www.nature.

com/npjcompumats

ARTICLE OPEN

Electron–phonon physics from first principles using the EPW

code
Hyungjun Lee1,2, Samuel Poncé 3, Kyle Bushick 4, Samad Hajinazar 5,11, Jon Lafuente-Bartolome 1,2, Joshua Leveillee1,2,
Chao Lian1,2, Jae-Mo Lihm 6, Francesco Macheda 7,8, Hitoshi Mori 5, Hari Paudyal 5, Weng Hong Sio1,9, Sabyasachi Tiwari1,2,
Marios Zacharias 10, Xiao Zhang4, Nicola Bonini7, Emmanouil Kioupakis 4, Elena R. Margine 5 and Feliciano Giustino 1,2 ✉

EPW is an open-source software for ab initio calculations of electron–phonon interactions and related materials properties. The
code combines density functional perturbation theory and maximally localized Wannier functions to efficiently compute
electron–phonon coupling matrix elements, and to perform predictive calculations of temperature-dependent properties and
phonon-assisted quantum processes in bulk solids and low-dimensional materials. Here, we report on significant developments in
the code since 2016, namely: a transport module for the calculation of charge carrier mobility under electric and magnetic fields
using the Boltzmann transport equation; a superconductivity module for calculations of phonon-mediated superconductors using
the anisotropic multi-band Eliashberg theory; an optics module for calculations of phonon-assisted indirect transitions; a module for
the calculation of small and large polarons without supercells; and a module for calculating band structure renormalization and
temperature-dependent optical spectra using the special displacement method. For each capability, we outline the methodology
1234567890():,;

and implementation and provide example calculations.

npj Computational Materials (2023)9:156 ; https://doi.org/10.1038/s41524-023-01107-3

INTRODUCTION Several methods have been proposed to address this chal-

The coupling between electrons and phonons is one of the most lenge1. Among those, one of the authors proposed to combine
intensively studied fermion–boson interactions in condensed DFPT with maximally localized Wannier functions (MLWFs)9 to
matter physics. It is responsible for a number of physical perform a physics-based interpolation via a generalized Fourier
phenomena in solids such as conventional superconductivity, transformation10. This method exploits the spatial localization of
temperature-dependent resistivity in metals and mobility in the electron-phonon matrix elements in the Wannier representa-
semiconductors, the formation of polarons, and phonon-assisted tion and enables efficient calculations of electron–phonon matrix
optical processes, to name a few1,2. Predictive non-empirical elements on ultra-dense momentum grids while retaining the
calculations of electron–phonon interactions play an essential role accuracy of DFPT.
The EPW code builds upon this methodology to compute a
in elucidating a variety of materials properties and their
number of properties relating to electron–phonon interactions
temperature dependence.
and temperature-dependent materials properties. EPW is the first
During the past three decades, first-principles calculations of
open-source ab initio software devoted to electron–phonon
electron–phonon interactions have been made possible by the
interactions and has been actively developed for over 16 years.
development of density functional theory (DFT)3,4 and density- It was publicly released in 2010 under a GNU General Public
functional perturbation theory (DFPT)5–7. However, the computa- License (GPL)11 and is being distributed within the Quantum
tional cost of these calculations is high, as they involve the ESPRESSO materials simulation suite12 since 2016. In addition to
evaluation of Brillouin zone integrals which typically require a very EPW, several software projects have been developed in recent
fine sampling of the crystal momenta of electrons and phonons. years to address electron–phonon interactions, including: Per-
For instance, in DFPT calculations, every single phonon wavevec- turbo13, elphbolt14, Phoebe15, and EPIq16. All these
tor and vibrational mode requires the solution of Sternheimer- packages rely on the same interpolation methodology employed
type equations that carry a computational complexity comparable by EPW10.
to a DFT total energy calculation8. As a result, calculating Two prior manuscripts17,18 describe the status of the EPW code
properties related to the electron–phonon interaction usually until 2016. The present manuscript aims to describe progress on
entails a computational workload that is orders of magnitude the code that has occurred since 2016 as well as ongoing efforts,
more costly than standard DFT calculations, thus rendering direct leading to the latest EPW release. Added functionalities involve: a
ab initio calculations of these quantities challenging. magneto-transport module for the calculation of the conductivity

1
Oden Institute for Computational Engineering and Sciences, The University of Texas at Austin, Austin, TX 78712, USA. 2Department of Physics, The University of Texas at
Austin, Austin, TX 78712, USA. 3European Theoretical Spectroscopy Facility, Institute of Condensed Matter and Nanosciences, Université catholique de Louvain, BE-1348 Louvain-
la-Neuve, Belgium. 4Department of Materials Science and Engineering, University of Michigan, Ann Arbor, MI 48109, USA. 5Department of Physics, Applied Physics and
Astronomy, Binghamton University-SUNY, Binghamton, NY 13902, USA. 6Department of Physics and Astronomy, Seoul National University, Seoul 08826, Korea. 7Department of
Physics, King’s College London, Strand, London WC2R 2LS, UK. 8Istituto Italiano di Tecnologia, Graphene Labs, Via Morego 30, I-16163 Genova, Italy. 9Institute of Applied Physics
and Materials Engineering, University of Macau, Macao, SAR 999078, P. R. China. 10University of Rennes, INSA Rennes, CNRS, Institut FOTON-UMR 6082, F-35000 Rennes, France.
11
Present address: Department of Chemistry, University at Buffalo, Buffalo, NY 14260, USA. We dedicate this manuscript to our dearest friend and beloved collaborator, Professor
Nicola Bonini, who passed away in October 2022. Nicola was a source of inspiration and joy to us all and played a key role in the development of EPW.
✉email: [email protected]

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
2
in metals as well as the drift and Hall mobility in semiconductors; a Phonon-limited carrier transport using the ab initio
module for the calculation of small and large polarons without Boltzmann transport equation
using supercells; a module for the solution of the full-band Background and formalism. The calculation of the electronic
Eliashberg equations for superconductors; a module for calcula- transport properties of metals and semiconductors is conveniently
tions of phonon-assisted indirect optical processes; and a module dealt with by the ab initio Boltzmann transport equation (aiBTE).
for calculations of finite-temperature electronic and optical The Boltzmann equation describes the non-equilibrium distribu-
properties via the special displacement method. The code has tion function of electrons and holes in the presence of external
been refactored to keep up with contemporary developments in electric or magnetic fields2. It carries strong predictive power as
high-performance computing (HPC) architectures, for example via recently demonstrated for many common semiconductors22.
hybrid two-level message passing interface (MPI19) and Open Although the Boltzmann formalism is usually derived within the
Multi-Processing (OpenMP20) parallelization, and parallel I/O via semi-classical approximation, the theory can rigorously be derived
parallel Hierarchical Data Format 5 (HDF521). from a non-equilibrium many-body Green’s function formalism
The manuscript is organized as follows. In the ‘Methods’ section and is understood as the quasiparticle approximation to the
(located at the end of the manuscript), we describe the notation Kadanoff–Baym theory23–26.
The EPW code implements the linearized Boltzmann transport
employed throughout this manuscript, we review the conceptual
equation, which describes the first-order response of the
basis of electron–phonon interpolation using DFPT and MLWFs, and
distribution function to external fields and is suitable for
we discuss the extension of this methodology to the case of polar investigating the conductivity of metals and the low-field mobility
materials with long-range Coulomb interactions. In the ‘Results’ of semiconductors. In this section, we focus on the case of
section, we describe the additional or expanded functionalities phonon-limited transport properties. The extension to include
available in the latest release of the EPW code. In particular, we charged defects is discussed in the next section.
begin by discussing calculations of carrier transport within the ab We consider homogeneous extended solids held at uniform
initio Boltzmann transport equation. We consider both electric and temperature, and carrier scattering by electron–phonon interac-
magnetic fields, as well as phonon-limited and charged defect- tions only, for now. Within the aiBTE, the linear response ∂Eβ f nk of
limited transport. Then we move to calculations of the super- the carrier distribution function to an external electric field E is
1234567890():,;

conducting critical temperature and superconducting gap function. given by22,27–41 Z

Here, we discuss elementary calculations based on the semi- ∂f 0nk P dq
∂Eβ f nk ¼ ev nkβ ∂εnk τ nk þ 2πτ_ nk jg ðk; qÞj2 ∂Eβ f mkþq
empirical McMillan equation, the isotropic Eliashberg theory, and mν ΩBZ mnν

the fully-anisotropic full-band Eliashberg theory. Next, we outline ´ ðnqν þ 1  f 0nk Þδðεnk  εmkþq þ _ωqν Þ
the methodology for computing small and large polarons from 
DFPT without resorting to large supercells. We then proceed with þðnqν þ f 0nk Þδðεnk  εmkþq  _ωqν Þ :
the formalism for the calculation of optical absorption spectra (1)
including phonon-assisted indirect transitions. After these applica- In this expression, fnk is the electron distribution function, n
tions which are based on Wannier–Fourier interpolation, we denotes a Kohn–Sham band index, and k is the electron
describe the special displacement method, which offers an wavevector. ∂Eβ f nk is a short-hand notation for (∂fnk/∂Eβ)∣E=0,
alternative and complementary strategy to computing vnkα = ℏ−1∂εnk/∂kα is the intra-band velocity matrix element for
electron–phonon properties without employing Wannier–Fourier the Kohn–Sham eigenvalue εnk, and δ denotes the Dirac delta
interpolation. At the end of this section, we discuss recent advances function. gmnν(k, q) is the electron–phonon matrix element
in computational algorithms, parallelization, and I/O, and we report describing the scattering from the state nk to the state mk + q
on parallel scaling benchmarks. In the ‘Discussion’ section we via the phonon of branch ν and wavevector q (see the “Methods”
discuss possible future directions for the EPW software project. section for more details). The integral is over the Brillouin zone.
The temperature enters this equation via the Fermi–Dirac and
Bose–Einstein equilibrium distribution functions f 0nk and nqν,
respectively. The quantity τnk in Eq. (1) is the carrier relaxation
time, and is obtained
Z from Fermi’s golden rule:
RESULTS P dq
In this section, we provide an overview of additional or expanded τ 1
nk ¼ _
2π
jg ðk; qÞj2
mν ΩBZ mnν
capabilities of the EPW code. In particular, we discuss how EPW h
´ ðnqν þ 1  f 0mkþq Þδðεnk  εmkþq  _ωqν Þ (2)
calculates carrier transport properties via the ab initio Boltzmann
transport equation (aiBTE), including both carrier-phonon and i
carrier-impurity scattering; how the superconducting critical þðnqν þ f 0mkþq Þδðεnk  εmkþq þ _ωqν Þ :
temperature and superconducting gap are computed via the The electrical conductivity tensor σαβ, which corresponds to the
solution of the Eliashberg equations; how we solve the ab initio variation of the current density with respect to the electric field, is
polaron equations to investigate small and large polarons without computed from ∂Eβ f nk using:
using large supercells; and how we perform calculations of optical Z
absorption spectra in indirect-gap semiconductors by including eX dk
σ αβ ¼  v nkα ∂Eβ f nk ; (3)
phonon-assisted optical transitions. We also describe an alternative Ω n ΩBZ
to Wannier interpolation to study electron–phonon interactions,
based on the special displacement method. Of these features, the and the drift mobility is obtained from this expression upon
dividing by the carrier density nc:
superconducting module has been enhanced and expanded with
respect to the previous release in 201618; all the other modules σαβ
described in the following have been developed ex novo since the μαβ ¼ : (4)
e nc
2016 release. The notations and conventions employed below are
These equations are valid for metals and for electrons or holes in
described in detail in the ‘Methods’ section. In the same section, we
semiconductors taken separately. A common approximation to
provide an overview of the Wannier–Fourier interpolation of the aiBTE is the relaxation time approximation, which consists of
electron–phonon matrix elements.

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
3
neglecting the second term on the right-hand side of Eq. (1). In electronic state lie within a user-defined energy window around
this case the mobility takes the simpler form: reference energy. In particular, the wavevectors k and q are
Z retained only if there exists a pair of bands m and n such that
e X dk ∂f 0nk
μSERTA
αβ ¼ v nkα v nkβ τ nk : (5) ∣εnk−Eref∣ < Δ and ∣εmk+q−Eref∣ < Δ, where Eref is reference energy
nc Ω n ΩBZ ∂εnk and Δ is the width of the window. In the case of metals, the
We refer to this approximation as the “self-energy relaxation time reference energy is set to the Fermi energy. In the case of
approximation” (SERTA) since the transport lifetime in Eq. (2) semiconductors, the Fermi energy is calculated from the user-
corresponds to the quasiparticle lifetime, which is proportional to specified temperature and carrier concentration, using the
the imaginary part of the electron-phonon self-energy1,22,24. bisection method.
The electrical conductivity and the drift mobility described by Magneto-transport calculations using Eq. (6) are harder to
Eq. (1) can be measured by time-of-flight measurements or by THz converge than calculations without magnetic fields. The iterative
photo-conductivity measurements. In the case of Hall and De solution is initialized by using the result of Eq. (1) as a first
Haas–Van Alphen measurements, an additional magnetic field B is approximation for ∂E β f nk ðBÞ, and the quantity ∇k ∂Eβ f nk ðBÞ
applied, and the resulting Lorentz force must be taken into appearing on the left-hand side of Eq. (6) is evaluated via finite
account in the aiBTE. In these cases, EPW solves the following differences. The correctness of the solution is tested by checking
aiBTE equation22,24,32,42: for the conservation
P of the carrier density, which corresponds to
the condition nk ∂Eβ f nk ðBÞ ¼ 0.
  ∂f 0
1  _e τ nk ðvnk ´ BÞ  ∇k ∂E β f nk ðBÞ ¼ ev nkβ ∂εnk τ nk To ensure numerically accurate evaluation of transport coeffi-
Z nk
cients with EPW, it is important to verify that results be converged
2πτ nk P dq 2
þ _ jg ðk; qÞj ∂E β f mkþq ðBÞ with respect to (i) the density of the Brillouin zone grids employed
mν ΩBZ mnν
  in Eqs. (1) and (6); (ii) the density of the coarse grids used for the
´ ðnqν þ 1  f nk Þδðεnk  εmkþq þ _ωqν Þ þ ðnqν þ f 0nk Þδðεnk  εmkþq  _ωqν Þ :
0
Wannier–Fourier interpolation of the DFT and DFPT data22; (iii) the
(6) Gaussian smearing employed to compute the Dirac delta functions
The variation ∂Eβ f nk ðBÞ is computed for magnetic fields sufficiently appearing in Eq. (1); and (iv) to check for the sensitivity of results to
small that their effect on the electronic and vibrational properties lattice parameters, exchange and correlation functionals, and
can be ignored; accordingly, the Kohn–Sham energies, phonons, pseudopotentials. To facilitate the convergence with respect to the
and their couplings in the above equation are all evaluated for Gaussian smearing and the k-point sampling, EPW offers the
B = 0. The application of a magnetic field in the direction B ^ results possibility of using the adaptive smearing method of ref. 44.
in an orthogonal flow of the charge carriers which can be
described by the linear response of the mobility to the field: Application example. To demonstrate the implementation of the
" Z # transport module in EPW, we investigate the electron and hole
ð2Þ ^ 1 1 X dk mobilities of cubic boron nitride, as well as the associated Hall
μαβ ðBÞ ¼  lim v nkα ∂Eβ f nk ðBÞ  μαβ : (7)
B!0 jBj nc Ω ΩBZ factors. We use the relativistic Perdew–Burke–Ernzerhof (PBE)
n
parametrization45 of the generalized gradient approximation to
From this expression, we obtain the Hall mobility in terms of the DFT. The pseudopotentials are norm-conserving, generated using
drift mobility as follows: the ONCVPSP code46, and optimized via the PseudoDojo
X
^ ¼
μHαβ ðBÞ ^ γβ ;
r Hαγ ðBÞμ initiative47. We consider room temperature and low carrier
(8) concentrations of 1013 cm−3. c-BN is a polar wide gap semicon-
γ
ductor with isotropic Born effective charges Z B ¼ 1:91 e and
where we have introduced the dimensionless Hall factor as22
Z N ¼ Z B ; isotropic dynamical quadrupoles Qκ,αβγ = Qκ∣ϵαβγ∣
X where ϵαβγ is the Levi–Civita tensor and Qκ an atom-dependent
^ ¼
r Hαγ ðBÞ μ1
ð2Þ ^ 1
αγ μγδ ðBÞμδβ : (9) scalar with values QB = 3.46ea0 and QN = −0.63ea0, where a0 is
γδ
the Bohr radius; and an isotropic high-frequency relative dielectric
Also in this case, one can simplify the solution of Eq. (6) by constant ϵ∞ = 4.5422. As will be discussed in Fig. 9, long-range
ignoring the second term on the right-hand side. This choice leads electron–phonon interactions are important in the case of c-BN.
to the SERTA approximation in the presence of a magnetic field. Wannier functions are calculated separately for the valence and
the conduction band manifold in order to reduce the
Computational considerations. EPW implements an iterative computational cost.
solver to obtain ∂Eβ f nk from Eq. (1), with the possibility of Broyden Figure 1a shows the convergence of calculated mobilities with
mixing43 to accelerate the convergence. The accurate evaluation the coarse grid of k-points used in the Wannierization procedure.
of Eqs. (3) and (4) is computationally demanding because it For this test, we employ a coarse q-point grid with half the
requires the knowledge of ∂Eβ f nk for a dense set of k-points in an number of the k-points; and we employ identical fine grids with
energy window of the order of a few tens of meV around the 503k- and q-points for the solution of Eq. (1); we use an energy
Fermi energy. In particular, since the scattering integral on the window Δ = 0.3 eV around the reference energy Eref set to the
right-hand side of Eq. (1) couples linear response coefficients at k band edge, and we employ adaptive smearing. This panel shows
and k + q points, the k-grid and the q-grid must be commensu- that the mobilities are converged with an accuracy of 1% when
rate. As this term is not evaluated in the SERTA approximation, coarse grids with 183 and 143 points are employed for electrons
SERTA calculations can instead be performed using incommensu- and holes, respectively. Figure 1b shows the convergence of
rate grids, including for example random and quasi-random calculated mobilities with respect to the fine grids. We see that
sampling31. It should be noted, however, that the SERTA both electron and hole mobilities are almost converged for grids
approximation tends to underestimate transport coefficients by consisting of 2503 points. Generally speaking, the fine grid
up to 50% as compared to the aiBTE22. required to achieve converged results depends on the effective
Regardless of the approximation chosen to calculate transport masses, with smaller masses requiring denser grids. As an
coefficients, EPW exploits crystal symmetry operations to eliminate example, accurate calculations of the electron mobility of GaAs
symmetry-equivalent wavevectors in the solution of Eq. (1); require homogeneous grids of order 4003 points, while in the case
accordingly, ∂E β f nk is evaluated within the irreducible wedge of of AlAs or GaP 803 points are sufficient22 (see Table 1).
the Brillouin zone. Furthermore, the solution of Eq. (1) is restricted In Fig. 1c we show convergence curves for the mobility
to the set of wavevectors for which both the initial and the final evaluated in the SERTA approximation. In this case, the

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
4

Fig. 1 Calculations of phonon-limited carrier mobility using EPW. a Convergence test for the electron and hole mobilities of c-BN with
respect to the number of k-points on the coarse grid. The number of q-points is set to half the number of k-points. b Convergence of
mobilities w.r.t. the number of grid points on the fine grids. The number of q-points is set to be the same as the number of k-points. c Same as
in b, but this time for calculations within the SERTA approximation. d and e Hall factors of c-BN for electrons and holes, as a function of the
number of points on the fine grids, for the full aiBTE and the SERTA approximation, respectively. The number of q-points is set to be the same
as the number of k-points. f Dependence of the electron drift mobility of c-BN with the Gaussian smearing.

convergence with the density of points in the fine grids is much respectively. The Hall factor in the SERTA approximation appears
slower than for complete aiBTE calculations. More importantly, to be close to the full aiBTE result, which is expected since it is
SERTA mobilities can differ by up to a factor of two from the defined as the ratio of two mobilities (see Eq. (9))48.
corresponding aiBTE results. The large difference between full In Fig. 1(f) we analyze the role of the Gaussian smearing parameter
aiBTE calculations and SERTA seems to be a common trend in in the calculations. To this end, we compute the electron mobility of
polar materials22. For the same k- and q-point grids, the c-BN as a function of grid size, for varying smearing parameter. It is
computational saving afforded by the SERTA approximation is seen that the size of the grid necessary for convergence increases
minimal; therefore, we recommend using SERTA only when the when decreasing the smearing in the few meV’s range. On the other
use of commensurate grids is too demanding, for example in the hand, the use of adaptive smearing affords fast convergence even
case of systems with many atoms in the unit cell. Figure 1d, e for smaller grid sizes. Based on this comparison, adaptive smearing is
show the computed Hall factors for electrons and holes in c-BN, the computationally most convenient strategy.

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
5
imp
Table 1. Typical size of fine grids employed in EPW calculations. partial scattering rate 1=τ nk!mkþq is calculated under the
following simplifying approximations: (i) each impurity is
EPW module Typical size of fine Brillouin zone described by an idealized point charge, embedded in the
grid dielectric continuum of the host material; (ii) the scattering rate
Minimum Maximum from each impurity is evaluated within the first Born approxima-
tion; (iii) defects are sufficiently diluted that the scattering rates
Phonon-limited transport 80 × 80 × 80 400 × 400 × 400 from different impurities are additive; (iv) impurities are randomly
Impurity-limited transport 100 × 100 × 100 180 × 180 × 180 distributed, and this random distribution is formally taken into
account by using the Kohn and Luttinger ensemble average58.
Superconductivity 60 × 60 × 60 120 × 120 × 120
Within these approximations, the charged impurity scattering rate
Polarons 3×3×3 50 × 50 × 50 reads57:
Phonon-assisted absorption 32 × 32 × 32 48 × 48 × 48  2
2π e2 4πZ X jhψmkþq jeiðqþGÞr jψnk ij
2
1
Special displacement method 3×3×3 10 × 10 × 10 ¼ Nimp 2 δðεnk  εmkþq Þ;
imp
τ nk!mkþq _ 4πϵ0 Ω G≠q jðq þ GÞ  ϵ  ðq þ GÞj
0

In the case where the fine grids for k- and q-points differ, we indicate the (12)
size of the denser grid. These values are only indicative and correspond to
the examples mentioned in the ‘Results’ section, which refer to simple where Nimp is the number of impurities per crystal unit cell
semiconductors, metals, and insulators with 2–4 atoms per unit cell. For (dimensionless), Ze is the charge of each impurity, and ϵ0 is the
systems with larger unit cells, the fine grids would have to be downscaled static relative dielectric constant tensor. The scattering rate in Eq.
accordingly. (12) takes into account the DFT electronic band structure and the
lattice screening, including possible dielectric anisotropy. The
main advantage of this model as compared to explicit calculations
Defect-limited carrier transport using the ab initio Boltzmann of charged defects in supercells is that it can be used system-
transport equation atically without requiring detailed knowledge of the defect
Background and formalism. Electron–phonon interactions are the physics and energetics in each material.
dominant scattering mechanism in high-purity single crystals with
low defect concentration, typically up to 1015 cm−3 49. In Computational considerations. In the EPW code, the scattering
tetrahedral semiconductors and a host of other materials, charged rate given in Eq. (12) is added to the electron–phonon scattering
defects are ubiquitous since dopant elements are employed to rate using Eqs. (11) and (10), and the aiBTE is solved as already
introduce free carriers50. Donors release electrons to the conduc- described for the phonon-limited mobility. The computational
tion band, thus becoming positively charged defects; similarly, overhead as compared to phonon-only calculations is minimal.
acceptors release holes into the valence band, and become In Eq. (12), the sum over G-vectors is handled via a Gaussian
negatively charged defects. In either case, the impurity generates filter, as described in the “Methods” section; the reader is
a long-ranged Coulomb potential that scatters charge carriers. This referred to that section for more details. The overlap integrals
scattering mechanism tends to dominate over electron–phonon between the initial and final Kohn–Sham states appearing in
processes at high doping concentrations. Eq. (12), 〈ψmk+q∣ei(q+G)⋅r∣ψnk〉, are evaluated by means of the
There exist popular semi-empirical relations to estimate the unitary Wannier function matrices described in the Methods.
effect of ionized impurity scattering on carrier transport in The integral over the scattering wavevectors q in Eq. (10) of
semiconductors, such as the Brooks–Herring51 and the impurity scattering rates given in Eq. (12) contains a singular
52
Conwell–Weisskopf formulas. However, these expressions rely ∣q∣−4 term that is not integrable. This issue is resolved by
on simplified parabolic band models and do not carry predictive introducing the screening of the defect potential by free carriers
power in the case of materials with multi-band or multi-valley released upon ionization. To this end, we replace ϵ0 in Eq. (12)
band extrema. First-principles calculations offer a modern alter- with the total dielectric function:
native to these earlier approaches, and have achieved consider- q2TF
able success in recent years53–56. The EPW code implements a ϵ0tot ¼ ϵ0 þ I; (13)
q2
module for charged defect scattering based on a randomized
distribution of point charges. This approach is described in detail where I denotes the identity matrix, and qTF is the Thomas–Fermi
in ref. 57 and summarized below. wavenumber obtained from the long-wavelength limit of the
For convenience we rewrite the aiBTE from Eq. (1) in the more Lindhard function49,56,59:
Z  
compact form: e2 4π X dk ∂f 0nk 
∂f 0 X Z dq   q2TF ¼
4πϵ Ω
2
Ω ∂ε 
: (14)
ev nkβ nk ¼ τ 1 ∂ E f mkþq  τ 1
∂E f nk ; 0 n BZ nk
∂εnk ΩBZ mkþq!nk β nk!mkþq β
m Temperature enters this expression via the equilibrium
(10) Fermi–Dirac distribution of the electrons or holes, f 0nk .
where the quantity τ 1 is the partial scattering rate from the The concentration of charged defects is an external input
nk!mkþq
Kohn–Sham state nk to the state mk + q. In the case of parameter in these calculations. In the case of ionized impurities in
electron–phonon scattering, τ 1 semiconductors, it is also possible to use a simple estimate for the
nk!mkþq is obtained from Eq. (2)
by removing the summation over m and the integral over the fraction fii of ionized impurities at a given temperature by using
P R the impurity energy level in the gap εd49,56,60.
Brillouin zone, i.e. m dq=ΩBZ . When both electron–phonon !
Z
scattering and charged impurity scattering are taken into account, 1 X dk
the partial scattering rate is written as the sum of the individual f ii ¼ max 1;
2
f ðεnk  εd Þ ; (15)
partial rates nd Ω n ΩBZ
1 1 1 where f is the Fermi–Dirac distribution and nd is the total
¼ þ imp ; (11) concentration of defects that may thermally ionize.
τ nk!mkþq τ ph τ
nk!mkþq nk!mkþq
Owing to the ∣q∣−4 scaling of the ionized-impurity scattering
where the superscripts refer to carrier-phonon (ph) and carrier- matrix elements, achieving convergence in the presence of
impurity (imp) scattering, respectively. In EPW, the carrier-impurity impurities is slightly more demanding than for calculations of

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
6

Fig. 2 Calculations of phonon- and impurity-limited carrier mobility using EPW. a Mobility of silicon as a function of temperature, for the
following three cases: phonon scattering only (black disks and lines); phonon and ionized impurity scattering, with an impurity concentration
of 1.75 × 1016 cm−3 (indigo disks and lines), and with a concentration of 1.3 × 1017 cm−3 (green disks and lines). Experimental data for the
same concentrations are shown as diamonds of the same color62,63,219. b Same as in (a) but for the hole mobility of silicon. The impurity
concentrations are 2.4 × 1016 cm−3 (indigo) and 2.0 × 1017 cm−3 (green), respectively. c Electron mobility of silicon at 300 K, as a function of
ionized impurity concentration: black disks and lines are calculations, black diamonds are experiments64,66. d Same as in (c) but for the hole
mobility.

phonon-limited mobility. For example, in recent calculations of Eq. (12). The calculations agree well with experiments up to an
carrier mobilities in Si, GaP, and SiC, fine grids with 1003, 1003, and impurity concentration around 1018 cm−364. Beyond this concen-
1803 points were necessary, respectively57 (see Table 1). tration, it is expected that additional mechanisms such as two-
impurity scattering and plasmon scattering65 will further reduce
Application example. To illustrate the impurity scattering cap- the mobility. The hole mobility, which is shown in Fig. 2d, follows a
ability of EPW, we calculate the mobility of electrons and holes in similar trend. The calculated hole mobility at low doping,
silicon as a function of both temperature and dopant concentra- 603 cm2 V−1 s−1, slightly overestimates the experimental range
tion. We consider coarse Brillouin zone grids with 123k-points and 450–500 cm2 V−1 s−1; this effect can be traced to the under-
63q-points, and fine grids with 1003 points. We use the PBE estimation of the heavy hole mass by DFT31. Upon increasing the
functional45 and ONCV pseudopotentials46,61, and we include impurity concentration, the mobility decreases following the same
spin–orbit coupling for the valence bands. We account for trend as for the electrons. These calculations agree well with
quadrupole corrections using the procedure described in the experiments64,66 and with previous first-principles calculations56.
Methods section.
Figure 2a shows the calculated electron mobility of silicon as a
function of temperature. When considering phonon scattering Phonon-mediated superconductivity using the ab initio
only, our calculations are in very good with measurements on Eliashberg theory
high-purity silicon62. Upon introducing ionized impurity scattering, Background and formalism. First-principles calculations of
the electron mobility at 100 K decreases from 11,813 to phonon-mediated superconductors are primarily based on three
4725 cm2 V−1 s−1 for a dopant concentration of 1.75 × 1016 cm−3, approaches, namely semi-empirical methods based on the
and to 1769 cm2 V−1 s−1 for a dopant concentration of McMillan formula67, the ab initio Eliashberg theory68,69, and the
1.3 × 1017 cm−3. These calculations agree well with experimental density-functional theory for superconductors70–73. The EPW code
data63. At higher temperatures, the reduction in the mobility due implements the former two approaches. In this section, we briefly
to impurity scattering is less significant since phonons provide the review methods based on the McMillan formula, and then we
dominant scattering mechanism. The hole mobility in Fig. 2b describe the Eliashberg formalism and its implementation.
exhibits a similar trend. Upon introducing ionized impurity The superconducting critical temperature Tc can be estimated
scattering, the hole mobility at 100 K decreases from 8877 to using the semi-empirical McMillan equation in Allen–Dynes
2884 cm2 V−1 s−1 for a dopant concentration of 2.4 × 1016 cm−3, form74:
and to 1056 cm2 V−1 s−1 for a dopant concentration of  
_ωlog 1:04ð1 þ λÞ
2.0 × 1017 cm−3. Also in this case, the agreement with experi- c ¼
k B T AD exp ; (16)
mental data is very good63. 1:2 λ  μ ð1 þ 0:62λÞ
Figure 2c shows the calculated electron mobility of silicon at where kB is Boltzmann’s constant, μ* is the semi-empirical
300 K, as a function of ionized impurity concentration. Up to a Coulomb pseudopotential, ωlog is the logarithmic average of the
dopant concentration of 1016 cm−3, phonon scattering dominates phonon frequencies, and λ is the electron–phonon coupling
and the mobility is relatively insensitive to impurity scattering. constant, as defined in ref. 1. The parameter μ* is largely material
Beyond this concentration, the mobility decreases sharply and independent, for a large class of superconductors a value in the
approximately as the inverse of the impurity density, in line with 0.1-0.15 range being typically required to fit the experimental

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
7
Tc74–76. Note that for very strong coupling (λ > 1.5), in ref. 74 two anisotropic electron–phonon coupling parameters λðnk; mk þ
correction prefactors f1 and f2 are introduced in Eq. (16) to better q; ωj  ωj0 Þ appearing in these equations are calculated as
estimate the critical temperature in this regime. NðεF Þ X 2ωqν
In addition, the code allows the user to calculate the critical λðnk; mk þ q; ωj  ωj0 Þ ¼ jgmnν ðk; qÞj2 :
_ ν ðωj  ωj0 Þ2 þ ω2qν
temperature using a more recent prescription based on a machine
learning approach77: (24)
¼ Equations (20)–(23) are supplemented by a statement of particle
T ML
c c ;
f ω f μ T AD (17)
number conservation, which determines the chemical potential84:
where the correction factors fω and fμ are given by " #
X Z dk X εnk  μF þ χ ðiωj Þ
λ þ ωlog =ω2  ðμ Þ1=3 6:86 expðλ=μ Þ Ne ¼ 1  2k B T nk
; (25)
f ω ¼ 1:92  0:08; fμ ¼ þ 1: ΩBZ θnk ðiωj Þ
1=2
λ expðωlog =ω2 Þ λ1  μ  ωlog =ω2 n j

(18) where Ne is the number of electrons per unit cell. Equations

(20)–(25) are referred to as the anisotropic full-bandwidth (FBW)
In these expressions, ω2 is the square root of the second moment
Eliashberg equations92 since they explicitly take into account
of the normalized weight function g(ω) = 2α2F(ω)/(ωλ), with α2F
scattering processes involving electrons with energies and
being the Eliashberg spectral function, as defined in ref. 1.
momenta that are not restricted to the vicinity of the Fermi
Equations (16) and (17) are useful for preliminary calculations but
surface. To find the temperature-dependent superconducting gap,
do not carry the predictive power of the ab initio Eliashberg
in EPW these equations are solved iteratively for different
theory.
temperatures; the highest temperature for which a non-trivial
The Eliashberg theory68,69 describes the superconducting phase
solution exists (ϕnk ≠ 0) is the superconducting critical tempera-
transition by means of finite-temperature Green’s functions. In this
ture Tc. Below this temperature, the superconducting gap Δnk is
theory, superconducting pairing arises from an attractive
given by
electron–electron interaction mediated by phonons, which is
partly compensated by the inter-electron Coulomb repulsion. The Δnk ðiωj Þ ¼ ϕnk ðiωj Þ=Z nk ðiωj Þ : (26)
Eliashberg theory can be formulated as a Dyson equation for a
The numerical solution of Eqs. (20)–(25) is computationally
generalized 2 × 2 matrix Green’s function via the Nambu-Gor’kov
demanding. A common simplification of these equations consists
formalism78,79. The off-diagonal elements of this matrix describe
of restricting the energy range close to the Fermi level75,80–86,93. In
Cooper-pair amplitudes in the superconducting state and are
this approach, it is assumed that the DOS within this energy
related to the superconducting gap function. These elements
window is constant. It can be shown that, within these
become nonzero below the critical temperature, marking the
approximations, the energy shift χnk vanishes and the requirement
transition to the superconducting state. It is standard practice to
in Eq. (25) is automatically satisfied. As a result, only two equations
expand the pairing self-energy using Pauli matrices ^τ i (i = 0, …, 3)
for Znk and ϕnk need to be solved self-consistently:
as follows75,80–86:
  P
Z
ωj0
^nk ðiωj Þ ¼ i_ωj 1  Z nk ðiωj Þ ^τ 0 þ χ nk ðiωj Þ^τ 3 þ ϕnk ðiωj Þ^τ 1 ;
Σ (19) πk B T
Z nk ðiωj Þ ¼ 1 þ Nðε
dq
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
F Þωj ΩBZ _2 ω20 þ Δ2
mkþq ðiωj 0 Þ
mj 0 (27)
where iωj = i(2j + 1)πT is the fermionic Matsubara frequency with j j

´ λðnk; mk þ q; ωj  ωj Þδðεmkþq  εF Þ; 0
being an integer, T is the absolute temperature, Znk(iωj) is the mass
renormalization function, χnk(iωj) is the energy shift, and ϕnk(iωj) is P
Z
dq Δmkþq ðiωj0 Þ
πk B T
the order parameter. This self-energy is expressed in terms of the Z nk ðiωj ÞΔnk ðiωj Þ ¼ Nðε FÞ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ΩBZ _2 ω20 þ Δ2
mkþq ðiωj 0 Þ
0
mj
electron Green’s function using the Migdal approximation87 for 
j

the electron–phonon contribution, and the GW approximation for ´ λðnk; mk þ q; ωj  ωj0 Þ  NðεF ÞV nk;mkþq δðεmkþq  εF Þ:
the electron–electron contribution88,89. Using Eq. (19) inside the (28)
Dyson equation for the electron Green’s function yields a set of
coupled equations for Znk, χnk, and ϕnk: These equations are referred to as the anisotropic Fermi surface
restricted (FSR) Eliashberg equations18,85.
Z To extract physical quantities of interest, such as the tunneling
kB T X dq ωj0 Z mkþq ðiωj0 Þ density of states and the heat capacity, the gap function on the
Z nk ðiωj Þ ¼ 1 þ λðnk; mk þ q; ωj  ωj0 Þ;
ωj NðεF Þ mj0 ΩBZ θmkþq ðiωj0 Þ real frequency axis is required75,81,82. In the EPW code, the
continuation of Δnk(iωj) from the imaginary to the real axis is
(20) performed either using Padé approximants94,95, or using the exact
Z
analytic continuation of ref. 96, as described in ref. 85.
kB T X dq εmkþq  μF þ χ mkþq ðiωj0 Þ Accounting for the Coulomb repulsion in the Eliashberg equations
χ nk ðiωj Þ ¼  λðnk; mk þ q; ωj  ωj0 Þ;
NðεF Þ mj0 ΩBZ θmkþq ðiωj0 Þ requires the evaluation of the matrix elements Vnk,mk+q. While this is
feasible in principle, it is common practice to replace the product
(21)
N(εF)Vnk,mk+q with the semi-empirical Morel–Anderson pseudopo-
Z tential μc 97. In this approach, the value of μc is linked to the cutoff
kB T X dq ϕmkþq ðiωj0 Þ   frequency ωc for the Matsubara summation on the imaginary axis,
ϕnk ðiωj Þ ¼ λðnk; mk þ q; ωj  ωj0 Þ  NðεF ÞV nk;mkþq ;
NðεF Þ mj0 ΩBZ θmkþq ðiωj0 Þ with ωc typically set to ten times the maximum phonon frequency
(22) ωph. Note that when converting between the μ* in Allen–Dynes
formula in Eq. (16) and μc , a ratio ωc/ωph of ten is often used74. In
having defined the auxiliary function: EPW, μc is specified by the user as an external parameter, and in
 2  2  2 many applications, values in the range μc ¼ 0:1-0.2 yield reasonable
θnk ðiωj Þ ¼ _ωj Z nk ðiωj Þ þ εnk  μF þ χ nk ðiωj Þ þ ϕnk ðiωj Þ :
agreement with experiments. More recently, first-principles calcula-
(23) tions of μc as a Fermi surface average of Vnk,mk+q have been used in
In Eqs. (20)–(23), N(εF) is the density of states (DOS) per spin at the conjunction with EPW91,98–100.
Fermi level, μF is the chemical potential, and the quantities Vnk,mk In the case of simple superconductors which do not exhibit
+q denote the matrix elements of the screened Coulomb significant anisotropy, it may be a good approximation to neglect the
interaction W between electron pairs, as given in refs. 90,91. The band and momentum dependence of the superconducting gap. In

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
8
these cases, instead of Eqs. (20)–(25), it is possible to solve a than FSR calculations. In particular, to determine the chemical
simplified version of the equations where all properties are averaged potential from Eq. (25) one needs a Matsubara frequency cutoff of at
over the Fermi surface76,101–103. The relations are referred to as the least a few times the Kohn–Sham energy window, leading to a
isotropic Eliashberg equations: considerable increase in computational cost. To circumvent this
Z X ωj0 Zðiωj0 Þ issue, the default behavior of EPW is to keep the chemical potential
kBT
Zðiωj Þ ¼ 1 þ dεNðεÞ λðωj  ωj0 Þ; (29) fixed at the Fermi level unless it is instructed by the user to solve Eq.
Nðε Þω F j θðε; iω 0 Þ j0 j (25). In this case, careful convergence tests with respect to the
Z Matsubara frequency cutoff are warranted.
kBT X ε  μF þ χðiωj0 Þ
χðiωj Þ ¼  dεNðεÞ λðωj  ωj0 Þ; (30)
NðεF Þ j0
θðε; iωj0 Þ Application example. To demonstrate the implementation of the
superconducting module in EPW, we investigate the super-
Z X ϕðiωj0 Þ  conducting properties of hexagonal niobium diselenide (2H-
kB T 
ϕðiωj Þ ¼ dεNðεÞ λðωj  ωj0 Þ  μc ; (31) NbSe2). 2H-NbSe2 exhibits a superconducting phase transition and
NðεF Þ j0
θðε; iωj 0Þ
a charge density wave instability below 7.2 and 33 K, respectively;
" # both phases are driven by a strong momentum-dependent
Z X ε  μF þ χðiωj Þ electron-phonon coupling107–112. For brevity, in this section, we
Ne ¼ dεNðεÞ 1  2k B T ; (32) focus on superconductivity and ignore the lattice instability.
θðε; iωj Þ
j We employ the PBE exchange and correlation functional45 and
where the counterpart of Eq. (23) is the optB86b-vdW van der Waals functional113–116, as well as ONCV
 2  2  2 pseudopotentials46,61 including semi-core electrons for Nb. We
θðε; iωj Þ ¼ _ωj Zðiωj Þ þ ε  μF þ χðiωj Þ þ ϕðiωj Þ : (33) use a plane waves kinetic energy cutoff of 80 Ry, Γ-centered
In Eqs. (29)–(31), λðωj  ωj0 Þ is a momentum-averaged version of 24 × 24 × 12 k-point and 6 × 6 × 4 q-point coarse grids, and
Eq. (24): Methfessel–Paxton smearing117 of 0.025 Ry. The optimized lattice
parameters are a = 3.46 Å and c = 12.57 Å, in good agreement
1 XZ dk
Z
dq
with the experimental data a = 3.43 Å and c = 12.55 Å108.
λðωj  ωj0 Þ ¼ λðnk; mk The Eliashberg equations are solved on a uniform 60 × 60 × 40
½NðεF Þ2 m;n
ΩBZ ΩBZ (34) k-point and 30 × 30 × 20 q-point fine grids. We consider energy
þq; ωj  ωj0 Þδðεnk  εF Þδðεmkþq  εF Þ: windows of 0.4 and 0.8 eV, Matsubara frequency cutoffs of 0.4, 0.8,
and 1.2 eV, and a Coulomb pseudopotential of μc ¼ 0:2. The Dirac
Similar to the anisotropic case, Eqs. (29)–(31) can also be reduced deltas in the FSR approach are replaced by Gaussians of width
to a set of two equations for Z(iωj) and ϕ(iωj), under the same 50 meV.
conditions as discussed for Eqs. (27) and (28). Figure 3a shows the band structure and DOS of 2H-NbSe2. Three
bands cross the Fermi level. This system exhibits soft phonons along
Computational considerations. Solving Eqs. (20)–(25) or the the ΓM direction109,111,112, which are stabilized by anharmonic
simplified version Eqs. (27) and (28) is computationally demanding effects109. Here, for simplicity, we avoid soft phonons by slightly
because a fine sampling of electron-phonon processes near the increasing the electronic smearing to 0.025 Ry in phonon calcula-
Fermi surface is required. In addition, due to the implicit form of tions, as shown in Fig. 3b. The same panel also shows the
these equations, the k- and q-point grids must be commensurate. To distribution of the electron–phonon coupling strength as described
accelerate calculations, EPW exploits crystal symmetries so that the by the Eliashberg spectral function α2F. Figure 3c shows the
gap function and related quantities are only computed for k-points anisotropic gap function color-coded on the Fermi surface of 2H-
in the irreducible wedge of the Brillouin zone, and only electronic NbSe2. We find a highly anisotropic two-gap structure: the high-gap
states within a small energy window centered at the Fermi energy region corresponds to the Fermi arcs around the K point, and the
are considered. Numerical tests show that converged results are low-gap region corresponds to the Γ-centered pockets. These results
obtained when this energy window is of the order of a few times the are similar to previous calculations for the related compound NbS298.
maximum phonon energy. In general, grids with an average point Figure 3d shows the energy distribution of the superconducting gap
density Δk ~ 0.02 Å−1 and Δq ~ 0.04 Å−1 are sufficient to converge function Δnk for various temperatures. The gap is seen to close
the superconducting gap85,91,104, but denser meshes may be gradually with increasing temperature, vanishing at the critical
required for superconductors with complex Fermi surfaces. As an temperature Tc = 19 K. The calculated critical temperature over-
example, 1202 and 403k-point grids and 602 and 203q-point grids estimates the experimental value of 7.2 K, consistent with prior
were used to describe the two-gap structure of bilayer C6CaC6 and calculations for bulk and monolayer transition metal dichalcogen-
the single-gap structure of bulk CaC6, respectively91 (see Table 1). ides98,109,112,118–120. The calculations in Fig. 3d are performed within
One further complication in the solution of the Eliashberg the FBW Eliashberg approach, with the chemical potential set to the
equations is that all quantities depend on Matsubara frequencies. Fermi energy. Calculated gap values for energy windows of 0.4 and
The Matsubara frequencies are proportional to the absolute 0.8 eV are very similar, therefore in the following we only consider
temperature, therefore superconductors with low Tc require a larger the former. Figure 3e shows the sensitivity of the superconducting
number of frequencies within the same energy range. While it is gap to the chemical potential: by requiring particle number
common to set the Matsubara frequency cutoff to ten times the conservation via Eq. (25), the chemical potential shifts by 25 meV
maximum phonon frequency85,86,101,105,106, we recommend always with respect to the Fermi energy, and the critical temperature
performing convergence tests with respect to this parameter. In EPW, slightly increases, by <1 K. In Fig. 3f we perform the same test, but
the number of Matsubara frequencies can be reduced using a sparse this time with a wider energy window, obtaining similar results.
sampling scheme whereby only a subset of frequencies is retained In Fig. 3g we also compare the FBW Eliashberg approach with the
with a density that decreases logarithmically with the Matsubara FSR approach, which is approximately 30% faster for the system
index. The grid density is controlled by an adjustable parameter; under consideration. The critical temperatures calculated from these
using the default setting for this parameter, EPW yields a ~30% approaches agree within 0.5 K. This finding relates to the DOS of 2H-
reduction of Matsubara frequencies, while keeping all the low- NbSe2 being slowly varying within a few hundred meV’s around the
est ~40% of frequencies. Numerical tests show that this approach Fermi energy. Larger differences are expected for systems with
preserves the accuracy of more expensive full-grid calculations. rapidly varying DOS, such as for example the high-Tc hydride
FBW Eliashberg calculations are computationally more demanding superconductors105.

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
9
a b 40
2
DFT FW = 0.4 eV
W90 FW = 0.8 eV 30
1
E-EF (eV)

ω (meV)
20
0

-1 10
phdos
α2 F
λ
-2 0
Γ M K ΓA L H A0 4 8 12 Γ M K ΓA L H A0 1 2 3
DOS (states/eV) phdos, α2F, λ
c 4.4 K d 5 fixed μF, ωc = 0.4 eV
e FW = 0.4 eV, ωc = 0.4 eV
FW = 0.4 eV fixed μF
3.9 M 4 FW = 0.8 eV variable μF

Δk (meV)
Δk (meV)

3.4 3
2.8 2
2.3
1
1.8
0
f 5 g h
FW = 0.4 eV, ωc = 1.2 eV FW = 0.4 eV, ωc = 0.4 eV, uniform FW = 0.4 eV, ωc = 0.4 eV, sparse
fixed μF FBW FBW
4 variable μF FSR FSR
Δk (meV)

0
0 4 8 12 16 0 4 8 12 16 0 4 8 12 16 20
Temp. (K) Temp. (K) Temp. (K)
Fig. 3 Eliashberg calculations for phonon-mediated superconductors in EPW. a Band structure and DOS of 2H-NbSe2. b Phonon dispersion
relations, phonon density of states (blue line), Eliashberg spectral function α2F (solid red line), and cumulative electron-phonon coupling
strength λ (dashed red line) of 2H-NbSe2. In (a), the bands obtained by Wannier interpolation are shown in dashed red lines. c Momentum-
resolved superconducting gap of 2H-NbSe2, color-coded on the Fermi surface220 and evaluated at 2 K [same dataset as in (d)]. d Energy
distribution of the superconducting gap of 2H-NbSe2 as a function of temperature, calculated using the FBW Eliashberg approach, chemical
potential set to the Fermi energy, and a Matsubara frequency cutoff of 0.4 eV. The energy window is 0.4 eV for the black line, and 0.8 eV for the
red line. e Gap function of 2H-NbSe2 obtained from the FBW approach using fixed chemical potential (black line) and variable chemical
potential (blue line). The energy window and Matsubara cutoff are both 0.4 eV. f Same as in (e), but using a Matsubara cutoff of 1.2 eV.
g Comparison between the gap functions obtained with the FBW approach (black line) and the FSR approach (blue). Both the energy window
and the Matsubara cutoff are 0.4 eV. h Same as in (g), but using a sparse Matsubara frequency grid. The dashed lines in d–h represent averages
of the gap distributions.

When using a logarithmic Matsubara frequency grid, the results sensitive to the choice of the exchange and correlation functional;
are essentially unaffected as shown in Fig. 3h. At the same time, the second, the size of the supercells required to obtain converged
sparse Matsubara grids afford a significant computational saving of energies and wavefunctions may be prohibitive.
almost a factor of two. In the EPW code, these challenges are overcome by recasting
the polaron problem into the solution of a coupled nonlinear
system of equations for the energy, wavefunction, and atomic
Polarons and electron self-trapping displacements associated with the polaron131–134. The founding
Background and formalism. An important manifestation of principle of this methodology is that the DFT formation energy of
electron-phonon interactions is the formation of polarons. The the polaron can be expressed as a self-interaction-free functional
polaron is a quasiparticle consisting of an electron or a hole of the polaron wavefunction ψ(r) and the atomic displacements in
dressed by a distortion of the crystal lattice. The lattice distortion the polaronic state Δτκαp132 (all symbols are defined in the
can be interpreted as a cloud of virtual phonons accompanying ‘Methods’ section):
the electron or hole. In the presence of strong electron–phonon Z XZ
^ 0 ψðrÞ þ ∂V 0KS
coupling, the electron or hole can become trapped in the lattice ΔE f ¼ dr ψ ðrÞH dr jψðrÞj2 Δτ καp
distortion that it produced; this configuration is referred to as a
KS
καp
∂τ καp
self-trapped polaron121–123. 1 X 0 (35)
In the context of first-principles calculations, the formation of þ C καp;κ0 α0 p0 Δτ καp Δτ κ0 α0 p0 ;
2
polarons has generally been investigated by means of DFT καp
calculations of an excess charge added to a large supercell. When
κ 0 α0 p 0
the calculation is initialized with a localized distortion of the
crystal lattice, it may be energetically favorable for the electron or where H ^ 0 and C 0 0 0 0 are the Kohn–Sham Hamiltonian and the
KS καp;κ α p
hole to localize around this distortion, leading to the formation of matrix of interatomic force constants in the ground-state structure
a polaron124–130. This approach faces two challenges: first, the without the polaron, and the integrals are over the Born–von
existence of polaronic states and their formation energy are very Kármán (BvK) supercell. Variational minimization of this energy

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
10
functional leads to the coupled system of equations: only half of the Bqν coefficients need to be evaluated at each
X ∂V 0 ðrÞ iteration. In addition, for parallel execution, we set a global gauge
^ 0 ψðrÞ þ
H KS
Δτ καp ψðrÞ ¼ ε ψðrÞ ; (36) for all interpolated Kohn–Sham wavefunctions and vibrational
KS
καp
∂τ καp eigenmodes. The diagonalization of the effective Hamiltonian is
performed via the Davidson method136 as implemented in
X Z ∂V 0KSðrÞ
1 Quantum ESPRESSO. The convergence of the iterative procedure
Δτ καp ¼  ðC 0 Þκαp;κ 0 α0 p0 dr jψðrÞj2 ; (37) is tested by evaluating the atomic displacements in real space via
κ 0 α0 p0
∂τ κ 0 α0 p0
Eq. (39) and comparing these real-valued vectors between
where V 0KS is the Kohn–Sham potential in the ground-state successive iterations. In the case of polaron calculations, choosing
structure without polaron, and ε represents the quasiparticle a fine Brillouin zone grid with N × N × N points implies that one is
excitation energy of the polaron133,134. To avoid performing studying polarons in a periodic BvK supercell consisting of
calculations in large supercells, in EPW Eqs. (36) and (37) are N × N × N unit cells. The formation energy of an isolated polaron
rewritten more conveniently in terms of Kohn–Sham states, is obtained by taking the limit of N → ∞. For large enough N, we
phonons, and electron–phonon matrix elements evaluated in the observe that the formation energy scales as 1/N, which is
primitive unit cell of the crystal. To this aim, the wavefunction is consistent with the Makov–Payne asymptotic expansion of the
expanded on the basis of Kohn–Sham states: energy of a point charge in a periodic cell132,137. Based on this
1 X observation, to obtain the energy of an isolated polaron, we
ψðrÞ ¼ pffiffiffiffiffiffiffi Ank ψnk ; (38) perform a Makov-Payne extrapolation to the limit of infinite
Nuc nk supercell size. Generally speaking, we find that for small polarons
and the atomic displacements are expanded on the basis of lattice the wavefunction shape is converged already for BvK supercells of
vibrational eigenmodes: the order of 5 × 5 × 5 unit cells, while for large polarons it is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi necessary to use supercells that are large enough to fully contain
2 X  _ the polaron density. While there is no formal upper limit to this
Δτ καp ¼  B eκα;ν ðqÞeiqRp : (39) size, the largest BvK supercell sizes that we employed are up to
Nuc qν qν 2Mκ ωqν
50 × 50 × 50 for the large electron polaron in NaCl (see Table 1).
With these definitions, Eqs. (36) and (37) become a nonlinear To visualize the polaron, EPW expresses the wavefunction as a
system of equations for the polaron coefficients Ank and Bqν: linear combination of maximally localized Wannier functions:
X
2 X ψðrÞ ¼ Amp wm ðr  Rp Þ;
Bqν gmnν ðk; qÞ Amkþq ¼ ðεnk  εÞAnk ; (40) (44)
Nuc qmν mp

where the coefficients Amp are obtained from the solution vectors
1 X  gmnν ðk; qÞ Ank via the Wannier–Fourier transformation:
Bqν ¼ A Ank : (41)
Nuc mnk mkþq _ωqν 1 X ikRp y
Amp ¼ e Umnk Ank ; (45)
These relations are referred to as the ab initio polaron equations. N uc nk
The polaron formation energy in Eq. (35) can be expressed in and Umnk is the unitary matrix that generates the smooth Bloch
terms of the polaron coefficients Ank and Bqν as132 gauge [Eq. (64) of the “Methods” section]. The inversion of this last
1 X 1 X relation also allows one to interpolate the coefficients Ank
ΔE f ¼ jAnk j2 ðεnk  εCBM Þ  jBqν j2 _ωqν ; (42)
Nuc nk Nuc qν throughout the Brillouin zone, which is useful to visualize how
specific bands contribute to the polaron wavefunction (Fig. 4).
where εCBM is the Kohn–Sham eigenvalue of the conduction band
minimum. This expression holds for electron polarons; in the case Application example. To demonstrate the implementation of the
of hole polarons, εCBM is replaced by the eigenvalue of the valence polaron module in EPW, we investigate the electron and hole
band maximum, and the first term on the right-hand side of Eq. polarons in rocksalt NaCl, a prototypical polar insulator that hosts
(42) acquires a minus sign. Equations (38), (39), and (42) provide small hole polarons called VK centers138,139. We perform calcula-
the real-space wavefunction of the polaron, the accompanying tions using the PBE exchange and correlation functional45, ONCV
atomic displacements, and its formation energy. pseudopotentials47, and a planewaves kinetic energy cutoff of
150 Ry. Ground-state and lattice-dynamical calculations are
Computational considerations. The solution of Eqs. (40) and (41) performed using a 12 × 12 × 12 uniform and unshifted Brillouin
require the knowledge of the Kohn-Sham energies εnk, the zone mesh for both k- and q-points. The Kohn–Sham states,
phonon frequencies ωqν, and the electron-phonon matrix phonons, and electron-phonon matrix elements that are needed
elements gmnν(k, q). All of these quantities are calculated by EPW to solve Eqs. (40) and (41) are generated on Brillouin zone grids
using Wannier–Fourier interpolation as described in the Methods. with up to 50 × 50 × 50 points by Wannier–Fourier interpolation.
The solution of the coupled system of equations is performed To this end, we use three Wannier functions to describe the Cl-3p
using an iterative procedure as follows. First, Ank is initialized in states in the valence band, and one Wannier function to describe
order to compute Bqν via Eq. (41). Then, Eq. (40) is solved by a single conduction band formed by the Na-3s states. The use of
constructing and diagonalizing the effective Hamiltonian additional valence or conduction bands only brings negligible
2 X changes to the polaron formation energy. In order to obtain the
Hnk;n0 k0 ¼ εnk δnk;n0 k0  B 0 g0 ðk; k0  kÞ :
ν k kν n nν
(43) polaron formation energies in the limit of infinite supercell size,
Nuc
we solve Eqs. (40) and (41) for N × N × N k- and q-point grids with
The set of solution coefficients Ank is used again in Eq. (41), and increasing N, and we perform a linear extrapolation of the energy
the process is repeated until convergence is achieved. vs. 1/N curve. Figure 4a shows an isosurface of the calculated
The initialization of the coefficients Ank is achieved by using a wavefunction of the electron polaron in NaCl. The formation
Gaussian function in reciprocal space centered around the band energy of this polaron in the infinite supercell limit is 37 meV; with
extremum. Alternatively, a polaron solution from a previous a radius of 147 Å, this polaron spans several tens of crystalline unit
calculation can be used. In the construction of the effective cells. The wavefunction is isotropic and can be described as a
Hamiltonian matrix in Eq. (43), we set the gauge condition combination of Na-3s orbitals modulated by an approximately
eκα;ν ðqÞ ¼ eκα;ν ðqÞ135. This condition ensures that Bqν ¼ Bqν , so Gaussian envelope.

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
11

Fig. 4 Calculation of small and large polarons with EPW. a Isosurface plot of the wavefunction of the electron polaron in NaCl. Na atoms are
in silver, Cl atoms are in green. b Displacements of the Cl atoms associated with the electron polaron shown in (a). The displacements are
exaggerated for clarity. c Isosurface plot of the hole polaron in NaCl, and associated atomic displacements. The displacements are
exaggerated for clarity. d Spectral weights of the electron (blue) and hole (orange) polaron of NaCl, superimposed to the band structure. The
size of the circles is proportional to ∣Ank∣2. e Spectral weights of the electron (blue) and hole (orange) polaron of NaCl, superimposed to the
phonon dispersion relations. The size of the circles is proportional to ∣Bqν∣2.

In Fig. 4b we show the atomic displacements associated with the optical transitions within second-order time-dependent perturba-
electron polaron of Fig. 4a. For clarity we only show the displacement tion theory. Phonon-assisted transitions involve two virtual
of the Cl ions. The ions tend to move away from the center of the processes, namely the absorption of a photon, and the absorption
electron wavefunction, consistent with the fact that the electron or emission of a phonon, in either order. In this section, we outline
polaron tends to repel anions. Figure 4c shows an isosurface of the the general formalism to describe these processes142,143.
calculated hole polaron of NaCl. The formation energy in the dilute We consider a linearly polarized electromagnetic wave with
limit is 820 meV, indicating a strongly bound polaron. Consistent with vector potential of amplitude A0, frequency ω, and polarization
the large formation energy, we find a very small polaron of radius vector e. In the following equations, the long-wavelength limit and
6 Å, which is comparable to the lattice parameter of 5.69 Å. This the corresponding electric dipole approximation are understood.
polaron consists primarily of a single 3p orbital centered on Cl and Second-order time-dependent perturbation theory states that the
oriented along the [100] direction. The accompanying atomic transition rate of an electron from an initial Kohn–Sham state nk to
displacements are shown in the same panel. As expected from the a final state mk + q involving a photon and a phonon qν is142:
negative charge of the polaron, the largest displacements are found P  2
for the Na cations which tend to move away from the polaron center. W mnν ðk; q; ωÞ ¼ 2π 2 2
_ e A0 e  ½S1;mnν ðk; qÞ þ S2;mnνβ ðk; qÞ
In Fig. 4d we overlay the coefficients ∣Ank∣2 with the band structure β¼ ± 1
in order to determine which electronic states contribute to the ´ δðεmkþq  εnk  _ω þ β_ωqν Þ:
electron (blue) and hole (orange) polaron. In the case of the electron (46)
polaron, only electrons near the bottom of the conduction band
contribute; the narrow distribution of Ank near k = 0 is consistent In this equation, β = +1 and −1 represent phonon emission and
with the large spatial extent of the electron polaron in real space. absorption processes, respectively. S1,mnν(k, q) and S2,mnνβ(k, q) are
Conversely, the hole polaron draws weight from the entire Brillouin the transition amplitudes for the processes illustrated in Fig. 5a:
zone, in line with the strong localization in real space. Similarly, in Fig. S1,mnν(k, q) refers to a process whereby the electron absorbs a
4e we show the coefficients ∣Bqν∣2 on the phonon dispersion photon, and then absorbs or emits a phonon; S2,mnνβ(k, q)
relations. The electron polaron is primarily driven by long-wavelength describes a process whereby phonon absorption/emission takes
longitudinal-optical phonons and longitudinal acoustic phonons, place, followed by the absorption of a photon. Explicit expressions
while the hole polaron is driven mostly by short-range optical modes. for these amplitudes are:
These spectral weight distributions suggest that the electron polaron X gmjν ðk; qÞvjn ðkÞ
in NaCl should be classified as a Fröhlich-type140 polaron, while the S1;mnν ðk; qÞ ¼ ; (47)
εjk  εnk  _ω þ iη
hole polaron should be classified as a Holstein-type polaron141. j

X vmj ðk þ qÞgjnν ðk; qÞ

Phonon-assisted indirect absorption S2;mnνβ ðk; qÞ ¼ ; (48)
j
εjkþq  εnk þ β_ωqν þ iη
Background and formalism. The EPW code has the capability to
compute, from first principles, optical absorption spectra in where vmn denotes velocity matrix elements between the
indirect band gap semiconductors by including phonon-assisted Kohn–Sham states and the sum extends to all possible occupied

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
12

Fig. 5 Indirect phonon-assisted optical absorption spectra with EPW. a Calculated band structure of silicon including GW quasiparticle
corrections (lines). The energy is referred to the top of the valence bands. The marks are measured energies of critical points: cross: ref. 158;
diamond: ref. 89; filled triangle: ref. 221; filled square: ref. 222; open triangle: ref. 157. b Calculated phonon dispersion relations of silicon (lines).
Neutron scattering data are shown as disks160 and filled triangles161. c Calculated imaginary part of the dielectric function of silicon (line)
including phonon-assisted processes. The squares represent experimental data from ref. 223. d Calculated absorption coefficient of silicon. The
circles are measurements from ref. 224.

and unoccupied states. The energy η is a small parameter to avoid The summation over virtual states in Eqs. (47) and (48) are restricted
singular denominators. From the transition rates Wmnν(k, q; ω), we to the manifold of valence and conduction bands included in the
obtain the imaginary part of the dielectric function by summing Wannierization procedure. The Dirac delta functions appearing in
over all possible transitions in the Brillouin zone and in the band Eq. (49) are replaced by Gaussian functions or Lorentzian functions
manifold142,143: with a finite broadening. The small parameter η in Eqs. (47) and (48)
Z Z
P dk dq are used to avoid singular denominators which arise if, for a given
Im ½ϵðωÞ ¼ 2 ϵπe0 Ω ω12
2
je  ½S1;mnν ðk; qÞ þ S2;mnνβ ðk; qÞj2 photon energy ℏω, direct transitions are resonant with indirect
mnν;β¼ ± 1 ΩBZ ΩBZ
´ Pmnνβ ðk; qÞδðεmkþq  εnk  _ω þ β_ωqν Þ;
transitions. These situations are encountered, for example, at the
onset of direct transitions, i.e. when the photon energy matches the
(49) direct gap. In these situations, the dielectric function and the
where the factor of two results from the electron spin in the case absorption spectrum become sensitive to the choice of η: too small
of spin-unpolarized systems; this factor is omitted in the case of an η leads to a divergence of the spectrum, and too large an η leads
calculations including spin–orbit coupling. The quantities Pmnνβ(k, to an excessive broadening. To probe the sensitivity of the spectra
q) contain the temperature-dependent Fermi–Dirac and to this parameter, the optics module of EPW calculates Im[ϵ(ω)] for a
Bose–Einstein distributions as follows: range of broadening parameters between 1 meV and 0.5 eV. Efforts
1þβ 1β are currently ongoing to eliminate these spurious singularities: we
Pmnνβ ðk; qÞ ¼ nqν þ f nk ð1  f mkþq Þ  nqν þ ð1  f nk Þf mkþq : believe that they arise from an intrinsic limitation of second-order
2 2
perturbation theory when direct and indirect transitions are in
(50) resonance; a more general theory that correctly describes these
In Eq. (49), the scalar quantity ϵ(ω) represents the projection of the resonances is under development145.
dielectric tensor along the polarization direction e of the electric In Eqs. (47) and (48), the velocity matrix elements are sensitive
field. For easier comparison with experimental measurements, to the details of the electronic structure. For accurate calcula-
after we obtain the imaginary part of the dielectric function from tions, it is preferable to include GW quasiparticle corrections to
EPW, we calculate the absorption coefficient using142: the Kohn–Sham eigenvalues89. These corrections require a
ω Im½ϵðωÞ renormalization of the velocity matrix elements to preserve
αabs ðωÞ ¼ ; (51) the f-sum rule of optical transitions146,147. In the EPW code, this
c nðωÞ
renormalization is carried out by evaluating the velocity via
where n(ω) is the real part of the refractive index and c is the finite differences. To this end, we consider quasiparticle energies
speed of light. n(ω) can be calculated from the standard relations and Kohn–Sham eigenvalues evaluated on k-point grids slightly
between the dielectric function and the refractive index144. In this offset along the Cartesian directions. The renormalized velocity
case, the real part of the dielectric function can be obtained by matrix elements are given by147
performing a Kramers–Kronig transformation of Im[ϵ(ω)] and
applying a rigid shift to match the value of ϵ(ω = 0) computed εQP
mkþδqα  εnkδqα
QP

from DFPT using Quantum ESPRESSO. Alternatively, n(ω) can be mn;α ðkÞ ¼
v QP mn;α ðkÞ;
v KS (52)
εKS
mkþδqα  εnkδqα
KS
taken from experiments143.
where δqα is a small wavevector along the Cartesian direction α, of
Computational considerations. In the optics module of EPW, the magnitude 10−3 × 2π/a (a is the lattice parameter).
imaginary part of the dielectric function is calculated using Eq. (49). To achieve convergence in calculations of indirect phonon-

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
13
assisted optical absorption spectra, Brillouin zone grids with equilibrium crystallographic sites. The displacements are chosen in
40 × 40 × 40 points have been employed in previous work on such a way that the corresponding atomic configuration
silicon143, as well as in more recent work on silicon and represents the best single-point approximant to the quantum
diamond145 (see Table 1). In addition, convergence of the thermal distribution of the atomic coordinates165.
phonon-assisted optical spectra for cubic boron arsenide148 and Under the approximations of adiabatic Born–Oppenheimer
cubic silicon carbide149 was achieved with Brillouin zone grids of decoupling and harmonic lattice, the quantum thermal average of
32 × 32 × 32. When free-carrier absorption is considered, denser an electronic or optical property described by the observable O
grids are required for converged spectra. For example, a grid of can be written as165
48 × 48 × 48 is needed for converged phonon-assisted free-carrier Y Z dx qν dy qν 2
ejzqν j =uqν Oðfτ καp gÞ;
2
absorption spectra in silicon150. OðTÞ ¼ (53)
qν
πu 2
qν

Application example. To demonstrate the implementation of the where q runs over the set of wavevectors in a uniform Brillouin zone
optics module in EPW, we examine phonon-assisted optical grid which excludes time-reversal invariant points and time-reversal
absorption in silicon. Gound-state DFT calculations are performed partners. In Eq. (53), zqν denote normal mode coordinates with real
using the PBE exchange and correlation functional45, ONCV part xqν and imaginary part yqν. u2qν ¼ ð_=2M0 ωqν Þð2nqν þ 1Þ is the
pseudopotentials46,61, a planewaves kinetic energy cutoff of mean-square displacement for the oscillator qν with Bose–Einstein
60 Ry, and a 6 × 6 × 6 Brillouin zone sampling of both the k-point occupation nqν; O({τ}) is the property of interest, such as for example
grid and the q-point grid. The optimized DFT lattice parameter is the Kohn–Sham eigenvalues, density of states, or optical absorption
a = 5.478 Å, similar to previous studies151–153. spectrum, calculated for the set of atomic coordinates {τκαp}. The
We evaluate quasiparticle corrections on a 6 × 6 × 6 k-point grid relation between these coordinates and the normal coordinates zqν,
within the GW method using the BerkeleyGW code154. The GW- which is required to carry out the integral in Eq. (53), is provided in
corrected band gap is 1.31 eV, which slightly overestimates the ref. 165.
experimental value of 1.12 eV at room temperature155, but agrees In the special displacement method, the configurational average
well with prior work using similar settings89. The resulting expressed by Eq. (53) is approximated by a single calculation for an
quasiparticle band structure is shown in Fig. 5a. Our calculations optimum configuration:
agree well with experimental measurements at several critical OðTÞ ’ Oðfτ 0καp þ Δτ ZG
καp gÞ; (54)
points89,156–158, shown by marks in the figure. In Fig. 5a we also
show representative optical transitions leading to the amplitudes in where τ 0καp
represent atomic coordinates in the DFT ground-state
Eqs. (47) and (48). The fundamental gap is indirect, between the top at zero temperature, and the optimum ‘ZG’ displacement is given
of the valence band at the Γ point and the bottom of the by165
conduction band near the X point. To correctly describe optical sffiffiffiffiffiffiffiffiffiffiffi
transitions near the fundamental gap, it is essential to include M0 X  
Δτ καp ¼
ZG
2 Sqν uqν Re eiqRp eκα;ν ðqÞ : (55)
phonon-assisted processes. Figure 5b shows calculated phonon Np Mκ qν
dispersion relations, which agree well with prior calculations by us18
and other groups159, as well as experimental data160,161 (solid As in Eq. (53), the summation is restricted to q-points from a
symbols). Figure 5c shows the imaginary part of the dielectric uniform Brillouin zone grid which are not time-reversal invariant
function, as obtained from Eq. (49) using 32 × 32 × 32 fine k- and q- and are not time-reversal partners. This partitioning is described in
point grids. The temperature is set to 300 K, and the Dirac delta Appendix B of ref. 1. The quantities Sqν appearing in Eq. (55) are
functions in Eq. (49) are approximated via Gaussians of width signs (±1) determined by the ZG module so as to guarantee that
50 meV. In this plot, a rigid shift of −0.19 eV is applied in order to the resulting displacements make Oðfτ 0καp þ Δτ ZG καp gÞ the best
match the GW band gap to the measured gap of silicon at room possible approximant to Eq. (53). In the thermodynamic limit of a
temperature (1.12 eV)155. Figure 5d shows the related optical large supercell, the ZG displacements reproduce the exact mean-
absorption coefficient from Eq. (51). This calculation requires the square anisotropic displacement tensors, which are given by167
real part of the frequency-dependent refractive index, for which we Z
used experimentally measured values from ref. 162 for simplicity. _ X dq
Uκ;αα0 ðTÞ ¼ eκα;ν ðqÞeκα0 ;ν ðqÞ ð2nqν þ 1Þ: (56)
Our calculated spectra agree very well with experiments in the 2Mκ ωqν ν ΩBZ
photon energy range between the indirect gap and the direct gap at In the same limit, a single evaluation of the property
3.3 eV. Beyond the lineshape, the magnitude also agrees well with Oðfτ 0καp þ Δτ ZG
the experiment over several orders of magnitude, with the theory καp gÞ tends to the exact thermal average in Eq. (53).
In the case of non-periodic systems, such as for example
underestimating the experimental data by 50% at most. This residual nanocrystals and quantum dots, these equations are replaced by
underestimation might be related to the rigid shift of the band gap, their Γ-point only versions164,168,169.
which we did not include in the velocity renormalization expressed The special displacement method is similar in spirit to
by Eq. (52), and to the fact that electron–phonon matrix elements are computing thermodynamic averages using path-integral molecu-
slightly too weak in DFT as a result of the band gap problem1,163. lar dynamics170, but it differs insofar as a single calculation is
required to evaluate the average instead of many molecular
Special displacement method dynamics snapshots. In this method, electron–phonon couplings
Background and formalism. The EPW package contains a set of are included non-perturbatively through the changes of the
standalone modules, the ZG toolkit, for calculating finite- Kohn–Sham energies and wavefunctions caused by the ZG
temperature properties including quantum zero-point effects via displacements. The method can be applied to compute any
the special displacement method (SDM)164–166. The special property that can be expressed by means of a Fermi Golden Rule,
displacement method is a supercell-based approach that is such as for example temperature-dependent band structures,
distinct from and complementary to the Wannier–Fourier inter- density of states, and optical spectra. The main limitation of this
polation method employed in all previous sections. The founding approach as compared to the Wannier–Fourier interpolation
principle of this method is that the effects of electron–phonon method employed in the preceding sections is that, being an
couplings on the electronic and optical properties of extended adiabatic theory, fine spectral features on the scale of the phonon
solids can be captured by performing calculations for a large energy are averaged out. Conversely, its main advantage is that it
supercell where the atoms have been displaced away from their is easy to use as it requires a single DFT calculation.

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
14
The Kohn–Sham energy eigenvalues generated by the special Application example. To demonstrate the implementation of the
displacement method capture temperature renormalization and ZG module, we investigate the temperature-dependent band
quantum zero-point effects at the same level as the adiabatic structure renormalization and phonon-assisted optical absorption
Allen–Heine theory of temperature-dependent band structures171. spectra of silicon and BaSnO3. We perform calculations using the
The calculation of the imaginary part of the dielectric function local density approximation (LDA) for the exchange and correla-
including temperature effects, zero-point corrections, and tion174,175 and ONCV pseudopotentials46,61. We use a planewaves
phonon-assisted indirect processes, is performed by evaluating: kinetic energy cutoff of 40 Ry for silicon and 120 Ry for BaSnO3,
Z and 6 × 6 × 6 uniform k-point grids for ground-state calculations.
2πe2 1 X dK
Im½ϵðωÞ ¼ ^ jψvK ij2 δðεcK  εvK  _ωÞ;
jhψ je  p With these settings, we obtain indirect and direct band gaps of
ϵ0 m2e Ωsc ω2 cv ΩBZ;sc cK
0.49 and 2.56 eV for silicon, respectively, and indirect and direct
(57) band gaps of 1.06 and 1.55 eV for BaSnO3, respectively. We
where Ωsc and ΩBZ,sc are the volumes of the supercell and the perform calculations of dynamical matrices on 4 × 4 × 4 q-point
supercell Brillouin zone, respectively, the summations over v, c grids in both cases and use the ZG module to generate
refer to valence and conduction states, and K is a wavevector of displacements in 3 × 3 × 3 supercells. The signs Sqν appearing in
the supercell Brillouin zone. This expression contains both direct Eq. (55) are determined by minimizing the error descriptor in Eq.
and indirect optical transitions: the indirect transitions are hidden (54) of ref. 165 with a dimensionless threshold η = 0.1, after
in the dependence of the wavefunctions ψvK and ψcK on the ZG enforcing a smooth Berry connection between vibrational
displacement. Similarly, this expression includes temperature eigenmodes across the Brillouin zone165. Kohn–Sham energies in
renormalization via the dependence of the energies εvK and εcK the structures with ZG displacements are calculated using a
on the ZG displacements164. 12 × 12 × 12 uniform K-point grid in the supercell Brillouin zone,
and 108 unoccupied states. For calculations of dielectric functions,
Computational considerations. Calculations using the special we use up to 200 randomly generated K-points in the supercell
displacement method proceed as follows. First, one performs Brillouin zone, as well as 27 and 135 conduction bands for silicon
phonon calculations for the crystalline unit cell (not the supercell) and BaSnO3, respectively. All Dirac deltas are replaced by
using standard DFPT on a coarse uniform Brillouin zone grid. From Gaussians of width 30 meV.
this calculation, the phonon frequencies and eigenmodes are Figure 6a shows the joint density of states (JDOS) of silicon at
computed on a finer Brillouin-zone grid with N1 × N2 × N3q-points 0 K (red) and 300 K (blue), as calculated using the special
using standard interpolation of the force constant matrix8. Then displacement method. For comparison, the JDOS computed for
the ZG displacement within a supercell consisting of N1 × N2 × N3 the DFT ground state structure is also shown in green. We see
unit cells is evaluated at the temperature T via Eq. (55). The desired that, upon including electron-phonon interactions via the special
property is finally computed with this supercell, with the atoms displacements, even at 0 K the JDOS is red-shifted with respect to
displaced according to Δτ ZG the DFT ground state. This is a manifestation of the zero-point
καp . The special displacements are
generated by the ZG module; the procedure is computationally band gap renormalization176. Upon increasing temperature, the
inexpensive and is performed serially. onset of the JDOS further red-shifts. This temperature-induced
The ZG toolkit provides several codes to analyze the results of band gap narrowing is referred to as the Varshni effect177. The
supercell calculations performed with special displacements, for horizontal offset between the JDOS for the DFT ground state
example, Brillouin-zone unfolding of temperature-dependent structure and the JDOS computed with special displacements at
band structures from the supercell to the unit cell, the density 0 K yields a zero-point renormalization of 50 meV, in good
of states, electronic spectra, and vibrational spectra. agreement with prior work178,179. More accurate values can be
Band unfolding is performed using the procedure outlined in calculated by increasing the supercell size164,165. The band gap
ref. 172, which consists of determining the spectral function in the renormalization can also be computed without using the JDOS; to
primitive Brillouin zone by projecting the wavefunctions of the this end, one needs to evaluate the Kohn–Sham eigenvalues at
supercell into the wavefunctions of the unit cell. In practice, the the supercell K-points that unfold onto the wavevectors of the
spectral density for the wavevector k at the energy ε is obtained band extrema in the Brillouin zone of the unit cell.
as Figure 6b and c report color maps of the electronic spectral
X functions of silicon and BaSnO3 at 0 K calculated using the special
Ak ðεÞ ¼ PmK;k δðε  εmK Þ; (58) displacement method. For comparison, the band structures in the
mK DFT ground state are overlaid with these color maps. These maps
where the spectral weights PmK,k are given by can directly be compared to angle-resolved photoelectron
X spectroscopy data. From these images, we identify numerically
PmK;k ¼ jcmK ðGsc þ k  KÞj2 : (59) the quasiparticle band structures by extracting the spectral peaks.
Gsc
In the case of silicon, the valence band maximum blue-shifts by
Here, cmK denotes planewaves coefficients of supercell wavefunc- 32 meV with respect to ground-state DFT, and the conduction
tions, and Gsc denotes reciprocal lattice vectors of the supercell. band red-shifts by 18 meV. The resulting gap renormalization of
Equation (59) refers to norm-conserving pseudopotential imple- 50 meV is in agreement with the value determined in Fig. 6a via
mentations. The ZG module includes additional terms that are the JDOS. In the case of BaSnO3, Fig. 6 shows a zero-point
required in this expression when using ultrasoft and PAW renormalization of the band gap of 10 meV; however, we
(projector augmented-wave) pseudopotentials165,169. A similar emphasize that this value is not fully converged, and larger
unfolding strategy is employed to analyze lattice dynamics, as supercells, as well as corrections for Fröhlich couplings180, are
well as X-ray and neutron diffuse scattering intensities accounting necessary to obtain accurate data.
for multiphonon interactions173. Figure 6d reports a convergence test of the imaginary part of
The supercell size required to achieve converged results in the dielectric function of silicon, as computed with the special
calculations employing the special displacement method depends displacement method. In this calculation, we keep the supercell
on the target property. For example, to compute optical spectra fixed, and we increase the number of random K-points in the
over a broad photon energy range, supercells of size 4 × 4 × 4 unit supercell Brillouin zone. It is seen that the dielectric function
cells are typically sufficient. However, to obtain accurate phonon- converges relatively rapidly with the number of points and full
induced band gap renormalization, say within a 10 meV accuracy, convergence is achieved with 200 points when the Dirac delta
supercells as large as 10 × 10 × 10 may be required165 (see Table 1). functions in Eq. (57) are replaced by Gaussians with 30 meV width.

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
15

Fig. 6 Applications of the special displacement method implemented in the ZG module. a Square root of the JDOS of silicon evaluated
with atoms in the DFT ground-state structure (red); using special displacements at 0 K (blue); and using special displacements at 300 K (green).
The red-shift of the JDOS with respect to the ground-state structure signals the quantum zero-point correction to the band gap (0 K curve)
and its temperature renormalization (300 K curve), respectively. b Electron spectral function of silicon calculated with the special displacement
method at 0 K, along the ΓX path in the Brillouin zone. The black disks indicate the band structures calculated in the crystalline unit cell with
the DFT ground-state geometry. c Electron spectral function of BaSnO3 was calculated with the special displacement method at 0 K, along the
ΓR path in the Brillouin zone. The black disks indicate the band structures calculated in the crystalline unit cell with the DFT ground-state
geometry. d Convergence of the imaginary part of the dielectric function of silicon at 0 K with respect to the number of random K-points used
to sample the Brillouin zone of the supercell. e Imaginary part of the dielectric function of silicon calculated with the atoms in the DFT ground-
state structure (red), and by using special displacements at 0 K (blue). A scissor shift of 0.64 eV is employed to match the experimental gap225.
f Imaginary part of the dielectric function of BaSnO3 calculated with the atoms in the DFT ground-state structure (red), and by using special
displacements at 0 K (blue) and 300 K (green). The vertical dashed lines indicate the direct and indirect band gap energies. A scissor shift of
1.91 eV is employed to match the experimental gap226.

Clearly, a larger smearing would require fewer K-points. OpenMP parallelization. We present benchmarking results which
In Fig. 6e we compare the imaginary part of the dielectric demonstrate that, with this improved parallelization strategy,
function of silicon computed in the DFT ground state structure EPW can run at nearly full scale on the pre-exascale HPC system
(red) with the special displacement method (blue). We see that the Frontera at the Texas Advanced Computing Center (TACC).
special displacements correctly capture phonon-assisted indirect Finally, we describe the I/O strategy employed in this explorative
optical transitions in the energy range between the indirect gap implementation, which is based on XML (eXtensible Markup
and direct gap of silicon. This approach provides an alternative Language) and HDF5 file formats. We emphasize that EPW
strategy for computing optical spectra to Eq. (49) implemented in continues to evolve with changes and improvements, and the
the indirect optics module of EPW. In the special displacement reader is referred to the EPW website for its most up-to-date
method, both temperature-dependent band structure renormali- features and functionalities182.
zation and phonon-assisted processes are included on the same
footing. Figure 6f shows the imaginary part of the dielectric Computational workflow and basic capabilities. The EPW code is
function of BaSnO3 at 0 K (blue) and 300 K (green), including logically partitioned into two components. The first component
phonon-assisted processes. For comparison, we also show the serves as an interface to the PWscf and PHonon codes of
calculation using the DFT ground-state structure (red), which Quantum ESPRESSO. The code imports ground-state charge
misses phonon-assisted transitions. The spectra compare well with density and wavefunctions with computational parameters from
prior work using the special displacement method181. prior runs with pw.x, and dynamical matrices, variations of the
Kohn–Sham potentials, and vibrational mode pattern files from a
prior phonon calculation with ph.x. All these quantities are
Implementation and HPC benchmarks evaluated on coarse k- and q-points grids, and the phonon
In this section, we describe some of the implementation and wavevectors are restricted to the irreducible wedge of the
parallel programming models of EPW. We first outline the Brillouin zone. Then the code calls Wannier90 in library mode
computational workflow and basic capabilities of EPW. Then we to calculate the unitary rotation matrices Umnk [Eq. (64) of the
describe our recent efforts to prepare EPW for exascale HPC “Methods” section] needed to transform Bloch states into
systems with a focus on an explorative implementation of a maximally localized Wannier functions. Subsequently, the code
highly scalable parallelization scheme, hybrid two-level MPI, and unfolds the irreducible wedge of the phonon Brillouin zone into

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
16
the full zone, evaluates electron–phonon matrix elements on grids (Fig. 7). These calculations are performed on the Frontera
coarse k- and q-point grids, and it transforms the Hamiltonian, supercomputer at TACC. Frontera consists of 8368 nodes, each
dynamical matrices, and electron–phonon matrix elements into equipped with two Intel Xeon Platinum 8280 ("Cascade Lake”)
the Wannier representation. At this point, the code can print processors with 28 cores per processor. To minimize statistical
relevant quantities for subsequent runs with epw.x, and it can fluctuations in these benchmarks, we execute each calculation four
continue to the interpolation part. The interpolation engine of times, and we average the resulting Wall times. Figure 7 shows that
EPW carries out Wannier interpolation of the Hamiltonian, using this parallelization scheme, we achieve ~92% of the ideal
dynamical matrices, and electron–phonon matrix elements on speedup up to 448,000 cores in the evaluation of the
fine k- and q-point grids after reading these quantities in the electron–phonon matrix elements.
Wannier representation, and then it calculates electron–phonon- The hybrid two-level MPI and OpenMP parallelization scheme is
related properties on these grids. controlled by three parameters that can be optimized to achieve
The current snapshot of the code supports spin-unpolarized maximum parallel scaling efficiency: the number of upper-level MPI
and non-magnetic spin–orbit calculations; work is currently in tasks, the number of lower-level MPI tasks, and the number of
progress to extend the code to the cases of spin-polarized and OpenMP threads per MPI task. In the benchmarks shown in Fig. 7,
non-collinear magnetic cases. Regarding pseudopotentials (PPs), we fix the number of k-point pools to 250, by varying the number of
norm-conserving (NC) PPs are supported. Since both ultrasoft PPs q-point pools depending on the number of cores. With this choice,
and the PAW method violate fundamental symmetry relations that we are able to achieve near-ideal speedup on extreme scaling tests;
the electron–phonon matrix elements must fulfill183, extending we expect that a speedup even superior to 92% could be achieved
the current NCPP implementation to these cases poses some upon further optimization of the parallelization parameters We also
challenges. We will proceed to these extensions in future releases. point out that, in large-scale runs, I/O time dominates the total Wall
time. To overcome this I/O bottleneck, we employ a low-I/O mode
Parallelization. At the time of writing of this manuscript, we whereby memory is exploited as much as possible and minimum
witness the launch of the first supercomputer entering the usage of storage media is made.
exascale computing era, namely Frontier at Oak Ridge National Our strategy for OpenMP parallelization is to use threaded
Laboratory; additional exascale systems are currently being versions of scientific libraries such as Intel MKL and Cray LibSci,
deployed worldwide. Exascale HPC systems are characterized by rather than explicitly using OpenMP directives throughout the code.
many-core and heterogeneous architectures. Since 2020, we have For the benchmarks reported in Fig. 7, we find that exceeding 7
made efforts to ready EPW for the exascale transition. In the OpenMP threads per MPI task worsens the performance. Therefore
following, we describe one strategy that we are currently we use 4 OpenMP threads per MPI task, and we make all threads
exploring, a hybrid two-level MPI/OpenMP parallelization scheme, belonging to the same MPI task reside in the same non-uniform
which makes EPW work seamlessly and efficiently on many-core memory access (NUMA) domain.
architectures.
Historically EPW has adopted a one-level parallelization over the Structured I/O. Recently we have been investigating the use of
electron wavevectors (k-point parallelization) via MPI. This strategy XML and the HDF5 data formats. Both formats can describe
has three shortcomings, which lead to an early saturation of the hierarchical data and are self-describing, flexible, and portable. For
speedup with the number of cores in large-scale calculations: (i) small to medium-sized data, the XML format is used; for large binary
There is an upper bound for the number of total MPI tasks which data, the HDF5 file format is used in parallel. In previous versions of
guarantees good scalability; this bound is determined by the EPW, several large arrays were read or written via per-process I/O or
number of k-points. (ii) As the number of MPI tasks increases, the parallel MPI I/O. The former quickly overwhelms the file systems in
overhead cost associated with MPI communications, in particular large-scale calculations by generating a large number of files; the
collective communications, increases. (iii) Due to the internal buffers latter is not easily portable. Both issues can be addressed for large-
of MPI, the memory overhead also increases with the number of MPI size binary data by employing parallel I/O via parallel HDF5.
tasks. To address these issues, we are exploring the extension of the
one-level parallelization over k-points to the hybrid two-level MPI
and OpenMP parallelization over both k- and q-points.
In the hierarchical two-level MPI and OpenMP parallelization
strategy, the total MPI tasks are partitioned into two levels of MPI
groups, each consisting of a set of MPI tasks; in the lowest level,
OpenMP parallelization is employed within each MPI task. For
example, the total MPI tasks are first divided into k- or q-point pools,
and each pool is further divided into q- and k-point pools,
respectively. Each q- or k-point pool contains a set of MPI tasks,
and OpenMP parallelization is employed within each MPI task. The
parallelization option employed in the upper-level pool (over k-
points or q-points) is chosen so as to maximize the calculation
efficiency.
As compared to the one-level parallelization, the hierarchical two-
level MPI and OpenMP parallelization can reduce the number of MPI
tasks in each MPI communicator group, thereby reducing the
overhead cost for collective MPI communication as well as the
memory footprint due to the internal buffers of the MPI library.
OpenMP parallelization can further reduce the memory footprint
associated with replicated memory allocations in the code. This
strategy enables increased flexibility in distributing the computa-
tional workload over a large number of cores. Fig. 7 Extreme scaling benchmarks of EPW. Strong-scaling test for
To demonstrate the scaling behavior of EPW, we perform strong- the evaluation of the electron–phonon matrix elements of MgB2.
scaling tests for the two-gap superconductor MgB2. We evaluate These tests were performed using EPW v6-alpha available at
electron-phonon matrix elements on 150 × 150 × 150 k- and q-point ref. 227.

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
17
We note that parallel I/O performance heavily depends on the potential for high modularity, extensibility, and reusability, which
details of file systems and the parallelization settings. For instance, in will facilitate flexible and sustained software development193. We
the case of the LUSTRE file system, the number and size of stripes expect that targeted use of OOP will greatly increase the efficiency
play an essential role in achieving high performance. In the case of and productivity in the development and maintenance of code194.
the ROMIO implementation of MPI-IO, identifying optimum para- Alongside these developments, we anticipate increased atten-
meters in large-scale runs is nontrivial and requires systematic tion to the issue of automation and the capability of performing
experimentation on HPC systems. electron–phonon calculations at scale for high-throughput
approaches, data science, and artificial intelligence/machine
learning (AI/ML) applications.
DISCUSSION In summary, in this manuscript, we provided a comprehensive
Since the last technical paper in 201618, the EPW code has update on the current status, functionalities, and performance of
considerably expanded in scope, to the point of becoming a the EPW code, we offered an overview of possible future directions
robust and efficient software platform for developing, testing, and for this and similar codes. We hope that, beyond capturing a
deploying methods addressing electron–phonon physics and snapshot of the current status of the EPW project, the present
related materials properties. It is then natural to ask which manuscript will contribute to making advanced electron–phonon
components and functionalities are still in need of improvement, calculations more accessible and more widely used in computa-
what are emerging directions in this area, and more generally tional materials discovery and design.
what comes next for the EPW project.
On the functionality front, we expect to continue expanding the
capabilities for transport, superconductivity, polarons, and optics METHODS
described in the ‘Results’ section. For example, the transport The key element of any calculation of electron-phonon interac-
module is based on the steady-state linearized Boltzmann tions and related materials properties is the electron–phonon
transport equation; here, it would be desirable to generalize the matrix element. The EPW code employs physics-based
methodology to high-field transport and to time-dependent Wannier–Fourier interpolation to compute electron-phonon
driving fields. Similarly, in the superconductivity module, the matrix elements accurately and efficiently and uses these matrix
treatment of electron–electron repulsion could be improved by elements as the starting point for advanced many-body calcula-
seamlessly integrating EPW with standard GW codes. The study of tions of electronic, optical, and transport properties of solids
polaron physics from an ab initio many-body perspective is just including temperature dependence and phonon-assisted quan-
beginning, therefore we can expect a number of developments in tum processes.
this area, from the study of polaron transport to their optical In the following we introduce the notations and conventions
properties and their relation with the theory of band structure employed throughout the manuscript, we provide a brief over-
renormalization. Calculations of phonon-assisted optical transi- view of the general principles of Wannier–Fourier interpolation of
tions are currently restricted to optical absorption spectra, but the electron–phonon matrix elements, and we outline the extension
formalism can just as well be employed for investigating related of this interpolation scheme to the case of polar materials. A more
phenomena such as for example Auger–Meitner recombina- comprehensive description of this methodology including
tion184. Beyond these core modules of the EPW project, we detailed derivations can be found in ref. 10.
anticipate growth in the calculation of electronic and optical
properties using the special displacement method within the
Notation and definitions
EPW/ZG code; for example, the current implementation focuses
on harmonic systems, but generalizations to strongly anharmonic In the study of the electronic structure and lattice dynamics of
systems have recently been proposed185. crystalline solids, infinitely extended crystals are described by
On the accuracy front, we anticipate that future work will focus considering a periodic BvK supercell. This supercell consists of
on improving the precision of the Wannier interpolation method Nuc = N1 × N2 × N3 primitive unit cells of the crystal and is subject
that is the basis for EPW. For example, methods that take to periodic boundary conditions. The pth unit cell within the BvK
dimensionality into account in the interpolation of the supercell is identified by the direct lattice vector Rp, with
electron–phonon matrix elements will be essential to perform p = 1, …, Nuc. The dual lattice of the direct lattice vectors Rp in
predictive and reliable calculations of transport, optics, super- reciprocal space defines a uniform grid of Nuc = N1 × N2 × N3
conductivity, and polarons in 2D materials186,187. Furthermore, it wavevectors in the Brillouin zone. Throughout the manuscript, we
would be highly desirable to improve the predictive power of the use k and q to denote electron and phonon wavevectors in the
DFPT electron–phonon matrix elements before even proceeding crystal Brillouin zone, respectively, and K to denote electron
to Wannier–Fourier interpolation; a recent proposal to employ the wavevectors in the Brillouin zone of the BvK supercell.
GW method to calculate many-body corrections to the DFPT We use ψnk to indicate the eigenfunctions of the DFT Kohn-
matrix elements offers a promising path toward this goal188. Sham Hamiltonian H ^ KS , and εnk the corresponding eigenvalues.
On the front of HPC, we envision continuing the current effort This wavefunction is periodic and normalized over the BvK
to enable EPW for exascale computing architectures. In addition to supercell, and can be expressed in the Bloch form:
the hybrid MPI/OpenMP parallelization, it will be necessary to 1
leverage diverse GPU architectures and to enable the code for ψnk ðrÞ ¼ pffiffiffiffiffiffiffi unk ðrÞeikr ; (60)
Nuc
large-scale runs that will become possible with the exascale
supercomputers that are being deployed worldwide. with unk being a lattice-periodic function normalized in the
Another important direction will be to enhance the interoper- primitive unit cell of volume Ω. The position vector of the atomic
ability of EPW with other major electronic structure software nucleus κ belonging to the unit cell identified by Rp is
packages such as Abinit189, VASP190, Siesta191, and GPAW192. τκp ¼ Rp þ τ κ , with τκ being the position vector within the
As several codes currently use the information generated by EPW primitive unit cell. The components of this vector along the
for a variety of post-processing tasks, it will make sense to develop Cartesian directions α = 1, 2, 3 are denoted by τκαp.
standardized data structures that adhere to the FAIR data Vibrational eigenmodes and eigenfrequencies are obtained by
principles. diagonalizing the dynamical matrix, which is computed via DFPT
Regarding the programming model, transitioning toward using Quantum ESPRESSO12. In turn, the dynamical matrix is the
object-oriented programming (OOP) will be useful due to the Fourier transform of the interatomic force constants C καp;κ0 α0 p0 , that

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
18
is the Hessian of the DFT total energy in the atomic coordinates. (except for low-lying semi-core states) as well as a few empty
The relation between interatomic force constants and dynamical conduction bands for metallic or semi-metallic systems. In systems
matrix is135 with medium-sized to large unit cells, the number of bands thus
1 X identified may be relatively large, and this will make EPW
Dκα;κ0 α0 ðqÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi eiqRp C κα0;κ0 α0 p ; (61) calculations more demanding because the number of operations
Mκ Mκ 0 p required to obtain electron-phonon matrix elements scales with
where Mκ is the mass of the κth nucleus. The eigenvalues and the square of the number of Wannier functions. Therefore, more
eigenvectors of the dynamical matrix are denoted by ω2qν and work will be needed in the future to achieve automation whilst
eκα,ν(q), respectively, and the index ν runs from 1 to 3M where M is preserving calculation efficiency.
the number of atoms in the primitive cell. ωqν corresponds to the
vibrational frequency, and eκα,ν(q) is the normal mode of vibration General principles of Wannier interpolation of
or polarization vector. electron–phonon matrix elements
The electron–phonon matrix element is defined by1,10 The localized nature of MLWFs provides the basis for accurate and
gmnν ðk; qÞ ¼ humkþq jΔqν v KS junk i; (62) efficient interpolation of Kohn–Sham wavefunctions and ener-
gies201. This is achieved by expressing the Kohn–Sham Hamilto-
where the integral is evaluated over the unit cell, and the lattice- nian in the Wannier representation, and noting that the
periodic component of the variation of the Kohn–Sham potential Hamiltonian matrix elements decrease rapidly with the distance
is given by between the Wannier function centers. The Hamiltonian matrix
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rffiffiffiffiffiffi elements in the Wannier representation are195
_ X M0 ∂ V KS ðrÞ
Δqν v ¼
KS
eiqðrRp Þ eκα;ν ðqÞ : (63) Z
2M0 ωqν καp Mκ ∂τ καp ^ np i ¼
Hmn ðRp Þ ¼ hwm0 jHjw ^ n0 ðr  Rp Þ :
dr wm0 ðrÞHw (66)
sc
In this expression, M0 is an arbitrary reference mass introduced for
convenience. MLWFs are defined in terms of Bloch states as9 Using Eq. (64), these matrix elements can be expressed in terms of
the Kohn–Sham Hamiltonian in the Bloch representation Hm0 n0 ðkÞ:
1 X ikðrRp Þ
wmp ðrÞ ¼ wm0 ðr  Rp Þ ¼ e Unmk unk ðrÞ; (64) 1 X ikRp y
N uc nk Hmn ðRp Þ ¼ e Umm0 k Hm0 n0 ðkÞUn0 nk ; (67)
N uc m0 n0 k
where Unmk is a unitary matrix in the indices m and n.
Throughout the manuscript, we used interchangeably the where the matrix Umnk is the same as in Eq. (64). Once Hmn(Rp) has
notation been determined, Eq. (67) can be inverted to generate Hm0 n0 ðkÞ
Z
1 X dk anywhere in the Brillouin zone.
or (65) Similarly to the electronic case, the dynamical matrix can be
Nuc k ΩBZ expressed in the phonon Wannier representation as10
to indicate summations or integrals over the Brillouin zone. The 1 X iqRp y
quantity ΩBZ on the right indicates the Brillouin zone volume. Dκα;κ0 α0 ðRp Þ ¼ e eκα;μ ðqÞDμν ðqÞeκ0 α0 ;ν ðqÞ ; (68)
N p qμν
The unitary matrix Unmk appearing in Eq. (64) is determined by
requiring that the functions wmp(r) be as localized as possible195, where Dμν(q) is the matrix element of the dynamical matrix in the
i.e. MLWFs9. The computation of these unitary matrices is Bloch representation for phonons. Once Dκα;κ 0 α0 ðRp Þ has been
performed by minimizing the spatial spread of the resulting determined, this relation can be inverted to obtain the phonon
Wannier functions9. This procedure is notoriously challenging eigenvectors and eigenvalues anywhere in the Brillouin zone; this
since the spread functional exhibits multiple local minima, is a standard procedure employed to compute phonon dispersion
especially in the case of entangled electron bands. In practice, relations7,202.
the construction of optimal MLWFs often requires fine-tuning of Equations (67) and (68) can be generalized to the case of
several computational parameters, such as the inner and outer electron–phonon matrix elements by considering Fourier trans-
energy windows, and the projections used to initialize the forms for both the Kohn–Sham states and for the vibrational
minimization algorithm. Historically, this optimization has often eigenmodes. The resulting electron–phonon matrix elements in
been performed by using chemical intuition, and sometimes a the Wannier representation are given by10:
trial-and-error approach. rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
In an effort to enable more systematic construction of MLWFs 1 X iðkRp þqRp0 Þ X 2Mκ ωqν y
gmnκα ðRp ; Rp0 Þ ¼ e eκα;ν ðqÞUymm0 kþq gm0 n0 ν ðk; qÞUn0 nk :
for high-throughput calculations and data-driven materials N2uc k;q m0 n0 ν
_
research, several strategies have been developed, such as for (69)
example those of refs. 196,197. These strategies fall into two
categories, one relying on the projectability concept198, and the Once the gmnκα ðRp ; Rp0 Þ have been computed, this relation can be
other on the selected-columns-of-the-density-matrix (SCDM) inverted to generate electron-phonon matrix elements anywhere
method199,200. The former approach relies on the projectability in the Brillouin zone.
of Bloch states on pseudo-atomic orbitals associated with the Equations (67)–(69) constitute the backbone of the interpola-
pseudopotentials; the projectability metric of each Bloch state is tion engine of EPW. In practice, the interpolation module of EPW
used to downselect the band manifold for subsequent minimiza- reads in the DFT electron density, dynamical matrices, and
tion of the spread functional. The latter approach does not rely on variations of the Kohn–Sham potential evaluated by Quantum
the pseudo-atomic orbitals or other choice of initial projections ESPRESSO on a coarse Brillouin zone grid; computes the
but generates an optimal set of states from the columns of the Kohn–Sham wavefunctions and electron–phonon matrix elements
density matrix via QR decomposition with column pivoting. on this grid; calls the Wannier90 code in library mode to obtain
Both approaches are very promising and bring Wannier the Wannier matrices in Eq. (64); transforms Hamiltonian,
functions one step closer to complete automation. Nevertheless, dynamical matrix, and electron–phonon matrix elements in the
some challenges remain. For example, to automatically construct Wannier representation; and interpolates all these quantities onto
Wannier functions and achieve high-quality Wannier interpolation, arbitrarily dense k- and q-point grids. A qualitative schematic of
both strategies require the use of all the occupied valence bands this process is shown in Fig. 8.

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
19

Transform electrons, Interpolate from Wannier

Calculation of matrix
phonons, matrix elements basis to Bloch basis on
elements on coarse grid
to Wannier basis ultra-fine grid

Initialization for Wannier Rotation matrices for Compute physical

function transform on Wannier functions quantities:
coarse grid Lifetimes, transport,
superconductivity,
polarons, optics

epw.x

Electron ZG.x
Phonon dynamical matrices
wavefunctions
& potential variations on Compute physical
& eigenvalues
coarse grid quantities using the special
on coarse grid
displacement method:
Temp.-dependent bands,
optical spectra, inelastic
scattering

Fig. 8 Schematic overview of the interpolation engine of EPW and its relation to the Quantum ESPRESSO and Wannier90 codes. The
EPW/ZG module is a stand-alone supercell-based set of codes, as described in the “Results” section.

Treatment of long-range electron–phonon interactions obtained following the standard procedure207,208 and coming from
In semiconductors and insulators, the ionic displacements the q + G → 0 limit of the overlap integrals between Kohn–Sham
P
associated with a phonon can induce a variation of the wavefunctions hψmkþq jeiqr jψnk i ¼ p Umpkþq Uypnk þ OðqÞ. This
Kohn–Sham potential that is long-ranged in nature. As a result, matrix element reduces to the standard Fröhlich interaction204
the electron–phonon matrix element gmnν(k, q) associated with when one considers parabolic electron bands and a dispersionless
longitudinal–optical (LO) phonons becomes singular in the long- LO mode in a cubic material186,205. This term is of the order of
wavelength limit q → 0. More specifically, in any material ∣q∣−1. The quadrupole contribution is the second term of the
exhibiting non-vanishing Born effective charges, the matrix multipole expansion of the Kohn–Sham potential and is of the
elements associated with LO phonons diverge as q/∣q∣2 at small order of ∣q∣0. The corresponding matrix element gL;Qmnν ðk; qÞ was
q. In these cases, the assumption of locality that underpins the derived in refs. 40,41,187,209,210 and reads:
methodology described in the previous section ceases to hold,
and the Wannier interpolation procedure needs to be modified to L;Q 2 P _
1=2 P
eiðqþGÞτκ eκα;ν ðqÞ
gmnν ðk; qÞ ¼ 4π e
Ω 4πϵ0 2Mκ ωqν ðqþGÞϵ1 ðqþGÞ
correctly capture the singularity. "
κα G≠q

The singularity in the matrix element can be dealt with by P P

´ 2 Qκ;αβγ ðqβ
1
þ Gβ Þðqγ þ Gγ Þ UmpkþqþG Uypnk
considering a multipole expansion of the Kohn–Sham potentials βγ p
#
resulting from individual atomic displacements. The first order in P  
 Z κ;αβ ðqβ þ Gβ Þ UmskþqþG ðqγ þ Gγ Þ Aspk;γ
W
þ husk jV jupk i Uypnk :
W Hxc;E γ W
this expansion is the dipole potential203, which scales as q/∣q∣2 and sp

is responsible for the well-known Fröhlich electron–phonon (72)

interaction204. The modification of the electron–phonon inter-
polation method to include long-range effects consists of In this expression, Qκ is the dynamical quadrupole tensor211,
separating short-range and long-range parts in the matrix V Hxc;E γ ðrÞ is the self-consistent potential induced by a uniform
elements, as follows: electric field E γ along the Cartesian direction γ41,187, Aspk;γ
W
¼
gmnν ðk; qÞ ¼ gSmnν ðk; qÞ þ gL;D L;Q
mnν ðk; qÞ þ gmnν ðk; qÞ; (70) W
sk
W
γ pk
W
ih∂u =∂k ju i is the Berry connection, and u is the smooth pk
where the first term on the right-hand side is the short-range Bloch basis187. In EPW, the calculation of the Berry connection
component, the second term is the dipole component, and the term is automatically activated for quadrupole calculations when
third term is the quadrupole component. Additional multipoles the velocity matrix is computed in the Wannier representation, cf.
could be considered, but the dipole and quadrupole terms already Eq. (27) of ref. 22. This term can be non-negligible in particular for
lead to very accurate results. The dipole matrix element gL;D
mnν ðk; qÞ
materials where the quadrupole tensor Qκ vanishes by symmetry,
was derived in refs. 205,206 and reads: such as SrO187. The V Hxc ;E term has been implemented in
4π e2 X _ 1=2 X
P y Abinit40,41,212 and EPW187,213 and has been tested for Si, GaP,
L;D p UmpkþqþG Upnk
gmnν ðk; qÞ ¼ i eiðqþGÞτκ ðq þ GÞ  Zκ  eκν ðqÞ: GaAs, SrO, h-BN monolayer, MoS2 monolayer, and InSe monolayer.
Ω 4πϵ0 κ 2Mκ ωqν G≠q
ðq þ GÞ  ϵ1  ðq þ GÞ
In all these cases, it was found that the contribution from this term
(71) represents <0.1% of the total quadrupole correction41,187. Since
In this expression, ϵ0 is the vacuum permittivity, ϵ1 is the high- this term is very small and its implementation is rather
frequency dielectric tensor of the material, G represents a cumbersome, it has not been included in public releases of
reciprocal lattice vector, Zκ is the Born effective charge tensor of these codes.
the atom κ, the bracket indicates the integral over the BvK The calculation strategy employed by EPW is as follows. First, the
supercell, and we use the notation a ⋅ B ⋅ c = ∑αβaαBαβcβ. The complete matrix elements gmnν(k, q) are evaluated on coarse k and q
unitary matrices Umnk in Eq. (71) are the Wannier transformation grids using DFPT. Second, the long-range contributions gL;D mnν ðk; qÞ

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
20

Fig. 9 Accurate computation of long-range electron–phonon matrix elements with EPW. a Brillouin zone of c-BN, the system used in this
example. b–d Full, long-range and short-range electron–phonon matrix elements, respectively, for the LO phonon mode along the q-path
shown by the green line in (a). The red lines are for the choice of including only the dominant reciprocal lattice vector in the G sums. The dark
blue lines correspond to the choice of applying a Gaussian filter, the procedure implemented in EPW. Circles are the reference data from
explicit DFPT calculations. e Sensitivity of the interpolation on the α parameter of the Gaussian filter. f Number of G vectors included in the
sum as a function of α, and g corresponding maximum relative error with respect to DFPT reference data. h Comparison between the
descriptor Dν(k = 0, q) obtained from direct DFPT calculations (white circles) and from Wannier interpolation: without including long-range
contributions (red line); including the dipole term (orange line); and including both dipole and quadrupole terms (blue line). In each case, the
corresponding corrections to the phonon dynamical matrix are also included. i Zoom over the area highlighted in (h), focusing on the LA
phonon mode.

and gL;Q
mnν ðk; qÞ are subtracted from the DFPT matrix elements using Brillouin zone boundaries, which in turn cause spurious oscilla-
Eqs. (71) and (72), leaving out the short-range component gSmnν ðk; qÞ tions in the interpolation of gSmnν ðk; qÞ. We illustrate this point in
on the coarse grids. Third, the standard Wannier electron–phonon Fig. 9a–d, focusing on the longitudinal-optical (LO) phonon of
interpolation of ref. 10 is applied to the short-range component only. cubic boron nitride (c-BN).
And fourth, the long-range contributions are added back using Eqs. An alternative strategy to avoid the sum over G is to cut off the
(71) and (72) on the fine k and q grids. interaction range in Eqs. (71) and (72) using a Gaussian filter,
The infinite sum over the G vectors in Eqs. (71) and (72) ensure exp½ðq þ GÞ  ϵ1  ðq þ GÞ=4α. This choice finds motivation in
the periodicity of the matrix elements in reciprocal space. In the Ewald summation method that is commonly employed to
practical calculations, this sum might be computationally evaluate the non-analytic contribution to the dynamical matrix in
demanding to converge. In principle, periodicity could be polar materials7,8,202. It was employed in previous versions of EPW as
enforced by including only one reciprocal lattice vector in the well as in other codes13,40,206, and it avoids derivative discontinuity.
sum, namely G ~ q such that jq þ G ~ q j ¼ minG jq þ Gj. However, this However, it does not preserve the periodicity of the matrix elements
choice introduces derivative discontinuities in gL;D mnν ðk; qÞ at the in reciprocal space, which is required by the periodic gauge. In EPW,

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
21
to maintain the periodicity of the matrix elements, we sum over a couplings in 2D by evaluating the average dielectric polarization
shell of G vectors centered around G ~ q , and then apply the Gaussian of a 2D slab enclosing the layer and removing the interaction
filter. Figure 9c shows how this procedure yields periodic and between periodic replicas via Coulomb truncation. In ref. 215, the
smooth long-range matrix elements with the correct behavior next authors additionally consider the contributions to the polar matrix
to the singularities, and enables accurate interpolation of the elements arising from out-of-plane dipoles and demonstrate an
complete matrix elements, as shown in Fig. 9b. improvement in the interpolation accuracy. In addition to these
The α parameter in the Gaussian filter is chosen in such a way as studies, more recent work by some of us focused on including
to ensure accurate interpolation (which requires large α) whilst quadrupole corrections for 2D materials187 and on establishing the
using as few G vectors as possible in Eqs. (71) and (72) (which connection between polar matrix elements in 3D and 2D
requires small α). In Fig. 9e we compare the interpolated matrix materials186. In particular, in ref. 187, the 2D long-range matrix
element gmnν(k, q) to explicit DFPT calculations, for different elements to quadrupolar order are given by
values of α, with the reciprocal space summation restricted to
_
1
X eκα;ν ðqÞ X X
those G vectors such that ðq þ GÞ  ϵ1  ðq þ GÞ<56α. With this
2
L;2D y
gL;2D
mnν ðk; qÞ ¼ pffiffiffiffiffiffi UmskþqþG huW
skþqþG jV qþGκα jupkþq iUpnk ;
W
2ωqν
choice, only G-vectors yielding a value of the Gaussian filter larger κα Mκ G≠q sp
than expð14Þ ’ 106 are included in the sum. We find that when (74)
α is chosen to match the size of the Brillouin zone [α = 1 (2π/a)2, a
is the lattice parameter], good interpolation is achieved. We note where the 2D long-range scattering potential to first order in q is
that, for layered materials, more G vectors are automatically " (
included in the sum along the perpendicular direction due to the πe f ðjqjÞ iqτκ
P P
V L;2D
qκα ðrÞ ¼ S jqj e
1
~ϵk ðqÞ
2iqβ Z κ;αβ þ qβ qγ Qκ;αβγ  jqj2 Qκ;αzz
anisotropy of the dielectric tensor. For completeness, in Fig. 9f we β γ
o n  oi
show how the number of required G vectors increases with α, and 2qβ Z κ;αβ qγ V Hxc;E γ ðrÞ þ ~ϵ?1ðqÞ 2jqj2 Z κ;αz z þ V Hxc;E z ðrÞ ;
in Fig. 9g we show the interpolation error in the matrix elements
as a function of α. We emphasize that, since EPW implements this (75)
Gaussian filter, caution should be used when extracting dipolar where S is the unit-cell area, and the dielectric functions appearing
and quadrupolar matrix elements from EPW for separate post- at the denominators are
processing, because gL;D L;Q
mnν ðk; qÞ and gmnν ðk; qÞ have the expected
q-dependence only at long wavelength. 2πf ðjqjÞ
~ϵk ðqÞ ¼ 1 þ q  αk  q; (76)
Figure 9h, i illustrate the importance of describing dipole and jqj
quadrupole interactions using Eqs. (71) and (72), for the case of
c-BN. To facilitate the comparison between the interpolation ~ϵ? ðqÞ ¼ 1  2πjqjf ðjqjÞ α? : (77)
results and explicit DFPT calculations for long-wavelength acoustic
k ⊥
modes, we use the following descriptor which removes the factor In these expressions, α and α are macroscopic polarizabilities in-
1=2
ωqν from the matrix element and averages over electronic and out-of-plane, respectively, and f ðqÞ ¼ 1  tanhðqL=2Þ is a
degeneracies: range-separation function and the distance L is determined
according to Eq. (63) of ref. 187. The 2D extension of the overlap
1 h X i1=2
Dν ðk; qÞ ¼ 2ρΩ_ωqν jgmnν ðk; qÞj2 : (73) integral is provided in Eqs. (46) and (47) of ref. 187. This formulation
_NW mn
also uses Coulomb truncation and generalizes the previous
Here, the sum over bands is carried over the NW states in the Wannier approaches of refs. 214,215 to the case of quadrupole corrections.
manifold and ρ is the mass density of the crystal. Dν(k, q) has units of In the approach of Ref. 186, the authors propose a unified
energy divided by length and can be thought of as a deformation formulation of polar electron–phonon interactions in 3D and 2D
potential of sort. In Fig. 9h, i, we choose the manifold composed by that reduces to the approach of refs. 206,217 in 3D, and to the
the top three valence bands of c-BN, and set k = 0. We see that the approach of refs. 214,215 in 2D. The theory is developed along the
dipole term is necessary to correctly describe the singular behavior of same lines as in ref. 217, by evaluating the potential of point
the LO phonon at long wavelength, and the quadrupole term is dipoles resulting from the atomic displacements. The key
necessary to correctly describe the discontinuous behavior of the difference is that, in this case, the dipoles belong to a supercell
longitudinal acoustic (LA) phonons at a long wavelength [see consisting of a dielectric stack where one component is the
expanded view in Fig. 9i]. The inclusion of both terms guarantees a material under consideration, and the other is a homogeneous
high-quality interpolation of the electron-phonon matrix elements dielectric (or vacuum). By solving this electrostatic problem, the
across the entire Brillouin zone. In particular, in this example, we find authors obtain the following matrix element:
that the deformation potential calculated by including dipole and  1=2 P
2
_ humkþqþG junk i P 1=2 iðqk þGk Þτκk
quadrupole terms agrees with direct DFPT calculations within gmnν ðk; qÞ ¼ 2ϵ0 ϵ1;env
e
Mκ e
G≠q j k
Ω 2ωqν q þ Gk j
κ
0.01 eV/Bohr (<1% of the maximum value) for the LA mode of P h i
∂Kðq þ G;τ κz Þ
c-BN along the ΓL line near the zone center. In refs. 22,187 a number of ´ Z κ;αβ eκβ;ν ðqÞ δα;k iðq þ GÞα Kðq þ G; τ κz Þ  δα;z ∂τ κz ;
αβ
additional comparisons between the present method and direct
DFPT calculations are discussed for standard three-dimensional (3D) (78)
and two-dimensional (2D) semiconductors. where ϵ∞,env is the dielectric constant of the homogeneous
The EPW code also implements long-range corrections for the dielectric surrounding the 2D material, and K(q + G, τκz) is a kernel
phonon dynamical matrix by computing dipole–dipole, that depends on the thickness of the 2D layer and the surrounding
dipole–quadrupole, and quadrupole–quadrupole terms as dis- medium, as well as their dielectric constants. The expression for K
cussed in refs. 7,8,202,211. Systematic tests of the interpolation is given by Eq. (A1) of ref. 186. The main practical difference
procedure described in this section are provided in ref. 22. between this formulation and other approaches is that it can be
The expressions for the long-range dipole and quadrupole used without the need for Coulomb truncation.
contributions to the electron–phonon matrix elements are given in
Eqs. (71) and (72) are for 3D bulk crystals. Several generalizations of
these expressions to the case of 2D materials and the transition
from 3D to 2D have been proposed during the past few DATA AVAILABILITY
years186,187,213–216. In the first of such generalizations214, the authors The data presented in each figure are available from the corresponding author upon
obtain the long-wavelength limit of polar electron–phonon reasonable request.

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
22
CODE AVAILABILITY 27. Restrepo, O. D., Varga, K. & Pantelides, S. T. First-principles calculations of
The codes used for the current study, namely Quantum ESPRESSO, Wannier90, and electron mobilities in silicon: phonon and Coulomb scattering. Appl. Phys. Lett.
EPW, are all open-source software distributed via their corresponding websites. In 94, 212103 (2009).
particular, the EPW code described in this manuscript can be downloaded from 28. Li, W. Electrical transport limited by electron–phonon coupling from Boltzmann
https://epw-code.org. The data presented in each figure are available from the transport equation: an ab initio study of Si, Al, and MoS2. Phys. Rev. B 92, 075405
corresponding author upon reasonable request. (2015).
29. Fiorentini, M. & Bonini, N. Thermoelectric coefficients of n-doped silicon from
first principles via the solution of the Boltzmann transport equation. Phys. Rev. B
Received: 15 February 2023; Accepted: 6 August 2023; 94, 085204 (2016).
30. Zhou, J.-J. & Bernardi, M. Ab initio electron mobility and polar phonon scattering
in GaAs. Phys. Rev. B 94, 201201 (2016).
31. Poncé, S., Margine, E. R. & Giustino, F. Towards predictive many-body calcula-
tions of phonon-limited carrier mobilities in semiconductors. Phys. Rev. B 97,
REFERENCES 121201(R) (2018).
1. Giustino, F. Electron–phonon interactions from first principles. Rev. Mod. Phys. 32. Macheda, F. & Bonini, N. Magnetotransport phenomena in p-doped diamond
89, 015003 (2017). from first principles. Phys. Rev. B 98, 201201(R) (2018).
2. Ziman, J. M. Electrons and Phonons (Oxford University Press, 1960). 33. Sohier, T., Campi, D., Marzari, N. & Gibertini, M. Mobility of two-dimensional
3. Hohenberg, P. & Kohn, W. Inhomogeneous electron gas. Phys. Rev. 136, materials from first principles in an accurate and automated framework. Phys.
B864–B871 (1964). Rev. Mater. 2, 114010 (2018).
4. Kohn, W. & Sham, L. J. Self-consistent equations including exchange and cor- 34. Ma, J., Nissimagoudar, A. S. & Li, W. First-principles study of electron and hole
relation effects. Phys. Rev. 140, A1133–A1138 (1965). mobilities of Si and GaAs. Phys. Rev. B 97, 045201 (2018).
5. Baroni, S., Giannozzi, P. & Testa, A. Green’s-function approach to linear response 35. Poncé, S., Schlipf, M. & Giustino, F. Origin of low carrier mobilities in halide
in solids. Phys. Rev. Lett. 58, 1861–1864 (1987). perovskites. ACS Energy Lett. 4, 456–463 (2019).
6. Savrasov, S. Y. Linear response calculations of lattice dynamics using muffin-tin 36. Poncé, S., Jena, D. & Giustino, F. Hole mobility of strained GaN from first prin-
basis sets. Phys. Rev. Lett. 69, 2819–2822 (1992). ciples. Phys. Rev. B 100, 085204 (2019).
7. Gonze, X. & Lee, C. Dynamical matrices, Born effective charges, dielectric per- 37. Poncé, S., Jena, D. & Giustino, F. Route to high hole mobility in GaN via reversal
mittivity tensors, and interatomic force constants from density-functional per- of crystal-field splitting. Phys. Rev. Lett. 123, 096602 (2019).
turbation theory. Phys. Rev. B 55, 10355–10368 (1997). 38. Lee, N.-E., Zhou, J. J., Chen, H. Y. & Bernardi, M. Ab initio electron-two-phonon
8. Baroni, S., de Gironcoli, S., Corso, A. D. & Giannozzi, P. Phonons and related scattering in GaAs from next-to-leading order perturbation theory. Nat. Com-
properties of extended systems from density-functional perturbation theory. mun. 11, 1607 (2020).
Rev. Mod. Phys. 73, 515 (2001). 39. Poncé, S. & Giustino, F. Structural, electronic, elastic, power, and
9. Marzari, N., Mostofi, A. A., Yates, J. R., Souza, I. & Vanderbilt, D. Maximally localized transport properties of β-Ga2O3 from first principles. Phys. Rev. Res. 2, 033102
Wannier functions: theory and applications. Rev. Mod. Phys. 84, 1419–1475 (2012). (2020).
10. Giustino, F., Cohen, M. L. & Louie, S. G. Electron–phonon interaction using 40. Brunin, G. et al. Phonon-limited electron mobility in Si, GaAs, and GaP
Wannier functions. Phys. Rev. B 76, 165108 (2007). with exact treatment of dynamical quadrupoles. Phys. Rev. B 102, 094308
11. GNU General Public License v3, Free Software Foundation. Gnu general public (2020).
license v3. https://www.gnu.org/licenses (2007). 41. Brunin, G. et al. Electron-phonon beyond Fröhlich: dynamical quadrupoles in
12. Giannozzi, P. et al. Advanced capabilities for materials modelling with QUAN- polar and covalent solids. Phys. Rev. Lett. 125, 136601 (2020).
TUM ESPRESSO. J. Phys.: Condens. Matter 29, 465901 (2017). 42. Macheda, F., Poncé, S., Giustino, F. & Bonini, N. Theory and computation of Hall
13. Zhou, J.-J. et al. Perturbo: A software package for ab initio electron–phonon scattering factor in graphene. Nano Lett. 20, 8861–8865 (2020).
interactions, charge transport and ultrafast dynamics. Comput. Phys. Commun. 43. Broyden, C. G. A class of methods for solving nonlinear simultaneous equations.
264, 107970 (2021). Math. Comput. 19, 577 (1965).
14. Protik, N. H., Li, C., Pruneda, M., Broido, D. & Ordejón, P. The elphbolt ab initio 44. Li, W., Carrete, J., Katcho, N. A. & Mingo, N. ShengBTE: a solver of the Boltzmann
solver for the coupled electron–phonon Boltzmann transport equations. npj transport equation for phonons. Comput. Phys. Commun. 185, 1747–1758
Comput. Mater. 8, 28 (2022). (2014).
15. Cepellotti, A., Coulter, J., Johansson, A., Fedorova, N. S. & Kozinsky, B. Phoebe: a 45. Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation
high-performance framework for solving phonon and electron Boltzmann made simple. Phys. Rev. Lett. 77, 3865 (1996).
transport equations. J. Phys. Mater. 5, 035003 (2022). 46. Hamann, D. R. Optimized norm-conserving Vanderbilt pseudopotentials. Phys.
16. Marchese, G. et al. Epiq (2022) (last accessed 27 January 2023); https://the-epiq- Rev. B 88, 085117 (2013).
team.gitlab.io/epiq-site. 47. van Setten, M. J. et al. The PseudoDojo: training and grading a 85 element
17. Noffsinger, J. et al. EPW: a program for calculating the electron–phonon cou- optimized norm-conserving pseudopotential table. Comput. Phys. Commun.
pling using maximally localized Wannier functions. Comput. Phys. Commun. 181, 226, 39–54 (2018).
2140–2148 (2010). 48. Di Paola, C., Macheda, F., Laricchia, S., Weber, C. & Bonini, N. First-principles
18. Poncé, S., Margine, E. R., Verdi, C. & Giustino, F. EPW: electron–phonon coupling, study of electronic transport and structural properties of Cu12Sb4S13 in its high-
transport and superconducting properties using maximally localized Wannier temperature phase. Phys. Rev. Res. 2, 033055 (2020).
functions. Comput. Phys. Commun. 209, 116–133 (2016). 49. Ashcroft, N. W. & Mermin, N. D.Solid State Physics (Saunders College Publishing,
19. Message Passing Interface Forum. MPI: A Message-Passing Interface Standard 1976).
Version 5.0 (Message Passing Interface Forum, 2022). 50. Lundstrom, M. Fundamentals of Carrier Transport 2nd edn (Cambridge University
20. Dagum, L. & Menon, R. Openmp: an industry standard API for shared-memory Press, 2000).
programming. Comput. Sci. Eng. 5, 46–55 (1998). 51. Brooks, H. Theory of the electrical properties of germanium and silicon. In
21. The HDF Group. Hierarchical Data Format, version 5 (1997-NNNN). https:// Advances in Electronics and Electron Physics (ed., Marton, L.), Ch. 2, 85–182
www.hdfgroup.org/HDF5/. (Academic Press, 1955).
22. Poncé, S. et al. First-principles predictions of Hall and drift mobilities in semi- 52. Debye, P. P. & Conwell, E. M. Electrical properties of n-type germanium. Phys.
conductors. Phys. Rev. Res. 3, 043022 (2021). Rev. 93, 693–706 (1954).
23. Kadanoff, L. P. & Baym, G. Quantum Statistical Mechanics: Green’s Function 53. Restrepo, O. D., Varga, K. & Pantelides, S. T. First-principles calculations of
Methods in Equilibrium and Nonequilibrium Problems. Frontiers in Physics. A electron mobilities in silicon: phonon and Coulomb scattering. Appl. Phys. Lett.
Lecture Note and Reprint Series (Benjamin, 1962). 94, 212103 (2009).
24. Poncé, S., Li, W., Reichardt, S. & Giustino, F. First-principles calculations of charge 54. Graziosi, P., Kumarasinghe, C. & Neophytou, N. Material descriptors for the
carrier mobility and conductivity in bulk semiconductors and two-dimensional discovery of efficient thermoelectrics. ACS Appl. Energy Mater. 3, 5913–5926
materials. Rep. Prog. Phys. 83, 036501 (2020). (2020).
25. Mahan, G. D. Quantum transport equation for electric and magnetic fields. Phys. 55. Ganose, A. M. et al. Efficient calculation of carrier scattering rates from first
Rep. 145, 251–318 (1987). principles. Nat. Commun. 12, 2222 (2021).
26. Macheda, F. Ab Initio Prediction of Thermomagnetic and Thermoelectric 56. Lu, I.-T., Zhou, J.-J., Park, J. & Bernardi, M. First-principles ionized-impurity scat-
Transport Phenomena in 3d and 2d Materials. Ph.D. thesis, King’s College tering and charge transport in doped materials. Phys. Rev. Mater. 6, L010801
London (2021). (2022).

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
23
57. Leveillee, J., Zhang, X., Kioupakis, E. & Giustino, F. Ab initio calculation of carrier 91. Margine, E. R., Lambert, H. & Giustino, F. Electron–phonon interaction and
mobility in semiconductors including ionized-impurity scattering. Phys. Rev. B pairing mechanism in superconducting Ca-intercalated bilayer graphene. Sci.
107, 125207 (2023). Rep. 6, 1–6 (2016).
58. Kohn, W. & Luttinger, J. M. Quantum theory of electrical transport phenomena. 92. Aperis, A. & Oppeneer, P. M. Multiband full-bandwidth anisotropic Eliashberg
Phys. Rev. 108, 590–611 (1957). theory of interfacial electron–phonon coupling and high Tc superconductivity in
59. Lindhard, J. On the properties of a gas of charged particles. Dan. Mat. Fys. Medd. FeSe/SrTiO3. Phys. Rev. B 97, 060501 (2018).
28, 3–58 (1954). 93. Allen, P. B. Fermi-surface harmonics: a general method for nonspherical pro-
60. Sanders, N. & Kioupakis, E. Phonon- and defect-limited electron and hole blems. application to Boltzmann and Eliashberg equations. Phys. Rev. B 13,
mobility of diamond and cubic boron nitride: a critical comparison. Appl. Phys. 1416–1427 (1976).
Lett. 119, 062101 (2021). 94. Vidberg, H. & Serene, J. Solving the Eliashberg equations by means of N-point
61. Schlipf, M. & Gygi, F. Optimization algorithm for the generation of ONCV Padé approximants. J. Low Temp. Phys. 29, 179–192 (1977).
pseudopotentials. Comput. Phys. Commun. 196, 36 (2015). 95. Leavens, C. & Ritchie, D. Extension of the N-point Padé approximants solution of
62. Canali, C., Jacoboni, C., Nava, F., Ottaviani, G. & Alberigi-Quaranta, A. Electron the Eliashberg equations to T ~ Tc. Solid State Commun. 53, 137–142 (1985).
drift velocity in silicon. Phys. Rev. B 12, 2265–2284 (1975). 96. Marsiglio, F., Schossmann, M. & Carbotte, J. P. Iterative analytic continuation of
63. Morin, F. J. & Maita, J. P. Electrical properties of silicon containing arsenic and the electron self-energy to the real axis. Phys. Rev. B 37, 4965–4969 (1988).
boron. Phys. Rev. 96, 28–35 (1954). 97. Morel, P. & Anderson, P. W. Calculation of the superconducting state parameters
64. Jacoboni, C., Canali, C., Ottaviani, G. & Alberigi Quaranta, A. A review of with retarded electron–phonon interaction. Phys. Rev. 125, 1263–1271 (1962).
some charge transport properties of silicon. Solid State Electron. 20, 77–89 98. Heil, C. et al. Origin of superconductivity and latent charge density wave in
(1977). NbS2. Phys. Rev. Lett. 119, 087003 (2017).
65. Caruso, F. & Giustino, F. Theory of electron–plasmon coupling in semi- 99. Heil, C., di Cataldo, S., Bachelet, G. B. & Boeri, L. Superconductivity in sodalite-like
conductors. Phys. Rev. B 94, 115208 (2016). yttrium hydride clathrates. Phys. Rev. B 99, 220502 (2019).
66. Misiakos, K. & Tsamakis, D. Accurate measurements of the silicon intrinsic carrier 100. Di Cataldo, S., Heil, C., von der Linden, W. & Boeri, L. LaBH8: towards high-Tc low-
density from 78 to 340 K. J. Appl. Phys. 74, 3293–3297 (1993). pressure superconductivity in ternary superhydrides. Phys. Rev. B 104, L020511
67. McMillan, W. L. Transition temperature of strong-coupled superconductors. (2021).
Phys. Rev. 167, 331 (1968). 101. Pickett, W. E. Generalization of the theory of the electron–phonon interaction:
68. Eliashberg, G. Interactions between electrons and lattice vibrations in a super- thermodynamic formulation of superconducting- and normal-state properties.
conductor. Sov. Phys.-JETP 11, 696 (1960). Phys. Rev. B 26, 1186–1207 (1982).
69. Eliashberg, G. Temperature Green’s function for electrons in a superconductor. 102. Ummarino, G. A. C. Eliashberg Theory, Vol. 3 of Lecture Notes of the Autumn
Sov. Phys.-JETP 12, 1000 (1961). School Correlated Electrons (Forschungszentrum Jülich GmbH Institute for
70. Oliveira, L. N., Gross, E. K. U. & Kohn, W. Density-functional theory for super- Advanced Simulation, 2013).
conductors. Phys. Rev. Lett. 60, 2430–2433 (1988). 103. Davydov, A. et al. Ab initio theory of plasmonic superconductivity within the
71. Lüders, M. et al. Ab initio theory of superconductivity. I. Density Eliashberg and density-functional formalisms. Phys. Rev. B 102, 214508 (2020).
functional formalism and approximate functionals. Phys. Rev. B 72, 024545 104. Kafle, G. P. et al. Ab initio study of Li–Mg–B superconductors. Phys. Rev. Mater. 6,
(2005). 084801 (2022).
72. Marques, M. A. L. et al. Ab initio theory of superconductivity. II. Application to 105. Sano, W., Koretsune, T., Tadano, T., Akashi, R. & Arita, R. Effect of van Hove
elemental metals. Phys. Rev. B 72, 024546 (2005). singularities on high-Tc superconductivity in H3S. Phys. Rev. B 93, 094525 (2016).
73. Sanna, A., Pellegrini, C. & Gross, E. K. U. Combining Eliashberg theory with 106. Schrodi, F., Oppeneer, P. M. & Aperis, A. Full-bandwidth Eliashberg theory of
density functional theory for the accurate prediction of superconducting superconductivity beyond Migdal’s approximation. Phys. Rev. B 102, 024503
transition temperatures and gap functions. Phys. Rev. Lett. 125, 057001 (2020).
(2020). 107. Valla, T. et al. Quasiparticle spectra, charge-density waves, superconductivity,
74. Allen, P. B. & Dynes, R. C. Transition temperature of strong-coupled super- and electron-phonon coupling in 2H-NbSe2. Phys. Rev. Lett. 92, 086401 (2004).
conductors reanalyzed. Phys. Rev. B 12, 905–922 (1975). 108. Weber, F. et al. Extended phonon collapse and the origin of the charge-density
75. Carbotte, J. P. Properties of boson-exchange superconductors. Rev. Mod. Phys. wave in 2H-NbSe2. Phys. Rev. Lett. 107, 107403 (2011).
62, 1027–1157 (1990). 109. Leroux, M. et al. Strong anharmonicity induces quantum melting of
76. Sanna, A. et al. Ab initio Eliashberg theory: making genuine predictions of charge density wave in 2H-NbSe2 under pressure. Phys. Rev. B 92, 140303(R)
superconducting features. J. Phys. Soc. Jpn. 87, 041012 (2018). (2015).
77. Xie, S. R. et al. Machine learning of superconducting critical temperature from 110. Yokoya, T. et al. Fermi surface sheet-dependent superconductivity in 2H-NbSe2.
Eliashberg theory. npj Comput. Mater. 8, 1–8 (2022). Science 294, 2518 (2001).
78. Gor’kov, L. P. Microscopic derivation of the Ginzburg–Landau equations in the 111. Anikin, A. et al. Ultrafast dynamics in the high-symmetry and in the charge
theory of superconductivity. Sov. Phys.-JETP 7, 505 (1958). density wave phase of 2H-NbSe2. Phys. Rev. B 102, 205139 (2020).
79. Nambu, Y. Quasi-particles and gauge invariance in the theory of super- 112. Sanna, A. et al. Real-space anisotropy of the superconducting gap in the charge-
conductivity. Phys. Rev. 117, 648–663 (1960). density wave material 2H-NbSe2. npj Quantum Mater. 7, 6 (2022).
80. Scalapino, D. J., Schrieffer, J. R. & Wilkins, J. W. Strong-coupling super- 113. Klimeš, J., Bowler, D. R. & Michaelides, A. van der Waals density functionals
conductivity. I. Phys. Rev. 148, 263–279 (1966). applied to solids. Phys. Rev. B 83, 195131 (2011).
81. Scalapino, D. J. The Electron–phonon Interaction and Strong-coupling Super- 114. Klimeš, J., Bowler, D. R. & Michaelides, A. Chemical accuracy for the van der
conductors, Vol. 1 (Dekker, 1969). Waals density functional. J. Phys.: Condens. Matter 22, 022201 (2010).
82. Allen, P. B. & Mitrović, B. Theory of Superconducting Tc, Vol. 37 of Solid State 115. Thonhauser, T. et al. Spin signature of nonlocal correlation binding in metal-
Physics (Academic Press, 1983). organic frameworks. Phys. Rev. Lett. 115, 136402 (2015).
83. Choi, H. J., Cohen, M. L. & Louie, S. G. Anisotropic Eliashberg theory of MgB2: Tc, 116. Thonhauser, T. et al. van der Waals density functional: self-consistent potential
isotope effects, superconducting energy gaps, quasiparticles, and specific heat. and the nature of the van der Waals bond. Phys. Rev. B 76, 125112 (2007).
Phys. C: Supercond. 385, 66–74 (2003). 117. Methfessel, M. & Paxton, A. T. High-precision sampling for Brillouin-zone inte-
84. Marsiglio, F. & Carbotte, J. P. Electron–Phonon Superconductivity (Springer, 2008). gration in metals. Phys. Rev. B 40, 3616 (1989).
85. Margine, E. R. & Giustino, F. Anisotropic Migdal–Eliashberg theory using Wannier 118. Zheng, F. & Feng, J. Electron-phonon coupling and the coexistence of super-
functions. Phys. Rev. B 87, 024505 (2013). conductivity and charge-density wave in monolayer NbSe2. Phys. Rev. B 99,
86. Marsiglio, F. Eliashberg theory: a short review. Ann. Phys. 417, 168102 (2020). 161119 (2019).
87. Migdal, A. Interaction between electrons and lattice vibrations in a normal 119. Wickramaratne, D., Khmelevskyi, S., Agterberg, D. F. & Mazin, I. I. Ising super-
metal. Sov. Phys.-JETP 34, 996 (1958). conductivity and magnetism in NbSe2. Phys. Rev. X 10, 041003 (2020).
88. Hedin, L. New method for calculating the one-particle Green’s 120. Das, S., Paudyal, H., Margine, E. R., Agterberg, D. F. & Mazin, I. I. Electron–phonon
function with application to the electron-gas problem. Phys. Rev. 139, A796 coupling and spin fluctuations in the Ising superconductor NbSe2. npj Comput.
(1965). Mater. 9, 66 (2023).
89. Hybertsen, M. S. & Louie, S. G. Electron correlation in semiconductors and 121. Alexandrov, A. S. & Devreese, J. T.Advances in Polaron Physics, Vol. 159 of Springer
insulators: Band gaps and quasiparticle energies. Phys. Rev. B 34, 5390 Series in Solid-State Sciences (Springer, 2010).
(1986). 122. Emin, D. Polarons (Cambridge University Press, 2012).
90. Lee, K.-H., Chang, K. J. & Cohen, M. L. First-principles calculations of the coulomb 123. Franchini, C., Reticcioli, M., Setvin, M. & Diebold, U. Polarons in materials. Nat.
pseudopotential μ*: application to Al. Phys. Rev. B 52, 1425–1428 (1995). Rev. Mater. 6, 560 (2021).

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
24
124. Deskins, N. A. & Dupuis, M. Electron transport via polaron hopping in bulk TiO2: 158. Madelung, O. Semiconductors: Group IV Elements and III–V Compounds (Springer,
a density functional theory characterization. Phys. Rev. B 75, 195212 (2007). 1991).
125. Franchini, C., Kresse, G. & Podloucky, R. Polaronic hole trapping in doped BaBiO3. 159. Petretto, G. et al. High-throughput density-functional perturbation theory
Phys. Rev. Lett. 102, 256402 (2009). phonons for inorganic materials. Sci. data 5, 1–12 (2018).
126. Lany, S. & Zunger, A. Polaronic hole localization and multiple hole binding of 160. Dolling, G. Inelastic scattering of neutrons in solids and liquids. IAEA 2, 37 (1963).
acceptors in oxide wide-gap semiconductors. Phys. Rev. B 80, 085202 (2009). 161. Nilsson, G. & Nelin, G. Study of the homology between silicon and germanium
127. Varley, J. B., Janotti, A., Franchini, C. & Van de Walle, C. G. Role of self-trapping in by thermal-neutron spectrometry. Phys. Rev. B 6, 3777–3786 (1972).
luminescence and p-type conductivity of wide-band-gap oxides. Phys. Rev. B 85, 162. Schinke, C. et al. Uncertainty analysis for the coefficient of band-to-band
081109(R) (2012). absorption of crystalline silicon. AIP Adv. 5, 067168 (2015).
128. Setvin, M. et al. Direct view at excess electrons in TiO2 rutile and anatase. Phys. 163. Antonius, G., Poncé, S., Boulanger, P., Côté, M. & Gonze, X. Many-body effects on
Rev. Lett. 113, 086402 (2014). the zero-point renormalization of the band structure. Phys. Rev. Lett. 112,
129. Kokott, S., Levchenko, S. V., Rinke, P. & Scheffler, M. First-principles supercell 215501 (2014).
calculations of small polarons with proper account for long-range polarization 164. Zacharias, M. & Giustino, F. One-shot calculation of temperature-dependent
effects. New J. Phys. 20, 033023 (2018). optical spectra and phonon-induced band-gap renormalization. Phys. Rev. B 94,
130. Falletta, S. & Pasquarello, A. Polarons free from many-body self-interaction in 075125 (2016).
density functional theory. Phys. Rev. B 106, 125119 (2022). 165. Zacharias, M. & Giustino, F. Theory of the special displacement method for
131. Sio, W. H., Verdi, C., Poncé, S. & Giustino, F. Polarons from first principles, without electronic structure calculations at finite temperature. Phys. Rev. Res. 2, 013357
supercells. Phys. Rev. Lett. 122, 246403 (2019). (2020).
132. Sio, W. H., Verdi, C., Poncé, S. & Giustino, F. Ab initio theory of polarons: 166. Zacharias, M. et al. Multiphonon diffuse scattering in solids from first principles:
Formalism and applications. Phys. Rev. B 99, 235139 (2019). Application to layered crystals and two-dimensional materials. Phys. Rev. B 104,
133. Lafuente-Bartolomé, J. et al. Unified approach to polarons and phonon-induced 205109 (2021).
band structure renormalization. Phys. Rev. Lett. 129, 076402 (2022). 167. Brüesch, P. Phonons, Theory and Experiments, Vol. 1 (Springer, 1982).
134. Lafuente-Bartolomé, J. et al. Ab initio self-consistent many-body theory of 168. Zacharias, M. & Kelires, P. C. Temperature dependence of the optical properties
polarons at all couplings. Phys. Rev. B 106, 075119 (2022). of silicon nanocrystals. Phys. Rev. B 101, 245122 (2020).
135. Maradudin, A. A. & Vosko, S. H. Symmetry properties of the normal vibrations of 169. Zacharias, M. & Kelires, P. C. Quantum confinement of electron–phonon cou-
a crystal. Rev. Mod. Phys. 40, 1–37 (1968). pling in graphene quantum dots. J. Phys. Chem. Lett. 12, 9940–9946 (2021).
136. Davidson, E. R. The iterative calculation of a few of the lowest eigenvalues and 170. Ramírez, R., Herrero, C. P. & Hernández, E. R. Path-integral molecular dynamics
corresponding eigenvectors of large real-symmetric matrices. J. Comput. Phys. simulation of diamond. Phys. Rev. B 73, 245202 (2006).
17, 87–94 (1975). 171. Allen, P. B. & Heine, V. Theory of the temperature dependence of electronic
137. Makov, G. & Payne, M. C. Periodic boundary conditions in ab initio calculations. band structures. J. Phys. C: Solid State Phys. 9, 2305 (1976).
!
Phys. Rev. B 51, 4014–4022 (1995). 172. Popescu, V. & Zunger, A. Extracting E versus k effective band structure from
138. Castner, T. G. & Känzig, W. The electronic structure of V-centers. J. Phys. Chem. supercell calculations on alloys and impurities. Phys. Rev. B 85, 085201
Solids 3, 178–195 (1957). (2012).
139. Jette, A. N., Gilbert, T. L. & Das, T. P. Theory of the self-trapped hole in the alkali 173. Zacharias, M. et al. Efficient first-principles methodology for the calculation of
halides. Phys. Rev. 184, 884–894 (1969). the all-phonon inelastic scattering in solids. Phys. Rev. Lett. 127, 207401
140. Fröhlich, H., Pelzer, H. & Zienau, S. XX. Properties of slow electrons in polar (2021).
materials. Lond. Edinb. Dublin Philos. Mag. J. Sci. 41, 221–242 (1950). 174. Perdew, J. P. & Zunger, A. Self-interaction correction to density-functional
141. Holstein, T. Studies of polaron motion: Part I. The molecular-crystal model. Ann. approximations for many-electron systems. Phys. Rev. B 23, 5048–5079
Phys. 8, 325–342 (1959). (1981).
142. Bassani, F. & Parravicini, G. P. Electronic States and Optical Transitions in Solids 175. Ceperley, D. M. & Alder, B. J. Ground state of the electron gas by a stochastic
(Pergamon Press, 1975). method. Phys. Rev. Lett. 45, 566–569 (1980).
143. Noffsinger, J., Kioupakis, E., Van de Walle, C. G., Louie, S. G. & Cohen, M. L. 176. Giustino, F., Louie, S. G. & Cohen, M. L. Electron–phonon renormalization of the
Phonon-assisted optical absorption in silicon from first principles. Phys. Rev. Lett. direct band gap of diamond. Phys. Rev. Lett. 105, 265501 (2010).
108, 167402 (2012). 177. Varshni, Y. P. Temperature dependence of the energy gap in semiconductors.
144. Giustino, F. Materials Modelling Using Density Functional Theory: Properties and Physica 34, 149–154 (1967).
Predictions (Oxford University Press, 2014). 178. Karsai, F., Engel, M., Kresse, G. & Flage-Larsen, E. Electron-phonon coupling
145. Tiwari, S. & Giustino, F. Ab-initio many-body calculations of optical absorption in semiconductors within the GW approximation. N. J. Phys. 20, 123008
spectra including both direct and indirect phonon-assisted processes (2023) (2018).
(submitted). 179. Poncé, S. et al. Temperature dependence of the electronic structure of semi-
146. Levine, Z. H. & Allan, D. C. Quasiparticle calculation of the dielectric response of conductors and insulators. J. Chem. Phys. 143, 102813 (2015).
silicon and germanium. Phys. Rev. B 43, 4187–4207 (1991). 180. Nery, J. P. & Allen, P. B. Influence of Fröhlich polaron coupling on renormalized
147. Rohlfing, M. & Louie, S. G. Electron-hole excitations and optical spectra from first electron bands in polar semiconductors: results for zinc-blende GaN. Phys. Rev. B
principles. Phys. Rev. B 62, 4927–4944 (2000). 94, 115135 (2016).
148. Song, B. et al. Optical properties of cubic boron arsenide. Appl. Phys. Lett. 116, 181. Kang, Y., Peelaers, H., Krishnaswamy, K. & Van de Walle, C. G. First-principles
141903 (2020). study of direct and indirect optical absorption in BaSnO3. Appl. Phys. Lett 112,
149. Zhang, X. & Kioupakis, E. Phonon-assisted optical absorption of SiC polytypes 062106 (2018).
from first principles. Phys. Rev. B 107, 115207 (2023). 182. Lee, H. et al. EPW Project (2023) (last retrieved 02-06-2023) https://epw-code.org.
150. Zhang, X., Shi, G., Leveillee, J. A., Giustino, F. & Kioupakis, E. Ab initio theory of 183. Engel, M. et al. Zero-point renormalization of the band gap of semiconductors
free-carrier absorption in semiconductors. Phys. Rev. B 106, 205203 (2022). and insulators using the projector augmented wave method. Phys. Rev. B 106,
151. Mo, Y., Tang, H., Bansil, A. & Tao, J. Accurate lattice geometrical parameters and 094316 (2022).
bulk moduli from a semilocal density functional. AIP Adv. 8, 095209 (2018). 184. Bushick, K. & Kioupakis, E. Phonon-assisted Auger-Meitner recombination in
152. Haas, P., Tran, F. & Blaha, P. Calculation of the lattice constant of solids with silicon from first principles. Phys. Rev. Lett. 131, 076902 (2023).
semilocal functionals. Phys. Rev. B 79, 085104 (2009). 185. Zacharias, M., Volonakis, G., Giustino, F. & Even, J. Anharmonic lattice
153. Favot, F. & Dal Corso, A. Phonon dispersions: Performance of the generalized dynamics via the special displacement method. Phys. Rev. B 108, 035155
gradient approximation. Phys. Rev. B 60, 11427–11431 (1999). (2023).
154. Deslippe, J. et al. BerkeleyGW: a massively parallel computer package for the 186. Sio, W. H. & Giustino, F. Unified ab initio description of fröhlich electron-phonon
calculation of the quasiparticle and optical properties of materials and nanos- interactions in two-dimensional and three-dimensional materials. Phys. Rev. B
tructures. Comput. Phys. Commun. 183, 1269–1289 (2012). 105, 115414 (2022).
155. Sze, S. M., Li, Y. & Ng, K. K. Physics and Properties of Semiconductors—A Review, 187. Poncé, S., Royo, M., Stengel, M., Marzari, N. & Gibertini, M. Long-range electro-
Ch. 1, 5–75 (Wiley, 2006). static contribution to electron–phonon couplings and mobilities of two-
156. Wachs, A. L., Miller, T., Hsieh, T. C., Shapiro, A. P. & Chiang, T. C. Angle-resolved dimensional and bulk materials. Phys. Rev. B 107, 155424 (2023).
photoemission studies of Ge(111)-c(2 × 8), Ge(111)-(1 × 1)H, Si(111)-(7 × 87), and 188. Li, Z., Antonius, G., Wu, M., da Jornada, F. H. & Louie, S. G. Electron–phonon
Si(100)-(2 × 1). Phys. Rev. B 32, 2326–2333 (1985). coupling from ab initio linear-response theory within the GW method:
157. Straub, D., Ley, L. & Himpsel, F. J. Conduction-band and surface-state critical correlation-enhanced interactions and superconductivity in Ba1−xKxBiO3. Phys.
points in Si: an inverse-photoemission study. Phys. Rev. Lett. 54, 142–145 (1985). Rev. Lett. 122, 186402 (2019).

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 H. Lee et al.
25
189. Gonze, X. et al. The Abinit project: Impact, environment and recent develop- 222. Zucca, R. R. L. & Shen, Y. R. Wavelength-modulation spectra of some semi-
ments. Comput. Phys. Commun. 248, 107042 (2020). conductors. Phys. Rev. B 1, 2668–2676 (1970).
190. Kresse, G. & Hafner, J. Ab initio molecular dynamics for liquid metals. Phys. Rev. B 223. Aspnes, D. E. & Studna, A. A. Dielectric functions and optical parameters of Si,
47, 558–561 (1993). Ge, GaP, GaAs, GaSb, InP, InAs, and InSb from 1.5 to 6.0 eV. Phys. Rev. B 27,
191. Soler, J. M. et al. The SIESTA method for ab initio order-N materials simulation. J. 985–1009 (1983).
Phys.: Condens. Matter 14, 2745 (2002). 224. Tiedje, T., Yablonovitch, E., Cody, G. & Brooks, B. Limiting efficiency of silicon
192. Enkovaara, J. et al. Electronic structure calculations with GPAW: a real-space solar cells. IEEE Trans. Electron Devices 31, 711–716 (1984).
implementation of the projector augmented-wave method. J. Phys.: Condens. 225. Green, M. A. & Keevers, M. J. Optical properties of intrinsic silicon at 300 K. Prog.
Matter 22, 253202 (2010). Photovolt. 3, 189–192 (1995).
193. Gamma, E., Helm, R., Johnson, R. & Vlissides, J. Design Patterns, Elements of 226. Kim, H. J. et al. Physical properties of transparent perovskite oxides (Ba,La)SnO3
Reusable Object-Oriented Software (Addison-Wesley, 1995). with high electrical mobility at room temperature. Phys. Rev. B 86, 165205
194. Ismail-Beigi, S. & Arias, T. A. New algebraic formulation of density functional (2012).
calculation. Comput. Phys. Commun. 128, 1–45 (2000). 227. Lee, H. et al. EPW v6-alpha (last accessed 5 November 2023). https://docs.epw-
195. Pizzi, G. et al. Wannier90 as a community code: new features and applications. J. code.org/doc/DownloadAndInstall.html (2023).
Phys.: Condens. Matter. 32, 165902 (2019).
196. Vitale, V. et al. Automated high-throughput wannierisation. npj Comput. Mater.
6, 66 (2020).
ACKNOWLEDGEMENTS
197. Qiao, J., Pizzi, G. & Marzari, N. Projectability disentanglement for accurate and
automated electronic-structure Hamiltonians. Preprint athttps://doi.org/ This research is supported by: the Computational Materials Sciences Program
10.48550/arXiv.2303.07877 (2023). funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences,
198. Agapito, L. A., Ferretti, A., Calzolari, A., Curtarolo, S. & Buongiorno Nardelli, M. under Award No. DE-SC0020129 (project coordination, scale-up, polaron module,
Effective and accurate representation of extended Bloch states on finite Hilbert transport module, optics module, special displacement module); the National
spaces. Phys. Rev. B 88, 165127 (2013). Science Foundation, Office of Advanced Cyberinfrastructure and Division of
199. Damle, A., Lin, L. & Ying, L. Compressed representation of Kohn–Sham orbitals Materials Research under Grants Nos. 2103991 and 2035518 (superconductivity
via selected columns of the density matrix. J. Chem. Theory Comput. 11, module, interoperability); the NSF Characteristic Science Applications for the
1463–1469 (2015). Leadership Class Computing Facility program under Grant No. 2139536 (prepara-
200. Damle, A. & Lin, L. Disentanglement via entanglement: a unified method for tion for LCCF); the Fond National de la Recherche Scientifique of Belgium (F.R.S.-
wannier localization. Multiscale Model. Simul. 16, 1392–1410 (2018). FNRS) and the European Union’s Horizon 2020 research and innovation program
201. Yates, J. R., Wang, X., Vanderbilt, D. & Souza, I. Spectral and Fermi surface under grant agreements No. 881603-Graphene Core3 (transport module); the NSF
properties from Wannier interpolation. Phys. Rev. B 75, 195121 (2007). DMREF award 2119555 (quasi-degenerate perturbation theory module). This
202. Giannozzi, P., de Gironcoli, S., Pavone, P. & Baroni, S. Ab initio calculation of
research used resources of the National Energy Research Scientific Computing
phonon dispersions in semiconductors. Phys. Rev. B 43, 7231–7242 (1991).
Center and the Argonne Leadership Computing Facility, which are DOE Office of
203. Vogl, P. Microscopic theory of electron–phonon interaction in insulators or
Science User Facilities supported by the Office of Science of the U.S. Department of
semiconductors. Phys. Rev. B 13, 694–704 (1976).
204. Fröhlich, H. Electrons in lattice fields. Adv. Phys. 3, 325–361 (1954). Energy, under Contracts No. DE-AC02-05CH11231 and DE-AC02-06CH11357,
205. Verdi, C. & Giustino, F. Fröhlich electron–phonon vertex from first principles. respectively. The authors acknowledge the Texas Advanced Computing Center
Phys. Rev. Lett. 115, 176401 (2015). (TACC) at The University of Texas at Austin for providing access to Frontera,
206. Sjakste, J., Vast, N., Calandra, M. & Mauri, F. Wannier interpolation of the Lonestar6, and Texascale Days, which have contributed to the research results
electron-phonon matrix elements in polar semiconductors: polar-optical cou- reported within this paper (http://www.tacc.utexas.edu); the Extreme Science and
pling in GaAs. Phys. Rev. B 92, 054307 (2015). Engineering Discovery Environment (XSEDE)218 which is supported by National
207. Marzari, N. & Vanderbilt, D. Maximally localized generalized Wannier functions Science Foundation grant number ACI-1548562, and in particular Expanse at the
for composite energy bands. Phys. Rev. B 56, 12847–12865 (1997). San Diego Supercomputer Center through allocation TG-DMR180071. S.P. acknowl-
208. Souza, I., Marzari, N. & Vanderbilt, D. Maximally localized Wannier functions for edges computational resources provided by the PRACE award granting access to
entangled energy bands. Phys. Rev. B 65, 035109 (2001). Discoverer in SofiaTech, Bulgaria (OptoSpin project id. 2020225411), and by the
209. Jhalani, V. A., Zhou, J.-J., Park, J., Dreyer, C. E. & Bernardi, M. Piezoelectric Consortium des Équipements de Calcul Intensif (CÉCI), funded by the FRS-FNRS
electron–phonon interaction from ab initio dynamical quadrupoles: Impact on under Grant No. 2.5020.11 and by the Walloon Region, as well as computational
charge transport in wurtzite GaN. Phys. Rev. Lett. 125, 136602 (2020).
resources awarded on the Belgian share of the EuroHPC LUMI supercomputer. K.B.
210. Park, J., Zhou, J.-J., Jhalani, V. A., Dreyer, C. E. & Bernardi, M. Long-range quad-
acknowledges the support of the U.S. Department of Energy, Office of Science,
rupole electron–phonon interaction from first principles. Phys. Rev. B 102,
Office of Advanced Scientific Computing Research, Department of Energy
125203 (2020).
211. Royo, M. & Stengel, M. First-principles theory of spatial dispersion: dynamical Computational Science Graduate Fellowship under Award Number DE-SC0020347.
quadrupoles and flexoelectricity. Phys. Rev. X 9, 021050 (2019). The authors wish to thank Zhenbang Dai, Nikolaus Kandolf, Viet-Anh Ha, and Amanda
212. Gonze, X. et al. Recent developments in the ABINIT software package. Comput. Wang for their contributions to the EPW project that are not discussed in this
Phys. Commun. 205, 106–131 (2016). manuscript; John Cazes and Hang Liu at TACC for their support with the
213. Poncé, S., Royo, M., Gibertini, M., Marzari, N. & Stengel, M. Accurate prediction of Characteristic Science Applications project, Paolo Giannozzi for his support with
Hall mobilities in two-dimensional materials through gauge-covariant quad- Quantum ESPRESSO; and Stefano Baroni for fruitful discussions.
rupolar contributions. Phys. Rev. Lett. 130, 166301 (2023).
214. Sohier, T., Calandra, M. & Mauri, F. Two-dimensional Fröhlich interaction in
transition-metal dichalcogenide monolayers: theoretical modeling and first- AUTHOR CONTRIBUTIONS
principles calculations. Phys. Rev. B 94, 085415 (2016).
H.L. coordinated the development of the code, developed hybrid parallelism, and
215. Deng, T. et al. Ab initio dipolar electron–phonon interactions in two-
performed scaling tests. S.P. co-developed the transport module and maintained the
dimensional materials. Phys. Rev. B 103, 075410 (2021).
216. Sio, W. H. & Giustino, F. Polarons in two-dimensional atomic crystals. Nat. Phys. test farm. K.B. co-developed the indirect absorption module. S.H. developed the full-
19, 629 (2023). band anisotropic Eliashberg module. J.L.B. tested and expanded the polaron module.
217. Verdi, C. & Giustino, F. Fröhlich electron–phonon vertex from first principles. J.L. developed the ionized-impurity scattering module. C.L. co-developed the polaron
Phys. Rev. Lett. 115, 176401 (2015). module. J.M.L. developed the Wannier-function perturbation theory module. F.M. co-
218. Towns, J. et al. XSEDE: accelerating scientific discovery. Comput. Sci. Eng. 16, developed the transport module. H.M. developed the intermediate representation for
62–74 (2014). the superconducting module. H.P. tested and expanded the anisotropic Eliashberg
219. Ottaviani, G., Reggiani, L., Canali, C., Nava, F. & Alberigi-Quaranta, A. Hole drift module. W.S. developed the initial version of the polaron module. S.T. developed the
velocity in silicon. Phys. Rev. B 12, 3318–3329 (1975). quasi-degenerate perturbation theory module. M.Z. developed the special displace-
220. Kawamura, M. Fermisurfer: Fermi-surface viewer providing multiple repre- ment module. X.Z. co-developed the indirect absorption module. N.B. supervised the
sentation schemes. Comput. Phys. Commun. 239, 197–203 (2019). development of the transport module. E.K. supervised the development of the
221. Spicer, W. E. & Eden, R. C. Photoemission investigation of the band structure of indirect absorption module. E.R.M. supervised the development of the super-
semiconductors. In Proc. Ninth International Conference on the Physics of Semi- conductivity module. F.G. supervised the project. All authors contributed to the
conductors, Vol. 1 (ed. Ryvkin, S. M.) p. 65 (Nauka Publishers, Moscow, 1968).
writing of the manuscript.

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2023) 156
 H. Lee et al.
26
COMPETING INTERESTS Open Access This article is licensed under a Creative Commons
The authors declare no competing interests. Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative
ADDITIONAL INFORMATION Commons license, and indicate if changes were made. The images or other third party
material in this article are included in the article’s Creative Commons license, unless
Correspondence and requests for materials should be addressed to Feliciano
indicated otherwise in a credit line to the material. If material is not included in the
Giustino.
article’s Creative Commons license and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly
Reprints and permission information is available at http://www.nature.com/
from the copyright holder. To view a copy of this license, visit http://
reprints
creativecommons.org/licenses/by/4.0/.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims
in published maps and institutional affiliations. © The Author(s) 2023

npj Computational Materials (2023) 156 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences