the
turbulent flows. Adv. Mech. 13(3/4), 135–169 (1990)
14. Babakov, A.V.: Program package FLUX for the simulation of fundamental and applied
problems of fluid dynamics. Comput. Math. Math. Phys. 56(6), 1151–1161 (2016)
15. Osher, S., Solomon, F.: Upwind difference schemes for hyperbolic systems of conservations
laws. Math. Comput. 38, 339–374 (1982)
16. Chakravarthy, S., Osher, S.: A new class of high accuracy TVD schemes for hyperbolic
conservation laws. AIAA Pap. 85(0363), 1–11 (1985)
17. Einfeldt, B.: On Godunov type methods for gas dynamics. SIAM J. Numer. Anal. 25, 294–318
(1988)
18. Kuznetsov, O.A.: Preprint No. 43. IPM RAN (Keldysh Institute of Applied Mathematics of
Russian Academy of Sciences), Moscow (in Russian) (1998)
19. Lugovskii, A.Y., Chechetkin, V.M.: The development of large-scale instability in Keplerian
stellar accretion disks. Astron. Rep. 56(2), 96–103 (2012)
20. Babakov, A.V.: On the possibility of the numerical modeling of non-stationary vortex structures
in a near wake. Comput. Math. Math. Phys. 28(1), 173–180 (1988)
21. Abakumov, M.V., Mukhin, S.I., Popov, Y.P., Chechetkin, V.M.: Studies of equilibrium con-
figurations for a gaseous cloud near a gravitating center. Astron. Rep. 40(3), 366–377
(1996)

[email protected]
Chapter 4
Numerical Modeling of the Kolmogorov
Flow in a Viscous Media

Svetlana V. Fortova and Alexey N. Doludenko

Abstract In this chapter, we consider the problem proposed by Kolmogorov to study

the causes of turbulence. We consider the action of the external periodic field alone
in one of the coordinates on a viscous conductive media in the two-dimensional case.
We propose a numerical study of this problem based on the solution of the systems
of Navier–Stokes equations. The calculations show that under certain conditions
periodic vortex structures may appear in the liquid similar to the “parquet” mode in
Kolmogorov task with the subsequent development of the self-similar regime.

4.1 Introduction

The study of the laminar flow stability with respect to small perturbations that always
exist in nature is of particular interest both for theoretical studies and for practical
applications [1]. This is explained by the fact that the explanation of diverse complex
fluid motions and problem of the occurrence of large-scale eddy currents as well are
associated with questions of stability [1, 2].
Despite the fact that three-dimensional turbulence is a diverse and essentially
non-linear phenomenon, interest to two-dimensional turbulence attracts the atten-
tion of many researchers [3–15]. Kreichnan and Batchelor [16, 17] established the
fact that two-dimensional turbulence is not a simplified model of three-dimensional
turbulence, but has its own unique properties. According to [18], the essential dif-
ference between two-dimensional and three-dimensional turbulence is as follows:
In three-dimensional case, motions are generated with scales smaller than the scale,

S. V. Fortova (B)
Institute of Computer Aided Design of the RAS, 19/18 Vtoraya Brestskaya ul., Moscow 123056,
Russian Federation
e-mail: [email protected]
A. N. Doludenko
Joint Institute for High Temperatures of the RAS, 13 Str.2, Izhorskaya ul., Moscow 125412,
Russian Federation
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 37

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_4

[email protected]
38 S. V. Fortova and A. N. Doludenko

at which turbulence is excited (the “pumping” scale) [1, 2]. In this case, the energy
is distributed in a direct cascade with Kolmogorov law –5/3 [18, 19] in the inertial
range of the kinetic energy spectrum. In the two-dimensional case, nonlinearity leads
to the appearance of motions with scales far exceeding the “pumping” scale with the
appearance of large coherent structures [18]. Energy can be distributed in a reverse
cascade (from small structures to large ones) with the –5/3 law of Kreichnan [16,
18] and is also transferred from the pump scale to small scales (direct cascade) with
the –3 law due to dissipation of enstrophy. Such movements occur in the formation
of cyclones and anticyclones, when the height of the atmosphere is about 10 km, the
cyclone size is about 100 km, and consequently, the atmospheric flow can be referred
to as quasi-two-dimensional [1].
The inverse cascade in two-dimensional turbulence was studied experimentally
[5] and numerically [6, 7]. The peculiarities of these studies are the emergence of
intensive large-scale movement including large eddies. In [8], a vortex dipole-stable
coherent structure was obtained numerically in a square cell with periodic boundary
conditions. In the field experiment [12], a stable coherent structure was also obtained.
In [2], the birth of a periodic self-oscillatory regime is considered as the first step
in the transition from a laminar flow to a turbulent one. It is known [1] that in the
flat case there is no strictly defined transition point: the laminar flow is transformed
into completely chaotic over a certain transition region, where secondary, as a rule,
oscillatory movements are superimposed on the main flow. It is of interest to obtain
general flow patterns in transition regions, study the dynamics of their behavior, deter-
mine the critical values of flow parameters, as well as, find self-oscillating periodic
flows and the possibilities of transition to chaos [20]. For geophysical applications,
it is of interest to study the stability of such a class of flows, for which the scales
of unstable perturbations are commensurate with the spatial scale of the main flow.
Such a study, apparently, can provide a key to explaining the evolution of cyclones
and anticyclones in the atmosphere and synoptic eddies in the ocean, etc. [1, 2, 18].
In this chapter, we consider the problem proposed by Kolmogorov to study the
causes of turbulence in the two-dimensional case and that was experimentally inves-
tigated in [21]. It is a study of the plane flow of an incompressible fluid under the
action of an external force, periodic in the transverse direction. In the linear formu-
lation, this problem was studied in [22, 23], where the fact of stability loss of the
main laminar flow with respect to spatially periodic perturbations with a long wave-
length along the flow was proved. The problem of the non-linear development of
perturbations and the occurrence of secondary stationary or periodic flows with loss
of stability of the laminar flow was described in [20, 23]. However, the issues related
to the existence of stable stationary or self-oscillatory regimes of the secondary flow,
as well as, the possibility of transition to chaos are still largely open.
We propose a numerical study of the problem of the flat flow of a viscous slightly
compressible fluid under the action of a periodic in the transverse direction force. The
flow parameters that lead to the appearance of a “vortex parquet” and the subsequent
loss of stability of the secondary flow have been determined.

[email protected]
4 Numerical Modeling of the Kolmogorov Flow in a Viscous Media 39

Fig. 4.1 Scheme of the

experiment

The chapter is structured as follows. Description of the experiment is given in

Sect. 4.2. Section 4.3 provides the problem statement and numerical method. Results
of the conducted experiment are presented in Sect. 4.4. Section 4.5 concludes the
chapter.

4.2 Description of the Experiment

Our numerical simulation is an interpretation of an experiment that can be found in

[21]. Scheme of the experiment is shown in Fig. 4.1.
A flat horizontal rectangular cuvette was filled with an electrically conductive
electrolyte aqueous solution (CuSO4). With the help of electrodes mounted on the
longitudinal sidewalls of the cell, a constant electric current was passed through the
electrolyte in the transverse direction. The cell with the electrolyte was mounted on
a sheet of magnetoelastic rubber, which served as the source of an external magnetic
field (see Fig. 4.1). By special magnetization, the magnetic field strength was created
with a profile close to sinusoidal. Thus, Lorentz electromagnetic force acted on the
moving fluid. As a result of this experiment, the authors managed to get a clear
picture of the self-oscillating regime of vortex structures called “vortex parquet”
or “Kolmogorov parquet.” We propose a numerical study of the flat flow problem
mainly of a viscous weakly compressible fluid under the action of a periodic force.

4.3 Problem Statement and Numerical Method

The problem statement and numerical method are considered in Sects. 4.3.1–4.3.2,
respectively.

[email protected]
40 S. V. Fortova and A. N. Doludenko

4.3.1 Problem Statement

Consider the problem of a flat flow of a viscous weakly compressible fluid under
the action of an external periodic force directed along OX-axis, which equals to
ρG sin(ky). Here, G = 0.01 N /kg is Lorentz force equaled to the vector product of
the strength of the current passed through the liquid, by the magnetic field strength,
k is the wavenumber that specifies the period of the force (in our calculation k =
1). This G force came from the experiments with the conducting fluid, which were
carried out by other researchers. The motion of the medium in this case is described
by Navier–Stokes equations in the form of Eq. 4.1.

∂ρ
+ ∇ · (ρV) = 0
∂t
∂ρu ∂p
+ ∇ · (ρuV) = − + ρG sin ky + μu
∂t ∂x
∂ρv ∂p
+ ∇ · (ρvV) = − + μv
∂t ∂y
p = ρ/β (4.1)

Here, V = (u, v)T , where u, v are the components of the velocity vector, along
OX- and OY- axes, respectively, P is the pressure, ρ is the density, β is the artificial
compressibility factor.
The flow is investigated in a rectangular area with periodic boundary conditions.
Size of this calculation domain equals to 8π × 4π along OX and OY axes, respec-
tively. Calculations were performed on a grid size of 200 × 100 and 800 × 400 cells
along OX- and OY- axes, respectively.
In the calculation, the initial condition was immediately set as the sum of the main
flow taken as u = sin(y) and superimposed small disturbances, which can be found
in Eqs. 4.2–4.3.

u(t = 0) = 0.1 sin(y) + 0.001 sin(x/2) (4.2)

v(t = 0) = 0.1 sin(y) + 0.001 sin(x/2) (4.3)

It is known that for an incompressible fluid, the long-wave perturbations imposed

on the main flow field are the most unstable. Therefore, such a superposition of
the main flow velocity and small perturbations superimposed on the main flow are
presented as the initial conditions for the velocity field. Such a statement of the initial
conditions made it possible to obtain an analogue of the self-oscillatory regime.
Other used initial conditions can be represented as follows:

P(t = 0) = P0 = 105 Pa, ρ = 1000 kg /m3 , μ = 2 Pas .

[email protected]
4 Numerical Modeling of the Kolmogorov Flow in a Viscous Media 41

Here μ is the viscosity of fluid.

4.3.2 Numerical Method

The calculation algorithm used for viscous medium modeling is based on MacCor-
mack explicit method, which is of second order in accuracy, in time, and in space
and well-proven in solving hyperbolic equations. Navier–Stokes equations are solved
using the method of artificial compressibility [24]. In this case, the hyperbolic part
of the equations is solved by MacCormack method, and the parabolic part is solved
using standard finite difference method.
The spectral representation of the kinetic energy will be obtained by decomposi-
tion in a two-dimensional Fourier integral. Each of the velocity components can be
expanded into a series of orthogonal harmonic functions:
  (1) (2)
vi (x, y) = [vi (k x , k y ) cos(k x x) cos(k y y) + vi (k x , k y ) cos(k x x) sin(k y y)
kx ky i = 1, 2
(3) (4)
+ vi (k x , k y ) sin(k x x) cos(k y y) + vi (k x , k y ) sin(k x x) sin(k y y)],
(4.4)

where εvi is one of the velocity components , k x and k y are the wave vector com-
( j)(i)
ponents along OX and OY, respectively, vi (k x , k y ) j = 1 ÷ 4 are the Fourier
coefficients obtained as

2π 2π
1
vi(1) (k x , k y ) = vi (x, y) cos(k x x) cos(k y y)d xd y, i = 1, 2,
π
0 0
2π 2π
1
vi(2) (k x , k y ) = vi (x, y) cos(k x x) sin(k y y)d xd y, i = 1, 2,
π
0 0
2π 2π
1
vi(3) (k x , k y ) = vi (x, y) sin(k x x) cos(k y y)d xd y, i = 1, 2,
π
0 0
2π 2π
1
vi(4) (k x , k y ) = vi (x, y) sin(k x x) sin(k y y)d xd y, i = 1, 2. (4.5)
π
0 0

Value
 2  2
v1 (k x , k y ) + v2 (k x , k y )
ε(k x , k y ) = , wher e
2

[email protected]
42 S. V. Fortova and A. N. Doludenko

 2  2  2  2
(1) (2) (3) (4)
vi (k x , k y ) = vi (k x , k y ) + vi (k x , k y ) + vi (k x , k y ) + vi (k x , k y ) , i = 1, 2
(4.6)

will be the desired image of the kinetic energy in the space of wave numbers.
So, calculating Fourier coefficients (Eq. 4.5), and, further, calculating Eq. 4.6, we
obtain a quantity ε(k x , k y ) depending on k x and k y . Going through all k x and k y in
both directions, we obtain the energy spectrum.

4.4 Results

We will give the views of arising characteristic flows of viscous fluid. From Fig. 4.2,
it can be seen that a “parquet” of not so big eddies first arises from the initial state
with the subsequent development of two relatively large vortices.
In the same Fig. 4.2, one can see the series of the vorticity magnitude taken in
different moments of time. It can be seen that from almost initial state (time = 60,
let name it phase 1) vortices form a parquet-like pattern (time = 200, let name it
phase 2), which over time combine at first into four vortices (time = 400, let name
it phase 3), and then into two vortices (time = 1000, 3140, let name it phase 4),
which rotate in different directions. Each of the pictures corresponds to more or less
stable pattern. These quasi-stable regimes exist on the time moments between peaks
of kinetic energy and enstrophy, which one can see in Fig. 4.3a. Thus, each peak
separates arising quasi-stationary flows. At the moments of local peaks, transient
regimes occur transforming one quasi-stationary state into another.
Turning to the consideration of the viscous case in detail, we can pay attention to
Fig. 4.3. One can see the graphs of kinetic energy and enstrophy on time normalized
on the maximum value of these parameters. The enstrophy can be described as the
integral of the square of the vorticity. In Fig. 4.3a, there are graphs for the case with
the region size equals to 200 × 100 computational cells. In Fig. 4.3b, there are graphs
for the case with the region size equals to 800 × 400 computational cells. Before
moment of time = 400 the behavior of fluid in two of these cases remains similar.
Even the time of transition between phase 1 and phase 2 is the same and equals to
approximately 70. Time of transition to the phase 3 is almost the same (t ≈ 300)
also but after the phase 3 in more rough mesh it goes into stable phase 4 during the
period of time from ≈520 to 1000. Not such a picture is observed in the case of a
finer grid. In this case, phase 3 seems to be stable, but nonetheless goes into a truly
stable phase 4 in the time zone from 1000 to 1500 (Fig. 4.2).
In the zone of time which equals to 1000 for rough mesh and 1500 for the finer
mesh after a long transient regime, a system consisting of two vortices rotating in
opposite directions begins to form. The conditional left vortex rotates counterclock-
wise, and the conditional right vortex rotates clockwise. After that, the system of two
vortices goes into the self-similar state. Each of the vortices from time to time forms
a “ring.” Further, this “ring” is divided into two vortices rotating in one direction and

[email protected]
4 Numerical Modeling of the Kolmogorov Flow in a Viscous Media 43

Fig. 4.2 Vorticity magnitude taken in different moments of time for different regions: a time = 60,
size = 200 × 100 cells, b time = 60, size = 800 × 400 cells, c time = 200, size = 200 × 100 cells,
d time = 200, size = 800 × 400 cells, e time = 400, size = 200 × 100 cells, f time = 400, size
= 800 × 400 cells, g time = 600, size = 200 × 100 cells, h time = 600, size = 800 × 400 cells,
i time = 1000, size = 200 × 100 cells, j time = 1000, size = 800 × 400 cells, k time = 3140, size
= 200 × 100 cells, l time = 2640, size = 800 × 400 cells

[email protected]
44 S. V. Fortova and A. N. Doludenko

Fig. 4.3 Graphs of kinetic energy and enstrophy on time normalized on the maximum value of
these parameters: a size of calculated region equals to 200 × 100 cells, b size of calculated region
equals to 800 × 400 cells

located close to each other. As a result of their interaction, a larger vortex is formed,
which later, in its turn, forms a “ring.” The process is repeated cyclically throughout
the observed time up to 12,800 time steps (for rough mesh). This self-similar mode
is observed with the minimum value of enstrophy. It can be also seen as a more
pronounced increase in kinetic energy in phase 4 for a finer grid.
At first glance, the behavior of a turbulent system, when energy enters, should
tend from order to chaos. This is how most systems behave. Over time, large vor-
tices should disintegrate into small ones. And this process is indeed observed when
studying 3D turbulence. However, in 2D case, everything happens the other way
around. The entry of energy into a chaotic mixture of small vortices (phase 2) in a

[email protected]
4 Numerical Modeling of the Kolmogorov Flow in a Viscous Media 45

Fig. 4.4 Kinetic energy spectrum at different moments of time, 200 × 100 cells: a linear scale,
b logarithmic scale

two-dimensional system leads to the fact that vortices rotating in the same direction
will form more and more large vortices (phase 3) with time. Those, in turn, interact
with each other until a pair of stable vortices remains in the system (phase 4). In this
case, the system becomes more orderly rather than chaotic. That means energy has
to flow from small structures to large ones with the –5/3 law of Kreichnan [16, 18].
If we consider the kinetic energy spectrum in our case (Fig. 4.4), then this kind of
spectrum is not clearly observed. In Fig. 4.4, one can find four graphs corresponding
to the above phases of fluid flow. Thus, we can observe different peaks on the kinetic
energy graphs (these peaks are better seen in Fig. 4.4 with linear scale).
Over time, some peaks disappear, for example, in the region of k = 28. Others
decrease in amplitude, for example, in the region of k = 56, 74. Still others are
formed in new locations, for example, peaks in the region of k = 38, 90. All this
says that there are different modes that carry most of the kinetic energy. And the
kinetic energy flows from one mode to another over time. All this happens because
we observe not the uniform turbulence, but phases passing into each other with more
or less regular structures (vortices). Moreover, during the transition from one phase
to another, the structures are combined or, in other words, they are enlarged.

4.5 Conclusions

The result of this research is a direct numerical simulation of the vortex flow formation
regime in a layer of weakly compressible medium based on Navier–Stokes equations.
Namely, a small perturbation of the velocity components leads to the appearance of a
“vortex parquet” and, further, the two rotating vortexes. In addition, when modeling
viscous fluid, several transient modes were found separated from each other by
maxima of kinetic energy or enstrophy. In the case of a finer computational grid,
phase 3 turned out to be more extended in time, and the transition to the self-similar

[email protected]
46 S. V. Fortova and A. N. Doludenko

regime took place somewhat later than in the case of a coarser grid. Besides that, the
self-similar regime is accompanied by the minimum value of enstrophy.

References

1. Belocerkovskij, O.M., Oparin, A.M.: Numerical Experiment: From Order to Chaos. Nauka,
Moscow (in Russian) (2000)
2. Landau, L.D., Lifshicz, E.M.: Continuum Mechanics. Gostex-teorizdat (In Russian) (1953)
3. Boffetta, G., Ecke, R.E.: Two-dimensional turbulence. Rev. Fluid Mech. 44, 427–451 (2012)
4. Kraichnan, R.H., Montgomery, D.: Two-dimensional turbulence. Rep. Prog. Phys. 43(5), 547–
619 (1980)
5. Sommeria, J.: Experimental study of the two-dimensional inverse energy cascade in a square
box. Fluid Mech. 170, 139–168 (1986)
6. Smith, L.M., Yakhot, V.: Finite-size effects in forced two-dimensional turbulence. J. Fluid
Mech. 274, 115–138 (1994)
7. Boffetta, G., Celani, A., Vergassola, M.: Inverse energy cascade in two-dimensional turbulence:
deviations from Gaussian behavior. Phys. Rev. E Stat. Phys. Plasmas Fluids Relat. Interdiscip.
Topics. E 61(1), R29–R32 (2000)
8. Chertkov, M., Connaughton, C., Kolokolov, I., Lebedev, V.: Dynamics of energy condensation
in two-dimensional turbulence. Phys. Rev. Lett. 99(8), 084501.1–084501.4 (2007)
9. Falkovich, G.: Symmetries of the turbulent state. J. Phys. A: Math. Theor. 42(12), 123001.1–
123001.18 (2009)
10. Kolokolov, I.V., Lebedev, V.V.: Profile of coherent vortices in two-dimensional turbulence.
JETP Lett. 101(3), 164–167 (2015)
11. Kolokolov, I.V., Lebedev, V.V.: Velocity statistics inside coherent vortices generated by the
inverse cascade of 2-D turbulence. J. Fluid Mech. 809, R2.1–R2.11 (2016)
12. Infeld, E., Rowlands, G.: Nonlinear Waves, Solitons and Chaos, Cambridge UP (2000)
13. Xia, H., Shats, M., Falkovich, G.: Spectrally condensed turbulence in thin layers. Phys. Fluids
21(12), 125101.1–125101.11 (2009)
14. Laurie, J., Boffetta, G., Falkovich, G., Kolokolov, I., Lebedev, V.: Universal profile of the vortex
condensate in two-dimensional turbulence. Phys. Rev. Lett. 113, 254503.1–254503.5 (2014)
15. Fortova, S.V., Oparina, E.I., Belotserkovskaya, M.S.: Numerical simulation of the Kolmogorov
flow under the influence of the periodic field of the external force. J. Phys.: Conf. Ser. 1128,
012089.1–012089.5 (2018)
16. Kraichnan, R.H.: Inertial ranges in two-dimensional turbulence. Phys. Fluids 10(7), 1417–1423
(1967)
17. Batchelor, G.K.: Computation of the energy spectrum in homogeneous two-dimensional
turbulence. Phys. Fluids 12(12), II-233–II-239 (1969)
18. Frik, P.G.: Turbulence: Models and Approaches. Lecture Course. Part 2. Perm (in Russian)
(1999)
19. Kolmogorov, A.N.: Local structure of turbulence in incompressible fluid at very high Reynolds
numbers. Dokl. Akad. Nauk SSSR 30(4), 299–303 (in Russian) (1941)
20. Belocerkovskij, S.O., Mirabel, A.P., Chusov, M.A.: On the construction of a supercritical
regime for a plane periodic flow. Izv. Acad. Sci., USSR, Atmos. Oceanic Phys. 14(1), 11–20
(in Russian) (1978)
21. Gledzer, E.B., Dolzhanskij, F.V., Obuxov, A.M.: Hydrodynamic Type Systems and Their
Application. Nauka, Moscow (in Russian) (1981)
22. Meshalkin, L.D., Sinaj, Ya.G.: Investigation of the stationary solution stability of one system
of equations of an incompressible viscous fluid plane motion. J. Appl. Math. Mech. 25(6),
1140–1143 (in Russian) (1961)

[email protected]
4 Numerical Modeling of the Kolmogorov Flow in a Viscous Media 47

23. Yudovich, V.I.: On the instability of parallel flows of a viscous incompressible fluid with respect
to spatially periodic perturbations. Zh. Vychisl. Mat. Mat. Fiz. 242–249 (in Russian) (1966)
24. Anderson, D., Tannehill, J, Pletcher, R.: Computational Fluid Mechanics and Heat Transfer.
Mir, Moscow (in Russian) (1990)

[email protected]
Chapter 5
On Structures of Supersonic Flow
Around Plane System of Cylindrical Rods

Sergey V. Guvernyuk and Fedor A. Maksimov

Abstract The chapter presents the results of numerical simulation of two-

dimensional laminar flows near a regular system of cylinders, forming a plane lattice
perpendicular to the velocity vector of the oncoming supersonic flow. A multiblock
computing technology is applied using local curvilinear grids adapted to the surface
of bodies and having finite areas of overlap with a global rectangular grid. The vis-
cous boundary layers are resolved on the local grids using Navier–Stokes equations.
The interaction of shock-wave structures and aerodynamic wakes behind the ele-
ments of the lattice is described within Euler equations. With a sequential increase
and decrease in Mach number of the oncoming flow, several rearrangements of the
flow structure near the grid are found. A multiple hysteresis was revealed, which
is expressed in the fact that the flow structure and aerodynamic loads on the lattice
elements depend not only on Mach number but also on the history of its change.

5.1 Introduction

The study of a supersonic flow around the bodies containing permeable structures
is important for a number of technical applications [1–3]. A supersonic flow around
a lattice of cylinders is of interest as a model problem, by the example of which it
is convenient to research the interrelation of the external large-scale and local near-
wall flows near the permeable bodies, such as mesh or perforated screens, distributed
systems of solid particles, etc. Very complex interaction of local compression shocks

S. V. Guvernyuk · F. A. Maksimov (B)

Institute of Mechanics, Lomonosov Moscow State University, 1, Michurinskii prt., Moscow
119192, Russian Federation
e-mail: [email protected]
S. V. Guvernyuk
e-mail: [email protected]
F. A. Maksimov
Institute of Computer Aided Design of the RAS, 19/18, Vtoraya Brestskaya ul., Moscow 123056,
Russian Federation

© Springer Nature Singapore Pte Ltd. 2020 49

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_5

[email protected]
50 S. V. Guvernyuk and F. A. Maksimov

occurs among themselves and with the aerodynamic wakes behind the lattice ele-
ments [4–6] in the case of sufficiently rarefied grids. Moreover, different schemes of
the resulting flows of such interaction are equally possible for some combinations of
the problem parameters. This indicates a possibility of the parametric hysteresis.
Hysteresis of the flow structure behind an infinite lattice of cylinders is revealed
in [5] under the conditions, when both regular and Mach schemes of intersection of
the oblique shocks propagating from the neighboring lattice elements, are equally
possible. Two types of hysteresis were identified by the lattice permeability in the
study [6], at fixed Mach number M = 6. One of them is associated with the destruction
of the collective flow around the lattice elements. The second one is connected with
the rearrangement of the near wake behind the lattice elements under the influence
of the local shock waves from the neighboring lattice elements.
The chapter considers the supersonic flow around a geometrically unchanged lat-
tice of cylinders (permeability 80%) with a variation in Mach number in the direction
of increasing and decreasing in the range from 2.0 to 4.5.
The chapter is structured in the following manner. Section 5.2 provides a for-
mulation of problem. Multiblock computing technology is discussed in Sect. 5.3.
The calculation results for lattice of 10 elements and infinite lattice with periodical
conditions are represented in Sects. 5.4 and 5.5, respectively. Conclusions are given
in Sect. 5.6.

5.2 Formulation of Problem

The flow around the grid of circular cylindrical rods with axes parallel to each other
and lying in a plane, perpendicular to the direction of a uniform supersonic stream
of a viscous perfect gas, is considered. It is assumed that the length of the rods is
larger than the transverse size of the lattice. Therefore, a flow field up to sufficiently
large distances in front and behind the plane lattice does not depend on the length of
the rods and is described by the system of two-dimensional Navier–Stokes equations
with boundary conditions of adhesion on the surface of the lattice elements.
The determining dimensionless parameters of the problem are: M is Mach number
of the unperturbed oncoming flow, Re is Reynolds number based on cylinders diam-
eter d, γ is the ratio of gas heat capacities (adiabatic exponent), n is the number of
rods in the lattice, h/d is the relative period of the lattice. The geometric permeability
of the lattice is defined as σ = (h − d)/d, 0 < σ < 1. The characteristic mode of a
supersonic flow around a permeable screen is a flow with a smooth detached shock
wave in front of the screen at small and moderate values of σ [1, 6, 7]. We will refer
to such a flow as mode A. The head shock wave ceases to be common for the entire
lattice and transforms into a system of local shock waves in the vicinity of cylinders
[6] for sufficiently large σ and M. This flow will be referred to as mode B.
The task of flow around a single lattice element cannot be set without taking into
account the location of this element relative to the edge of the lattice in the case of
mode A. On the other hand, a flow along the most elements of the lattice is local

[email protected]
5 On Structures of Supersonic Flow … 51

in mode B, independent of the location of the element relative to the edge of the
lattice. This allows us to use a simplified formulation of the problem considering a
flow around any of the rods as a fragment of a periodic flow near the infinite system
of cylinders [5, 6]. We use both of the described formulations of the problem in this
chapter.
A uniform supersonic flow is specified before the grid at the input boundary of
the rectangular computational domain. The lateral boundaries of the computational
domain are taken at a sufficiently large distance from the edges of the lattice and
non-reflective boundary conditions are set on them when calculating the joint flow
around all the elements of the lattice (mode A). The lateral boundaries pass along the
lines of symmetry between the rods at a distance of lattice period h and conditions
for the periodicity of the flow are set on them when calculating in mode B.

5.3 Multiblock Computing Technology

System of overlapping grids is used for numerical simulation of the flow: one global
and many local ones (by the number of elements in the lattice).
Description of the far field of the flow is carried out on a uniform global grid with
rectangular cells. This grid, strictly speaking, does not allow describing physical
dissipative processes in the wake. The application of Navier–Stokes equations with
a limited number of nodes used in real calculations on the global grid cannot be
justified. Therefore, Euler equations are applied (the dissipative term is assumed to
be zero) on this grid in numerical simulation of the flow. The local grids are adapted to
the surface of the lattice elements and have an exponential refinement, which makes
it possible to adequately simulate a viscous near-wall flow near these elements. In
this chapter, Navier–Stokes equations are used in the thin-layer approximation, i.e.,
only the second derivatives along the normal to the surface of the body are taken into
account when calculating the dissipative term, as in the theory of the boundary layer.
The nonstationary Navier–Stokes equations in a thin-layer approximation for a
two-dimensional plane flow of compressible gas in a dimensionless vector form in a
curvilinear coordinate system ξ = ξ (x, y) and η = η(x, y) are as follows:

∂ U ∂ ξx E + ξy F ∂ ηx E + ηy F ∂ S
+ + = ,
∂τ J ∂ξ J ∂η J ∂η J
⎧ ⎫ ⎧ ⎫ ⎧ ⎫ ⎧ ⎫
⎪ ρ ⎪ ⎪ ρu ⎪ ⎪ ρv ⎪ ⎪
⎪ 0 ⎪
⎪
⎪
⎨ ⎪ ⎬ ⎪
⎨ 2 ⎪
⎬ ⎪
⎨ ⎪
⎬ ⎪ ∂u ⎪
ρu ρu + p ρuv μ ⎨ m1 ∂η + m2 ηx ⎬
U= , E= , F= , S= ∂v ,
⎪
⎪ ρv ⎪ ⎪ ρuv ⎪ ⎪ ρv2 + p ⎪ Re ⎪ m1 ∂η + m2 ηy ⎪
⎩ ⎪ ⎭ ⎪
⎩ ⎪
⎭ ⎪
⎩ ⎪
⎭ ⎪
⎪
⎩
⎪
⎪
⎭
e (e + p)u (e + p)v m3
1 ∂u ∂v
m1 = ηx2 + ηy2 , m2 = ηx + ηy ,
3 ∂η ∂η

[email protected]
52 S. V. Guvernyuk and F. A. Maksimov

γ 1 ∂T ∂ u2 + v 2
m3 = m1 + + m2 ηx u + ηy v .
γ − 1 Pr ∂η ∂η 2

Here, t is the time, ρ designates the density, (u, v) are the components of the
velocity vector V in the respective directions (x, y) of Cartesian coordinate system,
p is the pressure, and e denotes the total energy of the unit
 gas volume, which for a
1 p
perfect gas can be represented as e = ρ ε + 21 u2 + v2 , where ε = γ −1 ρ
is the
internal gas energy, and γ is the adiabatic exponent.
The dimensionless variables are defined through dimensional quantities, which
are indicated by Eq. 5.1.
 
po t  X ρo  ρ p T μ
t= X = V= V ρ = p = T = μ = (5.1)
ρo L L po ρo po To μo

The subscript o in Eq. 5.1 means the value of the parameter in an undisturbed
flow. Here, L is the characteristic dimension, X = (x, y), and V = (u, v).
μc
It is assumed that the Prandtl number Pr = λ p is the constant, cp is the heat
capacity ratio,√λ stands for the coefficient of heat conductivity, μ is the viscosity
p  ρ  L
ratio. Re = o o
μo
is Reynolds number. The system of differential equations is
supplemented by the equation of state P = ρRT , where T is the temperature and R
denotes the gas constant (p = ρT in dimensionless form).
The coefficients of the transformation matrix can be calculated by the following
∂y ∂x ∂y ∂x
formulas: ξx = J ∂η , ξy = −J ∂η , ηx = −J ∂ξ , and ηy = J ∂ξ . Here, J is the Jacobian
∂x ∂y ∂x ∂y
of the transformation, which is determined by the formula J −1 = ∂ξ ∂η
− ∂η ∂ξ
.
The use of a generalized transformation makes it possible to construct a uniform
grid in the form of a unit square. The coefficients of the transformation matrix for
a given distribution of nodes in the physical domain are calculated using the central
differences.
In deriving the reduced system of equations, it is assumed that the coordinate
lines, ξ = const, are oriented along the normal to the surface of the body, and the
derivatives of the direction η actually correspond to the derivatives along the local
normal to the body surface. Due to this, the second derivatives along the normal to
the body surface are taken into account when calculating the dissipative term.
Figure 5.1a presents an example of a fragment of an external computational grid
with uniform rectangular cells superimposed by a body domain and the contour of
the external boundary of the grid near the body. Figure 5.1b presents a magnified
fragment of the calculation area near the edge of the body with a grid. First, coordinate
lines ξ = const are constructed (the conformal mapping ensures the orthogonality
of these lines to the body contour), and second, the grid nodes are arranged along
these lines using exponential refinement taking into account the minimum distance
between the body node and the nearest node to the body and the distance from the
body to the outer boundary of the computational domain. On considering the grid, the

[email protected]
5 On Structures of Supersonic Flow … 53

Fig. 5.1 Construction of computational grid: a fragment of an external computational grid,

b magnified fragment of calculation area, c interpolation of values

outer grid is superimposed by a set of n bodies, near each of which the corresponding
curvilinear grid is constructed.
The solution is obtained by the relaxation method. An explicit second-order
approximation scheme [8] is used. The specific feature of calculations using multi-
block technology is integration with a common time step on the outer computational
grid, i.e., the minimum integration time step is chosen from the stability condition
over the entire computational domain. This is not a significant limitation because the
grid is uniform. At the same time, integration on the grids near the bodies involves
the use of the local time step, i.e., the choice of the time step at each grid nose is
determined by the local conditions. This results in a faster disturbance propagation
and, therefore, relaxation.
In order to tie together the solutions on the outer grid and the grids around the
bodies into a single whole, after completing the integration step, the values of the gas-
dynamic functions on the outer boundary L of the grid near the body are determined
by the interpolation from the solution obtained on the outer grid. Since a two-step
difference scheme [8] is used, a similar procedure is also performed for the node
layer in the vicinity of the boundary L. At the same time, the solution at all nodes
at the outer grid that happens to be inside the domain of the solution determination
near the body is replaced by the solution obtained on this grid [9].
When recalculating the values of gas-dynamic functions from one grid to another,
interpolation is used. Interpolation is implemented in the following form (in the
considered two-dimensional case). At first, we define the cell ABCD, in which point
O is located (Fig. 5.1c). Then values of the functions at point O are determined by
the values of the functions at nodes A, B, C, and D. The value of function f at node O
can be obtained from its values at all three nodes, for definiteness, let them be node
B, A, and D, by the interpolation formula:

fO = fA + α · (fB − fA ) + β · (fD − fA ),

|AO×AB| |AO×AD|
where α = |AD×AB| and β = |AB×AD| .

[email protected]
54 S. V. Guvernyuk and F. A. Maksimov

In order to take into account the value of the function at node C, we can similarly
express f O in terms of its values at a point of another triple of nodes, for instance, D,
C, and B. The final expression for f O is taken as an arithmetic average of the values
for the four chosen variants of the corner point. The interpolation coefficients are
determined for all nodes on the contour L and adjacent to it.
For nodes of the uniform grid that are inside the grids around the bodies, for exam-
ple, the point O that happens to be in the cell A B C  D (Fig. 5.1c), the same procedure
is performed, and the corresponding interpolation coefficients are determined.
In addition to the area, in which solutions are conjugated on the outer uniform
grid and grids near bodies, the necessary boundary conditions ought to be set at other
boundaries. For grids near bodies, this is the boundary corresponding to the surface
of the body, on which the no-slip condition and the given surface temperature are
set.
In calculation, the grids with the following numbers of cells are used: the local
grid near the bodies includes 180 × 40 cells, the outer grid around body system
involves 2000 × 2400 cells, and the outer grid in the periodic conditions about one
lattice element consists of 8000 × 800 cells.

5.4 Calculation Results for Lattice of 10 Elements

The calculations were performed for fixed γ = 1.4, Re = 105 , n = 10, and h/d = 5 (σ
= 80%) with variation of M in the range 2.0 ≤ M ≤ 4.5. The symmetry of the flow
around the lattice was assumed, and in fact only the upper half of the flow region
with five cylinders located in it was calculated.
Figures 5.2, 5.3, 5.4, and 5.5 show the flow patterns obtained according to the
scenario with a sequential increase in Mach number and the use of the solutions from
the previous step in the parameter M as the initial field. The visualization of the flow
fields is represented by the levels of the density gradient modulus. The coordinate
axes are normalized to the diameter of the elements of lattice d. When constructing
all the drawings, the same variant of the palette settings is used, therefore, at higher
Mach numbers, the drawings are darker. First, up to M = 2.4, there is a smooth
detached shock wave (Fig. 5.2) before the lattice as before a solid screen. The flow
between this shock and the lattice is subsonic, and the sound speed is reached (mode
A) in the local gap between the adjacent rods.
Then, when M = 2.5 is reached, a sudden qualitative change in the flow pat-
tern occurs: the smoothness of the bow shock wave is disturbed, numerous branch
points appear, and a characteristic configuration of Mach intersection of shock waves
is formed between the neighboring lattice elements (Fig. 5.3). The reflected shock
waves interact with the subsonic region of the near wake, as a result of which pressure
rises there and the local areas of the separated flow behind the cylinders increase sub-
stantially. A similar phenomenon was observed with increasing lattice permeability
[6]. This mode is characterized by the fact that all elements of the finite lattice, except
for the extreme ones, are flown around in almost the same way. As Mach number

[email protected]
5 On Structures of Supersonic Flow … 55

Fig. 5.2 Flow around lattice as M increases, mode A: a M = 2.0, b M = 2.4

Fig. 5.3 Flow around lattice as M increases, mode B: a M = 2.5, b M = 2.7

[email protected]
56 S. V. Guvernyuk and F. A. Maksimov

Fig. 5.4 Flow around lattice as M increases, mode B: a M = 2.8, b M = 3.5

Fig. 5.5 Flow around lattice as M increases, mode B: a M = 3.6, b M = 4.0

increases, the reflected shock waves shift downstream, and the length of the separa-
tion areas behind the cylinders increases. In this case, the wake behind the lattice as
a whole has a fairly regular appearance, which persists up to M = 2.7 (Fig. 5.3).

[email protected]
5 On Structures of Supersonic Flow … 57

However, when reaching the value M = 2.8 (Fig. 5.4), the flow with a stationary
separation area behind each cylinder and a regular wake, as in Fig. 5.3, collapses.
Apparently, with such an extended separation region behind the cylinders, the
stationary balance of mass and momentum fluxes at the boundaries of the separation
regions ceases to be ensured, which leads to periodic self-oscillations in the position
and size of these regions and to a significant change in the structure of the entire
flow (Fig. 5.4). The visualized traces behind the elements of the lattice look like
Karman vortex street, however, it should be borne in mind that this happens against
the background of the generally supersonic flow behind the lattice.
The unsteady flow pattern is qualitatively preserved for all internal elements of
the lattice, when 2.8 ≤ M ≤ 3.5. At the same time, a transition from Mach to a regular
branching diagram of the bow shock occurs (Fig. 5.4). Then, starting with M = 3.6,
abrupt transitions to patterns with independent flow around lattice elements begin to
occur. The separation region sharply decreased near one of the internal elements of
the lattice (the second element from the plane of symmetry in Fig. 5.5) and became
the same as near the extreme element when M = 3.6.
The same thing happened near all elements, except for the one closest to the plane
of symmetry (Fig. 5.5), when M = 4.0. Finally, when M = 4.5, all lattice elements
are being flowed around without mutual influence on each other. The local flow is the
same as near a single cylinder in a supersonic flow. Shock waves from neighboring
lattice elements fall only on the supersonic part of the wakes, which affects the flow
characteristics downstream, but does not change the local flow around the elements
themselves. A similar flow around the lattice took place, when M = 4.0, while a
uniform field of the incoming flow was taken as the initial field (Fig. 5.6).

Fig. 5.6 Flow around lattice as M increases, mode B: a M = 4.0, b M = 3.2

[email protected]
58 S. V. Guvernyuk and F. A. Maksimov

Fig. 5.7 Flow around lattice as M decreases: a mode B, M = 2.6, b mode A, M = 2.3

The following are examples of the results of a series of calculations performed

in a different scenario: with a sequential decrease in Mach number from 4.5 to 2.0
(Figs. 5.6 and 5.7). All the flow patterns described above were also observed in this
case, but with a certain shift of the boundaries of the transition from one flow pattern
to another. The flow pattern is implemented without the influence of neighboring
elements on the flow around each other (Fig. 5.6) in the range of 4.0 ≥ M ≥ 3.2.
When 3.1 ≥ M ≥ 2.6, a flow pattern with a nonstationary wake behind the cylinders
and a significant effect of their interference on the parameters of the local separated
flows in the bottom area behind the lattice elements (Fig. 5.7, M = 2.6) is observed.
When 2.5 ≥ M ≥ 2.3, the flow is stabilized in the bottom area behind the cylinders,
while the length of the region of the separated flow is abnormally large (Fig. 5.7, M
= 2.3) as compared to other modes. The bottom pressure is increased on the lattice
elements. The latter means a decrease in the aerodynamic drag. When 2.2 ≥ M ≥ 2.0,
the flow around the grid returns to mode A (as in Fig. 5.2).
It should be stressed that a transition from one flow pattern around the lattice to
another occurs in a threshold manner in both considered scenarios of Mach number
changes. The transition takes place so that there are ranges of values of M, for which
different flow patterns can exist. The selection of an actual flow pattern is determined
by the pre-history of the change in Mach number.
The implemented flow patterns affect the aerodynamic drag F x of the grid. Accord-
ing to the calculation results, an estimate was obtained for the dependence of the
aerodynamic drag coefficient cx = 2F x /γ p∞ M 2 S x (p∞ is the static pressure in the
undisturbed flow and S x designates the drag area) on Mach number, Fig. 5.8. The

[email protected]
5 On Structures of Supersonic Flow … 59

Fig. 5.8 Dependence of

aerodynamic drag coefficient
of lattice element on Mach
number, curve 1—cx = f (M)
with consistent increase in
M, curve 2—with decrease in
M

unsteadiness of the wake is weakly reflected in the oscillations of cx (maximum

deviation of the average value is not more 6%).
In accordance with the four qualitatively different flow patterns listed above,
four levels of drag coefficient values are realized, the transitions between which are
characterized by three double-valued solution intervals.

5.5 Calculation Results for Infinite Lattice with Periodical

Conditions

As it was noted above, flows on mode B can be investigated under a simplified for-
mulation of the problem, considering one lattice element with periodicity conditions
on the side boundaries. This allows one to significantly reduce the computational
domain and increase the level of the solution refinement. On the other hand, the
correctness of the periodicity conditions in the case of nonstationary flow patterns
is not obvious, since the periodicity condition may impose restrictions on the nature
of the disturbance propagation in the flow region. Figure 5.9 shows the calculation
results for a single lattice element with periodicity conditions on the side boundaries.
The oncoming supersonic flow is directed from left to right in each of the presented
flow patterns. The distribution of the density gradient modulus is presented. The upper
and lower boundaries of the computational domain correspond to the period of the
infinite lattice. Transitions between the same flow patterns that were identified above
in the case of a lattice with a finite number of elements for mode B are observed.
Figure 5.10 shows a summary graph of the dependence of the drag coefficient
cx of a single lattice element. Line 1 corresponds to the scenario with a sequential
increase in M and line 2 indicates the scenario with a decrease in M for an infinite
lattice. The data of Fig. 5.8 obtained for a lattice of 10 elements (lines 1 and 2) are
shown in the same place in Fig. 5.10.
The calculation results using a finite-size lattice model and a simplified one for
a single lattice element with periodicity conditions do not qualitatively differ, but
there are some differences in estimating the hysteresis boundaries and the value of
aerodynamic drag in the mode with the formation of Karman vortex street. This may
be due to a virtually different resolution of the computational grid in different models.

[email protected]
60 S. V. Guvernyuk and F. A. Maksimov

Fig. 5.9 Evolution of flow structures around lattice element with continuous decrease and increase
in M

Fig. 5.10 Aerodynamics

drag coefficient of lattice
element with successive
increase (line 1, 1 ) and
decrease (line 2, 2 ) in Mach
number M

[email protected]
5 On Structures of Supersonic Flow … 61

However, it can also be associated with the presence of an extreme element in a

finite-size lattice, which affects indirectly all the lattice elements, which is not taken
into account in the simplified formulation of the problem with periodic boundary
conditions on the side boundaries.

5.6 Conclusions

The multiblock computing technology is used to calculate supersonic flow around a

system of cylindrical rods that form a plane lattice of finite width at various Mach
numbers and various scenarios of its change. Four flow patterns were revealed, the
new of which was a pattern with nonstationary periodic self-oscillations of the flow
in the near wake behind the lattice elements. The implemented flow patterns and the
boundaries of transition from one flow pattern to another depend on the pre-history
of the change in Mach number.
Using an example of the flow around the finite lattice of cylindrical rods with a
permeability of 80%, three ranges of flow ambiguity and the corresponding hysteresis
of Mach number characteristics were revealed. The first type is associated with
restructuring between the collective and local flow regimes of the lattice elements.
The second type is conditioned by the restructuring of the near wake behind the
cylindrical elements as a result of interaction with local shock-wave systems from
neighboring lattice elements. The third type of hysteresis is associated with the
occurrence of a nonstationary periodic flow in the wake of the lattice elements.

Acknowledgements The work was carried out with a partial financial support of the Russian
Foundation for Basic Research (Project No. 19-01-00242) and partially under the state task of the
ICAD RAS.
The calculations were carried out on MVS-100 K at Interdepartmental supercomputer center of
the Russian Academy of Sciences.

References

1. Guvernyuk, S.V.: On hypersonic flow around bodies with wire screens. Gas Wave Dyn. 4,
236–242 (in Russian) (2005)
2. Fomin, V.M., Mironov, S.G., Serdyuk, K.M.: Reducing the wave drag of bodies in supersonic
flows using porous materials. Tech. Phys. Lett. 35(2), 117–119 (2009)
3. Kirilovskiy, S.V., Maslov, A.A., Mironov, S.G., Poplavskaya, T.V.: Application of the skeleton
model of a highly porous cellular material in modeling supersonic flow past a cylinder with a
forward gas-permeable insert. Fluid Dyn. 53(3), 409–416 (2018)
4. Maksimov, F.A.: Supersonic body wrapping. Comput. Res. Model. 5(6), 969–980 (in Russian)
(2013)
5. Kudryavsev, A.N., Epshtein, D.B.: Hysteresis phenomenon in supersonic flow past a system of
cylinders. Fluid Dyn. 47(3), 395–402 (2012)

[email protected]
62 S. V. Guvernyuk and F. A. Maksimov

6. Guvernyuk, S.V., Maksimov, F.A.: Supersonic flow past a flat lattice of cylindrical rods. Comput.
Math. Math. Phys. 56(6), 1012–1019 (2016)
7. Guvernyuk, S.V., Savinov, K.G., Ul’yanov, G.S.: Supersonic flow round blunt perforated screens.
Fluid Dyn. 20(1), 124–129 (1985)
8. Maksimov, F.A., Churakov, D.A., Shevelev, Y.D.: Development of mathematical models and
numerical methods for aerodynamic design on multiprocessor computers. Comput. Math. Math.
Phys. 51(2), 284–307 (2011)
9. Isaev, S.A., Baranov, P.A., Usachov, A.E.: Multiblock Computing Technologies in VP2/3
Aerothermodynamics Package. LAP LAMBERT Academic Publishing, Saarbryuken (in Rus-
sian) (2013)

[email protected]
Chapter 6
Limiting Functions Affecting
the Accuracy of Numerical Solution
Obtained by Discontinuous Galerkin
Method

Marina E. Ladonkina , Olga A. Nekliudova and Vladimir F. Tishkin

Abstract In the numerical solution of hyperbolic systems of equations, Galerkin

method with discontinuous basic functions is proved to be very reliable. However, to
ensure the monotony of the solution obtained by this method, it is necessary to use a
smoothing operator, especially if the solution contains strong discontinuities. In this
chapter, we consider the classic Cockburn limiter, a moment limiter that preserves
the high order of the scheme, well-proven smoothing operator based on Weighted
Essentially Non-Oscillatory (WENO) reconstruction, the smoothing operator of a
new type based on averaging solutions, taking into account the rate of change of the
solution and the rate of change of its derivatives and slope limiter, preserving the
positivity of pressure. A comparison was made of the actions of these limiters on a
series of test problems. Numerical results show that using discontinuous Galerkin
method and applying moment limiter, slope limiter, WENO limiter, or limiter based
on averaging allows to obtain a high order of accuracy on smooth solutions, as well as
the clear, non-oscillating profiles on shock waves provided with appropriate constants
for the correct determinations of defective cells. In addition, slope limiter, WENO
limiter, and averaging limiter are simple enough to implement and be generalized on
the multidimensional unstructured grids.

M. E. Ladonkina (B) · O. A. Nekliudova · V. F. Tishkin

Keldysh Institute of Applied Mathematics of RAS, 4, Miusskaya Square, Moscow 125047,
Russian Federation
e-mail: [email protected]
O. A. Nekliudova
e-mail: [email protected]
V. F. Tishkin
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 63

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_6

[email protected]
64 M. E. Ladonkina et al.

6.1 Introduction

In the last two decades, Galerkin method with discontinuous basic functions has
been actively developed. This method was proposed more than fifty years ago to
solve the neutron transport equation [1]. Nowadays, an enormous amount of pub-
lications appears annually on the study and application of Discontinuous Galerkin
(DG) method for solving a wide class of applied problems.
Galerkin method with discontinuous basic functions is one of the numerical meth-
ods of high accuracy. This method has a number of advantages inherent in both finite-
element and finite-difference approximations. Notably, it provides a given order of
accuracy even on unstructured meshes [2] and can be used for meshes with an arbi-
trary cell shapes. This is especially relevant for solving complex multi-scale tasks.
It is well known that there are two approaches to improve the accuracy of the result-
ing solution. One of them is to grind up the grid in the areas of the existing fea-
tures of the solution, while the second approach is to increase the order of accuracy
of the scheme. The discontinuous Galerkin method allows one to implement both
approaches at once. Thus, the order of accuracy using the high-order polynomials is
increased making a local grid refinement (the so-called hp-adaptation) [3, 4] at the
same time. This is especially relevant for solving complex multi-scale problems.
The choice of a grid is one of the important issues of method implementation.
The undoubted advantage of discontinuous Galerkin method is a possibility of its
application on the grids of arbitrary structure. There are successful software imple-
mentations of Rectangular Mesh Generator (RMG) for solving three-dimensional
problems on unstructured grids containing as elements of only one type tetrahedral
[5, 6] or hexahedral [4, 7], as well as, for grids with different types of cells [8].
However, DG method has some implementation difficulties. First of all, to ensure
the monotony of solution obtained by this method, it is necessary to introduce so-
called slope limiters, especially if the solution contains strong discontinuities. Cock-
burn limiter [2] is the most widely used one on the tetrahedral grids. The concept
of this limiter can be implemented in the multidimensional problems with arbitrary
structured grids. However, this limiter, as all Total Variation Diminishing (TVD)
limiters, reduces the accuracy of the resulting solution [9].
Recently, various approaches have been developed to solve this problem. One of
them is to create a limiter of a higher order of accuracy. It was proposed in [10].
However, this limiter works well only on the structured grids. Another approach of
creating a limiter with a higher order of accuracy consists of using WENO limiter
[11]. In [12], the limiters were proposed that do not resort to the use of min mod
procedures and, accordingly, do not reduce the accuracy of the solution, which is a
great advantage of these limiters. In [13], a smoothing operator of a new type was
developed based on averaging of solutions, taking into account the rate of change of
the solution and the rate of change of its derivatives. It was shown that the application
of the proposed smoothing operator is not inferior to WENO limiter, and in some
cases exceeds the accuracy of the obtained solution, which is confirmed by numerical
studies.

[email protected]
6 Limiting Functions Affecting the Accuracy of Numerical Solution … 65

Another fundamentally different way of monotonizing a solution is based on intro-

ducing the artificial viscosity into a numerical scheme [14]. This approach involves
a use of empirical constants, which complicates its application to solve real prob-
lems. In [12], a modification of DG method for two-dimensional case was proposed
providing the possibility of a smooth transition from a high-order accuracy scheme
to a first-order monotonic scheme in flow singularity regions. The most simple and
effective implementation of the limiter was proposed in [8], but this limiter has its
flaws, since its implementation does not guarantee the suppression of nonphysical
oscillations.
Nowadays, a number of papers have shown that the use of limiters can adversely
affect the accuracy of the numerical solution [9, 15]. Therefore, investigations of
preserving the order of accuracy of the solution and ensuring its monotony remain
relevant.
In this chapter, we consider a model approach for DG method implementation.
Approximate solution is defined as the projection of a vector of unknown variables
onto a space of polynomials of degree p with time-dependent coefficients and coeffi-
cients of polynomials as unknowns. Also, we explore the limiting functions that are
most often used in this approach for real calculations.
In Sect. 6.2, we provide a description of DG method for Euler equations, as well
a description of tested limiters. Section 6.3 presents the test results and research on
solution of Einfeldt problem. Section 6.4 contains the conclusions for this chapter.

6.2 Discontinuous Galerkin Method for Euler Equations

We study Euler equation written in a conservative form:

∂t U + ∇ · F(U) = 0. (6.1)

It is supplemented with suitable initial-boundary conditions, the form of which

depends on the specific problem and will be specified later. The conservative variables
U and the components of the stream function F(U) are given as
⎧ ⎫ ⎧ ⎫
⎨ ρ ⎬ ⎨ ρu ⎬
U = ρ u , Fx (U) = ρ u 2 + p , (6.2)
⎩ ⎭ ⎩ ⎭
E (E + p)u

where ρ is the density of the fluid, u is the velocity, p is the pressure, ε is the
specific internal energy, and E = ρ ε + u 2 /2 is the total energy per volume unit.
To determine the pressure p, we use the ideal gas state equation:

p = (γ − 1)ρε,

[email protected]
66 M. E. Ladonkina et al.

where γ is the adiabatic index.

To apply the discontinuous Galerkin method, we cut the area, where the solution
is sought with a grid 0 = x1/2 ≤ x3/2 ≤ · · · ≤ x N +1/2 = L with grid spacing
xi = (xi+1/2 − xi−1/2 ). At each interval xi−1/2 ≤ x ≤ xi+1/2 , we search an
approximate solution of the system of equations (Eq. 6.1) in the form of a projection
of a vector of conservative variables U = (ρ, ρu, E) onto the space of polynomials
P(x) of degree p in the basis {φk (x)} with coefficients depending on time. Then, the
solutions will be
p
Uh (x, t) = Uk (t)φk (x),
k=0

where p is the degree of polynomials, and φk (x) is the corresponding basic function.
In this chapter, we will use Taylor basis.
An approximate solution of the system (Eq. 6.1) in the discontinuous Galerkin
method is sought as a solution of the following system [2]:

∂t Uh (x, t) · φk (x)dx − F(Uh (x, t))∂x φk (x)dx+

Ii Ii (6.3)
+Fi+1/2 φk (xi+1/2
l
) − Fi−1/2 φk (xi−1/2
r
) = 0,

where i = 0, …, N, k = 0, 1, 2.
In Eq. 6.3, Uh (x, t) = (ρh (x, t)ρu h (x, t)E h (x, t))T is the solution vector;
φk (xi+1/2
l
), φk (xi−1/2
r
) are the basis functions with the number k on the interval
I i calculated in points xi+1/2 ,xi−1/2 ; and Fi+1/2 , Fi−1/2 are the discrete flows, which
are monotonic functions of two variables:

Fi+1/2 = (Uh (xi+1/2

l
, t), Uh (xi+1/2
r
, t)),
Fi−1/2 = (Uh (xi−1/2
l
, t), Uh (xi−1/2
r
, t)),

for which the condition of approval is satisfied:

(Uh (xi , t), Uh (xi , t)) = F(Uh (xi , t)).

In this chapter, Rusanov–Lax–Friedrichs flows [16, 17] and Godunov flow [18]
are used as a numerical flux.
Hereinafter, Cockburn limiter, moment limiter, limiter based on WENO recon-
struction, limiter based on averaging the solution, and slope limiter are considered
in Sects. 6.2.1–6.2.5, respectively.

[email protected]
6 Limiting Functions Affecting the Accuracy of Numerical Solution … 67

6.2.1 Cockburn Limiter

A limiter is a certain operator acting on an approximate solution function on each

interval xi−1/2 , xi+1/2 . According to [2], we denote the action of this operator on the
function u by  h u.
Cockburn limiter is described in detail in [2]. For a linear function u = u i0 + u i1 φ1 ,
the action of the limiting operator can be written as

 h u = u i0 + ũ i1 φ1 , (6.4)

  
ũ i1 = min mod u i1 , α u i+1
0 − u i0 , α u i0 − u i−1
0 , (6.5)

min mod (a1 , . . . , a N )

sign(a1 ) min a j if sign(a1 ) = sign(a2 ) = · · · = sign(a N )
= 1≤ j≤N . (6.6)
0 otherwise

To use Cockburn limiter, in the case when the order of the polynomial is p > 1, we
project the function u onto the space of linear polynomials u l . Further, after applying
the limiter (Eqs. 6.4–6.6) to the linear function u l , the coefficients at the linear terms
ũ l1 and u l1 are compared, and, if they are equal, the original function is not being
changed after the action of the limiter. Otherwise, the result of the limiter will be a
linear function. When applying this limiter, the choice of the parameter 1 ≤ α ≤ 2
plays an important role. When α = 1, we get the most “hard” limiter ensuring the
monotony of the solution, when α = 2, we get the “less strict” limiter  h .

6.2.2 Moment Limiter

The next type of limiter investigated in this chapter is the “momentum” limiter
described in [10]. This limiter is characterized by the fact that it preserves the highest
possible order of the scheme.
The main idea of the method is that the solution is limited by limiting its coef-
ficients. The coefficient ũ ik corresponds to the kth derivative of the solution, and it
is compared with the alternative approximation of the kth derivative in terms of the
right and left differences of the (k−1)th derivative.
Starting with the highest coefficients k = p, we replace ũ ik in Eq. 6.7.
   i
ũ˜ ik = min mod ũ ik , αk ũ i+1
k−1 − ũ k−1 , αk ũ k−1 − ũ k−1 .
i i−1
(6.7)

The limiter is triggered, when ũ˜ ik = ũ ik . In the case of ũ˜ ik = ũ ik , limiting is

terminated, otherwise the coefficient ũ ik−1 is limited continuing as long as either k =

[email protected]
68 M. E. Ladonkina et al.

1 or the condition ũ˜ ik = ũ ik is fulfilled. In the limiter (Eq. 6.7), there is a parameter
αk , which value depends on the order of the coefficient k. In [17], it was shown that
the range of parameter αk variation is bounded below by a number 2(2k−1) 1
. It should
be noted that in the case of p = 1, the limiter completely coincides with the min mod
limiter.

6.2.3 Limiter Based on WENO Reconstruction

In [11], a limiter was proposed for the discontinuous Galerkin method based on
WENO reconstruction, which allows to preserve the high accuracy of the method
and does not distort the solution profile. At the first stage, the problem cells should
be identified, i.e., those cells in which limiting may be required. At the next stage, the
numerical solution in the problem cells is replaced with the reconstructed one, with
the polynomials obtained during the reconstruction retaining the original integral
average value in the cell and a high order of accuracy, but less prone to oscillations.
To identify the problem cells, we will use Total Variation Bounded (TVB) min mod
[11] limiter, where the min mod function is defined by Eq. 6.6 or through the function
converted by TVB min mod (Eq. 6.8):

a1 if |a1 | ≤ Mh 2
m̃(a1 , . . . , a N ) = , (6.8)
min mod(a1 , . . . , a N ) otherwise

where parameter M is chosen according to the solution of the problem.

The main idea of WENO limiter is that a new polynomial is built on the problem
cell, which is a convex combination of the original polynomial and polynomials on
neighboring cells with the necessary corrections to preserve the integral average in
the cell. In the calculations presented below, we used the coefficients indicated in
[11].

6.2.4 Limiter Based on Averaging the Solution

In [13], a different approach for limiter building was proposed. Let us consider a new
polynomial:
 
P jsm = P̄ j + P j0 1 − max α j+1/2 , α j−1/2 , P j0 = P j − P̄ j .

The criterion for selecting coefficients α j+1/2 , α j−1/2 is thoroughly described in

[13]. First, consider the order of deviation of neighboring polynomials from the
arithmetic mean of their integral means. As one of the coefficients, we will choose:

[email protected]
6 Limiting Functions Affecting the Accuracy of Numerical Solution … 69
x 
x j+3/2
 j+3/2  2 x j+3/2  2
2 av av
μ j+1/2 = P j − P j+1 dx P j − P j+1/2 dx + P j+1 − P j+1/2 dx ,
x j−1/2 x j−1/2 x j−1/2

av
P j+1/2 = 21 P̄ j + P̄ j+1 .

This coefficient does not reduce the order of the polynomial obtained.
Note that in the discontinuous Galerkin method, we can determine the coefficient
β j+1/2 in terms of the rate of change of the solution and its derivatives. We make the
coefficient:
c j+1/2 · t
α j+1/2 = μ j+1/2 β j+1/2 ,
x
proportional to the time step, but not violating the dimension. There is a multi-
1
plier x in the coefficient, which theoretically can reduce the accuracy of the solu-
tion, so replacing it with a multiplier ρρ̄1 , where ρ h (x) is the density, ρh (x) =
p
k=0 ρk (x)φk (x), ρ̄ is the integral average, and ρ1 the first derivative of ρh (x)
theoretically allows us to maintain the order of accuracy of the scheme:

c j+1/2 t
α j+1/2 = μ j+1/2 β j+1/2 .
ρ/ρ1

6.2.5 Slope Limiter

We describe a design of slope limiter according to [8]. To limit the polynomials in

the cell with the number K, it is necessary to go through all the neighboring cells
and calculate the integral average pressure value in each of them p̄i , i = 1, N ,
where N is the number of neighboring cells and choose the maximum and minimum
pressure values pmax = max( p̄i ), pmin = min( p̄i ), i = 1, N . Let us denote p + =
(1+ε) pmax , p − = (1−ε) pmin , where ε is some small positive constant. We calculate
the pressure in all quadrature points of the limited cell. Find the maximum and
K K
minimum values pmax , pmin . If at any quadrature point the pressure value exceeds
+
the value p , we multiply the original polynomials by some positive value α, so that
the pressure value at this quadrature point does not exceed p + . Similarly, we proceed
with the condition on p − . To get limited pressure values p̂ = (1 − α) p̄ + αp, we
recalculate the coefficients of all the original polynomials:

Û = (1 − α)Ū + αU = (1 − α)Ū + α Ujϕj,

j=0, p

where U = (ρ ρu E)T is the vector of the original polynomials, and Ū is the vector
of integral means of conservative variables. In addition, the same procedure is done
separately for density.

[email protected]
70 M. E. Ladonkina et al.

6.3 Numerical Experiments

Let us consider a series of test problems of one-dimensional unsteady gas dynamics.

Despite the simplicity of the statement, these problems reflect all the features of
gas-dynamic flows. The initial distribution of density, velocity, and pressure values
are presented in Table 6.1.
Test 1 (Sod problem). As a result of this test, a shock wave arises moving into the
low-pressure region and a rarefaction wave appears expanding into the high-pressure
region and a contact discontinuity (Fig. 6.1a).
Test 2 (Lax problem). In this problem, the same configuration of the solution
arises as in Sod problem, consisting of a shock wave and contact discontinuity with
large differences in gas-dynamic parameters than in the previous problem but, unlike
Sod, a less intense rarefaction (Fig. 6.1b).
Test 3 (supersonic shock tube). The configuration of the solution of this problem
is similar to the two previous ones. The shock wave, rarefaction wave, and contact
discontinuity arise. However, a solution of this problem allows us to estimate the
operation of computational schemes in the emerging regions of supersonic flows
(Fig. 6.1c).
Test 4 (Mach 3 problem). The solution is a contact discontinuity and two
rarefaction waves (Fig. 6.1d).
Test 5 (Einfeldt problem). As a result of this solution, two symmetric rarefaction
waves arise propagating in the opposite directions and fixed contact discontinuity
(Fig. 6.1).
Test 6 (shock entropy wave interaction) is the interaction of a shock wave with an
entropic perturbation. Mach number of the shock wave moving along the X-axis is M
= 3.5. After the passage of the shock wave behind the front, a complex flow is formed,
in which a series of smaller amplitude shock waves forms over time (Fig. 6.1e).
Test 7 (Woodward–Colella blast waves). This problem is a model of the interaction
of two shock waves and is one of the generally accepted tests for testing the operability
of numerical methods for solving gas dynamics problems. At the initial moment of
time, the density ρ = 1, velocity u = 0, the pressure is distributed as follows: p =

Table 6.1 Initial distribution of density, velocity, and pressure

No. Parameters to the left Parameters to the right Time
Density Velocity Pressure Density Velocity Pressure
1 1 0 1 0.125 0 0.1 2.0
2 0.445 0.698 3.528 0.5 0 0.571 1.3
3 1 0 1 0.02 0 0.02 0.15
4 3857 0.920 10.333 1 3.55 1 0.09
5 1 −2 0.4 1 2 0.4 0.15
6 3.857143 2.629369 10.3333 1 + 0.2sin(5x) 0 1 1.8

[email protected]
6 Limiting Functions Affecting the Accuracy of Numerical Solution … 71

Fig. 6.1 Density: a Test 1, b Test 2, c Test 3, d Test 4, e Test 6, and f Test 7

103, at 0 ≤ x ≤ 0.1, p = 102, at 0.1 ≤ x ≤ 0.9, and p = 102, at 0.9 ≤ x ≤ 1.

Estimated time t = 0.038 (Fig. 6.1f).
In Tests 1, 2, and 6, the computational domain is −5 ≤ x ≤ 5, while in Tests
3, 4, and 5, the computational domain is 0 ≤ x ≤ 1. The positions of the point of
discontinuity are −x0 = 0, −x0 = 0.5, and x0 = −4 in Tests 12, Tests 3–5, and Test

[email protected]
72 M. E. Ladonkina et al.

6, respectively. Calculations of Tests 1–5, Test 6, and Test 7 were carried out on the
grids of 100 cells, 200 cells, and 400 cells, respectively. In all calculations, gas is
assumed to be ideal, with an adiabatic exponent γ = 1, 4.
Note that for each limiter there are configurable parameters. We carry out a series
of calculations with the Cockburn limiter, moment limiter, and slope limiter with
parameters corresponding to weak limiting. Let us determine the optimal parameters
for WENO and Average Solution Meaning (AvSM) limiters. To identify defective
cells, you ought to define the parameter M in TVB min mod limiter. The first series
of calculations was carried out with the parameter M = 0.01. This parameter with
an excess determines the cells, in which it is necessary to carry out limiting. The
next step was to determine the critical value of the parameter M so that its further
increase would not affect the quality of the limitation that is, the calculation would be
performed without a limiter. The next series of calculations consisted in determining
the critical value of the parameter M. Thus, with an increase of the parameter M in
all tests, more distinct profiles are observed in the areas of solution discontinuities,
but at the same time, the appearance of oscillations is observed.
We present the results of calculations obtained with optimal parameters.
When solving problems 1–5, it is impossible to unambiguously determine the
advantages of using one or another limiter. It should be noted that the slope limiter
(green dots, Fig. 6.1) shows the most stable behavior. Although oscillations are
present in Test 3, it is easy to suppress them by decreasing the epsilon parameter.
Tests 4 and 5 turned out to be the most difficult for a numerical solution using WENO
(red dots, Fig. 6.1) and AvSM (black rounds, Fig. 6.1) limiters. Both represent the
decay of a gas-dynamic discontinuity in the form of a contact discontinuity and
two rarefaction waves. When solving Test 4, the best results were obtained using
Cockburn limiter (black crosses, Fig. 6.1), moment limiter (blue dots, Fig. 6.1), and
slope limiter (green dots Fig. 6.1). However, slope limiter did not cope well with the
solution of Test 5. For problem 7, errors were calculated in L2 norm with respect to the
“reference” solution from [19]. Errors in L2 norm have the following values: 2.09d-5,
1.04d-5, 1.25d-5, 3.94d-5, and 4.09d-5 for Cockburn limiter, moment limiter, slope
limiter, WENO limiter, and AvSM limiter, respectively.
When solving more complex problems 6 and 7, the use of limiters based on
averaging gives the best result with a coefficient k ∼ x. In [19], it was shown that
in some cases, the coefficient k ∼ ρ1 /ρ̄ when using quadratic functions makes the
limiter more dissipative.
It is worthwhile to dwell separately on solving problem 5. As is known, this
problem is often used when testing numerical methods, and one of the indicators
of a well-functioning scheme is the accuracy of the transfer of the contact discon-
tinuity region. In almost all calculations, a nonphysical surge of internal energy
is observed. We considered several numerical schemes for solving this problem.
We present a series of calculations performed according to the first-order Godunov
scheme (Fig. 6.2b) and the third-order discontinuous Galerkin method with quadratic
basis functions using moment limiter [10] with coefficients corresponding to “hard”
limiting (Fig. 6.2a).

[email protected]
6 Limiting Functions Affecting the Accuracy of Numerical Solution … 73

Fig. 6.2 Internal energy. Test 5. 500 points: a DG, P = 2, Godunov flux, moment limiter; b Godunov
scheme, P = 0

Adding the pseudo-heat conductivity to the numerical approximation of the flow

(Eq. 6.5), we obtain a clear decrease in the entropy release (Fig. 6.3).
Now let us pay attention to the definition of the initial data. In this problem, the
break point x0 = 0.5 falls on the boundary between the cells. In the region to the
left of the discontinuity point x ≤ x0 , the velocity is u = –0.2, and in the region on
the right, u = 0.2. We choose a grid so that the point x0 = 0.5 falls inside the cell
approximating the speed in this cell. Then we get u = 0. In Fig. 6.4, it is clearly seen
that there is practically no entropy surge.
Doing the same thing and by approximating the velocity gap in two cells, it is
possible to obtain a solution corresponding to the physical one. Thus, it was estab-
lished that the defect arising in the numerical solution of problem 5 is connected not
only with the choice of the numerical scheme but also with the method of specifying
the initial data.
Figure 6.5 shows the results of calculations of problem 5 with a corrected initial

Fig. 6.3 Internal energy including pseudo-heat conductivity. Test 5: a moment limiter, P = 2;
b Godunov scheme, P = 0

[email protected]
74 M. E. Ladonkina et al.

Fig. 6.4 Internal energy. Test 5. 501 points: a moment limiter, P = 2; b Godunov scheme, P = 0

Fig. 6.5 Internal energy. Test 5. 50,000 points: a Godunov scheme P = 0 (black dots), moment
limiter, P = 2 (white circles); b WENO limiter (white circles), AvSM limiter (solid line)

velocity profile. As can be seen from the graphs, we managed to improve the quality
of the result.
Note that in this case, the addition to the numerical scheme of the coefficient
corresponding to the pseudo-thermal conductivity leads to an additional smoothing
of the solution. The results of Test 5 without pseudo-heat conductivity and with
pseudo-heat conductivity are depicted in Fig. 6.6.
According to the results of the study, it is worth noting that all the limiters under
study showed themselves quite well in solving complex problems of one-dimensional
unsteady gas dynamics. The most stable solution is obtained when using Cockburn
limiter. This limiter can be implemented on unstructured grids with cells of various
shapes, but the use of this limiter does not guarantee the declared accuracy of the
discontinuous Galerkin method. In contrast to Cockburn limiter, moment limiter
retains the increased accuracy of the method, but is currently designed only for
structured grids. The most promising in our opinion are slope limiter, WENO limiter,

[email protected]
6 Limiting Functions Affecting the Accuracy of Numerical Solution … 75

Fig. 6.6 Internal energy. Test 5. 500 points: a without pseudo-heat conductivity, and b with pseudo-
heat conductivity

and limiter based on averaging, which are fairly simple to implement and generalize
to multidimensional unstructured grids.

6.4 Conclusions

Numerical results show that when solving problems using the discontinuous Galerkin
method, the application of moment limiter, slope limiter, WENO limiters, and limiter
based on averaging allows to obtain a high order of accuracy on smooth solutions, as
well as, the clear, non-oscillating profiles on shock waves provided with the appro-
priate constants for the correct definition of defective cells. In addition, slope limiter,
WENO limiter, and averaging limiter are simple enough to implement and generalize
to multidimensional unstructured grids.

References

1. Reed, W., Hill, T.: Triangular mesh methods for the neutron transport equation. Los Alamos,
Scientific Laboratory Report LA-UR-73-79, United States (1973)
2. Cockburn, B.: An introduction to the discontinuous Galerkin method for convection—dom-
inated problems. In: Quarteroni, A. (ed.) Advanced Numerical Approximation of Nonlinear
Hyperbolic Equations, LNM, vol. 1697, pp. 150–268 (1998)
3. Nastase, C., Mavriplis, D.: High-order discontinuous Galerkin methods using an hp-multigrid
approach. Comput. Phys. 213, 330–357 (2006)
4. Volkov, A.: Features of the use of the Galerkin method for the solution of the spatial Navier-
Stokes an-structured hexahedral equations. TsAGI Sci. J. XL(6), 695–718 (2009)
5. Bassi, F., Rebay, S.: Numerical evaluation of two discontinuous Galerkin methods for the
compressible Navier-Stokes equations. Int. J. Numer. Meth. Fluids 40, 197–207 (2002)

[email protected]
76 M. E. Ladonkina et al.

6. Krasnov, M., Kuchugov, P., Ladonkina, M., Lutsky, A., Tishkin, V.: Numerical solution of
the Navier-Stokes equations by discontinuous Galerkin method. J. Phys. Conf. Ser. 815(1),
012015.1–012015.9 (2017)
7. Bosnyakov, S., Mikhailov, S., Podaruev, V., Troshin, A.: Unsteady high order accuracy DG
method for turbulent flow modeling. Math. Model. 30(5), 37–56 (2018)
8. Yasue, K., Furudate, M., Ohnishi, N., Sawada, K.: Implicit discontinuous Galerkin method
for RANS simulation utilizing pointwise relaxation algorithm. Commun. Comput. Phys. 7(3),
510–533 (2010)
9. Ladonkina, M., Nekliudova, O., Ostapenko, V., Tishkin, V.: On the accuracy of the discontin-
uous Galerkin method in the calculation of shock waves. Comput. Math. Math. Phys. 58(8),
1344–1353 (2018)
10. Krivodonova, L.: Limiters for high-order discontinuous Galerkin methods. J. Comput. Phys.
226(1), 276–296 (2007)
11. Zhong, X., Shu, C.-W.: A simple weighted essentially non oscillatory limiter for Runge-Kutta
discontinuous Galerkin methods. J. Comput. Phys. 232, 397–415 (2013)
12. Volkov, A., Lyapunov, S.: Monotonization of the finite element method in gas dynamics
problems. TsAGI Sci. J. XL(4), 15–27 (2009)
13. Ladonkina, M., Nekliudova, O., Tishkin, V.: Construction of the limiter based on averaging of
solutions for discontinuous Galerkin method. Math. Model. 30(5), 99–116 (2018)
14. Persson, P.-O., Peraire, J.: Sub-cell shock capturing for discontinuous Galerkin method. AIAA-
2006-112.1–AIAA-2006-112.13 (2012)
15. Kovyrkina, O., Ostapenko, V.: On the practical accuracy of shock-capturing schemes. Math.
Model. 25(9), 63–74 (2013)
16. Rusanov, V.: Calculation of the interaction of non-stationary shock waves with obstacles. USSR
Comput. Math. Math. Phys. 2, 267–279 (1961)
17. Lax, P.: Weak solutions of nonlinear hyperbolic equations and their numerical computation.
Commun. Pure Appl. Math. 7(1), 159–193 (1954)
18. Godunov, S.: Difference method for the numerical calculation of discontinuous solutions of
the equations of hydrodynamics. Math. Notes 47(89)(3), 271–306 (in Russian) (1959)
19. Ladonkina, M., Neklyudova, O., Tishkin, V.: Application of averaging to smooth the solution
in DG method. Preprint KIAM 89, Moscow (in Russian) (2017)

[email protected]
 Part II
Numerical Simulation of Multiphase
Flows, Combustion, and Detonation

[email protected]
Chapter 7
Numerical Simulation of Detonation
Initiation: The Quest of Grid Resolution

Alexander I. Lopato , Artem G. Eremenko , Pavel S. Utkin

and Dmitry A. Gavrilov

Abstract The chapter is dedicated to the numerical investigation of the grid reso-
lution influence on the detonation initiation process in the multifocused system with
the profiled end-wall. Two-dimensional system of Euler equations coupled with the
single-step Arrhenius kinetic reaction mechanism was solved on completely unstruc-
tured triangular grid using the numerical scheme of second approximation order. The
important technical features of interaction with SALOME software used to build an
unstructured triangular computational grid, including differences in the results of the
triangulation algorithms, are discussed. The content of the structure elements of the
output file format CGNS of SALOME is considered. The mechanisms of detonation
initiation in the multifocused system are investigated. The grid convergence problem
and the influence of the resolution on flow structures are considered.

7.1 Introduction

Mathematical modeling of detonation processes is a significant tool for clarifying

the mechanisms, reducing the risks and costs of providing experiments, and supple-
menting the existing knowledge about the processes occurring during detonation.
As noted in the review [1], the majority of the works on numerical studies of gas

A. I. Lopato (B) · P. S. Utkin

Institute for Computer Aided Design of the RAS, 19/18, Vtoraya Brestskaya ul., Moscow 123056,
Russian Federation
e-mail: [email protected]
P. S. Utkin
e-mail: [email protected]
A. I. Lopato · A. G. Eremenko · P. S. Utkin · D. A. Gavrilov
Moscow Institute of Physics and Technology (National Research University), 9, Institutskiy per.,
Dolgoprudny, Moscow Region 141701, Russian Federation
e-mail: [email protected]
D. A. Gavrilov
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 79

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_7

[email protected]
80 A. I. Lopato et al.

detonation problems are related to the use of mainly structured computational grids.
At the same time, the problem of Detonation Wave (DW) initiation in complex struc-
tures, which include the curvilinear boundaries, protrusions, and bodies of various
forms in the geometry of the problem to be solved, occupies a special place in this
branch of science. The presence of such entities increases the role of unstructured
grids allowing not only to cover the entire computational domain under considera-
tion but also to adapt the computational grid in the areas, where it turns out to be
necessary. In addition, effective adaptation of the grid can also reduce the total CPU
cost in comparison with a uniform fine grid [1]. Finally, programming of adaptive
computational grids for unstructured grids in some cases is relatively easier than for
structured grids, as discussed in [1–3]. Although earlier works with calculations on
unstructured computational grids use computational methods of low approximation
order, gradually, methods and tools of the order increase (limiters, Weighted Essen-
tially Non-Oscillatory (WENO)-schemes [4], spectral volume method [5]) appear
and develop in literature. On the other hand, the unstructured grid method is more
diffusive than the structured one (see [1], for example). The accuracy of the gradient
calculation on the unstructured grid is less than on the structured one, especially in the
area of radically changing physical values. As a result, the solution of problems with
strong Shock Waves (SWs) including detonation requires a finer resolution, which
means a greater number of cells and computational sources and times in comparison
with structured grids.
Let us consider some works devoted to the study of DW on the unstructured
grids. In [1], attention was paid to the numerical study of the formation of cellular
detonation in H2 /air mixture in the tube in the two-dimensional case. The question
of the required resolution of the unstructured grid and comparison with the results
obtained using a structured grid in the simulation of detonation in the mixture was
considered. The series of calculations on unstructured grids with an average cell size
of 1, 2.5, 3, 5 μm and the calculation on a structured grid with a cell size of 5 μm
were carried out. The detailed model of Petersen and Hanson was used to model the
chemical reactions occurring in the mixture. The results show that the unstructured
grid simulation carried out with the identical grid resolution of structured grid study
could not capture the key DW features.
It was shown that in contrast to the calculation on the structured grid, some key
elements of the solution, such as triple point, are not fully and clearly captured using
the unstructured grid with cell size 5 μm due to the coarse grid resolution. The effect
of the grid resolution of the transverse waves depends not only on the resolution
around the detonation front but also on the resolution around the shear layer behind
the detonation front. The simulations that were carried out using cell size of 2.5 μm
and more could not capture the vortices in the area of shear layer. On the other hand,
the calculation with cell size 1 μm captures the vortices like 5 μm structured grid.
Thus, insufficient grid resolution leads to under-resolution of the key elements of DW
structure (shear layers, vortices), which leads to the serious qualitative differences
from the results obtained with sufficient resolution. The significant contribution to
this feature is the instability of the key elements of the solution. The authors noted that
the most important advantage of unstructured grids is easy application of adaptive

[email protected]
7 Numerical Simulation of Detonation Initiation … 81

grid refinement that allows adding/removing nodal points without modifying the
code drastically.
The work [6] focused on the numerical studies of DW structure in high and
low activation energy model mixtures characterized by their irregular and regular
detonation structure, respectively. The mathematical model was represented by 2D
Euler equations with the single-step Arrhenius kinetic reaction mechanism.
The ignition mechanism in irregular structure that corresponds to the high acti-
vation energy is due to the both shock compression behind the main front and by
turbulent mixing for unburned gas pockets of hot and cold gases at shear layers
associated with vortices characterized by hydrodynamic instabilities, mainly Richt-
myer–Meshkov instability, while the mechanism in regular structure that corresponds
to the low activation energy is completely defined by the shock compression. It was
also shown in [6] that DW structure depends on the grid resolution. The low reso-
lution (50 and 125 cells per half-reaction length) gives very poor insight about the
structure and instabilities behind DW front and due to this fact demonstrates very reg-
ular structure, while for finer resolution (300 and 600 cells per half-reaction length)
the structure is irregular. Moreover, there are no unburned gas pockets in the case of
low resolution of the grid. The lack of the pockets in case of coarse grid is associated
with insufficient resolution in numerical diffusion which artificially accelerated the
reaction rate.
The regular structure in low activation energy mixtures totally differs from the
irregular structure in considered problem. The number of elements including large
vortices of the detonation structure is much less than it was in irregular structure.
The whole gases which have passed through DW front are burned completely, and
no unburned gas pockets are apparent in the detonation structure. The results justify
that the grid resolution does not change qualitatively the detonation structure of the
mixture in low activation energy. This fact is associated with the absence of the
elements required for changes, mainly secondary triple points and hydrodynamic
instability. Thus, the work demonstrates that some differential problems may be
unstable depending on the dimensionless activation energy and may not have a grid
convergence in terms of classical theory.
Mentioned above researches demonstrate the possibility of productive study of
DW flows using unstructured computational grids. Among many problems devoted
to DW, an important place is occupied by the problem of optimization of detonation
initiation. The optimization can be reached with the help of SW reflection from
the focusing system. The system of elliptical reflectors was considered in works of
Vasil’ev and named “multifocused” system [7].
The motivation of this work is related to the study of detonation issues in multi-
focused systems and clarification of the basic mechanisms accompanying the deto-
nation process using unstructured computational grids, as well as, the study of the
manifested features using the unstructured grids approach for numerical studies in
this branch.
The chapter is organized as follows. The statement of the problem with the geom-
etry of the considered domain is represented in Sect. 7.2. The governing equations
for two-component fluid flows are introduced and approximated by the numerical

[email protected]
82 A. I. Lopato et al.

scheme of second approximation order on completely unstructured computational

grids in Sect. 7.3. Section 7.4 discusses the features of software used for grid gener-
ation and visualization of obtained results. Section 7.5 gives numerical experiments
with a discussion of the results, in particular, grid convergence problem. Finally, we
conclude this work in Sect. 7.6.

7.2 Problem Statement

Consider 2D channel with the elliptical end-wall shape filled with stoichiometric
hydrogen–oxygen mixture. The multifocused system with elliptical curves is rep-
resented schematically in Fig. 7.1. The values of semi-axis of considered elliptical
curves are equal to 5 and 3.5 mm. At the initial time moment, the incident SW is at
the distance of 28 mm from the left end of the channel. The considered Mach number
of the incident SW is 2.5. The pressure and temperature of the mixture in the area x
> 28 mm are 0.04 atm and 298 K. For the sake of computation cost diminishing, the
computational domain corresponds to the one half of the channel. The bottom half
part of the geometry is used in computations. As a result, the symmetry conditions
are set at the upper boundary, the slip-conditions at the right and bottom boundaries
and the inflow conditions with parameters of the incident SW at the left boundary.
Note that the sizes of the geometries of the considered computational domain
correspond to the geometry from the experimental work of Vasil’ev [7]. Elliptical
reflectors were fabricated by means of milling with milling cutter angle 45°.

Fig. 7.1 Schematic

statement of the problem

[email protected]
7 Numerical Simulation of Detonation Initiation … 83

7.3 Mathematical Model and Numerical Method

The mathematical model that produces the results discussed below includes 2D Euler
equations supplemented by one-step chemical reaction global model of hydrogen–
oxygen combustion described by the first-order Arrhenius kinetics. The governing
equations for considered two-component flows may be written in Cartesian frame
using Eq. 7.1.

∂U
+ ∂F + ∂G = S
⎡ ⎤ ⎡ ∂t∂ x⎤ ∂ y ⎡ ⎤ ⎡ ⎤
ρ ρu ρv 0
⎢ ρu ⎥ ⎢ ρu 2 + p ⎥ ⎢ ρvu ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ 2 ⎥ ⎢ ⎥
U = ⎢ ρv ⎥ F = ⎢ ρuv ⎥ G = ⎢ ρv + p ⎥ S = ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ e ⎦ ⎣ (e + p)u ⎦ ⎣ (e + p)v ⎦ ⎣ 0 ⎦
ρZ ρZ u ρZ v ρω
 
e = ρ2 u 2 + v2 + ρε, ε = ρ(γ−1)
p
+ Z Q, p = μρ RT, ω = −AρZ exp −E RT .
(7.1)

The nomenclature used here is the standard one. The specific heat ratio γ , molar
mass μ, heat release Q, and activation energy E are taken from the database [8]:

g MJ kJ m3
γ = 1.23, μ = 12 , Q = 7.37 , E = 76.2 , A = 9.16 · 108 .
mole kg mole kg s

Parameter A is calculated with the use of the reaction zone time, corresponded to
Konnov reaction mechanism:
1
A= exp(E RTvN ),
τρvN

where τ = 0.47 μs, ρvN = 0.541 kg/m3 , and T vN = 1682 K in accordance with [8].
The main feature of the computational technique is the numerical solution of the
governing equations on completely unstructured computational grids with triangular
cells. The computational algorithm is based on Strang splitting principle in terms of
physical processes called as the convection and chemical reactions. The spatial part
of Eq. 7.1 is discretized using the finite-volume method. The flux is calculated using
Advection Upstream Splitting Method (AUSM) [9] extended for the case of a two-
component mixture. Note that the chosen AUSM scheme of flux calculation is not a
necessary requirement. Thus, in [10], the calculation of DW flows was successfully
carried out with Courant–Isaacson–Rees flux scheme. For the approximation order
increase, the special reconstruction of the grid functions and minmod limiter are
applied [11]. The numerical flux in the local frame is computed using the parameters
from the left and the right sides of the edge. Time integration is advanced by the
second-order Runge–Kutta method [12]. The time step is chosen dynamically from
the stability condition. On the second stage of the algorithm, the system of ordinary

[email protected]
84 A. I. Lopato et al.

differential equations of chemical kinetics for Z variable and temperature in each

computational cell of the grid is solved.
The detail description of the numerical method with the verification of the
algorithm can be found in [13].

7.4 Technical Features

The realization of the computational algorithm includes three main steps called as
the filling of a given area with a set of computational cells with given properties or the
construction of a computation grid, carrying out computations in accordance with the
chosen computational algorithm, and visualizing and processing the obtained results.
In order to build a computational grid in this work, we use software SALOME [14],
which has ample opportunities for the construction of the geometry of the computa-
tional domain and filling it with cells of various types. Software can be downloaded
and installed not for Linux platforms but also has experimental build for Windows.
SALOME module of grid construction provides a wide range of algorithms partic-
ularly suited for finite-element and finite-volume methods. Group naming provides
the identification of local boundaries and contains various output formats of files
with results that can facilitate the visualization or other post-processing operations.
In grid construction module, it is possible to create a group of grid entities, in
particular, edges. Create a group with the name “in” for the edges with the inflow
boundary condition. For our computational domain, these are the edges that form the
left vertical boundary of the channel of the computational domain. The edges that
form the rest of the boundary will be combined into another group called “wall”.
For these edges, the slip-conditions are realized. Groups are exported along with
grid objects to a small number of file formats, one of which is Computational Fluid
Dynamics (CFD) General Notation System (CGNS) [15].
CGNS provides a standard for recording and recovering computer data associ-
ated with the numerical solution of equations of fluid dynamics [15]. The intent
is to facilitate the exchange of CFD data between sites, applications codes, and
across computing platforms, and to stabilize the archiving of CFD data. Application
Program Interface (API) is a platform-independent and includes the realization in
C/C++ application. The data are stored in a compact, binary format and are accessible
through a complete and extensible library of functions.
CGNS file containing a grid is a hierarchical data type having the following
structure. This entity is organized into a set of “nodes” in a tree structure according
to the certain rules that allow users to easily access the necessary information. The
rules are described in Standard Interface Data Structures (SIDS) [15]. The topmost
node is called the “root node”. Each node can be a “parent” for one or more “child”
nodes. A node can also have a child node reference to a node elsewhere in the file or
to a single node. The links are transparent to the user: the user “sees” the associated
child nodes as if they really exist in the current tree. The structure of the file with the
studied grid is shown in Fig. 7.2.

[email protected]
7 Numerical Simulation of Detonation Initiation … 85

Fig. 7.2 Hierarchical data type of CGNS file format

The first level of the hierarchical structure of CGNS file determines the version of
CGNS file—section CGNSLibraryVersion and the dimension of the task—section
SMESH_MESH. The CGNS version of the file defines API for interacting with the
file structure and possible types of nodes. Version 3.4.0 [16] is used in this work.
The next level characterizes the geometric model, number of the boundary vertices
(section Bar_2), number of cells (section Bar_3), coordinates of points in space
(section GridCoordinates), and boundary conditions (section ZoneBC). For various
problems and geometries, three types of boundary conditions are introduced: “in” cor-
responds to the inflow conditions, “out” corresponds to the extrapolation conditions,
and “wall” corresponds to the slip-conditions. Each node contains numbers of ver-
tices that are located on one or another boundary. Section GridCoordinates describes
directly the position of nodes in space on each of the axes: sections CoordinateX,
CoordinateY, and CoordinateZ.
A triangular calculation grid may be generated in several ways using SALOME.
In particular, there are three main algorithms of triangulation called as “Projection
1D-2D”, “Delaunay”, and “Frontal”. The computations in this work were carried out
using the Projection 1D-2D grid generation algorithm.
The algorithm is based on the advancing front method. The generation of the
computational grid starts from the boundary of the computational domain. On the
boundaries of the domain, nodes are searched for by recursively splitting the region
in half to the fineness limited by the parameters of the algorithm (the reference size of
the triangle edge). Further, the iterative process of searching for nodes is carried out,
which are connected to the edges of the boundary and form a new layer of triangles
and a new border separating the area with the generated grid from the area without
it. As a result, the state of the algorithm is always represented by the advancing
boundary front. The coordinates of the nodes of the new boundary are determined by
the area in front of the advancing boundary, in accordance with the rules and criteria
of [17].

[email protected]
86 A. I. Lopato et al.

The result of using this algorithm, as well as, two other algorithms is shown in
Fig. 7.3. As a region of triangulation, we consider a rectangle 0.1 × 0.02 m2 . We
specify in SALOME the edge length of 50 μm as the reference length of the edges
of triangles. Delaunay triangulation shows the distribution of the maximum lengths
of the triangles edges, which is close in shape to the normal Gauss distribution.
For the Frontal algorithm, all the constructed triangular cells have the same value
of maximum edge length. In the case of Projection 1D-2D triangulation, a part of
the cells has a maximum edge length that is equal to the selected reference value in
SALOME, the length of the remaining cells is higher and obeys a certain distribution,
and the overwhelming number of triangles (about 75%) has a maximum edge length
of about 87 μm. Thus, the selected algorithms give qualitatively different patterns
of distribution of triangles in terms of geometrical parameters. The question of the
optimal algorithm in the computations is quite complicated and determined by many
factors including the time of triangulation, quality of the computational grid, geom-
etry of the computational domain, and features of the computational algorithm when
dealing with unstructured grids.
To visualize the results obtained, Visit software [18] is used. Earlier versions of
this software had problems with the display of CGNS files, but since version 2.13.3
these problems have disappeared.

Fig. 7.3 The percentage

distribution of the maximum
lengths of triangles in the
case of Delaunay algorithm
(red color), Frontal (green
color), and Projection 1D-2D
(blue color). The horizontal
axis corresponds to the
maximum length of the
edges of the triangles in μm.
The vertical axis corresponds
to the percentage of triangles
with the selected maximum
edge length

[email protected]
7 Numerical Simulation of Detonation Initiation … 87

7.5 Numerical Experiments

A series of unstructured grids is prepared for resolution study. The average triangular
element sizes equal 12.5, 20, and 35 μm, while the approximate numbers of cells
are 2, 4, and 8 million cells. Consider the process of detonation initiation for the grid
with the average cell size of 12.5 μm.
Shock waves that are formed by the reflection of the incident SW from the plane
wall parts amplify the parameters inside the reflector cavity. The main reflected SW
emerging from the elliptical cavity is created by the reflected incident wave and Mach
stem. The combustion area gradually occupies the cavity of the reflector and spreads
outward primarily in the area between the shear layers, which separate completely
the burned gas of the jet flow with the compressed gases that have passed through the
reflected wave. Detonation initiates at about 41 μs in several points at the symmetry
axis of the channel outside the cavity of the reflector. First, the initiation of DW
occurs when the combustion front is reflected from the symmetry axis. Second, the
numerous collisions of the generated waves with the plane wall between the reflectors
leads to the increase in the gas pressure and temperature near the plane wall between
the reflectors and the appearance of the second place of DW initiation.
The main large-scale structures are successfully resolved using the relatively
coarse grid while the fine grid better resolves shear layers and vortices. Figure 7.4
confirms this fact. Some secondary modes of obtained hydrodynamic instabilities
are not apparent in case of low grid resolution. As a result, the present grid study
demonstrates less irregular structure for low considered resolution and more irreg-
ular one for high grid resolution. In addition, better resolution of the grid leads to
a smaller smearing of shocks. In the computations, these facts lead to the fact that
grid resolution affects the combustion process and moment of DW initiation. For
the coarsest grid, the detonation initiation occurs in the time range 40–41 μs. With
the refinement of the grid, the same moment occurs closer to 40 μs. Thus, the noted
peculiarities of computations can explain the weak change in the time moment of
initiation of detonation waves during the refinement of the grid.

7.6 Conclusions

The numerical investigation of mechanisms of detonation wave initiation is carried

out. The results of the computations show that the focusing of reflected waves, inter-
action with the generated waves, and generated hydrodynamic instabilities play an
important role in DW initiation in the considered problem. Mathematical model is
based on two-dimensional Euler equations written in Cartesian frame and supple-
mented by one-stage chemical reaction model. The numerical method of second
approximation order for integration of the governing equations on fully unstructured
triangular grid is proposed. The features of the technical component of the work

[email protected]
88 A. I. Lopato et al.

Fig. 7.4 Predicted temperature distributions at the time moment 40 μs for different grid resolution.
The considered average triangular element sizes are: a 35 μm, b 20 μm, and c 12.5 μm

responsible for the triangulation of the computational domain are considered. Con-
struction of the computational grid is carried out using the software SALOME. A
comparison of three algorithms for triangulation using the example of a test prob-
lem was made. The structure of the output file in CGNS format relevant for our
computations is described.
The series of unstructured grid is applied for resolution study. The computations
show that the initiation of detonation occurs later on the considered coarse grid than
on the detailed ones. The results obtained are in qualitative agreement with the results
of numerical studies by other authors. In particular, our calculations are confirmed
by the observations in [1]. The computations on unstructured grids require the use of
sufficiently detailed grids since the unstructured grid method is more diffusive than
the structured one. Insufficient grid resolution in the calculations causes the smearing
of shocks and under-resolution of hydrodynamic instabilities that lead to certain
differences in the flow patterns of the gas mixture and the weak change in the time
moment of detonation initiation. However, the use of unstructured grids allows us to

[email protected]
7 Numerical Simulation of Detonation Initiation … 89

explore the areas of arbitrary geometries, which is relevant in solving a number of

problems and expands the range of applicability of the computational method.

Acknowledgements The work of A. I. Lopato and P. S. Utkin is carried out under the state task
of the ICAD of the RAS.

References

1. Togashi, F., Lohner, R., Tsuboi, N.: Numerical simulation of H2 /air detonation using
unstructured mesh. Shock Waves 19, 151–162 (2009)
2. Loth, E., Sivier, S., Baum, J.: Adaptive unstructured finite element method for two-dimensional
detonation simulations. Shock Waves 8(1), 47–53 (1998)
3. Pimentel, C., Azevedo, J., Silva, L.: Numerical study of wedge supported oblique shock wave-
oblique detonation wave transitions. J. Braz. Soc. Mech. Sci. Eng. 24, 149–157 (2002)
4. Luo, H., Baum, J., Lohner, R.: A Hermite: WENO-based limiter for discontinuous Galerkin
method on unstructured grids. J. Comput. Phys. 225, 686–713 (2007)
5. Wang, Z.J.: Spectral (finite) volume method for conservation laws on unstructured grids: basic
formulation. J. Comput. Phys. 178, 210–251 (2002)
6. Mahmoudi, K., Mazaheri, K.: High resolution numerical simulation of the structure of 2-D
gaseous detonations. Proc. Combust. Inst. 33, 2187–2194 (2011)
7. Vasil’ev, A.A.: Cellular structures of a multifront detonation wave and initiation (Review).
Combust. Explos. Shock Waves 51(1), 1–20 (2015)
8. Schultz, E., Shepherd, J.: Validation of detailed reaction mechanisms for detonation simulation.
CalTech Explosion Dynamics Lab, Report No. FM99-5 (2000)
9. Liou, M.-S., Steffen Jr., C.J.: A new flux splitting scheme. J. Comput. Phys. 107, 23–39 (1993)
10. Lopato, A.I., Utkin, P.S.: The usage of grid-characteristic method for the simulation of flows
with detonation waves. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov, S.S.,
Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development of Grid-
characteristic Method, SIST, vol. 133, pp. 281–290. Springer (2019)
11. Chen, G., Tang, H., Zhang, P.: Second-order accurate Godunov scheme for multicomponent
flows on moving triangular meshes. J. Sci. Comput. 34, 64–86 (2008)
12. Shu, C.W., Osher, S.: Efficient implementation of essentially non-oscillatory shock-capturing
schemes. J. Comput. Phys. 77, 439–471 (1988)
13. Lopato, A.I., Utkin, P.S.: Numerical study of detonation wave propagation in the variable
cross-section channel using unstructured computational grids. J. Combust. 3635797, 1–8 (2018)
14. Salome. The open source integration platform for numerical simulation. https://www.salome-
platform.org. Accessed 20 June 2019
15. CGNS. CFD data standard. http://cgns.github.io. Accessed 20 June 2019
16. CGNS Version 3.4.0. https://github.com/CGNS/CGNS/releases/tag/v3.4.0. Accessed 20 June
2019
17. Schoberl, J.: An advancing front 2D/3D-mesh generator based on abstract rules. Comput. Vis.
Sci. 1, 41–52 (1997)
18. Visit. https://wci.llnl.gov/simulation/computer-codes/visit. Accessed 20 June 2019

[email protected]
Chapter 8
On the Stability of a Detonation Wave
in a Channel of Variable Cross Section
with Supersonic Input and Output Flows

Vladimir Yu. Gidaspov and Dmitry S. Kononov

Abstract The possibility of the formation of the stationary Shock Wave (SW) and
Detonation Wave (DW) in a variable cross section channel with the hydrogen–air and
hydrogen–oxygen mixtures in a quasi-one-dimensional non-stationary formulation
is investigated. The channel has a form of two successively arranged Laval nozzles.
A comparative analysis of the solutions in stationary equilibrium, stationary frozen,
non-stationary frozen, and non-stationary non-equilibrium formulations is presented.
Equilibrium initial approximation was proposed for non-equilibrium flows modeling.
Configurations of variable cross section channel with a stationary (detonation) wave
in the first expanded area are obtained. It is shown that non-equilibrium stationary
solutions in the first narrowing part of a dual Laval nozzle channel are unstable,
and non-equilibrium stationary solutions in the second expanded part are unstable
too, but they stabilize in the first one. The range of flow rates, at which a stationary
detonation wave exists, can be predicted with a high degree of accuracy by the
equilibrium stationary theory.

8.1 Introduction

Mathematical modeling of high-speed flows of combustible mixtures has great practi-

cal importance in the studying of processes of combustion and detonation in channels.
This approach substantially complementing the natural experiment makes possible
to increase a reliability of the designed power plants with lower material costs and
helps to understand the physical/chemical phenomena observed in them.
Possibility of the formation of the stationary SW and DW in a variable cross section
channel in a quasi-one-dimensional non-stationary formulation is investigated in this

V. Yu. Gidaspov · D. S. Kononov (B)

Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
V. Yu. Gidaspov
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 91

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_8

[email protected]
92 V. Yu. Gidaspov and D. S. Kononov

Fig. 8.1 The location of the

stationary DW in the channel

chapter. A comparative analysis of the obtained solutions and the results from [1], in
which the case of fuel combustion in Laval nozzle in the stationary DW is considered
with subsequent acceleration of combustion products to supersonic rates (Fig. 8.1) in
a quasi-one-dimensional stationary approximation, is presented. The gas flow before
DW is considered frozen, whereas after DW is equilibrium. In [1], the stationary
solution depends exclusively on the ratio of the channel areas in the current, inlet,
and minimum cross sections.
In [1–5], it is noted that in a frozen flow the stationary SW is stable in an expand-
ing area of channel and unstable in a narrowing one. To study DW stability in a
variable cross section channel, a numerical simulation of a frozen and chemically
non-equilibrium quasi-one-dimensional flow in a variable-section channel having
the form of two successively arranged Laval nozzles in series (Fig. 8.2) was carried
out. The contour is given by the analytical way. The radii of the critical sections of
the channel are R1 and R2 , respectively, with R1 > R2 . The radius of the inlet cross
section is R0 . The problem is studied in the non-stationary and stationary quasi-
one-dimensional formulations. Mathematical model does not take into account the
effects of viscosity, thermal conductivity, and diffusion. The flow at the channel inlet
was supersonic, and SW/DW was realized inside the channel, in which the flow
was braked to subsonic speeds. Approximately, at the critical section of the second
Laval nozzle, the flow rate passed through the sound speed, and in the expanding
part of the second nozzle the flow accelerated to supersonic rates. Hereinafter, the
narrowing part of the first Laval nozzle will be referred to as the first narrowing part,
the expanding part of the first Laval nozzle will be referred to as the first expanding
part, the narrowing part of the second Laval nozzle will be referred to as the second
narrowing part, and the expanding part of the second Laval nozzle will be referred
to as the second expanding part.
In [3], the supersonic inlet flow with the constant heat implementation was investi-
gated using the analytical and numeric methods. According to [3], DW stabilization

[email protected]
8 On the Stability of a Detonation Wave … 93

Fig. 8.2 Dependence of the radius of the investigated channel of variable cross section on the
longitudinal coordinate

is possible with non-constant heat release because of variable composition of the

combustion mixture.
In the equilibrium quasi-one-dimensional stationary formulation of the problem
according to the method from [1] for the above-described two-contour channel,
an analysis of possible flow variants with the stationary SW/DW was carried out
(Figs. 8.3 and 8.4). The relative radii of the location of SW/DW in the channel
(Fig. 8.3) and the temperature behind SW/DW (Fig. 8.4) are determined depending
on the flow rate at the channel inlet. It is shown (Fig. 8.3) that, in the absence of
chemical reactions in a hydrogen–air mixture, SW can be in the second tapering part
at flow rates of ~1300–1500 m/s, in a hydrogen–oxygen mixture at ~1800–2100 m/s,
and the first tapering and the first expanding parts at flow rates exceeding 1500 m/s
and 2100 m/s, respectively. In the case of a chemically reacting mixture (Fig. 8.3),
DW may be in the second tapering part for the hydrogen–air mixture at flow rates of
~2200–2600 m/s, in the hydrogen–oxygen mixture at ~3100–3700 m/s, and in the
first tapering and first expanding parts at flow rates exceeds 2600 m/s and 3700 m/s,
respectively. It should be noted that in the investigated range of parameters a chemi-
cally reacting flow is always realized since the temperature behind SW/DW exceeds
the auto-ignition temperature of the considered combustible mixtures (Fig. 8.4). Fur-
ther, the stable positions of SW/DW in the investigated channel (Fig. 8.2) were studied
by simulating a quasi-one-dimensional non-stationary flow of a multicomponent gas.
The chapter is organized as follows. Section 8.2 presents a mathematical model.
Testing methodology is described in Sect. 8.3. The results of mathematical modeling
are reported in Sect. 8.4. Section 8.5 concludes the chapter.

[email protected]
94 V. Yu. Gidaspov and D. S. Kononov

Fig. 8.3 Dependence of the relative cross-sectional radii of channel from the flow rate at the channel
inlet, in which the stationary SW/DW is implemented. Solid horizontal lines indicate the radius
values of the cross section of the channel respect to the local extrema of the profile, where curve
1—the frozen hydrogen–air mixture, curve 2—reacting hydrogen–air mixture, curve 3—frozen
hydrogen–oxygen mixture, and curve 4—reacting hydrogen–oxygen mixture

Fig. 8.4 Dependence of the temperature of the mixture beyond SW/DW from the flow rate at the
channel inlet, where curve 1—the frozen hydrogen–air mixture, curve 2—the reacting hydrogen–
air mixture, curve 3—the frozen hydrogen–oxygen mixture, and curve 4—the reacting hydrogen–
oxygen mixture

[email protected]
8 On the Stability of a Detonation Wave … 95

8.2 Mathematical Model

Non-stationary chemically non-equilibrium gas flow in the channel is described by

the system of Euler equations supplemented by thermal and caloric ones of state and
N equations of chemical kinetics [6] is provided by Eq. 8.1.
⎧ ∂ρ F
⎪
⎪ ∂t
+ ∂ρu∂x
F
=0
⎪
⎪ ∂ (ρu +P ) F
2
⎪
⎪
∂ρu F
+ = P ddFx
⎪
⎪  ∂t 2  ∂x 
⎪
⎪ ∂ρ E 2 F
u 2
∂ρu F E ρ + u2
P
⎪
⎪ + =0
⎨ ∂t ∂x
∂ρ Fγi ∂ρu Fγi
∂t
+ ∂x
= F Wi i = 1 . . . N (8.1)
⎪
⎪ 
⎪
⎪
N
⎪
⎪ P = ρ RT γi
⎪
⎪
⎪
⎪
i=1
⎪
⎪ 
N
⎩ E= γi E i (T )
i=1

Here, x is the longitudinal coordinate, t is the time, ρ is the density, u is the

mixture rate, P is the pressure, T is the temperature, H is the enthalpy, E i (T ) and
γi are the functions for internal energy from catalogs and concentration of the ith
component in the mixture, respectively, γ̄ is the vector of molar mass concentrations
of the components of the mixture, F = F(x) is the dependence of the cross-sectional
area of the channel on the longitudinal coordinate. When considering frozen flows,
it is assumed that W i = 0, i = 1 … N.
The quasi-one-dimensional stationary chemically non-equilibrium flow is
described by a similar system of equations, in which all the complexes ∂(·)∂t
are set
equal to zero.
A mixture of perfect gases is considered, whose thermodynamic properties are
described by defining the expression for Gibbs potential [7, 8]:

N
Pi
G(P, T, γ̄ ) = γi G i0 (T ) + RT ln , (8.2)
i=1
P0

where R is the universal gas constant, P0 is the normal pressure, G i0 (T ) are the known
dependences [7] of the temperature part of Gibbs molar potential of a separate com-
ponent of the mixture. The internal energy and density of the mixture are expressed
in terms of Gibbs potential and its partial derivatives provided by Eqs. 8.3–8.4.

N
∂G ∂G
E(T, γi ) = G − T −P = γi E i (T ) (8.3)
∂T P ∂T T i=1

[email protected]
96 V. Yu. Gidaspov and D. S. Kononov

∂G P
ρ(P, T, γ̄ ) = 1/ = (8.4)
∂P T N
RT γi
i=1

SW holds the conservation laws (Rankine–Hugoniot equations):

⎧
⎪
⎪ ρ1 v1 = ρ2 v2 ,
⎪
⎨ P1 + ρ1 v12 = P2 + ρ2 v22 ,
v2 v2 (8.5)
⎪
⎪ E 1 + ρP11 + 21 = E 2 + ρP22 + 22 ,
⎪
⎩
γ̄1 = γ̄2 ,

where the index “2” marks the values after SW, the index “1” is before v = D − u,
where D is the rate of SW.
To obtain a numerical solution of the system of Euler equations, the method of
Godunov [9] was used. For integrating the system of equations of the stationary
model, the difference approximation was taken from [1]. To integrate the system of
ordinary differential equations of chemical kinetics, the method of Pirumov from [6]
was used.

8.3 Testing

To test the numerical method of integrating the equations of chemical kinetics, a test
model of the adiabatic reaction at constant density provided by Eq. 8.6 is used.

N
γi E i (T ) = E
i=1
dγi
ρ = Wi (ρ, T, γ1 , . . . , γ N ) γi (0) = γi0 , i = 1, . . . , N
dt
ρ = const E = const (8.6)

The kinetic mechanism for this test consisted of 19 reversible stages involving 8
components [10]. The initial temperature was set at T0 = 1000 K, the initial density
was ρ0 = 0.1 kg/m3 , and a stoichiometric mixture of hydrogen and oxygen was taken
as the reacting gas. A characteristic ignition delay was obtained at the time t  50 µs,
the temperature of the combustion products was set at ≈ 3300 K (Fig. 8.5), and the
change in the concentrations of the mixture components during ignition (Fig. 8.6)
was similar to the results in [9, 10].
As a test for checking the numerical simulation of non-stationary flows, we took
the problem of the interaction of an isolated shock wave propagating in a channel
filled with a non-reacting ideal gas with an adiabatic parameter k = 1.4. It considers
the breakup of an arbitrary discontinuity at the point x = 0.2 with the specification

[email protected]
8 On the Stability of a Detonation Wave … 97

Fig. 8.5 Temperature distribution in the adiabatic reaction model at constant density

Fig. 8.6 The distribution of the concentrations of the components of the mixture in the model of
the adiabatic reaction at constant density, where curve 1—H2 , curve 2—O2 , curve 3—H2 O, curve
4—HO2 , curve 5—H2 O2 , curve 6—OH, curve 7—H, and curve 8—O

of the dimensionless initial parameters provided by Eq. 8.7.

3.857143 if x < 0.2

ρ=
1 + 0.2 sin(50(0.2 − x)) if x > 0.2

[email protected]
98 V. Yu. Gidaspov and D. S. Kononov

3.857143 if x < 0.2

u=
1 + 0.2 sin(50(0.2 − x)) if x > 0.2
10.33333 ifx < 0.2
P= (8.7)
1 ifx > 0.2

The calculation was carried out up to the time point t = 0.18 s. In the absence of an
analytical solution, the numerical solution obtained on such a detailed grid was taken
as an exact one, so that as the number of points increases, it remains unchanged. In
this chapter, this number of points is equal to 15,000 (Fig. 8.7). The area behind SW is
divided into subareas of high-frequency, low-frequency oscillations, and unperturbed
parameters, which are in good agreement with the results given in [11].
Testing of the simulation method of stationary quasi-one-dimensional non-
equilibrium flow was carried out by simulating the flow of a stoichiometric mixture of
hydrogen with oxygen (H2 + 0.5O2 ) in a constant section channel with the presence
of SW in the channel inlet section and without it with an initial temperature of 1000 K
and an initial pressure of 1 atm. Chemical transformations here and hereinafter were
modeled by 8 reversible stages (Table 8.1) [10].
The stoichiometric mixture of hydrogen and oxygen H2 + 0.5O2 with 6 react-
ing components H, O, H2 , O2 , H2 O, and OH and the hydrogen–air mixture H2 +
0.5O2 + 1.881N2 with 7 reacting components H, O, H2 , O2 , N2 , H2 O, and OH were
investigated.
The obtained results are presented on Pressure–Volume (PV)-diagram in the form
of Rayleigh–Michelson Lines (RMLs) (Fig. 8.8). Each of RMLs is limited by the

Fig. 8.7 Density distribution at time t = 0.18 s, where curve 1—the solution obtained by dividing
the computational domain into 15,000 points and curve 2—the solution obtained by dividing the
computational domain into 3,000 points

[email protected]
8 On the Stability of a Detonation Wave … 99

Table 8.1 Kinetic mechanism

Reaction A, mol, m3 , s, K n E, K
H2 + M = H + H + M 5.5 E18 −1.0 51.987
O2 + M = O + O + M 7.2 E18 −1.0 59.340
H2 O + M = OH + H + M 5.2 E21 −1.5 59.386
OH + M = O + H + M 8.5 E18 −1.0 50.830
H2 O + O = OH + OH 5.8 E13 0 9.059
H2 O + H = OH + H2 8.4 E13 0 10.116
O2 + H = OH + O 2.2 E13 0 8.455
H2 + O = OH + H 7.5 E13 0 5.586

Fig. 8.8 PV-diagram of stationary non-equilibrium flow in a constant shape channel, where curve
1—equilibrium DA, curve 2—SA, curve 3—RML for flow with u0 = 3600 m/s behind SW, curve
4—RML for flow with u0 = 2900 m/s behind SW, curve 5—RML for flow with u0 = 2661.3 m/s
behind SW, curve 6—RML for a flow with u0 = 4000 m/s, curve 7—RML for a flow with u0 =
3600 m/s, and curve 8—RML for a flow with u0 = 2664.7 m/s

Shock Adiabats (SA) and Detonation Adiabats (DA), which indicates the fulfillment
of conservation laws. Chapman–Jouguet mode (u0 = 2661.3 m/s) obtained by cal-
culation, in which RML DA touch (curve 5) is observed, with a decrease of rate, has
no solution. Unstressed flow regimes (curves 6–8) are received. It should be noted
that the minimum flow rate, at which there is a solution with u0 = = 2664.7 m/s,
is slightly greater than Chapman–Jouget rate. The calculation results are in good
agreement with the data of [9].

[email protected]
100 V. Yu. Gidaspov and D. S. Kononov

8.4 The Results of Mathematical Modeling

At first, the problem of modeling a frozen flow in a contour has been solved (Fig. 8.9).
As an initial approximation of the solution, it was assumed that in the whole compu-
tational domain uo = 4297 m/s, T 0 = 400 K, P0 = 1 atm, a stoichiometric mixture
of hydrogen and oxygen flows. The problem of obtaining a stable state flow was
solved, after which the emergence of SW was simulated in the expanding part, and
the search for the stable state solution was carried out again. The method of estab-
lishing a solution was obtained with a stationary SW in the first expanding part of the
investigated contour. It should be noted that the temperature behind the steady SW
≈ 3500 K is consistent with the dependence of the temperature behind a stationary
SW on the input flow rate shown in Fig. 8.4.
By varying the inlet flow rate in the channel, it was possible to obtain stable state
solutions with SW in the range of the initial rate from 2172 to 4297 m/s (Fig. 8.10).
Then, the hydrogen–air reacting gas was considered as an investigated mixture.
Based on the method given in [1] for a chemically equilibrium hydrogen–air mixture
in the investigated channel (R1 /R0 = 0.56, R2 /R0 = 0.493), an RR-diagram can be
constructed, according to which DW in this problem can be in the second narrowing
part of the channel in question at an input flow rate of more than 2200 m/s, in the
first tapering and first expanding parts of the channel in question at an input flow rate
of more than 2700 m/s (Fig. 8.11).
It should be noted that the equilibrium solution with a stationary SW depends
exclusively on the ratio of the channel radii in the current and inlet sections. Thus,

Fig. 8.9 Temperature distribution in time layers, frozen stoichiometric mixture of hydrogen with
oxygen, uo = 4297 m/s, where curve 1—channel contour shape, curve 2—start of propagation of
the implemented shock wave, t = 0.000255 s, curve 3—position of the shock wave at time t =
0.000375 s, and curve 4—position of the shock wave at time t = 0.02 s (stable state)

[email protected]
8 On the Stability of a Detonation Wave … 101

Fig. 8.10 The range of inlet rates in the channel and the corresponding stable state solutions, a
frozen stoichiometric mixture of hydrogen with oxygen, where curve 1—channel contour form,
curve 2—stable state solution, u0 = 4297 m/s, and curve 3—stable state solution, u0 = 2172 m/s

Fig. 8.11 RR-diagram for a chemically equilibrium hydrogen–air mixture at T 0 = 400 K

from the equilibrium stationary simulation it follows that there are three possible
positions of the stationary SW for the considered channel configuration (Fig. 8.12).
The stability of the position of SW/DW was investigated by the method of relax-
ation [3]. When setting the initial approximation with SW in the first narrowing
part of the channel with an input rate of 2750 m/s, DW turned out to be unstable

[email protected]
102 V. Yu. Gidaspov and D. S. Kononov

Fig. 8.12 Solutions with the stationary SW for a chemically equilibrium hydrogen–air mixture with
an inlet flow rate of 2750 m/s at R/R0 = 0.57, R is the section radius, in which SW is realized, where
curve 1—contour form, curve 2—solution with stationary SW in the second narrowing part, curve
3—solution with stationary SW in the first expanding part, and curve 4—solution with stationary
SW in the first narrowing part

and moved to the left, against the direction of flow leaving the computational area
(Fig. 8.13). This fact is consistent in RR-diagram depicted in Fig. 8.11.
When setting the initial approximation with SW in the first expanding part of the
channel with an inlet rate of 2750 m/s, DW was stable and only slightly shifted to
the left relative to the initial approximation (Fig. 8.14).
When setting the initial approximation with SW in the second narrowing part
of the channel in question with an input rate of the mixture of 2750 m/s, DW was
unstable in this position. It began to move to the left against the direction of flow
and stabilized with the same cross-sectional area of the channel as in the above case
(Fig. 8.15).
When specifying an initial approximation with SW in the second narrowing part
of the channel with an inlet rate of a mixture of 2650 m/s, DW is also unstable, but,
at the same time, unlike in the previous case it completely went beyond the limits of
the computational domain (Fig. 8.16).
The solutions obtained by the relaxation method were compared with the solu-
tions obtained by solving the direct problem of nozzle theory (quasi-one-dimensional
stationary case) with passing a singular point in the vicinity of the critical section
(Fig. 8.17) using the modified algorithm from [12]. The obtained results are in good
agreement in all computational areas with the exception of DW vicinity. This differ-
ence can be explained by the fact that to simulate the fine structure of DW, very small
discretization of the computational domain is required, which was implemented for
solving the problem with the stationary formulation only. When solving the problem

[email protected]
8 On the Stability of a Detonation Wave … 103

Fig. 8.13 Non-equilibrium non-stationary calculation of the flow of a hydrogen–air mixture with
DW in the first narrowing area, where curve 1—contour form, curve 2—initial approximation, curve
3—temperature distribution at time t = 0.000684 s, and curve 4—temperature distribution at time
t = 0.000776 s

Fig. 8.14 Non-equilibrium non-stationary calculation of the flow of a hydrogen–air mixture with
DW in the first expanding area, where curve 1—contour form, curve 2—initial approximation, and
curve 3—stable state solution: a temperature distribution in the channel, b temperature distribution
in the vicinity of the relaxation of DW

in the non-stationary formulation, the fine discretization in the vicinity of DW was

not done.

[email protected]
104 V. Yu. Gidaspov and D. S. Kononov

Fig. 8.15 Non-equilibrium non-stationary calculation of the flow of a hydrogen–air mixture with
DW in the second narrowing area, where curve 1—form contour, curve 2—initial approximation,
and curve 3—stable state temperature distribution

Fig. 8.16 Non-equilibrium nonstationary calculation of the flow of a hydrogen–air mixture with
DW in the second expanding area where curve 1—contour form, curve 2—initial approximation,
curve 3—temperature distribution at time t = 0.000811 s, curve 4—temperature distribution at time
t = 0.004928 s, and curve 5—temperature distribution at time t = 0.007188 s

[email protected]
8 On the Stability of a Detonation Wave … 105

Fig. 8.17 Chemically non-equilibrium mixture of hydrogen–air, inlet flow rate of 2750 m/s: a veloc-
ity distribution (curves 3 and 4) and sound velocity (curves 1 and 2) in the channel, b in the vicinity
of DW, c in the vicinity of the critical section (curves 1 and 4—method of relaxation) and (curves
2 and 3—direct problem of nozzle theory)

8.5 Conclusions

A quasi-one-dimensional non-stationary formulation of a chemically non-

equilibrium flow of a combustible hydrogen–air mixture in a channel consisting
of two consecutive Laval nozzles with fuel combustion in a stationary detonation
wave with supersonic flow and at the channel inlet and outlet was investigated. By
calculation, it was obtained that a stationary detonation wave is stable in the first
expanding part of the channel and unstable in narrowing parts. The range of the flow
rates at the channel inlet, at which the formation of a stationary detonation wave is
possible, is obtained. It is shown that for a hydrogen–air mixture in the investigated
channel, the range of flow rates, at which a stationary detonation wave exists, can be
predicted with a high degree of accuracy by the equilibrium stationary theory.

Acknowledgements This work was carried out within the state task no. 9.7555.2017/BCh.

[email protected]
106 V. Yu. Gidaspov and D. S. Kononov

References

1. Gidaspov, V.Yu.: Numerical simulation of one-dimensional stationary equilibrium flow in

engine detonation. Trudy MAI 83, 1–20 (in Russian) (2015)
2. Cherniy, G.G.: Gas dynamics. Moscow, Nauka Publ., (in Russian) (1988)
3. Levin, V.A., Manuilovich, I.S., Markov, V.V.: Stabilization of detonation waves in a supersonic
flow. Moscow Univ. Mech. Bull. 66(4), 77–82 (2011)
4. Kraiko, A.N., Shironosov, V.A.: Investigation of flow stability in a channel with a closing shock
at transonic flow velocity. J. Appl. Math. Mech. 40, 533–541 (1976)
5. Grin, V.T., Kraiko, A.N., Tillyaeva, N.I., Shironosov, V.A.: Analysis of one-dimensional flow
stability in a channel with arbitrary variation of stationary flow parameters between the closing
shock cross section and channel outlet. J. Appl. Math. Mech. 41, 651–659 (1977)
6. Pirumov, U.G., Roslyakov, G.S.: Gas Flow in Nozzles. Springer, Berlin, Heidelberg (1986)
7. Gurvich, L.V., Veyts, I.V., Alcock, C.B.: Thermodynamic Properties of Individual Sub-stances,
4th edn, vol. 1. Elements O, H(D, T), F, Cl, Br, I, He, Ne, Ar, Kr, Xe, Rn, S, N, P, and Their
Compounds. Part 2. New York, Hemisphere Pub. Corp. (1989)
8. Gidaspov, V.Yu., Severina, N.S.: Numerical simulation of the detonation of a propane-air
mixture, taking irreversible chemical reactions into account. High Temp. 55(5), 777–781 (2017)
9. Godunov, S.K.: Finite difference methods for numerical computations of discontinuous
solutions of the equations of fluid dynamics. Matematicheskii Sbornik 47, 271–306 (1959)
10. Gardiner, W.C. (ed.): Combustion Chemistry. Springer, New York (1984)
11. Shu, C.-W., Osher, S.: Efficient implementation of essentially non-oscillatory shock-capturing
schemes. J Comp. Phys. 83, 32–78 (1989)
12. Gidaspov, V.Yu.: Numerical simulation of chemically non-equilibrium flow in the nozzle of
the liquid-propellant rocket engine. Aerosp. MAI J. 20(2), 90–97 (2013)

[email protected]
Chapter 9
Physical and Kinematic Processes
Associated with Meteoroid When Falling
in the Earth’s Atmosphere

Viktor A. Andrushchenko , Vasiliy A. Goloveshkin

and Nina G. Syzranova

Abstract We consider the motion of a number of known meteor bodies in the Earth’s
atmosphere and their fall out in the form of meteorites on the Earth’s surface. Various
mechanisms of their destruction in the atmosphere are explored within the framework
of the extended theory of meteor physics. For some of them, the proposed hypothetical
model of the formation of the surface relief of falling meteoroids and, accordingly,
their fallen remains was tested, which turned out to correspond to the real structure
of the surfaces of the meteorites found—smooth in some cases and dotted with
irregularities in the form of rhegmaglypts in others.

9.1 Introduction

Meteor bodies are divided into three main classes: stone, iron, and ironstone. In the
percentage terms it is as follows: stone meteorites—92.5%, iron meteorites—5.7%,
ironstone meteorites—1.3%, and anomalous—0.5% [1]. The average number of
falling fireballs per year recorded by satellites and infrasound stations is more than 4
times higher than the average number of meteorites, which confirms the fact that most
of them “burns” in the atmosphere before reaching the Earth’s surface. Here, the term
“burns” refers to either an air explosion or progressive fragmentation with already
real combustion of small fragments left as a consequence of these phenomena.

V. A. Andrushchenko · N. G. Syzranova (B)

Institute of Computer Aided Design of the RAS, 19/18, Vtoraya Brestskaya ul., Moscow 123056,
Russian Federation
e-mail: [email protected]
V. A. Andrushchenko
e-mail: [email protected]
V. A. Goloveshkin
Institute of Applied Mechanics of the RAS, 7, Leningradsky prt, Moscow 125040, Russian
Federation
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 107

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_9

[email protected]
108 V. A. Andrushchenko et al.

Falling meteorites on the Earth is accompanied by a number of physical phenom-

ena: light, sound, and mechanical. The meteor body begins to glow at altitudes of
about 130–80 km, and at altitudes of 20–10 km its movement is significantly slowed
down. On this segment of the path called the region of delay, the heating and evapo-
ration of the meteor body (or fragments) stop, the luminosity disappears, and a thin
molten layer on the surface of the fragments quickly solidifies, forming a melting
crust on the surface. Under the influence of the temperature field depending on the
nature of the movement by this time, the surface takes a smooth or rhegmaglypts-
strewn structure of the relief. Having overcome the region of delay, fragments of
meteor bodies fall almost vertically under the influence of gravity and fall out in the
form of fragments—meteorites.
In this chapter, using the methods of mathematical modeling, we study the pro-
cesses that occur when moving at hypersonic speeds of meteoroids in the Earth’s
atmosphere, and assess their impact on the nature of the destruction of bodies. The
study of the movement and destruction of each particular meteoroid is an independent
task since each of them is significantly different: its form is arbitrary, and the struc-
ture is heterogeneous. As an example, the motion and destruction of several specific
meteorites significantly different in size, properties, composition of the material, and
trajectory parameters are analyzed.
The chapter has a following structure. Section 9.2 discusses a problem of determin-
ing heat fluxes to the surface of meteoroids. In Sect. 9.3, the model of meteoroid’s
fragmentation is described, and in Sect. 9.4 the mechanisms of destruction of the
bodies due to thermal stresses are presented. Section 9.5 concludes the chapter.

9.2 Heat Transfer to the Surface of Meteoroids

Problems of heat transfer of bodies moving at hypersonic speeds within the atmo-
sphere are quite fully studied. This is primarily due to the research of spacecraft
flights and the providing of their thermal protection. In meteor physics, the knowl-
edge of the heat transfer coefficient is necessary for the most realistic assessment of
the behavior of cosmic bodies at the entrance to the atmosphere.
Equations 9.1 of the extended physical theory of meteors in a form of their motion
in the exponential atmosphere [2] are used for the numerical study of the problem,
where V, M, and θ are the speed of the body, its mass, and angle of incidence of the
trajectory to the horizon, respectively, R E is the radius of the Earth, C D , C H , and Heff
are the coefficients of resistance, heat transfer to the surface of the body, and effective
enthalpy of vaporization of the meteoroid, respectively, Smid is the cross-sectional
area of the body, z is the height of the meteor body above the Earth’s surface, ρ0 is
the density of the atmosphere at z = 0, h is the characteristic scale of height.

= Mg sin θ − C D Smid ρV2 V dθ θ

2 2
M dV
dt dt
= g cos θ − VR Ecos
+z (9.1)
= −C H Smid ρV2 dz
3
Heff dM
dt dt
= −V sin θ ρ = ρ0 exp(−z/ h)

[email protected]
9 Physical and Kinematic Processes Associated with Meteoroid … 109

In the high-temperature gas stream, there are two heat transfer mechanisms from
the gas to the surface of the body: the convective and radiative heat transfers.
The following formula [3] is used for the convective heat flux at the critical point
of the spherical surface of the meteorite:
 ρ 1/2
∞
qc0 ≈ 3.3 × 10−5 3.2
V∞ , W/m2 .
R

Here, R is given in m, ρ∞ is given in kg/m3 , and V∞ is given in m/s. The index ∞

marks the parameters of the incident flux. ReVelle formula, the parameters of which
are presented in [3], is used for the radiation heat transfer coefficient at the critical
point provided by Eq. 9.2.

C Hr = f · e A1 ρ A2 +A3 V −1 R A4 +A5 V +A6 V V A7 +A8 V +A9 V −3

2 2
(9.2)

Accordingly, the heat flow at the critical point is written as qr 0 = 0.5ρ∞ V∞

3
C Hr .
The heat flux distribution along the spherical surface for the convective heat flux
is approximated by Eq. 9.3 [4], where β is the meridional cross-sectional angle
measured from the direction to the critical point, and for radiative flux we have [5]:
qr = qr 0 cosn β, n = 1/(0.051V − 0.43) + 1.811.

qc = qc0 (0.55 + 0.45 cos 2β) (9.3)

The total heat flux to the body surface is defined as q = qc + qr .

The calculated ratios of the finite mass M (near the Earth’s surface) to the initial
mass of the atmospheric entry Me for four meteoroids (Tunguska [6], Sikhote-Alin
[7], Kunya-Urgench [8], and Chelyabinsk [9]) are presented in Table 9.1. The calcu-
lations took into account only the ablation of bodies, and did not take into account
the process of fragmentation meteoroids.
The calculated data on the heat transfer to the surface of meteor bodies show that
for large meteoroids (Tunguska or Chelyabinsk), the radiative heat flux was several
orders of magnitude higher than the convective one along the entire trajectory of
the fall (high speed of flight and large body size). At the same time, for a relatively
small meteorite Kunya-Urgench (lower flight speed and small body size), the values
of radiation heat fluxes on almost the entire trajectory were significantly less than
convective. This fact was manifested in the process of mass loss.

Table 9.1 The parameters of meteoroids

Meteoroid M e (t) V e (km/s) Material, density (g/sm3 ) M/M e (%)
Tunguska (1908) 1,000,000 30 Ice, 1.0 1
Sikhote-Alin (1947) 500 15 Iron, 7.8 7
Kunya-Urgench (1998) 3 13 Stone chondrite, 3.3 84
Chelyabinsk (2013) 13,000 19 Stone chondrite, 3.3 5

[email protected]
110 V. A. Andrushchenko et al.

9.3 Fragmentation of Meteoroids

Calculation of the meteoric-body ablation requires us to take into account its frag-
mentation. Statistics of falls of meteoroids shows that most of them fell onto the
Earth as fragmented pieces.
In [2], one of the models of successive fragmentation that uses the statistical
theory of strength is considered. It is known that the structure of meteoric bodies
that penetrate the atmosphere has strength thin homogeneity. The statistical theory of
strength studies structurally inhomogeneous bodies [10]. In the context of this theory,
fragmentation occurs along with defects and cracks, which are available in space
bodies, i.e., the structural inhomogeneity of meteoric bodies influences the process
of fragmentation. As a result, the fragmentation appears as a process of successive
elimination of defects with increase in the load by means of body disruption along
these defects, so that the fragments created have a greater strength than the initial
body. In this connection, the fragmentation process will complete as soon as the
velocity of head begins to diminish.
According to this model, the fragment strength can be written as

σ ∗f = σe (Me /M f )α ,

where σe and Me are the strength and mass of the meteoroid before entering the
atmosphere, σ ∗f and M f are the same characteristics for the fragment, and α is
the exponent of the material nonhomogeneity (for larger α, the nonhomogeneity is
higher). The body fragmentation in the atmosphere takes place under condition:

ρ∗ V∗2 = σ ∗ ,

where the air pressure is equal to σ ∗ that is one of the strength characteristics
of the body material (compression, tensile, and shear strength). The fragmenta-
tion altitude z ∗ in the exponential atmosphere is determined from the condition
ρ∗ = ρ0 exp(−z ∗ / h). From this altitude, instead of a single body a swarm of cleaving
fragments with increasing number N falls that is accounted in the set of characteristic
variables. Assume that the formed fragments are spheres of equal mass their number
depending on the current values of the air pressure and the total mass of all frag-
ments, and the midsection area is obtained (determined on the assumption that the
fragments do not overlap) [2]. The problem of the fragmenting meteor body motion
is solved in three stages. At the first stage, the movement of a single body from the
entry altitude to the fragmentation altitude is considered. At the second stage, from
the fragmentation altitude to the maximum air pressure altitude the motion of swarm
of fragments is considered. At the third stage, the motion of only one fragment is
tracked since it is believed that the fragments are of the same size.
The strength of meteoroids varies widely. For a typical iron Sikhote-Alin meteorite
(Primorsky Krai, 1947) as a result of processing of observations [7], it was found
that the first fragmentation occurred at an altitude of ~58 km. According to these

[email protected]
9 Physical and Kinematic Processes Associated with Meteoroid … 111

data, the critical value of the strength parameter, at which the fragmentation process
began, is estimated by the value σe = 105 N/m2 . The variations in the total mass of
the meteor body depending on the altitude of the flight for the single body model and
for the fragmenting meteoroid at different values of parameter α are given in Fig. 9.1.
Parameter α depends on the degree of the inhomogeneity of the material and size
of the body. It should be noted that during the fragmentation of meteor bodies, the
removal of mass grows rapidly at first due to the increasing surface area in the flow.
However, this leads to an increase in the deceleration of the fragmenting body, i.e.,
to a decrease in the velocity of the meteor body and its fragments, which, in turn,
reduces the heat flow to the surface and slows the process of ablation. For this reason,
as the calculations show, the final mass of the fragmenting meteoroid, as a rule, turns
out to be greater than in the case of the single body model. For example, the final mass
for the fragmenting bolide with parameter α = 0.25 is ~90 t, while the final mass for
a single body is ~36 t. However, as the calculations show, at α = 0.125 and altitude
of 40 km, the meteoroid splits into a multitude N = 2 × 107 of small fragments,
which leads to intensive removal of mass even before their deceleration. In this case,
the final mass of the meteoroid will be approximately 10 t. It is well known that for
Sikhote-Alin meteorite, in total, several tens of thousands of fragments with a total
mass of 27 t were collected.
The cosmic body can collapse into several large fragments, which then fly
autonomously or split into a cloud of small fragments united by a common shock
wave and flying as a whole. This cloud usually expands rapidly and slows down dur-
ing flight causing a bright flash of radiation. When a large meteor body is destroyed,
both fragmentation scenarios can occur simultaneously. In both cases, the fragments
of the body acquire speeds in a direction perpendicular to the trajectory, which can
lead to transverse scattering of fragments.

Fig. 9.1 Variation in mass

of meteoroid along flight
trajectory for different values
of parameter α, where curve
1—single body (without
fragmentation), curve 2—α
= 0.5, N = 2.4×103 , curve
3—α = 0.25, N = 2.0×105 ,
and curve 4—α = 0.125, N
= 2.1×107

[email protected]
112 V. A. Andrushchenko et al.

In 1991, in the Czech Republic fell one of the largest bolide registered by the
European network—Benešov bolide [11]. It belonged to the intermediate and rela-
tively poorly studied class of cosmic bodies, which are almost completely “burned”
(or rather evaporated) in the atmosphere. Unique observational data including radia-
tion spectra were obtained for this bolide. The fragmentation of the Benešov bolide
was recorded starting from the height of 42 km. Observations [11] showed that rela-
tively small fragments were separated from the main body in the height range of 24
< H < 42 km. At an altitude of 24 km was the final crushing of the bolide, and its
complete extinction took place at an altitude of about 19 km. The dynamic model
of the gross-fragmentation [12] taking into account the crushing of the main body
gave a value for the mass entrance of 80–300 kg. Photometric mass of the bolide [13]
made up of 13,000 kg. The significant discrepancy between estimates of the mass of
the meteoroid calculated using the above model was explained in the following way.
Abnormally, high glow of the bolide was due to the movement of not a single body,
but a compact swarm of fragments with a large total area of the glowing surface.
Moreover, it was assumed that the process of fragmentation of the bolide began at
altitudes H = 60−50 km at pressures (1−5) × 105 N/m2 . Analysis of the dynamics
of the bolide and its radiation suggested that this meteoroid began to split up at high
altitude, about 60 km. Due to the relatively low speed in the direction perpendicular
to the main trajectory, the fragments at high altitude diverged at a distance no more
than the accuracy of photographic measurements. However, this leads to an increase
in the deceleration of the fragmenting body, i.e., to a decrease in the velocity of the
meteor body and its fragments, which, in turn, reduces the heat flow to the surface
and slows the process of ablation.
In work [14], the assessment of the possible expansion of fragments at altitudes
of ~60–40 km showed that due to the low density of air, the fragments did not fly
over long distances, but flew compactly that confirms the established fact that the
observed braking of the bolide belongs to the leading fragment. With a decrease in the
altitude of the flight, the scattering of individual fragments can be significant—up to
300 m. The comparison of the trajectory characteristics of the bolide obtained during
the computational experiment with the observational data showed good corrective
agreement.
A special place among the meteorites is occupied by carbonaceous chondrites,
fragments of which are very rare. Interesting is the study of a meteorite that fell with
an explosion on January 18, 2000 near Tagish lake in northwestern Canada [15]. The
mass of the meteoroid, officially called Tagish lake, reached 200 t before entering
the Earth’s atmosphere. It was found that the rate of entry of the meteoroid into the
atmosphere at an angle of ~16.5° to the horizon was ~15.8 km/s, and the transverse
body size to destruction—from 4 to 6 m. Unlike stone meteorites, the substance, of
which is similar to terrestrial rocks, such bodies are very fragile and resemble “dried
silt”. Entering the dense Earth atmosphere, they just crumble. The found fragments
of Tagish lake are rare specimens of one of the carbonaceous chondrites rich in
volatile substances. Scientists from the Southwest research Institute of the United
States made a sensational statement—they suggested that this meteorite came to us
from Kuiper belt. This is the conclusion they came to after careful consideration of

[email protected]
9 Physical and Kinematic Processes Associated with Meteoroid … 113

Fig. 9.2 The total mass

(curve 1) and kinetic energy
loss per unit length (curve 2)
with respect to altitude of
flight of Tagish lake
meteoroid

the composition of the meteorite. In most cases, meteoroids fall on our planet from
the main asteroid belt. The objects there are mostly composed of rocks and metals,
while the objects from Kuiper belt consist of volatile substances (ammonia, methane,
and water). If the theory is confirmed, the meteorite from Tagish lake will be the most
distant alien to our planet.
In Fig. 9.2, the calculation data showing the change in mass (curve 1) and the
loss of kinetic energy per unit length (curve 2) depending on the altitude of Tagish
lake meteoroid are presented. The model of progressive crushing with the degree of
heterogeneity of the material α = 0.25 was used. As calculations have shown, the
maximum loss of energy, the so-called explosion, happened at a height of H = 40
km with the number of fragments formed is ~105 .
It should be noted that when moving along the trajectory, the mass of the fractur-
ing Tagish lake meteoroid decreased by about 2 times that is the total mass of the
fragments of the meteoroids considered was much higher than the observations. This
fact is also confirmed in the study of motion and destruction and other meteoroids.
Thus, it is necessary to consider other mechanisms of destruction of meteoroids.

9.4 Thermal Stress

It should be noted that at the final stage of motion of meteor bodies, the process of
destruction may continue due to the temperature extension, which does not play a
significant role for large meteor bodies. However, if the sizes of fragments reach sev-
eral centimeters, the emerging temperature gradients may further destroy them to the
size of coarse dust, which rapidly melts and evaporates in air at a high temperature.
The estimates performed in [16] show that for a body with a radius of 10 cm, the
time needed to reach critical tension is ~4 s. Thus, over the time of passing through

[email protected]
114 V. A. Andrushchenko et al.

the atmosphere, such a body may be repeatedly destroyed due to the emerging tem-
perature tension. The calculations show that at the final stage of fragmentation at
α = 0.25, the number of fragments of Sikhote-Alin meteoroid reached 2 × 105 , and
the radius of an individual fragment was 3 cm. Thus, the process initiated by the
temperature extension leads to the additional removal of mass of the fragmenting
meteoroid.
This work analyzes also the influence of nonuniform temperature field on the
meteoroids stress-deformable state during their falling in the Earth atmosphere. As a
result of atmospheric interaction, the warming-up of the meteoroid thin near-surface
layer is formed. At this, if the object is rapidly rotated, its warming-up along the sur-
face will be naturally uniform. If the object is not rotated, then essential nonuniformity
of temperature field will occur, both in depth and over the surface.
To analyze the influence of induced temperature field on the iron meteoroid stress-
deformable state, the following problem is considered. The meteoroid is simulated by
the elastic infinite isotropic cylinder of radius R +h with the warmed-up near-surface
layer of thickness h (h  R). Suppose its velocity is perpendicular to the axis, and
in the first case the cylinder is rapidly rotated around the axis, and in the second
case the cylinder is moving without rotation. The initial equilibrium equations in the
fixed polar coordinates related to the body in partial derivatives for the stresses have
a view of Eq. 9.4, where σr , σϕ , σr ϕ are the components of the stress tensor.

∂σr 1 ∂σr ϕ σr − σϕ ∂σr ϕ 2 1 ∂σϕ

+ + =0 + σr ϕ + =0 (9.4)
∂r r ∂ϕ r ∂r r r ∂ϕ

According to theory of elasticity, the stress tensor components depend on

temperature and gradients of radial U (r, ϕ) and azimuthal W (r, ϕ) displacements:
 
∂U U 1 ∂W ∂U
σr = λ + + + 2μ − βT,
∂r r r ∂ϕ ∂r
   
∂U U 1 ∂W U 1 ∂W
σϕ = λ + + + 2μ + − βT,
∂r r r ∂ϕ r r ∂ϕ
 
1 ∂U ∂W 1
σr ϕ = μ + − W , (9.5)
r ∂ϕ ∂r r

where λ, μ are Lame constants, β = (2λ + 3μ)α1 , where α1 is the material thermal-
expansion coefficient, T is the temperature set by a known coordinate function:

0 < r < R : T = 0,
R < r < R + h : T = T (ϕ),

where T = T (ϕ) is an even function ϕ.

The system of equations (Eqs. 9.4–9.5) is solved at the boundary conditions r =
R (condition of continuity of normal and tangential stresses) and displacements
provided by Eq. 9.6.

[email protected]
9 Physical and Kinematic Processes Associated with Meteoroid … 115

r = R + h : σr (R + h, ϕ) = σr ϕ (R + h, ϕ) = 0 (9.6)

The system of equations (Eqs. 9.4–9.5) with boundary conditions (Eq. 9.6) is
solved by decomposition of the required functions into Fourier series by azimuthal
angle ϕ.
Decomposing the function T = T (ϕ) into a Fourier series, we obtain:
∞

T (ϕ) = T0 + Tn (ϕ) cos nϕ.
n=1

The solutions for the functions U (r, ϕ), W (r, ϕ) are sought in the form:
∞
 ∞

U (r, ϕ) = U0 (r ) + Un (r ) cos nϕ W (r, ϕ) = Wn (r ) sin nϕ.
n=1 n=1

After substitution of Fourier series system of equations (Eqs. 9.4–9.5) is trans-

formed to a system of ordinary differential equations along radial coordinate. Solving
this system and determining arbitrary constants using the boundary conditions, we
determine resulting stresses. Due to the bulkiness of the solution, only the finite for-
mulas for the maximum shear stresses required for the analysis of the solution results
are given below.
Two cases are considered: rapidly rotating cylinder and nonrotating cylinder. In
the first case, at the high speed of the object rotation, the thin near-surface layer will
be naturally heated uniformly in azimuth angle ϕ, viz., function T (ϕ) = θ , where θ
is constant. This means the maximum shear stress along the layer is equal to:

(1) μ
τmax ≈ βθ. (9.7)
(λ + 2μ)

In the second case, with no rotation, the temperature field in the near-surface layer
corresponds to the function T (ϕ) = θ + θ cos ϕ, viz., the maximum temperature is
realized in the head point of the falling object and is equal to 2θ , in the back is equal
to 0, and the average temperature over the surface is equal to θ . This means that the
maximum shear stresses are, respectively, equal to:
μ μ
τmax = βθ + βθ cos ϕ.
(λ + 2μ) (λ + 2μ)

The largest value of the maximum shear stress is reached in this case at ϕ = 0◦
and is equal to:

(2) 2μ
τmax ≈ βθ. (9.8)
(λ + 2μ)

[email protected]
116 V. A. Andrushchenko et al.

Suppose, to be definite, the meteoroid is an isotropic stone cylinder, for which

Lame constants are equal to: λ ≈ 22.5 × 109 N/m2 , μ ≈ 28.7 × 109 N/m2 , the
temperature expansion coefficient α1 ≈ 0.74×10−5 (grad)−1 and the strength critical
value σ ∗ ≈ 1.6 × 108 N/m2 [17].
In the case of the exceeding by any of the stresses of the strength critical value in
some points of the moving meteoroid, its destruction occurs in them. At the warm-up
of the cylinder near-surface layer, simulating the stone meteoroid to the temperature
θ = 500 ◦ C reached at the most moderate values of the air density in the atmosphere
at its initial falling velocity (hypervelocity) that means at sufficiently high altitudes,
the value of the maximum shear stress, in case of the cylinder rapid rotation (Eq. 9.7),
(1)
is equal τmax ≈ 1.8× 108 N/m2 . Thus, the exceeding by the maximum shear stress a
(1)
critical value of the material strength parameter of the stone meteoroid τmax > σ ∗ is
on the whole warmed-up near-surface layer h of the cylinder, and, so, the destruction,
delamination, and peeling near the surface are realized [18]. This effect is repeated
multiply; therefore, the external surface of the falling meteoroid takes a relatively
smooth form.
Now consider the case of the cylinder meteoroid falling without rotation (Eq. 9.8).
In the stone meteoroid, the maximum shear stress exceeds the strength critical value
(2)
τmax ≈ 3.6 × 108 N/m2 > σ ∗ , firstly, in the head point of the warmed-up near-
surface layer h at ϕ = 0◦ and in this point, the local cavity appears. This leads to the
local erosion of the cylinder flown by the hypervelocity air that creates Taylor-Görtler
flow instability [19] even for the single rough element. Following this instability in
the boundary layer of the incoming flow to the roughness areas, Görtler vortices are
formed [20], which are tornado-type vortices rotating at a great speed [13]. These
vortices increase substantially the pressure and heat exchange on surface of cylinder
that leads to intensification of local destructions of the falling meteoroid with the
rhegmaglypts formation on its surface [21]. During the rhegmaglypts formation (new
cavities), the tornado genesis process of Görtler vortices is intensified incrementally,
distributing from the head upstream part of the meteoroid-cylinder over its upper and
lower side parts to the back—leeward part.
Therefore, the calculations and analysis of the received results confirm the fact
that according to the type of the surface relief of the fall meteorites it is possible
to estimate the kinematics nature of the relative falling meteoroids. Their relatively
smooth surface corresponds to the fall of the rapidly rotating objects, and the surface
specked with rhegmaglypts corresponds to nonrotating ones.
In Fig. 9.3, there are photos of the corresponding stone meteoroids illustrating
these assertions.

[email protected]
9 Physical and Kinematic Processes Associated with Meteoroid … 117

Fig. 9.3 Reliefs of meteorite fragments: a Kunashak (1949) [22], b Chelyabinsk (2013) [23]

9.5 Conclusions

Thus, the main processes occurring during the motion from meteor bodies in the
Earth’s atmosphere have been considered. We have studied the mechanisms of
destruction of the bodies due to thermal stresses. The results obtained qualitatively
correctly reflect the process of destruction of bodies in the atmosphere. We have
theoretically studied the mechanisms of destruction of the bodies due to thermal
stresses. The results obtained qualitatively correctly reflect the observed processes
of destruction of bodies in the atmosphere.

References

1. Hughes, D.W.: Meteorite falls and finds: some statistics. Meteorites 19, 269–281 (1981)
2. Syzranova, N.G., Andrushchenko, V.A.: Simulation of the motion and destruction of bolides
in the Earth’s atmosphere. High Temp. 54(3), 308–315 (2016)
3. Andrushchenko, V.A., Syzranova, N.G., Shevelev, Yu.D.: An estimate of heat transfer to blunt
bodies moving with hypersonic velocity in the atmosphere. J. Appl. Math. Mech. 71(5), 747–
754 (2007)
4. Murzinov, I.N.: Laminar boundary layer on a sphere in hypersonic flow on a equilibrium
dissociating air. Fluid Dyn. 1(2), 131–132 (1966)
5. Apshtein, E.Z., Vartanyan, N.V., Sakharov, V.I.: Distribution of radiant heat flux over the surface
of three-dimensional and axisymmetric bodies in a supersonic ideal-gas flow. Fluid Dyn. 21(1),
78–83 (1986)
6. Korobeinikov, V.P., Shurshalov, L.V., Vlasov, V.I., Semenov, I.V.: Complex modeling of the
Tunguska catastrophe. Planet. Space Sci. 46(2/3), 231–244 (1998)
7. Fisher, D.E.: “Ages” of the Sikhote Alin iron meteorite. Science 139(3556), 752–753 (1963)
8. Mukhamednazarov, S.: Observation of a fireball and the fall of the first large meteorite in
Turkmenistan. J. Astron. Lett. 25(2), 117–118 (1999)
9. Emel’yanenko, V.V., Popova, O.P., Chugai, N.N., Shelyakov, M.A., Pakhomov, Yu.V., Shus-
tov, B.M., Shuvalov, V.V., Biryukov, E.E., Rybnov, Yu.S., Marov, M.Ya., Rykhlova, L.V.,
Naroenkov, S.A., Kartashova, A.P., Kharlamov, V.A., Trubetskaya, I.A.: Astronomical and
physical aspects of the Chelyabinsk event. Solar Syst. Res. 47(4), 240–254 (2013)

[email protected]
118 V. A. Andrushchenko et al.

10. Weibull, W.: A statistical theory of the strength of materials. Proc. Roy Swedish Inst. Eng. Res.
151, 1–45 (1939)
11. Spurny, P.: Recent fireballs photographed in central Europe. Planet. Space Sci. 42(2), 157–162
(1994)
12. Ceplecha, Z., Spurny, P., Borovička, J., Keclikovd, J.: Atmospheric fragmentation of mete-
oroids. Astron. Astrophys. 279, 615–626 (1993)
13. Borovička, J., Spurny, P.: Radiation study of two very bright terrestrial bolides and an
application to the Comet S-L 9 collision with Jupiter. Icarus 121, 484–510 (1996)
14. Andrushchenko, V.A., Maksimov, F.A., Syzranova, N.G.: Simulation of flight and destruction
of the Benešov bolid. Comput. Res. Model. 10(5), 605–618 (2018)
15. Grossman, J.N.: A meteorite falls on ice. Science 290, 283–285 (2000)
16. Andrushchenko, V.A., Syzranova, N.G., Shevelev, Yu.D., Goloveshkin, V.A.: Destruction
mechanisms of meteoroids and heat transfer to their surfaces. Math. Model. Comput. Simul.
8(5), 506–512 (2016)
17. Slyuta, E.N.: Physical and mechanical properties of stony meteorites. Sol. Syst. Res. 51(1),
64–85 (2017)
18. Kholin, N.N., Goloveshkin, B.A., Andruschenko, V.A.: Mathematic Simulation of the Wave
Activity in Condensed Medium and Meteoroids Dynamics. Lenand Publ., Moscow (in Russian)
(2016)
19. Chuvakhov, P.V., Borovoy, V.Y., Egorov, I.V., Radchenko, V.N., Olivier, H., Roghelia, A.:
Effect of small bluntness on formation of Görtler vortices in a supersonic compression corner
flow. J. Appl. Mech. Tech. Phys. 58(6), 975–989 (2017)
20. de la Chevalerie, D.A., Fontenean, A., de Luca, L., Cardone, G.: Görtler-type vortices in
hypersonic flows: the ramp problem. Exp. Therm. Fluid Sci. 15(2), 69–81 (1997)
21. Laganelli, A.L., Nestler, D.E.: Surface ablation patterns: a phenomenology study. AIAA J. 7,
1319–1325 (1969)
22. Vernadsky Geological Museum of the RAS. https://ok.ru/sgm_ran/topic/69255458122647.
Accessed 05 Oct 2019 (in Russian)
23. Vernadsky Geological Museum of the RAS. https://i.ytimg.com/vi/fKbt9Uxb8hI/
maxresdefault.jpg. Accessed 05 Oct 2019 (in Russian)

[email protected]
Chapter 10
Computational Modeling of Rarefied
Plasma and Neutral Gas Effusion
into Vacuum

Vadim A. Kotelnikov and Mikhail V. Kotelnikov

Abstract Physical, mathematical, and computational models of neutral gas and rar-
efied plasma effusion into vacuum space have been considered in this chapter. Gas
or plasma effusion is viewed as the process in which they escape from a container
through a relatively small hole shaped like a narrow long slot. Results of computa-
tional experiments have been provided: the distribution functions of charged particles
and neutral gas, their velocity field and concentration field. The impact of latitudinal
magnetic field on the distribution functions and momentums of charged particles
effusion has been studied. Parameters of evolutionary processes have been consid-
ered, from the moment, when the slot is formed through the transition of the effusion
process into stationary mode.

10.1 Introduction

Effusion is an outflow of neutral or ionized gases through relatively small holes

and is commonly found in various fields of modern technology. As spacecraft travel
through the Earth’s ionosphere or the space, they are at a risk of depressurization due
to the accidents, defects, welded seam deterioration, collisions with meteors or bits
of space junk, and a number of other causes.
Designing innovative aerospace engines, e.g., pulsed plasma thrusters, involves
designing vacuum benches, including the vacuum chambers, vacuum pumps, vacuum
furnaces, electric vacuum devices, vacuum insulation, and a number of other devices.
Any depressurization in those devices causes the effusion of neutral and ionized
gases that leads to their malfunction.

V. A. Kotelnikov (B) · M. V. Kotelnikov

The A.A. Blagonravov Institute of Machine Science of the RAS, 4 Maly Kharitonyevsky Per,
Moscow 101990, Russian Federation
e-mail: [email protected]
Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation

© Springer Nature Singapore Pte Ltd. 2020 119

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_10

[email protected]
120 V. A. Kotelnikov and M. V. Kotelnikov

This chapter presents the results of research of the neutral gas and plasma effusion
into vacuum space found by computational modeling using the kinetic theory. We
assume that the dimensions of the slot, through which a gas or plasma effuse into
vacuum space, are considerably shorter than free paths of particles effusing.
There exist multiple works dedicated to studies of gas effusion into vacuum
through relatively small holes [1–8]. However, we failed to find any work based
on the computational solutions referred to Vlasov kinetic equation [9, 10]. As we
studied plasma flows, we supplemented the above kinetic equation with Poisson’s
equation to find the self-consistent electrical field included in Vlasov equation.
Next, the chapter will consider the physical, mathematical, and computational
models of the problem, as well as, the results of computational experiments.
Section 10.2 discusses the physical, mathematical, and computational models.
Methodical calculations are described in Sect. 10.3. Results of computational
experiments are presented in Sect. 10.4, while conclusions are given in Sect. 10.5.

10.2 Physical, Mathematical, and Computational Models

of the Problem

Problems pertaining to the plasma and neutral gas effusion have a lot in common,
as they both use Vlasov kinetic equation. Both of the problems prove to be four-
dimensional problems in phase space. The geometrical shape of the effusion hole is
a narrow rectangle. It is assumed that there exists vacuum space within the compu-
tational domain at the initial time moment. The Distribution Functions (DFs) on the
cross-section of a hole are in both cases assumed to Maxwell–Bolztmann distribution
functions. Computational models of both of the problems are similar. Nonetheless,
there are substantial differences between them also, to name the principal ones:
• Neutral gas models use only one Vlasov equation, while plasma models use two
Vlasov equations, each for ions and electrons, respectively.
• Mathematical models of plasma are supplemented with Poisson’s equation for a
self-consistent electrical field.
• Scale systems used for non-dimensionalization of gas models differ from those of
plasma problems.
Here, the rarefied gas or rarefied plasma is considered that effuses from a tank
having a volume V through a small hole and into vacuum space, the hole’s dimensions
being shorter than the lengths of free paths of particles effusing. The gas or the
plasma contained in the tank is considered as being in a state of equilibrium. The
tank wall thickness is neglected. No particle collisions are assumed to occur within
the computational domain. The latter case can be often found in practice.
In a general case, the problem in question is a six-dimensional in phase space
(x, y, z, vx , vy , vz ) and non-stationary problem [11]. We propose considering the hole
as a rectangle having one side much longer than the other. Due to symmetry shear,
this modeled shape allows considerable reduction of the problem’s dimensions (in the

[email protected]
10 Computational Modeling of Rarefied Plasma and Neutral Gas … 121

Fig. 10.1 Geometry of the

problem

latter case, the problem depends on x, y, vx , vy in phase space) [10], and establishing
all the principal parameters of effusion phenomenon. The slot-shaped hole used
in our model is commonly found in real life and comes in the shape of cracks in
aircraft shells or in housings of vacuum devices. Please refer to Fig. 10.1 to see the
geometry of the problem, where d is the width of the hole, x 1 and x 2 are the respective
coordinates of the edges of the slot, x ∞ and y∞ are the coordinates of the outer edges
of the computational domain.
First and foremost, let us formulate a mathematical model of plasma effusion into
vacuum space. The system of equation is written down as shown below [10]:
 
∂ fα ∂ fα ∂ fα qα   ∂ fα   ∂ fα
+ vx + vy + E x + vy B + E y − vx B = 0, (10.1)
∂t ∂x ∂y mα ∂vx ∂v y
∂ 2ϕ ∂ 2ϕ 1 
+ 2 =− qα n α , E = −∇ϕ, α = i, e, (10.2)
∂x 2 ∂y ε0
  21 +∞ +∞
2kTα  
nα = f α x, y, vx , v y , t dvx dv y . (10.3)
mα
−∞ −∞

The mathematical model includes Vlasov equation for ions and electrons
(Eq. 10.1), Poisson’s equation for the self-consistent electrical field, a formula to
link strength and potential (Eq. 10.2), and a formula to link plasma concentrations
with component distribution (Eq. 10.3). Here, t is the time, E and ϕ are the strength
and potential of the electrical field, respectively, E x and E y are the strength compo-
nents along axes OX and OY, and f α , qα , mα , and nα are the distribution function, the
charge, the mass, and the concentration of charged particles, respectively. The index
i denotes ions and e denotes electrons.
Let us consider the system of initial and boundary conditions. It is assumed
that there exists vacuum space within the computational domain at the initial

[email protected]
122 V. A. Kotelnikov and M. V. Kotelnikov

time moment. The distribution function at the hole’s cross-section (at the “inflow”
boundary) is written down using Eq. 10.4.

f α = (n 0 /π )(m α /(2kTα ))3/2 exp[−m α vx2 + v2y /(2kTα )] (10.4)

Equation 10.4 is Maxwell function for ions and electrons. Here, n0 is the con-
centration of charged particles contained in plasma at the hole’s cross-section. The
plasma that effuses from the hole is assumed to be quasi-neutral, with the potential
at the hole’s cross-section assumed to be zero. “Softer” boundary conditions were
stipulated for the rest of the boundaries of the computational domain (the “outflow”
boundaries) and found through extrapolation of plasma parameters from adjacent
computational layers.
System (Eqs. 10.1–10.4) was reduced to a dimensionless form by means of a
system of scales below [9–11], where M n is the concentration scale, M n = n0 , M L
 1 2
is the length scale, M L = r D = ε0 kTi∞ n ∞ e2 / , Mϕ is the potential scale,
 1/2
Mϕ = kT |qe |, MVα is the velocity scale, MVα = 2kTα m α , α = i, e, M E is
the electrical field strength scale, M E = Mϕ /M L , M B is the electrical field induction
scale, M B = 2M E /MVi .
Here, rD is Debye length, and ε0 is the electrical constant. The rest of the scales
are found from dimension formulas. Dimensionless parameters are obtained as the
system is reduced to a dimensionless form, on which the solution to the problem
hinges:

r0 = r p /M, ϕ0 = ϕp /Mϕ , ε = Ti /Te , B0 = B/M B . (10.5)

The computational model of the problem is based on the iterative method, whereby
a transitional process from the initial to the final stationary state (finding the plasma
parameter distribution within the computational domain) is modeled. The method of
characteristics was used to solve Vlasov equations [12], while Poisson’s equations
were solved by means of spectral methods [11].
Now let us proceed to the mathematical model of neutral gas effusion into vacuum
space. The collisionless gas is expressed by Vlasov equation [9, 10] provided by
Eq. 10.6.

∂f ∂f ∂f
+ vx + vy =0 (10.6)
∂t ∂x ∂y

The gas particle concentration is found from Eq. 10.3, where α determines the
gas molecule concentration.
The boundary distribution functions at the hole’s cross-section (the “inflow”
boundary) are described by Eq. 10.7.

n0 m 3
2 m  2 
f zp = exp − vx + v2y (10.7)
π 2kT 2kT

[email protected]
10 Computational Modeling of Rarefied Plasma and Neutral Gas … 123

“Softer” boundary conditions were stipulated for the rest of the boundaries of the
computational domain (the “outflow” boundaries) and found through extrapolation
of plasma parameters from adjacent computational layers.
The calculation formulas for mean velocities of gas particles are as follows:

+∞ +∞ +∞ +∞
vx f dvx dv y v y f dvx dv y
−∞ −∞ −∞ −∞
vx mean = +∞ +∞
, v y mean = +∞ +∞
. (10.8)
f dvx dv y f dvx dv y
−∞ −∞ −∞ −∞

The formula for flow of particles from the hole is calculated as shown below:

  21 x2 +∞ +∞

2kT
J f r om ori f ice = v y f boundar y d xdvx dv y . (10.9)
m
x1 −∞ 0

In Eq. 10.9, x 1 and x 2 are the coordinates of the edges of the effusion hole.
The formula for a flow of particles over the outer boundary of the computational
domain can be written in a view of Eq. 10.10.

  21 x∞ +∞ 0
2kT  
Jouter boundar y = v y f t, x, y0 , vx , v y d xdvx dv y +
m
0 −∞ −∞
  1  ∞ +∞ +∞
x
2kT 2  
+ v y f t, x, y∞ , vx , v y d xdvx dv y +
m
0 −∞ 0
  1 y∞ 0 +∞
2kT 2  
+ v y f t, x0 , y, vx , v y dydvx dv y +
m
0 −∞ −∞
  1 y∞ +∞ +∞
2kT 2  
+ v y f t, x∞ , y, vx , v y dydvx dv y (10.10)
m
0 0 −∞

The mathematical model was reduced to a dimensionless form using a system of

scales, where M n is the concentration scale, M n = n0 , M L is the length scale, where d
is the short side of the rectangle, see Fig. 10.1, M L = d, M V is the velocity scale, M V
= (2RT /μ)1/2 , where R = 8.314472 J/(moles·K) is the gas constant, μ is the molar
mass of gas, M f is the distribution function scale, M f = M n /(M v )3 , M t is the time
scale, M t = M L /M v , M N is the scale of flow of particles per unit of the slot length,
MN = MnMvML .
As with plasma effusion, the computational model of the problem is based on
the iterative method, whereby a transitional process from the initial state to final
stationary state, during which we obtain the plasma parameter distribution within

[email protected]
124 V. A. Kotelnikov and M. V. Kotelnikov

the computational domain, is modeled. The method of characteristics algorithm was

used to solve Vlasov equation [12].

10.3 Methodical Calculations

It is worth noting that the problem to be solved was multidimensional and non-
stationary, and, in the case of plasma effusion, it also contained a considerable num-
ber of parameters, which resulted in heavy consumption of computational resource.
The computational domain, in the case of plasma effusion, contained 3,075,200 cells
of the computational mesh, the time step being 0.002 dimensionless units. The com-
putation time was approximately 10 h for desktops PCs, each having a quad-core
processor, with the clock rate of each core being 3.2 GHz, and the RAM being 3 GB.
The computation time was controlled visually, from a control page displayed
throughout the computation, on which the diagram of ionic and electronic currents
against time, inflowing from the hole and into the computational domain, and flowing
beyond its limits through outer boundaries, was being plotted and replotted at each
time step. The calculation was stopped whenever the ionic current flowing from the
hole became practically equal to the ionic current flowing through the limits of the
computational domain. Please refer to Fig. 10.2a, b for respective diagrams.
Those diagrams suggest that the calculation was stopped in 3.5 dimensionless
time units after the ionic current flowing through the computational domain limits
had become practically equal to the ionic current flowing from the nozzle, while
the electronic current flowing through the limits of the computational domain had
become practically equal to the electronic current flowing from the nozzle much
earlier than that.
In the case of neutral gas effusion, the computational algorithm was implemented
as a program in C++, with tools of OpenGL API used. The source code of the program
was designed using Visual Studio 2017 IDE. The dimensions of the computational

Fig. 10.2 Evolutions of ionic and electronic currents, as well as, of a flow of neutral gas particles
through the limits of the computational domain: a curve 1—electronic current flowing from the
hole and curve 2—the electronic current flowing from the computational domain, b curve 3—the
ionic current flowing from the hole and curve 4—the ionic current flowing from the computational
domain, c curve 5—a flow of gas particles flowing from the hole into the computational domain
and curve 6—a flow of gas particles flowing from the computational domain

[email protected]
10 Computational Modeling of Rarefied Plasma and Neutral Gas … 125

domain were 2 × 2 and contained 39,942,400 cells of the computational mesh, the
time step being 0.01 dimensionless units. The computation time was several hours
for a desktop PC having Intel Core i7-6700 K quad-core processor, the clock rate of
each core being 4 GHz, and the PC’s RAM being 32 Gb.
The computation stop time was controlled visually, from a control page displayed
throughout the computation, on which the diagram of gas particles flow from the
hole into the computational domain, and beyond its limits through outer boundaries,
was being plotted and replotted at each time step. Please refer to Fig. 10.2c above
for the diagram.
A moment, when a gas particle flow from the hole became practically equal to
the gas flow through the limits of the computational domain, was spotted. Next, the
following was visually observed and controlled on the monitor:
• The constant value of the gas particle flow crossing the computational domain,
which was indicative of the establishment of a stationary solution to the problem.
• Practically, full similarity was in between the gas flow from the hole into the
computational domain and the gas particle flow from the computational domain,
which indicated that the law of conservation of the mass of the gas worked within
the computational domain in an established stationary state.
Those two conditions having been met, the calculation was stopped with the
results being subject to further analysis.

10.4 Results of Computational Experiments

Distribution functions of charged and neutral particles at various points of the region
of interest, as well as, the momentums of those functions (the concentrations and
velocities fields) were found through computational experiments.
First, let us present the results for plasma effusion. Figure 10.3a, b, and c shows
isometric lines of the functions of distribution of ions recorded in the vicinity of
the hole’s cross-section at the time moment of 3.5 dimensionless units, which is

Fig. 10.3 Isometric lines of distribution functions of ions: a at the left-hand edge of the hole’s
cross-section, b in the middle of the hole, c at the right-hand edge of the hole’s cross-section

[email protected]
126 V. A. Kotelnikov and M. V. Kotelnikov

in accord with the established solution to the problem. The distribution function
shown in Fig. 10.3a was recorded at the left-hand edge of the hole, where there are
no ions traveling at positive-value V x or negative-value velocities V y . Accordingly,
the distribution function shown in Fig. 10.3c is determined by no ions traveling at
positive-value V x or negative-value velocities V y at the left-hand edge of the hole.
The distributions functions of ions in the vicinity of the middle of the cross-section
of the hole are shaped like a half-dome of Maxwells’ distribution function, and the
whole of it lies within the positive-value velocities region V y , while ions traveling at
negative-valued velocities V y are not present here either.
The calculations suggest that the distribution functions of electrons near the hole’s
cross-section have the same parameters as ions.
Figure 10.4a, b shows velocity fields of electrons and ions, respectively, their
respective vectors of mean velocities are plotted from the centers of cells of com-
putational meshes. The arrows of the vectors were omitted to avoid cluttering the
diagram with lines. Vectors of mean velocities for their respective plasma compo-
nents were only plotted wherever the condition ni ,e > 0.1 n0 was met, where n0 is
the concentration of charged particles at the time moment t = 0. The regions, where
the condition in question was met, are highlighted in light-gray, while the hole’s
cross-section is highlighted in dark-gray. The thermal velocity of ions was selected
as their velocity scale, while the thermal velocity of electrons was selected as their
velocity scale.
The above diagrams suggest that the effusion velocity of the electronic component
from the hole is much higher than that of the ionic component. As a result, a negative
volumetric charge forms within the computational domain at the initial moment of
evolution. The calculations suggest that the charge in question gradually decreases
in the process of evolution.
Figure 10.5a, b shows the isometric lines of concentrations of electrons and ions,
respectively. The parameters of the diagram in question are the same as those of
Fig. 10.4 above. Moreover, Fig. 10.5a suggests that the concentration of electrons near
the hole’s cross-section exceeds 0.5 dimensionless units. The calculation suggests
that the value exceeds that of the concentration of ions in the vicinity of the hole’s

Fig. 10.4 Velocity fields of plasma components: a velocity field of electrons, t = 0.032, b velocity
field of ions, t = 1.23

[email protected]
10 Computational Modeling of Rarefied Plasma and Neutral Gas … 127

Fig. 10.5 Isometric lines of concentrations of plasma components: a electrons, t = 0.032, b ions,
t = 1.23

cross-section at the same time moment. Thus, a low negative charge is present there,
and it forms because electrons that flow from the hole repel from the volumetric
negative charge located in the middle of the computational domain and concentrate
in the vicinity of the hole’s cross-section.
The impact of the axial magnetic field Bz on plasma effusing from the hole has also
been studied in this chapter. Figure 10.6a shows the velocity field of the electronic
component with the dimensionless magnetic induction value being B0 = 0.03. The
impact of Lorentz force on the electronic component can be clearly seen in Fig. 10.6a.
The calculations suggest that the above value of magnetic induction has no decisive
influence on the ionic component of plasma. Figure 10.6b shows the velocity field
of the ionic components with the dimensionless magnetic induction value being
B0 = 1.5. Here, the impact of the magnetic field on the ionic component becomes
substantial. Electrons, as they enter the computational domain, become magnetized,
and a positive volumetric charge forms within the computational domain.
Now, let us present the results of modeled effusion of neutral gas. Figure 10.7

Fig. 10.6 Velocity field of plasma when an axial magnetic field is present: a electrons, B = 0.03,
b ions, B = 1.5

[email protected]
128 V. A. Kotelnikov and M. V. Kotelnikov

Fig. 10.7 Distribution function of gas particles versus coordinate y (x = 5, t = 30): a y = 0.025,
b y = 1.5, c y = 3

shows the distribution function along the symmetry axis of the gas flow at the moment,
when solution is established depends on the coordinate y.
The dependence in question has been visualized in Fig. 10.8 in the shape of
isometric lines.
The above diagrams suggest that the distribution function changes its shape
when sheared from the hole to the boundary of the computational domain along
the symmetry axis of the flow.
The computational experiments suggest that any change in the shape of the dis-
tribution function due to its shift along axis OY causes a slight shift in its “center-
of-gravity” toward an increase in the component. As the “center-of-gravity” of the
distribution function corresponds to the vector of mean velocity of particles at a point
under study, the mean velocity of particles within the flow increases with increasing

Fig. 10.8 Isometric lines of distribution functions of gas particles versus coordinate y (x = 5, t =
30): a y = 0.025, b y = 1.5, c y = 3

[email protected]
10 Computational Modeling of Rarefied Plasma and Neutral Gas … 129

their distance from the hole, a phenomenon found through calculation, and is shown
in Fig. 10.9.
Figure 10.10 shows the distribution function versus the coordinate x as it shears
from the symmetry axis of the flow to the lateral boundary of the computational
domain. Those distribution functions were obtained on the 20th computational layer
from the hole.
Figure 10.11 visualizes the distribution functions from Fig. 10.10 in the shape of
isometric lines.
Figure 10.12 shows the concentration field of particles within the computational
domain, which was found through computational experiments. The diagram contains
the computational mesh, and the boundaries between the gray regions are isometric
lines. The concentration field has an axial symmetry.
The evolution of distribution of gas particles concentrations on axis OY along the
symmetry axis of the flow at various time moments is shown in Fig. 10.13. Due to

Fig. 10.9 Mean velocities of particles versus coordinate y along the symmetry axis of flow (t =
30)

Fig. 10.10 Distribution function of gas particles versus coordinate x (y = 1.5, t = 30): a x = 5,
b x = 5.75, c x = 6.5

[email protected]
130 V. A. Kotelnikov and M. V. Kotelnikov

Fig. 10.11 Isometric lines of distribution of gas particles versus coordinate x (y = 1.5, t = 30):
a x = 5, b x = 5.75, c x = 6.5

Fig. 10.12 Isometric lines of concentrations of gas particles within the computational domain:
a t = 1.5, b t = 30

dissipation of the effusion flow, the concentration of gas particles decreases smoothly
as the coordinate y increases.
Figure 10.14 shows a field of velocities of gas particles at the moment when
the flow establishes. The arrows of the vectors of mean velocities of particles were
omitted to avoid cluttering the diagram. The vectors of velocities were only plotted
at the nodes of the computational mesh, where the concentration of particles was n

[email protected]
10 Computational Modeling of Rarefied Plasma and Neutral Gas … 131

Fig. 10.13 Evolution of

distribution of gas
component concentration
along axis OY (x = 1), curve
1—t = 0.2, curve 2—t =
0.9, and curve 3—t = 10

Fig. 10.14 Velocity field of

gas particles (t = 30)

> 0.1n0 . The respective regions were highlighted in gray. The velocity field has an
axial symmetry. The dissipation of the flow increases toward the edges of the hole.

10.5 Conclusions

The research into gas and plasma effusion through a hole shaped as a long rectangle
provides a clear idea of the distribution functions, velocity fields, concentration fields,
potential of the self-consistent electrical field, and other parameters of effusing flows
of particles. The originally designed program codes allow further research into the
area, while taking into account

[email protected]
132 V. A. Kotelnikov and M. V. Kotelnikov

• The shape of the effusion hole.

• The thickness of the walls and material of the volume, from which effusion occurs.
• The ratio between the inner and outer pressures.
• The impact of self-fields and outer fields on effusion.
The research results presented here may be of use to designers of portable small
leak detectors for space stations and vacuum plants, in mass spectroscopy, and a
number of other application areas.

References

1. Saksagansky, G.L.: Molecular Flows in Complex Vacuum Structures. Atomizdat Publishing

House, Moscow (in Russian) (1980)
2. Rozanov, L.N.: Vacuum Technique. CRC Press, London, New York (2018)
3. Steckelmacher, W.: A review of the molecular flow conductance for systems of tubes and
components and the measurement of pumping speed. Vacuum 16(11), 561–584 (1966)
4. Semkin, N.D., Zanin, A.N., Piyakov, I.V., Voronov, K.E.: A time-of-flight mass spectrometer
for detecting a point of air leakage from a spacecraft. Instrum. Exper. Techn. 50(1), 108–112
(2007)
5. Holland, S.D., Roberts, R., Chimenti, D.E., Michael Strei, M.: Leak detection in manned
spacecraft using structure-borne noise. Appl. Phys. Lett. 86(17), 70–78 (2005)
6. Holland, S.D., Chimenti, D.E., Roberts, R., Michael Strei, M.: Locating air leaks in manned
spacecraft using structure-borne noise. J. Acoust. Soc. Am. 121(6), 3484–3492 (2007)
7. Nesterov, S.B., Astashina, M.A., Neznamova, L.O., Vasilyev, Y.K.: Problems and methods of
studies of rarefied gas medium near spacecraft. Vacuum Dev. Technol. 18(3), 183–186 (2007)
8. Rozanov, L.N., Skryabnev, A.Y.: Gas flow through circular duct at high pressure drops. Vacuum
Dev. Technol. 20(1), 3–8 (2010)
9. Kotelnikov, M.V.: The distribution functions of charged particles in the vicinity of a cylindrical
body in a flow of collisionless plasma in magnetic field. High Temp. 46(6), 757–762 (2008)
10. Kotelnikov, V.A., Kotelnikov, M.V.: Current–voltage characteristics of a flat probe in a rarified
plasma flow. High Temp. 54(1), 20–25 (2016)
11. Kotelnikov, V.A., Kotelnikov, M.V., Gidaspov, V.Y.: Computational modeling of flows of
collisional and collisionless plasma around body. FizMatLit, Mosow (in Rissian) (2010)
12. Kotelnikov, V.A., Kotelnikov, M.V.: An advanced method of characteristics. Matem. Mod
29(5), 85–95 (in Russian) (2017)

[email protected]
Chapter 11
Numerical Simulation of the Process
of Phase Transitions in Gas-Dynamic
Flows in Nozzles and Jets

Igor E. Ivanov , Vladislav S. Nazarov , Vladimir Yu. Gidaspov

and Igor A. Kryukov

Abstract The chapter presents a development of condensation and evaporation in

flows of two-phase gas-droplet mixture in the nozzles, jets, and external area in front
of the nozzle. Condensation of pure water vapor and condensation vapor into wet
stream mixture flow are considered. Two different models for modeling condensation
process are used. One of them is a quasi-chemical method. Another method is Method
Of Moments (MOM). Also, the task of gas mixture jump from metastable state to
stationary state and the task of flow of superheated steam are reviewed.

11.1 Introduction

Condensation process contains a lot of natural phenomena and modern technical

applications. Condensate can be formed in rarefaction areas around aerodynamic sur-
faces, when they maneuver in the engine blade or rocket nozzles, when gas intensely
expands, and so on. Condensation can lead to a shock wave that influences on the
airships aerodynamics, engine, and technical devices parameters. Sometimes in a

I. E. Ivanov · V. S. Nazarov (B) · V. Yu. Gidaspov

Moscow Aviation Institute (National Research University), 4, Volokolamskoe Highway, Moscow
125993, Russian Federation
e-mail: [email protected]
I. E. Ivanov
e-mail: [email protected]
V. Yu. Gidaspov
e-mail: [email protected]
I. E. Ivanov
Lomonosov Moscow State University, 1, Leninskie Gory, Moscow 119991, Russian Federation
I. A. Kryukov
Ishlinsky Institute for Problems in Mechanics of the RAS, 101-1, Pr. Vernadskogo, Moscow
119526, Russian Federation
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 133

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_11

[email protected]
134 I. E. Ivanov et al.

part of airship, where an expansion fan is located, the ice layer is formed. These phe-
nomenas change the airships aerodynamics to negative level. In some applications,
condensation is negative but for other applications it is positive. For example, there
are the vacuum deposition surface technology [1], planarization of surfaces with the
help of ion cluster beam [2], natural gas segregation, among others.
Homogeneous condensation process of substance can be divided into two sequen-
tial stages. The first one is nucleation (self-generation droplets) and the second one
is growing of droplets with the help of condensation coatings. Condensation pro-
cess intensity is determined by the saturation parameter S. The saturation parameter
represents the ratio of the vapor pressure of condensing substance to the saturation
pressure at a determined temperature. At S > 1, the formed nucleuses begin to grow
until they reach the chemical and thermodynamic equilibrium with the environment.
This study considers features of the condensation modeling methods. These
methods are used to study a process of vapor condensation in Laval nozzles.
Condensation of gas mixtures components in nozzle have been researched since
the middle of the last century [3, 4]. Calculation techniques of condensation vapor for
stream from supersonic nozzles, discharging to vacuum from sonic nozzle stream,
or discharging a stream to low-pressure ambient space mainly use the macroscopic
modeling method in case of continuum equations, as well as, in modifications of
classical nucleation theory [5–9]. Modern method, which fixes classical nucleation
theory problems close to the formation to saturation parameters area of liquid phase,
is considered in [10, 11].
Currently, microscopic (kinetic) approaches are widely used. In [12–14], a math-
ematical model of Monte-Carlo condensation process was considered. Clusters were
formed as a result of particle collisions. The elastic collision of molecules, recombi-
nation of molecules, association of a cluster and a monomer, association of clusters,
and evaporation of a monomer from a cluster were taken into account.
The quasi-chemical model of condensation is widely used [15, 16]. In the quasi-
chemical cluster model, it is assumed that the pair consists of monomers and molec-
ular aggregates—clusters formed from monomers connected by the forces of molec-
ular interaction. It is assumed that the growth of clusters occurs through the addition
of monomer to them, and their destruction happens through the loss of monomer.
However, two-particle reaction for small clusters is unlikely because there are prob-
lems with the removal of excess heat of reaction. In this case, a cluster growth occurs
at the expense of three-particle reaction. When a cluster becomes large enough to
absorb the impact energy and excess heat, the cluster growth becomes dominant due
to two-particle reactions. The simulation of the gas volume condensation process
can be carried out using the kinetic equations for the droplet size distribution func-
tion [17, 18]. However, a direct solution of the kinetic equations is possible only for
relatively simple model problems.

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 135

Method of moments is based on solving the equations for the moments of the
droplet size distribution function, which have recently become widely used [19–21].
The method requires relatively small computational costs and combines organically
with continuous modeling with Eulerian approach. The use of method of moments
is limited to the case of small droplets, where speed and temperature are not much
different from the corresponding parameters of the gas medium.
This study develops two ways for condensation modeling. The first one is a con-
tinual approach based on MOM. The second one is a kinetic approach based on a
quasi-kinetic model.
The chapter is structured as follows: Section 11.2 presents two methods for con-
densation simulating: method of moments and quasi-chemical model of condensa-
tion. In Sect. 11.3, there is the growth dynamics of clusters test. Also, Sect. 11.3
presents the results of modeling a jet flowing into external area. Section 11.4
concludes the chapter.

11.2 Mathematical Models

In this section, two methods are under consideration: method of moments and quasi-
chemical model of homogeneous condensation represented in Sects. 11.2.1 and
11.2.2, respectively.

11.2.1 Method of Moments

Two-phase substance is the multicomponent gas (carrier gas and vapor of conden-
sate substance) and clusters (droplets) of condensing substance. To construct the
mathematical model, following assumptions are introduced:
• The volume ratio of liquid phase is negligible.
• There is the mechanical and thermal equilibrium between the gas and liquid phase.
• There aren’t collisions between droplets.
The system of Navier–Stokes equations written in a weakly divergent form can
be adopted as a mathematical model of such a two-phase mixture:

∂U ∂(F − Fv ) ∂(G − Gv )
+ + = S, (11.1)
∂t ∂x ∂y

where

[email protected]
136 I. E. Ivanov et al.
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ρ ρu ρv
⎢ ρu ⎥ ⎢ ρu2 + p ⎥ ⎢ ρuv ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ρv ⎥ ⎢ ρuv ⎥ ⎢ ρv2 + p ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ρE ⎥ ⎢ (ρE + p)u ⎥ ⎢ (ρE + p)v ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
U =⎢ ρQ0 ⎥, F = ⎢ ρuQ0 ⎥, G = ⎢ ρvQ0 ⎥,
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ρQ1 ⎥ ⎢ ρuQ1 ⎥ ⎢ ρvQ1 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ρQ2 ⎥ ⎢ ρuQ2 ⎥ ⎢ ρvQ2 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ ρα ⎦ ⎣ ρuα ⎦ ⎣ ρvα ⎦
ραmax ρuαmax ρvαmax
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
0 0 0
⎢ τxx ⎥ ⎢ τyx ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ τxy ⎥ ⎢ τyy ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ uτxx + vτxy − qx ⎥ ⎢ uτxy + vτyy − qy ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
Fv = ⎢
⎢ 0 ⎥, Gv = ⎢
⎥ ⎢ 0 ⎥, S = ⎢
⎥ ⎢ J ⎥,
⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ r∗ J + ṙρQ0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ r∗2 J + 2ṙρQ1 ⎥
⎢ ⎥ ⎢ ⎥ ⎢  3 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
3 πρl r∗ J + 3ṙρQ2
4
0 0
0 0 0

where

∂u 2 ∂u ∂v ∂u ∂v
τxx = μ 2 − + , τyx = τxy = μ + ,
∂x 3 ∂x ∂y ∂x ∂y
∂v 2 ∂u ∂v ∂T ∂T
τyy = μ 2 − + , qx = −λ , qy = −λ .
∂y 3 ∂x ∂y ∂x ∂y

Here, ρ is the density, p is the pressure, T is the static temperature, u is the velocity
along the x-direction, v is the velocity along the y-direction, E is the total energy per
unit volume, μ is the viscosity coefficient, λ is the thermal conductivity coefficient.
First four equations (in Eq. 11.1) describe the dynamics of a two-phase mixture
in a two-dimensional volume, while other equations (the equations of moments)
characterize the evolution of changes in the parameters of the liquid-droplet phase.
Equations from the fifth to the seventh are obtained from the general equation of
dynamics describing the process of nucleation and dynamics of clusters (droplets)
with homogeneous condensation:

∂f ∂ ∂
+ (uf ) + (ṙf ) = δ(r − r∗ )J (11.2)
∂t ∂x ∂r

Here, the function f (x, t, r) is the size distribution function of the droplets, J is
the rate of homogeneous nucleation [20, 22], δ(r − r∗ ) is the delta-function, r∗ is the
droplet critical radius, ṙ is the droplet growth rate.
The derivation of the moment equations consists in the successive multiplication
of Eq. 11.2 by r k (k is the natural number) and integration over a radius in the range
from 0 to ∞. The result is an infinite chain of moment equations, Hill chain [3]:

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 137

∂ ∂
(ρQk ) + (ρUi Qk ) = (r∗ )k ρJ + kρQk−1 ṙ, k = 0, ∞, (11.3)
∂t ∂xi
∞
where ρQn = x∗ r n f (x, t, r)dr is the moments of the nth order.
In this case, for a unit mass of the mixture, the zero moment Q0 equals to the
number density of droplets per unit mass, Q1 is the sum of radii of all clusters, Q2 is
the sum of the squares of the radii of all clusters (estimate of the surface area of all
clusters), Q3 is the sum of cubes of radii of all clusters (estimation of the volume of
all clusters). Instead of the moment Q3 , it is convenient to use the mass fraction of
the liquid phase α:

4π
α= ρl Q3 ,
3
where ρl is the liquid phase density.
Thus, MOM describes the evolution of the liquid phase by a finite number of
moment equations derived from the general equation of the dynamics of the distri-
bution function f (x, t, r) [3]. In this research, a modification of MOM is used, at
which an additional equation is introduced that describes the dynamics of the mass
fraction of the condensing phase αmax (the sum of the mass fractions of droplets
and vapor of the condensing substance). This allows to extend the class of tasks to be
solved, for example, to consider problems with different contents of a condensable
substance in different zones of the computational domain.
Thermodynamics model. The thermophysical properties of the mixture and
equation of state for the mixture (caloric and thermal) are written using Eq. 11.4,
where CV a , CPa are the constant-volume and constant-pressure specific heats, respec-
tively, CV V , CPV are the specific heats of the condensing medium (vapor). CV mixt ,
CPmixt are the specific heats of the two-phase mixture, Cl is the specific heat of the
liquid water, Ra , RV , Rmixt are the individual gas constants of the carrier gas condens-
ing the medium and two-phase mixture, respectively, κf is the adiabatic exponent of
the mixture [19].

CV mixt = (1 − αmax )CV a + αmax CV V + α(Cl − CV V )

CPmixt = (1 − αmax )CPa + αmax CPV + α(Cl − CPV )
Rmixt = (1 − αmax )Ra + αmax RV − αRV )
CPmixt
κf = (11.4)
CV mixt

The caloric and thermal equations of state are mentioned below:

e = (1 − αmax )Cva T + αmax Cvv + α(Cl − Cvv ) + αL0 ,

p
p = ρTRmixt , af2 = κf , L = L1 T + L0 , L1 = CPV − Cl , (11.5)
ρ

[email protected]
138 I. E. Ivanov et al.

where T is the temperature of the mixture, af is the frozen velocity of sound of the
mixture, L is the latent heat of vaporization, e = e(T ) is the mixture internal energy.
The right sides of the moment equations in system (Eq. 11.1) are determined using
the parameters of the classical nucleation theory (J , ṙ = dr dt, r∗ , f (x, t, r)),

qc 2σ ρv2 4π r∗2 σ
J = exp −g ,
(1 + η) π m ρl
3 3 RV mT

1 (1 + η) is the corrective factor taking into account the non-stationarity of the

process [23], qc is the condensation coefficient (qc ≈ 1),

κf − 1 L L 1
η=2 − ,
κf + 1 Rv T Rv T 2

σ = kσ σ∞ , σ∞ is the flat film surface tension, kσ is the correction factor taking

into account the curvature of the drop, g is the nucleation correction factor multiplier,
S = pV pS is the saturation parameter.
To determine the magnitude of the growth rate of a drop, one of the following
models is used for the free-molecular and continual regime of flow of a condensable
substance around a cluster-drop.
One model is Hertz-Knuth model, where dr dt
is defined as

dr β pV − pS,r
= √ , (11.6)
dt ρl 2π RV T

where pS,r is the saturation pressure on the surface of a drop of average radius size,

2σ
pS,r = pS exp ,
ρl RV TrHill

pS is the surface saturation pressure, β is the evaporation coefficient [3],

 
Q2
if α > 10−6 2σ
ρl RV T ln S
if S > 1
rYill = Q0 , r∗ = . (11.7)
0 if α ≤ 10 −6 ∞ if S ≤ 1

dr
Another model is Hill–Young model [3, 24], where dt
is defined as

dr pv Cpv + Cvv
= √ (Ts (pv ) − T ),
dt ρl L 2π Rv T 2

Ts is the saturation temperature.

dr
One more model is Gyarmathy model [22, 25, 26], where dt
is defined as

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 139
 ∗

dr λv (Ts (pv ) − T ) 1 − rrhill
= , (11.8)
dt ρl L · rhill · (1 + 3.18Kn )

where λv is the thermal conductivity of vapor to active substance condensate, Kn is

the Knudsen number characterizing the flow around a drop of steam, Kn = 2rlhill , l is
the free length of the vapor molecule:
√
1.5μv Rv T
l= ,
pv

where μv is the dynamic viscosity of vapor of condensable substance.

The system of Eqs. 11.1–11.8 is solved by an explicit method of control volume
of the second order of accuracy in space, while for approximating inviscid flows
Godunov scheme is used in conjunction with the approximate Advection Upstream
Splitting Method (AUSM) and method for solving Riemann problem, and central-
difference approximation is used for approximating viscous flows.

11.2.2 Quasi-Chemical Model of Homogeneous

Condensation

The clustering process is described by the following, generally speaking, infinite-

dimensional kinetic model. It is assumed that the number of monomers per unit
volume is much more than dimers, trimers, etc. When this condition is fulfilled in the
kinetic mechanism, which determines the course of the whole condensation process,
the addition reactions of monomers to the cluster of the following type prevail:

Ki+(1)
Mi + M1 + ←→ Mi+1 , (11.9)
Ki−(1)

Ki+(2)
Mi + M1 + M1 ←→ Mi+1 + M1 , i = 1, 2, 3, . . . , (11.10)
Ki−(2)

Ki+(3)
Mi + M1 + MA ←→ Mi+1 + MA , (11.11)
Ki−(3)

where M1 , Mi , MA are the symbolic designation of the monomer, i-measure, and inert
gas molecules, Ki+(r) , Ki−(r) (r = 1, 2, 3) are the rate constants of the corresponding
requirements. The plus and minus markers correspond to cases of monomer addition
and detachment.
In accordance with the kinetic mechanism, the change in concentration over time
is described by a system of ordinary differential equations γi :

[email protected]
140 I. E. Ivanov et al.

d γi +(1) +(2) +(3)

ρ = (Ki−1 + Ki−1 ργ1 + Ki−1 ργA )ργ1 ργi−1
dt
− (Ki+(1) + Ki+(2) ργ1 + Ki+(3) ργA )ργ1 ργi − (Ki−1
−(1) −(2)
+ Ki−1 −(3)
ργ1 + Ki−1 ργA )ργi
−(1) −(2) −(3)
+ (Ki + Ki ργ1 + Ki ργA )ργi+1 .

By analogy with chemical kinetics, we can get a connection between the rate
constants of the forward and reverse reactions.
From the equilibrium condition of reactions (Eqs. 11.9–11.11), it follows:
ργi
(Ki−(1) + Ki−(2) ργ1 + Ki−(3) ργA ) = (Ki+(1) + Ki+(2) ργ1 + Ki+(3) ργA )ργ1 .
ργi+1

From the conditions of thermodynamic equilibrium follows the equality of

chemical potentials:

Gi (p, T ) + RT ln xi + G1 (p, T ) + RT ln x1 = Gi+1 (p, T ) + RT ln xi+1 .

It follows that

xi ργi Gi+1 (p, T ) − Gi (p, T ) − G1 (p, T ) − RT ln(γ1 m )

= = exp
xi+1 ργi+1 RT

or if we enter the value

Gi (p, T ) − iG1 (p, T )

εi = exp i ln x1 − ,
RT

then
ργi εi
= .
ργi+1 εi+1

Denoted by

νi = (Ki+(1) + Ki+(2) ργ1 + Ki+(3) ργA )ργ1 ,

we get

d γi εi−1 εi γi−1 γi γi γi+1

= νi−1 γi−1 − νi γi − νi−1 γi + νi γi+1 = νi−1 εi−1 − − νi εi − ,
dt εi εi+1 εi−1 εi εi εi+1

i = 2, 3, . . . , ∞.

Expressions of condensation reaction rates can be calculated by the formulas of

the liquid-drop theory [4, 15]. It should be noted that in practice the systems of
finite dimensionality N obtained from the source system are used by truncating it. In
this case, the system removes the equations for i > N, and in the amounts used, the

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 141

corresponding terms are removed. It is additionally assumed that clusters of a size

larger than N are not formed. In accordance with [4, 15], this system is approximated
by a finite system with dimension N (in real calculations N = 10,000–10,000,000):

d γi εi−1 εi
= νi−1 γi−1 − (νi + νi−1 )γi + νi γi+1 , i = 2, 3, . . . , N .
dt εi εi+1

The concentration of monomers in this case is obtained from the normalization

condition:


N
iγi = const = γ0 , γA = const.
i=1

The system of condensation kinetics equations is approximated by a semi-implicit

difference principle, the sweep method can be solved for each time step, and the
nonlinearity of γ1 is eliminated by the iteration method.
Comments. Also, in numerical modeling, an approximation of the source system
by a suitable finite-dimensional system of a similar type with effective values of
its coefficients is used. The latter are found using a transformation i = n(j) that
translates an infinite interval i = 1, 2, … into a finite j = 1, 2, …, N, N + 1. In this
case, the normalization condition is


N
dn(j)
nj nj γj = γ0 , nj = .
j=1
dj

Thermodynamic parameters of the mixture are found by formulas, in which it is

assumed that infinite sums with the participation of some thermodynamic parameter
Ai are approximated as follows:

∞
 
N
γi Ai ≈ nj γj Aj .
i=1 j=1

All parameters depending on the index j are calculated using the appropriate
formulas for i = n(j).
Thermodynamics model. The mixture of perfect gases and condensable com-
ponent described in terms of the model of perfect gases is considered. The thermo-
dynamic properties of such a mixture can be described by Gibbs specific potential
G:
∞

G(p, T , γ ) = γA (GA (p, T ) + RT ln xA ) + γi (Gi (p, T ) + RT ln xi ),
i=1

[email protected]
142 I. E. Ivanov et al.

∞

xi = γi m , xA = γA m , m = (γA + γi )−1 ,
i=1

N 
N
γA = γ̃i , GA (p, T ) = γ̃i G̃i (p, T )/γA .
i=1 i=1

The parts of the Gibbs molar potentials that are independent of the concentra-
tions for i parameters Gi (p, T ) (I = 1, 2, 3 …) and non-condensable molecules are
determined by the formulas:

Gj (p, T ) = RT ln(p/p0 ) + Gj◦ (T ), j = A, 1, 2, . . .

The thermodynamic properties of the clusters were calculated within the frame-
work of the liquid-drop model in its standard form. In particular, in order to write
Gi◦ (T ) for clusters, i.e., at i = 2, 3, … in the standard reference system of enthalpies,
one can use the expression:

Gi◦ (T ) = i(GL◦ (T ) − G1◦ (T )) + 4π ri2 NA σi (T ).

Here, r is the radius of the droplet which contains n molecules, σi (T) is the surface
tension of the i-measure. A cluster containing i gas molecules with molecular mass
m1 occupies the volume Vi = NAim 1
ρL (T )
= 43 π r 3 . Therefore,

1
3 im1 3
r= ,
4π NA ρL (T )

Gi◦ (T ) = i(GL◦ (T ) − G1◦ (T )) + Bi (T )i2/3 ,

1/3
Bi (T ) = σi (T )(36π )1/ 3 NA (m1 /ρL (T ))2/ 3 , i = 2, 3 . . .

According to Gibbs potential of the mixture, all other thermodynamic parameters

are determined.
Equilibrium distribution function. Consider the state of thermodynamic equi-
librium between the gaseous and liquid phase consisting of drops, including a spec-
ified number of molecules, with a known number of molecules γ0 condensing and
number of inert phase in a kilogram of the mixture molecules γA . We write the equi-
librium conditions between the gas and liquid phases at the phase transition, which
is the equality of the corresponding chemical potentials μi :

∂G
μi = = Gi (p, T ) + RT ln xi ,
∂γi P,T ,γj,j =i

μL = GL (p, T )

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 143

on the saturation curve

GL (pH , T ) = G1 (pH , T ) + RT ln x1 .

This equation contains three unknown parameters: the pressure, temperature, and
mole fraction of monomers in the mixture. The cluster size distribution function in
the state of thermodynamic equilibrium at the phase transition, in accordance with
the used model of thermodynamics and condensation kinetics, should be defined
from the following conditions:

G2 (pH , T ) + RT ln x2 = 2(G1 (pH , T ) + RT ln x1 ) = 2GL (pH , T ),

Gi (pH , T ) + RT ln xi = i(G1 (pH , T ) + RT ln x1 ) = iGL (pH , T ),

where
∞ −1

xi = γi γi + γA .
i=1

Parameter xi is the unknown mole fraction of the ith component of the

mixture, i = 1, 2, …, A is the unknown saturation pressure.
This system of equations can be supplemented by the normalization condition:
∞

xi + xA = 1.
i=1

This system is a system of nonlinear equations, from which the saturation curve,
and corresponding to it, the cluster size distribution function can be found. From
Eq. 11.2, we can get expressions for xi :

iGL (pH , T ) − Gi (pH , T )

xi = exp .
RT

From the relationship between molar fractions and molar mass concentrations,
the expression for the molar fraction of the inert component has a view:
∞ −1

xA = γA γi + γA ,
i=1
∞
 γA
γi + γA = .
i=1
xA

We write the law of conservation of the condensable component:

[email protected]
144 I. E. Ivanov et al.

∞ ∞
∞ −1
   γ0
iγi = γ0 => ixi = γ0 γi + γA = xA .
i=1 i=1 i=1
γA

The mole fraction of inert gas can be written in the following form:
∞
γA 
xA = ixi .
γO i=1

Thus, the saturation curve can be found from the normalization conditions:
∞

xi + xA = 1.
i=1

After substitution of molar fractions into it, we obtain the following espression:
∞
 γA iGL (p, T ) − Gi (p, T )
F(p, T ) = 1+ i exp − 1 = 0. (11.12)
i=1
γ0 RT

The nonlinear Eq. 11.12 is the relationship between the pressure and temperature
on the saturation curve. Accordingly, at the known temperature from Eq. 11.12, the
saturation pressure pH can be found. To solve Eq. 11.12, Newton method is used in
the work, the initial pressure approximation (pH0
) is calculated from the equality of
the chemical potentials of the monomers and the liquid phase:
0
pH GL0 (T ) − G10 (T )
= exp .
p0 RT

For practice, it is often of interest to find the equilibrium distribution function for
given two thermodynamic parameters, for example, p and T. The chemical potential
of the monomers (gas phase) is less than the chemical potential of the liquid phase:

G1 (p, T ) + RT ln x1 < GL (p, T ).

In the equilibrium state, for a cluster of any size i, the following equality holds:

Gi (p, T ) + RT ln xi = i(G1 (pH , T ) + RT ln x1 ).

Therefore,

iG1 (p, T ) − Gi (p, T )

xi = exp i ln x1 + .
RT

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 145

Fig. 11.1 Saturation curve

of water vapor in nitrogen
with different mass fractions
of water (1, 5, 10, 20, and
100%)

Similar to the above, the concentration of monomers can be found from a nonlinear
equation:
∞
 γA iG1 (p, T ) − Gi (p, T )
F(ln x1 ) = 1+ i exp i ln x1 + − 1 = 0.
i=1
γ0 RT
(11.13)

It should be noted that from this equation the cluster size distribution function can
be found in the neighborhood of the saturation curve on the left (in the metastable
region).
For water vapor in nitrogen, we present saturation curves on the P–T phase plane
depending on the mass fraction of water vapor (Fig. 11.1). As expected, the saturation
pressure increases significantly with decreasing mass fraction of water vapor.

11.3 Numerical Results

In this section, a special test with ideal constant volume adiabatic reactor is considered
in Sect. 11.3.1, while a wet steam flow with spontaneous condensation in Laval nozzle
is presented in Sect. 11.3.2.

[email protected]
146 I. E. Ivanov et al.

11.3.1 Test: Ideal Constant Volume Adiabatic Reactor

A special test was used to test the dynamics of gas transition from a metastable state
to an equilibrium state. At the initial moment of time, the parameters of the medium
in the volume correspond to the strong nonequilibrium (metastable) state. Medium
is motionless. In this chapter, we consider a mixture of nitrogen with water vapor.
Mass fraction of water vapor is 0.02. The pressure in the mixture is 100,000 Pa, the
temperature is 200 K. On the P–T diagram, this state corresponds to a point located
in the region to the left of the saturation curve for water. Over time, the system
will develop to an equilibrium state due to the formation of a liquid-droplet phase,
the release of heat of condensation and an increase in temperature and pressure
with a constant mixture density, but varying density of the gas phase. Thus, the
simulation result depends only on the macrostate of the system, in which it will
be after establishing equilibrium. This circumstance allows us to compare only the
processes of gas transition to the equilibrium state separately from the gas-dynamic
components of the system. In this test, the object of interest is both the final state of
the system (pressure and temperature, the mass fraction of the condensed phase) and
the dynamics of the transition from a nonequilibrium state to an equilibrium state
(transition time and system trajectory).
Comparison of the calculating results of the transition process to the equilibrium
state obtained by the quasi-chemical model and MOM is depicted in Fig. 11.2.
The dynamics of temperature change and pressure in an ideal adiabatic reactor of
constant volume is shown. For MOM, Eqs. 11.5–11.6 with different values of the
parameters β i g were used. For all values of these parameters, the final state of
the medium in the reactor coincides with the quasichemical model with an accuracy
of 0.3 K and 40.0 Pa. The dynamics of condensation for β = 1 and g = 1 are very
different in these two models (red dotted line and blue solid line in Fig. 11.2a, b).
In a quasi-chemical model, the growth process of large clusters (droplets) proceeds

Fig. 11.2 The evolution of parameters in an ideal adiabatic constant volume reactor: a temperature,
b pressure. Parameters of MOM (β = 1, 0.0015, 0.0025 and g = 1, 0.2, 0.2)

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 147

Fig. 11.3 Size distribution functions

more smoothly and nucleation occurs faster (Fig. 11.2). By accelerating the onset
of nucleation in MOM (g < 1) and slowing the growth of droplets (β < 1), one can
obtain a good correspondence of the dynamics for the two models under study.
Using a quasi-chemical model of condensation, the dynamics of the time vari-
ation of the cluster size distribution function were obtained (Fig. 11.3). The initial
cluster size distribution function (blue cross) was taken from the saturation curve and
contained clusters ranging in size from 1 to 10 with concentrations exceeding 10−10 .
The green curve corresponds to the equilibrium distribution function for a given spe-
cific volume and internal energy and contains in significant concentrations clusters
of sizes from 1 to 10, as well as, clusters with a size of 100,000, which in this model
correspond to the liquid phase. The violet, blue, red, and brown curves show the
dynamics of changes in the distribution function from the initial equilibrium state to
the final one. Initially, a condensation wave is formed (purple graph), which spreads
from small to large. After the condensation wave reaches the maximum of the cluster
sizes taken into account, the concentrations of clusters grow from the vicinity of the
maximum size, while the concentrations of clusters of intermediate sizes decrease
(blue, red, and brown curves). Concentrations of clusters of small sizes gradually
fall on the equilibrium distribution function (green curve).

11.3.2 Wet Steam Flow with Spontaneous Condensation

in Laval Nozzle

The flow of superheated water vapor (P0 = 124,000 Pa, T 0 = 391.55 K) in a flat
nozzle and jet flowing out into outlet space (P = 20,000 Pa) is considered. The
geometric parameters of the nozzle are taken from [27]. The height of the entrance
section is 0.06947 m, the critical one is 0.04 m, the output one is 0.04733 m, the length
of the supersonic part of the nozzle is 0.28625 m, the radius of the transonic part of

[email protected]
148 I. E. Ivanov et al.

Fig. 11.4 The distribution of parameters along the axis of the nozzle: a pressure, b radius of droplets

the nozzle is 0.27145 m. The flow calculations were performed by MOM in 1D and
2D formulations using various models of the growth rate of a drop (Eqs. 11.6–11.8).
The results of 1D calculations in comparison with experimental data from [27,
28] are shown in Fig. 11.4a, b. Figure 11.4a shows the pressure distribution along
OX axis in the plane of symmetry. The red curve corresponds to the calculations
by MOM with the model of the growth rate of a drop (Eq. 11.6) with parameters
β = 0.3 and the correction factor g = 1.5 in the ratio (Eq. 11.5). The orange curve
corresponds to the calculation by MOM with the model of the growth rate of the drop
(Eq. 11.10) with parameters β = 0.3 and g = 1.5. Figure 11.4b shows the distribution
along OX axis of the drop radius in the calculations (line) and experiment (circle).
The results of 2D calculations for a jet flowing into external area are shown in
Fig. 11.5. The mesh has about 100,000 cells. For the flow coming out of the nozzle,
the second condensation jump occurs as a result of the additional expansion of the gas
leaving the nozzle. The resulting shock wave behind Mach barrel causes a decrease
in concentration α. The maximum concentration of water is reached in the inner area
of the first barrel and is equal to 6.7%. Figure 11.6 shows Mach number contours.

Fig. 11.5 The distribution of water concentration in the nozzle

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 149

Fig. 11.6 Mach number contours

11.4 Conclusions

Two approaches to modeling the process of spontaneous condensation in gas-

dynamic flows have been developed. The first approach is a kinetic approach, which
is based on a quasi-chemical model. The second approach is a continual approach,
which is based on MOM. The model of gas dynamics is Navier–Stokes equation
in two-dimensional coordinate system and Euler equation in one-dimensional coor-
dinate system. To solve the gas-dynamic system, a numerical method of enhanced
order of accuracy based on Godunov method and AUSM method is used.
Using the two approaches developed, numerical simulation of condensation pro-
cesses was carried out in an ideal adiabatic reactor of constant volume and with a flow
of condensing matter in a nozzle using quasi-one-dimensional formulation and a jet
flowing into outlet space. To simulate a jet flowing into flooded space, Navier–Stokes
equations are used in two-dimensional formulation.
In the first case, spontaneous condensation is considered, when a mixture of nitro-
gen and water vapor (2% mass fraction) transfers from nonequilibrium (metastable)
to equilibrium. Both approaches give similar results on the final state, the difference
does not exceed 0.15% in temperature and 0.04% in pressure. However, the dynam-
ics of the process of transition to the equilibrium state may differ significantly for
the two approaches. Coherence high level of the dynamics of the process for the
two studied models is achieved by a special selection of correction factors in MOM.
Using a quasi-chemical model for water vapor in nitrogen, we obtained the saturation
curves in the P–T phase plane depending on the mass fraction of water vapor.
In the second test case, the flow of superheated steam in a flat Laval nozzle is
numerically investigated using MOM and the solution of Navier–Stokes equations.
In the supersonic part of the nozzle, spontaneous condensation of water occurs with
the formation of a region of increasing pressure and temperature (condensation jump)
directly near the critical section. A qualitative agreement was obtained between the
numerical and experimental pressure distributions on the plane of symmetry of the
nozzle, and the accuracy of 23% of the values of the average radii of droplets on the
nozzle section was obtained in the calculation and experiment.

[email protected]
150 I. E. Ivanov et al.

Corrective factors in MOM selected for the considered problems (in the first
problem due to match the quasi-chemical model and in the second one to match with
the experiment) have significantly different sets of values. This indicates the need to
improve MOM in terms of calculating the nucleation rate and rate of growth of the
droplets’ mass.

Acknowledgements This work was carried out within the state task no. 9.7555.2017/BCh.

References

1. Aksenov, I.I., Belous, V.A., Strel’nitskij, V.E., Aksyonov, D.S.: Vacuum-arc equipment and
coating technologies in KIPT. Phys. Radio Technol. Ion Plasma Technol. 58–71 (2016)
2. Ieshkin, A.E., Shemukhin, A.A., Ermakov, Y.A., Chernysh, V.S.: The influence of the gas
cluster ion beam composition on defect formation in targets. Mosc. Univ. Phys. Bull. 71(1),
87–90 (2016)
3. Hill, P.G.: Condensation of water vapour during supersonic expansion in nozzles. J. Fluid
Mech. 593–620 (1966)
4. Gorbunov, V.N., Ryzhov, YuA, Pirumov, U.G.: Noequilibrium condensation in high-speed gas
flows. Gordon and Breach Science Publishers, Institute of Mechanics, Moscow University,
USSR (1988)
5. Oswatitsch, K.: Kondensationserscheinungen in uberschallduzen. ZAMM 22, 1–14 (1942)
6. Stiver, H.: A condensation phenomenon in high velocity flows. In: Emmons, W. (ed.) Fun-
damentals of Gas Dynamics. Chapter 3. Princeton University Press, New Jersey, U.S.A.
(1958)
7. Saltanov, G.A., Seleznev, L.I., Tsiklauri, G.V.: Generation and growth of condensed phase in
high-velocity flows. Int. J Heat Mass Transf. 16, 1577–1587 (1973)
8. Kotake, S., Glass, I.I.: Flows with nucleation and condensation. Prog. Aerospace Sci. 19,
129–196 (1979)
9. Bakhtar, F., Young, J.B., White, A.J., Simpson D.A.: Classical nucleation theory and its appli-
cation to condensing steam flow calculations. In: Proceedings of the Institution of Mechanical
Engineers, vol. 219, Part C. J. Mech. Eng. Sci., pp. 1315–1333 (2005)
10. Gorbunov, A.A., Igolkin, S.I.: Statistic simulation of crystal grids growing at vapor condensa-
tion. Matem. Mod. 17(3), 15–22 (2005)
11. Bauer, S.Y., Frurip, D.J.: Homogeneous nucleation in metallic vapors. A self-consistent kinetic
model. J. Chem. Phys. 81(10), 1015–1024 (1977)
12. Bykov, N.Y., Gorbachev, Y.E.: Cluster formation in copper vapor jet expanding into vacuum:
the direct simulation Monte Carlo. Vacuum 163, 119–127 (2019)
13. Bykov, N.Y., Gorbachev, Y.E.: Mathematical models of water nucleation process for the direct
simulation Monte Carlo method. Appl. Math. Comput. 296, 215–232 (2017)
14. Bykov, N.Y., Safonov, A.I., Leshchev, D.V., Starinskiy, S.V., Bulgakov, A.V.: Gas-jet method
of metal film deposition: direct simulation Monte-Carlo of He-Ag mixture flow. Mater. Phys.
Mech. 38, 119–130 (2018)
15. Volkov, V.A., Muslaev, A.V., Pirumov, U.G., Rozovskii, P.V.: Non Equilibrium condensa-
tion of metal vapors/inert gas mixture during expansion through the nozzles of cluster-beam
generators. Fluid Dyn. 30(3), 399–408 (1995)
16. Egorov, B.V., Markachev, Y.E., Plekhanov, E.A.: Correlation of the quasi-chemical cluster
nucleation model when compared with the experimental data. Khimicheskaia Fizika 25(4),
61–70 (in Russian) (2006)

[email protected]
11 Numerical Simulation of the Process of Phase Transitions … 151

17. Sternin, L.E.: Fundamentals of gas dynamics of two-phase nozzle flows. Moscow, Izdatel’stvo
Mashinostroenie (in Russian) (1974)
18. Kortsenshteyn, N.M., Samuilov, E.V., Yastrebov, A.K.: Study of the volume condensation
process in supersaturated vapor by the direct numerical solution of the kinetic equation for the
droplet size distribution function. Colloid J. 69b(4), 488–295 (2007)
19. Gidaspov, V.U., Ivanov, I.E., Kryukov, I.A., Nazarov, V.S., Malashin F.A.: Study of the con-
densation process in nozzles with a large degree of expansion. Phys. Chem. Kinetics Gas Dyn.
19(2). Art. no. 737.1–737.17 (in Russian) (2018)
20. Luo, X., Cao, Y., Xie, H., Qin, F.: Moment method for unsteady flows with heterogeneous
condensation. Comput. Fluids 146, 51–58 (2017)
21. Wyslouzil, B.E., Heath, C.H., Cheung, J.L., Wilemski, G.: Binary condensation in a supersonic
nozzle. J. Chem. Phys. 113(17), 7317–7329 (2000)
22. Sova, L., Jun, G., Stastny, M.: Modifications of steam condensation model implemented in
commercial solver. AIP Conference Proceedings 1889, 020039.1–020039.8 (2017)
23. Kantorowitz, A.: Nucleation in very rapid vapor expansions. J. Chem. Phys. 19(9), 1097–1100
(1951)
24. Young, J.B.: The spontaneous condensation in supersonic nozzles. Phys. Chem. Hydrodyn.
3(1), 57–82 (1982)
25. Gyarmathy, G.: Grundlageiner Theorie der Nassdampfturbine. Dissertation, Juris Verlad,
Zurich (1960)
26. Choi, B., Shim, J., Kim, Ch., Park, J., You, D., Beak, J.: Numerical simulation of homogeneous
condensing wet-steam flow using an Eulerian–Lagrangian method. In: Proceedings of Shanghai
2017 Global Power and Propulsion Forum, GPPS AME, pp. 17–91 (2017)
27. Dykas, S., Majkut, M., Smolka, K., Strozik, M.: Experimental research on wet steam flow with
shock wave. Exp. Heat Transf. 28(5), 417–429 (2014)
28. Dykas, S., Majkut, M., Smolka, K., Strozik, M.: An attempt to make, a reliable assessment of
the wet steam flow field in the de Laval nozzle. Heat Mass Transf. 54(9), 2675–2681 (2018)

[email protected]
Chapter 12
Numerical Study of the Injection
Parameters Impact on the Efficiency
of a Liquid Rocket Engine

Yulia S. Chudina , Evgenij A. Strokach , Igor N. Borovik

and Vladimir Yu. Gidaspov

Abstract This chapter analyzes the effect of various parameters (such as droplet
injection velocity components by a swirl injector in a cylindrical coordinate system
and droplet size distribution parameters) of fuel injection in an oxygen-kerosene
rocket engine on the efficiency of the workflow. The study was conducted for two
cases of application of the wall film cooling of the combustion chamber and without it.
It is shown that the parameters for fuel injection in the case of using a wall film cooling
effect in an unobvious way. A description of the object of study, the main features of
the numerical experiment, and models used in the course of the study are given, as well
as, the results are analyzed and recommendations on their use and further research
in this area are formulated. The use of the obtained results allows to form a technical
task for the design of a mixing head that implements optimal combustion conditions
in the combustion chamber. Verification of the calculated data was carried out using
the results of experimental studies conducted at Moscow Aviation Institute (National
Research University) at the Department of Rocket Engines. This experimental work
was carried out using a specially designed DMTMAI-200OK rocket engine.

Y. S. Chudina · E. A. Strokach · I. N. Borovik (B) · V. Yu. Gidaspov

Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
Y. S. Chudina
e-mail: [email protected]
E. A. Strokach
e-mail: [email protected]
V. Yu. Gidaspov
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 153

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_12

[email protected]
154 Y. S. Chudina et al.

12.1 Introduction

The development of high-performance low-thrust liquid-propellant rocket engines is

among the most important trends in the development of space technology because an
increase in the efficiency of low-thrust engines gives an increase in the lifetime of the
spacecraft on the orbit. During the design process, the developer faces a number of
difficulties arising from the lack of information about the processes occurring in the
most thermally stressed area—Combustion Chamber (CC). In the limited volume of
CC, it is necessary to organize not only the high-quality mixture but also sufficient
film cooling. Another feature to complicate the performance is the pulsed operation
mode of the engine, where the non-stationary effects of wall heating, ignition, and
pulsations of fuel supply become decisive [1–4].
Taking into account the described difficulties, one can see how important it is for
the developer to obtain complete information about the working process in the hot
part of the engine in a relatively small amount of time and with the least computational
costs. The widely used modern numerical methods in fluid dynamics partially solve
this problem and help to understand what most affects the quality of the working
process [4].
The purpose of the presented numerical study is to determine the magnitude and
nature of the influence of the fuel injection parameters on engine performance.
Chapter is organized as follows. Section 12.2 provides a short review of features
of the working process in rocket engines combustion chambers. The experimental
test case chosen for this analysis is described in Sect. 12.3. Section 12.4 deals with the
numerical setup and models applied in the hot gas simulation, whereas the results of
the turbulent combustion of kerosene spray in gaseous oxygen with different injection
condition are presented in Sect. 12.5. Finally, Sect. 12.6 gives an overall conclusion
and summary of the results.

12.2 Features of the Working Process

The physical picture of the flow in CC is described in detail in [4]. This article
describes the most important processes for the development studies. The fundamental
role in the organization of the workflow assigns to the fuel injection through the
injector head. The main task is the complete and uniform mixing of components
across CC. The limited number of nozzle elements due to the small size of the
engine complicates sufficient mixing and evaporation. These considerations require
the development of nozzles for special structures, optimization of their location, and
selection of flow parameters.
Almost overwhelming effect on the completeness of combustion has the atomiza-
tion of the liquid fuel. Both the quality parameters, such as the type of atomization, the
configuration of the devices, and the quantitative parameters, such as the distribution
and size of the droplets, are of great importance.

[email protected]
12 Numerical Study of the Injection Parameters Impact … 155

Currently, the prevailing view says that there is a direct relationship between the
decrease in the diameter of the average droplet and the increase in the completeness of
combustion, which is explained by the rapid evaporation and burning of the droplets.
However, the influence of size can be ambiguous. For example, in the case of a
wide diameter distribution, the relatively high value of Sauter mean diameter may
correspond to a higher degree of combustion than at relatively low values of the
mean diameter. The main mechanism for increasing the completeness of combustion
when increasing the value of Sauter mean diameter is the action of relatively large
droplets that appeared in the spray spectrum. Large drops have more inertia compared
to small drops. Large drops stay longer in the combustion chamber and, therefore,
travel a longer distance. By evaporating, the drops make a trace of their vapors,
which are mixed with the vapors of the second fuel component due to turbulence
and diffusion. These phenomena increase the area of evaporation and combustion
of fuel. With monodisperous (narrow) distribution of droplets by diameter, small
droplets quickly evaporate near the injectors. Small drops make short tracks in the
combustion chamber, which may not intersect with each other at all. This means
that the fuel and oxidizer mix badly. The rapid evaporation of small droplets leads
to a stratification of evaporated propellant and combustion products the combustion
chamber volume, resulting in poor mixing of fuel and oxidizer, and, as a result,
low the completeness of combustion. This means that for each combustion chamber
design there is a distribution of droplets by diameter and a drop injection velocity
that ensures maximum combustion performance [5].
Considerable effect on the completeness of combustion has the spray angle and
fuel injection rate. If the spray angle is too small, the fuel and oxidizer do not mix
properly, and the already reacted mixture stratifies into a separate stream and does
not mix with the cold stream, i.e. regions with high or low component ratios appear
which reduces the combustion performance. If the spray angle is too wide, medium
and large droplets fall on both the mixing head and the wall, wherein the presence
of the oxidizer reactions occur, which leads to the wall material burnout.
The only widely used type of wall cooling in modern small size rocket engines is
film cooling. Either gaseous or liquid, reducing or oxidizing, its presence always leads
to a change in the oxidizer to fuel ratio, which reduces the combustion performance.
The use of a liquid film is the most difficult case to evaluate. First, at low feed rates, the
laminar (or nearly laminar) film experiences the effects originating in the core flow
as stretching and turbulent pulsations. This is reflected in the formation of Kelvin-
Helmholtz waves on the surface of the laminar film. Ultimately, the growth of these
instabilities can lead to the fragmentation of the film into droplets, thereby increasing
the heat transfer from the main flow to the wall. In such a case, it is needed to prevent
the early destruction of the film. Second, at higher flow rates a turbulent film, due to
the internal phenomena of instability, coupled with the aerodynamic influence of the
core flow, can break up into fine droplets faster than laminar film. The presence of a
large number of uncertainties leads to the need for constant optimization of the film
injection parameters.
This work is aimed to estimate the impact of fuel injection parameters on the effi-
ciency of the working process expressed by the incompleteness of combustion. The

[email protected]
156 Y. S. Chudina et al.

phenomena described above, which are associated with the ambiguity of parame-
ters, constitute the main difficulties in determining the optimal values of the variable
parameters. The proposed methodology, as well as, the recommendations obtained in
the modeling basis, is one of the steps to the improved understanding of the working
process in CC of small-sized rocket engines.

12.3 The Studied Object

In the working process study, an experimental kerosene-GOx DMTMAI-200 with a

nominal thrust of 200 N was selected as the test object (Fig. 12.1). The experimental
study of the engine was carried out earlier in [6].
The injector head contains one central and six peripheral open-type centrifugal
two-component spray nozzles, where the fuel and oxidizer nozzles are arranged so
that the fuel flowing out of the nozzle enters the oxidant flow, mixes, and flows into
the combustion chamber.
The wall in CC is protected from overheating by kerosene film cooling. The
cooling kerosene is fed through slots between the injector head and CC wall. The
liquid film forms a shroud that flows along the wall of the combustion chamber.

Fig. 12.1 View of the model of experimental rocket engine

[email protected]
12 Numerical Study of the Injection Parameters Impact … 157

12.4 Numerical Modeling of the Processes in Combustion

Chamber

Here, the commercial Computational Fluid Dynamics (CFD) code Ansys CFX is
used that supports a large number of mathematical models of physical processes. It
should be noted that it is widely used to evaluate the working process in engines and
power plants including rocket engines.
The numerical study was carried out using a mathematical model built on the
basis of Navier-Stokes equations averaged according to Favre [7].
Assumptions of the numerical model. The numerical model is based on the
following assumptions:
• The combustion products and fuel components are ideal gases. They have the
constant viscosities and heat capacities depending on temperature.
• To simulate the liquid film cooling, Euler-Lagrange method is used, in which the
liquid component is represented as a set of liquid drops injected from the film
slots.
• The calculation is carried out in a stationary problem formulation.
• The model takes the buoyancy into account.
• The walls in CC, nozzle, and walls of injector head are adiabatic.
• Turbulence model based on Boussinesq hypothesis is used.
• Simulation liquid fuel injection into the flow core is performed using Euler-
Lagrange approach with the initial parameters of size and distribution of the liquid
spray.
• The radiative heat transfer is not taken into account.
Methods for modeling of the droplets motion, evaporation, break-up, and
heat transfer. Modeling the movement of kerosene droplets is done using the clas-
sical approach. The approach for drag modeling (Schiller-Naumann (S-N) correla-
tion) uses the enhance to dynamically model the changing shape of the drops (Liu
correlation) [8], which is acceptable for the studied processes. As a model for heat
transfer, Ranz-Marshall correlation [9, 10] is included. Antoine equation for mass
transfer between the phases used reference factors for the n-decane C10 H22 since, in
the case of kerosene of JET-A type, it is assumed that under the observed conditions
the largest evaporating fraction is the n-decane [11].
Combustion simulation. In this study, the flamelet combustion model was used
[12]. The model is based on the assumption that the combustion process occurs in
thin layers with an internal structure. The flame itself is a complex joint structure of
one-dimensional stretched laminar flames called as flamelets. One of the assumptions
of this model is the constant Lewis number during the flamelet formation. CFX-RIF
library was used to create the flamelets. The reasons for using the flamelet model
were its relative robustness, a small number of applied equations, and notable practice
of application in CFD codes.
Secondary break-up models. The secondary break-up of liquid can strongly
influence the atomization parameters. Currently, there are many approaches to the

[email protected]
158 Y. S. Chudina et al.

calculation of such processes. Based on a preliminary study of the models available

in Ansys CFX and the recommendations in the open literature, it was decided to
use one of the most sophisticated models for the secondary breakup in CFX, CAB
model, to simulate the secondary fragmentation of liquid droplets [13–16].
General considerations. The main goal of the numerical study was to determine
the influence of various parameters on the engine performance, which is defined as
the pressure in CC and the characteristic velocity β defined by Eq. 12.1, where pcc
is the pressure in CC, F th is the throat area, ṁ Σ is the total mass flow rate [2].

pCC · Fth
β= (12.1)
ṁ Σ

Numerical domain and grid. The calculation domain is the 60° sector of the inter-
nal volume CC of the experimental engine. Gas-liquid centrifugal two-component
kerosene-oxygen nozzles are represented by circles on the surface of the injector
head. The diameter of the holes is 8 mm. The simulation domain and boundary con-
ditions are shown in Fig. 12.2. The properties of the fuel components are shown in
Table 12.1.
The numerical grid is shown in Fig. 12.3. A preliminary analysis of grid conver-
gence based on pressure values in CC and characteristic velocity revealed that the
grid convergence for these two parameters is achieved at the number of grid elements
of about 300,000.

Fig. 12.2 Simulation domain with boundary conditions. 1—the mass flow of the central nozzle,
2—the mass flow of the peripheral nozzle, 3—the mass flow rate of the liquid film, 4—cyclic
boundary conditions, 5—the boundary condition of the opening type with the setting of the ambient
pressure

[email protected]
12 Numerical Study of the Injection Parameters Impact … 159

Table 12.1 Properties of the propellant components [11]

Property Kerosene (liquid) Kerosene (gaseous) Oxygen
Density (kg/m3 ) 727 Ideal gas Ideal gas
Specific heat capacity 2192.4 Dependence from the Dependency from the
(J/kg K) NIST database [17] NIST database [17]
Dynamic viscosity 0.00212 5.28767e−06 2.06594e−05
(Pa s)
Surface tension (N/m) 0.027 – –
Thermal conductivity 0.1218 0.00907355 0.0254141
(W/m K)

Fig. 12.3 Numerical grid

12.5 Study of the Fuel Injection

Study of the fuel injection parameters influence on the core flow. At the first stage
of the study, only core-flow fuel injection simulations took place without accounting
for film cooling.
The effect of the injection parameters was determined by varying the components
of the velocity of the injected fuel droplets and flow of gaseous oxygen. The velocity
components are given in a cylindrical frame with a reference point located in the
center of the injector circle and represent the radial, axial, and tangential component
of the velocity vector.
The ratio of the axial and radial component determines the angle of the atomization
cone. In the case when the ratio of axial and radial components is much less than 1,
this will correspond to a wide atomization angle [>45° (angle between the side and the

[email protected]
160 Y. S. Chudina et al.

Table 12.2 Fuel injection

Parameter Value (kg/s)
parameters
Total kerosene mass flow 0.01827
Total oxygen mass flow 0.04972
Total mass flow (O + F) 0.068

axis)]. In the opposite situation, the flow will be stretched along the axis, which will
affect mixing and efficiency. The tangential component reflects the degree of swirling,
which due to high turbulent intensity promotes a mixing of the fuel components. In
addition, the evaluation of the effect of fuel injection at various ratios of the velocity
components is important in studying the effect of flow rate pulsations in the nozzle.
Such non-stationary phenomena may occur due to some instabilities, both in the
injector itself and in the feed system.
The ranges of fuel injection parameter variations are the following:
• The Rosin-Rammler size parameter [18] for the droplet diameter: 3–127 µm.
• The Rosin-Rammler spread parameter: 1.7–18.
• The droplet average Sauter diameter: 2.23–109 µm.
• The axial velocity component: 1.12–4.9 m/s.
• The radial velocity component: 1.12–4.3 m/s.
• The tangential velocity component: 1.12–4.3 m/s.
The fuel injection parameters are shown in Table 12.2.
The calculations were performed at a constant total mass flow rate of 0.068 kg/s
with an oxidizer excess ratio of 0.8.
In total, 125 simulations with five variable parameters were performed. An impor-
tant assumption used in the calculations was that the sprayed fuel droplets and oxygen
move with the same injection velocity. This assumption is introduced to simplify the
physical interpretation of the computational model and is based on a relatively high
flow rate of up to 21 m/s and a small diameter of the droplets. Earlier in [19], it was
shown that the diameter of droplets in nozzles of this type mainly depend on the
parameters of the carrier phase.
The results of the calculations are shown in Figs. 12.4, 12.5, and 12.6. The plots
show the increase of the characteristic velocity with the increase of the radial and
tangential components of the velocity can be observed. Moreover, in the range of
velocity components of 1.12–3 the characteristic velocity grows rapidly. The increase
of the radial component with a simultaneous decrease of the axial component leads to
the increase of the spraying angle. The increase of the tangential velocity projection
with the decrease of the axial velocity depicts the flow swirling degree. The described
effects have a significant impact on the processes occurring in the working volume.
This indicates a correlation with the physical picture of the workflow described in
previously published papers [20].
The dynamics of changes in the characteristic velocity with the increase of the axial
component also looks predictable—a small decrease. This is due to the deterioration
of mixing flows with increasing axial components.

[email protected]
12 Numerical Study of the Injection Parameters Impact … 161

Fig. 12.4 Impact of the tangential velocity on the characteristic velocity (the lines are the lines of
trend)

Fig. 12.5 Impact of the radial velocity on the characteristic velocity (the lines are the lines of trend)

The dependence of the characteristic velocity on the axial component is great,

but in this case, with an increase of the axial velocity, the values of the character-
istic velocity decrease. This is explained by the following: an increase of the axial
velocity reduces the residence time of the droplets in CC, which leads to a decrease
of the evaporated fuel fraction. On the other hand, the increase of the radial and
tangential velocity components leads to better mixing of the fuel components due to

[email protected]
162 Y. S. Chudina et al.

Fig. 12.6 Impact of the axial velocity component on the characteristic velocity (the lines are the
lines of trend)

the interaction between adjacent nozzles and the increase of the residence time of
the drops in CC due to the curvature and stretching of their traces. The variation of
other parameters, i.e. Rosin-Rammler size parameter and the spread parameter have
little effect.
The described phenomena will help in the formulation of recommendations for the
design of the elements of mixture formation and the geometric and modal appearance
of CC with regenerative cooling. For example, we can conclude that the growth of
the tangential and radial components with the decrease in the axial component allows
to increase greatly the combustion performance.
Study of the injection parameters influence with film cooling. The second part
of the research was aimed to model the working process with film cooling by varying
a large number of parameters. The liquid component of the fuel was fed through a
gap in the peripheral region of the injector head.
To simulate the film cooling, a large number of parcels was injected through the
film cooling slots. Thus, the approach is an assumption made to consider computa-
tional resources reduction. Moreover, such an approach still takes into account the
dynamic, thermal, and mass transfer phenomena. This greatly simplifies the calcu-
lation and improves the robustness and time. The choice of this method was made
both because of the computational requirements and complexity of other methods
of two-phase flow modeling. The disadvantage of this approach is the uncertainty in
the diameter of the droplets: the too small drops will evaporate quickly and the too
large ones may not evaporate at all. Both in the first and second cases, the working
process would be incorrectly interpreted.
The chapter discusses two parameters of fuel component injection into the film:
the diameter of kerosene droplets injected as a film cooler and their velocity. Fuel
injection parameters are shown in Table 12.3.
The ranges of fuel injection parameter variations:

[email protected]
12 Numerical Study of the Injection Parameters Impact … 163

Table 12.3 Fuel injection

Parameter Value (kg/s)
parameters
Total massflow of kerosene (per core flow) 0.01184
Total oxygen massflow (to the core flow) 0.03221
Kerosene massflow for the film cooling 0.02394
Total massflow (O + F) 0.068

• The Rosin-Rammler size: 12–109 µm.

• The Rosin-Rammler spread: 1–18.
• The axial velocity component: 0.9–3.9.
• The radial velocity component: 0.3–1.7.
• The tangential velocity component: 0.9–3.47.
• The diameter of the liquid droplets injected as the film: 53–250 µm.
• The velocity of the liquid droplets injected as the film: 0.3–2.19 m/s.
As an experimental design plan, a central composite plan extended by additional
points was used. The total number of points is 160. The size of the slot used for film
cooling component was 1 mm. The results of the study are shown in Figs. 12.7, 12.8,
12.9, 12.10 and 12.11.
According to the simulation results, there are significant differences in the depen-
dence of the completeness of combustion on the magnitude of the velocity compo-
nents of fuel injection compared to the results obtained in the simulations without
film cooling. Increasing the axial velocity component decreases the characteristic
velocity to a small extent. The growth of the tangential component also relatively
weakly increases the value of the characteristic velocity. The increase in the radial
component leads to a significant reduction in combustion performance.
In this case, the following phenomenon occurs. At some values of the radial
component, the oxidizer stream and liquid spray begin to fall directly into the film

Fig. 12.7 Impact of the axial

velocity component on the
characteristic velocity (the
lines are the lines of trend)

[email protected]
164 Y. S. Chudina et al.

Fig. 12.8 Impact of the

radial component on the
characteristic velocity (the
lines are the lines of trend)

Fig. 12.9 Impact of the

tangential component on the
characteristic velocity (the
lines are the lines of trend)

Fig. 12.10 Impact of the

diameter of the film droplets
on the characteristic velocity
(the lines are the lines of
trend)

[email protected]
12 Numerical Study of the Injection Parameters Impact … 165

Fig. 12.11 Impact of the

spread parameter of
Rosin-Rammler distribution
in the core flow on the
characteristic velocity (the
lines are the lines of trend)

area preventing the formation of a homogeneous mixture of fuel and oxidant in the
core flow, which reduces the efficiency of mixing, evaporation, and combustion.
In the given variation range, the size parameter of Rosin-Rammler diameter distri-
bution is insignificant. An increase of the spread parameter reduces the characteristic
velocity, which is consistent with the results of [5] and can be explained by the con-
clusions made in this work. An increase in the speed and diameter of the film drops
does not almost affect on the efficiency of the workflow in CC.
The results of the simulation of the workflow in CC with film cooling show that
the effect of the velocity and diameter of the drops of the film cooling (in these ranges
of variation) is not significant at the given flow rate in the film.
For the cases taking the film cooling into account, the characteristic velocity
reaches the highest values at low values of the spread parameter, which corresponds
to wide distribution. For cases without film cooling, the width of the distribution
has little effect on the efficiency of the workflow. This is most likely that the wide
distribution in the spectrum provides large and small droplets in the same propor-
tion, which allows to increase the evaporation/combustion areas and the average
temperature over the volume of CC. This effect shows itself more significantly in the
presence of film cooling, which creates thermal stratification over the cross-section
of the combustion chamber and significantly reduces the efficiency of the working
process.
To summarize, the numerical study illustrated the effect of various parameters
and indicated that mostly the core flow injection has a significant effect. However,
the coupled interaction of the effects of the parameters can strengthen or weaken the
resulting influence of a single parameter.

[email protected]
166 Y. S. Chudina et al.

12.6 Conclusions

Optimization of the fuel injection parameters can significantly improve engine effi-
ciency and increase combustion performance. Obviously, however, for each design
type of CC and the injector head, a certain most efficient spray droplet distribution
and combination of velocity components are presented.
An increase of the axial velocity component leads to a decrease in the characteristic
velocity. An increase in the radial component, with respect to the axial component,
in the case of the non-film cooling leads to the increase of the completeness of
combustion, and in the case of the film, to the decrease. The increase of the tangential
component allows to achieve a more complete mixing and efficiency of the workflow
regardless of the presence of the film.
An increase of the droplet diameter of the film and velocity of the liquid film has
little effect on the performance in the studied range.
An increase of the spread parameter reduces the characteristic velocity, which
is consistent with the results previously obtained by the authors. The large spread
parameter of the droplets diameter distribution means monodispersity of the droplets.
With the combustion chamber film cooling, there is a decrease in the performance
of combustion process, while increasing the monodispersity of the droplets, as well
as, without it. That is, poor mixing of the evaporated propellant leads to a decrease
combustion performance to an even greater extent, since film cooling also reduces
the combustion performance by itself.
The results of the velocity components impact can describe, among other phenom-
ena, the non-stationary operation, when the instability of the injection parameters
(mass flow distribution in the injectors, velocity pulsations, and spray parameters)
can significantly affect the efficiency of the engine.
The results obtained may stand as recommendations for the determination of the
requirements for mixing elements, the shape of CC, the mode of operation, etc. This is
closely related to the goal of creating a workflow modeling methodology, along with
obtaining complete information on the qualitative and quantitative parameters of the
process. Designers of rocket engines can use the results of this work for considerable
simplification of the design process.

Acknowledgements This work was supported by the Russian Ministry of Education and Science
(Project 13.7418.2017/8.9).

References

1. Kozlov, A.A., Vorobiev, A.G., Borovik, I.N.: Low-Thrust Liquid-Propellant Rocket Engines.
MAI Publishing House, Moscow (in Russian) (2013)
2. Alemasov, V.E., Dregalin, A.F., Tishin, A.P.: The Theory of Rocket Engines. Mashinostroenie,
Moscow (in Russian) (1980)

[email protected]
12 Numerical Study of the Injection Parameters Impact … 167

3. Vasiliev, A.P., Kudryavtsev, V.M., Kuznetsov, V.A., Kurpatenkov, V.D., Obelnitskii, A.M.:
Fundamentals of the Theory and Calculation of Liquid Rocket Engines. Higher School, Moscow
(in Russian) (1993)
4. Koroteev, A.S. (ed.): Work Processes in a Liquid Rocket Engine and Their Modeling.
Mashinostroenie, Moscow (in Russian) (2008)
5. Borovik, I.N., Strokach, E.A.: Spray diameter distribution influence on liquid rocket com-
bustion chamber performance, Bulletin PNRPU. Aerosp. Eng. 44, 45–61 (in Russian)
(2016)
6. Tashev, V.P.: Hydrocarbon fuel based on kerosene with additives to increase the energy
efficiency of LRE: Ph.D. theses (in Russian) (2013)
7. Yun, A.A., Krylov, B.A.: Calculation and Modeling of Turbulent Flows with Heat Exchange,
Mixing, Chemical Reactions and Two-Phase Flows in the Fastest-3D Software Package: Study
Guide. MAI Publishing House, Moscow (in Russian) (2007)
8. Liu, B., Mather, D., Reitz, R.D.: Modeling the effects of drop drag and breakup on fuel sprays.
SAE Technical Paper 930072 (1993)
9. Ranz, W.E., Marshall, W.R.: Evaporation from drops. Part I. Chem. Eng. Prog. 48(3), 141–146
(1952)
10. Ranz, W.E., Marshall, W.R.: Evaporation from drops. Part I and Part II. Chem. Eng. Prog.
48(4), 173–180 (1952)
11. Reid, R.C., Prausnitz, J.M., Sherwood, T.K.: The Properties of Gases and Liquids. McGraw-
Hill, New York (1977)
12. Peters, N.: Laminar diffusion flamelet models in non-premixed turbulent combustion. Prog.
Energy Combust. Sci. 10, 319–339 (1984)
13. Spalding, D.B.: Mixing and chemical reaction in steady confined turbulent flames. In:
Symposium (Int.) on Combustion, vol. 13, no. 1, pp. 649–657 (1971)
14. Frank, Th., Kumzerova, E. Esch, Th.: Validation of Lagrangian spray formation for use in
internal combustion engines. In: 5th Joint FZR & ANSYS Workshop “Multiphase Flows:
Simulation, Experiment and Application”, Dresden, Germany, pp. 1–27 (2007)
15. Kumzerova, E., Esch, Th., Menter, F.: Spray simulations: application of various droplet breakup
models. In: 6th International Conference on Multiphase Flow, Leipzig, pp. 1–10 (2007)
16. Kumzerova, E., Esch, T.: Extension of the cab for the wide range of Weber number range.
In: 22nd European Conference Liquid Atomization & Spray Systems, pp. ILASS08-4-5.1–
ILASS08-4-5.7 (2008)
17. NIST (National Institute of Standards and Technology, USA) Chemistry Webbook. http://
webbook.nist.gov. Accessed 08 Sept 2019
18. Rosin, P., Rammler, E.: The laws governing the fineness of powdered coal. J. Inst. Fuel 7,
29–36 (1933)
19. Vasil’ev, A.Y., Maiorova, A.I.: Physical features of liquid atomization when using different
methods of spraying. High Temp. 52(2), 250–258 (2014)
20. Kozlov, A.A., Strokach, E.A.: Investigation of a workflow simulation method in the combustion
chamber of liquid rocket motors of small thrust based on the Eddy Dissipation Model, Bulletin
PNRPU. Aerosp. Eng. 44, 27–44 (in Russian) (2016)

[email protected]
 Part III
Computational Solid Mechanics

[email protected]
Chapter 13
Methods for Calculating the Dynamics
of Layered and Block Media
with Nonlinear Contact Conditions
Ilia S. Nikitin , Nikolay G. Burago , Vasily I. Golubev
and Alexander D. Nikitin

Abstract Continual models of solid media with a discrete set of slip planes (layered,
block media) and with nonlinear type slip conditions at the contact boundaries of
structural elements are constructed. The constitutive equations of the resulting sys-
tem of equations contain a small viscosity parameter in the denominator of nonlinear
free terms. For a stable numerical solution of a system of differential equations, an
explicit–implicit method is proposed with the explicit approximation of the equations
of motion and implicit approximation of the constitutive relations containing a small
parameter. From implicit nonlinear difference approximations analytically using the
perturbation method, various effective formulas for the correction of stress compo-
nents after an “elastic” time step are obtained. To calculate the “elastic” step, we used
a grid-characteristic method on hexahedral grids, which allowed us to increase sig-
nificantly the speed of calculations and simulate a non-stationary three-dimensional
problem of generating a response from an oriented layered or block cracked cluster
located in a homogeneous medium.

I. S. Nikitin (B) · A. D. Nikitin

Institute of Computer Aided Design of the RAS, 19/18, Vtoraya Brestskaya ul., Moscow 123056,
Russian Federation
e-mail: [email protected]
A. D. Nikitin
e-mail: [email protected]
N. G. Burago
Ishlinsky Institute for Problems in Mechanics of the RAS, 101, b1, pr. Vernadskogo, Moscow
119526, Russian Federation
e-mail: [email protected]
V. I. Golubev
Moscow Institute of Physics and Technology (National Research University), 9, Institutsriy per.,
Dolgoprudny, Moscow Region 141700, Russian Federation
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 171

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_13

[email protected]
172 I. S. Nikitin et al.

13.1 Introduction

Continual models of deformable solid media with a discrete set of slip planes (layered,
block media) and with nonlinear (dry friction or viscous-plastic types) slip conditions
at contact boundaries can be obtained with the discrete variant of the slippage theory
[1] or with the method of the asymptotic homogenization [2–4]. In all of these
cases, the constitutive system of equations with the nonlinear free term and the small
stress relaxation time are included. For the stable numerical solution of differential
equations, the explicit–implicit method with the explicit approximation of motion
equations and the implicit approximation of constitutive equations containing small
parameter in the denominator of the free term was proposed. A set of effective
formulas for the stress tensor correcting at the “elastic” step was obtained.
The developed model can be used for the numerical simulation of the seismic
survey process in complex geological fractured media. Nowadays, a lot of different
approaches are created: the finite difference method, Galerkin method, finite ele-
ment method, etc. [5–7]. There are also various numerical methods actively used to
construct hybrid calculation algorithms [8]. Also note that, as the defining system
of equations of elasticity that describes the propagation of seismic waves is hyper-
bolic, its numerical solution can be carried out by the grid-characteristic method.
Apparently, the characteristic method was proposed firstly in [9]. It was described in
detail for a one-dimensional case in [10] and later generalized for a multidimensional
case in [11]. Recently, it was successfully applied for the simulation of waves in the
acoustic [12], elastic [13], and fractured [14–17] media. It was also used to solve
the inverse geophysical problems, like migration algorithms [18, 19]. Recently, some
steps were done to construct the compact schemes with narrow spatial stencils [20].
In this work, it was used to obtain a numerical solution of elastic part of the problem
considered. Numerical simulations of the dynamic scattering process for subsurface
layered and block objects in elastic 2D and 3D media were carried out using modern
high-performance computing systems.
The chapter is organized as follows: Section 13.2 discusses a mathematical model.
The layered model system of equations and block model system of equations are
presented in Sects. 13.3–13.4, respectively. Numerical method is considered in
Sect. 13.5. Section 13.6 provides the simulation results. Section 13.7 concludes the
chapter.

13.2 Mathematical Model

Nonlinear interaction conditions between contacting structural elements may be for-

mulated. In Cartesian system xi (i = 1, 2, 3), the unbounded elastic medium with
the oriented system of periodic parallel slip planes is considered. The orientation is
set with the unit vector of normal n. The distance between slip planes is constant
and equal to ε. The density of the material ρ and Lame moduli λ and μ are known

[email protected]
13 Methods for Calculating the Dynamics of Layered … 173

constants. The stress state of the medium is described with the stress tensor σ. The
shear stress vector at the slip plane is equal to τ = σ · n − (n · σ · n)n and the
normal stress is equal to σn = n · σ · n. We define the slip velocity vector γ and the
delamination velocity vector ω = ωn based on the discontinuity of tangential [Vτ ]
and normal [Vn ] velocity at contact boundaries: γ = [Vτ ]/ε, ω = [Vn ]/ε.
We assume the presence of thick interlayers between elastic layers d  ε; however,
we will not take it into account explicitly, but using slip conditions at compressed
layers’ boundaries. At the contact boundaries the special conditions are specified.
(1) Piecewise linear condition of slip and weak delamination.
When σn < 0 (compressed contact),

γ = τ̇/kγ , ω = 0. (13.1)

When σn ≥ 0 (weak delamination),

γ = τ̇/kγ , ω = σ̇n /kω , kγ /μ  1, kω /μ  1. (13.2)

(2) The slip condition for Coulomb friction with a low viscous additive and weak
delamination.

When σn < 0 (compressed contact), τ = q|σn |(γ/|γ| + ηγ) or expressing γ

through tensions τ provided by Eq. 13.3.

1 τ |τ|
γ=  − 1 , ω = 0 (13.3)
η |τ| q|σn |

When σn ≥ 0 (weak delamination),

γ = τ̇/kγ , ω = σ̇n /kω , kγ /μ  1, kω /μ  1. (13.4)

Here, kγ i kω are the coefficient of weak elastic tangential and normal bond of
layers, η is the viscosity coefficient, q is the Coulomb friction coefficient, F(y) =
F(y)H (y), H (y) is the Heaviside function, H (y) = 0 if y < 0, H (y) = 1 if
y ≥ 0. The contact plane with the interaction condition defined is called the slip-
delamination plane.
It should be noticed that previously in [2] for the delamination regime the “full
delamination condition” was used ≥ 0: τ = σn = 0 that is the asymptotic

case of the “weak delamination” when kω → 0, kγ → 0, = [u n ] ε is the
normalized discontinuity of normal displacements at the contact boundary defined
by the equation ˙ = ω . Also, for the case of the weak delamination ω = σ̇n /kω ,
inequalities σn ≥ 0 and ≥ 0 are equivalent.
From the numerical point of view, small parameters kγ i kω are regularizations that
allow us to prevent the oscillations occurrence, when sharply changing the compress
boundary to the full delamination condition.

[email protected]
174 I. S. Nikitin et al.

To construct the continuum model with a set of these slip-delamination planes,

we are going to deal with γ and ω as discontinuous functions of space and time.
Also, we will use main relationships from the slippage theory as many other authors.
It allows us to take into account contributions from γ and ω into nonelastic strain
tensors eγ and eω , respectively:

eγ = (n ⊗ γ + γ ⊗ n)/2, γ · n = 0, (13.5)

eω = (n ⊗ ω + ω ⊗ n)/2 = ωn ⊗ n, ω = ωn. (13.6)

The full strain tensor e equals to the sum of elastic and nonelastic parts provided
by Eq. 13.7.

e = ee + eγ + eω , e = (∇v + ∇v T )/2 (13.7)

Here, v is « macroscopic » velocity of medium particles, ee is the elastic strain

tensor accordingly to Hooke’s law:

σ̇ = λ(ee : I)I + 2μee . (13.8)

The final equation is the motion equation in the view of Eq. 13.9.

ρ v̇ = ∇ · σ (13.9)

13.3 Layered Model System of Equations

In layered medium containing a set of elastic layers, the only system of slip-
delamination planes are possible with the normal n. If the normal to contact
boundaries n is oriented along the x 2 , then its components are n j = δ 2j .
Through conditions for γ and ω corresponding to the local contact conditions
(Eq. 13.1), we can write Eq. 13.10.

γ j = σ̇2 j /kγ , ω = σ̇22 H (σ22 )/kω (13.10)

Through conditions for γ and ω corresponding to the local contact conditions

(Eq. 13.2), we can write Eqs. 13.11–13.12.

1 σ2 j |τ|
γj =  − 1(1 − H (σ22 )) + σ̇2 j H (σ22 )/kγ (13.11)
η |τ| q|σ22 |

[email protected]
13 Methods for Calculating the Dynamics of Layered … 175

ω = σ̇22 H (σ22 )/kω , |τ| = σ2k σ2k (13.12)
k =2

Based on the chosen normal direction, the final system of equations for this model
is represented by Eqs. 13.13–13.15.

ρ v̇i = σi j, j , σ̇ii = λvk,k + 2μvi,i − λω (13.13)

i =2

σ̇22 = λvk,k + 2μv2,2 − (λ + 2μ)ω, σ̇i j = μ(vi, j + v j,i ) (13.14)

i, j =2

σ̇2 j = μ(v2, j + v j,2 ) − μγ j , i = j (13.15)

j =2

13.4 Block Model System of Equations

The block medium consists of parallelepiped elastic elements with three possible
slip-delamination planes. These planes are defined with normals n(s) , s = 1,2,3. In
this case, the nonelastic strain tensors are calculated by Eqs. 13.16–13.17.


3
eγ = (n(s) ⊗ γ(s) + γ(s) ⊗ n(s) )/2 , γ(s) · n(s) = 0 (13.16)
s=1

eω = (n(s) ⊗ ω(s) + ω(s) ⊗ n(s) )/2 = ω(s) n(s) ⊗ n(s) , ω(s) = ω(s) n(s) (13.17)

If three normals to slip-detachment planes are oriented along the coordinate axis
of Cartesian system, then n (s)
j = δ j , where δ j is the Kronecker’s symbol.
s s
(i) (i)
Through conditions for γ and ω corresponding to the local contact conditions
(Eq. 13.1), formulae included in Eq. 13.18 are given as follows:

γ j(i) = σ̇i j /kγ , i = j, ω(i) = σ̇ii H (σii )/kω . (13.18)

Through conditions for γ(i) and ω(i) corresponding to the local contact conditions
(Eq. 13.2), formulae included in Eqs. 13.19–13.20 are given as follows:
 
1 σi j τ(i) 
γ j(i) =    − 1(1 − H (σii )) + τ̇i j H (σii )/kγ , i = j, (13.19)
η τ(i)  q|σii |

 
ω(i) = σ̇ii H (σii )/kω , τ(i)  = σik σik . (13.20)
k =i

[email protected]
176 I. S. Nikitin et al.

As for the layered medium, we can rewrite the main system in the suitable form
of Eqs. 13.21–13.22.

ρ v̇i = σi j, j , σ̇ j j = λvk,k + 2μv j, j − λ ω(l) − (λ + 2μ)ω( j) (13.21)
l= j

( j)
σ̇i j = μ(vi, j + v j,i ) − μγ j(i) − μγi ,i = j (13.22)

13.5 Numerical Method

Both formulated systems are semi-linear hyperbolic systems, and the numerical solu-
tion can be obtained with different explicit schemes. However, the slippage process
switches on the nonlinear free term with small viscosity parameter in the denomina-
tor. The system transforms into the form with small parameter and ordinary explicit
schemes will not be stable. To overcome this problem, the use of explicit–implicit
method is proposed. The implicit approximation is used only for equations that con-
tain small term in the denominator. All other equations are approximated with the
explicit scheme.
Let us describe this approach for σ̇2 j for compressed contact case, σ22 < 0 for
the layered medium using Eq. 13.23.
 
σ̇2 j = μ(v2, j + v j,2 ) − μσ2 j |τ| (q|σ22 |)−1 (η|τ|) (13.23)
j =2

Implicit approximation with first-order time approximation has a view of

Eqs. 13.24–13.25.
   n+1  
(σ2n+1
j − σ2 j )
n
t = μ(v2,
n+1
j + v j,2 ) − μσ2 j 
n+1 n+1 n+1
e (q σ22 )−1 (η n+1
e )
(13.24)

 n+1 2  n+1 2
n+1
e = σ12e + σ32e (13.25)

Here, indices n +1 and n mean the current and next time layers, respectively, t is
the time step. We assume that values vin+1 and σ jn+1
j were calculated with the explicit
schemes for elastic equations. Solving this equation and also the same equation for
normal tensions we can obtain correcting formulas.
When σ22en+1
< 0 (compressed contact),

σ22
n+1
= σ22e
n+1
, σiin+1 = σiie
n+1
, ω = 0, i = 1, 3. (13.26)
 n+1 
When n+1
e ≥ q σ22e ,

[email protected]
13 Methods for Calculating the Dynamics of Layered … 177
 n+1  n+1     n+1 
σi2n+1 = q σ22e  σ
i2e
n+1
e 1+δ n+1
e 1 + δq σ22e  , (13.27)

γi = (σi2e
n+1
− σi2n+1 )/(μ t), i = 1, 3. (13.28)
 n+1 
When n+1
e < q σ22e ,

σi2n+1 = σi2e
n+1
, γin+1 = 0, i = 1, 3, δ = η/(μ t). (13.29)

When σ22e
n+1
≥ 0 (delamination process),

1 1 λ(σ22e
n+1
− σ22
n
)
σ22
n+1
= (βσ22e
n+1
+ σ22
n
), σiin+1 = σiie
n+1
− ,
(1 + β) (1 + β) (λ + 2μ)
1 (σ22e
n+1
− σ22
n
)
ω= , i = 1, 3, (13.30)
(1 + β) (λ + 2μ) t
1 1 (σi2e
n+1
− σi2n )
σi2n+1 = (ασi2e
n+1
+ σi2n ), γi = , i = 1, 3. (13.31)
(1 + α) (1 + α) μ t

Coefficients for weak shear α = kγ /μ and stretching β = kω /(λ + 2μ), where

j e = σi j + λvk,k δi j + μ(vi, j + v j,i )
α, β < < 1, σin+1 n n+1 n+1 n+1
t is the stress value after
the elastic step. This formula, in fact, is the adjustment of “elastic” stresses for the
“friction cone” with viscous corrections. We used here the simplest formula for σ2n+1 je
only to illustrate the derivation process.
Let us describe precisely the elastic step of the calculation algorithm. The linear
dynamic elasticity equations are given by Eq. 13.32.

ρ v̇i = σi j, j , σ̇i j = λvk,k δi j + μ(vi, j + v j,i ) (13.32)

Here, λ and μ are the Lamé constants and δi j is the Kronecker delta.
The first line in the system presents three equations of motion, while the sec-
ond line presents six rheological relations. The vector of variables consists of nine
components and has the form of Eq. 13.33.

u = (v1 , v2 , v3 , σ11 , σ12 , σ13 , σ22 , σ23 , σ33 )T (13.33)

Note that the solid mechanic system can be written in matrix form as Eq. 13.34,
where Aj are 9 × 9 matrices and (x1 , x2 , x3 ) is the orthonormal system of coordinates.

∂u 
3
∂u
= Aj (13.34)
∂t j=1
∂ xj

[email protected]
178 I. S. Nikitin et al.

This system is solved using the grid-characteristic method on parallelepiped

meshes. It splits up into three one-dimensional systems of equations:

∂u ∂u
= Aj , j = 1, 2, 3. (13.35)
∂t ∂x j

Each system is a hyperbolic and possesses a complete set of eigenvectors with

real eigenvalues. Each of the systems can be rewritten by Eq. 13.36, where the matrix
j is composed of the eigenvectors and j is a diagonal matrix.

∂u ∂u
= −1
j jj (13.36)
∂t ∂ξ j

At the splitting step in the fixed direction, the matrix j is given by

 j = diag(c1 , −c1 , c2 , −c2 , c2 , −c2 , 0, 0, 0), (13.37)

where

c1 = (λ + 2μ)/ρ, c2 = μ/ρ. (13.38)

After changing to the variables v =  u, each of the systems splits into nine inde-
pendent scalar advection equations (in what follows, the index j is omitted wherever
possible):

∂v ∂v
+ = 0. (13.39)
∂t ∂ξ j

The one-dimensional advection equations are solved using the method of charac-
teristics. After all the components of v are advected, the solution is recovered using
Eq. 13.40.

un+1 = −1 vn+1 (13.40)

The underlying computer code involves schemes of second to fourth orders of

accuracy. In this study, we used the fourth-order accurate scheme (ζ = t/ h, h is
spatial coordinate step):

vmn+1 = vmn − ζ ( 1 − ζ( 2 − ζ( 3 −ζ 4 ))),

1 = (−2vm+2
n
+ 16vm+1
n
− 16vm−1
n
+ 2vm−2
n
)/24,

2 = (−vm+2
n
+ 16vm+1
n
− 30vmn + 16vm−1
n
− vm−2
n
)/24, (13.41)

[email protected]
13 Methods for Calculating the Dynamics of Layered … 179

3 = (2vm+2
n
− 4vm+1
n
+ 4vmn − 2vm−2
n
)/24,

4 = (vm+2
n
− 4vm+1
n
+ 6vmn − 4vm−1
n
+ vm−2
n
)/24.

Additionally, we used a grid-characteristic monotonicity criterion. For positive

components of diagonal matrix  j , it has the form of Eq. 13.42.

min{vmn , vm−1
n
} ≤ vmn+1 ≤ max{vmn , vm−1
n
} (13.42)

For negative components of diagonal matrix  j , it is symmetric. In the simplest

case when this criterion is violated, the solution is corrected as follows:
⎧
⎨ max{vmn , vm−1
n
}, vmn+1 > max{vmn , vm−1
n
},
vmn+1 = min{vm , vm−1 }, vm < min{vm , vm−1 },
n n n+1 n n (13.43)
⎩ n+1
vm , min{vmn , vm−1
n
} ≤ vmn+1 ≤ max{vmn , vm−1
n
}.

This limiter preserves the fourth order of the scheme in domains, where the solu-
tion is fairly smooth (the characteristic criterion is satisfied). In the case of high
solution gradients, the order of the scheme is reduced to the third. Parallel algorithms
for high performance computing systems were used [21].

13.6 Simulation Results

In Fig. 13.1, two 2D wave fields of scattered seismic waves on the layered structure
with different orientations for contact condition (Eq. 13.2) are shown.
The influence of the layer orientation on the scattered wave field is clearly seen.
Seismic response is nonsymmetric for non-vertical layers orientation. Based on
this fact, it is possible to make some assumptions about the main direction in the
subsurface fractured medium.
In Figs. 13.2 and 13.3, 3D wave fields of scattered seismic waves on layered and
block structures for contact condition (Eq. 13.1) are shown. The whole medium was a

Fig. 13.1 2D-scattered seismic waves on the fractured cluster: a vertical cracks orientation, b 45°
cracks orientation

[email protected]
180 I. S. Nikitin et al.

Fig. 13.2 3D-scattered seismic waves for time moment t = 0.56 s: a on the block, b on the layered
clusters

Fig. 13.3 3D-scattered seismic waves for time moment t = 0.8 s: a on the block, b on the layered
clusters

parallelepiped with sizes 10 × 10 × 3 km. P-wave velocity was equaled to 4500 m/s,
S-wave velocity was equaled to 2250 m/s, density was equaled to 2500 kg/m3 . The
day surface condition was used on the upper side. In the center of the medium along
axes OX and OY with the depth 50 m the point source with the 30 Hz Ricker time
dependence function was applied. At the 2 km depth, the object was set with sizes 3
× 3 × 0.2 km. Two different models were compared: the block and layered with the
normal vector along the axis OX. The parallelepiped mesh with 5 m cells containing
2.4 billion of nodes was constructed. The time step was 1 ms, and totally 2,000 steps
were done. To visualize results we save each 40 time steps: 3 slices, 2 main vertical
slices, and one horizontal equals to the day surface. The total computation time was
5.5 h on 100 computation cores for both mathematical models.

[email protected]
13 Methods for Calculating the Dynamics of Layered … 181

As expected, the amplitude of the seismic response for the block medium is
significantly higher due to the existence of horizontal contact planes that reflect the
incident wave (see Figs. 13.2a, 13.3a). Also, 3D asymmetry of the response from
the layered medium is clearly seen (see Figs. 13.2b, 13.3b). It is explained by the
presence of the preferred direction along the normal vector.
We carried out the dynamic analysis of the seismic signal registered at the day sur-
face for both models. Spatial distributions for first, second, third, and fourth responses
are depicted in Figs. 13.4, 13.5, and 13.6. The first response is the P-wave reflected
from the top boundary of the cluster. It has the smallest arrival time. The second
response is the S-wave generated due to the interaction of the initial P-wave with the
cluster. The third and fourth responses are the P-wave with two times more depth
path length and the S-wave generated while interacting with cluster, respectively.
Due to the specific parameters of this model, these two signals are very close, so
it is quite difficult to split them. One of the major features of all responses is the
symmetry for block model and asymmetry for layered model. It is clearly seen at the
wave patterns with OXY plane.

Fig. 13.4 Registration (modulus of velocity) of the first seismic response at the day surface: a OXZ
slice, b OXY slice for block medium, c OXY slice for layered medium

Fig. 13.5 Registration (modulus of velocity) of the second seismic response at the day surface:
a OXZ slice, b OXY slice for block medium, c OXY slice for layered medium

[email protected]
182 I. S. Nikitin et al.

Fig. 13.6 Registration (modulus of velocity) of the third and fourth seismic response at the day
surface: a OXZ slice, b OXY slice for block medium, c OXY slice for layered medium

13.7 Conclusions

Continual models of solid media with a discrete set of slip planes (layered, block
media) and with nonlinear slip conditions at the contact boundaries of structural
elements were constructed. For a stable numerical solution of a system of differential
equations, an explicit–implicit method is proposed with an explicit approximation of
the equations of motion and an implicit approximation of the constitutive relations
containing a small parameter in the denominator. A set of effective formulas for the
stress tensor correcting at the “elastic” step was obtained. Numerical simulations of
the dynamic scattering process for subsurface layered and block objects in elastic 2D
and 3D media were carried out using modern high-performance computing systems.

Acknowledgements This work has been carried out using computing resources of the federal
collective usage center Complex for Simulation and Data Processing for Mega-science Facilities
at NRC “Kurchatov Institute”, http://ckp.nrcki.ru/. This work was carried out with the financial
support of the Russian Science Foundation, project no. 19-71-10060.

References

1. Nikitin, I.S.: Dynamic models of layered and block media with slip, friction and separation.
Mech. Solids 43(4), 652–661 (2008)
2. Burago, N.G., Zhuravlev, A.B., Nikitin, I.S.: Continuum model and method of calculating for
dynamics of inelastic layered medium. Math. Model. Comput. Simul. 11(3), 59–74 (2019)
3. Burago, N.G., Nikitin, I.S.: Improved model of a layered medium with slip on the contact
boundaries. J. Appl. Math. Mech. 80(2), 164–172 (2016)
4. Nikitin, I.S., Burago, N.G., Nikitin, A.D.: Continuum model of the layered medium with
slippage and nonlinear conditions at the interlayer boundaries. Solid State Phenom. 258, 137–
140 (2017)
5. Virieux, J., Calandra, H., Plessix, R.E: A review of the spectral, pseudo-spectral, finite-
difference and finite-element modelling techniques for geophysical imaging. Geophys.
Prospect. 59(5), 794–813 (2011)
6. Carcione, J.M., Herman, C.G., Kroode, A.P.E: Seismic modeling. Geophysics 67(4), 1304–
1325 (2002)

[email protected]
13 Methods for Calculating the Dynamics of Layered … 183

7. Burago, N.G., Nikitin, I.S., Yakushev, V.L.: Hybrid numerical method with adaptive over-
lapping meshes for solving nonstationary problems in continuum mechanics. Comput. Math.
Math. Phys. 56(6), 1065–1074 (2016)
8. Lisitsa, V., Tcheverda, V., Botter, C.: Combination of the discontinuous Galerkin method with
finite differences for simulation of seismic wave propagation. J. Comput. Phys. 311, 142–157
(2016)
9. Massau, J.: Memoire sur L’integration Graphique des Equations aux Derivess Partielles. F.
Meyer-van Loo, Ghent (1899)
10. Zhukov, A.I.: Using the Method of Characteristics for the numerical solution of one-
dimensional problems of gas dynamics. Tr. Mat. Inst. Akad. Nauk SSSR 58, 4–150
(1960)
11. Butler, D.S.: The numerical solution of hyperbolic systems of partial differential equations of
three independent variables. Proc. Roy. Soc. London. Ser. A 255(1281), 232–241 (1960)
12. Beklemysheva, K.A., Vasyukov, A.V., Golubev, V.I., Zhuravlev, Y.I.: On the estimation of
seismic resistance of modern composite oil pipeline elements. Dokl. Math. 97(2), 184–187
(2018)
13. Golubev, V.I., Golubeva, YuA: Full-wave simulation of the earthquake initiation process. CEUR
Work. Proc. 2267, 346–350 (2018)
14. Golubev, V., Khokhlov, N., Grigorievyh, D., Favorskaya, A.: Numerical simulation of destruc-
tion processes by the grid-characteristic method. Procedia Comput. Sci. 126, 1281–1288
(2018)
15. Muratov, M.V., Petrov, I.B.: Application of fractures mathematical models in exploration seis-
mology problems modeling. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov,
S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 years of the Development of
Grid-characteristic Method, SIST, vol. 133, pp. 120–131. Springer, Berlin (2019)
16. Favorskaya, A.V., Zhdanov, M.S., Khokhlov, N.I., Petrov, I.B.: Modeling the wave phe-
nomena in acoustic and elastic media with sharp variations of physical properties using the
grid-characteristic method. Geophys. Prospect. 66(8), 1485–1502 (2018)
17. Favorskaya, A.V., Kabisov, S.V., Petrov, I.B.: Modeling of ultrasonic waves in fractured rails
with an explicit approach. Dokl. Math. 98(1), 401–404 (2018)
18. Golubev, V.I.: The usage of grid-characteristic method in seismic migration problems. In:
Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov, S.S., Jain, L.C. (eds.) Proceedings
of the Conference on 50 years of the Development of Grid-characteristic Method, SIST, vol.
133, pp. 143–155. Springer, Berlin (2019)
19. Golubev, V.I., Voinov, O.Y., Petrov, I.B.: Migration of seismic data for multi-layered fractured
geological media using elastic approach. In: 19th Science and Applied Research Conference Oil
and Gas Geological Exploration and Development, pp. 43756.1–43756.5 (in Russian) (2017)
20. Khokhlov, N.I., Golubev, V.I.: On the class of compact grid-characteristic schemes. In: Petrov,
I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov, S.S., Jain, L.C. (eds.) Proceedings of the
Conference on 50 years of the Development of Grid-characteristic Method, SIST, vol. 133,
pp. 64–77. Springer, Berlin (2019)
21. Ivanov, A.M., Khokhlov, N.I.: Efficient inter-process communication in parallel implemen-
tation of grid-characteristic method. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N.,
Simakov, S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 years of the Development
of Grid-characteristic Method, SIST, vol. 133, pp. 91–102. Springer, Berlin (2019)

[email protected]
Chapter 14
Algorithms for Calculating Contact
Problems in the Solid Dynamics

Nikolay G. Burago , Ilia S. Nikitin and Alexander D. Nikitin

Abstract The chapter discusses the explicit and implicit non-matrix finite ele-
ment algorithms for calculating contact interactions between elastic–plastic bodies.
We consider Lagrangian contact algorithms that are based on Lagrange multipli-
ers (explicit methods) and penalty functions (implicit methods). Examples of the
calculation of contact interactions during high-speed processes of the collision of
elastic–plastic bodies and explosion welding of tubular samples are presented.

14.1 Introduction

The development of contact algorithms for elastic–plastic bodies started more than
half a century back together with the advent of computers and numerical methods.
Reviews of contact algorithms can be found in [1–9].
Here, Lagrangian contact algorithms are considered. Such contact algorithms are
the parts of solution method, which are responsible for detection, tracking, and cal-
culation of contact interaction between deformable bodies. In most of the cases, the
boundary conditions for problems in solid mechanics belong to the following three
types: the boundary conditions with predefined displacements or velocities, boundary
conditions with predefined loads from external bodies that are not involved in calcu-
lation, and boundary contact conditions that define contact boundaries, their motion,
and contact loads during interaction of considered deformable bodies between each

N. G. Burago (B)
Ishlinsky Institute for Problems in Mechanics of the RAS, pr. Vernadskogo 101, b1, Moscow
119526, Russian Federation
e-mail: [email protected]
I. S. Nikitin · A. D. Nikitin
Institute of Computer Aided Design of the RAS, Vtoraya Brestskaya ul., 19/18, Moscow 123056,
Russian Federation
e-mail: [email protected]
A. D. Nikitin
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 185

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_14

[email protected]
186 N. G. Burago et al.

other. In most cases, a contact boundary is not predefined in advance. It can be vari-
able in time and ought to be found and calculated along with and as a part of general
problem about deformation of bodies.
Most of proposed contact search algorithms are overviewed in [4]. Search algo-
rithms often use quite a significant part of calculations required to solve the problem.
We tested many of such algorithms and the one that was chosen in our practice is
considered further.
In most of dynamic incremental (step by step in time) Lagrange algorithms,
the contact boundaries are detected by the penetration of boundary nodes through
alien boundary cells. That may happen after predictor time step, which is calculated
neglecting the contact. The contact zone detection is based on selection of the con-
tacting pairs “boundary cell—alien boundary node” if they are close enough to each
other (much less than local spatial cell size) or if a boundary cell is intersected by
a track of alien node motion. After contact pairs are found, some so-called “mas-
ter–slave” algorithm is used during correction stage to prevent penetration with help
of normal contact load. In explicit schemes, the correction stage is implemented
iteratively by means of circumventing the contact pairs calculating the contact loads
from the magnitude of penetration for eliminating it. At each time step for such cor-
rection, two run-arounds of boundary are enough. In implicit schemes, the transition
from Lagrange multipliers to penalty functions is done by representing Lagrange
multiplier (normal contact load) as a product of the distance of penetration by the
penalty factor. As a result, the positive definiteness of the discrete operator (“stiffness
matrix”) is preserved, but it is necessary to take into account that too large values of
the penalty factor lead to a worsening of the conditionality of the problem. There-
fore, to prevent loss of accuracy and provide convergence of iterative solutions, a
preconditioning should be used. For this purpose, the residuals of equations used in
iterations are multiplied by an approximate inverse matrix of the system of algebraic
equations. In order not to perform time-consuming operations of matrix inversion
for preconditioning, it is sufficient to use the matrix of inverse diagonal elements of
stiffness matrix, that is, in other words, to use the scaling of unknowns.
Contact pairs algorithms, in our opinion, are the simplest. They are very easily
implemented in the case of complex geometry of the solution domain when using a
system of Cartesian rectangular coordinates. The implementation within the frame-
work of Galerkin–Petrov variational formulation with reduced requirements for the
smoothness of a generalized solution is especially simple (when the equations con-
tain only the first derivatives of the desired functions). Then the simplest piecewise
linear approximation on finite elements in the form of rectangular parallelepipeds
gives quite good results.
In this chapter, we consider in detail the algorithms of Lagrange multiplier meth-
ods for explicit schemes, and algorithms of penalty functions for implicit schemes
because in combination with a finite element piecewise linear approximation of
the solution. These algorithms are very easy to implement and provide good quality
solutions even on personal computers.
This chapter is organized as follows. Statement of general contact problem is
considered in Sect. 14.2. Numerical algorithm based on matrix-free finite element

[email protected]
14 Algorithms for Calculating Contact Problems … 187

method is described in Sect. 14.3. Contact calculation using Lagrange multipliers

technique is highlighted in Sect. 14.4. Application of penalty functions to contact
computations is presented in Sect. 14.5. The effectiveness of the algorithms is illus-
trated by two non-trivial examples in Sect. 14.6: the impact of two plates at an
angle and axisymmetric welding of two dissimilar tube samples under the action of
a detonation wave. Section 14.7 concludes the chapter.

14.2 Mathematical Model

The system of equations describing the behavior of an elastic–plastic medium is used

here in the simplified variant [10]. The laws of conservation of mass and momentum
as well as the kinematic relations are written as Eq. 14.1.

du
ρ = ρ0 det(F−1 ), ρ = ∇ · σ + ρg
dt
1 1
F−1 = ∇x0 ε = (I − F−T · F−1 ) e = (L + LT )
2 2
dε dx
e= + ε · L + LT · ε L = ∇u =u (14.1)
dt dt

Here, ρ is the mass density, u is the velocity, t is the time, x and x0 are the radius
vectors of material point in actual and initial states, F is the strain gradient, L is the
velocity gradient, ε is the Almansi strain tensor, e is the Eulerian strain rate tensor,
σ is the Cauchy stress tensor, d/dt is the material time derivative, ∇ is the spatial
differentiation operator in actual state, I is the unity tensor, and g is the body force
density.
Constitutive equations are used according to Prandtl–Reuss plastic flow theory
and Mises plasticity condition and expressed by Eq. 14.2.

σ = − pI + σ σ = 2μ(ε − εp )
ρ ρ σ : e
p=K ln ep = H ( p ) 2 σ (14.2)
ρ0 ρ0 ks

Here, p is the pressure, σ is the deviatoric stress, ε , εp , and ep are the deviatoric
strain, plastic strain, and plastic strain rate tensors, respectively,  p = σ : σ − ks2 ≤
0, H is the Heaviside function, ks is the yield radius, K and µ are the elastic moduli.
The solution domain in the general case consists of several (spaced apart in space
for the problems of dynamics) subdomains corresponding to the deformable bodies
under consideration. On a part of the surface, interaction with external bodies is
specified in the form of loads:

t ≥ 0, x ∈ S p : σ · n = P∗ (x, t) (14.3)

[email protected]
188 N. G. Burago et al.

or velocities

t ≥ 0, x ∈ Su : u = u∗ (x, t). (14.4)

On the rest of the surface, so-called contact surface Scont = (S\Su )\S p , velocities
and loads are due to the interaction of the bodies in question with each other. Unknown
in advance contact surface Scont is the set of all points that have different initial
(Lagrangian) coordinates and coincident actual coordinates:

x− , x+ ∈ Scont : ∃x−
0
= x+
0
|x− = x(x−
0
, t) = x+ = x(x+
0
, t).

Loads P and velocities u on the contact surface Scont are determined by the contact
conditions that make up the continuity of the normal velocity, Newton’s third law
and the law of friction, respectively (Eq. 14.5).

(u+ − u− ) · n+ = 0
P+ = −P−
Pτ i = f (Pn , (u+ − u− ) · τi+ ) (14.5)

Here, n, τi are the normal and tangent unity vectors to the surface S, respectively,
and

n+ = −n− , Pn = P · n, τi+ = −τi− , Pτ i = P · τi , (i = 1, 2).

Initial conditions are the following:

x ∈ V, t = 0 : x = x0 , u = u0 , ε p = 0. (14.6)

Thus, it is required to solve the initial boundary value problem for the system of
equations (Eqs. 14.1, 14.2) with the boundary (Eqs. 14.3–14.5) and initial (Eq. 14.6)
conditions.

14.3 Numerical Method

The variation form of the equations of motion (the equation of virtual works) has the
form of Eq. 14.7.
 
du
ρ( − g) · δud V + σ : ∇δud V
dt
V V

[email protected]
14 Algorithms for Calculating Contact Problems … 189

  2 

= p · δud S − δ Pn (x∗ − x) · nd S − Pτ i (δx∗ − δx) · τi d S (14.7)
Sp Scont i=1 S
cont

Conditions (Eq. 14.3) are taken into account in the integral over the boundary with
a given surface load (the first integral in the right-hand side). Conditions (Eq. 14.4)
on a part of the surface with given velocities are taken into account as complemen-
tary boundary conditions. The conditions on the contact boundary (Eq. 14.5) are
introduced into the variational equation of virtual works using Lagrange multipliers
method (the next integrals in the right-hand side). In the integrals over the contact
surface in (Eq. 14.7), x∗ = x(x0∗ , t) denotes the intersection point of the continuation
 
of the outer normal (x + αn, α ≥ 0) with the surface point x = x(x0 , t) x0 = x0∗ .
If such point x∗ does not exist, i.e., if the line that continues outside normal does not
cross S, then assume x∗ = x + n. For all x ∈ S, the following inequality must be
satisfied:

(x∗ − x) · n ≥ 0. (14.8)

This inequality expresses the condition of non-penetration of one part of the

surface into another. Equality corresponds to the points of the contact surface x∗ , x ∈
Scont . The role of Langrage multiplier for inequality (Eq. 14.8) is played by the normal
contact load Pn (x, t) to be determined.
In the solution domain, we introduce a grid of elements consisting of tetrahe-
dra, prisms, or parallelepipeds in the three-dimensional case and of triangles and
quadrilaterals in the two-dimensional case. Let xi (i = 1, 2, . . . , Nv ) be the nodal
radius vectors, C(k, l) (k = 1, 2, . . . , Nc ; l = 1, 2, . . . , Mc ) be the nodal numbers
in elements, and B(k, l) (k = 1, 2, . . . , Nb ; l = 1, 2, . . . , Mb ) be the nodal num-
bers in boundary elements. Let the bypass of the boundaries of the surface elements
take place clockwise for the external observer. On the time layer, n introduces the
notation: uin , xin are the velocity and nodal radius-vector, and [ε p ]nk is the plastic
strain tensor in the element center. Let ω be the set of nodal numbers, ωu be the
set of boundary nodal numbers with predefined velocities, and  be the set of finite
element numbers.
Piecewise linear finite element approximation of coordinates, displacements, and
velocities is used. The stress, strain, and strain rate tensors are calculated in the cen-
ters of the elements ([σ]nk , [ε]nk , [e p ]nk , and [e]nk ). Integrals of a variational equation
that do not contain time derivatives are replaced by sums of integrals over finite ele-
ments by quadrature formulas with Gaussian points at the centers of finite elements.
For integrals with time derivatives in finite elements, the quadrature formulas with
Gaussian points at the nodes of the grid are used. This makes the mass matrix diago-
nal for fast processes and provides the use of explicit finite element time integration
schemes. Thus, we obtain the discrete form of Eqs. 14.1, 14.2 provided by Eq. 14.9.

Mi (uin+1 − uin γin − (1 − γin )

uin ) = fin tn i ∈ ω\ωu

[email protected]
190 N. G. Burago et al.

uin+1 = u∗i
n+1
i ∈ ωu
xin+1 = xin + uin+1 tn i ∈ ω
[εp ]n+1
k = [ε ]n+1
k − [σ ]n+1
k /2/μ k ∈  (14.9)

To calculate new stress–strain state in each finite element k ∈ , the operations

are performed using Eq. 14.10.


MC
1
[F−1 ]nk = [∇]nk j xC(k,
0
j) [ε]k =
n
(I − [F−T ]n+1
k · [F−1 ]n+1
k )
j=1
2


MC
1
[L]nk = [∇]nk j uC(k,
n
j) [e]k =
n
([L]nk + [LT ]nk )
j=1
2


MC
1
[L]nk = [∇]nk j uC(k,
n
j) [e]k =
n
([L]nk + [LT ]nk )
j=1
2

[ε ]nk= [ε]nk − I[ε]nk : I/30 [ρ]nk = ρ0 det([F−1 ]nk )

[ρ]nk [ρ]nk
[ p]nk = K ln 0 [σ ]nk = 2μ([ε ]nk − [εp ]nk )
ρ0 ρ0
[σ]nk = −[ p]nk I + [σ ]nk [σ̃ ]n+1
k = [σ ]nk + 2μ([e ]nk − [ep ]nk ) tn

βkn = ks /max{ks , [σ̃ ]n+1k : [σ̃ ]n+1
k } [σ ]k
 n+1
= βkn [σ̃ ]n+1
k


MC
1
[L]n+1
k = [∇]n+1
k j uC(k, j) [e]k
n+1 n+1
= ([L]nk + [LT ]nk )
j=1
2

[ε ]n+1
k = [ε]n+1
k − I[ε]n+1
k : I/3 (14.10)

Here, the plastic flow is calculated by using method of Wilkins [1]. As an additional
viscosity included on shock waves, a variable Lax viscosity is used in the equation
of motion. The average speed on the old time layer n is calculated as follows:

i ∈ ω, j ∈ ωi : 
uin = 0.5(max unj + min unj ).
j∈ωi jinωi

Here ωi are the numbers of neighbors of node i. The hybrid viscosity parameter
that regulates Lax viscosity is given by γin = min{1, γ0 + κ0 |uin − uin |}, where
−1
γ0 = u s /c, κ0 = (0.2u s ) , u s = max ui − min ui , c = d P/dρ + 4/3μ/ρ.
n n 2
i∈ω i∈ω
Parameter γkn includes Lax viscosity in the vicinity of shock waves and increases
it to the nominal value with increasing impact velocity. Value Mi is nodal mass
calculated by the following expression:

[email protected]
14 Algorithms for Calculating Contact Problems … 191


NC 
MC
Mi = [V ]nk ρk MC−1 H̃ (i − C(k, l)), i ∈ ω.
k=1 l=1

The right sides of the discrete equations of motion are determined by the con-
tributions from the internal forces, external loads, and contact loads provided by
Eq. 14.11.


Nc 
Mc 
Nb 
Mb
fin = [g1 ]nkl H̃ (i − C(k, l)) + [g2 ]nkl H̃ (i − B(k, l))+
k=1 l=1 k=1 l=1


ND 
MD
+ [g3 ]nkl H̃ (i − D(k, l)) (14.11)
k=1 l=1

Here, function H̃ is equal to one for zero argument and zero otherwise, N D , M D are
the numbers of boundary contact elements and the number of nodes in the contact
elements. The contributions from the internal and specified external loads are as
follows:

[g1 ]nkl = −[V ]nk [σ]nk · ∇kln , (k = 1, 2, . . . , NC , l = 1, 2, . . . , MC ),

[g2 ]nkl = Mb−1 [P]nk Skn , (k = 1, 2, . . . , Nb , l = 1, 2, . . . , Mb ).

The calculation of contributions from contact loads [g3 ]nkl is considered further.
Explicit method is stable under usual Courant stability condition:
⎛ ⎞
1
tn = min ⎝ n ⎠, (14.12)
k in ck (max(|[∇x ]nkl |, |[∇ y ]nkl |, |[∇z ]nkl |)))−1
l

where [∇x ]nkl ex + [∇ y ]nkl e y + [∇z ]nkl ez is the discrete spatial differentiation operator.
At high impact velocities (of the order of the speed of sound), to ensure accuracy,
an additional time step restriction was introduced expressing the requirement that
deformation increments be sufficiently small:
−1
 1/2
tn ≤ εmax max [e]nk : [e]nk ) . (14.13)
k∈

Here, εmax = 0.1εs is the maximum permissible deformation increment at a

time step, εs is the yield point deformation.

[email protected]
192 N. G. Burago et al.

14.4 Use of Lagrange Multipliers

Let the surface of the bodies be represented by boundary cells: segments in two-
dimensional case and triangles in three-dimensional case. Let the local numbering
of the nodes in the boundary cells in three-dimensional case be taken clockwise if
viewed from the outside of the body, and in two-dimensional case the local numbering
corresponds to the bypass of the boundary clockwise. This is necessary to uniquely
determine the direction of the outer normal to the boundary.
At the beginning of each time step, a preliminary calculation of the new position
of the boundary without taking into account the contact is made. Then, among the
“boundary cell—boundary node” pairs, contact ones are selected, that is, such that
the normal omitted from the boundary node on the boundary cell plane, crosses this
boundary cell, and the countable penetration occurs. If there are several candidates for
a given boundary node from the boundary elements for the role of partner in a contact
pair, the boundary element that is intersected by the trajectory of this boundary node
is uniquely selected. The mathematical record of these selection conditions is given
below.
To speed up the process of selecting contact pairs, firstly we reject too far from
each other situated potential contact partners. Such selection is carried out first by
the difference of coordinates (more economical check), then by distance and only
then later by projection.
The contact pair forms a triangle in two-dimensional problems and a tetrahedron
in three-dimensional case. That is, the number of nodes in it MC equals to 3 or
4, respectively. The numbers of these nodes are stored in the information array of
contact pairs C(k, l), (k = 1, . . . , NC ; l = 1, . . . , MC )). For each contact pair k,
first MC − 1 numbers of information array C(k, l) correspond to nodes of contact
boundary element and the last MC th number corresponds to alien contact boundary
node.
The algorithm for calculating the contact load is implemented iteratively. In each
contact pair, the contact load normal to the boundary cell is determined from the
condition that the volume of the contact pair vanishes. In the process of sequential
bypass of contact pairs, the contact load and coordinates of the nodes get better until
volumes of all the contact pairs become zero. In explicit schemes when implement-
ing Lagrange multipliers method, this is achieved, as a rule, in two run-arounds of
the border (the second run-around is done for control). In implicit schemes when
implementing the penalty method (it is described in the next section), the velocity
field is iterated to the convergence in the entire solution domain.
Below we write out the formulas for calculating the contact pair. Let xin be the
current radius vectors of the contact pair of nodes on the “old” time layer, uin+1 and
(xin+1 )s = xin + (uin+1 )s tn are the desired node values of the speeds and radius
vectors on the new time layer, respectively, and s = 0, 1, 2, . . . is the iteration
number. The value (uin+1 )0 corresponds to the preliminary calculation of the new
time layer without taking into account the contact.

[email protected]
14 Algorithms for Calculating Contact Problems … 193

In three-dimensional case, the outer normal to the boundary cell is determined by

the ratio:

(n)s = ((x3n+1 )s − (x1n+1 )s ) × (((x2n+1 )s − ((x1n+1 )s )/L .

Here, L = |((x3n+1 )s − (x1n+1 )s ) × (((x2n+1 )s − ((x1n+1 )s )|. A pair of boundary

node—the boundary cell is a contact pair if the normal projection of the node on the
plane of the cell belongs to the cell:

L 1 + L 2 + L 3 = 1, (14.14)

d = n · ((x4n+1 )s − (x1n+1 )s ) ≤ 0. (14.15)

Here, values L i , (i = 1, 2, 3) are L-coordinates of projection (xcn+1 )s = (x4n+1 )s −

d(n) of boundary node MC onto the plane of boundary cell 1-2-3:

(L 1 )s = ((x3n+1 )s − (xcn+1 )s ) × (((x2n+1 )s − ((xcn+1 )s )/L ,

(L 2 )s = ((x3n+1 )s − (x1n+1 )s ) × (((xcn+1 )s − ((x1n+1 )s )/L ,
(L 3 )s = ((xcn+1 )s − (x1n+1 )s ) × (((x2n+1 )s − ((x1n+1 )s )/L .

For contact cell L-coordinates are non-negative. The contact integrals in the right
side of the equation of virtual works are represented in the following form:

 
Nc 
Mc
δ (Pn )n (u+ − u− ) · nn d S = (Pn )rn (L i )rn δu K (r,i) · nrn Srn
Scn r =1 i=1


Nc 
Mc
+ δ(Pn )r (L i )rn un+1
K (r,i) · nr Sr ,
n n

r =1 i=1
 
2 
Nc 
2 
Mc
(Pτ α )n (δu+ − δu− ) · τnα d S = (Pτ α )rn (L i )rn δu K (r,i) · (τα )rn Srn .
Scn α=1 r =1 α=1 i=1

Vectors of contact loads in discrete equations of motion are calculated by formulas:

 

Nc 
Mc 
2
(fc )in = (Pn )rn nrn + (Pτ α )rn (τα )rn Srn L rl
n
H̃ (i − K (r, l)).
r =1 l=1 α=1

Here, function H̃ is equal to one for zero argument and zero otherwise. A
consequence of the modified equation of virtual work is also the equality:

[email protected]
194 N. G. Burago et al.


Mc
(L i )rn un+1
K (r,i) · nr = 0
n

i=1

that is providing continuity of normal velocities at the contact boundary. Accounting

for contact leads to the emergence of a new group of unknown quantities—Lagrange
multipliers (Pn )rn (r = 1, . . . , Nc ) that satisfy NC additional algebraic relations lim-
iting the possible movements of contact nodes and characterized by a non-diagonal
matrix, and, hence, introducing an element of implicitness into explicit schemes.
Also the mutual dependence of the unknown contact zone and contact loads makes
the contact problem nonlinear and leads to the need for their iterative determination
(iteration by nonlinearity).
Normal contact loads must be compressive:

(Pn )rn ≤ 0, r = 1, . . . , NC .

Contact normal loads should act against the direction of the outer normal to the
boundary. If this condition is violated, the corresponding contact pair is excluded
from the set of contact pairs and recalculation is performed. To solve the system of
equations on the contact boundary, the simple Gauss–Seidel-type iterative process
is applied.
The algorithm described above is used as an addition to the explicit schemes for
calculating the deformation. Note that the contact algorithm does not depend on the
choice of a particular explicit scheme. At each time step, only two border bypasses
are made (the second is for control). We emphasize that all parts of the border are
equal (no master-slave borders). Therefore, the contact of some parts of the body
boundary with other parts of the same body boundary is calculated in the same way
as the contact with other bodies.
To improve the property of the scheme to maintain symmetry and ensure the
independence of the result of the order of bypassing contact pairs, the corrected
values of contact loads and speeds are accumulated in a separate array, and the main
arrays of coordinates and speeds are updated at the end of the calculation of contact
pairs.

14.5 Using Penalty Functions

In the case of implicit schemes, the contact interaction is accounted using the penalty
function method. A modified equation of virtual work (Eq. 14.7) is used, in which
it is assumed:

Pn = λ̃(x+ − x− ) · n+ , λ̃ >> 1.

[email protected]
14 Algorithms for Calculating Contact Problems … 195

Here, λ̃ is the penalty factor. In practical calculations, the penalty factor is taken to
be equal to the reciprocal of the “machine epsilon” (for four-byte arithmetic, this is
equal approximately 106 ). The approximation of contact additional terms in Eq. 14.7
is carried out in the same way as in Lagrange multipliers method. The influence of
contact members is taken into account in the iterative process of conjugate gradient
method (a matrix-free implementation of implicit schemes is used [11]). Before cal-
culating the new time layer, the contact zone is determined by conditions (Eqs. 14.14,
14.15), but the condition (Eq. 14.15) is weakened and replaced with the condition of
sufficient proximity of the boundaries:

d = n · ((x4n+1 )s − (x1n+1 )s ) ≤ 0.1h min . (14.16)

Here, h min is the minimum length of the boundary segment of the grid. Account-
ing for contact interactions in the case of an implicit scheme did not complicate the
solving process and was implemented simply as an additional subroutine defining
the contact area at each step and calculating contact terms of the equations of motion
in iterations using the conjugate gradient method.

14.6 Examples

In this section, the impact of two plates at an angle and axisymmetric welding of
two dissimilar tube samples under the action of a detonation wave are considered in
Sects. 14.6.1 and 14.6.2, respectively.

14.6.1 The Impact of Two Bodies at an Angle

Let the plate-impactor hits the plate-barrier with velocity u 0 /c(1) = 0.2, where c(1)
is a sound velocity in material of impactor that has the following properties:

K (1) = 975, μ(1) = 369, σs(1) = 1.0, dσs(1) /da p = 0.0, c(1) = 1,
kω(1) = 0.0, kθ(1) = 103 , F = εmax − εlim
(1) (1)
, εlim = 0.01.

Barrier material has the following properties:

K (2) = 243, μ(2) = 92, σs(2) = 0.25, dσs(2) /da p = 0, c(2) = 1,

kω(2) = 0.0, kθ(2) = 106 , F = εmax − εlim
(2) (2)
, εlim = 0.01.

Here, εmax is the maximal principal strain.

In Fig. 14.1a, the initial grid is visible. Figure 14.1b, c shows the contact interac-
tion. The target has fractured into three parts that are moving independently. Ricochet

[email protected]
196 N. G. Burago et al.

Fig. 14.1 Impact of plates: a initial grid, b contact evolution, c fracture evolution. The target has
fractured into three parts. The impactor is destroyed

of impactor that is completely destroyed is seen clearly. The method for calculating
destruction is described in [12].

14.6.2 Explosion Welding Problem

Earlier in [13], the problem of explosion welding was numerically solved with
allowance for large plastic deformations in order to study the wave formation process
at the boundary. Here, axisymmetric welding of two dissimilar metal tube samples
under the action of a detonation wave propagating along the outer surface of the tubes
is considered. The speed of the traveling detonation wave is comparable to the speed
of elastic waves, for example, in titanium and steel, which are used to make bimetallic
billets, but it is subsonic, providing a favorable impact mode for welding [14]. The
pressure in the detonation wave considerably exceeds the static yield strength of these
metals, which makes it possible to use the hydrodynamic approach for approximate
solutions [14, 15]. However, it should be noted that the dynamic yield strength at
high strain rate can be several times higher than the static one, and the neglect of the
strength effects of solids leads to a distorted estimate of the stress–strain state. When
modeling the contact interaction, the conditions for Coulomb friction were chosen
into which the welding criterion depending on the level of plastic deformations was
entered. Such a criterion does not contradict the well-known jet criterion of welding
since the formation of a cumulative jet is possible only in the case of developed
plastic flows in the vicinity of the collision front.
Figure 14.2 presents the results of the calculation of high-speed collisions of
titanium (external) and steel (internal) tubular samples. For better visibility of the
process, a regime with a large peak pressure in the detonation wave was chosen,

[email protected]
14 Algorithms for Calculating Contact Problems … 197

Fig. 14.2 Contour lines: a for σrr , b for σr z , c characteristic function Sv along z

which led to severe plastic deformations of the work-pieces, especially in the end
zone of the external titanium work-piece.
The following input data were taken. The maximum load in the detonation wave
σ0 = 6.0 · 10−2 , the velocity of the detonation wave cd = 0.45, the duration of
the loading pulse dim = 0.075, static yield strengths τ Fe = 0.0013, τT i = 0.0009,
relative thicknesses d Fe /R = 0.1, dT i /R = 0.075, the gap between the work-pieces
d0 /R = 0.05, the critical value of the intensity of plastic deformations γcr = 0.2.
Figure 14.2a, b shows the contour lines of the stress components σrr and σr z .
Quality of welding can be judged by the function Sv (Fig. 14.2c), which is defined
at the contact boundary of two bodies and takes values Sv = 1 in the contact zone
with welding, Sv = 0.5 in the contact zone without welding, and Sv = 0 outside
the contact zone. The function Sv shows that quality of welding of the examined
samples is good everywhere except for a small end zone, which is often observed in
the experiment (the “melting” of the end, which is usually cut).

14.7 Conclusions

The chapter discusses in detail contact algorithms based on Lagrange multiplier

method for explicit finite element schemes and penalty function method for implicit
finite element schemes. Examples of the calculation of contact interactions during
high-speed processes of the collision of elastic–plastic bodies and explosion welding
of tubular samples are presented.

Acknowledgements The study was supported by the Russian Government programs in IPMech
RAS and ICAD RAS.

[email protected]
198 N. G. Burago et al.

References

1. Wilkins, M.L.: Computer Simulation of Dynamic Phenomena. Springer, New York (1999)
2. Benson, D.J.: Computational methods in Lagrangian and Eulerian hydrocodes. Comput. Meth.
Appl. Mech. Engng. 99, 235–394 (1992)
3. Bourago, N.G.: A survey on contact algorithms. In: International Workshop on Grid Generation
and Industrial Applications, Computing Centre of the RAS, pp. 42–59. Moscow (2002)
4. Bourago, N.G., Kukudzhanov, V.N.: A review of contact algorithms. Mech. Solids 40(1), 35–71
(2005)
5. Fomin, V.M., Gulidov, A.I., Sapozhnikov, G.A., Shabalin, I.I., Babakov, V.A., Kuropatenko,
V.F., Kiselev, A.B., Trishin, YuA, Sadyrin, A.I., Kiselev, S.P., Golovlev, I.F.: High velocity
interaction of bodies. SB RAS Publ, Novosibirsk (in Russian) (1999)
6. Petrov, I.B., Kholodov, A.S.: Numerical study of some dynamic problems in the solid mechanics
by the grid-characteristic method. Comput. Math. Math. Phys. 24(3), 61–73 (2016)
7. Wriggers, P., Panagiotopoulos, P. (eds.): New developments in contact problems. Springer,
Vien GmbH (1999)
8. Barber, J.R., Ciavarella, M.: Contact mechanics. Int. J. Solids Struct. 37, 29–43 (2000)
9. Golubev, V., Khokhlov, N., Grigorievyh, D., Favorskaya, A.: Numerical simulation of destruc-
tion processes by the grid-characteristic method. Procedia Comput. Sci. 126, 1281–1288
(2018)
10. Burago, N.G., Nikitin, I.S., Nikitin, A.D., Stratula, B.A.: Algorithms for calculation damage
processes. Frattura ed Integrità Strutturale 13(49), 212–224 (2019)
11. Burago, N.G., Nikitin, I.S.: Matrix-free conjugate gradient implementation of implicit schemes.
Comput. Math. Math. Phys. 58(8), 1247–1258 (2018)
12. Burago, N.G.: Modeling of damage in elastic plastic bodies. Comput. Continuum Mechanics
1(4), 5–20 (2008)
13. Annin, B.D., Sadovskaya, O.V., Sadovskiy, V.M.: Numerical simulation of oblique collision
of plates in elastoplastic formulation. Phys. Mesomech. 3(4), 23–28 (2000)
14. Deribas, A.A.: Physics of Hardening and Explosion Welding. Nauka, Novosibirsk (in Russian)
(1980)
15. Godunov, S.K., Zabrodin, A.V., Ivanov, M.Y., Kraiko, A.N., Prokopov, G.P.: Numerical
Solution of Multidimensional Problems of Gas Dynamics. Nauka, Moscow (in Russian) (1976)

[email protected]
Chapter 15
Different Approaches for Solving Inverse
Seismic Problems in Fractured Media

Vasily I. Golubev , Maxim V. Muratov and Igor B. Petrov

Abstract The inverse seismic problem for oil and gas exploration is investi-
gated. Three different approaches based on the same fundamental grid-characteristic
method for solving the direct wave problem were successfully examined. The
dynamic behavior of the geological medium is described by the full-wave linear elas-
tic system of equations. Finite difference method on structured meshes was applied.
The migration problem in the fractured elastic medium was successfully solved and
positions of cracks were recovered. First method calculates the global minimum of
the specially constructed functional. Second method uses the modern machine learn-
ing approaches for reconstructing the geological model. Third method is based on
the traditional migration algorithm with the adjoint operator. It was shown that all
examined approaches may be used to solve inverse problems of the seismic survey
process.

15.1 Introduction

Information about the properties of geological layers and positions of fractures under
the day surface is very important when choosing the places for residential and indus-
trial buildings, exploring oil, and gas deposits [1–3]. Usually methods of acoustics
are used to determine these parameters. These methods so-called seismic survey
process are cheaper than well drilling.
One of the founders of the seismic migration theory was Claerbout [4, 5]. After the
creation of modern high-performance computing systems, significant efforts were

V. I. Golubev (B) · M. V. Muratov · I. B. Petrov

Moscow Institute of Physics and Technology (National Research University), 9, Institutskii per.,
Dolgoprudny, Moscow Region 141700, Russian Federation
e-mail: [email protected]
M. V. Muratov
e-mail: [email protected]
I. B. Petrov
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 199

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_15

[email protected]
200 V. I. Golubev et al.

made for the development of new precise methods [6, 7]. Initially, all methods were
based on the acoustic approach that doesn’t take into account the initiation of shear
waves and treats the geological medium as the fluid. It was shown, that it leads
to the lack of subvertical layer boundaries on migration images. To overcome this
drawback, full-wave elastic model was used [8]. Nowadays, the great interest for
seismic engineers is the identification of fractured zones. It is associated with the high
permeability of the medium and possible availability of hydrocarbons. Mathematical
models were developed to take into account the complex structure of media [9–11].
Diffracted waves are under comprehensive study of exploration geophysicists. A lot
of research works are devoted to the numerical simulation of seismic responses from
fractured media [12–15].
In recent years, machine learning techniques and, in particular, deep neural net-
works have shown impressive results in many areas, such as computer vision, speech
recognition, and machine translation. For example, in the field of computer vision, it
was possible to solve many problems previously unsolved, such as the classification
problem [16], recognition problem [17], and problem of image generation [18]. One
of the significant advantages of deep learning methods is that these methods can be
transferred to many other areas related to processing of large amounts of data. One
such area is the exploration seismology problems. Several works in this field have
already been carried out. In [19], the problem of fault detection in 2D was solved
using a deep convolutional neural network. As data for training the neural network,
we used synthetic data obtained by solving large direct problems. In [20], a similar
problem was solved in 3D. The great advantage that these papers draw attention to
the input data for deep learning algorithms, which do not require special processing
and, therefore, such methods can be simpler to use than standard exploration seis-
mology methods. Flexibility and relative simplicity make such methods effective for
solving practical problems. Thus, in [21] deep neural networks are used to detect
CO2 emissions, and in [22] these methods are used to detect and classify defects in
composite materials. The results show that the use of machine learning methods in
exploration seismology is important topic for research.
In this research, we investigated three different approaches for solving inverse
problems of the seismic survey. Firstly, we constructed the functional of minimization
based on synthetic responses from layered and fractured media. All parameters of
the model may be estimated by the appropriate minimization procedure. Secondly,
we used modern machine learning techniques to reconstruct the fractured structure
of the geological medium. Thirdly, we solved the classic migration problem using
adjoint operators and the grid-characteristic method on structured meshes.
Chapter is organized as follows. Direct seismic problem solution is discussed in
Sect. 15.2. Section 15.3 presents inverse problem solution. Section 15.4 gives the
conclusions.

[email protected]
15 Different Approaches for Solving Inverse Seismic Problems … 201

15.2 Direct Seismic Problem Solution

Direct problem solution is an important step in the process of the inverse problem
investigation. It allows us to obtain synthetic responses from models with known
structures. And it is necessary to have direct solver as a temporary step in the inverse
algorithm.
The defining system of equations of a linearly elastic medium can be represented
as
   
∂ Vi ∂σ ji ∂σi j  ∂ Vk ∂ Vi ∂Vj
ρ = , =λ Ii j + μ + , (15.1)
∂t ∂ x j ∂t k
∂ xk ∂x j ∂ xi

where Vi is the velocity component, σi j is the stress tensor, ρ is the density of the
medium, λ and μ are the Lame coefficients, Ii j is the component of the unit tensor.
 Consider the two-dimensional
 case. Let us introduce the vector of variables u =
Vx , Vy , σx x , σ yy , σx y . Then the system Eq. 15.1 is reduced to the form:

∂ u  ∂ u
+ Ai = 0. (15.2)
∂t i=1,2
∂ξ i

The numerical solution is found using the grid-characteristic method [23, 24].
Approximation is carried out on a structural rectangular mesh. Values at each point
are found using values at mesh reference points u ri jkl and weights of these points
pi jkl (
r ):

u(
r) = pi jkl ( u ri jkl .
r ) (15.3)
i, j,k,l

The boundary condition can be written in the general form as


Du(ξ1 , ξ2 , t + τ ) = d, (15.4)

where D is a certain matrix of size 9 × 3 for the three-dimensional case (5 × 2 for

two-dimensional case), d is the vector, u(ξ1 , ξ2 , t + τ ) is the value of the required
velocity values and components of the stress tensor at the boundary point at the next
time step.
At the top boundary of the computational domain, the condition of a free boundary
has a view:

T n = 0. (15.5)

To specify the fracture, an infinitely thin fracture model with the condition of fluid-
filled fracture was used [25, 26]. Such a fracture is defined as a contact boundary

[email protected]
202 V. I. Golubev et al.

with the condition of free sliding:

va · n = vb · n, fna = − fnb , fτa = fτb = 0. (15.6)

Such a contact boundary completely passes longitudinal oscillations without

reflection and fully reflects transverse waves.
The underlying computer code involves schemes of second to fourth orders of
accuracy. In this study, we used the fourth-order accurate scheme (ζ = t/ h, h is
the spatial coordinate step):

vmn+1 = vmn − ζ ( 1 − ζ ( 2 − ζ ( 3 − ζ 4 ))),

1 = (−2vm+2 + 16vm+1 − 16vm−1 + 2vm−2 )/24,
n n n n

2 = (−vm+2
n
+ 16vm+1
n
− 30vmn + 16vm−1
n
− vm−2
n
)/24,
3 = (2vm+2 − 4vm+1 + 4vm − 2vm−2 )/24,
n n n n

4 = (vm+2
n
− 4vm+1
n
+ 6vmn − 4vm−1
n
+ vm−2
n
)/24. (15.7)

Additionally, we used the grid-characteristic monotonicity criterion. For positive

components of diagonal matrix of eigenvalues, it has the form of Eq. 15.8.

min{vmn , vm−1
n
} ≤ vmn+1 ≤ max{vmn , vm−1
n
} (15.8)

For negative components it is symmetric. In the simplest case when this criterion
is violated, the solution is corrected as follows:
⎧
⎨ max{vmn , vm−1
n
}, vmn+1 > max{vmn , vm−1
n
},
vmn+1 = min{vm , vm−1 }, vm < min{vm , vm−1 },
n n n+1 n n (15.9)
⎩ n+1
vm , min{vmn , vm−1
n
} ≤ vmn+1 ≤ max{vmn , vm−1
n
}.

This limiter preserves the fourth-order of the scheme in domains, where the solu-
tion is fairly smooth (the characteristic criterion is satisfied). In the case of high
solution gradients, the order of the scheme is reduced to the third order.

15.3 Inverse Problem Solution

Initially, the inverse problem of determination of number and thicknesses of layers

was considered. We assumed that there are a set of layers (from one to three) with
known physical properties but unknown thicknesses. Then the vector of parameters
includes the number of layers N = 1, 2, 3 and the thickness of each layer Hmin ≤
h k ≤ Hmax , k = 1, 2, 3. Geological parameters of the first layer are given as follows
Cp = 2000 m/s, Cs = 1400 m/s, ρ = 2000 kg/m3 , the parameters of the second layer
are given as follows Cp = 2400 m/s, Cs = 1600 m/s, ρ = 2500 kg/m3 , and the

[email protected]
15 Different Approaches for Solving Inverse Seismic Problems … 203

Fig. 15.1 The dependence

of I(z) on two-layer heights.
One global and many local
extremes

parameters of the third layer are given as follows Cp = 2600 m/s, Cs = 1700 m/s, ρ
= 2800 kg/m3 . Berlage source was used with the main frequency of 30 Hz.
Peculiarity of the problem is the fact that information can be obtained only from
acoustic measurements. Let us consider 1D case. On the day surface, one geophone
is situated which measures vertical velocity Ṽy (xi , t j ) of the ground resulting from
wave reflections from layers boundaries. And we can try to found the value of z
when the numerical response Vy (z , xi , t j ) will be as close as possible to Ṽy (xi , t j ).
Seismic registrations were carried out with spatial step of 10 m.
The mathematical problem can be formulated as the optimization problem of least
squares:

minI (z ), z ∈ {1, 2, 3} × [Hmin , Hmax ]{1,2,3} , (15.10)

 2
I (z ) = Vy (z , xi , t j ) − Ṽy (xi , t j ) . (15.11)
i j

Equation 15.11 has no analytical form, and each its value can be obtained only as
a result of rather time-consuming numerical calculations. For solving the problem
(Eqs. 15.10 and 15.11), we use the direct methods of sorting. The results are presented
in Fig. 15.1.
We have found that functional (Eq. 15.11) has one strong global minimum and
some small local minimums. This functional is more sensitive to the thickness of the
first layer than to the second layer. The main reason for this fact is that the amplitude
of the first response exceeds the amplitude of other responses.
The same approach was applied to the seismic survey problem of fractured media.
We used the homogeneous half-space model with the free day surface. It contained
a single linear fluid-filled crack. Then the vector of parameters includes the crack
depth h, h 1 ≤ h ≤ h 2 , and the orientation angle α, α1 ≤ α ≤ α2 . The illustration is
represented in Fig. 15.2.
The homogeneous medium had density 2500 kg/m3 , P-wave velocity 3000 m/s,
S-wave velocity 1700 m/s. The length of the crack was 100 m. The initial perturbation
was P-wave with the length of 100 m. Vertical component of velocity was registered
with the spatial distance of 10 m at the day surface.

[email protected]
204 V. I. Golubev et al.

Fig. 15.2 The experiment

for the crack identification

Fig. 15.3 The cross section

of minimization functional
for fixed crack angle

Fig. 15.4 The minimization

functional as a function of
crack angle (alpha) and crack
depth (h)

The mathematical problem can be formulated as the optimization problem of least

squares (Eq. 15.11) on the special domain:

minI (z ), z ∈ D = [h 1 ; h 2 ] × [α1 ; α2 ]. (15.12)

[email protected]
15 Different Approaches for Solving Inverse Seismic Problems … 205

In Figs. 15.3 and 15.4, the cross section as 2D plot and 3D plot are represented.
As in the previous example, a single global minimum exists.
Another approach based on machine learning methods was applied to the problem
of cracks identification. The process of solving the problem of recognizing the spatial
position of a fracture in the elastic medium consists of two stages: training of the
neural network and recognizing of a control sample of seismic data. To create a
training sample, the direct problems are solved with different fracture parameters.
A two-dimensional problem is considered, in which it is proposed to find the spa-
tial position and angle of inclination of a single fracture of a fixed size (100 m) using
seismic data. The fracture is in a homogeneous elastic medium with the following
elastic characteristics: Cp = 4500 m/s, Cs = 2500 m/s, ρ = 2500 kg/m3 . The size of
the computational domain is 2 km × 2 km (Fig. 15.5). The position of the fracture
varies in the range of 1000 m vertically and horizontally. The angle of inclination is
in the range of ±15° (subvertical fracture). In the middle of the border of the study
area, a sinusoidal elastic pulse consisting of 5 periods (a wavelength of 100 m) is
excited. The values of the vertical component of the velocity of the reflected waves
are recorded on seismic receivers uniformly located on the excitation surface of the
wave pulse (65 receivers in total).
Keras deep learning library based on Tensorflow library and CUDA parallel com-
puting architecture was used. Keras library was chosen because of its simplicity of use
and possibilities sufficient to solve the problem. In this chapter, a two-dimensional
inverse problem was solved, but the proposed method can be extended to the case of
3D. To solve the problem, a neural network consisting of 3 convolutional layers and
two fully connected layers was proposed. The training set consisted of pairs (X, y),

Fig. 15.5 The scheme of

fracture placement

[email protected]
206 V. I. Golubev et al.

where X is a seismogram (a matrix of real numbers 65 × 65 in size) and y is a set of

parameters defining the position of the fracture. In the case under consideration, y
was given by 4 real numbers—the coordinates of the ends of the fractures (the height
of the fracture is remained by constant). The network has the folloing architecture:
the first convolutional layer (63, 63, 64), the second convolutional layer (29, 29, 128),
the third convolutional layer (12, 12, 256), the first fully connected layer (9216, 256),
and the second fully connected layer (256, 4). The dimensions of all filters in the
convolutional layers are 3 × 3, the activation function is ReLU. The total number of
network parameters: 2.7 × 106. As the optimizer of the neural network, Adam was
chosen with a learning rate of 0.001.
The recognition of a control sample is the process of minimizing the functional:


 r eal

pr ed 
J= yi − yi  . (15.13)
L2
i

With each new era in learning, the value of the functional decreases and tends to
a certain value (Fig. 15.6). Therefore, the method can be used to solve this class of
problems.
The results of recognition of a single fracture are shown in Fig. 15.7, and Fig. 15.8
shows the seismogram of the wave response obtained by seismic receivers on the
upper surface of the studied region. Figure 15.8 shows the location and orientation
of the real position of the crack and predicted. It is seen quite a good match.
The last problem investigated in this research was the classic 2D migration prob-
lem. The goal was to compare two different approaches taken into account the depen-
dency of P-wave and S-wave velocities on the depth. Firstly, a separate grid was
constructed for each geological layer. And each grid had its own elastic properties
(Lame parameters). The special glue condition was consistently applied to ensure the
continuity of the stress tensor and the velocity vector over the contact boundary. In
the second approach, we blurred parameters of all layers along the vertical axis and
used the numerical schemes with the matrix elements with the spatial dependency.

Fig. 15.6 The graph of functional J dependence on period of study

[email protected]
15 Different Approaches for Solving Inverse Seismic Problems … 207

Fig. 15.7 Control seismograms for recognition: a vertical cracks, b subvertical left rotated crack,
c subvertical right rotated crack

Fig. 15.8 Spatial placement of fractures, real (dotted line), and predicted (solid line): a vertical
cracks, b subvertical left rotated crack, c subvertical right rotated crack

With the help of curvilinear structured meshes, we put inside the model sub-
vertical fluid-filled cracks. Eliminating the drastic increase of the problem com-
putational complexity, we decided to use specific boundary conditions rather than
solving directly the acoustic equation inside the crack volume. This approach
was successfully verified previously on direct seismic problems for different crack
orientations.
To obtain the migration image of the layered fractured medium the method pro-
posed in [8] was used. We analyzed the kernel in a form of Eq. 15.14, where v is
the velocity vector obtained as a solution of the direct problem and v† is the velocity
vector obtained as a solution of the adjoint problem.

K imp (x) = −ρ(x) v† (x, −t)v(x, t)dt (15.14)

Both direct and adjoint problems were solved numerically with the grid-
characteristic method.

[email protected]
208 V. I. Golubev et al.

Table 15.1 Elastic properties

Number of Width, m P-wave S-wave
of geological layers
layer velocity, m/s velocity, m/s
1 500 3500 1750
2 1000 4500 2250
3 300 5000 2500
4 100 4000 2000
5 300 5500 2750
6 800 5500 2750

In this research, we extended our previous numerical experiments with the migra-
tion process of seismic data from layered fractured media presented by [27–29]. We
used the same 2D layered geological reference model consisted of six elastic hor-
izontal layers with different geometrical and physical properties. The density was
constant along with the depth and was equal to 2500 kg/m3 . We enumerated all lay-
ers from 1 (top) to 6 (bottom). It is illustrated at Table 15.1 and Fig. 15.9. Identical
elastic parameters for two last layers were used to illustrate the correctness of our
special glue contact conditions. At the depth of 1650 m, two fluid-filled subvertical
cracks were placed. The length of both cracks was 100 m, and the angle was 10° from
vertical. The curvilinear structured grid with the spatial step approximately 5 m and
with ~106 nodes was used. Ricker point sources with central frequency 30 Hz were
placed at each 10 m on the day surface. Three-component receivers were placed at
the same places. Totally 3.3 s of physical time were simulated with the time step
0.45 ms.
At the initial step, the direct problem was numerically solved for the reference
layered fractured medium. Obtained synthetic seismograms were used in the fur-
ther processing chain. According to the described above procedure, the migration
image was obtained with the layered background model in [27] (Fig. 15.10). Cracks
positions are clearly seen on it, and all layer boundaries were filtered naturally.

Fig. 15.9 Two-dimensional model of the geological layered medium extended with two subvertical
fluid-filled cracks (inside the blue layer). Contact boundaries are depicted with black horizontal lines

[email protected]
15 Different Approaches for Solving Inverse Seismic Problems … 209

Fig. 15.10 Migration image, obtained with the layered background model [27]. Cracks positions
can be identified

Unfortunately, the precise positions of layer boundaries are not known in most
cases. After the seismic inversion procedure, the background model is the 3D massive
with the dependency of elastic properties on coordinates. To estimate the influence of
the background model inaccuracy on the migration image, a new numerical exper-
iment was carried out. We blurred our layered model with Gauss filter along the
vertical axis with the 20 m window and used it in the migration pipeline (Fig. 15.11).
Due to the discrepancy of background models, the obtained migration image contains
not only correct crack positions but layer boundaries too. And their amplitudes are
significantly higher than the signal from cracks. The second problem is the presence
of two false boundaries with the large enough signal on the final migration image.

Fig. 15.11 Migration image obtained with the blurred layered background model. Gauss impulse
with the 20 m window was applied. True positions of layer boundaries are depicted with white lines

[email protected]
210 V. I. Golubev et al.

15.4 Conclusions

The migration problem in the fractured elastic medium was successfully solved and
positions of cracks were recovered. First method calculated the global minimum of
the specially constructed functional. Second method used modern machine learning
approaches for reconstructing the geological model. Third method was based on
the traditional migration algorithm with the adjoint operator. It was shown that all
examined approaches may be used to solve inverse problems of the seismic survey
process.
This research is the continuation of the work dedicated for proving the applica-
bility of the grid-characteristic numerical method to migration problems of fractured
media. The full-wave simulation with the linear elastic approach was carried out. The
blurred background model constructed from the precise layered model was used. The
signal from both subvertical fluid-filled cracks was successfully registered. Unfortu-
nately, its amplitude on the migration image is significantly lower than the response
from true and false obtained boundaries. The further research may be concentrated
on the development of the filtration procedure for eliminating the spatially correlated
signal from true layer boundaries and false boundaries too.

Acknowledgements This research has been carried out using computing resources of the federal
collective usage center Complex for Simulation and Data Processing for Mega-science Facilities
at NRC “Kurchatov Institute”, http://ckp.nrcki.ru/. This research was carried out with the financial
support of the Russian Science Foundation, project no. 19-11-00023.

References

1. Binnani, N., Khare, R.K., Golubev, V.I., Petrov, I.B.: Probabilistic seismic hazard analysis
of punasa dam site in India. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov,
S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development of
Grid-Characteristic Method, SIST, vol. 133, pp. 105–119. Springer (2019)
2. Stognii, P.V., Khokhlov, N.I.: 2D seismic prospecting of gas pockets. In: Petrov, I.B.,
Favorskaya, A.V., Favorskaya, M.N., Simakov, S.S., Jain, L.C. (eds.) Proceedings of the
Conference on 50 Years of the Development of Grid-Characteristic Method, SIST, vol. 133,
pp. 156–166. Springer (2019)
3. Golubev, V.I., Golubeva, Y.A.: Full-wave simulation of the earthquake initiation process. In:
CEUR Workshop Proceedings, vol. 2267, pp. 346–350 (2018)
4. Claerbout, J.F.: Coarse grid calculations of waves in inhomogeneous media with application
to delineation of complicated seismic structure. Geophysics 36(3), 407–418 (1970)
5. Claerbout, J.F., Doherty, S.M.: Downward continuation of moveout-corrected seismograms.
Geophysics 37(5), 741–768 (1972)
6. Etgen, J., Gray, S., Zhang, Y.: An overview of depth imaging in exploration geophysics.
Geophysics 74, WCA5–WCA17 (2009)
7. Jiao, K., Huang, W., Vigh, D., Kapoor, J., Coates, R., Starr, W.E., Cheng, X.: Elastic migration
for improving salt and subsalt imaging and inversion. In: SEG Las Vegas Annual Meeting,
pp. 1–5 (2012)

[email protected]
15 Different Approaches for Solving Inverse Seismic Problems … 211

8. Luo, Y., Tromp, J., Denel, B., Calandra, H.: 3D coupled acoustic-elastic migration with topog-
raphy and bathymetry based on spectral-element and adjoint methods. Geophysics 78(4),
S193–S202 (2013)
9. Burago, N.G., Nikitin, I.S., Yakushev, V.L.: Hybrid numerical method with adaptive overlap-
ping meshes for solving nonstationary Problems in Continuum Mechanics. Comput. Math.
Math. Phys. 56(6), 1065–1074 (2016)
10. Nikitin, I.S., Burago, N.G., Nikitin, A.D.: Explicit-Implicit schemes for solving the problems of
the dynamics of isotropic and anisotropic elastoviscoplastic media. In: IOP Conference Series:
Journal of Physics: Conference Series, vol. 1158, pp. 032039.1–032039.8 (2019)
11. Burago, N.G., Nikitin, I.S.: Algorithms of through calculation for damage processes. Comput.
Res. Model. 10(5), 645–666 (2018)
12. Fang, X., Zheng, Y., Fehler, M.C.: Fracture clustering effect on amplitude variation with offset
and azimuth analyses. Geophysics 82(1), N13–N25 (2017)
13. Muratov, M.V., Petrov, I.B.: Application of fractures mathematical models in exploration seis-
mology problems modeling. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov,
S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development of
Grid-Characteristic Method, SIST, vol. 133, pp. 120–131. Springer (2019)
14. Zheng, Y., Fang, X., Fehler, M.C., Burns, D.R.: Seismic characterization of fractured reservoirs
using 3D double beams. In: SEG Technical Program Expanded Abstracts, pp. 1–6 (2012)
15. Hu, H., Zhengm, Y.: 3D seismic characterization of fractures in a dipping layer using the
double-beam method. Geophysics 83, 123–134 (2018)
16. Krizhevsky, A., Sustskever, I., Hinton, G.E.: ImageNet classification with deep convolutional
neural networks. In: 25th International Conference on Neural Information Processing Systems,
vol. 1, pp. 1097–1105 (2012)
17. Szegedy, C., Toshev, A., Erhan, D.: Deep neural networks for object detection. In: 26th
International Conference on Neural Information Processing Systems, vol. 2, pp. 2553–2561
(2013)
18. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.,
Bengio, Y.: Generative adversarial nets. In: 27th International Conference Neural Information
Processing Systems, vol. 2, pp. 2672–2680 (2014)
19. Zhang, C., Frogner, C., Araya-Polo, M., Hohl, D.: Machine-learning based automated fault
detection in seismic traces. In: EAGE Conference and Exhibition, pp. 1–5 (2014)
20. Araya-Polo, M., Dahlke, T., Frogner, C., Zhang, C., Poggio, T., Hohl, D.: Automated fault
detection without seismic processing. Lead. Eadge 36(3), 194–280 (2017)
21. Wu, Y., Lin, Y., Zhou, Z., Delorey, A.: Seismic-net: a deep densely connected neural network
to detect seismic events, pp. 1–8. arXiv:1802.02241 (2018)
22. Menga, M., Chua, Y.J., Woutersonb, E., Ong, C.P.K.: Ultrasonic signal classification and imag-
ing system for composite materials via deep convolutional neural networks. Neurocomputing
257, 128–135 (2017)
23. Ivanov, A.M., Khokhlov, N.I.: Efficient inter-process communication in parallel implemen-
tation of grid-characteristic method. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N.,
Simakov, S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development
of Grid-Characteristic Method, SIST, vol. 133, pp. 91–102. Springer (2019)
24. Khokhlov, N.I., Golubev, V.I.: On the class of compact grid-characteristic schemes. In: Petrov,
I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov, S.S., Jain, L.C. (eds.) Proceedings of the
Conference on 50 Years of the Development of Grid-Characteristic Method, SIST, vol. 133,
pp. 64–77. Springer (2019)
25. Golubev, V.I.: The usage of grid-characteristic method in seismic migration problems. In:
Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov, S.S., Jain, L.C. (eds.) Proceedings
of the Conference on 50 Years of the Development of Grid-Characteristic Method, SIST, vol.
133, pp. 143–155. Springer (2019)
26. Golubev, V., Khokhlov, N., Grigorievyh, D., Favorskaya, A.: Numerical simulation of destruc-
tion processes by the grid-characteristic method. Procedia Comput. Sci. 126, 1281–1288
(2018)

[email protected]
212 V. I. Golubev et al.

27. Golubev, V.I., Voinov, O.Y., Petrov, I.B.: Migration of seismic data for multi-layered fractured
geological media using elastic approach. In: 19th Science and Applied Research Conference Oil
and Gas Geological Exploration and Development, pp. 43756.1–43756.5 (in Russian) (2017)
28. Golubev, V.I., Voinov, O.Y., Zhuravlev, Y.I.: On seismic imaging of fractured geological media.
Dokl. Math. 96(2), 514–516 (2017)
29. Golubev, V.I., Voinov, O.Y., Petrov, I.B.: Seismic imaging of fractured elastic media on the
basis of the grid-characteristic method. Comput. Math. Math. Phys. 58(8), 1309–1315 (2018)

[email protected]
Chapter 16
Elastic Wave Scattering on a Gas-Filled
Fracture Perpendicular to Plane P-Wave
Front

Alena V. Favorskaya

Abstract This chapter discusses the features of the scattering of plane P-waves on
gas-filled fractures located along with the motion of the incident wave front. This
problem has practical significance in the areas of nondestructive testing and seismic
exploration, primarily in the area of railway nondestructive testing. This type of
fractures falls into the blind zone. In this chapter, such types of reflected waves are
considered that can be registered, and, thus, the blind zone of the recording equipment
can be avoided. Analytical expressions for reflected waves amplitudes and scattering
angles are obtained. To obtain these expressions, the Wave Logica approach was
used. This approach combines the advantages of the analytical study of wave fields
and study of the computational solution of the elastic wave equation. Comparison
of the analytical expressions with visualized wave fields (wave patterns) at the stage
of derivation of these analytical expressions greatly facilitates the study that allows
to avoid mistakes and also demonstrates the accuracy of the applied computational
method. In this chapter, the grid-characteristic numerical method was used for the
numerical solution of the elastic wave equation.

16.1 Introduction

Exact analytical solutions of the wave equation, which are commonly called waves of
various types, such as longitudinal P-waves, shear S-waves, Rayleigh, Love, Stone-
ley, Krauklis, exchangeable, etc., only approximately describe the wave processes
observed in real substantially heterogeneous media, as they occur interaction of

A. V. Favorskaya (B)
Moscow Institute of Physics and Technology (National Research University), 9, Institutsky per.,
Dolgoprudny, Moscow 141700, Russian Federation
e-mail: [email protected]
National Research Centre “Kurchatov Institute”, 1, Akademika Kurchatova pl., Moscow 123182,
Russian Federation
Scientific Research Institute for System Analysis of the RAS, 36(1), Nahimovskij pr., Moscow
117218, Russian Federation

© Springer Nature Singapore Pte Ltd. 2020 213

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_16

[email protected]
214 A. V. Favorskaya

waves (multiple reflections, interference, etc.). Obtaining an appropriate analytical

solution, for example, for Krauklis waves, can be found in [1]. Also, a rigorous
analytical approach aimed at studying solutions of hyperbolic systems of equations,
most fully presented in [2], is known. Another approach is a full-wave numerical
modeling of wave effects [3–7]. In [8–11], a combined approach for studying spatial
dynamic wave fields called Wave Logica was proposed.
Note that the approach used in the analysis of spatial dynamic wave fields calcu-
lated by full-wave computer simulation describes the dynamics of wave propagation
and structure of wave fronts in more detail than the classical approach, which implies
the explicit selection of particular analytical solutions of the wave equation and
more accurately than the geometric approximation and ray-tracing method [12, 13].
Accordingly, the proposed approach allows to identify the patterns and relationships
more accurately that can later be used in applications.
For numerical simulation of ultrasonic nondestructive railway testing, the finite
element method [7], the mass-spring-lattice model [14], the finite-difference method
[15], and the quasi-analytical finite element method (semi-analytical finite element
method) [16] are used.
The infinitely thin gas-filled and fluid-filled fractures are particular limiting cases
of Schoenberg family of models [17]. Experimental confirmation of Schoenberg
theory was given in [18]. Finite-difference methods have gained wide popularity due
to ease of implementation [19]. In this work, the so-called homogeneous approach
to finite-difference modeling of dynamic processes in an elastic body was used. In
this approach, the boundary conditions for inhomogeneities are explicitly specified,
which allows one to simulate an open fracture with a discontinuous (as it passes
through its plane) displacement and a continuous stress. The effect of fracture size
and density on seismic wave propagation was investigated in [3], while analyzing
sections of wave fields were not significantly represented, and a number of qualitative
but non-detailed results were made. Also, in the work there is no information about the
analysis of the wave front in the dynamics, and a comparison of the sections was made
only at a fixed point in time. The effect of the ratio of the wavelength to the length
of the fracture on the reflection amplitude was numerically demonstrated in [20].
Three-dimensional calculations in a medium containing two-dimensional fractures
with explicit formulation of the boundary conditions were performed [21]. In [22],
the dependence of the amplitude of reflection and diffraction on the characteristics
of fractures was investigated.
The grid-characteristic method was proposed in [23]. Later a family of numerical
methods was developed [24] to simulate the process of ultrasound nondestructive
testing of railway tracks [25, 26], composite materials [27], modeling the process of
seismic exploration of fractured zones [6, 28], etc.
This chapter is structured as follows. Section 16.2 presents the problem statement
and mechanical mathematical model of the problem. Section 16.3 shows the elastic
wave patterns, which are visualized elastic wave fields. Section 16.4 deals with the
amplitudes of velocity and Cauchy stress tensor components for the wave types under
consideration. The derivation of the formulae of scattering amplitudes and angles is
discussed in Sect. 16.5. Section 16.6 concludes the chapter.

[email protected]
16 Elastic Wave Scattering on a Gas-Filled Fracture … 215

16.2 Problem Statement

In this section, the problem statement is discussed. The calculations were carried out
in mechanical mathematical statements, the scheme of which is shown in Fig. 16.1.
The gas-filled fracture is modeled as two free boundaries along OY axis with
conditions according to Eqs. 16.1–16.2.

σx x = 0 (16.1)

σx y = 0 (16.2)

In Eqs. 16.1–16.2, σx x and σx y are the Cauchy stress tensor components.

Equations 16.1–16.2 are similar with the following formula:

f ≡ σ · n = 0. (16.3)

In Eq. 16.3, f is the traction, σ is the Cauchy stress tensor, n is the normal being
out to the borders of the fracture.
Grid-characteristic computational method [24] was used to solve elastic wave
equation given by Eqs. 16.4–16.5.

∂
ρ v(r, t) = (∇ · σ(r, t))T (16.4)
∂t
∂    
σ(r, t) = ρcP2 − 2ρcS2 (∇ · v(r, t))I + ρcS2 ∇ ⊗ v(r, t) + (∇ ⊗ v(r, t))T
∂t
(16.5)

In Eqs. 16.4–16.5, v(r, t) is the velocity vector-function, σ(r, t) is the symmetric

Cauchy stress tensor-function, ρ is the material density, cP is the P-wave speed, cS is
the S-wave speed, t is the time, r is the radius vector.
The size of the integration area and the time were chosen in such a way that it
corresponded to the infinite space in Fig. 16.1, and the responses from non-reflecting
boundary conditions would not have time to reach the area near the fracture, around
which wave processes were studied.

Fig. 16.1 Scheme of

mechanical mathematical
models, incident P-wave, and
gas-filled fracture

[email protected]
216 A. V. Favorskaya

Table 16.1 Parameters of

Model P-wave speed S-wave speed Density
elastic media in mechanical
(m/s) (m/s) (kg/m3 )
mathematical models
“cs_norm” 6,250 3,200 7,800
“cs_mini” 6,250 2,000 7,800

Different elastic parameters of the media for two models under consideration are
shown in Table 16.1.
The parameters of the computational grid in time and coordinates were chosen in
such a way that the visualized wave fields (wave patterns) would coincide with each
other upon further refinement of the computational grid step.

16.3 Elastic Wave Patterns

In this section, the elastic wave patterns of various components and different time
moments are presented. Figures 16.2 and 16.3 show the results of computational
experiments for “cs_mini” and “cs_norm” models, respectively.
In Figs. 16.2, 16.3, the maximum value of the velocity modulus is in red color,
middle value is in white color, and zero value is in blue color, while the maximum
positive value of velocity and Cauchy stress tensor components is in purple color,
maximum negative is in green color, and zero value is in gray color. Note that as max-
imum values it is not meant the maximum values of the functions, but the maximum
values on the scale, in accordance with the principle set forth in the work [29].
Figure 16.4 shows the wave traveling and reflection scheme. This scheme was
obtained by analyzing the dynamics of the snapshots of wave fields, the so-called
analysis of wave patterns.
In Fig. 16.4 and further in the text, L is the fracture length, D is the scattered
S-waves’ fronts length, hP is the incident P-wave wavelet length, hS is the scattered
S-waves wavelet length.
Thus, the scattered S-waves were identified, which are of primary interest. There
are also the cylindrical P- and S-waves scattered from the upper and lower edges of
the fracture, but they have smaller amplitude.

16.4 Amplitudes of Private Solutions Components

In this section, a derivation of the formulae for the amplitudes of velocity and Cauchy
stress tensor components for different private solutions of elastic wave equations
figured in the problem are considered. In accordance with Wave Logica approach
[8–11], the obtained amplitudes of particular solutions were compared with elastic
wave patterns.

[email protected]
16 Elastic Wave Scattering on a Gas-Filled Fracture … 217

Fig. 16.2 Wave patterns for “cs_mini” model: a 21st snapshot, velocity module, b 21st snapshot,
horizontal component of velocity, c 21st snapshot, vertical component of velocity, d 35th snap-
shot, velocity module, e 35th snapshot, horizontal component of velocity, f 35th snapshot, vertical
component of velocity, g 21st snapshot, main horizontal component of Cauchy stress tensor XX,
h 21st snapshot, main vertical component of Cauchy stress tensor YY, i 21st snapshot, tangential
component of Cauchy stress tensor XY, j 35th snapshot, main horizontal component of Cauchy
stress tensor XX, k 35th snapshot, main vertical component of Cauchy stress tensor YY, l 35th
snapshot, tangential component of Cauchy stress tensor XY

In accordance with Riemann invariants, for the incident P-wave (Fig. 16.4)
traveling oppositely axis OY there are Eqs. 16.6–16.10.

vxP = 0 (16.6)

vPy = VP (16.7)

[email protected]
218 A. V. Favorskaya

Fig. 16.3 Wave patterns for “cs_norm” model: a 21st snapshot, velocity module, b 21st snapshot,
horizontal component of velocity, c 21st snapshot, vertical component of velocity, d 35th snap-
shot, velocity module, e 35th snapshot, horizontal component of velocity, f 35th snapshot, vertical
component of velocity, g 21st snapshot, main horizontal component of Cauchy stress tensor XX,
h 21st snapshot, main vertical component of Cauchy stress tensor YY, i 21st snapshot, tangential
component of Cauchy stress tensor XY, j 35th snapshot, main horizontal component of Cauchy
stress tensor XX, k 35th snapshot, main vertical component of Cauchy stress tensor YY, l 35th
snapshot, tangential component of Cauchy stress tensor XY

Fig. 16.4 Wave traveling

and reflection scheme,
maximum value of the
velocity modulus is in
turquoise, middle value is in
black, and zero value is in
yellow

[email protected]
16 Elastic Wave Scattering on a Gas-Filled Fracture … 219

  2 
cS
σPx x = 1−2 ρcP VP (16.8)
cP

σPyy = ρcP VP (16.9)

σPx y = 0 (16.10)

In Eqs. 16.6–16.10 and further in the text, vx , v y are the velocity components, σx x ,
σ yy , and σx y are the Cauchy stress tensor components, ρ is the material density, cP is
the P-wave speed, cS is the S-wave speed, V is the proportional to wave amplitude.
Note that these expressions are consistent with the signs of wave patterns.
Note that for the models under consideration, a rotation matrix is written as follows
for the “right” angle α:

sin α cos α
. (16.11)
− cos α sin α

Then the velocity vector and Cauchy stress tensor will be transformed by
Eqs. 16.12–16.16.

vx = sin α · vxW + cos α · vW

y (16.12)

v y = − cos α · vxW + sin α · vW

y (16.13)

σx x = sin2 α · σW
x x + cos α · σ yy + 2 sin α cos α · σx y
2 W W
(16.14)

σ yy = cos2 α · σW
x x + sin α · σ yy − 2 sin α cos α · σx y
2 W W
(16.15)

 
σx y = − sin α cos α · σW
x x − σ yy − cos 2α · σx y
W W
(16.16)

In Eqs. 16.12–16.16, the index W corresponds to the components of the velocity

and Cauchy stress tensor in the coordinate system, in which the wave moves along
the OX axis.
In accordance with Riemann invariants, for S-wave traveling along axis OX there
are Eqs. 16.17–16.19.

vSy = VS (16.17)

σSx y = −ρcS vSy = −ρcS VS (16.18)

vxS = σSx x = σSyy = 0 (16.19)

[email protected]
220 A. V. Favorskaya

Substituting Eqs. 16.17–16.19 based on Riemann invariants into Eqs. 16.12–

16.16, one can obtain the formulae for the shear SR-wave (Fig. 16.4) scattered to the
right side of the fracture expressed by Eqs. 16.20–16.24.

vxSR = cos α · VSR (16.20)

y = sin α · VSR
vSR (16.21)

σSR
x x = −2 sin α cos α · ρcS VSR (16.22)

σSR
yy = 2 sin α cos α · ρcS VSR (16.23)

σSR
x y = cos 2α · ρcS VSR (16.24)

After substitution, a comparison is made with the wave patterns in accordance

with the signs of the components of the velocity and Cauchy stress tensor.
Substituting Eqs. 16.17–16.19 based on Riemann invariants into Eqs. 16.12–
16.16, one can obtain the formulae for the shear SL-wave (Fig. 16.4) scattered to the
left side of the fracture expressed by Eqs. 16.24–16.29.

vxSL = cos α · VSL (16.25)

y = − sin α · VSL
vSL (16.26)

σSL
x x = 2 sin α cos α · ρcS VSL (16.27)

σSL
yy = −2 sin α cos α · ρcS VSL (16.28)

σSL
x y = cos 2α · ρcS VSL (16.29)

After substitution, a comparison is made with the wave patterns in accordance

with the signs of the components of the velocity and Cauchy stress tensor.

16.5 Scattering Amplitudes and Angles

In this section, a derivation of formulae for scattering amplitudes and angles is dis-
cussed. When P-wave moves along the fracture border, the P- and S-waves from
the source moving along the fracture appear. This imaginary source moves with the
incident P-wave speed and compensate the XX-component of Cauchy stress tensor
provided by Eqs. 16.30–16.31.

[email protected]
16 Elastic Wave Scattering on a Gas-Filled Fracture … 221
     
cS 2 y
fL (y, t) = − 1 − 2 ρcP VP G t − n (16.30)
cP cP
  2   
cS y
fR (y, t) = 1 − 2 ρcP VP G t − n (16.31)
cP cP

In Eqs. 16.30–16.31, n is the normal outer to the left border of the fracture,
fL (y, t) and fR (y, t) are the volume density of force for these two sources moving
alone left and right borders of the fracture, respectively, G(t) is the shape in the
incident P-wave, y is Y coordinate.
That is, the same as in the case of half-space, P-wave reflected from the boundary,
and the head H-wave are formed. However, since here the free boundary is localized
in the problem, one can obtain the expressions, relating the size of the fracture to
the size of the resulting superposition of cylindrical shear waves propagating from a
point source, provided by Eqs. 16.32–16.34.
 
cS
α = arcsin (16.32)
cP
cS
hS = hP (16.33)
cP

D = L · sin α (16.34)

In Eqs. 16.32–16.34, α the angle between S-waves fronts and fracture, L is the
fracture length, D is the S-waves fronts extension, hP and hS are the P-wave and
S-wave fronts’ width, respectively. These values are marked in Fig. 16.4.
Let us single out plane S-waves from this solution. For the right S-wave, one can
substitute Eqs. 16.8 and 16.22 into Eq. 16.1 and obtain the following formula:
  2 
cS
−2 sin α cos α · ρcS VSR + 1 − 2 ρcP VP = 0. (16.35)
cP

From Eq. 16.35, one can find the amplitude using Eq. 16.32 for the angle α:
 
2
1−2 cS
cP
cP
cS cos 2α
VSR = VP = − VP . (16.36)
2 sin α cos α 2 sin α
One can also find a nonzero component of shear stresses along the border, which
appeared due to the fact that this wave is moving along the free boundary:

cos2 2α
xy = −
σSR · ρcS VP . (16.37)
2 sin α

[email protected]
222 A. V. Favorskaya

This shear stress will already be compensated by a moving point source of shear
stresses since for waves from this point source at the contact boundary; both the shear
and main components of the stress tensor will be equal to zero.
Now one can find a similar equation for the left S-wave. Substituting Eqs. 16.8
and 16.27 into Eq. 16.1, one can obtain:
  2 
cS
2 sin α cos α · ρcS VSL + 1 − 2 ρcP VP = 0. (16.38)
cP

From Eq. 16.38, one can find the amplitude using Eq. 16.32 for the angle α:
 
2
1−2 cS
cP
cP
cS cos 2α
VSL = − VP = VP . (16.39)
2 sin α cos α 2 sin α
One can also find a nonzero component of shear stresses along the border, which
appeared due to the fact that this wave is moving along the free boundary:

cos2 2α
σSL
xy = · ρcS VP . (16.40)
2 sin α
Note that since Lame parameter λ is positive, the angle α lies in the range from 45°
to 90°, which means that the amplitude of the right S-wave has the same sign as the
amplitude of the incident P-wave, and the amplitude of the left S-wave is opposite.
This corresponds to the calculated wave fields discussed in Sect. 16.3.
Note that S-waves scattered on the fracture do not occur due to scattering only the
incident P-wave. These S-waves occur as a result of the interference of the incident
P-wave, the cylindrical P-wave scattered from the upper point of the fracture, and
waves from a point source of force:
  2 
cS
fc = ρcP 1−2 G(t). (16.41)
cP

In Eq. 16.41 and further in the text, G(t) is the wavelet shape function in the
incident P-wave, which is equal to zero outside the wave front.
Also note that the scattered S-waves interfere with the cylindrical S-wave from the
same point source and with the cylindrical S-wave, which occurs when the result of
the interfering of the incident P-wave and the cylindrical P-wave travel from the low
point of the fracture. These cylindrical P- and S-waves are formed due to the fact that
at the time of the incident plane P-wave travel the upper point of the fracture, for this
case a gap in vertical main component of Cauchy stress tensor occurs (Eq. 16.42).
  2 
cS
σ yy = ρcP 1−2 G(t) (16.42)
cP

[email protected]
16 Elastic Wave Scattering on a Gas-Filled Fracture … 223

A similar situation is observed for the lower point of the fracture. Thus, the wavelet
shape in the generated S-waves is not constant along their front and differs somewhat
from the original pulse shape, and not only because these scattered S-waves interfere
with the wave from a moving source of shear stresses.

16.6 Conclusions

In the chapter, the analytical formulas for the scattering of plane P-wave on a gas-
filled fracture located along the motion of the front of this wave are derived. Scattered
S-waves have been identified and studied. These S-waves can be useful for nonde-
structive testing of fractures perpendicular to the incident P-waves. At the stage of
expressions derivation, a comparison was made with visualized elastic wave fields
(wave patterns). The effectiveness of Wave Logica approach for applied study of
wave fields and obtaining the corresponding analytical expressions is shown.

Acknowledgements The reported study was funded by RFBR according to the research Project
No. 18-31-20063.

References

1. Krauklis, P., Krauklis, V.: One type of waves in media with loosely bonded interface. J. Math.
Sci. 55(3), 1725–1732 (1991)
2. Hörmander, L.: The Analysis of Linear Partial Differential Operators I: Distribution Theory
and Fourier Analysis. Springer-Verlag (2009)
3. Vlastos, S., Liu, E., Main, I.G., Li, X.-Y.: Numerical simulation of wave propagation in media
with discrete distributions of fractures: effect of fracture size and spatial distributions. Geophys.
J. Int. 152(3), 649–668 (2003)
4. Burago, N.G., Nikitin, I.S.: Improved model of a layered medium with slip on the contact
boundaries. J. Appl. Math. Mech. 80(2), 164–172 (2016)
5. Glushko, A.I., Nikitin, I.S.: One method of calculating brittle fracture waves. Mech. Solids
21(3), 130–135 (1986)
6. Muratov, M.V., Petrov, I.B.: Application of fractures mathematical models in exploration seis-
mology problems modeling. In: Petrov, I.B., Favorskaya, A.V., Favorskaya, M.N., Simakov,
S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 Years of the Development of
Grid-Characteristic Method, SIST, vol. 133, pp. 120–131. Springer (2019)
7. Leger, A., Deschamps, M. (eds.): Ultrasonic Wave Propagation in Non Homogeneous Media,
vol. 128. Springer Science & Business Media (2009)
8. Favorskaya, A., Kabanova, A., Petrov, I.: Applying Wave Logica method for geophysical
prospecting. In: 20th Conference Oil and Gas Geological Exploration and Development, pp. 1–4
(in Russian) (2018)
9. Favorskaya, A.V., Zhdanov, M.S., Khokhlov, N.I., Petrov, I.B.: Modeling the wave phe-
nomena in acoustic and elastic media with sharp variations of physical properties using the
grid-characteristic method. Geophys. Prospect. 66(8), 1485–1502 (2018)
10. Favorskaya, A., Golubev, V., Grigorievyh, D.: Explanation the difference in destructed areas
simulated using various failure criteria by the wave dynamics analysis. Procedia Comput. Sci.
126, 1091–1099 (2018)

[email protected]
224 A. V. Favorskaya

11. Favorskaya, A.V.: The use of multiple waves to obtain information on an underlying geological
structure. Procedia Comput. Sci. 126, 1110–1119 (2018)
12. Glassner, A.S. (ed.): An Introduction to Ray Tracing. Elsevier (1989)
13. Julian, B.R., Gubbins, D.: Three-dimensional seismic ray-tracing. J. Geophys. 43, 95–113
(1977)
14. Zak, A., Krawczuk, M., Ostachowicz, W.: Propagation of in-plane waves in an isotropic panel
with a crack. Finite Elem. Anal. Des. 42(11), 929–941 (2006)
15. Ludwig, R., Lord, W.: A finite-element formulation for the study of ultrasonic NDT systems.
IEEE Trans. Ultrason. Ferroelectr. Freq. Control 35(6), 809–820 (1988)
16. Bartoli, I., Marzani, A., Lanza di Scalea, F., Viola, E.: Modeling wave propagation in damped
waveguides of arbitrary cross-section. J. Sound Vibr. 295(3–5), 685–707 (2006)
17. Schoenberg, M.: Elastic wave behaviour across linear slip interfaces. J. Acoust. Soc. Am. 68,
1516–1521 (1980)
18. Pyrak-Nolte, L.J., Myer, L.R., Cook, N.G.W.: Anisotropy in seismic velocities and amplitudes
from multiple parallel fractures. J. Geophys. Res. 95, 11345–11358 (1990)
19. Kelly, K.R., Ward, R.W., Treitel, S., Alford, R.M.: Synthetic seismograms finite-difference
approach. Geophysics 41(1), 2–27 (1976)
20. Kruger, L., Oliver, S., Saenger, E.H., Oates, S.J., Shapiro, S.A.: A numerical study on reflection
coefficients of fractured media. Geophysics 72(4), D61–D67 (2007)
21. Zhang, J., Gao, H.: Elastic wave modelling in 3-D fractured media: an explicit approach.
Geophys. J. Int. 177, 1233–1241 (2009)
22. Silvestrov, I., Reda, B., Evgeny, L.: Poststack diffraction imaging using reverse-time migration.
Geophys. Prospect. 64, 129–142 (2016)
23. Magomedov, K.M., Kholodov, A.S.: The construction of difference schemes for hyperbolic
equations based on characteristic relations. USSR Comput. Math. Math. Phys. 9(2), 158–176
(1969)
24. Favorskaya, A.V., Petrov, I.B. Grid-characteristic method. In: Favorskaya, A.V., Petrov, I.B.
(eds.) Innovations in Wave Processes Modelling and Decision Making, SIST, vol. 90, pp. 117–
160. Springer (2018)
25. Favorskaya, A.V., Kabisov, S.V., Petrov, I.B.: Modeling of ultrasonic waves in fractured rails
with an explicit approach. Doklady Math. 98(1), 401–404 (2018)
26. Favorskaya, A., Khokhlov, N.: Modeling the impact of wheelsets with flat spots on a railway
track. Procedia Comput. Sci. 126, 1100–1109 (2018)
27. Beklemysheva, K.A., Vasyukov, A.V., Petrov, I.B.: Numerical modeling of delamination
in fiber-metal laminates caused by low-velocity impact. In: Petrov I.B., Favorskaya, A.V.,
Favorskaya, M.N., Simakov, S.S., Jain, L.C. (eds.) Proceedings of the Conference on 50 Years
of the Development of Grid-Characteristic Method, SIST, vol. 133, pp. 132–142. Springer
(2019)
28. Favorskaya, A., Petrov, I., Grinevskiy, A.: Numerical simulation of fracturing in geological
medium. Procedia Comput. Sci. 112, 1216–1224 (2017)
29. Favorskaya, A.V., Petrov, I.B.: The use of full-wave numerical simulation for the investigation
of fractured zones. Math. Models Comput. Simul. 11(4), 518–530 (2019)

[email protected]
Chapter 17
Discrete Element Method Adopting
Microstructure Information

Andrew A. Zhuravlev , Karine K. Abgaryan

and Dmitry L. Reviznikov

Abstract Multiscale discrete element model adapting information on material

microstructure is introduced. Modeled structure is represented by a set of tetrahe-
dral elements bound by their faces. Each element has an associated atomic sample,
which represents atomic structure of the element. All element properties are deter-
mined from molecular dynamics simulation of its associated sample. Therefore, none
of the specific properties of the material are needed aside from its atomic composition
and microstructure. The chapter focuses on elastic behavior of modeled structures.
Comparisons of the results obtained using multiscale discrete element simulation
with molecular dynamics data and known macroscopic material properties show
fairly good accuracy of the proposed model.

17.1 Introduction

Growth of available computational resources enables modeling of systems with a

high-level time and space resolution. One of the most promising ways in this direction
is multiscale modeling [1, 2]. Nevertheless, the traditional methods of computational
solid mechanics make use of macroscopic material properties without considering
the microstructure. This fact limits applicability of these methods to design new
materials with required properties.

A. A. Zhuravlev · K. K. Abgaryan · D. L. Reviznikov (B)

Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
A. A. Zhuravlev
e-mail: [email protected]
K. K. Abgaryan
e-mail: [email protected]
Federal Research Centre “Information and Control” of the RAS, 44, Vavilova ul, Moscow
119333, Russian Federation

© Springer Nature Singapore Pte Ltd. 2020 225

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_17

[email protected]
226 A. A. Zhuravlev et al.

Numerical methods for structures simulation can be classified into two major
groups. First group contains methods based on continuum equations [3]. This group
can be divided into two subgroups. The first one contains methods solving the equa-
tions on a mesh. The use of adaptive grids allows to achieve a high spatial resolution
and get the detailed pictures of the process [4]. Typical representative of this group is
the finite element method. Some approaches have been proposed to make this class
of methods applicable to processes with breaking objects [5].
Second subgroup contains meshless methods [6]. These methods search the solu-
tion of continuum equations as a combination of basic functions. As an example let
us refer to the meshless method [7]. One of the popular methods of this class is the
smoothed-particle hydrodynamics, which is based on a set of computational particles
approximating the solution with a smoothing kernel function [8, 9]. At the begin-
ning, it was mainly used for fluid simulation but later it was adopted for simulation
of solids. Methods of this group are generally computationally expensive but more
flexible as it comes to complex processes like those with failure of solids.
Second big group contains the discrete element methods. These methods come in
many different variants, a comprehensive review can be found in [10]. Basic meth-
ods of this type model the objects as a set of closely packed particles interacting
with some potential and moving according to laws of classical dynamics. These
methods are relatively simple but in the case of wise potential choice and careful
parameters tuning they can be used to simulate quite complex processes [11, 12].
Movable cellular automatons use different approaches for linking and unlinking ele-
ments that enable fine-tuning of a model to some known material properties [13].
Dissipative particles dynamics method is one of the first methods trying to adopt infor-
mation about atomic structure of material through additional dissipative and random
forces [14]. Extended discrete element method allows to include the additional prop-
erties in simulation like thermodynamic state of element [15]. In general, these meth-
ods can work with less information about simulated materials properties compared
to the methods based on continuum equations.
The purpose of the presented chapter is to introduce a novel multiscale method for
materials modeling, which requires information only from the atomic level (atomic
structure and potential of atomic interaction). The latest can be obtained from first
principles calculations. Thus, the proposed approach creates the ground for design
of new materials with the required properties.
This chapter is structured as follows. Section 17.2 presents the mathematical
model and deals with choice of computational parameters required by model. In
Sect. 17.3, the results obtained from the proposed model are compared with the
results of molecular dynamics simulation and macroscale parameters of selected
materials. Section 17.4 concludes the chapter.

[email protected]
17 Discrete Element Method Adopting Microstructure Information 227

17.2 Multiscale Discrete Element Model

Modeled structures are virtually divided into tetrahedral elements. Pairs of elements
are stiffly bound by their faces. Each element contains a small but representative
sample of atomic structure. For example, for a crystalline material it will be a sample
of crystalline lattice oriented in certain direction. The whole system evolution is
governed by equations of motion for every element vertex.
Element properties are determined by the atomic sample associated with the
element (Fig. 17.1):

AS = (B, M, P, V, Q, E),

where B ∈ R3×3 is the matrix composed of basis vectors of atomic sample, the sample
is considered to be a parallelepiped generated by these vectors, M ∈ R N ×N is the
diagonal matrix of atomic masses, P ∈ R3×N is the matrix of atom positions in local
(fractional) coordinates, V ∈ R3×N is the matrix of atom velocities, Q ∈ R3×3 is the
lattice rotation matrix, E : R3×N → R is the potential energy function. Positions of
atoms in the element can be calculated using the following formula:

PLocal = Q · B · P.

The general scheme of computational process is presented in Fig. 17.2. Let us

consider the process in detail starting from element deformation. The deformation
for tetrahedral elements is always linear and can be restored using the following
formula:

Fig. 17.1 Atomic sample

associated with the element

[email protected]
228 A. A. Zhuravlev et al.

Fig. 17.2 General modeling scheme

 −1
DElement = (v1 − v0 |v2 − v0 |v3 − v0 ) · v1∗ − v0∗ |v2∗ − v0∗ |v3∗ − v0∗ ,

where DElement is the deformation tensor of the element, vi∗ are the positions of vertices
at the beginning of simulation, vi are the current positions of vertices.
Then, associated atomic sample for the element is considered to be subjected to
the same deformation (Fig. 17.3):

DSample = DElement .

Fig. 17.3 Atomic sample deformation induced by element deformation: a element deformation,
b atomic sample deformation

[email protected]
17 Discrete Element Method Adopting Microstructure Information 229

To get a sample’s response to deformation, it is deformed and then a few steps

of molecular dynamics are performed in the sample to calculate the resulting lattice
structure change due to the deformation. Then real atom positions in global reference
frame can be calculated by the following formula:

PGlobal = DSample · Q · B · P.

Note that atomic samples should not be very large, as they will be involved in
computationally expensive molecular dynamics simulations, neither they should be
too small, as they must be able to accurately represent properties of material. Sam-
ples associated with the element represent its internal structure and can be consid-
ered continuously repeated in every direction. To model such behavior, the periodic
boundary conditions are imposed on sample, and pairs of atoms are interacting only
by minimum distance:
 
δi j = arg min ri + Bδ − r j ,
δ∈(−1,0,1)3
ri j = ri + Bδ i j − r j .

Motion of atoms in the sample is governed by Newton’s law of motion, where

forces acting on particles are derived from interatomic potential:

V̇ = FM−1 = −∇ E(PGlobal )M−1
.
ṖGlobal = V

For this chapter, the embedded-atom method is used to compute forces between
atoms. It is a many-body potential that gives potential energy in the following form:
 
     
Ei = F ρ ri j + 21 ϕ ri j ,
j=i j=i

E= Ei ,
i

   
where ϕ ri j is the pair-wise potential function, ρ ri j is the electron density contri-
bution of atom j in position of atom i, F is the embedding function or energy of atom
at the point with given electron density. It is a very generic form of potential energy.
All functions are tabulated and then interpolated with splines. Forces are calculated
according to the potential energy gradient:

 ∂F(ρ) ∂F(ρ) ∂ρ(ri j ) ∂ϕ (ri j )

∇ri E = ∂ρ ρ=ρ
+ ∂ρ ρ=ρ ∂ri
+ ∂ri
r̂i j ,
j=i
  
i j

ρi = ρ ri j .
j=i

[email protected]
230 A. A. Zhuravlev et al.

The forces can be calculated in two steps. At the beginning, the electron density
is calculated for every atom in system evaluating density contribution from every
neighboring atom. Then, a resulting force can be calculated in a single pass over
all neighbors for every particle. It is one of the least computationally expensive
potentials aside from pair potentials, and it represents a good trade-off between the
computational complexity and physical accuracy of molecular dynamics model.
Equations of motion are integrated by velocity Verlet integration method, which
is a common choice for molecular dynamics simulations. For atomic simulations, a
general practical rule is to use femtosecond time steps:

P(t + dt) = P(t) + V(t)dt − ∇ E(P(t))M−1 dt2
2

.
V(t + dt) = V(t) − (∇ E(P(t)) + ∇ E(P(t + dt)))M−1 dt2

Sample’s response on a deformation is given by a virial stress of a molecular

system:
 
   T 
SSample = |D 1
− V − V̄ M V − V̄ + 1
ri j fiTj ,
Sample ·Q·B| 2
i, j=i
V·1 N ×1 ·11×N
V̄ = N
,

where SSample is the stress tensor of the sample, DSample · Q · B is the sample volume,
V̄ is the average sample atom velocity.
Considering that associated sample is a representative fragment of the element and
taking into account that inner volume of the element is much greater than boundary
volume, we conclude that properties of the element are completely determined by
corresponding properties of the associated sample. Therefore, stress tensor of the
element is equal to stress tensor of its associated sample:

SElement = SSample .

To determine forces acting on every vertex from the element first of all we ought
to determine element response to deformation.
Having a stress tensor for every element, we can calculate their dynamics.
Dynamics of every element is fully determined by dynamics of their vertices:

flm = Sl · nlm ,
nlm = 1
2
vlm 1 − vlm 0 × vlm 2 − vlm 0 ,
flmk = flm aalmk
lm
,

where Sl is the stress tensor of element l, nlm is the normal to face m of element l,
alm is the face area, vlm k is the vertex k of face m of element l, almk is the part of face
m of element l area closest to vertex vlm k , flmk is the force acting on vertex vlm k from
face m of element l (Fig. 17.4a).

[email protected]
17 Discrete Element Method Adopting Microstructure Information 231

Fig. 17.4 Area and volume partition between vertices: a assigned area, b assigned volume

Having forces acting on element vertices, we ought to assign masses to vertices

to calculate their accelerations. The part of the element closest to the corresponding
vertex is associated with the vertex and the element mass is distributed between
vertexes according to these volume parts (Fig. 17.4b).
After calculation of accelerations equations of motion are integrated for every
vertex using velocity Verlet method:

rk (t + dt) = rk (t) + vk (t)dt + fkm(t)
2
dt
k 2 .
vk (t + dt) = vk (t) + fk (t)+fk (t+dt)
2m k
dt

To keep the calculations correct an admissible time step is ought to be selected.

Time step selection can be guided by the following algorithm:
1. Estimate Young’s modulus (E) of material from small molecular dynamics
simulation.
2. Estimate speed of sound in material from Young’s modulus:

E
c= .
ρ

3. Select a time step in such a way that no perturbation travels through an entire
element in a single step:

c · dt < L,
dt < Lc = L Eρ ,

where L is the characteristic length of elements in model.

There are a few things left to consider before the beginning of simulation. First
of all, we have to find out what the minimum possible sample size of the element

[email protected]
232 A. A. Zhuravlev et al.

is. Having periodic boundary conditions imposes the strong restrictions on the sam-
ple size. Most interatomic potentials have so called cutoff distance rcut , a maximum
distance between interacting atoms; atoms at a larger distance are considered nonin-
teracting. Periodic boundary conditions in three-dimensional space make 27 virtual
copies of each atom that can interact with other atoms, but we select only a closest
one. Thus, we ought to ensure that for every atom there is no more than one copy of
every other atom in cutoff sphere. This is provided by creating a sample large enough
to contain cutoff sphere. For cubic lattice it is expressed in the following way:
 
k = α 2racut ,
L = ka,

where rcut is the cutoff distance depending on the potential, a is the size of crystal
lattice unit cell, α is an expected compression of element, k is the number of crystal
unit cells in the sample, and L is the length of the sample side.
In this research for all experiments except the first one, the base rotation of crystal
lattice in every element is generated from uniform random distribution over rotation
group SO(3). At first, a unit quaternion denoting a chosen rotation is generated in
the following way:

u 1 , u 2 , u 3 ∼ U (0, 1),
√
√ s = 1 − u 1 sin 2π u 2 , 
√ √
v= 1 − u 1 cos 2π u 2 , u 1 cos 2π u 3 , u 1 cos 2π u 3 ,
q = (s, v).

Then, this quaternion is converted into a rotation matrix for convenience.

Initial velocities are assigned from Maxwell–Boltzmann distribution for selected
temperature. Assignment of some initial velocities for atoms in the sample is
important due to abnormal behavior of crystals at close to zero temperatures.
Since the proposed method is computationally expensive, it is highly important
to use a parallelization. Most expensive parts of computation are handed over to
GPU. In this work, all molecular dynamics simulations of samples associated with
elements are conducted on GPU using CUDA technology. In CUDA, a computation
model the whole GPU task is divided into blocks of threads. Threads in one block
can communicate with each other but threads from different blocks have no simple
and fast way of communication. Therefore, a computation of molecular dynamics for
one element is fully conducted by threads of one block as synchronization between
time steps and exchange data on particles movement are needed. Computations in
different elements are completely independent of each other, so threads from different
blocks do not need any means of communication. This type of computations is well
suited for GPU architecture.

[email protected]
17 Discrete Element Method Adopting Microstructure Information 233

17.3 Computational Experiments

Three sets of computational experiments have been performed to validate multiscale

discrete element model. The first group of experiments is directed on comparison
of the proposed model with direct molecular dynamics simulation. The problem
formulation is the following: an ideal crystal of selected material was periodically
continued in one direction and stretched in that direction with constant speed. Exper-
imental sample had the size of 120 unit cells in every direction and was stretched
to 145% of its initial length in 0.15 ns. Experiments were conducted on the cop-
per and aluminum samples. Both materials have face-centered lattice. Initial sample
temperature was set to 300 K. Stress in crystal was studied in this experiment.
Molecular dynamic model contained around 7 millions of atoms. Simulation step
was set to 1 femtosecond. Each experiment required around 12 h on Nvidia Tesla
P100 GPU, which is modern server GPU. The multiscale discrete element model
used 384 elements. The further increase of this value has little effect on the results.
All atomic samples contained around 330 thousands of atoms. Note that the amount
of atoms in samples is the smallest possible to keep periodic images of every atom
from interacting with atom itself. The time step was set to 50 fs. Each experiment
required around 2.5 h on Nvidia GeForce GTX 780 GPU, which is a rather old GPU
for workstations.
Computational experiments show a good correspondence of the results obtained
from proposed model with classical molecular dynamics results, which can be con-
sidered as the exact solution (Fig. 17.5). Proposed model reproduces well stresses
during elastic deformations and stress after the beginning of sample yielding. How-
ever, the model gives slightly different stress at the start of yielding process. This can
be explained by the absence of real boundary in the proposed method. All boundary
conditions are taken into account only on the top level of model, but atomic sam-
ples have no information about sample boundary. This prevents samples from easier
dislocations accumulation and breakdown of boundaries.

Fig. 17.5 Comparison of stresses obtained from molecular dynamics and proposed models for:
a aluminum crystals, b copper crystals

[email protected]
234 A. A. Zhuravlev et al.

The second group of experiments focused on comparison of discrete element

modeling results with macroscopic material properties. Static material loading is
considered. The problem formulation is the following. Polycrystalline sample of
selected material was periodically continued in one direction and stretched in that
direction with constant speed. Three materials were selected: copper and aluminum
having a face-centered crystalline lattice and iron having a body-centered crystalline
lattice. Simulation was carried out in quasi-static way; every few steps sample was
deformed and then stabilized for a few steps. The object contained 1,536 elements.
Each element contained an atomic sample of around 850 atoms for copper and alu-
minum and around 420 atoms for iron. Stress-deformation dependence was studied
to determine Young’s modulus. Reference values of Young’s modulus for studied
materials are: aluminum—70 GPa, copper—130 GPa, and iron—210 GPa.
Experimental results show good correspondence of calculated stress with linear
estimation from Young’s modulus (Fig. 17.6).
The third group of experiments focused on simulation of dynamic processes in
materials. All experimental models contained 3,840 elements. Each element con-
tained an atomic sample of around 850 atoms. In the first computational experiment,
a polycrystalline copper plate was affected by instantaneous impulse in a direction
normal to its free surface. Pressure wave propagation is presented in Fig. 17.7. The

Fig. 17.6 Comparison of stress obtained from the model with stress obtained from linear elasticity
theory for examined materials: a aluminum, b copper, c iron

[email protected]
17 Discrete Element Method Adopting Microstructure Information 235

Fig. 17.7 Pressure wave propagation in copper plate at: a 0.25 ms, b 0.5 ms

reference sound speed here is 4,725 m/s. In the second experiment, a polycrystalline
aluminum rod was affected by instantaneous impulse in a direction parallel to it
(Fig. 17.8, the reference value is 5,091 m/s). In the third experiment, a polycrystalline
aluminum rod was affected by instantaneous impulse in a direction perpendicular to
it. Shear wave propagation was studied in this experiment (Fig. 17.9, the reference
value is 3,103 m/s).
The expected position of the sound wave calculated taking the reference sound
speed in material is given in Figs. 17.7, 17.8, 17.9 by the vertical line. In all cases,
the results of computational experiment are in good correspondence with reference
data.

Fig. 17.8 Pressure wave propagation in aluminum rod at: a 0.3 ms, b 0.55 ms

[email protected]
236 A. A. Zhuravlev et al.

Fig. 17.9 Shear wave propagation in aluminum rod at: a 0.5 ms, b 0.75 ms

17.4 Conclusions

A novel discrete element method for solids simulation has been presented. The main
feature of the proposed method is a combination of macroelements and atomic sam-
ples associated with discrete elements. It enables adopting material microstructure
into the mathematical model. The proposed method does not need knowledge of
any macroscopic properties of material. The required information comes from the
atomic level. The results of computational experiments showed a good agreement
with molecular dynamics results and the reference data for considered materials. The
multiscale discrete element modeling can be applied for design of new materials with
required properties.

Acknowledgements The reported study was funded by RFBR, project number 18-08-00703.

References

1. Tadmor, E.B., Miller, R.E.: Modeling Materials. Continuum, Atomistic and Multiscale
Techniques. Cambridge University Press, Cambridge (2014)
2. Abgaryan, K.K.: Multiscale Modeling in Material Science Problems. MAKS Press, Moscow
(in Russian) (2017)
3. Shabana, A.A.: Computational Continuum Mechanics, 3rd edn. Wiley, Hoboken (2018)
4. Burago, N.G., Nikitin, I.S., Yakushev, V.L.: Hybrid numerical method with adaptive over-
lapping meshes for solving nonstationary problems in continuum mechanics. Comput. Math.
Math. Phys. 56(6), 1065–1074 (2016)
5. Moës, N., Dolbow, J., Belytschko, T.: A finite element method for crack growth without
remeshing. Int. J. Numer. Meth. Engng. 46, 131–150 (1999)
6. Youping, C., Lee, J.D., Eskandarian, A.: Meshless Methods in Solid Mechanics, 1st edn.
Springer, New York (2009)
7. Vasilyev, A.N., Kolbin, I.S., Reviznikov, D.L.: Meshfree computational algorithms based on
normalized radial basis functions. In: Cheng, L., Liu, Q., Ronzhin, A. (eds.) Advances in Neural
Networks, pp. 583–591. Springer International Publishing (2016)

[email protected]
17 Discrete Element Method Adopting Microstructure Information 237

8. Stellingwerf, R.F., Wingate, C.A.: Impact modeling with smooth particle hydrodynamics. Int.
J. Impact Eng. 14, 707–718 (1993)
9. Connolly, A., Iannucci, L., Hillier, R., Pope, D.: Second order Godunov SPH for high velocity
impact dynamics. In: Papadrakakis, M., Kojic, M., Tuncer, I. (Eds.) Proceedings of the 3rd
South-East European Conference Computational Mechanics, pp. 13–35 (2013)
10. Chen, Y., Zimmerman, J., Krivtsov, A., McDowell, D.L.: Assessment of atomistic coarse-
graining methods. Int. J. Eng. Sci. 49, 1337–1349 (2011)
11. Abgaryan, K.K., Zhuravlev, A.A., Zagordan, N.L., Reviznikov, D.L.: Discrete-element simu-
lation of a spherical projectile penetration into a massive obstacle. Comput. Res. Model. 7(1),
71–79 (2015)
12. Abgaryan, K.K., Eliseev, S.V., Zhuravlev, A.A., Reviznikov, D.L.: High-speed penetration.
Discrete-element simulation and experiments. Comput. Res. Model. 9(6), 937–944 (2017)
13. Psakhie, S., Shilko, E., Smolin, A., Astafurov, S., Ovcharenko, V.: Development of a formalism
of movable cellular automaton method for numerical modeling of fracture of heterogeneous
elastic-plastic materials. Frattura ed Integrita Strutturale 24, 26–59 (2013)
14. Groot, R.D., Warren, P.B.: Dissipative particle dynamics: bridging the gap between atomistic
and mesoscopic simulation. J. Chem. Phys. 107(11), 4423–4435 (1997)
15. Peters, B.: Measurements and application of a discrete particle model (DPM) to simulate
combustion of a packed bed of individual fuel particles. Combust. Flame 131(1–2), 132–146
(2002)

[email protected]
Chapter 18
Durability Evaluation of Bonded Repairs
for the Damaged Metallic Structures
Subjected to Mechanical and Thermal
Loads

Alexey A. Fedotov and Anton V. Tsipenko

Abstract The analytical calculation model was developed to determine the repair
joint parameters and to study the relation between repair design parameters stage
by stage. The first stage of calculation is devoted to defining the stress–strain con-
ditions of the joint without influence of the damage—the method of eigenstrains
and eigencurvatures based on Eshelby’s theory of ellipsoidal inclusion is used. The
computations on the second stage include damage geometric parameters and the cal-
culation of the stress intensity factors at the key points of joint is performed. The third
stage is to evaluate the damage growth rate within the repair joint taking into account
the material properties change under cyclic loads. To figure out the real degradation
behavior of carbon fiber plastic material that can be used for a bonded repair patch
preparation, the experimental research was performed to study the variation of the
longitudinal and transversal elastic moduli at −60, +23, and +80 °C and the varia-
tion of Poisson ratio under the cyclic loads at the same values of temperature. The
developed model allows to analyze the effectiveness of the load-transfer abilities of
the bonded patch and estimate the damage growth rate in the repairable structure.
The model can be used as starting point for the analysis of wide number of bonded
repair techniques and design variants.

18.1 Introduction

Airplane maintenance process assumes that the airframe structural element is

repairable if it has damage that leads to or may lead to decreasing of the residual
strength below the defined limits. Cracks and corrosion damage are the widespread
types of damage on metallic aircraft structures. The availability of the robust,

A. A. Fedotov (B) · A. V. Tsipenko

Moscow Aviation Institute (National Research University), 4, Volokolamskoe shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
A. V. Tsipenko
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 239

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_18

[email protected]
240 A. A. Fedotov and A. V. Tsipenko

reliable, and justified, technically and economically, repair techniques influences

on the effectiveness of serviceability of technical systems, where the aircraft plays
the primary role.
Usually, a repair of detected damage is represented by reinforcing metallic patch
installed at damage location to restore the mechanical properties of the initial struc-
ture: the residual strength, stiffness, fatigue resistance, and level of allowable damage.
For commercial airplanes, the basic repair methods are described within Structural
Repair Manuals (SRM) delivered with airplane. These methods mainly include the
application of bolts and rivets as fastening parts.
At the time of damage detection on the airframe structure based on collected
information, the selection of the desired repair alternative is performed:
• The repair procedure is not necessary (damage is inside the outer finish coatings).
• The cosmetic repair procedure (without any patch) and sealing process are required
(the size of damage is within the allowable scale defined in SRM).
• The patch repair is required (the damage size will have significant influence on
the structural residual strength during the continuous service life).
• The repair is not efficient (the damage component should be replaced).
In general, the repair schema for the restoration of the structural properties should
be simple for realization and should have minimal effect on surrounding components.
The installed repair parts should not block the movable structural and system parts
and should not spoil the aerodynamics and aeroelastics (if it is critical for damaged
parts) beyond the established limits.
The typical requirements for repair design are mentioned below [1, 2]:
• Strength restoration of the damaged structure.
• Restriction of the damage growth rate (if the entire damage elimination is not
possible or ineffective).
• Minimal change of the initial local stiffness and stress distribution.
• High durability of the installed repair joint subject to design loads and environ-
mental factors.
• Tolerance to additional mechanical damage of the initial structure during the
service life.
• Reliable inspection process for installation quality and in-service conditions of the
installed repair joint.
• Account of the functional requirements for the structure (aerodynamics, fire
protection, electromagnetic effects, etc.).
In addition to listed above, the acceptable repair procedure should meet some extra
requirements, such as minimal time of repair installation (i.e., minimal time of the
airplane extraction from flight schedule), utilization of the cheap and widespread
material and tools, application of simple process steps, and minimal damage to
adjacent structural elements.
The mentioned requirements can be met by repair procedures using not only bolted
patches but bonded patches as well. Composite bonded repair procedures have some
advantages in comparison with bolted repair techniques [1]:

[email protected]
18 Durability Evaluation of Bonded Repairs for the Damaged … 241

• High stiffness along the defined direction—this property allows us to apply thin
patches and to orient them along the load path only.
• Superior fatigue life of composite materials—exceeding the fatigue life of the
parent structure.
• Low weight of patch—minimal influence on balancing characteristics for repairs
on control surfaces and high lift elements.
• Manufacturing performance to easily produce the patches of different shape, size,
and curvature.
Besides the advantages, the composite patch application for bonded repairs has
its own problems that need to be addressed at the maintenance planning of aircraft
service. One of the primary strength problems is the difference in thermal expansion
values between the composite patch material and damaged initial metallic structure.
If the adhesives of high curing temperature are used, the local thermal stresses must be
evaluated during stress calculation of bonded joints. In many cases of cyclic thermal
loading, the resulting stresses may lead to contact failure in the adhesive layer and
induce the growth of the existing damage. To minimize the negative effect of the
thermal expansion difference between composite patch and metallic parent structure
the hybrid metal-polymeric materials (like SIAL or GLARE) may be applied. These
materials consist of alternate plies of fiberglass and aluminum and demonstrate the
advanced fracture toughness and durability properties [3–5]. The right selection of
the adhesive may compensate the difference in thermal strains of the joined dissimilar
materials [6–10].
Based on described features, the scope of the bonded repair techniques may be
marked out as follows [11, 12]:
1. To decrease the stress intensity:
• In the area of fatigue crack appearance and propagation.
• In the area of corrosion cracking and structural materials.
• In the area, where the allowable damage limits need to be increased (reinforcing
patch installation).
2. To restore the strength and stiffness:
• After the corrosion damage removal beyond SRM limits.
• After the removal of defects in the material bulk.
• After machining the surfaces for the reduction of stress intensity.
• After the thermal damage of structure.
3. To reinforce the weak structure:
• Deformation decreasing at the locations of concentrated stresses.
• Decreasing the level of secondary bending moments.
• Decreasing the vibration level and acoustic damage prevention.
To narrow the search diapason of optimal variants of composite bonded repair
patches, it is necessary to have the flexible calculation module to compute the stress–
strain condition of the bonded joint at every case of damage and assess the probability

[email protected]
242 A. A. Fedotov and A. V. Tsipenko

of damage growth and damage growth rate taking into account the degradation of
material properties used for repair. The computation module described below is built
using the inclusion theory [13, 14]. This applied computational module constructs
the relatively simple but accurate analytical model suitable enough for preliminary
design of the bonded repair joints.
The presented calculations are described in separate sections. Section 18.2 deals
with the core analytical model of the bonded joint. This model utilizes the inclu-
sion methodology for heterogeneous materials in the joint and Hart-Smith model of
adhesive layer. Section 18.3 contains data achieved for carbon fiber material after the
series of tests, where the specimens were loaded by cyclic mechanical force at three
values of temperature. Section 18.4 demonstrates the work results for the bonded
joint model from Sect. 18.2 augmented by test data from Sect. 18.3. The conclusive
graphics in Sect. 18.4 illustrate the variability of the actual stresses at different values
of temperature at the edge of repair patch due to the change of the patch material
properties. Section 18.5 gives the conclusions.

18.2 Analytical Model of the Bonded Repair

To analyze the bonded repair, the repair schema is split into two areas (Fig. 18.1):
the metallic skin with a cutout of the shape same as a patch and inclusion—hybrid
plate “patch-adhesive-skin”.
For the algorithmization purposes, it is assumed that the first stage of calculation is
to assess the patch effect on stress–strain condition of the skin and on the distribution
of the internal forces between the skin and patch, while the damage is absent. The
second stage defines the damage growth characteristics as a reaction for new stress
condition due to the bonded patch existence: Stress Intensity Factors (SIFs) are
evaluated by means of fracture mechanics methods.
In this section, two stages are considered. Section 18.2.1 describes stress com-
putation for bonded joint without damage (Stage 1), and computation of SIFs in the
skin with damage (Stage 2) is discussed in Sect. 18.2.2.

Fig. 18.1 Inclusion schema

[email protected]
18 Durability Evaluation of Bonded Repairs for the Damaged … 243

18.2.1 Stage 1. Stress Computation for Bonded Joint Without

Damage

Stress computation is performed separately for skin with a cutout and for inclusion
with a subsequent junction of solutions per strain compatibility conditions. Stresses
and strains are defined by external mechanical and thermal loads. One-sided installa-
tion of the bonded patch leads to the appearance of the additional bending moment in
the bonded joint. Bending moment governs the variable mechanical stresses through
the material thickness, this variability is considered as superposition of bending and
membrane components.
The computation process is divided into three steps in order to find the stresses in
the one-sided bonded system under thermomechanical loads [15].
The first step calculates stresses in the skin at the patch location and stresses in the
patch due to temperature differences T after cooling the system “patch-adhesive-
skin” down to minimal service temperature. The method of eigenvalues of strains
and curvatures by Eshelby [16] and Beom [17] is used to define the fields of strains
and curvatures in the skin in the inclusion area and out of it.
The second step allows evaluating the dominance of loading factors within the
bonded joint and potentially reducing the amount of calculations. This step finds
f∗
the virtual value of applied stress σi j to nullify thermal bending stresses in the skin
f∗
along the patch. Further, the value σi j is compared with initial loads in the skin σ∞i j .
f∗
If the condition σ∞i j ≤ σi j becomes true, then the case of thermal loads is assumed
critical, and SIFs are calculated based on stress found at the first step. Otherwise,
f∗
the additional third step is performed, and stresses from loads of σ∞i j − σi j are
calculated.
f∗
At the third step, previously calculated load of σ∞i j − σi j is assumed as initially
sstep2 f∗
applied mechanical load, and σi j = σi j (0) represents a thermal load in the center
of skin under the patch. Calculation is reduced to two-dimensional case loaded with
f∗
specific force P = (σ∞22 −σ22 )ts . The patch is assumed as free of mechanical stress,
sstep2
and the skin is additionally loaded with internal stress of σi j (that is matched to
sstep2
force of P0 = σ22 (0)ts ) [18]. The mean and bending stresses in the skin under
the patch directly affecting the damage growth rate are calculated at this step.

18.2.2 Stage 2. Computation of Stress Intensity Factors

in the Skin with Damage

The stress values acquired at the previous stage are used to define SIFs at the crack
tip and evaluate the crack growth parameters. The patch effect is simulated by means
of a set of elastic springs bridging the crack. Mean and bending stresses σ is j i σ is j


are assumed constant along the crack line and equal to values in the center of skin
under the patch.

[email protected]
244 A. A. Fedotov and A. V. Tsipenko

SIF at the crack tip is determined by membrane and bending stress contribution:
√ √
2z E s πa 3E s πa
K I (z) = K mem − K b , K mem = h 1 (1), K b = h 2 (1),
ts 2 2
1,2
where h 1,2 = √h1−r 2
, r = x/a, and 
h 1,2 are the solutions of the following normalized
integral equations [19, 20]:

1 
1 h 1 (η) σ0
− dη + (ktt a)
h 1 (r ) + (ktb a)
h 1 (r ) = m ,
2π (r − η) 2 Es
−1
1   2 1  
3 h 2 (η) 15 a √ a 

− dη − L 10 |r − η| h 2 (η)dη
2π (r − η)2 2π (1 + νs ) ts ts
−1 −1

σ0
+ (kbt a)
h 1 (r ) + (kbb a)
h 2 (r ) = b ,
Es

where ki j are the elastic constants of the springs bridging the crack: ktt = Edsttts ,
ktb = E6ds tbt 2 , kbt = E6ds btt 2 , kbb = 36d bb
E s ts3
, and di j are the elements of the inverted compliance
s s
matrix ci j described in detail in paper [21], E s and ts are the skin elastic modulus
and skin thickness, respectively, a is a half of crack length.
The elements of the compliance matrix can be found based on well-known method
of bonded joint analysis, where the equations of strains in the adhesive layer are solved
(A index is referred to adhesive layer properties) [22] provided by Eq. 18.1.
 
d 3γ A GA 1 1 dγ A
−4· · +  · =0
dy 3 tA E s · ts E p · tp dy
d 4ε A E 1 1
4
+ A · + · εA = 0 (18.1)
dy tA Ds Dp

Here, Ds, p is the bending stiffness of the skin and the patch, respectively, E s, p =
E s, p /(1 − νs,2 p ), E A = 2G A /(1 − ν A ).
The rotation of crack surfaces  θ0 and the opening displacement along the skin
centerline  ν0 can be found from Eq. 18.1 taking into account the conditions at the
boundaries between the adhesive layer and coupling parts.
The link of loading factors and resulting strains in the adhesive layer is described
per matrix expression utilizing the compliance parameters of the system of bridging
springs:

[email protected]
18 Durability Evaluation of Bonded Repairs for the Damaged … 245

ṽ0 ctt ctb n0

= ,
θ̃0 cbt cbb m0

where n 0 = N p y=0
= −N0 , m 0 = Ns | y=0 = −M0 .

18.3 Elastic Properties Degradation of the Repair Patch

Material

The damage propagation in the aluminum skin is defined by applied cyclic load and
the material fatigue properties evaluated during the series of standardized tests. In
the repair joint, the portion of the applied load is transferred through the repair patch
and, thus, the patch loading will be cyclic. In this case, one of the primary factors
affecting the repair effectiveness is the elastic properties degradation of the patch
material subject to applied cyclic load.
Practical feasible methods to define and control the properties degradation and
Polymer Cement Mortars (PCM) fracture parameters are intensively developed [23].
Nevertheless, the researchers are coming to a conclusion that it is necessary to per-
form the series of experiments with the specimens of required layup and made of
required material to get the authentic information about structure behavior under
fatigue (cyclic) load taking into account the big scatter of the tracked and logged
parameters.
The change of external thermal, moisture, and other climate conditions affects
the PCM properties’ degradation substantially. This effect noticeably complicates
the simulation process of the real PCM service conditions at research facilities. It is
noted that even at low temperature in the very beginning of climatic exposure with
the case of no mechanical loads, the properties of composites may change due to the
relaxation of the material’s initial structural nonequilibrium obtained at specimen
manufacturing [24–26].
A large number of studies are devoted to alteration of strength limits of compos-
ites under the variety of external load factors, while the elastic modulus variation
sometimes stands out of scope. For the woven composites, the change of the elastic
modulus will be much more visible than for composites made of unidirectional tapes
[27]. Patch material elastic modulus is one of the determinant parameters for the
bonded repair procedures design. To figure out the degradation behavior of carbon
fiber plastic material that is proposed to use for the bonded repair patch preparation,
the current research is studied: the variation of the longitudinal and transversal elastic
moduli of the carbon fiber plastic specimens at temperatures −60, +23, and +80 °C
and variation of Poisson ratio under cyclic load at the same values of temperature.
Elastic moduli degradation was evaluated in relation to elastic moduli at the first load
cycle at the temperatures mentioned above.

[email protected]
246 A. A. Fedotov and A. V. Tsipenko

Hereinafter, Sect. 18.3.1 provides a description of the testing procedure and mate-
rials. Results of fatigue tests are presented in Sect. 18.3.2. Elastic modulus change
evaluation based on fatigue test data is discussed in Sect. 18.3.3.

18.3.1 Testing Procedure and Materials

The study was performed in the laboratory of Kazan National Research Techni-
cal University using the servo-hydraulic testing machine ITW BiSS with a climate
chamber. Each specimen for fatigue tests was subjected to cyclic tension load with
ratio R = 0.1. The amplitude of tension load was defined as 67% of mean fracture
load for such specimens at current temperature. Bonded strain gauges were used for
gathering the strain data due to longitudinal and transversal stretching of the loaded
specimens. Tests were performed at three values of temperature: −60, +23, and +
80 °C. Series of 10 specimens were tested at each temperature. General testing pro-
cedure matched ASTM D3039/D3479 requirements, specimen geometry for fatigue
tests, and location of the strain gauges are shown in Fig. 18.2.
Test specimens were made of plain weave carbon fabric ECC 450 and epoxy resin
CHS Epoxy 520 as a matrix. The specimens for fatigue tests were flat plates with
6 plies of carbon fabric with layup [0/90]0 , the specimens had the preformed tabs
to clamp specimens in the grips of the testing machine. The properties of composite
material are shown in Table 18.1.
The specimens were manufactured per Resin Transfer Molding (RTM) process
with curing in the oven in accordance with curing cycle recommended by the resin
system supplier. The general view of specimens with bonded strain gauges is pre-
sented in Fig. 18.3. Since the object of current research was to analyze the elastic
properties but not a fracture of specimens, the specimens for fatigue tests were
manufactured without stress concentrators in the working zone.

Fig. 18.2 The sketch of the specimen for fatigue test (dimensions are in mm)

[email protected]
18 Durability Evaluation of Bonded Repairs for the Damaged … 247

Table 18.1 The properties of

Property Value
the ECC 450 material
Number of filaments per yarn 3000
Areal weight, kg/m3 0.200
Ply thickness, mm 0.327
Fiber type Torayca T300 J
Fiber ultimate strength, MPa 4210
Fiber elastic modulus, GPa 230
Fiber ultimate strain 1.8%
Fiber density, kg/m3 1780
CTE, 1/°C −0.43 × 10−6

Fig. 18.3 General view of the specimens for fatigue tests

18.3.2 Results of Fatigue Tests

Fatigue tests of carbon specimens were split into two stages. A quasi-static fracture
load F fracture at three values of temperature was defined at the first stage. The spec-
imens were elongated with constant load growth rate of 1 mm/min until specimen
failure. Fracture stress values for each temperature tested are shown in Fig. 18.4.
The maximum strength (604 ± 30 MPa) was observed for the specimens tested
at room temperature. The maximum scatter of measurements (approx. 51 MPa) was
found for specimens tested at −60 °C. The increasing influence of random defects
in material structure on its strength due to embrittlement during the cooling may
account for this effect.
The second stage contained the fatigue tests of specimens under the tension load
with 5 Hz frequency and number of cycles 105 . The frequency value was selected
to minimize the time spent for one specimen testing and to reduce the heating level
of specimens during the tests [28]. The number of load cycles is stated as a minimal
requirement to estimate the fatigue strength by means of fatigue rate. The cyclic load
amplitude was set at 0.67 F fracture for each test temperature. This load value matches
the patch design ultimate load at the bonded repair procedure creation.

[email protected]
248 A. A. Fedotov and A. V. Tsipenko

Fig. 18.4 Fracture stresses at static load (vertical lines represent the error bars of measurements)

18.3.3 Elastic Modulus Change Evaluation Based on Fatigue

Test Data

The analytical relations satisfying the test data at three values of temperature were
constructed after the data statistical processing. Since all series depicted the fast
growth of the longitudinal strains on the starting period of tests within 500 load
cycles and gently sloping strain growth for the rest time of tests, the results of the
experiments are approximated by the following relation with the minimal dispersion:
 
E 1 (N ) = E 1500 + E 11 − E 1500 k −N , for 1 ≤ N < 500,
E 1 (N ) = E 1500 + k1 N + k2 N 2 , for N ≥ 500,

where N is the number of load cycles, E 11 is the mean longitudinal elastic modulus at
the first load cycle, E 1500 is the mean longitudinal elastic modulus at the 500th load
cycle, k, ki are the relation empirical parameters. Parameter values for the approxi-
mating relation are defined in Table 18.2, the approximating relation curves calcu-
lated per Table 18.2 data and reduced to elastic modulus at the first load cycle E 11
are plotted in Fig. 18.5a, b.
The variation of the transversal elastic modulus was evaluated by change of Pois-
son ratio on the number of load cycles. Poisson ratio change is well approximated
by linear relation:

[email protected]
18 Durability Evaluation of Bonded Repairs for the Damaged … 249

Table 18.2 Parameters for

Parameter −60 °C +23 °C +80 °C
approximating relation
E 1 (N ) 1 ≤ N < 500
E 11 , GPa 110.223 106.594 53.402
E 1500 , GPa 74.016 70.841 42.508
k 1.056 1.033 1.014
N ≥ 500
k1 6.965 × 10−7 −4.416 × −1.458 ×
10−7 10−4
k2 −6.965 × 3.382 × 7.940 ×
10−10 10−10 10−10

ν12
final
− ν12
1
ν12 (N ) = ν12
1
+ N,
105

where ν121
is the Poisson ratio at the first load cycle, ν12
final
is the Poisson ratio at the
5
final 10 th load cycle. Poisson ratio curves are plotted in Fig. 18.6.

18.4 Bonded Repair Effectiveness Calculation Results

The computation efforts of the repair bonded joint in accordance with the pro-
posed analytical model allow to estimate the effectiveness of the selected repair
configuration and narrow the search area to find the optimal design of the system
“skin-adhesive-patch”. The repair configuration effectiveness estimation assumes the
calculation of the following values:
• SIF at the crack tip is directly affecting the rate of the crack growth in the skin
under the patch.
• Stress concentration and the skin near the patch edge—additional crack growth
condition in the skin along the patch edge will depend on this value.
Figure 18.7 shows the relation a = f (N ) between the crack length and num-
ber of load cycles for the 2.0 mm thick aluminum skin made of 7075-T6 alloy.
The repair patch diameter is 100 mm, and the patch is made of composite materi-
als (carbon fiber, fiberglass, and boron-epoxy fabrics) of quasi-isotropic layup and
GLARE 2-3/2-0.2 hybrid metal-polymeric material. Initial crack length is 10 mm.
Load diapason is σmax = 0.4σ B Al , σmin = 0.04σ B Al . Adhesive material is Cytec
FM-73, the mechanical properties are found in the official technical data supplied by
manufacturer [29]. Material properties used for calculation are shown in Table 18.3.
The analyzed fiberglass patch showed the negative results as it was not able to
block the propagation of crack. The only material demonstrating the crack stopping
abilities is boron plastic B(4)/5505.

[email protected]
250 A. A. Fedotov and A. V. Tsipenko

Fig. 18.5 Relationship of the longitudinal elastic modulus on the number of load cycles: a the
range of cycles shown is 0 to 500, b the range of cycles shown is 0 to 105

If the patch is made of carbon fiber plastic ECC 450/ CHS Epoxy 520 tested in
Sect. 18.2, the damage propagation features will substantially depend on temperature
of the bonded joint as shown in Fig. 18.8.
It is evident that the repair patch does not restrict the damage growth in the skin
material at +80 °C. Thus, this patch should not be used to repair the structure pri-
marily loaded at the elevated temperature. It is noted that the a = f (N ) behavior is
changed significantly: the crack growth is intensively increasing at the stating period

[email protected]
18 Durability Evaluation of Bonded Repairs for the Damaged … 251

Fig. 18.6 Relationship of Poisson ratio on the number of load cycles: a absolute values of Poisson
ratios, b Poisson ratio change relative to the value at the first loading cycle

Fig. 18.7 A comparison of the repair effectiveness for the patches made of different materials

of loading unlike the charts shown in Fig. 18.7. Thus, the elastic properties degra-
dation is an essential factor influencing the damage propagation and the inspection
interval criteria for the current repair location.
Figures 18.9 and 18.10 show the variation of the stress concentration K t in the
E t  E S,P
skin along the patch edge depending on stiffness ratio S (S = EP tSP , E S,P = 1−ν S,P
)
S
at different B/A ratios of the elliptic patches (A and B are the lengths of the ellipse
semi-axis).
The higher patch stiffness and the lower patch radius provide the significant
increasing of the stress concentration in the skin. This leads to reduction of the repair

[email protected]
252 A. A. Fedotov and A. V. Tsipenko

Table 18.3 Patch material properties

E x , GPa E y , GPa Gxy , GPa α x , 1/°C α y , 1/°C t, mm
Carbon T300/934 74 74 19.89 1.335 × 24.07 × 1.58
10−6 10−6
S-glass/epoxy 43 8.9 5.89 11.23 × 11.23 × 1.95
fiberglass 10−6 10−6
B(4)/5505 boron 204 18.5 5.59 10.28 × 10.28 × 1.04
10−6 10−6
GLARE 2-3/2-0.2 68.9 53.8 15.2 16.38 × 24.48 × 1.10
10−6 10−6
7075-T6 sheet 73.1 73.1 28.0 23.2 × 23.2 × 2.0
(reference) 10−6 10−6

Fig. 18.8 A comparison of the repair effectiveness for the patch made of carbon ECC 450/CHS
Epoxy 520 at various temperatures

configuration effectiveness due to additional areas of critical stresses, where the val-
idating stress calculations are required. For the patch geometry within the range
0.5 < B/A < 2, the utilization of the stiffness value for the degraded patch material
gives the underestimation of the stress concentration in the skin near the edge of the
patch.

[email protected]
18 Durability Evaluation of Bonded Repairs for the Damaged … 253

Fig. 18.9 Stress concentration in the skin near the patch edge—static load

Fig. 18.10 Stress concentration in the skin near the patch edge—cyclic load

18.5 Conclusions

The following statements may conclude the presented chapter:

[email protected]
254 A. A. Fedotov and A. V. Tsipenko

1. The performed literature survey demonstrates the potential ability to install

bonded repairs on the damaged metallic structures of the commercial aircraft. It is
possible to develop the repair process that can be realized by airline maintenance
personnel in the actual conditions of the design, technological, and environmental
restrictions.
2. The experimental study of the fatigue degradation for the composite material
mechanical properties was performed. The test specimens were subjected to
cyclic loads at various temperatures. The effect of material stiffness degradation
on the bonded repair performance was evaluated.
3. The analytical technique for parameter calculation of the bonded repair joint was
developed. This technique is suitable for preliminary design and concept review
of the repair configurations.
4. The proposed analytical technique can be tuned for specific structural and repair
materials and solutions. This technique has a potential for further development
and improvement to account several extra features including (but not limited to):
a. Adhesive behavior under cyclic and long-term loads.
b. The character of adhesion between the adhesives and coupling surfaces.
c. Combined influence of the climatic factors on the bonded joint strength and
the material mechanical properties.
Also it can be interesting to expand the scope of the proposed analytical technique
to the damaged composite structures of modern commercial airplanes like Yak-
242, Airbus A350, or Boeing 787 and 777-9X. Such expansion requires revision of
the computational core of the technique to simulate the damage growth processes
within the complex heterogeneous composite material. The expansion will require
performing the intensive experimental and theoretical research.

References

1. Baker, A.A.: Repair of metallic airframe components using fibre-reinforced polymer (FRP)
composites. In: Karbhari, V.M. (ed.) Rehabilitation of Metallic Civil Infrastructure Using Fiber
Reinforced Polymer (FRP) Composites, pp. 11–59. Elsevier Ltd (2014)
2. Baker, A.A.: A proposed approach for certification of bonded composite repairs to flight-critical
airframe structure. Appl. Compos. Mater. 18, 337–369 (2011)
3. Kablov, E.N., Antipov, V.V., Klochkova, Y.Y.: Aluminium-lithium alloys of new generation
and aluminum fiberglass laminates on their basis. Tsvetnye Metally 86–91 (2016)
4. Antipov, V.V., Serebrennikova, NYu., Senatorova, O.G., Morozova, L.V., Lukina, N.F., Nefe-
dova, YuN: Hybrid laminated materials with slow fatigue-crack development. Russian Eng.
Res. 37(3), 195–199 (2017)
5. Shestov, V.V., Antipov, V.V., Ryabov, D.R.: Corrosion resistance and mechanical properties
of layered structural material based on aluminum alloy and fiberglass thin sheets. Metallurgist
60(11–12), 1191–1196 (2017)
6. Petrova, A.P., Lukina, N.F., Dement’eva, L.A., Anikhovskaya, L.I.: Film structural adhesives.
Polym. Sci. Ser. D 8(2), 138–143 (2015)

[email protected]
18 Durability Evaluation of Bonded Repairs for the Damaged … 255

7. Petrova, A.P., Lukina, N.F., Sharova, I.A.: Assessment of strength of adhesive joints made with
epoxy adhesives under the influence of different factors. Polym. Sci. Ser. D 7(3), 228–232
(2014)
8. Anikhovskaya, L.I., Batizat, D.V., Petrova, A.P.: The VK-50 elastomeric structural film
adhesive. Polym. Sci. Ser. D 9(3), 251–254 (2016)
9. Lukina, N.F., Kotova, E.V., Chursova, L.V., Kirienko, T.A.: Adhesive binders for layered
alumopolymer composite materials. Polym. Sci. Ser. D 9(4), 392–395 (2016)
10. Petrova, A.P., Sharova, I.A., Lukina, N.F., Buznik, V.M.: The applicability of adhesives in
Arctic conditions. Polym. Sci. Ser. D. 9(2), 188–194 (2016)
11. Oakafor, A.C., Bhogpurapu, H.: Design and analysis of adhesively bonded thick composite
patch repair of corrosion grind out in 2024T3 clad aging aircraft structures. Compos. Struct.
76(1–2), 138–150 (2006)
12. Baker, A.A.: Fatigue life recovery in corroded aluminum alloys using bonded composite
reinforcements. Appl. Compos. Mater. 13(3), 127–146 (2006)
13. Rose, L.R.F.: An application of the inclusion analogy. Int. J. Solids Struct. 17(8), 827–838
(1981)
14. Lee, Y.-G., Zou, W.-N., Ren, H.-H.: Eshelby’s problem of inclusion with arbitrary shape in an
isotropic elastic half-plane. Int. J. Solids Struct. 81, 399–410 (2016)
15. Duong, C.N.: An engineering approach to geometrically nonlinear analysis of a one-sided
composite repair under thermo-mechanical loading. Compos. Struct. 64(1), 13–21 (2004)
16. Eshelby, J.D.: The determination of the elastic field of an ellipsoidal inclusion and related
problems. Proc. Roy. Soc. (London) A241, 376–396 (1957)
17. Beom, H.G.: Analysis of a plate containing an elliptic inclusion with eigencurvatures. Arch.
Appl. Mech. 68(6), 422–432 (1998)
18. Rose, L.R.F.: Theoretical analysis of crack patching. In: Baker, A.A., Jones R. (eds) Bonded
Repair of Aircraft Structure, pp. 77–106. Kluwer Academic Publisher (1988)
19. Joseph, P.F., Erdogan, F.: Plates and shells containing a surface crack under general loading
conditions. NASA Contractor Report 178323 (1987)
20. Joseph, P.F., Erdogan, F.: Surface crack problems in plates. Int. J. Fract. 41(2), 105–131 (1989)
21. Wang, C.H., Rose, L.R.F.: A crack bridging model for bonded plates subjected to tension and
bending. Int. J. Solids Struct. 36(13), 1985–2014 (1999)
22. Hart-Smith, L.J.: Adhesive-bonded single-lap joints. NASA report CR-112236 (1973)
23. Eremin, A.V., Byakov, A.V., Lyubutin, P.S., Panin, S.V.: Development of acoustic-optical
approach for structural health monitoring of composite materials under cyclic loading. Russ.
Phys. J. 59(7–2), 49–54 (2016)
24. Kablov, E.N., Startsev, O., Krotov, A.S., Kirillov, V.N.: Climatic aging of composite aviation
materials: I Aging mechanisms. Russ. Metallurgy (Metally) 2011(10), 993–1000 (2011)
25. Kablov, E.N., Startsev, O., Krotov, A.S., Kirillov, V.N.: Climatic aging of composite aviation
materials: II. Relaxation of the initial structural nonequilibrium and through-thickness gradient
of properties. Russian Metall. (Metally) 2011(10), 1001–1007 (2011)
26. Kablov, E.N., Startsev, O., Krotov, A.S., Kirillov, V.N.: Climatic aging of composite aviation
materials: III Significant aging factors. Russ. Metall. (Metally) 2012(4), 323–329 (2012)
27. Daggumati, S., De Baere, I., Van Paepegem, W., Degrieck, J., Xu, J., Lomov, S.V., Verpoest,
I.: Fatigue and post-fatigue stress–strain analysis of a 5-harness satin weave carbon fibre re-
inforced composite. Compos. Sci. Technol. 74, 20–27 (2013)
28. De Baere, I., Van Paepegem, W., Degrieck, J.: On the design of end tabs for quasistatic and
fatigue testing of fibre-reinforced composites. Polym. Compos. 30(4), 381–390 (2009)
29. FM-73 Epoxy Film Adhesive. Technical Data Sheet AEAD-00019. Cytec Engineering Mate-
rials, https://docplayer.net/37348033-Fm-73-epoxy-film-adhesive.html. Last accessed 10 Aug
2019

[email protected]
Chapter 19
Parametric Identification of Tersoff
Potential for Two-Component Materials

Karine K. Abgaryan and Alexander V. Grevtsev

Abstract The chapter is dedicated to the study of the parametric identification of

Tersoff potential for one-component and two-component materials. The chapter
features a comparison of minimization methods in terms of speed and accuracy,
and the results of the implemented software operation for one-component and two-
component materials. It was shown that implemented software can do the identifi-
cation of two-component materials but with slightly less accurate results. However,
two-component parametric identification requires more time for computation. All
results were compared to the results of the experimental and quantum-mechanical
modeling.

19.1 Introduction

Molecular-Dynamic (MD) modeling is applied when a natural experiment is impos-

sible, very complicated, or highly expensive. The molecular-dynamic approach is
one of the fields of mathematical modeling that is currently rather frequently used in
the problems of material science, which accounts for the relevancy of the work. In
MD modeling, the behavior of the interacting atoms of a system is described within
the framework of classical dynamics [1]. Their location and speed are determined by
means of integration of a system of ordinary differential equations. At the same time,
the forces affecting the atoms are determined by interatomic interaction. Usually, its
description in a system has a rather complicated form, as it can include interactions of

K. K. Abgaryan (B) · A. V. Grevtsev

Federal Research Center “Information and Control” of the RAS, 44/2 Vavilova ul, Moscow
119333, Russian Federation
e-mail: [email protected]
A. V. Grevtsev
e-mail: [email protected]
K. K. Abgaryan
Moscow Aviation Institute (National Research University), 4, Volokolamskoe shosse, Moscow
125993, Russian Federation

© Springer Nature Singapore Pte Ltd. 2020 257

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_19

[email protected]
258 K. K. Abgaryan and A. V. Grevtsev

various types. The main types of interatomic interactions include the ionic, covalent,
metallic, and van-der-Waals interactions. The ion interaction is commonly consid-
ered to be caused by a displacement of valence electrons from one atom to another,
between which the electrostatic attraction occurs. Covalent interaction occurs when
a covalent chemical bond is formed. At the same time, consolidation and concen-
tration of the electron pair occur on the molecular orbital. A metallic bond occurs
when shared electron gas interacts with the ion core of a crystalline structure. Van-
der-Waals interactions are applied in order to describe intermolecular interaction, in
a series of organic compounds, etc.
The chapter considers one- and two-component compounds with the covalent
type of chemical bonding described using Tersoff potential, which was proposed in
[2] and considered for one-component materials. Parameters of Tersoff potential for
silica, which is a one-component material, were calculated in [3].
In this chapter, Tersoff potential was modified for use with two-component mate-
rials. Each potential has a certain set of parameters, the values of which are unique
for each material. The process of finding such sets represents the parametric iden-
tification of the potential. It should be understood that the problem of parametric
identification is multiextremal, and, hence, it is necessary to find the global min-
imum. The work presents a comparison of Monte Carlo and simulated annealing
methods for global minimization and Hooke–Jeeves and Radial Granular Search
(RGS) methods for local minimization.
Software for parametric identification of certain materials was used to perform
calculations shown in the chapter. The software uses parallel calculations in order to
reduce the identification time. Various parallelization methods are compared.
Section 19.2 includes the statement of problem of parametric identification of
interatomic potential. Section 19.3 presents the comparison of optimization meth-
ods used for parametric identification. Section 19.4 describes the computational
results for different materials. Section 19.5 includes a description of the implemented
software. Section 19.6 presents the conclusion of the chapter.

19.2 Problem Statement

Calculation of the total energy of the modeled material atom system is carried out
using Eq. 19.1.

1   
n n
U = V rij (19.1)
2 i=0 j=0

In Eq. 19.1, rij is the distance between atoms i and j. The interaction energy V (rij )
is calculated using Eq. 19.2.
       
V rij = fc rij VR rij − bij VA rij (19.2)

[email protected]
19 Parametric Identification of Tersoff Potential … 259
 
Parameter VR rij is the atom j to atom i repulsion potential, which is calculated
by Eq. 19.3.
 
  De √  
VR rij = exp −β 2S rij − re (19.3)
S −1
 
Parameter VA rij is the atom j to atom i attraction potential, which is calculated
using Eq. 19.4.
 
  SDe 2 
VA rij = exp −β rij − re (19.4)
S −1 S

Here, fc is the function of cutoff around the ith atom, which is continuous and dif-
ferentiable for all rij . This function allows to reduce a complexity of the calculations
and is described by Eq. 19.5.
⎧
⎪
⎨  1  if r < (R − Rcut )
π(r−R)
fc (r) = 1
1 − sin if R − Rcut < r < R + Rcut (19.5)
⎪
⎩
2 2Rcut
0 if r > (R + Rcut )

Equation 19.5 has two parameters. Parameter R is the distance parameter, which
is measured in angstroms. The cutoff function has the limits of 0 and 1 at the distance
of Rcut from the value of R in the positive and the negative directions. Parameter bij
is the order of bond between atoms i and j, which depends on the bond angle of atom
i and the nuclear environment. The bond order is described according to Eq. 19.6.
  n − 2n1
bij = 1 + γ ζij (19.6)

Parameter ζij provides weighted calculations for bonds other than i-j. Calculations
are carried out for each kth atom using Eq. 19.7.


n
   
ζij = fc (rik )g θjik ωijk (19.7)
k=i,j

Parameter ωijk provides a calculation of the efficiency of a coordinate system,

which is affected by the distances of various neighboring atoms. Parameter ωijk can
be calculated using Eq. 19.8.
  3 
ωijk = exp λ3 rij − rik (19.8)
 
Parameter g θjik defines the dependence on the bond angle of atom θjik with
vertex in ith atom, which can be calculated using Eq. 19.9.

[email protected]
260 K. K. Abgaryan and A. V. Grevtsev

  c 2 c2
g θjik = 1 + −    (19.9)
d d 2 + h − cos θjik

This calculation allows to stabilize the geometry of crystalline grid atoms in

shearing operations.
The parametric identification of Tersoff potential requires the determination of
the parameter vector ξ = (ξ1 , . . . , ξk ) ∈ X , X ⊆ Rk , at which the lowest value of
the minimization function (Eq. 19.10) is achieved.

(fn (ξ ))2 
n
(f1 (ξ ))2
F(ξ ) = ω1 + · · · + ω , ωi = 1 (19.10)
f˙ 2 f˙ 2
n
1 n i=1

In Eq. 19.10, ωi are the weighting coefficients, the sum of which equals to 1, f˙n is
the reference value of the ith studied material property, fi (ξ ) is the value of the ith
studied material property, which is obtained using the specified set of parameters.
The weight factors are usually calculated as ωi = 1n , where n is the number of stud-
ied material properties. Additionally, the squares of the difference between the refer-
ence and calculated values of the characteristics are normalized in the minimization
function. It allows to obtain more adequate values of the function.
The problem of parametric identification can be formally described as shown in
Eq. 19.11.

ξ = arg min F(ξ ) (19.11)

ξ ∈X

It should be noted that there is a vast number of local minima in these types of
problems. It is very important that the determined minimum is the global minimum.
For this, it is important to ensure the use of minimization method that allows to
find the global minimum among the local minima. An increase in the number of
potential parameters results in a significant increase in the difficulty of finding a
global minimum.
There is a number of minimization methods, which may resolve these types of
problems, for example, the simulated annealing or Monte Carlo methods. Each
of these methods has a different way of treating the local minima. It is impor-
tant to understand that such methods have a lower accuracy compared to the local
minimization methods.
In a parametric minimization problem, methods allow to find a global minimum
with low accuracy should be used first. This provides a possibility to find an area,
where the solution is located, faster. Then, high accuracy local minimization methods
should be used.

[email protected]
19 Parametric Identification of Tersoff Potential … 261

19.3 Comparison of Optimization Methods

Various minimization methods should be compared based in terms of the parameter-

ization time and value of the minimization function. Global minimization methods
should be compared first.
Monte Carlo method and simulated annealing method are rather different in terms
of their essence. In the first method, numerous sets of parameters are randomly
calculated, and the best one of the sets is selected. The accuracy of the result is directly
dependent on the number of generated sets. The simulated annealing method has a
limited number of iterations, which depends on the temperature reduction function.
In the implemented software, the number of iterations of the simulated annealing
method is 112. In order to correctly compare the methods in terms of accuracy and
time, several runs of each method are required, and the total execution time and the
obtained value of the minimization function should be considered.
For comparison, each method was run ten times. Monte Carlo method generates
1,120 sets, and the simulated annealing method includes a total of 1,120 iterations.
Table 19.1 demonstrates that Monte Carlo method is approximately 3.5 times faster.
Table 19.2 demonstrates that the behavior of the simulated annealing method
is inferior to that of Monte Carlo method. The average value of the minimization
function of the simulated annealing method is 0.0839228511, while the average value
of the minimization function of Monte Carlo method is 0.0496128700.
The local minimization methods are compared using the resulting sets of param-
eters obtained under global minimization provided by Monte Carlo method, as this

Table 19.1 Comparison of

Method Time, ms
the time required for global
minimization methods Monte Carlo 66,981
Simulated annealing 238,210

Table 19.2 Comparison by

Set Monte Carlo method, F(ξ ) Simulated annealing
the final value of the
method, F(ξ )
minimization function
1 0.0206441 0.185270050549
2 0.026144 0.056386588562
3 0.040383 0.108290095692
4 0.0462078 0.065527332121
5 0.0512995 0.083363670061
6 0.0535324 0.077168913690
7 0.0585524 0.131231663262
8 0.0653591 0.042978009482
9 0.0654594 0.025928630353
10 0.068547 0.063083557049

[email protected]
262 K. K. Abgaryan and A. V. Grevtsev

Table 19.3 Comparison of

Method Time, ms
the time required for local
minimization methods RGS 12,508
Hooke–Jeeves 26,516

Table 19.4 Comparison by

Set RGS method, F(ξ ) Hooke–Jeeves method, F(ξ )
the final value of the
minimization function 1 0.0206441 0.185270050549
2 0.026144 0.056386588562
3 0.040383 0.108290095692
4 0.0462078 0.065527332121
5 0.0512995 0.083363670061
6 0.0535324 0.077168913690
7 0.0585524 0.131231663262
8 0.0653591 0.042978009482
9 0.0654594 0.025928630353
10 0.068547 0.063083557049

method proved itself to be the better of two tested methods. RGS method and Hooke–
Jeeves method are compared. Table 19.3 demonstrates that RGS method is almost
two times faster than Hooke–Jeeves method.
Table 19.4 demonstrates that the behavior of Hooke–Jeeves method is inferior in
terms of the obtained value of the minimization function. The average value of the
minimization function using RGS method is 0.0003019762, while the average value
of the minimization function using Hooke–Jeeves method is 0.0006873606.

19.4 Results

The computational results for silicon, germanium, aluminum nitride, and boron
nitride are represented in Sects. 19.4.1–19.4.4, respectively.

19.4.1 Silicon

Silicon is the most commonly used material in semiconductor devices. Its oxide is
rather easily obtained in furnaces with the formation of semiconductor interfaces.
Silica has a crystalline structure of a diamond. The valence band in the crystal is
completely filled. Silicon retains operation stability at high temperatures. Planetary
silica reserves are almost limitless.

[email protected]
19 Parametric Identification of Tersoff Potential … 263

Table 19.5 Identified

Parameter Value
parameters of silicon
De 5.14592
S 4.69548
β 1.44156
re 2.23518
R 2.85
Rcut 0.15
c 24,697.3
d 103.132
h −0.273259
n 1.21149
γ 0.953207
λ 0.00156529

For this material, identification of the potential parameters was carried out using
Monte Carlo method. Eight sets with the lowest value of the minimization function
are selected from the four thousand randomly generated sets. Then, the selected
parameters were refined using RGS method. The calculated parameters are shown
in Table 19.5.
The reference values of the properties were calculated on a computer using VASP
software package. The experimental data was taken from [1]. The characteristics
calculation results with the given potential parameters are provided in Table 19.6.

Table 19.6 Characteristics for silicon

Characteristic Experiment VASP Identification
Ecoh −4.63 −4.617291375 −4.59422
a 5.431 5.465408871 5.58357
B 0.9783 0.8936352 0.914167
C 0.509 0.46909605 0.52863
C11 1.657 1.5190966 1.61829
C12 0.639 0.5809045 0.561208
C44 0.796 0.6277141 0.624638
ζ 0.524 0.522770577 0.517098

[email protected]
264 K. K. Abgaryan and A. V. Grevtsev

19.4.2 Germanium

Germanium is the 32nd element in the periodic table. Under normal conditions,
the crystalline grid takes the shape of a diamond and represents a semiconductor
material.
Germanium was used in microelectronics for the manufacture of transistors and
diodes, until it was replaced by silica. Nevertheless, germanium is still used. Accord-
ing to certain musicians, the use of germanium-based transistors results in a better
sound. Currently, germanium is mainly used in UHF devices as an element of the
silicon and germanium alloy. This alloy makes it possible to achieve subterahertz
frequencies.
With the parameters shown in Table 19.7, the obtained properties deviate by no
more than one-tenth from the experimental values. The experimental data was taken
from [2]. The characteristics given in Table 19.8 were rather close to those obtained
in the VASP software package.

Table 19.7 Identified parameters of germanium

Parameter Value
De 5.60702
S 6.09551
β 1.4484
re 2.24235
R 2.95
Rcut 0.15
c 33,739
d 125.481
h −0.524262
n 1.57274
γ 1.6957
λ 17.5114

Table 19.8 Characteristics

Characteristic Experiment VASP Identification
for germanium
Ecoh −3.85 −3.78168 −3.77807
a 5.658 5.645 5.6753
B 0.7516 0.745452 0.746909
C 0.403 0.3469163 0.348045
C11 1.2889 1.208007 1.2102
C12 0.4829 0.514174 0.514251
C44 0.671 0.604663 0.602228
ζ 0.521 0.561069 0.562381

[email protected]
19 Parametric Identification of Tersoff Potential … 265

19.4.3 Aluminum Nitride

Aluminum nitride is mainly used with a crystalline grid of a wurtzite shape under
standard temperature and pressure. Aluminum nitride of a sphalerite shape is also
interesting in terms of its optical and physical properties. This type of material is
very difficult to manufacture, as it is very reactive and requires a special sterility of
raw materials.
This is a two-component material. The number of parameters depends on the
number of interactions between various elements. There are three types of interac-
tions for the two elements. Aluminum interacts with aluminum, aluminum interacts
with nitrogen, and nitrogen interacts with nitrogen.
Thus, there are three times more parameters. Table 19.9 demonstrates all 3 sets
of 12 parameters. R and Rcut for each interaction are not identified, but specified at
the initial stage.
The obtained characteristics given in Table 19.10 are close to those obtained using
VASP software package.

Table 19.9. Identified

Parameter Al-Al value Al-N value N-N value
parameters of aluminum
nitride De 3.35512 3.91278 2.71983
S 0.835993 1.09552 3.10157
β 1.05919 0.509301 1.5406
re 2.13729 2.57726 3.19281
R 2.335 2.335 2.335
Rcut 0.8 0.8 0.8
c 85871.1 3201.3 746,383
d 158.496 31.1845 30.3546
h −0.273204 −0.658417 7.15848
n 20.9305 5.40596 6.0999
γ 0.0416951 0.0623046 0.0649637
λ 1.05287 1.48318 0.585351

Table 19.10 Characteristics

Characteristic Experiment VASP Identification
for germanium
Ecoh −5.76 −5.75281 −5.73307
a 4.38 4.38 4.22682
B 2.08 2.22859 2.24682
C 0.72 0.791064 0.794709
C11 3.04 3.283342 3.3049
C12 1.60 1.701214 1.71347
C44 1.93 1.909453 1.8594
ζ 0.55 0.640232 0.627198

[email protected]
266 K. K. Abgaryan and A. V. Grevtsev

The experimental data was taken from [4]. It should be noted that the deviation of
VASP characteristics from the experimental values is rather substantial. It should be
noted that the number of identifiable parameters increased three-fold. Due to that, a
higher deviation of characteristics may be observed.

19.4.4 Boron Nitride

Boron nitride with a sphalerite-type crystalline grid is a stable structure under normal
conditions. This material is characterized by exceptional physical properties, excel-
lent strength, and chemical inertness. A wide bandgap, high melting temperature,
and low dielectric constant make boron nitride a very good material for microelec-
tronic devices. The use of Tersoff potential for molecular-dynamic modeling of boron
nitride is described in [5, 6].
Boron nitride is mainly used as a boundary layer in the growing of GaN on SiC.
This material is of great interest for nanotube growing. The optical properties and
high transparency make boron nitride a useful material for optical windows and X-ray
diaphragms.
At the end of the identification, potential parameters became equal to the values
given in Table 19.11. The characteristics for such values are shown in Table 19.12.
The experimental data was taken from [4].

Table 19.11 Identified

Parameter B-B value B-N value N-N value
parameters of boron nitride
De 47.3631 5.99477 2.29434
S 1.9582 6.03235 3.69895
β 0.0337772 1.26559 1.44987
re 2.5521 1.6491 2.97387
R 1.95 1.95 1.95
Rcut 0.75 0.75 0.75
c 2579.2 34907.1 787946
d 160.381 102.02 26.1181
h −0.388802 −0.615747 −5.13706
n 53.5382 64.6551 6.47006
γ 0.391766 0.402921 0.0875969
λ 0.217494 0.0118349 0.182635

[email protected]
19 Parametric Identification of Tersoff Potential … 267

Table 19.12 Characteristics for boron nitride

Characteristic Experiment VASP Identification
Ecoh −6.68 −6.728191125 −6.56826
a 3.6157 3.6157 3.79437
B 4.00 4.206769067 4.10621
C 3.15 3.3215539 3.3132
C11 8.20 8.6355076 8.63461
C12 1.90 1.9923998 2.00906
C44 4.80 4.7791548 4.87595
ζ – 0.381397288 0.382457

19.5 Description of Software

The chapter features the results of calculations performed on the developed software,
which is designed for the parametric identification of Tersoff potential parameters for
various materials. The programming language applied was C++. OpenMP package
was used for parallel calculations. The implemented software can be installed on
Windows or Linux operating systems.
For clarity of the results, we present the parameters of the processor of the sys-
tem, on which the calculations were performed: AMD Ryzen 7 1800X Eight-Core
Processor 3.60 GHz.

19.6 Conclusions

The chapter presents a study of the molecular-dynamic modeling stage referred

to as the parametric identification of potential. A software product for parametric
identification was implemented. Tersoff potential was modified, so that it could be
used for two-component materials. Due to modification, the number of potential
parameters increased three-fold, which in turn led to the decline of the identification
accuracy.
Monte Carlo method and simulated annealing method were compared for global
minimization. As demonstrated by the results, Monte Carlo method is more accurate
and faster than simulated annealing method. Hooke–Jeeves method and RGS method
were compared for local minimization. Based on the provided results, RGS method
is better.

[email protected]
268 K. K. Abgaryan and A. V. Grevtsev

The work also features the results of implemented software operation. The pro-
gram is suitable for both one-component and two-component materials. The char-
acteristics calculated with the obtained sets of parameters for one-component and
two-component materials are the same or feature a minor deviation from the char-
acteristics calculated in VASP software package. It implies that the parameters
were correctly identified.

References

1. Abgaryan, K.K.: Multiscale Modeling in the Problems of Structural Materials Science. MAKS
Press (in Russian) (2017)
2. Powell, D.: Lattice dynamics and parameterisation techniques for the Tersoff potential applied
to elemental and type III-V semiconductors. Diss. for the degree of Doctor of Philosophy, The
University of Sheffield (2006)
3. Abgaryan, K.K., Posypkin, M.A.: Optimization methods as applied to parametric identification
of interatomic potentials. Comput. Math. Math. Phys. 54(12), 1929–1935 (2014)
4. Shimada, K., Sota, T., Suzuki, K.: First-principles study on electronic and elastic properties of
BN, AlN, and GaN. J. Appl. Phys. 84(9), 4951–4958 (1998)
5. Los, J.H., J. Kroes, M.H., Albe, K., Gordillo, R.M., Katsnelson, M.I., Fasolino, A.: Extended
Tersoff potential for boron nitride: energetics and elastic properties of pristine and defective
h-BN. Phys. Rev. B 96(18), 184108.1–184108.11 (2017)
6. Albe, K., Möller, W.: Modelling of boron nitride: atomic scale simulations on thin film growth.
Comput. Mater. Sci. 10(1–4), 111–115 (1998)

[email protected]
 Part IV
Numerical Study of Dynamic Systems

[email protected]
Chapter 20
Multi-agent Optimization Algorithms
for a Single Class of Optimal
Deterministic Control Systems

Andrei V. Panteleev and Maria Magdalina S. Karane

Abstract The algorithms and software of three metaheuristic multi-agent methods:

fish school search, krill herd, and imperialist competitive algorithm are considered.
Recommendations on the parameter selection for each method are given. On the
basis of krill herd and imperialist competitive algorithm, a hybrid extremum search
algorithm is formulated. An algorithm for finding open-loop control for a single
class of dynamic systems based on the use of the described multi-agent algorithms is
also formed. Optimal open-loop control has the form of a step function with a given
switching set. Software that allows to find the optimal open-loop control, criterion
value, and coordinates of switching points of the control law on the basis of the
suggested algorithms was formed. A specially selected set of test open-loop optimal
control problems was solved. The obtained results confirmed that the numerical
solution is close to the optimal one.

20.1 Introduction

The development of multi-agent optimization algorithms has been going on for many
years in order to improve the existing classical optimization methods that do not
always give a solution that is close to optimal or even replaces them with new
algorithms. A feature of multi-agent algorithms is the use of a group of individ-
uals (agents) on a certain set that performs certain actions (operations) in order to
reach an extreme point. Using this approach, we can optimize not only multiextremal
functions of many variables, but also find a solution for optimal open-loop control
problems [1, 2], which is also important due to the practical importance of such
problems in aviation and space technology.

A. V. Panteleev (B) · M. M. S. Karane

Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
M. M. S. Karane
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 271

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_20

[email protected]
272 A. V. Panteleev and M. M. S. Karane

The Chapter is organized as follows. Section 20.2 provides a description of multi-

agent methods. Application of multi-agent methods for optimal open-loop control
problems is discussed in Sect. 20.3. Section 20.4 concludes the chapter.

20.2 Description of Multi-agent Methods

Multi-agent methods for solving the optimization problem are described in this
section. Section 20.2.1 provides an optimization problem. Fish school search algo-
rithm, krill herd algorithm, imperialist competitive algorithm, and hybrid multi-agent
algorithm are introduced in Sects. 20.2.2–20.2.5, respectively.

20.2.1 Optimization Problem

Given the objective function f (x) = f (x1 , x2 , . . . , xn ) defined on the set of admissible
solutions D ⊆ Rn . It is required to find the constrained global maximum of a function
f (x) on set D, i.e. ,such a point x∗ ∈ D, that

f (x∗) = max f (x), (20.1)

x∈D

where x = (x1 , x2 , . . . , xn )T , D = { x|xi ∈ [ai , bi ], i = 1, 2, . . . , n }.

The task of finding the minimum of a function f (x) is reduced to the task of
finding the maximum by replacing the sign before the function with the opposite:
f (x∗) = min f (x) = −max[ − f (x)]. Function f (x) can be multiextremal, so the
x∈D x∈D
required solution in the general case is not unique.

20.2.2 Fish School Search Algorithm

Solution search strategy. Fish school search algorithm [3–5] uses the results of
studying the behavioral features of certain fish species that can exist only within the
flock, which reduces the individual freedom of their movements, but increases the
intensity of the competition for food. Such a union of fish, as shown by observing
them in the oceans and rivers, confirms that the benefits greatly exceed the drawbacks.
The algorithm uses the following main features of the behavior of schools of fish:
• Food (imitation of the natural instinct of fish consisting in the search for food).
It is necessary because fish must be fed in order to grow strong and be capable
of reproduction. When they get food, the fish gain weight, and, when they swim,
they lose weight.

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 273

• Swimming (this function is implemented collectively by all fish in flock to search

for food).
• Reproduction (imitation of a natural selection mechanism creating new objects to
support the search process).
Each fish from the flock has an internal “memory” of the success in the search
for food (approaching the extreme point) enclosed in the weight of the fish. A flock
evolves by exchanging information between parents as a result of reproduction and
also as a result of collective movement. The aquarium is a set of acceptable solutions.
The presence of food shows the fish of the aquarium area that defines good regions
for finding a solution. In the process of swimming, the idea of a global redirection
of all fish to that part of the aquarium, which is considered by all flock fish as the
most preferable from the point of view of food search, is realized. Reproduction of
fish, in turn, allows to move from a comparative study of the areas of the aquarium
to a process that refines the solution within the framework of the found area. For the
process of reproduction, fish with the greatest weight are only allowed. In the process
of finding solutions, an initial population NP of fish is first generated—a school on
the set D using a uniform distribution.
Next, fish when searching for food (determining the extremum point) use Food
operation, Swimming operation, and Reproduction operation discussed below.
Food operation. Fish are looking for food that is scattered in the aquarium in
various concentrations. To search for food, fish perform individual movements. As
a result, the fish may gain weight or vice versa depending on the search result. It is
assumed that the increment of fish weight is proportional to the normalized increment
of the value of the objective function:

f (xj,k+1 ) − f (xj,k )
W j,k+1 = W j,k +   , (20.2)
max f (xm,k+1 ) − f (xm,k )
m=1,...,NP

where W j,k+1 , W j,k are the fish weights with number j at the (k + 1)th and kth itera-
tions, respectively, f (xj,k+1 ), f (xj,k ) are the objective function values corresponding
to the new xj,k+1 and current xj,k fish position, respectively. The weight of the fish
usually varies from 1 to Wscale . For k = 0, the weight of all fish is the same and
equals Wscale
2
.
Swimming operation.
(a) Each fish moves in a random direction so that the value of the objective function
increases. If a new position of the fish is not included in the set D (aquarium),
then the movement does not occur. After completing the movement, the fish
feeds, the new position of each fish, j = 1, . . . , NP, is calculated according to
Eq. 20.3, where δ ∼ R[−0.5, 0.5] and stepind k
decrease linearly with the increase
of iterations number (an individual step is a vector with the same coordinates).

[email protected]
274 A. V. Panteleev and M. M. S. Karane

xj,k+1 = xj,k + stepind

k
·δ (20.3)

New position is accepted if value of the objective function increases (when the
restriction on belonging to the set D is fulfilled). Otherwise, the fish does not move.

(b) After each fish has moved, the weighted average of the individual shifts is
calculated by Eq. 20.4, where j is the fish number.

j
xind = xj,k+1 − xj,k (20.4)

It means that the fish that made a successful movement determines the resulting
direction of movement more than others. Each fish in the school changes its position
according to Eq. 20.5.


NP  
xind
m
· f (xm,k+1 ) − f (xm,k )
m=1
xj,k+1 = xj,k + NP 
j = 1, . . . , NP (20.5)
 
f (xm,k+1 ) −f (xm,k )
m=1

(c) Collective movement is based on the movement of the whole flock. If the
flock increased the weight in case of a successful move, the radius of the flock
increases. Otherwise, it decreases. The position of each fish varies with respect
to the center of gravity of the flock (barycenter) determined at each iteration k:


NP
xj,k · W j,k
j=1
Bari(k) = . (20.6)

NP
W j,k
j=1

Besides, the increment vector stepvol (vector with the same coordinates linearly
decreasing with increasing number of iterations) is used.
If the total weight of the flock has increased, then the fish move toward the
barycenter:
 j,k 
x − Bari(k)
x j,k+1
=x j,k
− stepvol · rand ·  j,k . (20.7)
x − Bari(k)

If the total weight of the flock has decreased, then the fish move away from the
barycenter:

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 275
 j,k 
x − Bari(k)
x j,k+1
=x j,k
+ stepvol 
· rand · j,k , (20.8)
x − Bari(k)

where rand ∼ R[0, 1].

Reproduction operation. Fishes that have reached the threshold weight (W i,k >
thr) are selected. For each selected fish, a pair is sought that is, a fish, which corre-
sponds to the maximum ratio of its weight to the distance to the applicant. Each pair
of fish with numbers i and j generates a descendant with weight and position:

W i,k + W j,k
W child ,k = , (20.9)
2
xi,k + xj,k
xchild ,k = . (20.10)
2
Next, parents and descendants are ranked by weight. To preserve the size of the
population, all fish that have the lowest weight (the population with size NP remains
in the flock) are removed. The process ends, when a certain number of iterations
ITER is reached. The solution is to choose the position of the fish with the highest
weight.
Recommendations on the selection of parameters. The population size NP
determines the number of calculations of the objective function at each iteration. For
a problem with a large range of admissible solutions, it is recommended to take a
larger parameter value NP. Recommended parameter values NP ∈ [30, 40].
The number of iterations ITER determines how long the search for new solutions
will continue. As a rule, the more ITER, the more accurate the answer. For a standard
set of functions, the recommended values depend on the complexity of the function
ITER ∈ [1000, 10000].
The maximum weight Wscale is that an individual can gain. With the passing
of the search time, the fish with the maximum weight Wscale is selected, and its
position is taken as the answer. The recommended value of this parameter Wscale ∈
[1000, 5000]. The threshold weight W determines those fish that will be allowed to
reproduce. Recommended value W ∈ [900, 4500].
Individual step stepind 0 is a vector with the same components, which decreases
linearly with increasing number of iterations to the value stepind 1 , stepvol is the
vector with the same coordinates linearly decreasing with increasing number of
iterations. Recommended values are the following: stepind 0 = 0.1, stepind 1 = 0.01,
and stepvol = 0.001.

20.2.3 Krill Herd Algorithm

Solution search strategy. Krill herd method [3, 6, 7] refers to bioinspiration because
it is based on the results of the analysis of the behavior of krill packs resembling
shrimps. Their positions change under the influence of three factors: the presence

[email protected]
276 A. V. Panteleev and M. M. S. Karane

of other members of the population, need to search for food, and random walks.
Usually. the movement of krill population is determined by two goals: the increase
in the density of krill and attainment of food.
At the beginning of the process, a population NP is generated from individuals
on a set of admissible solutions D using a uniform distribution. It is assumed that the
motion of the jth member of the population occurs according to Eq. 20.11, where xj
is the position, V j is the speed, which consists of three components.

dxj
= Vj (20.11)
dt
The first component is determined by the influence of neighbors (members of the
population that belong to a certain neighborhood of jth element of a certain radius),
the best element in the entire population, and information about its former speed.
The second component is determined by the movement toward the food source (the
“center of mass” of the population is taken for it), information about the former
speed in search of food, and memory of its best result for all the iterations. The third
component imitates the random walks of the individual, which decreases with the
increasing number of iterations. To revive the search process, the cross and mutation
operations used in other evolutionary methods and method of differential evolution
are applied. The search procedure ends when the specified number of iterations is
reached.
Solution search algorithm. Solution search algorithm includes the following
steps.
Step 1. Set the method parameters: NP is the number of krill in the population, Smax
is the maximum krill speed, μ is the small positive number, Imax is the maximum
number of iterations, Vf is the maximum speed of movement to the food source,
Dmax is the maximum diffusion speed of krill. Let I = 1 (iteration count).
Step 2. Generate the initial krill population on a set D using the uniform distribution
law: x1 , . . . , xNP . Calculate the values of the objective function f (x1 ), . . . , f (xNP ).
Find the best and worst solutions xbest , xworst .
Step 3. For each element of the population, j = 1, . . . , NP, perform the following
steps.
Step 3.1. Find the radius of the neighborhood around the solution xj :

NP
1
dεj = dj,k , (20.12)
5 · NP
k=1


n
j
2
where dj,k = xi − xik . Determine the number of neighbors Sj of the
i=1
solution xj from the condition dj,k ≤ dεj .

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 277

Step 3.2. Find:

xk − xj
x̄j,k = , k = 1, . . . , Sj (for all neighbors), (20.13)
dj,k + μ
Nj
f (xj ) − f (xk )
f¯ j,k = , k = 1, . . . , Nj , αjlocal = f¯ j,k · x̄j,k .
f (xwrost ) − f (xbest )
k=1
(20.14)

Step 3.3. Find:


I f (xj ) − f (xbest )
cbest = 2 · rand + , rand ∼ U [0; 1], f¯ j,best = ,
Imax f (xwrost ) − f (xbest )
(20.15)
xbest − xj
= cbest · f¯ j,best · x̄j,best .
target
x̄j,best = ,α (20.16)
dj,best + μ j

Step 3.4. Find:

target
α j = αjlocal + αj . (20.17)

Step 3.5. Define:

S j,New = Smax · α j + ω · S j,old , (20.18)

where ω ∼ U [0; 1], S j,old is the old speed generated by the other members of the
population (under I = 0, the speed is assumed to be equal to the zero vector).
Step 4. For any element of the population, j = 1, . . . , NP, perform the following
steps.
Step 4.1. Find the position of the food source:


NP
xj
f (xj )
j=1
xfood = . (20.19)

NP
1
f (xj )
j=1

Step 4.2. Find:

f (xj ) − f (xfood ) xfood − xj

f¯ j,food = , x̄ j,food
= , (20.20)
f (xwrost ) − f (xfood ) dj,food + μ

[email protected]
278 A. V. Panteleev and M. M. S. Karane

I
= cfood · f¯ j,food · x̄j,food .
food
cfood = 2 · 1 − , βj (20.21)
Imax

Step 4.3. Find:

f (xj ) − f (xjbest )
f¯ j,jbest = , (20.22)
f (xwrost ) − f (xjbest )

where xjbest is the best position of jth population element,

xjbest − xj
x̄j,jbest = , β best = f¯ j,jbest · x̄j,jbest . (20.23)
dj,jbest + μ j

Step 4.4. Find:

food
β j = βj + βjbest . (20.24)

Step 4.5. Define:

F j,New = Vf · β j + ωf · F j,old , (20.25)

where Vf is the maximum speed of movement to the food source, ωf ∼ ∪[0, 1],
F j,old is the old speed generated by the movement to the food source (under I = 0,
the speed is assumed to be equal to the zero vector).
Step 5. Define:

I
Dj = Dmax · 1 − · δ, (20.26)
Imax

where δ is n-dimensional vector with components δi ∼ U [−1, 1].

Step 6. For any j = 1, . . . , NP define:

V j = S j,New + F j,New + Dj . (20.27)

Step 7. For any j = 1, . . . , NP define:

xj,New = xj,old + V j · t, (20.28)


∞
j,New
where t = ci · (bi − ai ), ci is the number on the interval [0, 2]. If xi ∈
/
i=1
j,New
[ai , bi ], then let
xi = ai + χ · [bi − ai ], χ ∼ U [0, 1].
Step 8. Crossbreeding. For any j = 1, . . . , NP perform the following steps:
Step 8.1. Find:

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 279

Cr = 0.2 · f¯ j,best . (20.29)

Step 8.2. Define:


j,New
j,Cr xr if randi < Cr
xi = j,New r ∈ {1, . . . , j − 1, j + 1, . . . , n}, i = 1, . . . , n,
xi overwise
(20.30)

where r is the random integer from set {1, . . . , j − 1, j + 1, . . . , n}.

As a result, solutions x1,Cr , . . . , xNP,Cr are found.
Step 9. Mutation. For any j = 1, . . . , NP perform.
Step 9.1. Find:

Mu = 0.05 · f¯ j,best . (20.31)

Step 9.2. Define:

  p q
j,Mu xibest + v xi + xi if randi < Mu
xi = j,Cr i = 1, . . . , n, (20.32)
xi else

Step 9.3. Let xj = xj,Mu , j = 1, . . . , NP.

j,Mu j,Mu
where v ∼ U [0, 1]. If xi ∈
/ [ai , bi ], then let xi = ai + χ · [bi − ai ].
Step 10. Calculate the values of the objective function. Find the best and worst
solutions xbest , xworst . For each solution xj , find the best position xjbest for all past
iterations.
Step 11. Check fulfillment of termination condition. If I < Imax , then let I = I + 1
and go to Step 3. If I = Imax , then process is complete. As an approximate solution
of the problem, select xbest , f (xbest ).
Recommendations on the selection of parameters. The population size NP
determines the number of calculations of the objective function at each iteration. For
a problem with a large range of admissible solutions, it is recommended to take a
larger parameter value NP. Recommended parameter values NP ∈ [40, 50].
The number of iterations Imax determines how long the search for new solutions
will continue. As a rule, the more Imax , the more accurate the answer. For a standard
set of functions, the recommended values depend on the complexity of the function,
i.e., Imax ∈ [1000, 10,000].
The maximum krill speed Smax is used to determine the speed of each member
of the flock (see Step 3.5). The recommended value of this parameter Smax = 0.01.
Small positive number μ corrects the change in the position of thejth member of the
population (see Steps 3.2, 3.3, 4.3). The recommended value μ = 0.3.
The maximum speed of movement to the food source Vf is used to determine
the speed of movement to the food of each member of the flock (see Step 4.5). The

[email protected]
280 A. V. Panteleev and M. M. S. Karane

recommended value of this parameter Vf = 0.02. The maximum diffusion speed of

krill is Dmax . The recommended value of this parameter Dmax ∈ [0.002, 0.01].

20.2.4 Imperialist Competitive Algorithm

Solution search strategy. The strategy of imperialist competitive algorithm [3, 8,

9] uses observations of the behavior of empires in the fight for spheres of influ-
ence. Imperialism is a policy of expanding the government’s administrative influence
beyond the borders of the country, which is realized both through direct manage-
ment and indirectly through influence on the markets of food, goods, materials, etc.
Thus, all countries are divided into empire and colony. The empires seek to use the
resources of other countries or simply to influence their policies by opposing other
empires. Regardless of the motivating reasons, empires seek to increase the number
of their colonies and extend their influence to the whole world.
The method uses ideas, both evolutionary algorithms, and “swarm intelligence”
methods. It begins with the formation of the initial population—countries in the world
(solutions on a set of admissible solutions). Some of the best countries (by the size
of the objective function) are selected for the role of the imperialist countries, while
the rest form colonies. All colonies are assigned to the imperialist states, and their
number is determined by the strength of such a state, inversely proportional to the
value of the objective function. This is how empires are formed: the imperialist state
and its colonies. The largest number of colonies corresponds to the most powerful
imperialist state. Then each colony begins to move toward its imperialist state. The
strength of the empire is determined by the strength of the imperialist state and its
colonies (the share of the average strength of the colonies is added to the strength of
the state). The competition between empires leads either to an increase (at least, to
non-decreasing) of the strength of the empire or to a decrease in it. Weak empires
disappear with time. The described mechanisms should lead to a situation, where
there is only one empire in the world, and all other countries are its colonies (this is
the condition for the end of the process). The position of the imperialist state is taken
as an approximate solution of the problem.
Solution search algorithm. Solution search algorithm involves the following
steps.
Step 1. Set the method parameters: the size of population (the number of countries)
Npop , the number of imperialist countries Nimp , parameters of colony shift β, γ , and
colony influence parameter ξ .
Step 2. Generate Npop countries (solutions from set D) using uniform distribution
low. Calculate the value of objective function and order solutions (countries) by
increasing the objective function: x1 , x2 , . . . , xNpop (the solution x1 corresponds to
the smallest value of the objective function).
Step 3. Choose imperialist states. Select Nimp solutions (countries) from among the
first in the list of solutions. Nimp corresponds to the best values of the objective

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 281

function. Further, they will be called imperialistic. Count the number of colonies
Ncol = Npop − Nimp .
Step 4. Form empires.
Step 4.1. Calculate the normalized value of each imperialist state:

f˜ (ximp ) = f (ximp ) − f (xNpop ), j = 1, . . . , Nimp .

j j
(20.33)

Step 4.2. Find the normalized strength of each imperialist state:

 
 
 
 ˜ j 
 f (x ) 

P = N
j imp , j = 1, . . . , Nimp . (20.34)

 ˜ j 
imp
 f (ximp ) 
 j=1 

Step 4.3. Find the number of colonies of each imperialist state:

Nimp −1
  N
Ncj = P j · Ncol , j = 1, . . . , Nimp − 1, Nc imp = Ncol − Ncj , (20.35)
round
j=1

where round [·] is the round-off operation.

j
Step 4.4. For any imperialist state, select Nc countries randomly from among the
colonies. The imperialist state and chosen colonies form an empire.
Step 5. The shift of the colonies of the empire to the imperialist state (assimilation
procedure). For any empire, consistently fulfill, j = 1, . . . , Nimp .
Step 5.1. Choose the first colony in the empire with number j randomly.
Step 5.2. Find new colony locations xnew :

xnew = xold + U (0; β · d ) · V1 + d · tgθ · V2 , (20.36)

where β, θ are the parameters, d is the distance from the colony to the imperialist
state, U (a, b) is the random variable uniformly distributed on [a, b], V1 is the iden-
tity vector directed from the colony to the imperialist state, V2 is the random iden-
tity vector perpendicular V1 : {V2 } = {{rand }|V1 · V2 = 0, V2 = 1, V1 = 1},

n
j j
θ = U (−γ , γ ), d = (ximp,i − xc,i )2 , j is the imperialist state number, ximp is
i=1
the position of the imperialist state, xc is the colony position. Thus, vectors V1 and
V2 can be found as follows:

1 j j
V1 = x − xc,1 ; . . . ; ximp,n − xc,n , (20.37)
d imp,1

[email protected]
282 A. V. Panteleev and M. M. S. Karane

V1 · Ṽ2 = V1,1 · Ṽ2,1 + . . . + V1,n · Ṽ2,n = 0 (orthogonality condition),

(20.38)

where components Ṽ2,1 , . . . , Ṽ2,n−1 can be generated randomly on a segment

[−1, 1], and Ṽ2,n find from the orthogonality condition. Next, normalize the vector
Ṽ2 : V2 = Ṽ2  .
Ṽ2 
Step 5.3. Calculate the value of the objective function f (xnew ).
j
If f (xnew ) ≥ f (ximp ), then go to Step 5.4.
j j j
If f (xnew ) < f (ximp ), then let xc = ximp , ximp = xnew (the colony becomes an
imperialist state, and the former imperialist state becomes a colony).
Go to Step 5.1.
j
Step 5.4. If the number of colonies that have changed position has reached Nc ,
then complete the procedure. Otherwise, randomly select the next colony, which
has not yet changed position, and go to Step 5.2.
Step 6. Competition between empires.
Step 6.1. Find the total cost of the empire:


j
Nc
f (xci )
j i=1
TC j = f (ximp ) +ξ · j
, j = 1, . . . , N , (20.39)
Nc

where ξ is the positive number less than 1, xci is the position of ith colony of jth
j
empire determined by the position ximp of the imperialist state.
Step 6.2. Find normalized total cost of empire:

NTC j = TC j − max{TC i } . (20.40)

i∈{1,...,Nimp }

Step 6.3. Find the level of influence of each empire:

 
 
 
 NTC j 
 
P = N
j
, j = 1, . . . , Nimp . (20.41)
 imp 
 NTC 
i

i=1

Step 6.4. Find the weakest empire with the lowest value P j , j ∈ {1, . . . , Nimp }.
Step 6.5. In the found empire, find the weakest colony with the highest value of
the objective function.
Step 6.6. Generate vectors:

P = (p1 , p2 , . . . , pNimp ); R = (r 1 , r 2 , . . . , r Nimp ), (20.42)

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 283

where r j = U (0, 1),

D = P − R = (d 1 , d 2 , . . . , d Nimp ) = (p1 − r 1 , . . . , pNimp − r Nimp ). (20.43)

Step 6.7. The colony defined in Step 6.5 should be included in the empire
corresponding to the largest value among the components of the row vector D.
Step 7. Disintegration of the weakest empires.
If there are no colonies in the empire, it will cease to exist (the country is included
in the empire, as defined in Step 6.7). There is a new number Nimp .
Step 8. If Nimp = 1, then process is complete. The solution of the problem is to
consider the position of the imperialist state. Else, go to Step 5.
Recommendation on the parameters selection. The number of countries Npop
determines the number of calculations of the objective function at each iteration. For
a problem with a large range of admissible solutions, it is recommended to take a
larger parameter value Npop . Recommended parameter values Npop ∈ [50, 300].
The number of imperialist countries Nimp determines how long the search for
new solutions will continue. At the end of the search, only one empire remains.
Recommended values for a standard set of functions Nimp ∈ [5, 30].
The colony shift parameters β, γ determine the movement of the colonies to its
empire (see Step 5.2). Recommended parameter values β = 2, γ = π4 .
Colony influence parameter ξ . Recommended parameter value ξ = 0.1.

20.2.5 Hybrid Multi-agent Algorithm

After analyzing the test function optimization results of the methods described, it can
be seen that the result obtained by the imperialist competition algorithm differs from
the other two. The members of the population at the last iteration are scattered around
the maximum point, while in the other two methods, all the individuals merged into
one point, and the result obtained by this method gives a large deviation from the
exact solution.
Imperialist competition algorithm does not stop with the achievement of a given
number of iterations (like the other two methods), but stops when all empires are
dissolved except for one. The position of this empire is the solution of the problem.
As a result, the idea arose to create a hybrid algorithm [3], which includes the
imperialist competition algorithm and a method of a krill herd. By combining these
two methods, one can achieve a shorter algorithm time and obtain a more accurate
result, since at the beginning, the method that imitates imperialist competition will
help to reduce the scope of the solution search by obtaining the final population in
some neighborhood of the extreme point, and the following method will help clarify
the solution in the found area.

[email protected]
284 A. V. Panteleev and M. M. S. Karane

20.3 Application of Multi-agent Methods for Optimal

Open-Loop Control Problems

In this section, the statement of the optimal open-loop control problem is given
in Sect. 20.3.1. In Sect. 20.3.2, optimal control problem solutions are proposed.
Section 20.3.3 provides the designed software, while Sect. 20.3.4 describes the
statement of optimal open-loop control test problems and their solutions.

20.3.1 Statement of the Problem

Let the behavior of the control object model be described by an ordinary differential
equation:

ẋ(t) = f (t, x(t), u(t)) , (20.44)

where x is the system state vector, x = (x1 , . . . , xn )T ∈ Rn , u is the control vector,

u = (u1 , . . . , uq )T ∈ U ⊆ Rq , U is some given set of admissible control values
determined by the direct product of segments [a1 , b1 ]×· · ·×[aq , bq ], t ∈ T = [t0 , t1 ]
is the system time interval, start time t0 and terminal time t1 are set, f (t, x, u) is the
continuous vector function, Rn is n-dimensional Euclidean space.
The initial condition x(t0 ) = x0 sets the initial state of the system.
We define the set of admissible processes D(t0 , x0 ) as a set of pairs d = (x(·), u(·))
that include trajectory x(·) and control u(·) (where ∀t ∈ T : x(t) ∈ Rn , u(t) ∈
U , functions x(·) are continuous and piecewise-differentiable, and u(·) piecewise-
continuous), satisfying Eq. 20.44 with the initial condition.
On the set D(t0 , x0 ), we define the cost functional

I (d ) = F(x(t1 )). (20.45)

Need to find such a pair d ∗ = (x∗ (·), u∗ (·)) ∈ D(t0 , x0 ) that I (d ∗ ) = min I (d ).
d ∈D(t0 ,x0 )
We consider Eq. 20.44 linear in control, which has the form:

ẋ(t) = A(x(t)) + B(t)u(t) , (20.46)

where A(x) is the nonlinear function and B(t) is the matrix n × q depending on time.
In Eq. 20.46, the structure of optimal open-loop control is relay according to the
maximum principle. Therefore, it is proposed to look for an approximate solution
in a parametric form determined by the number of control switching moments and
their values.

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 285

20.3.2 Search Algorithm of Optimal Open-Loop Control

Search algorithm includes the following steps.

Step 1. Initialization. Select a method from the group of multi-agent algorithms
and set its parameters. Set a switching number p = 0 in the control u(t), wherein
tΠ0 ∈ {t0 , t1 }.
Step 2. Generate the initial population (controls) of NP individuals on the time
interval t ∈ [t0 , t1 ]. The resulting 1, . . . , NP sequences of values are switching
points tR ∈ [t0 , t1 ] in the control u(t).
Step 3. Generate control by generated switching point values
p
upj (t) = ap χ (t0 ) + (ap − bp ) (−1)k χ (t − tRk ), (20.47)
k=0


0 if t ≤ 0
where χ (t) = , j ∈ 1, NP, p ∈ 1, q, ap ≤ u ≤ bp .
1 if t > 0
Step 4. Integrate NP systems of differential equations (Eq. 20.46) with controls
u1 (t), . . . , uNP (t) using the fourth order Runge–Kutta method. For any individual,
obtain the corresponding trajectories x11 , . . . , x1NP , . . . , xn1 , . . . , xnNP and calculate the
values of the cost functional I 1 , . . . , I NP .
Step 5. Fulfill the next iteration of the selected method of minimizing Eq. 20.45.
Obtain new positions of individuals 1 , . . . , NP (switching point values). Go to Step
3.
Step 6. The loop (Step 3–Step 5) ends, when a certain number of iterations are reached.
The best individual is selected (set of control switching points). The corresponding
control and trajectory, as well as, the value of the cost functional Ip∗ , are taken as an
approximate solution of the problem with the switching number equaled to p.
∗
Step 7. If Ip∗ < Ip−1 (condition is checked under p ≥ 1), then let p = p + 1 and go
∗
to Step 2. If Ip∗ ≥ Ip−1 , then the search procedure for optimal open-loop control is
completed, and control with p switching is selected.

20.3.3 Software

Software [3] was developed based on all methods described above. To create it, we
used Microsoft Visual Studio development environment, the programming language
is C #.
With the help of the developed software, the effectiveness of the described algo-
rithms on a standard set of test examples (for example: Shaffer function, root function,
trapfall, and Rosenbrock function.) was explored. Also, the set of problems on find-
ing the optimal open-loop control was solved (Tables 20.1, 20.2, 20.3, 20.4, 20.5 and

[email protected]
286 A. V. Panteleev and M. M. S. Karane

Table 20.1 Formulation of

Parameters Values
the task 1
The dimension of the state vector n=2
Time interval t ∈ [0, 1]
Control constraint −1 ≤ u ≤ 1
Initial value x(0) = (0, 0)

System of differential equations ẋ1 = x2 + sin x1 + u
ẋ2 = x1 cos x2 u

Cost functional I (u) = x2 (1)

Table 20.2 The results of solving the task 1

Optimization method Coordinates of points Switching point The value of the
(x1 (1), x2 (1)) coordinates functional I
Fish school search (0.43494, –0.13587) 0.49 −0.13587
Krill herd (0.42459, –0.13478) 0.48 −0.13571
Imperialist (0.44665, –0.13598) 0.5 −0.13598
competitive
algorithm
Known solution [10] (0.440804, –0.13593) 0.5 −0.13599

Table 20.3 Formulation of the task 2

Parameters Values
The dimension of the state vector n=2
Time interval t ∈ [0, 2]
Control constraint −1 ≤ u ≤ 2
Initial value x(0) = (−1, 0)T

System of differential equations ẋ1 = x22 + u
ẋ2 = 8 sin x1 + x1 − x2 − u
Cost functional I (u) = −x2 (2)

Table 20.4 The results of solving the task 2

Optimization method Coordinates of points Switching point The value of the
(x1 (2), x2 (2)) coordinates functional I
Fish school search (16.50987, 6.37294) (0.62, 1.52, 1.74, −16.50987
1.93)
Krill herd (15.93114, 6.18436) (0.57, 0.58, 0.58, −15.93114
1.85)
Imperialist (16.67731, 6.55339) (0.57, 1.35, 1.52, −16.67731
competitive 1.86)
algorithm
Known solution [10] (16.76268, 6.35095) (0.5, 1.25, 1.5, 1.8) −16.76268

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 287

Table 20.5 Formulation of the task 3

Parameters Values
The dimension of the state vector n=2
Time interval t ∈ [0; 1, 6]
Control constraint −2 ≤ u ≤ 1
Initial value x(0) = (1, 0)T
⎧
System of differential equations ⎨ ẋ1 = 1
+ 3 sin x2 + u
cos x1 + 2
⎩
ẋ2 = x1 + x2 + u
Cost functional I (u) = −x1 (1, 6) + 21 x2 (1, 6)

Table 20.6 The results of solving the task 3

Optimization method Coordinates of points Switching point The value of the
(x1 (1.6), x2 (1.6)) coordinates functional I
Fish school search (3.43034, 12.81994) 1.24 −2.97963
Krill herd (3.52562, 13.00372) 1.27 −2.97624
Imperialist (3.93412, 13.53938) 1.39 −2.83557
competitive
algorithm
Known solution [10] (3.46114, 12.884) 1.26 −2.98086

20.6).
On the initial form of the software (Fig. 20.1), the user can select a task to find
the optimal open-loop control, set the switching number in the control, select the
optimization method, and specify its parameters.
The result of the program are the coordinates of the points x1 (t1 ), x2 (t2 ), the
optimal value of the cost functional I , and the coordinates of the switching points.
After finding the optimal control, the program displays the graphs of the control
function and trajectories.

20.3.4 Solving the Problem of Finding Optimal Open-Loop

Control

Task 1. Formulation of the task (Table 20.1) [3].

The best switching number: p = 1.
Optimization method and its parameters: fish school search (NP = 5, ITER =
100, Wscale = 150, W = 100, stepvol = 0.1, stepind = 0.01), krill herd (NP = 10,

[email protected]
288 A. V. Panteleev and M. M. S. Karane

Fig. 20.1 Software interface

ITER = 100, Smax = 0.01, mu = 0.5, Vf = 0.02, Dmax = 0.005, ci = 0.2), and
imperialist competitive algorithm (Npop = 150, Nimp = 15, ITER = 500, β = 0.2,
γ = 0.02, ξ = 0.01).
The results of solving the task are presented in Table 20.2.
Graphs of optimal trajectories and controls are shown in Fig. 20.2.
Task 2. Formulation of the task (Table 20.3) [3].
The best switching number: p = 4.
Optimization method and its parameters: fish school search (NP = 30, ITER =
500, Wscale = 5000, W = 4500, stepvol = 0.1, stepind = 0.01), krill herd (NP = 40,
ITER = 1000, Smax = 0.01, mu = 0.5, Vf = 0.02, Dmax = 0.005, ci = 0.2), and
imperialist competitive algorithm (Npop = 150, Nimp = 15, ITER = 500, β = 0.2,
γ = 0.02, ξ = 0.01).
The results of solving the task are presented in Table 20.4.
Graphs of optimal trajectories and controls are shown in Fig. 20.3.
Task 3. Formulation of the task (Table 20.5) [3].
The best switching number: p = 1.
Optimization method and its parameters: fish school search (NP = 5, ITER =
100, Wscale = 300, W = 250, stepvol = 0.1, stepind = 0.01), krill herd (NP = 10,
ITER = 100, Smax = 0.01, mu = 0.5, Vf = 0.02, Dmax = 0.005, ci = 0.2), and
imperialist competitive algorithm (Npop = 40, Nimp = 4, ITER = 500, β = 0.2,
γ = 0.02, ξ = 0.01).
The results of solving the task are presented in Table 20.6.
Graphs of optimal trajectories and controls are shown in Fig. 20.4.

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 289

Fig. 20.2 Trajectories and control

Fig. 20.3 Trajectories and control

[email protected]
290 A. V. Panteleev and M. M. S. Karane

Fig. 20.4 Trajectories and control

20.4 Conclusions

The chapter outlined multi-agent methods for finding a constrained global extremum:
fish school search, krill herd, and imperialist competitive algorithm. Based on these
methods, an algorithm has been developed for finding the optimal open-loop control
of deterministic systems, which is looking for control in a relay mode with a certain
switching number. Using the described methods, the software was created that allows
to find the solution to the optimal open-loop deterministic control problem. The
software draws the graphics of the trajectories and controls and also calculates the
optimal value of the criterion and coordinates of the switching points in the control.

References

1. Panteleev, A.V., Metlitskaya, D.V.: An application of genetic algorithms with binary and real
coding for approximate synthesis of suboptimal control in deterministic systems. Autom.
Remote Control 72(11), 2328–2338 (2011)
2. Panteleev, A.V., Metlitskaya, D.V.: Using the method of artificial immune systems to seek
the suboptimal program control of deterministic systems. Autom. Remote Control 75(11),
1922–1935 (2014)
3. Karane, M.M.C.: Comparative analysis of multi-agent methods for constrained global optimiza-
tion. In: IV International Conference on Information Technologies in Engineering Education,
pp. 128–133 (2018)

[email protected]
20 Multi-agent Optimization Algorithms for a Single Class … 291

4. Bastos Filho, C.J.A., de Lima Neto, F.B., Lins, A.J.C.C., Nascimento, A.I.S., Lima, M.P.: A
novel search algorithm based on fish school behavior. In: 2008 IEEE International Conference
on Systems, Man and Cybernetics, pp. 2646–2651 (2008)
5. Bastos Filho, C.J.A., de Lima Neto, F.B., Lins, A.J.C.C., Nascimento, A.I.S., Lima, M.P.: Fish
school search: overview. In: Chiong, R. (Ed.) Nature-Inspired Algorithms for Optimisation.
SCI, vol. 193, pp. 261–277. Springer, Heidelberg (2009)
6. Bacanin, N., Pelevic, B., Tuba, M.: Krill herd (KH) algorithm for portfolio optimization. In:
Mathematics and Computers in Business, Manufacturing and Tourism, pp. 39–44 (2013)
7. Gandomi, A.H., Alavi, A.H.: Krill herd: a new bio-inspired optimization algorithm. Commun.
Nonlinear Sci. Numer. Simul. 17(12), 4831–4845 (2012)
8. Atashpaz-Gargari, E., Lucas, C.: Imperialist competitive algorithm: an algorithm for optimiza-
tion inspired by imperialist competition. In: IEEE Congress on Evolutionary Computation,
pp. 4661–4667 (2007)
9. Kaveh, A., Talatahari, S.: Imperialist competitive algorithm for engineering design problems.
Asian J. Civ. Eng. (Build. Hous.) 11(6), 675–697 (2010)
10. Finkelstein, E.A.: Computational technologies of approximation of the reachable set of a con-
trolled system. Dissertation for the degree of Canadian Technological Sciences. Institute of
System Dynamics and Control Theory, Irkutsk (in Russian) (2018)

[email protected]
Chapter 21
Spectral Method for Analysis
of Diffusions and Jump Diffusions

Gevorg Y. Baghdasaryan , Marine A. Mikilyan , Andrei V. Panteleev

and Konstantin A. Rybakov

Abstract The chapter discusses the use of the spectral form of mathematical descrip-
tion, or the spectral method, for the statistical analysis of stochastic dynamical sys-
tems: diffusions and jump diffusions, i.e., for solving Fokker–Planck–Kolmogorov
equation and Kolmogorov–Feller equation for the probability density of the state
vector for these dynamical systems. The spectral form of mathematical description
allows to transform linear partial differential equations or partial integro-differential
equations into a system of linear algebraic equations, which determines coefficients
according to orthogonal series expansions for the probability density with respect to
an arbitrary orthonormal system of functions. As an example for testing, the Dryden
wind turbulence model and its modification, allowing to take into account not only
continuous random effects but also impulse ones, are considered.

21.1 Introduction

The spectral form of mathematical description, or the spectral method, is used for
solving various problems of the control theory [1]. At the beginning, it was applied to
solve the output processes analysis problem of linear nonstationary control systems.

G. Y. Baghdasaryan · M. A. Mikilyan
Institute of Mechanics, National Academy of Sciences of Armenia, 24, Marshal Baghramyan
Ave., 0019 Yerevan, Armenia
e-mail: [email protected]
M. A. Mikilyan
e-mail: [email protected]
Russian-Armenian University, 123, Hovsep Emin St, 0051 Yerevan, Armenia
A. V. Panteleev · K. A. Rybakov (B)
Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
A. V. Panteleev
e-mail: [email protected]
© Springer Nature Singapore Pte Ltd. 2020 293
L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_21

[email protected]
294 G. Y. Baghdasaryan et al.

Further, its scope was significantly expanded [2–6]. Now it is used for the analysis,
synthesis, and identification of control systems, which are described by ordinary dif-
ferential equations and partial differential equations, integro-differential equations,
equations with fractional derivatives, etc.
The use of the spectral form of mathematical description implies the representa-
tion of input and output signals as sequences of coefficients according to orthogonal
series expansions. Such signal characteristics are called the spectral characteristics.
The characteristics of linear dynamical systems in the spectral form of mathematical
description are the spectral characteristics of linear operators [7]. Elementary and
specialized algorithms to use the spectral form of mathematical description are devel-
oped for computer algebra systems and applications [8]. These algorithms include
subroutines for calculating the spectral characteristics of typical input signals, spec-
tral characteristics of multiplication operators with typical multipliers, differentiation
and integration operators, shift operators, and operators of fractional integration and
differentiation with respect to various orthonormal systems of functions such as Leg-
endre, Chebyshev, Laguerre, and Hermite polynomials, trigonometric and complex
exponential functions, Walsh and Haar functions, functions defined by wavelets or
splines, etc.
In this chapter, the output processes analysis problem for nonlinear stochastic
dynamical control systems is concerned. The results obtained in [4] for diffusions
are expanded to jump diffusions. As an example for testing the spectral method, the
Dryden wind turbulence model and its jump diffusion modification are considered.
The remainder of this chapter is organized as follows. In Sect. 21.2, the spec-
tral method formalism is proposed. The spectral method for analysis problem of
diffusions is described in Sect. 21.3. Section 21.4 provides the spectral method for
analysis problem of jump diffusions. The chapter is summarized in Sect. 21.5.

21.2 Spectral Method Formalism

Hereinafter, the necessary definitions for multidimensional matrices as well spectral

characteristics of functions and linear operators are given in Sects. 21.2.1–21.2.2,
respectively.

21.2.1 Multidimensional Matrices

Multidimensional matrices are needed for further presentation of main results.

Therefore, the necessary definitions for multidimensional matrices are given below.
We will denote ( p + q)-dimensional matrix with entries ai1 ...i p j1 ... jq by A( p, q),
i 1 , . . . , i p , j1 , . . . , jq = 0, 1, 2, . . . The separation of indices into two groups
i 1 , . . . , i p and j1 , . . . , jq allows to attribute the structure of a multidimensional matrix
[7], it is important to define the product of multidimensional matrices [9].

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 295

1. Let α, β ∈ R and let A( p, q) = [ai1 ...i p j1 ... jq ] and B( p, q) = [bi1 ...i p j1 ... jq ] be the
infinite ( p + q)-dimensional matrices. The expression α A( p, q) + β B( p, q) is
the infinite ( p + q)-dimensional matrix C( p, q) = [ci1 ...i p j1 ... jq ] if

ci1 ...i p j1 ... jq = αai1 ...i p j1 ... jq + βbi1 ...i p j1 ... jq , i 1 , . . . , i p , j1 , . . . , jq = 0, 1, 2, . . .

2. Let A( p, r ) = [ai1 ...i p k1 ...kr ] and B(r, q) = [bk1 ...kr j1 ... jq ] be the infinite ( p + r )-
dimensional and (r +q)-dimensional matrices, respectively. The product A( p, r )·
B(r, q) is the infinite ( p + q)-dimensional matrix C( p, q) = [ci1 ...i p j1 ... jq ] if

∞

ci1 ...i p j1 ... jq = ai1 ...i p k1 ...kr bk1 ...kr j1 ... jq < ∞,
k1 ,...,kr =0

i 1 , . . . , i p , j1 , . . . , jq = 0, 1, 2, . . .

An infinite 2p-dimensional matrix E( p, p) is said to be the identity matrix

if A( p, p) · E( p, p) = E( p, p) · A( p, p) = A( p, p) for each 2p-dimensional
matrix A( p, p).
3. Let A( p, p) be an infinite 2p-dimensional matrix. An infinite 2p-dimensional
matrix B( p, p) is said to be the two-sided inverse of A( p, p) if A( p, p) ·
B( p, p) = B( p, p) · A( p, p) = E( p, p). We will use the notation A−1 ( p, p)
to denote the two-sided inverse of A( p, p).
4. Let A( p, q) = [ai1 ...i p j1 ... jq ] and B(r, s) = [bk1 ...kr l1 ...ls ] be the infinite ( p + q)-
dimensional and (r + s)-dimensional matrices, respectively. The tensor product
A( p, q) ⊗ B(r, s) is the infinite ( p +r + q + s)-dimensional matrix C( p +r, q +
s) = [ci1 ...i p k1 ...kr j1 ... jq l1 ...ls ] if

ci1 ...i p k1 ...kr j1 ... jq l1 ...ls = ai1 ...i p j1 ... jq bk1 ...kr l1 ...ls ,
i 1 , . . . , i p , k1 , . . . , kr , j1 , . . . , jq , l1 , . . . , ls = 0, 1, 2, . . .

5. Let A( p, q) = [ai1 ...i p j1 ... jq ] be an infinite ( p +q)-dimensional matrix. An infinite

(q + p)-dimensional matrix B(q, p) = [b j1 ... jq i1 ...i p ] is said to be the transpose
of A( p, q) if

b j1 ... jq i1 ...i p = ai1 ...i p j1 ... jq , i 1 , . . . , i p , j1 , . . . , jq = 0, 1, 2, . . .

We will use the notation AT ( p, q) to denote the transpose of A( p, q).

[email protected]
296 G. Y. Baghdasaryan et al.

21.2.2 Spectral Characteristics of Functions and Linear

Operators

In this section, we introduce necessary definitions and propositions for the spectral
characteristics and the spectral transform.
Let {qi0 (t)}i∞0 =0 be an orthonormal basis of L 2 (T) and let { pi1 ...in (x)}i∞1 ,...,in =0 be
an orthonormal basis of L 2 (Rn ) [10], where t ∈ T = [t0 , T ] and x ∈ Rn . Then

{ei0 i1 ...in (t, x) = qi0 (t) pi1 ...in (x)}i∞0 ,i1 ,...,in =0 (21.1)

is the orthonormal basis of L 2 (T × Rn ).

Definition 1 An infinite (n + 1)-dimensional matrix H (n + 1, 0) = [h i0 i1 ...in ] is

called the spectral characteristic of a square-integrable function h(t, x), i.e., h(t, x) ∈
L 2 (T × Rn ), if
 
h i0 i1 ...in = (ei0 i1 ...in (t, x), h(t, x)) L 2 (T×Rn ) = ei0 i1 ...in (t, x)h(t, x)dxdt,
T Rn
i 0 , i 1 , . . . , i n = 0, 1, 2, . . . ,

and
∞

h(t, x) = h i0 i1 ...in ei0 i1 ...in (t, x), (t, x) ∈ T × Rn . (21.2)
i 0 ,i 1 ,...,i n =0

Thus, H (n + 1, 0) = S[h(t, x)] and h(t, x) = S−1 [H (n + 1, 0)], where S and

S−1 denote the spectral transform and the spectral inversion, respectively.
Similarly, the spectral characteristic of a square-integrable function h(t) ∈ L 2 (T)
with respect to the orthonormal basis {qi0 (t)}i∞0 =0 and the spectral characteristic of
a square-integrable function h(x) ∈ L 2 (Rn ) with respect to the orthonormal basis
{ pi1 ...in (x)}i∞1 ,...,in =0 can be defined. For example, an infinite n-dimensional matrix
H (n, 0) = [h i1 ...in ] is called the spectral characteristic of a function h(x) ∈ L 2 (Rn )
if

h i1 ...in = ( pi1 ...in (x), h(x)) L 2 (Rn ) = pi1 ...in (x)h(x)dx, i 1 , . . . , i n = 0, 1, 2, . . .
Rn

and
∞

h(x) = h i1 ...in pi1 ...in (x), x ∈ Rn . (21.3)
i 1 ,...,i n =0

Definition 2 An infinite 2(n + 1)-dimensional matrix A(n + 1, n + 1) =

[ai0 i1 ...in j0 j1 ... jn ] is said to be the spectral characteristic of a linear operator A : DA ⊆

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 297

L 2 (T × Rn ) → L 2 (T × Rn ) if

ai0 i1 ...in j0 j1 ... jn = (ei0 i1 ...in (t, x), Ae j0 j1 ... jn (t, x)) L 2 (T×Rn )
 
= ei0 i1 ...in (t, x)Ae j0 j1 ... jn (t, x)dxdt,
T Rn
i 0 , i 1 , . . . , i n , j0 , j1 , . . . , jn = 0, 1, 2, . . .

The spectral transform of linear operators is also denoted by S, therefore, S[A] =

A(n + 1, n + 1).
Basic properties of the spectral transform for functions and linear operators are
as follows:
1. For any h l (t, x) ∈ L 2 (T × Rn ) and αl ∈ R, l = 1, 2, . . . , L , the following
relation is satisfied (the linearity property of the spectral transform):
 

L 
L
S αl h l (t, x) = αl S[h l (t, x)].
l=1 l=1

2. If the function h(t, x) ∈ L 2 (T × Rn ) such that h(t ∗ , x) = h ∗ (x) ∈ L 2 (Rn ),

t ∗ ∈ T, q(1, 0; t ∗ ) is the infinite column matrix with entries qi0 (t ∗ ), i.e.,

q(1, 0; t ∗ ) = [ q0 (t ∗ ) q1 (t ∗ ) q2 (t ∗ ) . . . ]T ,

H (n + 1, 0) and H ∗ (n, 0) are spectral characteristics of h(t, x) and h ∗ (x),

respectively, then

(q T (1, 0; t ∗ ) ⊗ E(n, n)) · H (n + 1, 0) = H ∗ (n, 0),

where E(n, n) is the 2n-dimensional identity matrix.

3. If A : DA ⊆ L 2 (T × Rn ) → L 2 (T × Rn ) is a linear operator, h(t, x) ∈ DA ,
and A(n + 1, n + 1) is the spectral characteristic of A, then

S[Ah(t, x)] = A(n + 1, n + 1) · S[h(t, x)].

4. If A : DA ⊆ L 2 (T × Rn ) → L 2 (T × Rn ) and B : DB ⊆ L 2 (T × Rn ) → RB ⊆
DA are linear operators, C = A◦B is a composition of A and B, A(n +1, n +1),
B(n + 1, n + 1), and C(n + 1, n + 1) are spectral characteristics of A, B, and
C, respectively, then

C(n + 1, n + 1) = A(n + 1, n + 1) · B(n + 1, n + 1).

Spectral characteristics of functions with a similar properties can also be defined

for elements, which do not belong to L 2 (T × Rn ) (e.g., for elements of L p (T × Rn ),
where p < 2, or for distributions [1, 11]).

[email protected]
298 G. Y. Baghdasaryan et al.

21.3 Spectral Method for Analysis of Diffusions

In this section, the statement of the problem is given in Sect. 21.3.1. Section 21.3.2
provides the detailed description of the spectral method for solving Fokker–Planck–
Kolmogorov equation, and Sect. 21.3.3 contains numerical results for the analysis
of Dryden wind turbulence model.

21.3.1 Problem Statement

Let X (t) be an Rn -valued random process that satisfies Itô Stochastic Differential
Equation (SDE):

d X (t) = f (t, X (t))dt + σ (t, X (t))dW (t), X (t0 ) = X 0 , (21.4)

where t ∈ T = [t0 , T ], f (t, x):T × Rn → Rn is the n-dimensional function, σ (t, x):

T × Rn → Rn×s is the (n × s)-dimensional matrix function, W (t) is the standard
s-dimensional Wiener process, X 0 is the initial state with a given probability density
ϕ0 (x) (X 0 and W (t) are independent).
Functions f (t, x) and σ (t, x) satisfy the conditions for the existence and unique-
ness of the strong or weak solution of SDEs [12], and E|X 0 |2 < +∞, where E is the
expectation or mean.
For any t ∈ T, the most comprehensive statistical characteristic of X (t) is the
probability distribution function F(t, x) = F(t, x1 , . . . , xn ), x ∈ Rn :

F(t, x) = Pr{X 1 (t) < x1 , . . . , X n (t) < xn },

where Pr{ · } is the probability. This characteristic can be expressed by an integral

of the probability density ϕ(t, x) = ϕ(t, x1 , . . . , xn ) as follows:
 x1  xn
F(t, x1 , . . . , xn ) = ... ϕ(t, x1 , . . . , xn )dx1 . . . dxn .
−∞ −∞

Thus,

∂ n F(t, x1 , . . . , xn )
ϕ(t, x1 , . . . , xn ) = .
∂ x1 . . . ∂ xn

It is known that if the probability density ϕ(t, x) exists then it satisfies Fokker–
Planck–Kolmogorov equation or Kolmogorov’s forward equation [7, 12–15]:

∂ϕ(t, x) n
∂ 1   ∂2
n n
=− [ f i (t, x)ϕ(t, x)] + [gi j (t, x)ϕ(t, x)],
∂t i=1
∂ xi 2 i=1 j=1 ∂ xi ∂ x j

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 299

ϕ(t0 , x) = ϕ0 (x) (21.5)

or
∂ϕ(t, x)
= Aϕ(t, x), ϕ(t0 , x) = ϕ0 (x), (21.6)
∂t
where A is the forward diffusion operator defined by:

n
∂ 1   ∂2
n n
Aϕ(t, x) = − [ f i (t, x)ϕ(t, x)] + [gi j (t, x)ϕ(t, x)].
i=1
∂ xi 2 i=1 j=1 ∂ xi ∂ x j
(21.7)

We assume that
(a) There exists the probability density ϕ(t, x) for the random process X (t).
(b) Probability densities ϕ(t, x) and ϕ0 (x) are the square-integrable functions, i.e.,
ϕ(t, x) ∈ L 2 (T × Rn ), ϕ0 (x) ∈ L 2 (Rn ).
(c) For any ξ(t, x) ∈ C0∞ (T × Rn ), the following equation is satisfied:
  
n  
∂ϕ(t, x) ∂ξ(t, x)
ξ(t, x) dxdt = f i (t, x)ϕ(t, x)dxdt
T Rn ∂t i=1T Rn ∂ xi
n  
1 
n
∂ 2 ξ(t, x)
+ gi j (t, x)ϕ(t, x)dxdt,
2 i=1 j=1 T Rn ∂ xi ∂ x j

where C0∞ (T × Rn ) denotes the space of functions that have the compact support
and continuous derivatives of all orders, i.e., there exists a weak solution [16] of the
equation for the probability density ϕ(t, x).
In relations given above, gi j (t, x) are entries of the (n×n)-dimensional symmetric
matrix function g(t, x) = σ (t, x)σ T (t, x), i, j = 1, 2, . . . , n.
Thus, the analysis problem of diffusions (Eq. 21.4) is formulated as follows. Given
functions f (t, x), σ (t, x), defining the Itô SDE and probability density ϕ0 (x) of the
initial state X 0 , find the probability density ϕ(t, x).

21.3.2 Spectral Method for Solving

Fokker–Planck–Kolmogorov Equation

Apply the spectral transform to the left-hand and right-hand sides of Eq. 21.5 using
the linearity property of the spectral transform (see Sect. 21.2.2). Then

[email protected]
300 G. Y. Baghdasaryan et al.

  n  
∂ϕ(t, x) ∂
S =− S [ f i (t, x)ϕ(t, x)]
∂t i=1
∂ xi
 
1 
n n
∂2
+ S [gi j (t, x)ϕ(t, x)] . (21.8)
2 i=1 j=1 ∂ xi ∂ x j

Further, we will use new notations (i, j = 1, 2, . . . , n):

(i) P(n + 1, n + 1) is the spectral characteristic of the differentiation operator
∂/∂t.
(ii) Pi (n + 1, n + 1) and Pi j (n + 1, n + 1) are the spectral characteristics of
differentiation operators ∂/∂ xi and ∂ 2 /∂ xi ∂ x j , respectively.
(iii) Fi (n + 1, n + 1) and G i j (n + 1, n + 1) are the spectral characteristics of
multiplication operators with multipliers f i (t, x) and gi j (t, x), respectively.
Let (n+1, 0) and 0 (n, 0) be the spectral characteristics of probability densities
ϕ(t, x) and ϕ0 (x), respectively, i.e.,
 
(n + 1, 0) = [ϕi0 i1 ...in ], ϕi0 i1 ...in = ϕ(t, x)ei0 i1 ...in (t, x)dxdt, (21.9)
T Rn

0 (n, 0) = [ϕ0i1 ...in ], ϕ0i1 ...in = ϕ0 (x) pi1 ...in (x)dx, (21.10)
Rn

i 0 , i 1 , . . . , i n = 0, 1, 2, . . .

Then we have
 
∂ϕ(t, x)
S = P(n + 1, n + 1) · (n + 1, 0) − q(1, 0; t0 ) ⊗ 0 (n, 0),
∂t
 
∂
S [ f i (t, x)ϕ(t, x)] = Pi (n + 1, n + 1) · Fi (n + 1, n + 1) · (n + 1, 0),
∂ xi
 
∂2
S [gi j (t, x)ϕ(t, x)] = Pi j (n + 1, n + 1) · G i j (n + 1, n + 1) · (n + 1, 0),
∂ xi ∂ x j

where

P(n + 1, n + 1) = P(n + 1, n + 1) + (q(1, 0; t0 ) · q T (1, 0; t0 )) ⊗ E(n, n),

(21.11)

Pi j (n + 1, n + 1) = Pi (n + 1, n + 1) · P j (n + 1, n + 1), (21.12)

i, j = 1, 2, . . . , n.

To prove this, it is necessary to use properties of the spectral transform for

functions and linear operators from Sect. 21.2.2. In particular,

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 301
 
∂ϕ(t, x)
S = P(n + 1, n + 1) · (n + 1, 0),
∂t

where the spectral characteristic P(n + 1, n + 1) can be represented in the form

P(n + 1, n + 1) = P(n + 1, n + 1) − (q(1, 0; t0 ) · q T (1, 0; t0 )) ⊗ E(n, n)

by the property 3. Using the property 2, we obtain

P(n + 1, n + 1) · (n + 1, 0)
= (P(n + 1, n + 1) − (q(1, 0; t0 ) · q T (1, 0; t0 )) ⊗ E(n, n)) · (n + 1, 0)
= P(n + 1, n + 1) · (n + 1, 0)
− (q(1, 0; t0 ) ⊗ E(n, n)) · (q T (1, 0; t0 ) ⊗ E(n, n)) · (n + 1, 0)
= P(n + 1, n + 1) · (n + 1, 0) − (q(1, 0; t0 ) ⊗ E(n, n)) · 0 (n, 0)
= P(n + 1, n + 1) · (n + 1, 0) − q(1, 0; t0 ) ⊗ 0 (n, 0).

Properties 3 and 4 imply relations for first-order and second-order derivatives of

the probability density ϕ(t, x) with multipliers f i (t, x) and gi j (t, x).
Thus, Eq. 21.8 reduces to

P(n + 1, n + 1) · (n + 1, 0) − q(1, 0; t0 ) ⊗ 0 (n, 0)


n
=− Pi (n + 1, n + 1) · Fi (n + 1, n + 1) · (n + 1, 0)
i=1

1 
n n
+ Pi j (n + 1, n + 1) · G i j (n + 1, n + 1) · (n + 1, 0), (21.13)
2 i=1 j=1

consequently, the spectral characteristic (n + 1, 0) satisfies the equation

P(n + 1, n + 1) · (n + 1, 0) − q(1, 0; t0 ) ⊗ 0 (n, 0)

= A(n + 1, n + 1) · (n + 1, 0), (21.14)

where


n
A(n + 1, n + 1) = − Pi (n + 1, n + 1) · Fi (n + 1, n + 1)
i=1

1 
n n
+ Pi j (n + 1, n + 1) · G i j (n + 1, n + 1). (21.15)
2 i=1 j=1

In fact, the 2(n + 1)-dimensional matrix A(n + 1, n + 1) is the spectral charac-

teristic of the forward diffusion operator A defined by Eq. 21.7. Also note that the

[email protected]
302 G. Y. Baghdasaryan et al.

spectral characteristic (n + 1, 0) is called the generalized characteristic function

(the ordinary characteristic function is Fourier transform of the probability density),
and Eq. 21.13 is called the generalized characteristic function equation [2, 4, 7].
Thus, the analysis problem of diffusions (Eq. 21.4) is reduced to solving the infinite
system of linear algebraic equations (Eq. 21.13) with unknown entries of the spectral
characteristic (n + 1, 0). Some aspects of solving the infinite system of linear
algebraic equations are given in [17, 18]. The exact solution of Eqs. 21.13–21.14 is
defined by:

(n + 1, 0) = (P(n + 1, n + 1) − A(n + 1, n + 1))−1 · (q(1, 0; t0 ) ⊗ 0 (n, 0)).

(21.16)

Then, the probability density ϕ(t, x) is given by Eq. 21.2:

∞

ϕ(t, x) = S−1 [(n + 1, 0)] = ϕi0 i1 ...in ei0 i1 ...in (t, x), (t, x) ∈ T × Rn ,
i 0 ,i 1 ,...,i n =0
(21.17)

where ϕi0 i1 ...in are entries of the spectral characteristic (n + 1, 0).
To find the approximate solution of the analysis problem of diffusions (Eq. 21.4)
all spectral characteristics should be truncated on all dimensions. The methodical
inaccuracy caused by the spectral characteristic truncation is described in [1, 2]. In
this case, we obtain the probability density ϕ(t, x) as a partial sum

0 −1 L
L  1 −1 n −1
L
ϕ(t, x) ≈ ... ϕi0 i1 ...in ei0 i1 ...in (t, x), (t, x) ∈ T × Rn , (21.18)
i 0 =0 i 1 =0 i n =0

where L 0 , L 1 , . . . , L n are the truncation orders of spectral characteristics.

The algorithm for solving the analysis problem of diffusions (Eq. 21.4) by the
spectral method is given below:
1. Specify basis systems {qi0 (t)}i∞0 =0 and { pi1 ...in (x)}i∞1 ,...,in =0 for L 2 (T) and L 2 (Rn ),
respectively. Form the basis system {ei0 i1 ...in (t, x)}i∞0 ,i1 ,...,in =0 for L 2 (T × Rn ) by
Eq. 21.1. Specify truncation orders L 0 , L 1 , . . . , L n for all spectral characteristics.
2. Find the column matrix q(1, 0; t0 ) with entries qi0 (t0 ), i.e.,

q(1, 0; t0 ) = [ q0 (t0 ) q1 (t0 ) . . . q L 0 −1 (t0 ) ]T .

3. Find spectral characteristics P(n+1, n+1) and Pi (n+1, n+1) of differentiation

operators ∂/∂t and ∂/∂ xi , respectively, i = 1, 2, . . . , n. Find the matrix P(n +
1, n +1) and spectral characteristics Pi j (n +1, n +1) of differentiation operators
∂ 2 /∂ xi ∂ x j using Eqs. 21.11–21.12, i, j = 1, 2, . . . , n.
4. Find spectral characteristics Fi (n+1, n+1) and G i j (n+1, n+1) of multiplication
operators with multipliers f i (t, x) and gi j (t, x), respectively, i, j = 1, 2, . . . , n.

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 303

5. Find the spectral characteristic A(n + 1, n + 1) of the forward diffusion operator

A using Eq. 21.15.
6. Find the spectral characteristic 0 (n, 0) of the probability density ϕ0 (x) for the
initial state X 0 using Eq. 21.10.
7. Find the solution (n + 1, 0) of the generalized characteristic function Eq. 21.14
using Eq. 21.16.
8. Find the approximate solution ϕ(t, x) of Fokker–Planck–Kolmogorov equation
by Eq. 21.18.

21.3.3 Dryden Wind Turbulence Model

Consider one-dimensional Dryden wind turbulence model that is a mathematical

model of continuous gusts [19–22]. It is given by the linear SDE:

Vt 
d X (t) = −μX (t)dt + σ dW (t), X (0) = X 0 , μ = , σ = 2μσ0 ,
Lt
(21.19)

where t ∈ T = [0, T ], Vt is the longitudinal flight velocity, L t is the turbulence scale

length, σ0 is the standard deviation of the wind velocity, W (t) is the one-dimensional
standard Wiener process, i.e.,

n = s = 1, t0 = 0, f (t, x) = −μx, σ (t, x) = σ, g(t, x) = σ 2 .

Here, X (t) is the wind velocity and X 0 is the initial wind velocity. If X 0 is a
constant or X 0 is a random variable having a normal distribution (Gaussian distribu-
tion), then X (t) is Gaussian process (more precisely, Ornstein–Uhlenbeck process
[12]) with the probability density ϕ(t, x) that is defined by the mean m(t) = EX (t)
and the second-order moment M(t) = EX 2 (t) or the mean m(t) = EX (t) and the
variance D(t) = E(X (t) − m(t))2 = M(t) − m 2 (t):

σ 2 −2μt σ2
m(t) = m 0 e−μt (m 0 = EX 0 ), M(t) = M0 − e + (M0 = EX 02 ),
2μ 2μ
σ 2 −2μt σ2
D(t) = D0 − e + (D0 = E(X 0 − m 0 )2 = M0 − m 20 ).
2μ 2μ

Functions m(t) and M(t) satisfy the following Ordinary Differential Equations
(ODEs):

ṁ(t) = −μm(t), m(0) = m 0 , (21.20)

[email protected]
304 G. Y. Baghdasaryan et al.

Ṁ(t) = −2μM(t) + σ 2 , M(0) = M0 = EX 02 . (21.21)

The mean m(t) and the second-order moment M(t) also satisfy ODEs (Eqs. 21.20–
21.21), when X 0 has a non-Gaussian distribution, but in this case X (t) is the non-
Gaussian process.
Fokker–Planck–Kolmogorov equation corresponding to SDE (Eq. 21.19) can be
written as

∂ϕ(t, x) ∂ σ 2 ∂ 2 ϕ(t, x)
= μ [xϕ(t, x)] + , ϕ(0, x) = ϕ0 (x), (21.22)
∂t ∂x 2 ∂x2

where ϕ0 (x) is the probability density for the initial wind velocity X 0 . Let X 0 be a
random variable having a standard normal distribution, i.e.,

1
ϕ0 (x) = √ e− 2 x .
1 2

2π

Further, we apply the algorithm for solving analysis problem of diffusions

(Eq. 21.4) by the spectral method.
1. Let {qi0 (t)}i∞0 =0 be Legendre polynomials { P̂i0 (t)}i∞0 =0 and let { pi1 (x)}i∞1 =0 be
Hermite functions { ˆ i1 (x)}i∞=0 , where t ∈ T and x ∈ R, then {ei0 i1 (t, x) =
1

qi0 (t) pi1 (x) = P̂i0 (t)ˆ i1 (x)}i∞,i =0 . Thus,

0 1

2i + 1 
i
tk k!
P̂i (t) = (−1)i−k Ci+k
i
Cii−k k , Cki = ,
T k=0 T i! (k − i)!

i! −x 2 /2  (−1)k (2x)i−2k
i/2
ˆ i (x) =
 √ e , i = 0, 1, 2, . . .
2i π k=0
k! (i − 2k)!

Let us assume that L 0 = L 1 = 32 (truncation orders).

2. The column matrix q(1, 0; 0) with entries qi0 (0) is

2i + 1
q(1, 0; 0) = [ P̂0 (0) P̂1 (0) . . . P̂L 0 −1 (0) ]T , P̂i (0) = (−1)i .
T

3. Spectral characteristics P(2, 2) and P1 (2, 2) of differentiation operators ∂/∂t

and ∂/∂ x are defined by:
 
∂e j0 j1 (t, x)
P(2, 2) = [ pi0 i1 j0 j1 ], pi0 i1 j0 j1 = ei0 i1 (t, x) dxdt
T R ∂t
  
d P̂ j0 (t) d P̂ j0 (t)
= P̂i0 (t) dt ˆ ˆ
i1 (x) j1 (x)dx = P̂i0 (t) dt · δi1 j1 ,
T dt R   T dt
∂e j0 j1 (t, x)
P1 (2, 2) = [ p1i 0 i 1 j0 j1 ], p1i 0 i 1 j0 j1 = ei 0 i 1 (t, x) dxdt
T R ∂x

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 305

  ˆ j (x)  ˆ j (x)
d d
= P̂i 0 (t) P̂ j0 (t)dt ˆ i (x)
 1
dx = δi 0 j0 · ˆ i (x)
 1
dx,
1 dx 1 dx
T R R

where δi0 j0 and δi1 j1 are the Kronecker deltas. Consequently,

P(2, 2) = P(1, 1) ⊗ E(1, 1), P1 (2, 2) = E(1, 1) ⊗ P1 (1, 1),

and

P(2, 2) = P(2, 2) + (q(1, 0; 0) · q T (1, 0; 0)) ⊗ E(1, 1)

= (P(1, 1) + q(1, 0; 0) · q T (1, 0; 0)) ⊗ E(1, 1),
P11 (2, 2) = P12 (2, 2) = E(1, 1) ⊗ P12 (1, 1),

where E(1, 1) is the two-dimensional identity matrix, P(1, 1) is the matrix with
entries

 √
d P̂ j (t) 1 2 (2i + 1)(2 j + 1) for i < j and ( j − i) mod 2 = 1
pi j = P̂i (t) dt =
T dt T 0 otherwise,

and P1 (1, 1) is the matrix with entries

⎧ √
 2 j for j − i = 1
ˆ j (x)
d 1⎨ √
p1i j = − p1 ji = ˆ i (x)
 dx = − 2i for i − j = 1
R dx 2⎩
0 otherwise.

The formulas for entries pi j and p1i j have been derived in [1, 7] using defi-
nitions and properties for Legendre polynomials and Hermite polynomials [23,
24].
4. Spectral characteristics F(2, 2) and G(2, 2) of multiplication operators with
multipliers f (t, x) and g(t, x) are defined by:
 
F(2, 2) = [ f i0 i1 j0 j1 ], f i0 i1 j0 j1 = f (t, x)ei0 i1 (t, x)e j0 j1 (t, x)dxdt
T R
  
= −μ P̂i0 (t) P̂ j0 (t)dt ˆ i1 (x)
x ˆ j1 (x)dx = −μδi0 j0 · xˆ i1 (x)
ˆ j1 (x)dx,
T R R
 
G(2, 2) = [gi0 i1 j0 j1 ], gi0 i1 j0 j1 = g(t, x)ei0 i1 (t, x)e j0 j1 (t, x)dxdt
T R
 
= σ2 P̂i0 (t) P̂ j0 (t)dt ˆ i1 (x)
ˆ j1 (x)dx = σ 2 δi0 j0 δi1 j1 .
T R

Consequently,

F(2, 2) = μ · E(1, 1) ⊗ X (1, 1),

G(2, 2) = σ 2 · E(1, 1) ⊗ E(1, 1) = σ 2 · E(2, 2),

[email protected]
306 G. Y. Baghdasaryan et al.

where E(2, 2) is the four-dimensional identity matrix, X (1, 1) is the matrix with
entries
⎧√
 j + 1 for j − i = 1
1 ⎨√
χi j = χ ji = ˆ i (x)
x ˆ j (x)dx = √ i + 1 for i − j = 1
R 2⎩ 0 otherwise.

This relation for entries χi j has been derived in [7] using definitions and
properties for Hermite polynomials [23, 24].

5. The spectral characteristic A(2, 2) can be expressed as

σ2
A(2, 2) = −P1 (2, 2) · F(2, 2) + · P12 (2, 2).
2
6. The spectral characteristic 0 (1, 0) of the probability density ϕ0 (x) for the initial
wind velocity X 0 is defined by
 
ˆ
1 1 for i = 0
0 (1, 0) = [ϕ0i ], ϕ0i = i (x)ϕ0 (x)d x =  √
R 2 π 0 otherwise
 √
because ϕ0 (x) = (1/ 2 π ) ˆ 0 (x).
7. The solution (2, 0) of the generalized characteristic function equation has the
following form:

(2, 0) = (P(2, 2) − A(2, 2))−1 · (q(1, 0; 0) ⊗ 0 (1, 0)).

8. The approximate solution ϕ(t, x) of Fokker–Planck–Kolmogorov equation is

defined by:

0 −1 L
L  1 −1

ϕ(t, x) ≈ ˆ i1 (x), (t, x) ∈ T × R,

ϕi0 i1 · P̂i0 (t) · 
i 0 =0 i 1 =0

where ϕi0 i1 are entries of the matrix (2, 0).

The approximate solution of the analysis problem for T = 10 s, Vt = 60 m/s,
L t = 1000 m, and σ0 = 1.5 m/s is presented in Fig. 21.1.

21.4 Spectral Method for Analysis of Jump Diffusions

This section structure is similar to previous Sect. 21.3, i.e., it includes Sect. 21.4.1
with the statement of the problem, Sect. 21.4.2 with detailed description of the spec-
tral method for solving Kolmogorov–Feller equation, and Sect. 21.4.3 with numerical
results for the analysis of Dryden wind turbulence model with jumps.

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 307

Fig. 21.1 The probability density ϕ(t, x) of the wind velocity X (t)

21.4.1 Problem Statement

Let X (t) be an Rn -valued random process that satisfies Itô SDE with a compound
Poisson process:

d X (t) = f (t, X (t))dt + σ (t, X (t))dW (t) + d P c (t), X (t0 ) = X 0 , (21.23)

where all notations except P c (t) have been introduced in Sect. 21.3.1.
In Eq. 21.23, P c (t) is the compound Poisson process, which can be defined in
different ways [14, 25, 26]. Let λ(t, x):T × Rn → R+ denote the jump rate (or
intensity) and let ρ(t, δ) denote the probability density for jumps (random increments
of X (t)). These two functions specify Poisson process P(t) so that

Pr(P(t + t) − P(t) = 1|X (t) = x) = λ(t, x)t + o(t)

for small t, and X (τ j ) = X (τ j− ) +  j , where jumps  j ∈ Rn are random vectors

distributed with probability density ρ(τ j , δ), j = 1, 2, . . ., and {τ j } are points of the
Poisson process P(t), τ0 = t0 , i.e.,


P(t)
P (t) =
c
j.
j=1

Functions f (t, x), σ (t, x), λ(t, x), ρ(t, δ) satisfy the conditions for the existence
and uniqueness of the strong or weak solution of SDEs with a compound Poisson
process [25], and E|X 0 |2 < +∞.

[email protected]
308 G. Y. Baghdasaryan et al.

The probability density ϕ(t, x) satisfies Kolmogorov–Feller equation or Kol-

mogorov’s forward equation [25, 26]:

∂ϕ(t, x)  ∂ n
1   ∂2
n n
=− [ f i (t, x)ϕ(t, x)] + [gi j (t, x)ϕ(t, x)]
∂t i=1
∂ xi 2 i=1 j=1 ∂ xi ∂ x j

− λ(t, x)ϕ(t, x) + λ(t, ξ )ρ(t, x − ξ )ϕ(t, ξ )dξ , ϕ(t0 , x) = ϕ0 (x).
Rn
(21.24)

Equation 21.24 can be expressed in the form of Eq. 21.6 for the forward jump
diffusion operator A defined by:

n
∂ 1   ∂2
n n
Aϕ(t, x) = − [ f i (t, x)ϕ(t, x)] + [gi j (t, x)ϕ(t, x)]
i=1
∂ xi 2 i=1 j=1 ∂ xi ∂ x j

− λ(t, x)ϕ(t, x) + λ(t, ξ )ρ(t, x − ξ )ϕ(t, ξ )dξ . (21.25)
Rn

Here, we will suppose that assumptions (a) and (b) from Sect. 21.3.1 are satisfied,
and the assumption similar to (c) with respect to Eq. 21.24 is also satisfied.
Thus, the analysis problem of jump diffusions (Eq. 21.23) is formulated as follows.
Given functions f (t, x), σ (t, x), λ(t, x), ρ(t, δ), defining Itô SDE with a compound
Poisson process, and the probability density ϕ0 (x) of the initial state X 0 , find the
probability density ϕ(t, x).

21.4.2 Spectral Method for Solving Kolmogorov–Feller

Equation

Apply the spectral transform to the left-hand and right-hand sides of Eq. 21.24 using
the linearity property of the spectral transform (see Sect. 21.2.2). Then
 
∂ϕ(t, x)
S
∂t
 n  
∂
=− S [ f i (t, x)ϕ(t, x)]
i=1
∂ xi
 
1 
n n
∂2
+ S [gi j (t, x)ϕ(t, x)]
2 i=1 j=1 ∂ xi ∂ x j
 
− S[λ(t, x)ϕ(t, x)] + S λ(t, ξ )ρ(t, x − ξ )ϕ(t, ξ )dξ . (21.26)
Rn

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 309

We will use notations (i)–(iii) from Sect. 21.3.2 as well new notations:
(iv) (n + 1, n + 1) is the spectral characteristic of the multiplication operator with
the multiplier λ(t, x).
(v) R(n + 1, n + 1) is the spectral characteristic of the linear integral operator R
with the kernel ρ(t, x − ξ ), i.e.,

Rψ(t, x) = ρ(t, x − ξ )ψ(t, ξ )dξ , ψ(t, x) : T × Rn → R.
Rn

Therefore,

S[λ(t, x)ϕ(t, x)]

= (n + 1, n + 1) · (n + 1, 0),
 
S λ(t, ξ )ρ(t, x − ξ )ϕ(t, ξ )dξ
Rn
= R(n + 1, n + 1) · (n + 1, n + 1) · (n + 1, 0),

where spectral characteristics (n + 1, 0) and 0 (n, 0) have been defined by

Eqs. 21.9–21.10.
Thus, Eq. 21.26 reduces to

P(n + 1, n + 1) · (n + 1, 0) − q(1, 0; t0 ) ⊗ 0 (n, 0)


n
=− Pi (n + 1, n + 1) · Fi (n + 1, n + 1) · (n + 1, 0)
i=1
1 
n  n
+ Pi j (n + 1, n + 1) · G i j (n + 1, n + 1) · (n + 1, 0)
2
i=1 j=1
− (n + 1, n + 1) · (n + 1, 0) + R(n + 1, n + 1) · (n + 1, n + 1) · (n + 1, 0), (21.27)

consequently, the spectral characteristic (n + 1, 0) satisfies the equation that coin-
cides with Eq. 21.14, but in this case A(n + 1, n + 1) is the spectral characteristic of
the forward jump diffusion operator A defined by Eq. 21.25:


n
A(n + 1, n + 1) = − Pi (n + 1, n + 1) · Fi (n + 1, n + 1)
i=1

1 
n n
+ Pi j (n + 1, n + 1) · G i j (n + 1, n + 1)
2 i=1 j=1
− (n + 1, n + 1) + R(n + 1, n + 1) · (n + 1, n + 1).
(21.28)

[email protected]
310 G. Y. Baghdasaryan et al.

The spectral characteristic (n +1, 0) is also called the generalized characteristic
function, it satisfies the generalized characteristic function equation (Eq. 21.27).
Thus, the analysis problem of jump diffusions (Eq. 21.23) is reduced to solving
the infinite system of linear algebraic system (Eq. 21.27) with unknown entries of
the spectral characteristic (n + 1, 0). The exact solution of Eq. 21.27 is defined by
Eq. 21.16, the probability density ϕ(t, x) is given by Eq. 21.17 or Eq. 21.18 for the
approximate solution of the analysis problem of jump diffusions.
The algorithm for solving analysis problem of jump diffusions (Eq. 21.23) by the
spectral method is given below (items 1–4 are identical to the algorithm for solving
analysis problem of diffusions (Eq. 21.4) in Sect. 21.3.2):
5. Find the spectral characteristic (n + 1, n + 1) of the multiplication operator
with the multiplier λ(t, x) and the spectral characteristic R(n + 1, n + 1) of the
linear integral operator R with the kernel ρ(t, x − ξ ).
6. Find the spectral characteristic A(n + 1, n + 1) of the forward jump diffusion
operator A using Eq. 21.28.
7. Find the spectral characteristic 0 (n, 0) of the probability density ϕ0 (x) for the
initial state X 0 using Eq. 21.10.
8. Find the solution (n + 1, 0) of the generalized characteristic function equation
(Eq. 21.14) using Eq. 21.16.
9. Find the approximate solution ϕ(t, x) of Kolmogorov–Feller equation by
Eq. 21.18.

21.4.3 Dryden Wind Turbulence Model with Jumps

Consider the modified one-dimensional Dryden wind turbulence model with jumps.
It is given by the linear SDE with a compound Poisson process:

d X (t) = −μ∗ X (t)dt + σ∗ dW (t) + d P c (t), X (0) = X 0 , (21.29)

where t ∈ T = [0, T ], μ∗ and σ∗ are constants that will be specified below.

The compound Poisson process P c (t) is defined by the constant jump rate λ and
the probability density for jumps  j (see Sect. 21.4.1) that equals the probability
density for the initial wind velocity X 0 given in Sect. 21.3.3, i.e.,

n = s = 1, t0 = 0, f (t, x) = −μ∗ x, σ (t, x) = σ∗ , g(t, x) = σ∗2 ,

λ(t, x) = λ, ρ(t, δ) = ϕ0 (δ).

The mean m(t) = EX (t) and the second-order moment M(t) = EX 2 (t) satisfy
the following ODEs:

ṁ(t) = −μ∗ m(t) + λm  , m(0) = m 0 , (21.30)

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 311

Ṁ(t) = −2μ∗ M(t) + σ∗2 + 2λm  m(t) + λM  , M(0) = M0 = EX 02 , (21.31)

where m  and M  are the mean and the second-order moment of jumps. Thus, m  =
0 and M  = 1, because jumps have a standard normal distribution. To compensate
the compound Poisson process with respect to the mean and the second-order moment
the constants μ∗ and σ∗ should be specified as follows:

μ∗ = μ, σ∗2 + λM  = σ 2 (σ∗ = 2μσ02 − λ, 2μσ02 ≥ λ),

where μ , σ and σ0 have been defined in Sect. 21.3.3.

Hence, Eqs. 21.30–21.31 coincide with Eqs. 21.20–21.21, respectively.
Kolmogorov–Feller equation corresponding to SDE (Eq. 21.29) can be written
as:

∂ϕ(t, x) ∂ σ 2 ∂ 2 ϕ(t, x)
= μ∗ [xϕ(t, x)] + ∗
∂t ∂x  2 ∂x2
− λϕ(t, x) + λ ϕ0 (x − ξ )ϕ(t, ξ )dξ , ϕ(t0 , x) = ϕ0 (x). (21.32)
R

Further, we apply the algorithm for solving analysis problem of jump diffusions
(Eq. 21.23) by the spectral method (items 1–4 should be used from the solving
analysis problem for Dryden turbulence wind model (Eq. 21.19) in Sect. 21.3.3).
5. The spectral characteristic (2, 2) of the multiplication operator with the
multiplier λ(t, x) is defined by:
 
(2, 2) = [λi0 i1 j0 j1 ], λi0 i1 j0 j1 = λ(t, x)ei0 i1 (t, x)e j0 j1 (t, x)dxdt
  T R
= λ P̂i0 (t) P̂ j0 (t)dt ˆ i1 (x)
 ˆ j1 (x)dx = λδi0 j0 δi1 j1 ,
T R

where δi0 j0 and δi1 j1 are the Kronecker deltas. Consequently,

(2, 2) = λ · E(1, 1) ⊗ E(1, 1) = λ · E(2, 2),

where E(1, 1) and E(2, 2) are two-dimensional and four-dimensional identity

matrices, respectively.
The spectral characteristic R(2, 2) of the linear integral operator R with the kernel
ϕ0 (x − ξ ) satisfies the following relations:

R(2, 2) = [ri0 i1 j0 j1 ],
  
ri0 i1 j0 j1 = ei0 i1 (t, x) e j0 j1 (t, ξ )ϕ0 (x − ξ )dξ dxdt
T R R
 
= P̂i0 (t) P̂ j0 (t)dt ˆ
i1 (x)  ˆ j1 (ξ )ϕ0 (x − ξ )dξ dx
T R R

[email protected]
312 G. Y. Baghdasaryan et al.

= δi0 j0 ˆ i1 (x)[
 ˆ j1 (x) ∗ ϕ0 (x)]dx,
R

ˆ j1 (x) ∗ ϕ0 (x) is the convolution of  j1 (x) and ϕ0 (x). Therefore,

where 

R(2, 2) = E(1, 1) ⊗ R(1, 1),

where entries

i j = ˆ i (x)[
 ˆ j (x) ∗ ϕ0 (x)]dx
R

of the matrix R(1, 1) can be calculated by a recurrence relation based on properties

for the Hermite polynomials [23, 24]. Here, they are calculated numerically.
6. The spectral characteristic A(2, 2) is expressed as:

σ2
A(2, 2) = −P1 (2, 2) · F(2, 2) + · P12 (2, 2) − (2, 2) + (2, 2) · R(2, 2).
2
The spectral characteristic 0 (1, 0) of the probability density ϕ0 (x) for the initial
wind velocity X 0 , the relation for the solution (2, 0) of the generalized character-
istic function equation and the approximate solution ϕ(t, x) of Kolmogorov–Feller
equation are similar to ones given in Sect. 21.3.3 for the solving analysis problem
for Dryden turbulence wind model (Eq. 21.19) (see items 6–8).
The approximate solution of the analysis problem for T = 10 s, Vt = 60 m/s,
L t = 1000 m, σ0 = 1.5 m/s, and λ = 0.2 is presented in Fig. 21.2. In fact, this solu-
tion has minimal differences compared with the approximate solution of the analysis

Fig. 21.2 The probability density ϕ(t, x) of the wind velocity X (t)

[email protected]
21 Spectral Method for Analysis of Diffusions and Jump Diffusions 313

problem from Sect. 21.3.3. This is a consequence of the selecting of compound

Poisson process parameters so that wind velocities X (t) defined by linear SDEs
(Eqs. 21.19 and 21.29) have the zero means and the equal second-order moments.

21.5 Conclusions

In this chapter, the output processes analysis problem for nonlinear stochastic dynam-
ical control systems is concerned. The results from [4] for diffusions (the spectral
method for solving Fokker–Planck–Kolmogorov equation) are expanded to jump dif-
fusions (the spectral method for solving Kolmogorov–Feller equation). A detailed
description of proposed methods is supplemented by step-by-step algorithms for
solving analysis problem and numerical experiments. The spectral method can also
be used for other problems if mathematical model includes the partial differential
equations [16, 24, 27, 28].

References

1. Solodovnikov, V.V., Semenov, V.V., Peschel, M., Nedo, D.: Design of Control Systems on Dig-
ital Computers: Spectral and Interpolational Methods. Mashinostroenie, Moscow (in Russian),
Verlag Technik, Berlin (in German) (1979)
2. Semenov, V.V., Sotskova, I.L.: The spectral method for solving Fokker–Planck–Kolmogorov
equation for stochastic control system analysis. In: 2nd IFAC Symposium on Stochastic Control,
pp. 503–508. Pergamon Press, Oxford (1987)
3. Rybakov, K.A., Sotskova, I.L.: Spectral method for analysis of switching diffusions. IEEE
Trans. Autom. Control 52(7), 1320–1325 (2007)
4. Panteleev, A.V., Rybakov, K.A.: Analyzing nonlinear stochastic control systems in the class of
generalized characteristic functions. Autom. Remote Control 72(2), 393–404 (2011)
5. Rybin, V.V.: Modeling of Nonstationary Integer-Order and Fractional-Order Control Systems
by Grid-Projection Spectral Method. MAI Publication, Moscow (in Russian) (2013)
6. Kozhevnikov, A.S., Rybakov, K.A.: Analysis of nonlinear stochastic systems with jumps
generated by Erlang flow of events. Open J. Appl. Sci. 3(1), 1–7 (2013)
7. Panteleev, A.V., Rybakov, K.A., Sotskova, I.L.: Spectral Method of Nonlinear Stochastic
Control System Analysis. Vuzovskaya kniga, Moscow (in Russian) (2015)
8. Rybakov, K.A., Rybin, V.V.: Algorithms and software for calculating automated control systems
in the spectral form of mathematical description. In: Modern Science: Theoretical, Practical
and Innovative Aspects of Progress, vol. 2, pp. 171–199. Scientific Cooperation Publication,
Rostov-on-Don (in Russian) (2018)
9. Sokolov, N.P.: Operations on multidimensional matrices. Sov. Math., Dokl. 6, 1115–1118
(1965)
10. Hutson, V., Pym, J.S., Cloud, M.J.: Applications of Functional Analysis and Operator Theory.
Elsevier, Amsterdam (2005)
11. Antosik, P., Mikusiński, J., Sikorski, R.: Theory of Distributions. The Sequential Approach.
Elsevier Scientific, Amsterdam (1973)
12. Øksendal, B.: Stochastic Differential Equations. An Introduction with Applications. Springer,
Berlin (2000)

[email protected]
314 G. Y. Baghdasaryan et al.

13. Cerrai, S.: Second order PDE’s in Finite and Infinite Dimension. A Probabilistic Approach.
Springer, Berlin (2001)
14. Pugachev, V.S., Sinitsyn, I.N.: Stochastic Systems: Theory and Applications. World Scientific,
Singapore (2002)
15. Risken, H.: The Fokker-Planck Equation: Methods of Solution and Applications. Springer,
Berlin (1996)
16. Ladyženskaja, O.A., Solonnikov, V.A., Ural’ceva, N.N.: Linear and Quasi-Linear Equations
of Parabolic Type. AMS, Providence (1968)
17. Cooke, R.G.: Infinite Matrices and Sequence Spaces. MacMillan & Co., London (1950)
18. Lindner, M.: Infinite Matrices and Their Finite Sections: An Introduction to the Limit Operator
Method. Birkhäuser, Basel (2006)
19. Dryden, H.L.: A review of the statistical theory of turbulence. Quart. Appl. Math. 1(1), 7–42
(1943)
20. Dobrolensky, Y.: Flight Dynamics in Turbulent Atmosphere. Mashinostroenie, Moscow (in
Russian) (1969)
21. Flying qualities of piloted aircraft: MIL-HDBK-1797. U.S. Department of Defense, Washing-
ton (1997)
22. Kulikov, V.E.: Forming filter for differentiable turbulent wind simulation. In: Proceedings of
the Moscow Institute of Electromechanics and Automation, vol. 7, pp. 36–42 (in Russian)
(2013)
23. Bateman, H., Erdélyi, A.: Higher Transcendental Functions, vol. 2. McGraw-Hill Book
Company, New York (1953)
24. Korn, G.A., Korn, T.M.: Mathematical Handbook for Scientists and Engineers. Dover
Publication, New York (2000)
25. Øksendal, B., Sulem, A.: Applied Stochastic Control of Jump Diffusions. Springer, Cham
(2005)
26. Hanson, F.B.: Applied Stochastic Processes and Control for Jump-Diffusions: Modeling,
Analysis, and Computation. SIAM, Philadelphia (2007)
27. Baghdasaryan, G., Mikilyan, M.: Effects of Magnetoelastic Interactions in Conductive Plates
and Shells. Springer, Cham (2016)
28. Baghdasaryan, G., Danoyan, Z.: Magnetoelastic Waves. Springer, Singapore (2018)

[email protected]
Chapter 22
Long-Period Lunar Perturbations
in Earth Pole Oscillatory Process: Theory
and Observations

Sergej S. Krylov , Vadim V. Perepelkin and Alexandra S. Filippova

Abstract In this chapter, the dynamic effects of the Earth pole motion in the
celestial-mechanical problem statement as the “deformable Earth–Moon problem in
the gravitational field of the Sun” are discussed. The orbits of the Moon and the Earth–
Moon barycenter are assumed as known and given ones. Combination harmonics in
the Earth pole motion are found and their connection with perturbations caused by
the Moon’s orbit precession is shown. Applying a numerical–analytical approach,
the additional components of the Earth pole motion model were determined in an
explicit form.

22.1 Introduction

Creating dynamic models of the Earth pole motion, which allow to determine its
position on the Earth surface with high accuracy is fundamental in solving a number
of problems in astrometry, navigation, and geophysics.
It is known [1] that the two main components of the Earth pole motion are two
harmonics with periods of 365 and 433 days, respectively, with relatively slowly
varying parameters. Also in the polar motion, a trend and high-frequency oscillations
can be identified that have mostly irregular nature. Oscillations with periods of 365
and 433 days are mainly due to the orbital motion, deformability of the Earth’s figure,
and the influence of mobile geomedia, leading to variations in the Earth dynamic
characteristics—its figure and the vector of the Earth’s own angular momentum.

S. S. Krylov · V. V. Perepelkin (B) · A. S. Filippova

Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
S. S. Krylov
e-mail: [email protected]
A. S. Filippova
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 315

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_22

[email protected]
316 S. S. Krylov et al.

Considering the celestial-mechanical problem statement as a basis for building a

more complex model of the Earth’s rotational motion that includes mobile media, it
is of a significant interest to take into account the influence of the perturbations from
the Earth–Moon system spatial motion on the Earth pole oscillatory process.
High-precision astrometric measurements of the Earth Orientation Parameters
(EOP) [2, 3], such as variations in the Earth pole coordinates, allow us to conclude
that there are interrelated dynamic processes in the Earth–Moon–Sun system, the
description and explanation of which in the scientific literature is mostly missing.
For example, a transformation of the coordinate system can be found, depending on
the annual Earth motion around the Sun and the Chandler wobble (associated with
the Earth’s internal structure and large-scale geophysical phenomena), that result in
the detection of the Earth pole oscillatory process synchronous with the precession
motion of the lunar orbit. This leads to the need for the development in the existing
theory of the deformable Earth motion relative to its center of mass and for the
construction of a refined mathematical model of EOP forecast. The latter can be
implemented by an adequate choice of the model complexity and disturbing factors
that are taken into account. The complexity of the model must correspond to the
measurement accuracy and the duration of the data processing interval [4]. The latter
is achieved by extensively analyzing the basic functions composition, their number
and by adjusting of parameters. The justification of the developed model is carried
out by performing numerical experiments using the least squares method and spectral
correlation analysis.
The aim of this research is to study the effect of lunar–solar long-period dis-
turbances on the Earth pole motion. A mathematical description of the Earth pole
motion model and gravitational-tidal lunar–solar disturbances is given in Sect. 22.2.
In Sect. 22.3, the gravitational-tidal mechanism of the Earth pole motion with a fre-
quency close to the frequency of the lunar orbit precession is discussed. Section 22.4
concludes the chapter.

22.2 Mathematical Description of the Earth Pole

Oscillatory Processes

In Sect. 22.2.1, a model of the pole motion is considered that is based on the dynamic
Euler–Liouville equations and takes into account the gravitational-tidal disturbing
moments and angular momentum of the atmosphere. In Sect. 22.2.2, the tidal lunar–
solar potential is described. Variations in geopotential coefficients that appear due to
the gravitational-tidal perturbations are discussed in Sect. 22.2.3.

[email protected]
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory … 317

22.2.1 Celestial-Mechanical Model of the Earth Pole Motion

Using the developed in [4] dynamic model of the Earth pole oscillations for a long-
term forecast, it is possible to achieve high accuracy of motion approximation up
to 10–12 years (for ~103 –104 measurements quantity) and make a forecast within 6
years that correspond to observation data.
According to [4], the differential equations for the model of Earth pole motion
can be obtained from the dynamic Euler–Liouville equations with a variable inertia
tensor:
d
J ω + ω × J ω = M, ω = ( p, q, r )T , J = J ∗ + δ J, J ∗ = const, (22.1)
dt
 
J ∗ = diag(A∗ , B ∗ , C ∗ ), δ J = δ J (t), δ J    J ∗ , (22.2)

.
M = M S + M L − h −ω × h. (22.3)

Here, J is the matrix of the variable inertia tensor, ω is the angular velocity vector
in the Earthbound coordinate system [1, 4], which is qualitatively and quantitatively
corresponds with the International Terrestrial Reference Frame (ITRF) [1]. The axes
of the chosen coordinate system approximately coincide with the main central axes
of inertia of the “frozen” Earth figure taking into account “equatorial bulge”. It
is assumed that small variations of the inertia tensor δ J include various harmonic
components that resulted from the regular perturbation effects of gravitational diurnal
tides from the Sun and the Moon and, possibly, from other perturbations, for example,
annual, monthly, etc. M S and M L are the disturbing moments from the Sun and the
Moon, respectively, and h is the vector of the angular momentum of the atmosphere.
Due to the smallness of p, q ( p, q  r ), in the first approximation in p, q
from Eqs. 22.1–22.3 after averaging over the spin angle we obtain the system of
equations:

C−B Mp Jqr 2
ṗ + rq = − r , (22.4)
A A A
C−A Mq J pr 2
q̇ − rp = + r , (22.5)
B B B
B−A Mr
ṙ + pq = , (22.6)
C C
where M p , Mq , Mr are the disturbing moments of forces, J pr , Jqr are the Earth
centrifugal moments of inertia. Here, it is assumed that

A = A∗ (1 + α p ), B = B ∗ (1 + αq ), C = C ∗ (1 + αr ),
(22.7)
δ A(ϕ2 ) = A∗ α p , δ B(ϕ2 ) = B ∗ αq , δC(ϕ2 ) = C ∗ αr ,

[email protected]
318 S. S. Krylov et al.

where α p,q,r describes daily tidal “bulge”, ϕ2 is the spin angle.

Then from Eqs. 22.4–22.6 with r = r0 = const, we get the equations for the
components p, q of the Earth pole motion:
 
C ∗ − B∗ C∗ B∗ C ∗ − B∗ −Jqr r 2 + M p
ṗ + rq = − αr + αq + α p rq + ,
A∗ A∗ A∗ A∗ A∗
(22.8)
 
C ∗ − A∗ C∗ A∗ C ∗ − A∗ J pr r 2 + Mq
q̇ − r p = αr − α p − αq r p + . (22.9)
B∗ B∗ B∗ B∗ B∗

Since the daily tidal bulges have a daily mean equal to zero, after averaging over
ϕ2 the equations take the form:

ṗ + N p q = jqr
0 2
r + μ p , p(t0 ) = p0 ,
q̇ − Nq p = − j pr0 2
r + μq , q(t0 ) = q0 , (22.10)
C ∗ −B ∗ C ∗ −A∗
N p = A∗ r0 , Nq = B ∗ r0 , r0 = 7.29 × 10−5 rad/s,
 
where μ p , μq are the perturbing torque–weight ratios of forces, j pr
0
= − J pr /B ∗ ϕ =
  
0
0, jqr = − Jqr /A∗ ϕ = 0 are the tidal “bulges”, and N = N p Nq ∼ = 0.84 ÷ 0.85
cycles per year is the Chandler frequency.
The achievement of a prediction high accuracy of the Earth pole motion is con-
nected on the one hand with taking into account various disturbing factors, and, on
the other hand, with the construction of a generalizing dynamic model that allows
one to analyze subtle effects in the Earth pole oscillatory process on a qualitative
level.

22.2.2 Lunar–Solar Perturbations in the Model of Earth Pole

Motion

Lunar–solar gravitational-tidal forces have the potential, which is represented as

a series of spherical functions [5]. For example, in this case, when the Moon is
assumed as a gravitating point-like body or a sphere, the Moon tidal potential U M
can be represented using Eq. 22.11.
∞

UM = U Mn (22.11)
n=2

The first term of the series (for n = 2) corresponds to zonal, tesseral, and sectorial
lunar tides on the Earth’s surface. The expression U M2 in the tidal potential series
can be represented as:

[email protected]
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory … 319

U M2 = P20 (cos θ )a20 (t) + P21 (cos θ )(a21 (t) cos ϕ + b21 (t) sin ϕ)
+ P22 (cos θ )(a22 (t) cos 2ϕ + b22 (t) sin 2ϕ), (22.12)

where P2m (cos θ ) are the associated Legendre functions, θ and ϕ are the geographic
coordinates of a point, r is the distance between a point and the Earth’s center of
mass. The coefficients a2m (t), b2m (t) depend on time and are determined by the Moon
position relative to the Earth. Denoting θ M , ϕ M as the latitude and longitude of the
Moon, respectively, the decomposition coefficients of Eq. 22.12 can be expressed in
the following form:
2
a20 = 21 g ξ 3 Rr ERE m M
R̄ E M m E
(3 cos2 θ M − 1),
2 3
a21 = −3 8π
5
N212
g ξ 3 Rr E R̄R E mM
mE
cos θ M sin θ M cos ϕ M ,
EM
2 3
b21 = −3 8π
5
N212
g ξ 3 Rr E R̄R E mM
mE
cos θ M sin θ M sin ϕ M ,
EM
2 3 (22.13)
a22 = −3 8π
5
2
N22 g ξ 3 Rr E R̄R E mM
mE
sin2 θ M cos 2ϕ M ,
EM
2 3
b22 = −3 8π
5
2
N22 g ξ 3 Rr E R̄R E mM
mE
sin2 θ M sin 2ϕ M ,
EM
5 (2−m)!
ξ= R̄ E M
RE M
, N2m = (−1)m 4π (2+m)!
,

where R̄ E M is the average distance between the Earth’s center of mass and the Moon,
R E M is the current distance between the Earth’s center of mass and the Moon, R E is
the average Earth radius (R E ∼ = 6.38 × 106 m), m E is the mass of the Earth, m M is
the mass of the Moon, g is the gravitational acceleration.
The classical theory of the lunar motion for the problem “Earth–Moon system
in the gravitational field of the Sun” allows one to take into account a number of
dynamic effects of variations in the Earth axial rotation velocity including also the
main inequalities of the Moon’s motion [5, 6].
For the spatial version of the bounded three-body problem Earth–Moon–Sun, the
perturbed motion equation for the lunar orbit node  M and inclination of I of the
lunar orbit plane to the ecliptic have the form [7]:

3 n 2S
˙M =−
 [1 − cos 2(l M −  M ) − cos 2(l S −  M ) + cos 2λ], (22.14)
4 nM
3 n 2S
I˙ = − sin I [sin 2(l S −  M ) − sin 2(l M −  M ) + sin 2λ]. (22.15)
4 nM

Here, n M , n S are the sidereal mean motion of the Moon and the Sun, respectively,
fluctuations of the angle I occur with the lunar orbit node period 18.61 years, l M , l S
are the mean longitude of the Moon and the Sun, respectively, a M is the semi-major
axis of the Moon orbit, (l M −  M ) is the angle between the Moon and the ascending
node of the lunar orbit, and λ ∼ = (n M − n S )t + λ0 .

[email protected]
320 S. S. Krylov et al.

The tidal potential contains harmonics with different periods [1]. The main com-
ponents (with the largest amplitudes) have periods of a year, half a year, 13.66 and
26.73 days, as well as, less significant 9.1 days and a third of the year. Along with the
main frequencies of the Moon’s motion, there is a stable harmonic with an argument
of 2λ and the period of half of the synodic month, and a high frequency harmonic
with an argument of (2λ + M) and the period of 9.56 days, where M is the mean
Moon anomaly, which is affected by changes in the mean longitude and perigee
displacement. The lunar inequality associated with the argument (2λ − M) is an
eviction with a period of 31.81 days. The presence of these components is due to the
perturbation corresponding to the second zonal harmonic of the tidal potential.
A quasiperiodic lunar perturbation (the precessional motion of the lunar orbit and
small variations in the inclination of its plane) corresponding to Eqs. 22.14–22.15 is
included in the additional tidal potential.
The tidal potential U M2 is expressed by the sum of harmonic terms with
combination frequencies [7]:

1 r2 0 n
U M2 = − κg (1 − 3 cos2 θ ) Ai cos(νi t + ψi0 )
4 RE i

1 r2 1 n
− κg sin 2θ Bi cos(νi t + νϕ t + ψi1 )
2 RE i

1 r2  n
2

− κg sin2 θ Ci cos(νi t + 2νϕ t + ψi2 ),

2 RE i
 3
3 m M RE
κ= = 0.843 × 10−7 , 0 ≤ r < R E . (22.16)
2 m E RE M

Here, t is the Greenwich mean solar time, n 0 , n 1 , n 2 are the number of zonal,
tesseral, and sectorial harmonics, respectively, and coefficients Ai , Bi , Ci are their
amplitudes. The values of νi = ai l˙M + bi l˙S + ci ṗ M + di ṗ S + ei  ˙ M are lin-
ear combinations of the angular parameters l M,S , p M,S ,  M derivations with integer
coefficients. The parameters l M and l S are the mean longitudes of the Moon and the
Sun, respectively, with periods of 27.55 and 365.25 sidereal days. The value of p M is
the mean longitude of the Moon perigee varying with a period of 8.85 years, and p S
is the mean longitude of the Sun perigee changing with a period of 25,700 years. The
parameter  M determines the longitude of the Moon ascending node and it varies
with a period of 18.61 years.

[email protected]
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory … 321

22.2.3 Variations of the Geopotential Coefficients

Due to the presence of perturbations of the tidal potential for the figure of the Earth, an
additional perturbing potential δW (t) arises, which depends on time. With the expan-
sion of the potential δW, the largest term in terms of magnitude is the perturbation
from the second harmonic δW2 :

f m E R 2E
δW2 = Y2 (θ, ϕ), (22.17)
r3
where Y2 (θ, ϕ) is the change in the normalized spherical function, and is expressed
in terms of the coefficients of the second order of the geopotential series [8]:

Y2 (θ, ϕ) = δc20 P20 (cos θ ) + [δc21 cos ϕ + δs21 sin ϕ]P21 (cos θ )
+ [δc22 cos 2ϕ + δs22 sin 2ϕ]P22 (cos θ ). (22.18)

The variations of the geopotential coefficients caused by the perturbations from

the Sun and the Moon are obtained from Eq. 22.16 and have the following form:

n0 n1
δ c20 = ∑a
j =1
0j cos Θ0 j , δ c21 = ∑a
j =1
1j cos Θ1 j ,

n1 n2
δ s21 = ∑a
j =1
1j sin Θ1 j , δ c22 = ∑a
j =1
2j cos Θ 2 j ,

n2
(22.19)
δ s22 = ∑a
j =1
2j sin Θ 2 j ,

3
⎛ R ⎞ R2 m
= χζ g ⎜ E ⎟ E M ,
⎝ REM ⎠ fmE

where i j = {iMj , iSj } are linear combinations of the angles τ0 , l M,S , p M,S , Ω M
due to the Moon and the Sun perturbations, respectively.
The harmonics cos(i j ), when i equals to 0, 1, and 2, correspond to zonal, tesseral,
and sectorial tides, respectively. The value of τ0 = t −l M −l S is the Greenwich mean
lunar time, ζ is the ratio of the vertical tidal displacement of the Earth’s surface to
the displacement of the equipotential surface of the tidal potential, the coefficient χ
equals to 1 for a model of the homogeneous Earth and 0.843 for a real Earth.
The variations of the coefficients of the second zonal harmonic of the geopotential
can be expressed in terms of the variations of the inertia tensor components

[email protected]
322 S. S. Krylov et al.
⎛ ⎞
δ A −δ J pq −δ J pr
δ J = ⎝ −δ J pq δ B −δ Jqr ⎠ (22.20)
−δ J pr −δ Jqr δC

in the following way:

δ A + δ B − 2δC δ J pr δ Jqr
δc20 = , δc21 = , δs21 = ,
2m E R 2E m E R 2E m E R 2E
δB − δA δ J pq
δc22 = , δs22 = . (22.21)
4m E R 2E 2m E R 2E

To further describe the variations of the inertia tensor and geopotential coefficients,
it is convenient to adopt the following notation in the expansion of the inertia tensor
δ Ji j :

(ϕ) (2ϕ)
δ Ji j = δ Ji(t)
j + δ Ji j + δ Ji j ,

(ϕ) (2ϕ)
where δ Ji(t)
j are the intra-annual and interannual variations, δ Ji j , δ Ji j are daily
and semidiurnal variations, respectively.
Intra-day variations δ Ji(ϕ) (2ϕ)
j , δ Ji j contain oscillation components with combina-
tional frequencies of the νi spatial variant of the problem “deformable Earth–Moon
in the gravitational field of the Sun”.

22.3 Gravitational and Tidal Perturbations in the Model

of the Earth Pole Motion

In this section, expressions for lunar–solar gravitational-tidal moments are given in

Sect. 22.3.1. In Sect. 22.3.2, more subtle effects of the disturbed oscillatory process
of the Earth pole are considered in the canonical action-angle variables.

22.3.1 Gravitational-Tidal Lunar–Solar Moment of Forces

Considering the gravitational-tidal perturbations from the Moon and the Sun, we
consider the moments of gravitational forces structure that are included in the right-
hand side of Eqs. 22.1–22.3. The expressions for the components of the moments of
gravitational forces, for example, from the Sun, were obtained in [4, 9–11] and have
the following form:

[email protected]
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory … 323

M pS = 3ω2 [(C ∗ + δC − (B ∗ + δ B))γq γr + δ Jqr (γq2 − γr2 ) + δ J pq γ p γr − δ J pr γ p γq ],

MqS = 3ω2 [(A∗ + δ A − (C ∗ + δC))γ p γr + δ J pr (γr2 − γ p2 ) + δ Jqr γ p γq − δ J pq γq γr ],
MrS = 3ω2 [(B ∗ + δ B − (A∗ + δ A))γ p γq + δ J pq (γ p2 − γq2 ) + δ Jqr γ p γr − δ J pr γq γr ],
ω = ω∗ (1 + e S cos ν S )3/2 ,
γ p = sin θ S sin ϕ S , γq = sin θ S cos ϕ S , γr = cos θ S ,
(22.22)

where ω∗ is a constant determined by gravitational and focal parameters, ν S is a

true anomaly of the Earth when moving in an elliptical orbit with eccentricity e S ,
γ p , γq , γr are the direction cosines of the radius vector in the bound coordinate
system, ψ S , θ S , ϕ S are Euler angles defining the orientation of that coordinate system
relative to the orbital one, where z axis is directed to the attracting center, namely, the
Sun, A∗ , B ∗ , C ∗ are the effective main central moments of inertia with regard to the
“frozen” Earth deformations. They can be calculated with sufficient accuracy. The
coefficients δ A, δ B, δC, δ J pq , δ Jqr , and δ J pr are due to tidal diurnal and semidiurnal
gravitational influences of the Moon and the Sun.
After averaging over the fast variable ϕ S (ϕ S is the angle of Earth’s spin) for M p,q,r
S

simple expressions are obtained:

   
M pS = 3ω02 χ1Sp sin2 θ S + χ2Sp sin θ S cos θ S − δ Jqr ϕS cos 2θ S ,
   
MqS = 3ω02 χ1qS
sin2 θ S + χ2q
S
sin θ S cos θ S + δ J pr ϕS cos 2θ S , (22.23)
 
MrS = 3ω02 χ1rS sin2 θ S + χ2rS sin θ S cos θ S .

The values of χ1Sp,1q,1r , χ2Sp,2q,2r are the tidal coefficients due to semidiurnal and
diurnal tides. They are obtained by averaging coefficients at sin2 θ S and sin θ S cos θ S
in the components of the gravitational moment from the Sun by ϕ S :
   
χ1Sp = − δ Jqr sin2 ϕ S ϕS − 21 δ J pr sin 2ϕ S ϕS ,
 
χ2Sp = 21 (δC − δ B) cos ϕ S ϕS + δ J pq sin ϕ S ϕS ,
   
χ1q
S
= δ J pr cos2 ϕ S ϕS + 21 δ Jqr sin 2ϕ S ϕS ,
  (22.24)
χ2qS
= 21 (δ A − δC) sin ϕ S ϕS − δ J pq cos ϕ S ϕS ,
 
χ1rS = 21 (δ B − δ A) sin 2ϕ S ϕS − δ J pq cos 2ϕ S ϕS ,
   
χ1rS = 21 δ Jqr sin ϕ S ϕS − δ J pr cos ϕ S ϕS .

The values of the coefficients χ1Sp,1q,1r , χ2Sp,2q,2r are small values that can be
determined based on observational data.
The expressions of the direction cosines of the Sun’s radius vector are written
using Euler kinematic equations, which specify the orientation of the related axes
relative to the orbital coordinate system [4]:
. .
p = ψ S sin θ S sin ϕ S + θ S cos ϕ S + ω0 (ν S )(sin ψ S cos ϕ S

[email protected]
324 S. S. Krylov et al.

+ cos ψ S sin ϕ S cos θ S ),

. .
q = ψ S sin θ S cos(ϕ S ) − θ S sin ϕ S + ω0 (ν S )(− sin ψ S sin ϕ S +
+ cos ψ S cos ϕ S cos θ S ),
. .
r = ϕ S + ψ S cos θ S − ω0 (ν S ) cos ψ S sin θ S ,
.
ω0 (ν S ) = ν S = ω∗ (1 + e S cos ν S )2 , (22.25)

where e S = 0.0167 is the orbit eccentricity, ν S (t) is the true anomaly. Conducting
some transformations in Eq. 22.23 and integrating them in the first approximation,
we obtain the following expressions:

r = r 0 , ϕ S ≈ r 0 t + ϕ 0 , ν S = ω∗ t + ν S0 , cos θ S (ν S ) = a cos ν S ,
sin θ S cos θ S = b cos ν S + d cos 3ν S + . . . , (22.26)
◦
where the initial values are θ S0 = 66 33 , ψ S0 ∈ [0, 2π ], coefficients for cos ν S are in
the intervals a ∈ [0.4, 1], b ∈ 0.4, 3π 4
and depend on the initial values θ S0 , ψ S0 , and
the coefficient d is much less than one [4].
In a simplified version of the problem for the stationary lunar orbit (Keplerian
orbit) the disturbing moment from the Moon has a structure similar to Eq. 22.22.
The perturbing moment from the Moon M L leads to small-scale tidal changes in the
speed of the Earth axial rotation and the position of the Earth pole at relatively short
time intervals. In particular, in order to select a component with a period of 9.1 days
in the expansion of the lunar moment M L it is necessary to keep the third harmonic
in the expression sin θ M cos θ M , i.e.,

sin θ M cos θ M = b(θ M

0
, ψM
0
) cos ν M + d cos 3ν M + . . . ,

where ν M is true anomaly of the Moon.

In the general case, when the Moon orbit performs a known precessional-nutation
motion, the perturbing moment from the Moon is determined by the expression:

M L = −m M R E M × ∇W. (22.27)

Here, R E M is the radius is the vector of the center of mass of the Moon relative
to the Earth’s center of mass, W is the Earth external gravitational potential. The
decomposition of a geopotential in a series of spherical functions has the form [8]:
∞ n  
f m E   R E (n+1)
W = Pnm (cos θ M )(cnm cos mϕ M + snm sin mϕ M ).
R E n=0 m=0 r
(22.28)

As noted above, the largest harmonic in the geopotential decomposition is the sec-
ond one. Then the expressions for the components of the gravitational-tidal moment

[email protected]
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory … 325

take the form:

3 f m E m M R 2E 
M pL = −c20 sin 2θ M sin ϕ M + s21 (3 cos2 θ M − 1)
2 r3
+ sin2 θ M [s21 cos 2ϕ M − c21 sin 2ϕ M ]
+ 2 sin 2θ M [s22 cos ϕ M − c22 sin 2ϕ M ]),
3 f m E m M R 2E 
MqL = c20 sin 2θ M cos ϕ M − c21 (3 cos2 θ M − 1)
2 r3
+ sin2 θ M [c21 cos 2ϕ M + s21 sin 2ϕ M ]
−2 sin 2θ M [c22 cos ϕ M + s22 sin 2ϕ M ]),
3 f m E m M R 2E
MrL = (sin 2θ M [c21 sin ϕ M − s21 cos ϕ M ]
2 r3 
+ 4 sin2 θ M [c22 sin 2ϕ M − s22 cos 2ϕ M ] . (22.29)

∗ ∗
Coefficients of the geopotential c2m , s2m consist of constant coefficients c2m , s2m
and periodic tidal variations of δc2m , δs2m . After multiplying the tidal components in
L
M p,q,r , the expressions containing long-period terms and terms resulting in diurnal
librations, are obtained.
Expressions of the components of the δ M p,q,r L
additional gravitational-tidal
moment from the Moon are


n1 
n0 
n2
δ M pL = −2δc20 b0Mj sin 1Mj + δs21 b0Mj cos 0Mj + δs21 b2Mj cos 2Mj
j=1 j=1 j=1


n2 
n1 
n1
− δc21 b2Mj sin 2Mj − 4δc22 b1Mj sin 1Mj + 4δs22 b1Mj cos 1Mj ,
j=1 j=1 j=1


n1 
n0 
n2
δ MqL = 2δc20 b0Mj cos 1Mj − δc21 b0Mj cos 0Mj + δs21 b2Mj sin 2Mj
j=1 j=1 j=1


n2 
n1 
n1
+ δc21 b2Mj cos 2Mj − 4δc22 b1Mj cos 1Mj − 4δs22 b1Mj sin 1Mj ,
j=1 j=1 j=1


n1 
n1
δ MrL = δc21 b1Mj sin 1Mj + δs21 b1Mj cos 1Mj
j=1 j=1


n2 
n2
+ δc22 b2Mj sin 2Mj + δs21 b2Mj cos 2Mj , (22.30)
j=1 j=1

which can be decomposed into the sum of the combinational harmonics of the spatial
variant of the problem “deformable Earth–Moon in the gravitational field of the Sun”.

[email protected]
326 S. S. Krylov et al.

Now in the expressions of the gravitational moment from the Moon (Eq. 22.30) let
us take into account variations of the geopotential decomposition coefficients caused
only by solar disturbances, and vice versa: in expressions of the gravitational moment
from the Sun—the variations of the geopotential decomposition coefficients caused
only by lunar disturbances. For example, substituting the variations δciMj caused by
the Moon influence in the expressions of additional moment of gravitational forces
from the Moon will result in δ M p,q,r
L
= 0 for the elastic Earth model, and for real
Earth δ M p,q,r
L
will be a small value.
The total lunar–solar additional gravitational-tidal moment, for example, for
δ MqSL = δ MqS + δ MqL will be


n1 
n0 
n1
δ MqSL = 2δc20
S
b0Mj cos 1Mj − δc21
S
b0Mj cos 0Mj + 2δc20
M
b0S j cos 1S j
j=1 j=1 j=1


n0 
n2 
n2
− δc21
M
b0S j cos 0S j + δs21
S
b2Mj sin 2Mj + δc21
S
b2Mj cos 2Mj
j=1 j=1 j=1


n2 
n2 
n1
+ δs21
M
b2S j sin 2S j + δc21
M
b2S j cos 2S j − S
4δc22 b1Mj cos 1Mj
j=1 j=1 j=1


n1 
n1 
n1
− 4δs22
S
b1Mj sin 1Mj − 4δc22
M
b1S j cos 1S j − 4δs22
M
b1S j sin 1S j ,
j=1 j=1 j=1
(22.31)

and in the case of the elastic Earth δ M p,q,r

SL
= 0.
Taking into account the mantle viscosity the tidal bulge from the attracting body
in its diurnal motion will be “ahead” of the tidal bulge phase for the absolutely
elastic Earth model, which will lead not to the zero value, but to the negative one
for the additional gravitational-tidal moment, and to the secular slowing down of the
Earth’s rotation [12, 13]. In this case, the total moment (Eq. 22.31) will be nonzero.
The latter will lead to small-scale librations with combination harmonics. Among
these harmonics those associated with the spatial motion of the Earth–Moon system
can be distinguished.

22.3.2 The Oscillatory Process of the Earth Pole

at the Frequency of the Moon’s Orbit Precession

The actual problem of studying the irregular behavior of the main components of the
Earth pole oscillations is currently poorly understood. Of a significant interest, there
are studies aimed at establishing the geophysical and celestial-mechanical reasons for
such behavior of the Chandler and 1-year components, and the construction of refined

[email protected]
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory … 327

prediction models for EOP required for solving high-precision satellite navigation
problems [8].
In modern works, for example, [14, 15] the problem of the how astronomical fac-
tors influence on the planetary scaled geophysical processes and fluctuations of EOP
is considered. In particular, the spectral analysis results of EOP measurement data
show the presence of harmonics with frequencies close to those of the Earth–Moon
system motion. It should be noted that astronomical factors affecting geophysical pro-
cesses can be identified within the framework of a generalizing celestial-mechanical
model of the Earth motion relative to its center of mass.
Now let us assume that the perturbing factors are caused by the deformations of the
viscoelastic mantle (the outer deformable layer) due to gravitational perturbations
caused by the influence of the Moon and the Sun. The main goal of constructing
a model of the Earth pole perturbed motion is to identify the parameters Chandler
wobble and to predict its trajectory. It is of interest to study the motion at various
time intervals available in observations from intra-year intervals to the period of pole
beats.
The perturbed Chandler wobble dynamics of the instantaneous axis is associated,
in particular, with a change in the angle δ2 [4], which determines the change in the
amplitude of Chandler wobble. The angular variable δ2 is the angle between the Earth
figure axis and the angular momentum vector.
For a more detailed analysis of the perturbed Chandler wobble of the asymmet-
ric Earth pole, it is convenient to use the canonical action-angle variables [4]. The
perturbed Routh functional of the problem in question can be represented using
Eq. 22.32 as in [3].

R = R0 + ε R1 (I1 , I2 , I3 , w1 , w2 , w3 , u, u̇) + ε2 . . . (22.32)

Here, ε R1 is the perturbing functional due to gravitational tides, u is the displace-

ment vector of the viscoelastic medium points in the Earth’s mantle, I1 , I2 , I3 are
the action variables (I2 is the value of the deformable Earth angular momentum, and
I1 , I3 are its projections on the axis close to the figure axis of the terrestrial and
celestial geocentric coordinate systems, respectively), and angles w1 , w2 , w3 are the
canonically conjugate variables corresponding to the phase of the Earth pole motion,
the Earth’s own rotation angle, and the precession angle, respectively, ε > 0 is a
small dimensionless parameter characterizing the relative magnitude of the disturbing
factors in Eq. 22.32.
It can be shown [3] that the equation for δ2 in the action-angle variables will be
 
1 1 1 + κ 2 sn 2 (w1 , λ) ∂ R1 ∂ R1
δ̇2 = ε − − δ2 . (22.33)
I2 sin δ2 κ∗ dn(w1 , λ) ∂w1 ∂w2

Here, I2 is the magnitude of the Earth own angular momentum, the angular variable
w1 determines the phase of the Earth pole motion, so that ẇ1 has the meaning of
the Earth pole instant frequency turning around the “mean pole” (the “mean pole”
determines the long-period motion of the Earth pole), the phase angle w2 is associated

[email protected]
328 S. S. Krylov et al.

with the Earth axial rotation, sn, dn are Jacobi elliptic functions, λ ∼ 10−2 is a small
parameter, the values of κ, κ∗ are the determined by the inertia moments of the Earth:

C ∗ (A∗ − B ∗ ) 
κ2 = ∗ ∗ ∗
, κ∗ = 1 + κ 2 .
A (B − C )

From Eq. 22.32, we obtain the equation for δ2 :

 2  
λ 2 + κ2 4
δ̇22 ≈ 2 sin 2w1 + cos 2w1 · [a1 (δ1 , i, σ∗ ) cos  M
κ I 2 κ∗ I2

+ a2 (w3 , δ1 , i, σ∗ ) cos 2 M + a2s (w3 , δ1 , i, σ∗ ) sin 2 M ,
c
(22.34)

where a1 , a2s , a2c are unknown coefficients determined by the parameters of the Earth
deformable layer, the values of which can be estimated from EOP astrometric mea-
surements data, δ1 is the angle between the angular momentum vector and the axis
of the geocentric celestial coordinate system that is orthogonal to the ecliptic (its
change is related to the precession and nutation of the Earth), w3 is the Earth pre-
cession angle, i is the angle of inclination of the Moon’s orbit plane to the ecliptic,
 M is the longitude of the ascending node of the Moon’s orbit, that is defined by
Eq. 22.14, σ∗ is a small dissipation coefficient of the viscoelastic layer of the Earth’s
mantle.
Equation 22.34 leads to the presence in the oscillatory process of the Earth pole
of combinational harmonics associated with the precession of the Moon’s orbit. Let
us denote a(t) as the variable amplitude of the pole motion, which varies with a
period of 6.45 years. Then the equations of motion of the Earth pole in the first
approximation in σ∗ can be represented as

x p = a(t) cos w1 + δ2 cos w1 ,

y p = a(t) sin w1 + δ2 sin w1 . (22.35)

The additional terms in Eq. 22.35 are the Chandler wobble modulated by
harmonics with the frequency of the precession of the lunar orbit.
Now, having the obtained decomposition of the gravitational-tidal moment, we
will seek a solution to Eq. 22.10 in the form of Eq. 22.35 with known perturbations
of μ p , μq . To do this, they can be determined using the measurement data from
International Earth Rotation and Reference Systems Service (IERS) of the trajectory
of the Earth pole from Eq. 22.10. Putting them into Fourier series, we leave only the
considered combination harmonics w1 +  M , w1 −  M and define the corresponding
pole oscillations in an explicit form. In Fig. 22.1, the additional components of model
(Eq. 22.35) are shown.
Next, perform the transformation of the coordinates of the Earth pole:

[email protected]
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory … 329

Fig. 22.1 Comparison of x p , y p obtained by processing IERS observations data (discrete

points) with a curve obtained using lunar ephemeris: a x p values, b y p values

⎛ ⎞
   
ηp ξp xp
δϕ = arctan⎝ ⎠, = (wh − w1 ) (w1 ) .
ξp + ξ p2 + η2p ηp yp
(22.36)

The following notation is introduced into Eq. 22.36:  is the plane rotation matrix,
a0 is the average value of the amplitude of the pole oscillations around the “mean
pole” (that is without the trend component), cx , c y are the set the position of the
“mean pole” and contain constants, secular terms and variations with periods longer
than 6 years, ẇh − ẇ1 = νT is the frequency of the 6-year amplitude modulation of
the Earth pole oscillatory motion, x p , y p are the second terms on the right-hand
side of Eq. 22.35 for x and y, respectively.
In Fig. 22.2, the polar angle δϕ is compared with the oscillations of the intersection
point between the equator and the lunar orbit. The units of the oscillation amplitudes
are radians, and τ is the time in standard years. In new coordinate system, it is possible
to illustrate synchronous oscillations of the Earth pole with the precessional motion

[email protected]
330 S. S. Krylov et al.

Fig. 22.2 Comparison of the polar angle variations δϕ (bottom line) with the oscillations along the
equator of the point of intersection of the lunar orbit and the equator (upper line), constructed using
the lunar ephemeris

of the lunar orbit and to determine the regular component of δϕ phase variation,
which makes it possible to use lunar ephemeris when predicting additional terms in
the model of Earth pole motion.
To do this, making the inverse transformation
   
xp −1 −1 cos δϕ − 1
= ξ p2 + η2p  (w1 )  (wh − w1 ) , (22.37)
yp sin δϕ

we get additional terms in the oscillations of the Earth pole leading to amplitude
modulation of its main motion. In Fig. 22.1, the calculated curves are compared with
those extracted from the data from IERS. In this case, expressions depending on
lunar ephemeris are used in the obtained numerical-analytical model.

22.4 Conclusions

The developed celestial-mechanical model that takes into account the gravitational-
tidal lunar–solar perturbations allows to assume the presence of a specific oscillatory
process in the Earth pole motion. The perturbation from the Moon leads to the
additional combinational harmonics modulated by a harmonic at a frequency close
to the lunar orbit precession frequency, and the perturbation from the Sun makes this
process nonstationary. More precisely, it can be said that the process will be quasi-
stationary until the ratio of the amplitudes of the Chandler and annual components
goes through 1 (becomes more than one or, on the contrary, less). Using the data
processing of IERS observations over the Earth pole trajectory, a method is proposed
for identifying the discovered oscillations.

[email protected]
22 Long-Period Lunar Perturbations in Earth Pole Oscillatory … 331

Acknowledgements This work was carried out within the state task no. 9.7555.2017/BCh.

References

1. International Earth Rotation and Reference Systems Service—IERS Annual Reports, http://
www.iers.org. Last accessed 09 Oct 2019
2. Perepelkin, V.V., Rykhlova, L.V., Filippova, A.S.: Long-period variations in oscillations of the
Earth’s pole due to lunar perturbations. Astron. Rep. 63(3), 238–247 (2019)
3. Markov, YuG, Perepelkin, V.V., Filippova, A.S.: Analysis of the perturbed Chandler wobble of
the Earth pole. Dokl. Phys. 62(6), 318–322 (2017)
4. Akulenko, L.D., Markov, YuG, Rykhlova, L.V.: Motion of the Earth’s poles under the action
of gravitational tides in the deformable Earth model. Dokl. Phys. 46(4), 261–263 (2001)
5. Munk, W.H., MacDonald, G.J.F.: The Rotation of the Earth. Cambridge University Press, New
York (1961)
6. Sidorenkov, N.S.: The Interaction Between Earth’s Rotation and Geophysical Processes. Wiley-
VCH Verlag GmbH and Co. KGaA (2009)
7. Schubert, G.: Treatise on Geophysics, vol. 3. Elsevier, Geodesy (2007)
8. Xu, G.: Sciences of Geodesy—I: Advances and Future Directions. Springer, Berlin, Germany
(2010)
9. Kumakshev, S.A.: Gravitational-tidal model of oscillations of Earth’s poles. Mech. Solids 53(2),
159–163 (2018)
10. Kumakshev, S.A.: Model of oscillations of Earth’s poles based on gravitational tides. In:
Karev, V., Klimov, D., Pokazeev, K. (eds.) Physical and Mathematical Modeling of Earth
and Environment Processes. PMMEEP, pp. 157–163 (2017)
11. Klimov, D.M., Akulenko, L.D., Kumakshev, S.A.: The main properties and peculiarities of the
Earth’s motion relative to the center of mass. Dokl. Phys. 59(10), 472–475 (2014)
12. Zlenko, A.A.: A celestial-mechanical model for the tidal evolution of the Earth–Moon system
treated as a double planet. Astron. Rep. 59(1), 72–87 (2015)
13. Zlenko, A.A.: The force function of two rigid celestial bodies in Delaunay–Andoyer variables.
Astron. Rep. 60(1), 174–181 (2016)
14. Bizouard, C., Remus, F., Lambert, S., Seoane, L., Gambis, D.: The Earth’s Variable Chandler
Wobble Astronomy and Astrophysics, vol. 526, pp. A106.1–A106.4 (2011)
15. Zotov, L., Bizouard, C., Shum, C.K.: A possible interrelation between Earth rotation and
climatic variability at decadal time-scale. Geodesy Geodyn. 7(3), 216–222 (2016)

[email protected]
Chapter 23
Application of Modified Fireworks
Algorithm for Multiobjective
Optimization of Satellite Control Law

Andrei V. Panteleev and Alexander Yu. Kryuchkov

Abstract Application of modified metaheuristic global optimization algorithm

“fireworks” to solve problems of multiobjective optimization is discussed. All objec-
tives are numeric and have the same importance. A possible solution to the problem
is a vector of real numbers. Generally, the solution to the problem is a set of Pareto
optimal possible solutions. The application of the algorithm of finding an effective
programmed control in the problem of stabilizing a satellite in a circular orbit is
considered. We study a problem with the fixed right end and known finite time,
where control is a function in the class of piecewise constant functions and satisfies
the constraints. The control must minimize the values of the quadratic objective and
the objective of describing fuel consumption simultaneously. Searching of control is
divided into two stages. At the first stage, the optimization problem is solved for each
of the objectives with penalties. The values of the penalties are selected to satisfy
the terminal constraints. At the second stage, the penalties found are used to solve a
multiobjective optimization problem.

23.1 Introduction

In the modern world, designing new technical systems is becoming more complex.
Requirements for systems in various areas are growing and many different factors
have to be taken into account [1, 2]. These factors may reflect the opposite goals. In
such circumstances, the search for an acceptable solution becomes a difficult task.
As a rule, the solution should be a compromise between conflicting requirements.

A. V. Panteleev (B) · A. Yu. Kryuchkov

Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation
e-mail: [email protected]
A. Yu. Kryuchkov
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 333

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_23

[email protected]
334 A. V. Panteleev and A. Yu. Kryuchkov

When designing various systems, there is always a question about the choice of
parameters that affect the final result. To select the optimal parameters, we can for-
mulate an optimization problem. Each solution to the problem is a set of parameters
most suitable from the point of view of the selected objectives. Comparison of solu-
tions is based on the values of the objectives, and each objective can correspond to
a particular factor and reflects the degree of influence of the solution on this factor.
For example, the smaller objective value corresponds to higher savings in energy
consumption. If one objective is selected, then one objective optimization problem is
obtained. Many different algorithms have been developed for this problem solution
[3]. When there are several objectives, the problem becomes more complicated. In a
problem with one objective, an optimal solution is a solution that minimizes or max-
imizes the objective value. In a multiobjective problem, it is possible to give several
concepts of the optimal solution. For example, if the objectives are ordered by impor-
tance, then to compare solutions, we can compare the values of objectives based on
the lexicographic ordering. The solution will be optimal if it gives the minimum value
to the most important objective. Optimal solutions can be infinitely many. There are
definitions of optimality according to Geoffrion [4] and Pareto [5]. When choosing
the definition of optimality, a specific system of preferences is also fixed. Of course,
one can try to reflect the influence of a solution on various factors in one objective, for
example, using the scalarization of objectives. It simplifies the optimization problem
because it becomes a single-objective optimization problem. There is a difficulty:
it is hard to describe the influence of a decision on all factors within one objective,
since it is possible to lose some of the information about the links between factors
describing different goals. If a multiobjective problem reduced to a single-objective,
one has certain disadvantages, for example, there are Pareto-optimal solutions that
cannot be found through the scalarization of objectives [6].
In control theory, when finding the optimal control, one objective is usually con-
sidered [7, 8]. In modern conditions, it is necessary to solve problems when one
objective is not enough to describe all the requirements. Thus, a problem with several
objectives appears. It can be solved as a problem with one objective if we combine
several objectives into one. The process of combining the objectives itself is not
unambiguous. Therefore, it is necessary to develop algorithms for solving multiob-
jective optimization problems, which allow to find a set of Pareto-optimal solutions,
but not requiring scalarization. Some ideas for solving problems in control theory
with several objectives were considered in [9].
In the general case, it is impossible to describe Pareto optimal solutions analyti-
cally. Therefore, the developing of algorithms for finding an approximate solution is
needed. An approximate solution is a finite set of solutions close to the exact solution.
Various algorithms based on different ideas were proposed to solve the problems.
In the chapter, a modification of the one-objective optimization “fireworks” algo-
rithm and its application to finding programmed control, which stabilizes a satellite,
is considered. The algorithm belongs to metaheuristic algorithms and does not guar-
antee to find an exact solution. The control is sought in the class of piecewise constant
functions, due to the linearity of the dynamic system. The initial problem is reduced
to the parametric optimization problem, in which the solution is a vector of real

[email protected]
23 Application of Modified Fireworks Algorithm … 335

numbers. Numerical global optimization algorithms may be required to solve the

problem.
A new problem was formed with modified initial objectives to solve the multiob-
jective problem and find the control law. Penalties were added to each objective in the
new problem. They ought to be found when solving optimization problems for each
objective. The next step is to solve the multiobjective problem with modified objec-
tives, but with fixed penalties. Since the problem is solved numerically, therefore,
the control is only suboptimal. The solution can be considered acceptable depending
on the requirements in applied problems, where finding the exact analytical solution
is impossible.
The chapter is organized as follows. Section 23.2 provides the model of a dynamic
system. Section 23.3 discusses the idea of solving a multiobjective control problem.
Section 23.4 considers a modification of the “fireworks” algorithm. Section 23.5
presents the results of numerical experiments, and Sect. 23.6 gives the conclusions.

23.2 Dynamic System Model

Consider the problem of the stability of the stationary motion of a satellite relative
to an axis orthogonal to its circular orbit plane [10]. If we neglect the gravitational
moments in comparison with the control moments, then Euler equations are replaced
by the following:
⎧
⎪
⎪ ω̇x = a × ωz + u1
⎨
ω̇z = −a × ωx + u2
. (23.1)
⎪
⎪ ω (t ) = ωx0
⎩ x 0
ωz (t0 ) = ωz0

where ωx , ωz are the angular rotation velocity of the satellite, a = const, t0 is the start
time. The controls u1 , u2 equal to the ratio of the corresponding torque relative to the
axis Cx , Cz by the moment of inertia I = Ix = Iz , Cx , Cz are the axes associated
with the center of mass of the satellite, where Ix , Iz are the moments of inertia
about the axis Cx , Cz . Jet engines control rotation, so the control is constrained:
|ui | ≤ hi , i = 1, 2.
The control ought to minimize two objectives:

T 
I1 = ∫ u12 (t) + u22 (t) dt → min ,
t0 u1 ∈UN1 ,u2 ∈UN2
T
I2 = ∫[|u1 (t)| + |u2 (t)|]dt → min ,
t0 u1 ∈UN1 ,u2 ∈UN2

j
where T is the known terminal time, UN , j = 1, 2 are the sets of admissible controls:

[email protected]
336 A. V. Panteleev and A. Yu. Kryuchkov
N −1
j
UN = bi × χ (t − ti ):|bi | ≤ hj ,
i=0

where t0 < t1 < · · · < tN −1 < T are the switching times, N is the number of switch-
ings, χ (·) is the Heaviside step function. Terminal constraints ωx (T ) = 0, ωz (T ) = 0
should be satisfied. Minimizing the objective I2 corresponds to the task of optimizing
fuel consumption.

23.3 Sketch of Solution

It is necessary to formulate a multiobjective optimization problem to find a con-

trol that minimizes two objectives simultaneously. Since the problem is solved
numerically, not analytically, the following approach is proposed:
1. Modify the definitions of the objectives to take into account the terminal
constraints by introducing penalties λi > 0, i = 1, . . . , 4:

T 
J1 = ∫ u12 (t) + u22 (t) dt + λ1 × ωx2 (T ) + λ2 × ωz2 (T ) → min ,
t0 u1 ∈UN1 , u2 ∈UN2
T
J2 = ∫[|u1 (t)| + |u2 (t)|]dt + λ3 × ωx2 (T ) + λ4 × ωz2 (T ) → min .
t0 u1 ∈UN1 , u2 ∈UN2

2. Select the number of control switchings N as the parameter describing the control
laws u1 (t), u2 (t).
3. Find penalties values λi > 0, i = 1, . . . , 4 and controls u1 (t), u2 (t). The
terminal conditions must be satisfied with certain accuracy: ωx (T ) ≈ 0, ωz (T ) ≈
0. The solution is encoded as a numerical vector x ∈ R2N +4 : the control u1 (t) is
encoded first, then u2 (t) and the penalties λi , i = 1, . . . , 4. We use the assumption
that control takes constant values over equal lengths of time:

xi , t ∈ [h × (i − 1); h × i),
u1 (t) =
xN , t ∈ [h × (N − 1); h × N ]

xj , t ∈ h × (j − N − 1); h × (j − N )),
u2 (t) =
x2N , t ∈ [h × (N − 1); h × N ]

where xi ∈ [−h1 ; h1 ], i = 1, . . . , N −1, xj ∈ [−h2 ; h2 ], j = N +1, . . . , 2N −1, h =

T /N . The penalties λi , i = 1, . . . , 4 are the last two components of the vector
x:x2N +i = λi for the first objective J1 and x2N +i = λi+2 for the second objective
J2 , i = 1, 2. The optimization problems are

J1 (x) → min, (23.2)

x∈D

[email protected]
23 Application of Modified Fireworks Algorithm … 337

J2 (x) → min, (23.3)

x∈D

where D = {x ∈ R2N +2 | − h1 ≤ xi ≤ h1 , i = 1, . . . , N , −h2 ≤ xi ≤ h2 , i =

N + 1, . . . , 2N ,a1 ≤ x2N +1 ≤ b1 , a2 ≤ x2N +2 ≤ b2 }.
4. After finding the penalty parameters for each objective, they are fixed, and the
multiobjectives optimization problem is solved:

J1 (x)
J (x) = → min, (23.4)
J2 (x) x∈D1

where D1 = {x ∈ R2N | − h1 ≤ xi ≤ h1 , i = 1, . . . , N , −h2 ≤ xi ≤ h2 , i =

N + 1, . . . , 2N }.

23.4 Solution of Multiobjective Problem

Hereinafter, Sect. 23.4.1 includes a mathematical formulation of the problem.

Section 23.4.2 contains a general description of algorithm. Section 23.4.3 provides
a detailed description.

23.4.1 Multiobjective Optimization Problem

The multiobjective optimization problem with constraints is considered. It is assumed

that all objectives have the same importance and reducing the value of one objective
with fixed values of the other objectives is preferable. A vector of real numbers
represents the solution (all objectives are numeric):
⎛ ⎞
f1 (x)
⎜ ⎟
F(x) = ⎝ ... ⎠ → min, (23.5)
x∈D
fm (x)

where m ≥ 2 is the number of objectives, D is the set of acceptable solutions,

fj : D → R, j = 1, . . . , m :
 
D = x ∈ Rn |ai ≤ xi ≤ bi , ai < bi , i = 1, . . . , n .

Need to find an approximation of the set of Pareto optimal solutions. We will need
to give several definitions for the further description of the algorithm.

[email protected]
338 A. V. Panteleev and A. Yu. Kryuchkov

Definition 1 Let F(x) ∈ Rm :F(x) = (f1 (x), . . . , fm (x))T . It is a vector of objectives

of a solution x ∈ D.
   
Definition 2 Let F 1 = F x1 , F 2 = F x2 . They are vectors of objectives of x1 ∈
D, x2 ∈ D.F 1 dominates F 2 :F 1 ≺ F 2 if ∀i ∈ {1, . . . , m}, Fi1 ≤ Fi2 and ∃ j ∈
{1, . . . , m}:Fj1 < Fj2 .

Definition 3 A solution x1 ∈ D is preferable than x2 ∈ D:x1 ≺ x2 ⇔ F(x1 ) ≺

F(x2 ).
 
Definition 4 A set P = {x ∈ D| x ∈ D:F x ≺ F(x)} is a set of Pareto optimal
solutions.

Definition 5 A set F(P) = {F(x)|x ∈ P} is Pareto front.

The result of solving the problem expressed by Eq. 23.5 will be a finite set of
solutions, in which each element is close to some element of P.

23.4.2 Modification of Multiobjective Fireworks Algorithm

The modification of “fireworks” algorithm [11] is based on an imitation of the fire-

works process. The firework is accompanied by a cloud of luminous fragments filling
the vicinity of an exploding charge. This process is associated with a local search
procedure in optimization problems.
Each firework volley determines the transition from one iteration of the search
to another (from one generation of solutions to another). Points (decisions) in the
set of feasible solutions D are determined for the first volley in an amount NP. An
explosion occurs, generating debris scattering from the points of the explosion. The
radius of the explosion is determined for each point separately.
Next is the process of forming a new generation of solutions. Non-dominated
sorting of solutions is performed based on their vectors of objectives. Let I = {xp |xp ∈
D, p = 1, . . . , NP} be a set of solutions on the current iteration, where NP ≥ 1. The
result of the non-dominated sorting is a partition of the set I into k disjoint subsets
Fi , i = 1, . . . , k, 1 ≤ k ≤ NP,k is the number of the last subset in the partition
mentioned below.


k
I= Fi , Fi ∩ Fj = ∅, i = j
i=1
   
F1 = x ∈ I |x ∈ I : F x ≺ F(x)
..
.
 

l−1  
l−1
 

Fl = x ∈ I \ Fi x ∈ I \ Fi : F x ≺ F(x)

i=1 i=1

[email protected]
23 Application of Modified Fireworks Algorithm … 339

..
.

k−1
Fk = I \ Fi
i=1

In other words, non-dominated sorting is a repeating procedure to take the pre-

ferred solutions. In the first step, the preferred solutions are selected from I . Further,
these preferred solutions are removed from I , and the procedure is repeated to the
remainder.
The solutions corresponding to the points of the explosion and the resulting debris
are selected from the sets F1 , F2 , . . . , Fk until the number of selected solutions is
equal to NP. If in some subset of solutions there are more than necessary to take, then
part of the solutions are chosen randomly based on the distances to other solutions
in this subset in the objective space.
The search process stops when a specified number of iterations is reached.
Fast non-dominated sorting algorithms were proposed in [12, 13]. The algorithm
having complexity O(n logm−1 n) in the worst case is presented in [13].
Software implementations of the non-dominated sorting algorithm are available
on GitHub for the following programming languages: C# [14], Python [15], and
Java [16].

23.4.3 Algorithm

Step 1. Choose parameters:

• Number of charges at each iteration NP ∈ N.
• Parameter m > 0 controlling the number of debris.
• Parameters smin , smax ∈ N, smin ≤ smax are the minimum and maximum number
of debris for each charge.
• Maximum explosion amplitude Amax > 0.
• Maximum number of iterations itermax ∈ N.
Step 2. Let iter = 1. Generate NP solutions in the set of feasible solutions
D,I iter = x1,1 , . . . , xNP,1 :

p,1
xi = ai + Urand (0; 1) × (bi − ai ),

where i = 1, . . . , n, p = 1, . . . , NP,Urand (0, 1) is continuous uniform random

variable on [0, 1].
Step 3. Non-dominated sorting of I iter . Partition I iter into F1 , . . . , Fl , 1 ≤ l ≤ NP.
Step 4. Explosion and debris generation.
Step 4.1. For each p = 1, . . . , NP calculate:
1. Subset number q: xp,iter ∈ Fq .

[email protected]
340 A. V. Panteleev and A. Yu. Kryuchkov

2. Number of debris:
   
|F |
sp,iter = m × log2 1 + ql × 1 − NPq ,
⎧  
⎨ smin ifsp,iter  ≤ smin
ŝp,iter = smax if sp,iter ≥ smax ,
⎩  p,iter 
s otherwise

where [·] is the integer part, l is the number of the last subset in the partition, ŝp,iter
is the number of debris generated by the explosion at xp,iter .
Step 4.2. Determine the position of debris. For each p = 1, . . . , N , find the
position of debris with numbers s = 1, . . . , ŝp,iter :

1. Find q : xp,iter ∈ Fq .
2. Let x̂p,iter,s = xp,iter .
3. For each debris with number s :

a. Let ξ = Urand (0, 1).

b. Randomly select dimensions:

n̂ = [n × ξ ].

c. If ξ < 0.5, then apply the first algorithm for determining the position of the
debris:
(1) Calculate the magnitude of the explosion:
 
 q   Fq 
A p,iter
= Amax × log2 1 + × .
l NP

(2) For each coordinate number i selected from n̂:

i. Calculate the displacement:

hsi = Ap,iter × Urand (−1, 1).

ii. Calculate the debris coordinate value:

iii. hsi = Ap,iter × Urand (−1, 1)
d. If ξ ≥ 0.5, then apply the second algorithm for determining the position of
the debris:
(1) For each coordinate number i selected from n̂ :
p,iter,s p,iter
x̂i = xi × Nrand (1, 1),

where Nrand (1; 1) is the normal random variable, 1 is the mean, and 1 is the
standard deviation.
Step 4.3. Checking the boundary of the set of acceptable solutions D. For each
p = 1, . . . , NP:

[email protected]
23 Application of Modified Fireworks Algorithm … 341

1. For each s = 1, . . . , ŝp,iter check:

p,iter,s
If x̂i ∈
/ [ai ; bi ], i = 1, . . . , n, then:

p,iter,s
p,iter,s Urand (ai ; 0.5 × (ai + bi ))if x̂i < ai
x̂i = .
Urand (0.5 × (ai + bi ), bi )otherwise
 
2. Append x̂p,iter,s to the I iter : I iter = I iter ∪ x̂p,iter,s .

Step 5. Creating a new population.

Step 5.1. Non-dominated sorting I iter : F1 , . . . , Fl , where l is the number of the
last subset in the partition I iter . Let iter = iter + 1, I iter = ∅.
Step 5.2. Find umin :
  u 
 
 
umin = min u :  Fi  ≥ NP .
1≤u≤l  
i=1

u 
 min  u
min −1
umin

If  Fi  = NP, then I iter = Fi , K = ∅, otherwise I iter = Fi ,
i=1 i=1 i=1
  −1
umin
K = Fumin . If umin = 1 and Fumin  > NP, then Fi is empty set. All solutions
i=1
fall into the set K.
Step 5.3. For each point xw ∈ K, calculate R(xw ) that is the sum of distances to
other points and p(xw ) that is the probability of explosion:

R(xw ) = ρ(F(xw ), F(xq )),
xq ∈K
R(xw )
p(xw ) = 
R(xq )
,
xq ∈K

where ρ(·, ·) is Euclidean distance.

Step 5.4. Randomly choose from K the set of points (solutions) in the number of
NP − I iter  based on probability p(xw ) and append it to the set I iter .
Step 5.5. If iter ≤ Itermax , then go to Step 3, otherwise I iter is an approximate
solution.
It is not necessary to perform non-dominated sorting on Step 3 after the first iter-
ation. Information about the partition can be taken after Step 5.1. Such optimization
can reduce computational costs.

23.5 Numerical Experiments

The following parameters were selected in the numerical √experiments: t0 =

0, ωx (t0 ) = 0.01, ωz (t0 ) = 0.1, a = 0.00007292123518 × 3, T = 1.5, h1 =

[email protected]
342 A. V. Panteleev and A. Yu. Kryuchkov

h2 = 10, a1 = a2 = 5000, b1 = b2 = 30, 000 for the problem (Eq. 23.2),

a1 = a2 = 7000, b1 = b2 = 50, 000 for the problem (Eq. 23.3).
The problem was solved 5 times for each objective using three algorithms: “big
bang-big crunch”, “fireworks”, “grenade explosion method” [17–19]. The values of
the penalties λi , i = 1, . . . , 4 were chosen based on the results of the solutions. The
five best solution results by objectives are listed in tables mentioned below.
The best results were selected using the non-dominated sorting. The selection
were based on the three objectives: the final value of the objective in the problem and
absolute values |ωx (T )|, |ωz (T )|. The links between the parameters of the algorithm
and the results obtained are made through the Id column.
First, the problem expressed by Eq. 23.2 was solved. The number of switches
N = 8, 10, 15. The parameters of the algorithms are listed in Tables 23.1, 23.2,
and 23.3. They were the same for all values of N . Table 23.4 shows the results
of the problem solution expressed by Eq. 23.2, while Table 23.5 shows the results
of the problem solution provided by Eq. 23.3. BBBC name means the “big bang-
big crunch” algorithm, FW is “fireworks”, GEM is “grenade explosion method” in
Tables 23.4 and 23.5. Parameters α and β are given in the “fireworks” algorithm [19]
instead of Smin and Smax . Last of them are used for the algorithm of “fireworks” in
the latest version of the software that implements all of the above algorithms [20].
The purpose of the parameters remains the same, but the range has changed. If α and
β limited the number of debris by multiplying by m, then parameters Smin , Smax take
values from the set of the natural numbers and limit the number of debris explicitly.
The system of differential equations (Eq. 23.1) was integrated using the fourth
order Runge–Kutta algorithm with a step by time is equal 0.006. The results of the

Table 23.1 Parameters of BBBC algorithm

Id itermax NP α β
1 500 900 0.4 0.5
2 400 300 0.2 0.1
3 600 200 0.5 0.5
4 400 300 0.5 0.9
5 1000 200 0.5 0.5

Table 23.2 Parameters of FW algorithm

Id itermax NP Smin Smax Amax m
1 300 100 10 30 2.5 20
2 900 60 10 50 1 20
3 150 20 10 50 2 20
4 130 30 10 50 9 20
5 200 200 10 50 0.9 2

[email protected]
 Table 23.3 Parameters of GEM algorithm
Id Ngr Nq itermax Rt−initial mmin mmax Psin pts Rrd Ndm
1 1 100 500 8.485 0.1 0.9 5 0.8 10 1
2 1 100 500 8.485 0.1 0.9 5 0.8 150 1
3 2 75 250 3 0.1 0.9 5 0.8 10 1
23 Application of Modified Fireworks Algorithm …

4 2 50 900 3 0.1 0.9 5 0.8 150 1

5 2 100 600 3 0.1 0.9 5 0.8 200 1

[email protected]
343
344 A. V. Panteleev and A. Yu. Kryuchkov

Table 23.4 Results of solution problem expressed by Eq. 23.2

Id Algorithm N J1 ωx (T ) ωz (T ) λ1 λ2
1 BBBC 8 0.007604 −5.8801E−5 −5.0031E−5 6020.491 13228.306
1 BBBC 8 0.007686 −4.7398E−5 1.2918E−4 9156.242 10734.697
5 BBBC 8 0.007785 7.9434E−5 3.3302E−5 7995.033 8813.139
2 FW 8 0.015567 −2.0242E−5 −9.1201E−5 11831.137 15142.621
1 FW 8 0.024222 1.6105E−5 −4.2268E−4 9151.26 7584.314
1 FW 8 0.024643 −1.0386E−4 2.7879E−5 13696.774 28017.862
4 GEM 8 0.556426 −1.3949E−4 9.907E−5 6736.098 5083.258
2 GEM 8 0.612563 5.9419E−5 −2.7177E−4 7162.864 5493.837
2 GEM 8 0.638406 5.0606E−5 −1.4113E−5 5179.473 6840.771
1 BBBC 10 0.008995 −8.3586E−6 3.4592E−5 21240.517 5288.833
3 BBBC 10 0.011183 2.475E−4 2.7719E−6 5115.562 5502.813
5 BBBC 10 0.009019 1.0354E−4 −4.8657E−5 7962.194 7662.075
2 FW 10 0.015147 8.9975E−5 6.9862E−5 15727.856 10001.139
2 FW 10 0.016794 3.6849E−5 −1.6649E−4 6051.207 22030.654
2 FW 10 0.016905 1.8238E−4 6.1224E−5 10031.998 15250.165
2 GEM 10 0.736677 −5.123E−4 2.1993E−4 6161.452 5426.359
2 GEM 10 1.440087 −6.0246E−4 −1.5134E−4 5054.838 9103.089
2 GEM 10 1.531033 −3.3369E−5 −9.9396E−5 9821.737 5487.826
5 BBBC 15 0.00811 −5.7092E−6 −4.5947E−5 7372.927 5518.607
5 BBBC 15 0.075976 7.9654E−5 1.0107E−5 14552.243 18685.553
1 BBBC 15 0.009598 −5.5698E−5 −1.6512E−4 18727.757 5001.869
2 FW 15 0.033248 7.2419E−5 −1.3879E−4 24315.069 5036.152
2 FW 15 0.034547 −8.2631E−5 −6.5246E−5 15625.11 11576.285
2 FW 15 0.0392 −5.9139E−5 −5.1834E−4 29219.92 9410.349
2 GEM 15 1.384027 −1.2177E−3 −6.8241E−6 9554.58 5195.936
5 GEM 15 1.464559 −3.8043E−5 5.2589E−4 5557.229 6332.396
2 GEM 15 1.764458 −1.4677E−4 −2.849E−4 5904.814 7288.434

solution for different values of the number of control switchings N = 8, 10, 15 are
listed below in Tables 23.1, 23.2, and 23.3.
Values of the control for N = 8 and the problem (Eq. 23.2) are presented in
Table 23.6.
Values of the control for N = 8 and the problem (Eq. 23.3) are presented in
Table 23.7.
The following values of the penalties were selected based on results of problems
solutions expressed by Eqs. 23.2 and 23.3. They are listed in Table 23.8.

[email protected]
23 Application of Modified Fireworks Algorithm … 345

Table 23.5 Results of solution problem expressed by Eq. 23.3

Id Algorithm N J2 ωx (T ) ωz (T ) λ3 λ4
1 BBBC 8 5.160511 1.2955E−4 −2.9862E−4 8617.706 30659.494
5 BBBC 8 7.793863 6.4492E−4 −2.808E−4 14617.899 12032.31
1 BBBC 8 9.523325 6.8663E−4 −1.2991E−5 25216.057 49486.752
2 FW 8 0.110096 4.4293E−5 3.5374E−5 7398.211 23228.186
2 FW 8 0.110418 1.7397E−5 −1.8073E−5 11301.078 26449.15
2 FW 8 0.110385 −5.1218E−5 3.8691E−5 23538.786 7480.885
1 GEM 8 4.272012 3.2021E−3 6.9987E−3 12832.101 15461.899
1 GEM 8 4.274881 5.0938E−3 −5.7929E−4 7186.127 13161.372
5 GEM 8 4.432162 1.3039E−5 −3.0173E−5 10185.876 7386.642
5 BBBC 10 4.714524 −1.8E−4 −3.7626E−4 8597.463 12434.566
1 BBBC 10 6.252822 −1.0633E−4 −2.3524E−4 8816.932 10224.87
1 BBBC 10 6.468312 −2.3239E−6 −6.1775E−4 11334.467 8479.597
2 FW 10 0.110101 3.0433E−5 2.4713E−5 7544.905 49805.834
2 FW 10 0.110188 2.3023E−6 8.403E−6 15014.967 12347.239
2 FW 10 0.110671 5.8278E−5 −5.6617E−6 7719.76 9349.415
1 GEM 10 3.836151 −2.6624E−3 1.428E−3 20703.729 11021.126
5 GEM 10 4.11288 −4.4488E−6 −2.0417E−4 8004.571 8102.188
2 GEM 10 4.280879 2.9121E−4 7.374E−5 7000.455 13175.199
1 BBBC 15 5.827723 2.4443E−4 −2.5118E−4 22853.125 9440.198
1 BBBC 15 7.344124 −7.1157E−4 6.4369E−5 11760.751 31003.365
1 BBBC 15 7.746715 −1.9718E−4 −4.4188E−4 26858.854 13654.348
2 FW 15 0.10872 5.7673E−5 6.8015E−5 11968.315 8067.865
2 FW 15 0.112241 1.2268E−5 −5.6859E−6 17727.378 46942.4
2 FW 15 0.112687 1.4441E−4 −4.8097E−5 10930.355 29076.976
2 GEM 15 3.409267 −4.4071E−4 1.4896E−4 9431.92 7369.179
2 GEM 15 4.315806 2.599E−4 −5.5759E−7 7721.063 7441.143
4 GEM 15 4.417192 −9.7125E−6 −2.986E−5 7574.725 8838.705

Table 23.6 Values of the

t u1 (t) u2 (t)
control
0 −0.00151397 −0.02574828
0.1875 −0.00132441 −0.09630083
0.375 −0.00467101 −0.08455635
0.5625 −0.01891687 −0.06688008
0.75 0.00663547 −0.06081657
0.9375 −0.0212785 −0.06917927
1.125 −0.00204489 −0.04615376
1.3125 −0.01087784 −0.08420601

[email protected]
346 A. V. Panteleev and A. Yu. Kryuchkov

Table 23.7 Values of control

t u1 (t) u2 (t)
0 −0.00014781 −0.50271218
0.1875 −9.58977344E−5 −0.01173894
0.375 −0.05076043 −0.00054373
0.5625 −0.00075133 2.06553592E−9
0.75 −2.08503817E−6 −0.01919405
0.9375 3.78612438E−7 5.06952171E−7
1.125 −0.0006604 −0.00032599
1.3125 −9.60528929E−6 1.27306602E−6

Table 23.8 The penalties for various values of the number of switchings
Penalty N =8 N = 10 N = 15
λ1 6020.49142775647 21240.5168329197 7372.92658580968
λ2 13228.3055394025 5288.83298590383 5518.60746104122
λ3 7398.210744558931 15014.9665908026 17727.3778533999
λ4 23228.1859177857 12347.2391412813 46942.3996244755

The problem expressed by Eq. 23.4 was solved with fixed penalties from
Table 23.8 and different values of the number of switchings. The algorithm
parameters are presented in Table 23.9.
The values of the objectives are presented in Fig. 23.1 for N = 8.
The values of the objectives are presented in Fig. 23.2 for N = 10.
The values of the objectives are presented in Fig. 23.3 for N = 15.

Table 23.9 Parameters of the modified “fireworks” algorithm

Id Itermax NP Smin Smax Amax m
1 900 300 5 20 0.25 5
2 500 500 5 20 0.9 15
3 600 600 10 30 0.8 10
4 1000 200 5 20 0.2 10
5 2000 300 5 20 0.1 4
6 2000 300 5 20 0.05 4

[email protected]
23 Application of Modified Fireworks Algorithm … 347

Fig. 23.1 The values of the

objectives for N = 8

Fig. 23.2 The values of the

objectives for N = 10

23.6 Conclusions

The numerical algorithm for solving multiobjective optimization problems is pro-

posed. The main difficulties in developing multiobjective optimization algorithms
are the convergence and uniformity of solution distribution along Pareto front.
Pareto dominance is based on a comparison of vectors of objectives. In practice,
it can cause difficulties. Rounding errors complicate the comparison. For example,
let two vectors of objectives are equal, except for two components differ from each
other by a small value. Does one vector of objectives dominate another? We can
get different results depending on the choice: the approximate Pareto front, where it
is not there, or not to find part of Pareto front. The solution may be the application

[email protected]
348 A. V. Panteleev and A. Yu. Kryuchkov

Fig. 23.3 The values of the

objectives for N = 15

of dominance rules based on fuzzy logic. For example, an approach for solving
multiobjective optimization problems based on fuzzy logic is proposed in [21].
An essential challenge is the solution of multiobjective optimization problems
with a large number of objectives (three or more). NSGA-3 algorithm for solving
problems with a large number of objectives (from three to fifteen) is demonstrated in
[22]. However, the authors did not provide a software implementation of their algo-
rithm. The algorithm has different implementations. Comparison between NSGA-2
and NSGA-3 algorithms is presented [23]. It has been shown that NSGA-3 is not
always better than NSGA-2.
Nevertheless, the advances in solving multiobjective optimization problems allow
us to consider applied problems. For example, the problem of finding programmed
control was considered. The control is sought in the class of piecewise constant
functions. The initial problem became the parametric optimization problem.
A new problem was formed with modified initial objectives to solve the multi-
objective problem. The penalties were added to each objective. The algorithm for
solving multiobjective optimization problems is proposed. It is numerical and does
not guarantee finding the exact solution. The found control is only suboptimal.

References

1. Arias-Montano, A., Coello, C.A.C., Mezura-Montes, E.: Multiobjective evolutionary algo-

rithms in aeronautical and aerospace engineering. IEEE Trans. Evol. Comput. 16(5), 662
(2012)
2. Rangaiah, G.: Multi-Objective Optimization. World Scientific, Singapore (2017)
3. Nocedal, J., Wright, S.: Numerical Optimization. Springer, New York (2006)
4. Geoffrion, A.M.: Proper efficiency and the theory of vector maximization. J. Math. Anal. Appl.
22, 618–630 (1968)

[email protected]
23 Application of Modified Fireworks Algorithm … 349

5. Collette, Y., Siarry, P.: Multiobjective Optimization. Springer, Berlin Heidelberg, Berlin (2004)
6. Talbi, E.-G.: Metaheuristics: From Design to Implementation. Wiley, Hoboken, NJ, USA (2009)
7. Panovskiy, V., Panteleev, A.: Meta-heuristic interval methods of search of optimal in average
control of nonlinear determinate systems with incomplete information about its parameters. J.
Comput. Syst. Sci. Int. 56(1), 52–63 (2017)
8. Panteleev, A., Pis’mennaya, V.: Application of a memetic algorithm for the optimal control
of bunches of trajectories of nonlinear deterministic systems with incomplete feedback. J.
Comput. Syst. Sci. Int. 57(1), 25–36 (2018)
9. Prasad, U., Sarma, I.: Multiobjective optimal control problems: game cooperative solution
according to Nash-Harsani. Autom. Remote Control 6, 99–105 (1975)
10. Babadzanjanz, L.K., Pototskaya, I.Y.: Control in Mechanical Systems with Expenditure
Criteria. Izd-vo SPbSU, Saint-Petersburg (in Russian) (2003)
11. Tan, Y., Zhu, Y.: Fireworks algorithm for optimization. In: Tan, Y., Shi, Y., Tan, K.C. (eds.)
Advances in Swarm Intelligence, pp. 355–364. Springer, Berlin Heidelberg, Berlin, Heidelberg
(2010)
12. Fortin, F.-A., Grenier, S., Parizeau, M.: Generalizing the improved run-time complexity algo-
rithm for non-dominated sorting. In: Proceeding of the Fifteenth Annual Conference on Genetic
and Evolutionary Computation Conference. ACM Press, New York, USA (2013)
13. Buzdalov, M., Shalyto, A.: A provably asymptotically fast version of the generalized jensen
algorithm for non-dominated sorting. In: Bartz-Beielstein, T., Branke, J., Filipič, B., and Smith,
J. (eds.) 13th International Conference on Parallel Problem Solving from Nature—PPSN XIII,
pp. 528–537. Springer International Publishing, Cham (2014)
14. KernelA/nds. C# implementation of the non-dominated sorting. https://github.com/
KernelA/nds. Last accessed 17 Oct 2019
15. KernelA/nds-py. A Python implementation of the non-dominated sorting. https://github.com/
KernelA/nds-py. Last accessed 17 Oct 2019
16. Mbuzdalov/non-dominated-sorting. This repo contains implementations of algorithms for non-
dominated sorting and a benchmarking suite. https://github.com/mbuzdalov/non-dominated-
sorting. Last accessed 17 Oct 2019
17. Ahrari, A., Atai, A.: Grenade explosion method—a novel tool for optimization of multimodal
functions. Appl. Soft Comput. 10(4), 1132–1140 (2010)
18. Erol, O., Eksin, I.: A new optimization method: big bang-big crunch. Adv. Eng. Softw. 37(2),
106–111 (2006)
19. Panteleev, A., Kryuchkov, A.: Metaheuristic optimization methods for parameters estimation
of dynamic systems. Civ. Aviat. High Technol. 20(2), 37–45 (2017)
20. KernelA/eoptimization-library. The library for constrained optimization. Implemented three
algorithms: big bung-big crunch, Fireworks, Grenade explosion. https://github.com/KernelA/
eoptimization-library. Last accessed 17 Oct 2019
21. Jamwal, P., Abdikenov, B., Hussain, S.: Evolutionary optimization using equitable fuzzy sorting
genetic algorithm (EFSGA). IEEE Access. 7, 8111–8126 (2019)
22. Deb, K., Jain, H.: An evolutionary many-objective optimization algorithm using reference-
point-based nondominated sorting approach. Part I: solving problems with box constraints.
IEEE Trans. Evol. Comput. 18(4), 577–601 (2014)
23. Ishibuchi, H., Imada, R., Setoguchi, Y., Nojima, Y.: Performance comparison of NSGA-II and
NSGA-III on various many-objective test problems. In: 2016 IEEE Congress on Evolutionary
Computation, pp. 3045–3052 (2016)

[email protected]
Chapter 24
Approximate Filtering Methods
in Continuous-Time Stochastic Systems

Konstantin N. Chugai, Ivan M. Kosachev and Konstantin A. Rybakov

Abstract The goal of this chapter is to consider algorithms based on the parti-
cle method for solving the optimal filtering problem for nonlinear continuous-time
stochastic observation systems not only by the minimum mean squared error esti-
mate, but also by the maximum a posteriori estimate. Particle filters are proposed on
the basis of Duncan–Mortensen–Zakai equation, as well as, on the basis of robust
Duncan–Mortensen–Zakai equation. To find the mode of the conditional distribu-
tion approximately, Edgeworth series is used for the conditional probability density
expansion. This approach allows to reduce significantly the computation time in
contrast to finding the mode by estimating the conditional probability density, for
example, the histogram or kernel estimations.

24.1 Introduction

Methods and algorithms for solving the optimal filtering problem have a lot of prac-
tical applications [1–9], such as the radio signal detection from a noise, navigation,
and telemetry information processing for moving objects, parameter identification,
etc.
We suggest filtering algorithms based on the particle method for nonlinear
continuous-time stochastic observation systems for the unbiased estimate with a

K. N. Chugai
Research Institute of the Belarusian Armed Forces, Minsk 220103, Republic of Belarus
e-mail: [email protected]
I. M. Kosachev
Military Academy of the Republic of Belarus, 220, pr. Nezavisimosti, Minsk 220057, Republic of
Belarus
e-mail: [email protected]
K. A. Rybakov (B)
Moscow Aviation Institute (National Research University), 4, Volokolamskoe shosse, 125993
Moscow, Russian Federation
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 351

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_24

[email protected]
352 K. N. Chugai et al.

minimum mean squared error and the maximum a posteriori estimate. For obtaining
the maximum a posteriori estimate, we apply Edgeworth series [10] for the expansion
of marginal conditional probability densities. This approach allows to estimate the
mode of the conditional distribution approximately, since a partial sum of Edgeworth
series is used for marginal conditional probability densities (conditional probabil-
ity densities for each component of the state) instead of the conditional probability
density of the state vector. However, this approach has an important advantage such
that it significantly reduces the computation time in contrast to finding the mode by
consistently estimating the conditional probability density.
Traditionally, the particle method is associated with Duncan–Mortensen–Zakai
equation (the equation for the unnormalized conditional probability density of the
state) [11]. Here, we use not only this equation, but also robust Duncan–Mortensen–
Zakai equation (the equation for the unnormalized conditional probability density of
the special vector state) [12]. As a rule, the robust Duncan–Mortensen–Zakai equation
has been studied for stationary observation models [12–15]. In this chapter, the
robust Duncan–Mortensen–Zakai equation for nonstationary observation models is
concerned. Some preliminaries have been obtained in [16]. According to the proposed
form in contrast [17], the optimal filtering problem for the nonstationary case can be
solved by applying the particle method [18–21].
The rest of this chapter is structured as follows. Section 24.2 provides the statement
of the optimal filtering problem. Equations for the conditional probability density
and conditional central moments of the state are given in Sect. 24.3. New filtering
algorithms based on the particle method for nonlinear continuous-time stochastic
observation systems are described in Sect. 24.4. Section 24.5 presents the conclusions
for this chapter.

24.2 Optimal Filtering Problem

Consider a signal observation model described by Itô Stochastic Differential

Equations (SDEs) [22, 23]:

d X (t) = f (t, X (t))dt + σ (t, X (t))dW (t), X (t0 ) = X 0 , (24.1)

dY (t) = c(t, X (t))dt + ζ (t)d V (t), Y (t0 ) = Y0 = 0, (24.2)

where X ∈ Rn is the state, Y ∈ Rm is the observation, t ∈ T = [t0 , T ], f (t, x) :

T × Rn → Rn is the n-dimensional function, σ (t, x) : T × Rn → Rn×s is the
(n × s)-dimensional matrix function, c(t, x) : T × Rn → Rm is the m-dimensional
function, ζ (t) : T → Rm×d is the (m × d)-dimensional matrix function such that the
symmetric matrix η(t) = ζ (t)ζ T (t) is nonsingular, i.e., det η(t) = 0 for any t ∈ T.
Then, W (t) and V (t) are the standard s-dimensional and d-dimensional Wiener
processes, respectively, X 0 is the initial state with a probability density ϕ0 (x) such

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 353

that E|X 0 |2 < ∞, where E is the expectation or mean. The initial state X 0 and Wiener
processes W (t), V (t) are independent.
Functions f (t, x), σ (t, x), c(t, x), and ζ (t) are given, they satisfy the conditions
on the existence and uniqueness of the solution of SDEs [22], i.e., there exist two
positive constants c1 and c2 such that for all (t, x), (t, x  ) ∈ T × Rn

| f (t, x)|2 + |σ (t, x)|2 + |c(t, x)|2  c1 (1 + |x|2 ),

| f (t, x) − f (t, x  )|2 + |σ (t, x) − σ (t, x  )|2 + |c(t, x) − c(t, x  )|2  c2 |x − x  |2 ,

where


n 
n
|x| =
2
xi2 , | f (t, x)| =
2
f i2 (t, x),
i=1 i=1

n 
s 
m
|σ (t, x)|2 = σil2 (t, x), |c(t, x)|2 = c2j (t, x).
i=1 l=1 j=1

Such conditions can be weakened, especially for functions f (t, x) and c(t, x). In
fact, it is sufficient to use weakened existence and uniqueness conditions for SDEs
solution from [24, 25].
The optimal filtering problem is to find an estimate X̂ (t) given the observations
Y0t = {Y (τ ), τ ∈ [t0 , t]} so that X̂ (t) = ψ(t, Y0t ), where the function ψ(t, ·) satisfies
the following condition:

EΠ (E(t)) → min ∀t ∈ T, (24.3)

ψ(t,·)

in which E(t) = X (t) − X̂ (t) is the estimation error and Π (ε) is the loss function
[26].
If Π (ε) = εT Lε, where L is the (n ×n)-dimensional positive semidefinite matrix,
i.e., Π (ε) is the quadratic loss function, then

X̂ MMSE
(t) = ψ MMSE
(t, Y0t ) = E[X (t)|Y0t ] = x p(t, x|Y0t )d x,
Rn

and, if Π (ε) = 1 − δ(ε), where δ(ε) is the Dirac delta function, i.e., Π (ε) is the
simple loss function, then

X̂ MAP (t) = ψ MAP (t, Y0t ) = arg maxn p(t, x|Y0t ). (24.4)
x∈R

In the relations given above, p(t, x|Y0t ) is the conditional probability density of
the state X [2, 19, 23]. Equation 24.3 defines the unbiased estimate X̂ MMSE (t) with
a minimum mean squared error, and Eq. 24.4 defines the maximum a posteriori

[email protected]
354 K. N. Chugai et al.

estimate X̂ MAP (t), i.e., the mode of the conditional distribution with the probability
density p(t, x|Y0t ).

24.3 Equations for Conditional Probability Density

The optimal estimate X̂ (t) defined by Eqs. 24.3–24.4 is expressed in terms of the con-
ditional probability density p(t, x|Y0t ). This density satisfies Stratonovich–Kushner
equation [27, 28]:

∂ p(t, x|Y0t )   
= A p(t, x|Y0t ) + λ(t, x, Ẏ (t)) − λ(t, X (t), Ẏ (t)) p(t, x|Y0t ),
∂t
p(t0 , x) = ϕ0 (x), (24.5)

where A is the forward diffusion operator [22]:

n
∂
A p(t, x|Y0t ) = − [ f i (t, x) p(t, x|Y0t )]
i=1
∂ x i

1   ∂2
n n
+ [gi j (t, x) p(t, x|Y0t )]
2 i=1 j=1 ∂ xi ∂ x j
1
= − ∇ T ( f (t, x) p(t, x|Y0t )) + tr[∇∇ T (g(t, x) p(t, x|Y0t ))],
2

g(t, x) is the (n × n)-dimensional symmetric matrix function σ (t, x)σ T (t, x), tr[·]
is the trace of the matrix, and the function λ(t, x, z) is specified as follows:


m 
m 
1
λ(t, x, z) = ck (t, x)qkr (t) zr − cr (t, x)
k=1 r =1
2

1
=cT (t, x)q(t) z − c(t, x) , q(t) = η−1 (t), (24.6)
2

and

 
λ(t, X (t), Ẏ (t)) = E[λ(t, X (t), Ẏ (t))|Y0t ] = λ(t, x, Ẏ (t)) p(t, x|Y0t )d x.
Rn

Equation 24.5 is the nonlinear Stochastic Partial Differential Equation (SPDE)

in Stratonovich interpretation. It can be used to derive equations for conditional
moments of the state X .
Let μi1 ...i R (t) be the Rth conditional central moment of the state X , i.e.,

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 355

◦ ◦
μi1 ...i R (t) = X i1 (t) . . . X i R (t) ,

◦
where X i (t) = X i (t) − X̂ iMMSE (t), i = 1, 2, . . . , n. Then differential equations for
conditional central moments are as follows [29]:


R
◦
μ̇i1 ...i R (t) = f iq (t, X (t))( X i1 ...iq−1 iq+1 ...i R (t) − μi1 ...iq−1 iq+1 ...i R (t))
q=1


R−1 
R
◦
+ giq is (t, X (t)) X i1 ...iq−1 iq+1 ...is−1 is+1 ...i R (t)
q=1 s=1+q


R
◦
+ X iq (t)λ(t, X (t), Ẏ (t)) μi1 ...iq−1 iq+1 ...i R (t)
q=1
  ◦ ◦
+ λ(t, X (t), Ẏ (t)) μi1 ...i R (t) − X i1 (t) . . . X i R (t)λ(t, X (t), Ẏ (t)) ,

where
◦ ◦ ◦ ◦ ◦
X i1 ...iq−1 iq+1 ...i R (t) = X i1 (t) . . . X iq−1 (t) X iq+1 (t) . . . X i R (t),
◦ ◦ ◦ ◦ ◦ ◦ ◦
X i1 ...iq−1 iq+1 ...is−1 is+1 ...i R (t) = X i1 (t) . . . X iq−1 (t) X iq+1 (t) . . . X is−1 (t) X is+1 (t) . . . X i R (t).

These equations can be rewritten in a brief form using the multi-index notation
such that


R
◦ 
R−1 
R
◦
μ̇ī (t) = f iq (t, X (t))( X ī(q) (t) − μī(q) (t)) + giq is (t, X (t)) X ī(q,s) (t)
q=1 q=1 s=1+q
⎛ ⎞ 

R
◦ ◦
+ ⎝ X iq (t)μī(q) (t) + μī (t) − X ī (t)⎠λ(t, X (t), Ẏ (t)) , (24.7)
q=1

where ī = (i 1 . . . i R ), ī(q) = (i 1 . . . i q−1 i q+1 . . . i R ), ī(q, s) =

(i 1 . . . i q−1 i q+1 . . . i s−1 i s+1 . . . i R ).
A methodical approach to the high-precision filtering for the signal observation
model (Eqs. 24.1–24.2) based on a truncation of moments by cumulant closed tech-
niques [30–32] is considered in [29]. The high accuracy of developed algorithms
for the optimal nonlinear filtering problem is due to the use of conditional higher
order central moments defined by differential Eq. 24.7, the adaptability of the high-
precision filtering is provided by calculating in real time the conditional skewness
and excess kurtosis for all components of the state. The most effective computa-
tion of expectations · in differential Eq. 24.7 can be carried out using the method
developed in [33] for the statistical approximation of an arbitrary nonlinearity.

[email protected]
356 K. N. Chugai et al.

The unnormalized conditional probability density ϕ(t, x|Y0t ) satisfies Duncan–

Mortensen–Zakai equation [11, 12, 17, 19]:

∂ϕ(t, x|Y0t )
= Aϕ(t, x|Y0t ) + λ(t, x, Ẏ (t))ϕ(t, x|Y0t ), ϕ(t0 , x) = ϕ0 (x). (24.8)
∂t
Equation 24.8 is the linear SPDE in Stratonovich interpretation. The function
λ(t, x, z) is called an absorption and recovering intensity [21, 34] or a potential
function [20].
Let us assume that

− −λ(t, x, z) λ(t, x, z) < 0
λ (t, x, z) = ,
0 λ(t, x, z)  0

+ λ(t, x, z) λ(t, x, z) > 0
λ (t, x, z) = ,
0 λ(t, x, z)  0

i.e., λ(t, x, z) = −λ− (t, x, z) + λ+ (t, x, z). Using such representation for the func-
tion λ(t, x, z), we can rewrite Duncan–Mortensen–Zakai equation as the generalized
Fokker–Planck–Kolmogorov equation with absorption and recovering functions:

∂ϕ(t, x|Y0t )
= Aϕ(t, x|Y0t ) − λ− (t, x, Ẏ (t))ϕ(t, x|Y0t ) + λ+ (t, x, Ẏ (t))ϕ(t, x|Y0t ),
∂t
ϕ(t0 , x) = ϕ0 (x).

Note that the generalized Fokker–Planck–Kolmogorov equations have been intro-

duced, e.g., in [34, 35], for the switching diffusions or stochastic dynamical systems
with variable or random structure.
Then we can define a special random process with terminating and branching
paths [21, 34]. Paths of such process are completely determined by SDE (Eq. 24.1),
and the observations described by SDE (Eq. 24.2) affect on the terminating and
branching rates (or intensities). The probabilities of terminating and branching on
the time interval [t, t + t] at X (t) = x and Ẏ (t) = z for small t are Pr − (t, t) =
λ− (t, x, z) t + o( t) and Pr + (t, t) = λ+ (t, x, z) t + o( t), respectively. Such
probabilistic interpretation has been used for solving approximately the filtering and
prediction problems [21, 34, 36].
To find the conditional probability density p(t, x|Y0t ), we should use the
normalized representation:

ϕ(t, x|Y0t )
p(t, x|Y0t ) =  , t ∈ T.
Rn ϕ(t, x|Y0 )d x
t

Rewrite Eq. 24.8 as

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 357

∂ϕ(t, x|Y0t )
= Lϕ(t, x|Y0t ) + cT (t, x)q(t)Ẏ (t),
∂t
where
1
Lϕ(t, x|Y0t ) = Aϕ(t, x|Y0t ) − cT (t, x)q(t)c(t, x)ϕ(t, x|Y0t ).
2
Then define functions


m
h k (t, x) = qkr (t)cr (t, x), k = 1, 2, . . . , m,
r =1

and a new unnormalized conditional probability density

 

m
ρ(t, x|Y0t ) = exp − h k (t, x)Yk (t) ϕ(t, x|Y0t )
k=1
 
= exp −h (t, x)Y (t) ϕ(t, x|Y0t ).
T
(24.9)

The unnormalized conditional probability density ρ(t, x|Y0t ) satisfies the robust
Duncan–Mortensen–Zakai equation [17, 21]:

∂ρ(t, x|Y0t )  m
= Lρ(t, x|Y0t ) − Yk (t)Lk ρ(t, x|Y0t )
∂t k=1

m 
m 
m
∂h k (t, x)
+ Yk (t)Yr (t)Lkr ρ(t, x|Y0t ) − Yk (t)ρ(t, x|Y0t ), (24.10)
k=1 r =1 k=1
∂t

where Lk = [Hk , L] and Lkr = 21 [Hk , Lr ] = 21 [Hk , [Hr , L]], Hk are the
multiplication operators with multipliers h k (t, x), [·, ·] denotes Lie bracket.
Expressions for operators Lk and Lkr have been obtained in [21] for the stationary
observation model

dY (t) = c(X (t))dt + d V (t)

instead of the nonstationary model (Eq. 24.2), where c(x) : Rn → Rm is the m-

dimensional function, V (t) is the m-dimensional standard Wiener process.
Let us obtain expressions for operators Lk and Lkr with respect to nonstationary
observation model (Eq. 24.2). Firstly,

Lk ρ(t, x|Y0t ) =[Hk , L] ρ(t, x|Y0t ) = [Hk , A] ρ(t, x|Y0t )

=(Hk ◦ A)ρ(t, x|Y0t ) − (A ◦ Hk )ρ(t, x|Y0t )

[email protected]
358 K. N. Chugai et al.

n
∂
= − h k (t, x) [ f i (t, x)ρ(t, x|Y0t )]
i=1
∂ x i

1 n  n
∂2
+ h k (t, x) [gi j (t, x)ρ(t, x|Y0t )]
2 i=1 j=1
∂ x i ∂ x j

n
∂
+ [ f i (t, x)h k (t, x)ρ(t, x|Y0t )]
i=1
∂ x i

1   ∂2
n n
− [gi j (t, x)h k (t, x)ρ(t, x|Y0t )].
2 i=1 j=1 ∂ xi ∂ x j

Further, we will use the following properties. If u(t, x) and v(t, x) are twice
differentiable functions with respect to x, then
   
∂ (uv) ∂u ∂v ∂ 2 (uv) ∂ 2u ∂ ∂v ∂ ∂v ∂ 2v
= v+u , = v+ u + u −u
∂ xi ∂ xi ∂ xi ∂ xi ∂ x j ∂ xi ∂ x j ∂ x j ∂ xi ∂ xi ∂ x j ∂ xi ∂ x j

for all i, j = 1, 2, . . . , n. The matrix function g(t, x) is symmetric, i.e., gi j (t, x) =

g ji (t, x). Consequently,


n
∂h k (t, x)
Lk ρ(t, x|Y0t ) = f i (t, x) ρ(t, x|Y0t )
i=1
∂ x i
 
1  ∂
n n
∂h k (t, x)
− gi j (t, x) ρ(t, xY0 )
t
2 i=1 j=1 ∂ x j ∂ xi
 
1  ∂
n n
∂h k (t, x)
− gi j (t, x) ρ(t, x|Y0 )
t
2 i=1 j=1 ∂ xi ∂x j

1 
n n
∂ 2 h k (t, x)
+ gi j (t, x) ρ(t, x|Y0t )
2 i=1 j=1 ∂ xi ∂ x j

n
∂h k (t, x)
= f i (t, x) ρ(t, x|Y0t )
i=1
∂ x i
⎡ ⎤
 n
∂ ⎣
n
∂h k (t, x)
− gi j (t, x) ρ(t, x|Y0t )⎦
i=1
∂ x i j=1
∂ x j

1 
n n
∂ 2 h k (t, x)
+ gi j (t, x) ρ(t, x|Y0t )
2 i=1 j=1 ∂ xi ∂ x j

or

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 359

n
∂ k
Lk ρ(t, x|Y0t ) = f k (t, x)ρ(t, x|Y0t ) − [gi (t, x)ρ(t, x|Y0t )] + h k (t, x)ρ(t, x|Y0t ),
i=1
∂ x i

where


n
∂h k (t, x)
f (t, x) =
k
f i (t, x) = ∇ T h k (t, x) f (t, x), gik (t, x)
i=1
∂ xi

n
∂h k (t, x)
= gi j (t, x) , i = 1, 2, . . . , n,
j=1
∂x j

1 
n n
∂ 2 h k (t, x) 1
h k (t, x) = gi j (t, x) = tr[g(t, x)∇∇ T h k (t, x)].
2 i=1 j=1 ∂ xi ∂ x j 2

Secondly,

Lkr ρ(t, x|Y0t ) = [Hk , Lr ] ρ(t, x|Y0t )

= (Hk ◦ Lr )ρ(t, x|Y0t ) − (Lr ◦ Hk )ρ(t, x|Y0t )
1 n
∂ r
= − h k (t, x) [gi (t, x)ρ(t, x|Y0t )]
2 i=1
∂ x i

1 ∂ r
n
+ [g (t, x)h k (t, x)ρ(t, x|Y0t )]
2 i=1 ∂ xi i
1 r
n
∂h k (t, x)
= gi (t, x) ρ(t, x|Y0t )
2 i=1 ∂ xi

or
1 kr
Lkr ρ(t, x|Y0t ) = g (t, x)ρ(t, x|Y0t ),
2
where


n
∂h k (t, x)
g (t, x) =
kr
gir (t, x) = ∇ T h k (t, x)gr (t, x)
i=1
∂ xi
= ∇ h k (t, x)g(t, x)∇h r (t, x) (g k (t, x) = g(t, x)∇h k (t, x)).
T

Thus,

∂ρ(t, x|Y0t ) m n
∂ k
= Aρ(t, x|Y0t ) + Yk (t) [gi (t, x)ρ(t, x|Y0t )]
∂t k=1 i=1
∂ x i

[email protected]
360 K. N. Chugai et al.


m
− Yk (t)( f k (t, x) + h k (t, x))ρ(t, x|Y0t )
k=1

1 
m m
+ Yk (t)Yr (t)g kr (t, x)ρ(t, x|Y0t )
2 k=1 r =1
1 
m m
∂h k (t, x)
− h k (t, x)ck (t, x)ρ(t, x|Y0t ) − Yk (t) ρ(t, x|Y0t ).
2 k=1 k=1
∂t

Let us introduce new notations


m
f˜i (t, x, y) = f i (t, x) − yk gik (t, x), i = 1, 2, . . . , n,
k=1

m
1 
m m
ν(t, x, y) = − yk ( f k (t, x) + h k (t, x)) + yk yr g kr (t, x)
k=1
2 k=1 r =1
1  ∂h k (t, x)
m m
− h k (t, x)ck (t, x) − yk
2 k=1 k=1
∂t

or
 
∂h(t, x) T
f˜(t, x, y) = f (t, x) − g(t, x) y, (24.11)
∂x
∂h(t, x) 1
ν(t, x, y) = − y T f (t, x) − tr[g(t, x)∇∇ T (y T h(t, x))]
∂x 2
 
1 T ∂h(t, x) ∂h(t, x) T
+ y g(t, x) y
2 ∂x ∂x
1 ∂h(t, x)
− h T (t, x)c(t, x) − y T . (24.12)
2 ∂t
Consequently, Eq. 24.10 can be rewritten in the form:

∂ρ(t, x|Y0t ) 
= Aρ(t, x|Y0t ) + ν(t, x, Y (t))ρ(t, x|Y0t ), (24.13)
∂t
where

n
∂ ˜

Aρ(t, x|Y0t ) = − [ f i (t, x, Y (t))ρ(t, x|Y0t )]
i=1
∂ x i

1   ∂2
n n
+ [gi j (t, x)ρ(t, x|Y0t )]
2 i=1 j=1 ∂ xi ∂ x j

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 361

1
= −∇ T ( f˜(t, x, Y (t))ρ(t, x|Y0t )) + tr[∇∇ T (g(t, x)ρ(t, x|Y0t ))].
2
The initial condition for Eqs. 24.10 and 24.13 is determined by ρ(t0 , x) = ϕ0 (x),
since exp{−h T (t0 , x)Y0 } = 1. Equation 24.13 can also be represented as the general-
ized Fokker–Planck–Kolmogorov equation with absorption and recovering functions
[21, 34]:

∂ρ(t, x|Y0t ) 
= Aρ(t, x|Y0t ) − ν − (t, x, Y (t))ρ(t, x|Y0t ) + ν + (t, x, Y (t))ρ(t, x|Y0t ),
∂t
where

− −ν(t, x, y) ν(t, x, y) < 0
ν (t, x, y) = ,
0 ν(t, x, y)  0

+ ν(t, x, y) ν(t, x, y) > 0
ν (t, x, y) = ,
0 ν(t, x, y)  0

i.e., ν(t, x, y) = −ν − (t, x, y) + ν + (t, x, y).

The function ν(t, x, y) is the absorption and recovering intensity or a potential
function similar to Eq. 24.6. This equation describes an evolution of the unnormalized
conditional probability density of the special state X̃ ∈ Rn defined by the following
Itô SDEs:

d X̃ (t) = f˜(t, X̃ (t), Y (t))dt + σ (t, X̃ (t))d W̃ (t), X̃ (t0 ) = X 0 , (24.14)

where t ∈ T, f˜(t, x, y) is introduced by Eq. 24.11, W̃ (t) is the s-dimensional

standard Wiener process.
Then we can define a special random process with terminating and branch-
ing paths determined by SDE (Eq. 24.14) and the observations described by SDE
(Eq. 24.2) affect on the terminating and branching rates (or intensities). The proba-
bilities of terminating and branching on the time interval [t, t + t] at X (t) = x and
Y (t) = y for small t are Pr − (t, t) = ν − (t, x, y) t + o( t) and Pr + (t, t) =
ν + (t, x, y) t + o( t), respectively.

24.4 Particle Filters

To solve approximately Duncan–Mortensen–Zakai Eq. 24.8, we can use the particle

method or sequential Monte Carlo method [18, 19]. Let ω(t) denote the weight
function defined by the following equation:

[email protected]
362 K. N. Chugai et al.
 t 
ω(t) = exp λ(τ, X (τ ), Ẏ (τ ))dτ
t
 0t  t 
1
= exp c (τ, X (τ ))q(τ )dY (τ ) −
T
c (τ, X (τ ))q(τ )c(τ, X (τ ))dτ .
T
t0 2 t0

So, the estimate X̂ MMSE (t) is the normalized weighted mean [19]:

E[ω(t)X (t)]
X̂ MMSE (t) = .
Eω(t)

To find the estimate X̂ MMSE (t), it is necessary to simulate M sample paths X j (t)
of the random process X (t) and corresponding paths ω j (t) of the weight function
ω(t) by a numerical method for Itô SDE (Eq. 24.1) (a pair (X j (t), ω j (t)) is called
a particle), j = 1, 2, . . . , M. For example, using Euler–Maruyama method [37], we
have
√
X k+1 = X k + h f (tk , X k ) + hσ (tk , X k ) Wk ,
cT (tk ,X k )q(tk )(Y (tk+1 )−Y (tk ))− 21 cT (tk ,X k )q(tk )c(tk ,X k )h
ωk+1 = ωk e , ω0 = 1,

where h = (T − t0 )/N is the time discretization step, tk = t0 + kh, Wk is the s-

dimensional random vector with independent components having a standard normal
distribution, k = 0, 1, . . . , N − 1. Thus,

1  j j 
M M
j
X̂ MMSE (tk ) ≈ X̂ k = ωk X k , Ωk = ωk . (24.15)
Ωk j=1 j=1

The unnormalized conditional probability density ϕ(t, x|Y0t ) and the conditional
probability density p(t, x|Y0t ) can be represented as


M
j j 1  j
M
j
ϕ(tk , x|Y0tk ) ≈ ωk δ(x − X k ), p(tk , x|Y0tk ) ≈ ω δ(x − X k ), (24.16)
j=1
Ωk j=1 k

where δ(x − x ∗ ) is the Dirac delta function concentrated at x ∗ [18, 19].

Similarly, the particle method can be used to solve approximately the robust
Duncan–Mortensen–Zakai Eq. 24.13. Define the weight function as follows:
 t 
ω̃(t) = exp ν(τ, X (τ ), Y (τ ))dτ ,
t0

where ν(t, x, y) is introduced in Eq. 24.12. Then, we can simulate M sample paths
X̃ j (t) of the random process X̃ (t) and corresponding paths ω̃ j (t) of the weight
function ω̃(t) also using Euler–Maruyama method [37] for Itô SDE (Eq. 24.14) (a

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 363

pair ( X̃ j (t), ω̃ j (t)) is also a particle), i = 1, 2, . . . , M:

√
X̃ k+1 = X̃ k + h f˜(tk , X̃ k , Y (tk )) + hσ (tk , X̃ k ) W̃k ,
ω̃k+1 = ω̃k eν(tk , X̃ k ,Y (tk ))h , ω̃0 = 1.

Here, W̃k is the s-dimensional random vector with independent components

having a standard normal distribution, k = 0, 1, . . . , N − 1.
Hence, the unnormalized conditional probability density ρ(t, x|Y0t ) can be
represented in the form similar to Eq. 24.16:


M
j j
ρ(tk , x|Y0tk ) ≈ ω̃k δ(x − X̃ k ).
j=1

Also note that the function exp{−h T (t, x)Y (t)} in Eq. 24.9 is the additional
weight function, which can be used to approximate conditional probability densities
ϕ(t, x|Y0t ) and p(t, x|Y0t ). Thus, we can conclude that

1  
M M
ω̃k X̃ k , Ω̃k∗ =
j∗ j j∗
X̂ MMSE (tk ) ≈ X̂ k = ω̃k , (24.17)
Ω̃k∗ j=1 j=1

j∗ j j
where ω̃k = ω̃k exp{h T (tk , X̃ k )Y (tk )}, and


M
j∗ j 1 
M
j∗ j
ϕ(tk , x|Y0tk ) ≈ ω̃k δ(x − X̃ k ), p(tk , x|Y0tk ) ≈ ω̃k δ(x − X̃ k ).
j=1
Ω̃k∗ j=1

We can also use different representations for conditional probability densities

p(t, x|Y0t ), ϕ(t, x|Y0t ), and ρ(t, x|Y0t ). For example, it is possible to construct the
histogram or another density estimates using particles. The kernel estimation [38]
can also be applied for the conditional probability density p(t, x|Y0t ):
!
1  j
M j
x − Xk
p(tk , x|Y0tk ) ≈ ω K
Ωk h nx j=1 k
hx

or
!
1 
M
j∗
j
x − X̃ k
p(tk , x|Y0tk ) ≈ ω̃k K ,
Ω̃k∗ h nx j=1
hx

where K (x) is the kernel, i.e., some probability density function, h x > 0 is
the smoothing parameter. For instance, K (x) is the probability density for the

[email protected]
364 K. N. Chugai et al.

n-dimensional normal distribution:

1
e− 2 |x| .
1 2
K (x) =
(2π ) n/2

To find the estimate X̂ MAP (t) or mode of the conditional distribution, the kernel
estimations for the conditional probability density p(t, x|Y0t ) can be used, but the
calculation time increases in this case. Furthermore, this estimate depends on the
kernel and the smoothing parameter. As an alternative, it is suggested to use Edge-
worth series [10] for the expansion of marginal conditional probability densities. This
approach allows to reduce significantly the computation time in contrast to finding
the mode by consistent estimating the conditional probability density.
Let us denote marginal conditional probability densities by pi (t, xi |Y0t ):

pi (t, xi |Y0t ) = p(t, x|Y0t )d x1 . . . d xi−1 d xi+1 . . . d xn , i = 1, 2, . . . , n.
Rn−1

For"the normalized random value ξi = (X i − X̄ i )/σi , where X̄ i = EX i and

σi = E(X i − X̄ i )2 , we can write the marginal probability density as
 
1 μ3,i (3) 1 μ4,i
pi∗ (ξ ) = φ(ξ ) − φ (ξ ) + − 3 φ (4) (ξ )
3! σi3 4! σi4
 
10 μ23,i (6) 1 μ5,i μ3,i (5)
+ φ (ξ ) − − 10 3 φ (ξ )
6! σi6 5! σi5 σi
 
35 μ3,i μ4,i (7) 280 μ33,i (9)
− − 3 φ (ξ ) − φ (ξ ) + . . .
7! σi3 σi4 9! σi9
 
1 μ3,i 1 μ4,i 10 μ23,i
= φ(ξ ) 1 + H3 (ξ ) + − 3 H4 (ξ ) + H6 (ξ )
3! σi 3 4! σi 4 6! σi6
    !
1 μ5,i μ3,i 35 μ3,i μ4,i 280 μ33,i
+ − 10 3 H5 (ξ ) + − 3 H7 (ξ ) + H9 (ξ ) + . . . ,
5! σi5 σi 7! σi3 σi4 9! σi9

where φ(ξ ) is the probability density for the standard normal distribution and Hk (ξ )
is the Hermite polynomial of degree k:

1 d k φ(ξ )
φ(ξ ) = √ e− 2 ξ , φ (k) (ξ ) =
1 2
= (−1)k φ(ξ )Hk (ξ ),
2π dξ k

and μ R,i are the Rth central moments of the random value X i , i.e., μ R,i = E(X i −
X̄ i ) R .
According to [10], the mode of the distribution with probability density pi∗ (ξ ) can
be approximately calculated by

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 365

1 μ3,i
ξi∗ = − ,
2 σi3

consequently,

1 μ̂3,i,k
X̂ MAP (tk ) ≈ Ξk , (Ξk )i = ( X̂ k )i − , i = 1, 2, . . . , n, (24.18)
2 σ̂i,k
2

where σ̂i,k is the estimate of the conditional standard deviation and μ̂3,i,k is the
estimate of the conditional third central moment for the ith component X i of the
state X at time t = tk . These statistics can be obtained similar to X̂ k , i.e.,
#
$
$ 1  M
1  j
M
=%
j j j
σ̂i,k ωk ((X k )i − ( X̂ k )i )2 , μ̂3,i,k = ω ((X k )i − ( X̂ k )i )3 ,
Ωk j=1 Ωk j=1 k

or
#
$
$ 1  M
1  j∗
M
=% ∗
j∗ j j
σ̂i,k ω̃k (( X̃ k )i − ( X̂ k )i )2 , μ̂3,i,k = ∗ ω̃k (( X̃ k )i − ( X̂ k )i )3 ,
Ω̃k j=1 Ω̃k j=1

where Ωk and Ω̃k∗ have been defined in Eqs. 24.15 and 24.17, respectively.
We can also obtain another mode approximation based on necessary conditions
for extrema of the marginal probability density approximation using a partial sum of
Edgeworth series. Thus,

1 μ3,i (3) 1 μ3,i
pi∗ (ξ ) ≈ φ(ξ ) − φ (ξ ) = φ(ξ ) 1+ H3 (ξ ) .
3! σi3 3! σi3

The mode ξi∗ is a root of the quartic equation:

6σi3
ξ 4 − 6ξ 2 + ξ + 3 = 0, μ3,i = 0 (24.19)
μ3,i

that is a consequence of the equation

dpi∗ (ξ )
=0
dξ

and the recurrence relation for Hermite polynomials

d Hk (ξ )
= ξ Hk (ξ ) − Hk+1 (ξ ), H0 (ξ ) = 1.
dξ

[email protected]
366 K. N. Chugai et al.

All roots of Eq. 24.19 can be found by Descartes–Euler method or Ferrari’s method
(Descartes–Euler method is more preferable because Eq. 24.19 has a “reduced”
∗
form [39]), and the marginal mode approximation ξi,k is the root of Eq. 24.19 with
substitutions σi = σ̂i,k and μ3,i = μ̂3,i,k if μ̂3,i,k = 0, and for this root the probability
density pi∗ (ξ ) has the largest value. Consequently,

∗
X̂ MAP (tk ) ≈ Ξk , (Ξk )i = ( X̂ k )i + σ̂i,k ξi,k , i = 1, 2, . . . , n, (24.20)

∗
and ξi,k = 0, (Ξk )i = ( X̂ k )i if μ̂3,i,k = 0.
The algorithms for solving approximately the optimal filtering problem that pro-
vide two estimates for the state are given below. These estimates are the unbiased esti-
mate X̂ MMSE (t) with a minimum mean squared error and the maximum a posteriori
estimate X̂ MAP (t) (see Eqs. 24.3–24.4 for details). Algorithm 1 is based on the sample
paths simulation of the random process X (t), and Algorithm 2 is based on the sample
paths simulation of the random process X̃ (t). These algorithms correspond to Dun-
can–Mortensen–Zakai equation and the robust Duncan–Mortensen–Zakai equation,
respectively.
Algorithm 1
1. Specify M, the number of paths for the random process (X (t), ω(t)) to be simu-
lated (sample size), specify h, the time discretization step such that there exists the
j
natural number N for which h = (T −t0 )/N . Generate initial states X 0 according
j
to a given distribution with the probability density ϕ0 (x) and let k = 0, ω0 = 1,
j = 1, 2, . . . , M.
2. Estimate the conditional mean of the state and estimate the mode of the con-
j j M
ditional distribution at time t = tk using the sample ({X k } M j=1 , {ωk } j=1 ) as
follows:

1  j j
M
1 μ̂3,i,k
( X̂ kMMSE )i = ( X̂ k )i = ω (X )i , ( X̂ kMAP )i = ( X̂ k )i − ,
Ωk j=1 k k 2 σ̂i,k
2

where σ̂i,k is the estimate of the conditional standard deviation and μ̂3,i,k is the
estimate of the conditional third central moment for the ith component X i of the
state X at time t = tk (i = 1, 2, . . . , n), i.e.,
#
$
$ 1  M
1  j
M
=%
j j j
σ̂i,k ωk ((X k )i − ( X̂ k )i )2 , μ̂3,i,k = ω ((X k )i − ( X̂ k )i )3 ,
Ωk j=1 Ωk j=1 k

and

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 367


M
j
Ωk = ωk .
j=1

If T − tk = 0, terminate the estimation process. Otherwise, let j = 1.

3. Obtain a realization of the state
√ and corresponding weight at time t = tk + h:
j j j j j
X k+1 = X k + h f (tk , X k ) + hσ (tk , X k ) Wk ,
j j j
j j T
(tk ,X k )q(tk )(Y (tk+1 )−Y (tk ))− 21 cT (tk ,X k )q(tk )c(tk ,X k )h
ωk+1 = ωk ec ,
j
where Wk is the realization of the s-dimensional random vector with indepen-
dent components having a standard normal distribution.
4. If j = M, let tk+1 = tk + h and k := k +1, go to Step 2. Otherwise, let j := j +1
and go to Step 3.
Algorithm 2
1. Specify M, the number of paths for the random process ( X̃ (t), ω̃(t)) to be sim-
ulated (sample size), specify h, the time discretization step such that there exists
the natural number N for which h = (T − t0 )/N . Generate initial special states
j
X̃ 0 according to a given distribution with the probability density ϕ0 (x) and let
j
k = 0, ω̃0 = 1, j = 1, 2, . . . , M.
2. Estimate the conditional mean of the state and estimate the mode of the con-
j j M
ditional distribution at time t = tk using the sample ({ X̃ k } M j=1 , {ω̃k } j=1 ) as
follows:

1 
M
j∗ j 1 μ̂3,i,k
( X̂ kMMSE )i = ( X̂ k )i = ω̃k X̃ k , ( X̂ kMAP )i = ( X̂ k )i − ,
Ω̃k∗ j=1
2 σ̂i,k
2

where σ̂i,k is the estimate of the conditional standard deviation and μ̂3,i,k is the
estimate of the conditional third central moment for the ith component X i of the
state X at time t = tk (i = 1, 2, . . . , n), i.e.,
#
$
$ 1  M
1  j∗
M
=% ∗
j∗ j j
σ̂i,k ω̃k (( X̃ k )i − ( X̂ k )i )2 , μ̂3,i,k = ∗ ω̃k (( X̃ k )i − ( X̂ k )i )3 ,
Ω̃k j=1 Ω̃k j=1

and


M
Ω̃k∗ =
j∗ j∗ j j
ω̃k , ω̃k = ω̃k exp{h T (tk , X̃ k )Y (tk )}.
j=1

If T − tk = 0, terminate the estimation process. Otherwise, let j = 1.

[email protected]
368 K. N. Chugai et al.

3. Obtain a realization of the special state and corresponding weight at time t =

tk + h:

√
X̃ k+1 = X̃ k + h f˜(tk , X̃ k , Y (tk )) +
j j j j j
hσ (tk , X̃ k ) W̃k ,
j
ω̃k+1 = ω̃k eν(tk , X̃ k ,Y (tk ))h ,
j j

j
where W̃k is the realization of the s-dimensional random vector with independent
components having a standard normal distribution.
4. If j = M, let tk+1 = tk + h and k := k +1, go to Step 2. Otherwise, let j := j +1
and go to Step 3.
Note that in algorithms given above, Eq. 24.18 is used for obtaining approximately
the mode of the conditional distribution. Equation 24.20 can also be applied to this.
Thus, we will have
∗
( X̂ kMAP )i = ( X̂ k )i + σ̂i,k ξi,k ,

∗
where ξi,k is a root of the quartic equation

6σ̂i,k
3
ξ 4 − 6ξ 2 + ξ + 3 = 0, μ̂3,i,k = 0,
μ̂3,i,k

for which the function

!
1 μ̂3,i,k
φ(ξ ) 1 + H3 (ξ ) , H3 (ξ ) = ξ 3 − 3ξ,
3! σ̂i,k
3

has the largest value. If μ̂3,i,k = 0, then ( X̂ kMMSE )i = ( X̂ kMAP )i for the considered
mode approximation.
Also note that various methods for solving SDEs numerically, such as Runge–
Kutta type methods [37, 40, 41], Rosenbrock type methods [40, 42], Platen’s methods
[37], Milshtein’s methods [43], and Kuznetsov’s methods [44, 45] instead of the
Euler–Maruyama method should be applied to increase the estimation accuracy.

24.5 Conclusions

The filtering algorithms based on the particle method for nonlinear continuous-time
stochastic observation systems for the unbiased estimate with a minimum mean
squared error and the maximum a posteriori estimate are suggested in this chapter.

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 369

For obtaining the maximum a posteriori estimate, Edgeworth series for the expan-
sion of marginal conditional probability densities is applied. This approach allows to
estimate the mode of the conditional distribution approximately, but it significantly
reduces the computation time in contrast to finding the mode by consistent estimat-
ing the conditional probability density. Particle method is used not only for Dun-
can–Mortensen–Zakai equation, but also for the robust Duncan–Mortensen–Zakai
equation.

Acknowledgements This work is partially supported by the Russian Foundation for Basic
Research, project no. 17-08-00530.

References

1. Yavin, Y.: An alternative approach to non-linear filtering: maximizing the probability of hitting
a target set. Int. J. Syst. Sci. 13(3), 289–299 (1982)
2. Bar-Shalom, Y., Li, X.R., Kirubarajan, T.: Estimation with Applications to Tracking and
Navigation: Theory, Algorithms and Software. Wiley, New York (2001)
3. Pany, T.: Navigation Signal Processing for GNSS Software Receivers. Artech House, Boston
(2010)
4. Stepanov, O.A., Koshaev, D.A., Motorin, A.V.: Identification of gravity anomaly model parame-
ters in airborne gravimetry problems using nonlinear filtering methods. Gyroscopy Navig. 6(4),
318–323 (2015)
5. Khalaf, W., Chouaib, I., Wainakh, M.: Novel adaptive UKF for tightly-coupled INS/GPS
integration with experimental validation on an UAV. Gyroscopy Navig. 8(4), 259–269 (2017)
6. Shavrin, V.V., Tislenko, V.I., Lebedev, VYu., Filimonov, V.A., Konakov, A.S.: Sigma-point
Kalman filter algorithm in the problem of GNSS signal parameters estimation in non-coherent
tracking mode in spacecraft autonomous navigation equipment. Gyroscopy Navig. 9(4), 255–
266 (2018)
7. Rudenko, E.A.: Autonomous path estimation for a descent vehicle using recursive Gaussian
filters. J. Comput. Syst. Sci. Int. 57(5), 695–712 (2018)
8. Rybakov, K.A.: Solving the nonlinear problems of estimation for navigation data processing
using continuous particle filter. Gyroscopy Navig. 10(1), 27–34 (2019)
9. Stepanov, O.A., Vasiliev, V.A., Toropov, A.B., Loparev, A.V., Basin, M.V.: Efficiency anal-
ysis of a filtering algorithm for discrete-time linear stochastic systems with polynomial
measurements. J. Franklin Inst. 356(10), 5573–5591 (2019)
10. Cramér, H.: Mathematical Methods of Statistics. Princeton University Press, Princeton (1999)
11. Zakai, M.: On the optimal filtering of diffusion processes. Zeitschrift für Wahrscheinlichkeit-
stheorie und Verwandte Gebiete 11(3), 230–243 (1969)
12. Hazewinkel, M.: Lectures on linear and nonlinear filtering. In: Schiehlen, W.O., Wedig, W.
(eds.) Analysis and Estimation of Stochastic Mechanical Systems. International Centre for
Mechanical Sciences (Courses and Lectures), vol. 303, pp. 103–136. Springer, Vienna (1988)
13. Baras, J.S., Blankenship, G.L., Mitter, S.K.: Nonlinear filtering of diffusion processes. In: 8th
IFAC Congress, article id 23.1 (1981)
14. Baras, J.S., Blankenship, G.L., Hopkins, W.E.: Existence, uniqueness, and asymptotic behavior
of solutions to a class of Zakai equations with unbounded coefficients. IEEE Trans. Autom.
Control 28(2), 203–214 (1983)
15. Yau, S.-T., Yau, S.S.-T.: Real time solution of Duncan-Mortensen-Zakai equation without
memory. In: 47th IEEE Conference on Decision and Control, pp. 5086–5091 (2008)

[email protected]
370 K. N. Chugai et al.

16. Rybakov, K.A.: Robust Duncan-Mortensen-Zakai equation for non-stationary stochastic sys-
tems. In: International Multi-Conference on Engineering, Computer and Information Sciences
(SIBIRCON), pp. 151–154 (2017)
17. Luo, X., Yau, S.S.-T.: Complete real time solution of the general nonlinear filtering problem
without memory. IEEE Trans. Autom. Control 58(10), 2563–2578 (2013)
18. Crisan, D.: Exact rates of convergence for a branching particle approximation to the solution
of the Zakai equation. Ann. Probab. 31(2), 693–718 (2003)
19. Bain, A., Crisan, D.: Fundamentals of Stochastic Filtering. Springer, New York (2009)
20. Del Moral, P.: Feynman-Kac Formulae: Genealogical and Interacting Particle Systems with
Applications. Springer, New York (2004)
21. Rybakov, K.A.: Statistical Methods of Analysis and Filtering for Continuous Stochastic
Systems. MAI Publ, Moscow (in Russian) (2017)
22. Øksendal, B.: Stochastic Differential Equations. An Introduction with Applications. Springer,
Berlin (2000)
23. Pugachev, V.S., Sinitsyn, I.N.: Stochastic Systems: Theory and Applications. World Scientific,
Singapore (2002)
24. Veretennikov, A.J.: On strong solutions and explicit formulas for solutions of stochastic integral
equations. Math USSR Sbornik 39(3), 387–403 (1981)
25. Veretennikov, A.J.: On stochastic equations with degenerate diffusion with respect to some of
the variables. Math USSR Izvestiya 22(1), 173–180 (1984)
26. Jazwinski, A.H.: Stochastic Processes and Filtering Theory. Academic Press, New York (1970)
27. Stratonovich, R.L.: Conditional Markov Processes and Their Application to the Theory of
Optimal Control. Elsevier (1968)
28. Kushner, H.J.: On the differential equations satisfied by conditional probability densities of
Markov processes, with applications. J. SIAM Ser. A: Control 2(1), 106–119 (1964)
29. Kosachev, I.M.: Methodology of high-precision nonlinear filtering of random processes in
stochastic dynamical systems with a fixed structure. Mil. Acad. Republic of Belarus Proc. 4,
125–161 (un Russian) (2014)
30. Dashevskii, M.L.: Cumulant method for closing moment equations in analyzing nonlinear
systems. Prob. Control Inf. Theory 4, 317–328 (in Russian) (1975)
31. Kashkarova, A.G., Shin, V.I.: Modified cumulant methods of analyzing stochastic systems.
Autom. Remote Control 2, 69–79 (in Russian) (1986)
32. Socha, L.: Linearization Methods for Stochastic Dynamic Systems. Springer, Berlin (2008)
33. Kosachev, I.M., Eroshenkov, M.G.: Analytical Modeling of Stochastic Systems. Nauka i
tekhnika Publ, Minsk (in Russian) (1993)
34. Rybakov, K.A.: Solving approximately an optimal nonlinear filtering problem for stochastic
differential systems by statistical modeling. Numer. Anal. Appl. 6(4), 324–336 (2013)
35. Kazakov, I.E., Artemiev, V.M., Bukhalev, V.A.: Analysis of systems with random structure.
Fizmatlit, Moscow (in Russian) (1993)
36. Averina, T.A., Rybakov, K.A.: An approximate solution of a prediction problem for stochastic
jump-diffusion systems. Numer. Anal. Appl. 10(1), 1–10 (2017)
37. Kloeden, P.E., Platen, E., Schurz, H.: Numerical Solution of SDEs Through Computer
Experiments. Springer, Berlin (1994)
38. Silverman, B.W.: Density Estimation for Statistics and Data Analysis. Chapman & Hall/CRC,
London (1986)
39. Korn, G.A., Korn, T.M.: Mathematical Handbook for Scientists and Engineers. Dover Publ,
New York (2000)
40. Artemiev, S.S., Averina, T.A.: Numerical Analysis of Systems of Ordinary and Stochastic
Differential Equations. VSP, Utrecht (1997)
41. Roberts, A.J.: Model Emergent Dynamics in Complex Systems. SIAM, Philadelphia (2014)
42. Averina, T.A., Karachanskaya, E.V., Rybakov, K.A.: Statistical modeling of random processes
with invariants. In: International Multi-Conference on Engineering, Computer and Information
Sciences (SIBIRCON), pp. 34–37 (2017)

[email protected]
24 Approximate Filtering Methods in Continuous-Time … 371

43. Milshtein, G.N., Tretyakov, M.V.: Stochastic Numerics for Mathematical Physics. Springer,
Berlin (2004)
44. Kuznetsov, D.F.: On numerical modeling of the multidimensional dynamic systems under
random perturbations with the 1.5 and 2.0 orders of strong convergence. Autom. Remote
Control 79(7), 1240–1254 (2018)
45. Kuznetsov, D.F.: On numerical modeling of the multidimensional dynamic systems under
random perturbations with the 2.5 order of strong convergence. Autom. Remote Control 80(5),
867–881 (2019)

[email protected]
Chapter 25
Essentials of Fractal Programming

Alexander S. Semenov

Abstract Fractal programming is a programming paradigm based on the concept of

“elastic objects”, which can transformed (unfolded and folded) dynamically at run
time using strategy planning model and production rules. These rules are keeping the
object structure self-similar, thus it has fractal property: parts similar to the whole. The
paradigm aims to optimize searching of suitable workflow structure of the object at
run time. The elastic object models, production rules integrated by iterated algebraic
system, and adjustment strategy planning model are introduced.

25.1 Introduction

In response to fluctuating demand of the society and the increased connectivity of

people, things, and services, IT systems themselves are becoming highly dynamic.
Run time factors on demand are increasingly determining the elasticity of a system.
The elasticity is becoming main characteristic of the technologies: cloud computing,
blockchain, Internet of things, digital platforms, robotics, biometrics, persuasive
technology, and augmented reality.
The term elasticity means “the degree to which a system is able to adapt to
workload changes by provisioning and de-provisioning resources in an autonomic
manner, such that at each point in time, the available resources match the current
demand as closely as possible” [1]. The elastic computing is one of the most important
design goals facing software developers.
The traditional static analysis is based on the assumption that a system consists of
objects with unchanged structure. If a system is expected to be elastic at run time, its
ability to elasticity must be considered, when it is designed. Hence, the elasticity of
object must be modeled and be built into it at design time. All this means an evolved

A. S. Semenov (B)
Moscow Aviation Institute (National Research University), 4, Volokolamskoe Shosse, Moscow
125993, Russian Federation
e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 373

L. C. Jain et al. (eds.), Advances in Theory and Practice of Computational
Mechanics, Smart Innovation, Systems and Technologies 173,
https://doi.org/10.1007/978-981-15-2600-8_25

[email protected]
374 A. S. Semenov

“shift” [2] from the static analysis to analysis with the elastic dynamicity of objects
at run time.
In this chapter, a technique that is able to evaluate the effects of elasticity at design
time is introduced. This technique is based on the concept of elastic objects and fractal
Iterated Algebraic System (IAS). Elastic objects are transformed dynamically at run
time by fractal algebra operations. The term fractal highlights the properties of elastic
objects: autoscaling (ability for unfolded and folded structure of an object), pattern
ability, similarity, and symmetry. The operation of replication (prototyping) makes
possible prototype elastic objects in the manner of prototype-based programming.
An elastic object structure is optimized by the planning model. This underpins the
relevance of the ideas presented in this chapter.
The chapter is organized as follows. In Sect. 25.2, the elastic object model is
considered. The elastic object model is based on a system of predefined scaling
conditions that automatically handles IT resources from resource pools [3]. Elastic
objects can be scalable with respect to its size, meaning that they automatically add
objects to the system. Modeling replication is a technique for achieving scalability
[4]. Replication is a key for providing high availability and fault tolerance in elastic
objects [5].
In Sect. 25.3, models of fractal elastic objects, such as container–component,
architectural model based on fractal graphs, and fractal Petri nets, are considered.
All these elastic objects models are based on IAS. Operations of IAS are defined
and exemplified for each model of elastic object. The basic idea of the fractal [6]
scalability in this context took into account. Fractal programming is a technique
for achieving the goal of elasticity objects and their computing. Each elastic object
changes its structure on demand by the fractal rules. The model of strategy planning
of elasticity as a composite part of elastic object is included. Assumption is made
that adding elastic objects to a system independently increases a processing potential
[5, 7]. Lastly, Sect. 25.4 concludes the chapter.

25.2 The Elastic Object Model

The elasticity is a paradigm that initially known in physics and widely used in eco-
nomics. In physics, it means the property of some material returning to an initial
form or state after the following deformation. For computer systems, this means that
with the increased workload of the service take place “deformation”, i.e., it nega-
tively impacted the users. Due to this impact, the capacity of the service needs to be
increased by IT resources. This may be repeated several times over a certain period of
time until the workload returns to its initial value. The service workload and capacity
are at the core of the elastic computing.
For coordinating elasticity, a system needs concepts of elastic objects—funda-
mental building blocks for engineering an end-to-end elasticity. Modern software
development techniques evolve around the use of object-orientation approach [8].

[email protected]
25 Essentials of Fractal Programming 375

Service-oriented approach [9] is an evolution of the object-oriented approach. Hence,

designing elastic object is considered as an integral part of a service.
To design an elastic object, it is necessary to analyze it from the point of view
of autoscaling (replicating and composition of objects) that depends on the work-
load of planning strategy, provisioning and binding of IT resources, and synthesis.
Summarizing these points of view, the next formula is written as follows:

Elasticity = Autoscaling + Planning strategy + Binding + Synthesis (25.1)

Automatic scaling (autoscaling) is a prerequisite for the elastic object. Method

autoscaling of the object is the ability to handle fluctuated workload (without adding
IT resources to the object) or by adding IT resources to the object for extending a
system’s capacity. Scalability of the object needs to be built into it at the design time.
The elasticity is a degree, to which a service autonomously adapts capacity to a
workload over time. The elasticity needs to be considered over time for changing
workloads. Such an adaptation must be timely minimized [10]. The timeliness entails
that adaptation process is autonomous [11]. The workload characterizes the data to
be processed by a service’s operations on fluctuating users demand during the day,
see a typical example Table 25.1.
Let a service has enough capacity at a workload of 2,000 users (see Fig. 25.1).
Thus, the service’s capacity defines the maximum workload, which the system can
handle in terms of number of users. If the workload during some period of time
increased up to 3,000 users, then the capacity of the elastic computer system must
be transformed dynamically.
There are two main inheritance models in the object-oriented approach: class-
based [8] and prototype-based [12]. In prototype-based model, only objects exist.
New objects are produced by prototyping and modification of prototypes. Prototyping
objects and mechanisms for scalable data replication are the important techniques of

Table 25.1 Timetable of

S Time Workload
workload observations
1 2 1,000
2 4 2,000
3 6 3,000
4 8 4,000
5 10 5,000
6 12 6,000
7 14 7,000
8 16 8,000
9 18 9,000
10 20 6,000
11 22 4,000
12 24 3,000

[email protected]
376 A. S. Semenov

Object
capacity
10 Planning Strategy

8 Folding
Unfolding
7

2 Time (h)

Initial 2 4 6 8 10 12 14 16 18 20 22 24
object e0 - prototype of the elastic object e0
- delete prototype of e0 object
- connection or composition of the elastic objects

Fig. 25.1 Strategic planning model

achieving structure of services with good performance, high availability, and fault
tolerance in a distributed system [13].
When an application requires more resources than are available, it negatively
impacted the users. Because of it, elastic computing must be provided for IT
resources: hardware and software entities, such as servers, CPU cores, memory,
program components, and modules. If a workload was changed, the capacity must
be transformed dynamically at run time on demand (see Fig. 25.1) by prototyping
the initial object e0 .
Thus, if a workload was decreased, for example, after 9,000 users, the capacity of
the system must be decreased (see Fig. 25.1) by deleting the prototyped objects. The
elastic computing acquires and releases IT resources, when demand changes that
requires support for late binding of IT resources. This technique has been success-
fully applied in programming language environments but also for operating systems,
where modules can be loaded and unloaded at will. An application eliminates the
binding between software and hardware through virtualization. The elastic applica-
tions would benefit if the underlying infrastructure provided the binding and replica-
tion for the elasticity. Operations must be themselves elastic. There are the follow-
ing binding techniques: resource-to-node, client-to-server, process-to-resource, and
geographical binding [14].
The granularity is usually used to characterize how many objects must be pro-
totyped with defined capacity at a certain time. The level of granularity should be
sufficient. The granularity trade-off is important at the cost of precision, accuracy,
and scope [2].
The cost-efficiency is closely related to the optimal provisioning. The optimal
provisioning strikes a balance between over-provisioning and under-provisioning.

[email protected]
25 Essentials of Fractal Programming 377

Transformations of the elastic object depend on provisioning of IT resources. It must

be taken into account by planning strategy.
Figure 25.2 generalizing Eq. 25.1, the model of the elastic object, is introduced.
The elastic object is a container which consists of components: attributes, autoscal-
ing operations (prototyping, composition), model of planning strategy, synthesis or
modification models, and pool of IT resources for the object (software and hardware).
The container is an object (in terms of object-oriented approach), which aggre-
gates components or containers. A component is an object, which aggregates by the
container.
In Fig. 25.3, the elastic computing platform is presented as the elastic object. At
any given level of abstraction, meaningful collections of elastic objects collaborate
to achieve some higher level behavior. This is exactly the fractal organization of
complexity [15].

Fig. 25.2 The elastic object Elastic object e0

model Attributes

Autoscaling operations:
1. prototyping
Components
2.composition

Planning
Containers Strategy

Synthesis model

run-time binding

Software library Hardware Component

Component

IT resources for object e0

Fig. 25.3 The elastic Demand

computing platform is the
elastic object

Elastic computing platform:

architectural elastic object

IT resources

Software resources:
Hardware Resources
elastic objects,
planning strategies,
synthesis models

[email protected]
378 A. S. Semenov

In the next section, the following models of elastic objects are defined: the con-
tainer–component model, the architectural model based on graphs [16], and Petri
nets model. These models are essentials of fractal programming.

25.3 Fractal Programming

Fractal programming is a programming paradigm based on the concept of “elastic

objects”, which can be transformed (unfolded and folded) dynamically at run time
using strategy planning model and production rules. These rules are keeping the
object structure self-similar, thus it has the fractal property: parts similar to the
whole. The paradigm aims to optimize searching suitable workflow structure of the
object at run time. The elastic object models, production rules integrated by IAS,
and adjustment strategy planning model are introduced. IAS is used for autoscaling:
unfolding and folding elastic objects mainly by operation prototyping. In dependence
on the model of elastic object, the operation prototyping is overloaded, and additional
operations are included in the fractal algorithm.
The fractal algorithms are based on constructing objects with requirement prop-
erties, such as self-similarity, scaling, and fractal dimension [6, 17]. The term fractal
is used to highlight the properties of elastic objects: abstraction, ability for unfolding
and folding, pattern ability, similarity, symmetry, and scalability. In many works on
fractals, self-similarity is used as defining property. In this chapter, the ordered bag
of self-similar prototyping objects is used. The ordered bag makes possible to define
the position of object in it.
The bag (for example, implemented by container) could include repeatable objects
the big number of times. The assumption was made that self-similarity describes the
objects, in which the same objects are repeated over and over again on different levels
of the scalability, for example, binary tree (the part looks like the whole).
Hereinafter, a model of container–component elastic objects is considered in
Sect. 25.3.1, while a model of architectural elastic objects is represented in
Sect. 25.3.2. Section 25.3.3 provides a model of fractal Petri nets as elastic objects.

25.3.1 The Model of Container–Component Elastic Objects

In all self-similar constructions, there is a relationship between the scaling factor and
the number of parts that the original object is divided into. The relationship is the
power law:

Nr d = 1, (25.2)

[email protected]
25 Essentials of Fractal Programming 379

(a) (b)
inheritance aggregation
N containers
≡ ↓
e0
container
2
≡ ↑ r scale factor
component 1 2 3

Fig. 25.4 The elastic object as recursive relationships between containers and components:
a inheritance and aggregation relationships, b the elastic object e0 (r = 1/3, N = 2, k = 1)

where N is the number of self-similar objects, 1/r d -, is the reduction scaling factor,
d is the fractal dimension.
Let E be the container–component elastic object. The operation prototyping and
composition (aggregation of prototyping objects) are carried out repeatedly; the out-
put of iteration is the input for the next one. For components, operations prototyping
and composition are undefined. As a whole, the container–component elastic object
is an object based on recursive relationships between containers and components.
Definition 1 Elastic object E = < e0 , ≡ (1/r, N), ↓ > , where e0 is an initial con-
tainer, ≡ : e0 −→ E is the operation prototyping object e0 in an ordered multi
bag E = [e1 , e2 , o1 , o2 ,…, or-N ,…,eN ], which consists from N containers (eN are
parts) and 1/r–N components (o1/r −N ) (here, the index shows the position of each
self-similar object in the bag), ↓: E −→ e0 is the operation aggregation. An ordered
multi bag E aggregates by container e0 (it is whole).
In Fig. 25.4a, the relationships of inheritance and aggregation between containers
and components are shown. In Fig. 25.4b, an object e0 with scale factor r = 1/3, two
containers N = 2, and one component 1/r−N is shown.
Algorithm 1 The scaling of the elastic object.

E = {e0 }, n = 1

∀e ∈ E repeat ↓ (≡ (1/r, N)e)n times.

Graphically, container is depicted as a rectangle. Nested containers are depicted

as nested rectangles. Component is depicted as gray rectangle. Positions of the
containers are ordered by indexes.
In Fig. 25.5, the second iteration k = 2 (level 2) of the elastic object e0 is shown.
IAS generalizes the composition of elastic objects and makes possible unfolded and
folded its structure [14].
Definition 2 Fractal algebra is a tuple fa = <E, > , where E is the ordered multi
bag of elastic objects,  is the set of uniquely invertible operations.  = { ≡ (1/r,

[email protected]
380 A. S. Semenov

e0
e11 e12 e13
e21 e22 e23 e21 e22 e23

Fig. 25.5 Elastic object: r = 1/3, N = 2, k = 2

N), ↓}, where ≡ (1/r, N) is an operation of prototyping (see Definition 1) ≡ (1/r,

N)−1 is an invertible operation, ↓ is the operation of aggregation E containers (see
Definition 1), ↓−1 is the invertible operation.
Let ƒ : E −→ E be a recursive mapping of an ordered multi-bag E in self-similar
an ordered multi bag E by fa operations.
Definition 3 Iterated Algebraic System (IAS) ƒn : E −→ E is the recursive mapping
of an ordered multi-bag E, where n = 0,1,…,k is the step of mapping, n++ is the
unfolding operation, n = n + 1, n−− is the folding operation, n = n−1.
Definition 4 Fractal algorithms, in which the fa operations presented by rules, have
the following notation:

F = f n (E, [R]), (25.3)

where F is the ordered multi bag or fractal object (result of mapping), R = [r 1 , r 2 ,…,
r m ] is the ordered set of rules. Rules consist of fa operations and conditions.
For consideration, the example of the elastic object algorithm is written as follows,
k = 3 (level):

e0 = f n (E = {e0 }, [n + +, ∀e ∈ E n , e ↓ (E n =≡ (3, 2)e)|0 < n =< k]). (25.4)

In all the above examples, the elastic object unfolded its structure. For the purpose
of folding the elastic object, the uniquely invertible operations were included in fractal
algebra. For generalized unfolded and folded elastic object, Eq. 25.5 is used.

e0 = f n (E = {e0 } [n + +, ∀e ∈ E n , e ↓ (E n =≡ (3, 2)e)|0 < n =< k]

[n − −, ∀e ∈ E n , e ↓−1 (E n =≡−1 (3, 2)e)|0 < n >= k])
(25.5)

The container–component model of the elastic object is based on fractal inter-

pretation of constructing the container and component. It makes possible to build a
model of adjustment capacity planning.
Adjustment planning strategy is adding or reducing the capacity of IT system in
small or large amounts due to consumer’s demand.
Let w be the workload and c be the capacity needed to handle workload w. A
timetable of observations of wn is given in Table 25.1. Our objective is to build a

[email protected]
25 Essentials of Fractal Programming 381

Fig. 25.6 The elastic object

as a tree: r = 1/3, N = 2,
k=2

model using fractal algorithm that captures the relationship between changes in the
workload and capacity requirement.
Let e0 be a server implements by elastic object. In general, IT capacity plan-
ning involves estimating the storage, computer hardware, software, and connection
infrastructure resources required over some future period of time. Let the server has
a capacity c = 1 unit for 3,000 workload and each prototyping container of the server
has the capacity of 1 unit.
In Fig. 25.6, a container–component tree of the elastic object with marked capacity
is depicted, k = 2. Components are used for replication the shared data, which, in
its turn, are applied by container. For workload of 9,000 will need k containers, k =
wn=18 /c:

e0 = f n (E = {e0 }, [n + +, ∀e ∈ E n , e ↓ (E n =≡ (3, 2)e)|0 < n =< wn /c],

[n − −, ∀e ∈ E n , e ↓−1 (E n =≡−1 (3, 2)e)|0 < n >= wn /c]).
(25.6)

The elastic computing is a run time optimization of elastic objects. One of the
potential problems is that the elasticity takes time.

25.3.2 The Model of Architectural Elastic Objects

Fractal graphs by applying elementary operations of fractal algebra to the initial

graph g are defined axiomatically [15]. A set of elementary operations forms the
basis of fractal algebra.

Definition 5 Fractal algebra Δ = {g | ≡ , ↔, ÷} for graphs over g is called the

set of uniquely invertible operations: prototyping (≡−1 ), connection (↔−1 ), and
subdivision (÷−1 ).

Definition 6 Fractal graph g is a graph generated by the operations of fractal algebra.

Subdivision (denoted by ÷) edges {v, v } of the graph g connecting vertices v,v by
inserting an additional vertex v and two edges incident to v {v,v }, {v ,v } is
denoted by (v ∈ g) ÷ (v ∈ g). As a result, the graph g and g are homeomorphic.

[email protected]
382 A. S. Semenov

Fig. 25.7 Operations of the

fractal algebra Δ for graphs:
a operation on subdivision,
b operation connection,
c operation prototyping

The operation of the subdivision is illustrated in Fig. 25.7a. Let the graph g has
two vertices, then g1 = (v ∈ g) ÷ (v ∈ g). The result of the inverse operation applied
to the graph g1 is the graph g = (v ∈ g1 ) ÷−1 (v ∈ g1 ).
Connection (denoted by ↔) of the distinguished vertex v of the graph g and its
prototype v in the graph g by the edge {v, v } is denoted by (v ∈ g) ↔ (v ∈ g ). The
connection operation is illustrated in Fig. 25.7b. The result of the connection graph
g1 = (v ∈ g) ↔ (v ∈ g ). The result of the inverse operation applied to the graph g1
is {g, g } = (v ∈ g1 ) ↔−1 (v ∈ g1 ).
Prototyping (denoted by≡) making a copy g of a connected graph-sample g.
Figure 25.7c illustrates the prototyping operation. Prototype g of a graph-sample
g is created. The result of the inverse operation applied to the graph g ≡ −1 is the
graph g.
The class of the graph is determined by rules that implement autoscaling: line
(L), mesh (M), hypercube (H), and tree (T ) graphs.
The following algorithms are based on IAS (see Eq. 25.3), operation prototyping
omitted, as well as, a graph g is a prototype of a graph g.

Algorithm 2 L = ƒn (g0 , [(v ∈ g) ↔ (v ∈ g )]) is the class of linear graphs see
(Fig. 25.8a).

Algorithm 3 M = ƒn (g0 , [(vi ∈ g) ↔ (vi  ∈ g ), (vj ∈ g) ↔ (vj  ∈ g )]) is the class

of lattices, where vi , vj , i = j are two pairs of isomorphic vertices of the graph g and
the prototype graph g (see Fig. 25.8b).

Algorithm 4 H = ƒn (g0 , [∀ v,v ((v ∈ g) ↔ (v ∈ g ))]) is the class of hypercubes, the
connection operation is performed for all isomorphic vertices of the graphs g and g
(see Fig. 25.8c).

[email protected]
25 Essentials of Fractal Programming 383

Fig. 25.8 Autoscaling of the fractal graphs: a line, b mesh, c hypercube, d tree

[email protected]
384 A. S. Semenov

Algorithm 5 T = ƒn (g0 , [(v ∈ g)↔(v ∈ g ), (v ∈ g) ÷ (v ∈ g )]) is the class of trees,

where v, v are the vertices of the graph g and its prototype (see Fig. 25.8d).

For the presented classes of graphs, the initial graph is the graph g0 with one
vertex.
In Fig. 25.8, the elements of fractal graphs depicted by the operations of connection
and subdivision are highlighted by a dotted line.
Uniquely invertible operations are used for destructing subgraphs of initial graph.
The elastic object planning strategy can be nested in these algorithms, for example,
class line (the operation prototyping omitted):

L = f n (g0 , [n + +, (v ∈ g) ↔ |(v  ∈ g  )|0 < n =< wn /c],

[n − −, (v ∈ g) ↔−1 (v  ∈ g  )|0 < n >= wn /c]) (25.7)

25.3.3 The Model of Fractal Petri Nets as Elastic Objects

Fractal Petri (FP) nets dynamically synthesized from self-similar Place/Transition

subnets on the base of IAS have been introduced in [18, 19] and analyzed in [20].
Here, the example consumer–producer modeled by FP-net is considered as elastic
object [19]. FP-net consists of one producer, one consumer, and buffer between them:
1. The buffer (BUF) may contain at most three tokens (messages).
2. The producer PRODUCE two tokens in each step and SEND them into buffer.
The production steps of the producer are counted.
3. Consumer RECIEVE one token when accessing the buffer and CONSUME it.
A demand of the consumer is modeled by N 1 , a capacity of the producer is modeled
by N 2 . FP-nets N 1 and N 2 are the elastic objects.
In Fig. 25.9, FP-net composed of two elastic objects (net N 1 and N 2 with shared
resource place—buffer) is shown. For producer and consumer, each of these objects
has elasticity: unfolded or folded.

Fig. 25.9 A
Producer–consumer FP-net t12 SEND 2 t 22
CONSUME
2
p13 p23

p12 3 p22
BUF
2
PRODUCE RECIEVE t 21
t11
p11 p21

N1 N2

[email protected]
25 Essentials of Fractal Programming 385

There are three cases of elasticity in this FP-net: 1:M (one producer to many
consumers), M:1 (many producers to one consumer), M:M (many producers to many
consumers).
FP-net for 1:M (one producer to many consumers) system can be written:

f n (F P, [N n =≡ N 2 , p31 ∈ N 1 # p3n ∈ N n |n > 1]), (25.8)

where # is the operation overlapping places that makes possible to control and
planning capacity of the producer.
FP-net for M:1 (many producers to one consumer) system can be written:

f n (F P, [N n =≡ N 1 , p3n ∈ N n # p32 ∈ N 2 |n > 1]). (25.9)

FP-net for M:M (many producers to many consumers) system can be written:

f n,m (F P, [N n = ≡ N 1 , p3n ∈ N n # p32 ∈ N 2 |n > 1, N m

= ≡ N 2 , p3m ∈ N m # p31 ∈ N 1 |m > 1]). (25.10)

Uniquely invertible operations are used for destructing prototyping subnets of

FP-net. The elastic object planning strategy can be nested in these algorithms, for
example, producer N 1 must change capacity depends on consumer workload N 2 :

f n (F P, [n + +, N n =≡ N 1 , p31 ∈ N 2 # p3n ∈ N n |0 < n =< wn /c],

[n − −, N n =≡−1 N 1 , p31 ∈ N 2 #−1 p3n ∈ N n |0 < n >= wn /c]). (25.11)

25.4 Conclusions

The main contribution of the chapter is the integration theory of fractal objects by
means of IAS, the elastic object models, and object-oriented approach into a new
paradigm called as Fractal-Oriented Programming (FOP). FOP aims to optimize
searching suitable workflow structure of the object at run time and consolidates dif-
ferent statements to elastic systems. FOP is a way of programming that takes into
account the automatic elasticity of objects at run time based on fractal scaling. It does
so by adding additional behavior to existing code modifying the code itself by strat-
egy planning model. Moreover, the properties of the result code remain predictable
due to the fractal properties of elastic objects. This gives a new type of abstraction,
encapsulation, inheritance, modularity, and concurrency of objects, which are indi-
cated in this chapter. FOP forms a basis for fractal-oriented analysis and development
of elastic systems.

[email protected]
386 A. S. Semenov

References

1. Herbst, N.R., Kounev, S., Reussner, R.: Elasticity in cloud computing: what It is, and what it is
not. In: 10th International Conference Autonomic Computing, San Jose, CA, pp. 23–27 (2013)
2. Becker, S., Brataas, G., Lehrig, S. (Eds.): Engineering Scalable, Elastic, and Cost-efficient
Cloud Computing Applications: the Cloudscale Method. Springer International Publishing
AG (2017)
3. Erl, T., Mahmood, Z., Puttini, R.: Cloud Computing: Concepts, Technology & Architecture.
Prentice Hall, New Jersey (2013)
4. Tanenbaum, A.S., Steen, M.: Distributed Systems: Principles and Paradigms. Martin van Steen
(2016)
5. Coulouris, G., Dollimore, J., Kindberg, T., Blair, G.: Distributed systems concepts and design.
Addison Wesley, Boston, Massachusetts (2012)
6. Crownover, R.: Introduction to Fractals and Chaos. Jones and Bartlett Publishers, Inc. (1995)
7. Antonopoulos, N., Gillam, L. (eds.): Cloud Computing Principles, Systems and Applications.
Springer, London Limited (2010)
8. Booch, G., Jacobson, I., Rumbaugh, J. (eds.): Object-Oriented Analysis and Design with
Applications, 3rd edn. Addison-Wesley, Pearson Education, Inc (2007)
9. Erl, T.: Service-Oriented Architecture: Concepts, Technology, and Design. Prentice Hall PTR,
Publisher (2005)
10. Furht, B., Escalante, A. (eds.): Handbook of Cloud Computing. Springer Science-Business
Media, LLC (2010)
11. Wilder, B.: Cloud Architecture Patterns. O’Reilly Media (2012)
12. Noble, J., Taivalsaari, A., Moore, I. (eds.): Prototype-Based Programming: Concepts, Lan-
guages and Applications, 1st edn. Springer (1999)
13. Fehling, C., Leymann, F., Retter, R., Schupeck,W., Arbitter, P. (eds.): Cloud Computing
Patterns: Fundamentals to Design, Build, and Manage Cloud Applications. Springer, Wien
(2014)
14. Mahmood, Z., Hill, R. (eds.): Cloud Computing for Enterprise Architectures. Springer, London
Limited (2011)
15. Semenov, A.S.: Simulation Self-Organised Evolution Processes: Fractoid-Oriented Approach.
Moscow Aviation Institute press, Moscow (in Russia) (2013) Please make this Ref. as [16]
16. Warnecke, H.J.: The Fractal Company: A Revolution in Corporate Culture. Springer, Berlin
(1993) Please make this Ref. as [15]
17. Peitgen, H., Jurgens, H., Saupe, D.: Chaos and Fractals. New Frontiers of Science. Springer,
New York (2004)
18. Semenov, A.S.: Fractal Petri nets. In: 4th International Conference Control, Decision and
Information Technologies, pp. 1174–1179. Barcelona, Spain (2017)
19. Semenov, A.S.: Pattern-type reachability analysis of distributed systems based on fractal
Petri nets. In: 5th International Conference Control, Decision and Information Technologies,
pp. 346–351. Thessaloniki, Greece (2018) Please make this Ref. as [20]
20. Semenov, A.S.: Pattern recognition technique for synthesis fractal Petri nets. In: 6th Interna-
tional Conference Control, Decision and Information Technologies. Paris, France, pp. 1798–
1803 (2019) Please make this Ref. as [19]

[email protected]