Abstract
EPW is an opensource 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 temperaturedependent properties and phononassisted quantum processes in bulk solids and lowdimensional 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 phononmediated superconductors using the anisotropic multiband Eliashberg theory; an optics module for calculations of phononassisted indirect transitions; a module for the calculation of small and large polarons without supercells; and a module for calculating band structure renormalization and temperaturedependent optical spectra using the special displacement method. For each capability, we outline the methodology and implementation and provide example calculations.
Similar content being viewed by others
Introduction
The coupling between electrons and phonons is one of the most intensively studied fermion–boson interactions in condensed matter physics. It is responsible for a number of physical phenomena in solids such as conventional superconductivity, temperaturedependent resistivity in metals and mobility in semiconductors, the formation of polarons, and phononassisted optical processes, to name a few^{1,2}. Predictive nonempirical calculations of electron–phonon interactions play an essential role in elucidating a variety of materials properties and their temperature dependence.
During the past three decades, firstprinciples calculations of electron–phonon interactions have been made possible by the development of density functional theory (DFT)^{3,4} and densityfunctional perturbation theory (DFPT)^{5,6,7}. However, the computational cost of these calculations is high, as they involve the evaluation of Brillouin zone integrals which typically require a very fine sampling of the crystal momenta of electrons and phonons. For instance, in DFPT calculations, every single phonon wavevector and vibrational mode requires the solution of Sternheimertype equations that carry a computational complexity comparable to a DFT total energy calculation^{8}. As a result, calculating properties related to the electron–phonon interaction usually entails a computational workload that is orders of magnitude more costly than standard DFT calculations, thus rendering direct ab initio calculations of these quantities challenging.
Several methods have been proposed to address this challenge^{1}. Among those, one of the authors proposed to combine DFPT with maximally localized Wannier functions (MLWFs)^{9} to perform a physicsbased interpolation via a generalized Fourier transformation^{10}. This method exploits the spatial localization of the electronphonon matrix elements in the Wannier representation and enables efficient calculations of electron–phonon matrix elements on ultradense momentum grids while retaining the accuracy of DFPT.
The EPW code builds upon this methodology to compute a number of properties relating to electron–phonon interactions and temperaturedependent materials properties. EPW is the first opensource ab initio software devoted to electron–phonon interactions and has been actively developed for over 16 years. It was publicly released in 2010 under a GNU General Public License (GPL)^{11} and is being distributed within the Quantum ESPRESSO materials simulation suite^{12} since 2016. In addition to EPW, several software projects have been developed in recent years to address electron–phonon interactions, including: Perturbo^{13}, elphbolt^{14}, Phoebe^{15}, and EPIq^{16}. All these packages rely on the same interpolation methodology employed by EPW^{10}.
Two prior manuscripts^{17,18} describe the status of the EPW code until 2016. The present manuscript aims to describe progress on the code that has occurred since 2016 as well as ongoing efforts, leading to the latest EPW release. Added functionalities involve: a magnetotransport module for the calculation of the conductivity in metals as well as the drift and Hall mobility in semiconductors; a module for the calculation of small and large polarons without using supercells; a module for the solution of the fullband Eliashberg equations for superconductors; a module for calculations of phononassisted indirect optical processes; and a module for calculations of finitetemperature electronic and optical properties via the special displacement method. The code has been refactored to keep up with contemporary developments in highperformance computing (HPC) architectures, for example via hybrid twolevel message passing interface (MPI^{19}) and Open MultiProcessing (OpenMP^{20}) parallelization, and parallel I/O via parallel Hierarchical Data Format 5 (HDF5^{21}).
The manuscript is organized as follows. In the ‘Methods’ section (located at the end of the manuscript), we describe the notation employed throughout this manuscript, we review the conceptual basis of electron–phonon interpolation using DFPT and MLWFs, and we discuss the extension of this methodology to the case of polar materials with longrange Coulomb interactions. In the ‘Results’ section, we describe the additional or expanded functionalities available in the latest release of the EPW code. In particular, we begin by discussing calculations of carrier transport within the ab initio Boltzmann transport equation. We consider both electric and magnetic fields, as well as phononlimited and charged defectlimited transport. Then we move to calculations of the superconducting critical temperature and superconducting gap function. Here, we discuss elementary calculations based on the semiempirical McMillan equation, the isotropic Eliashberg theory, and the fullyanisotropic fullband Eliashberg theory. Next, we outline the methodology for computing small and large polarons from DFPT without resorting to large supercells. We then proceed with the formalism for the calculation of optical absorption spectra including phononassisted indirect transitions. After these applications which are based on Wannier–Fourier interpolation, we describe the special displacement method, which offers an alternative and complementary strategy to computing electron–phonon properties without employing Wannier–Fourier interpolation. At the end of this section, we discuss recent advances in computational algorithms, parallelization, and I/O, and we report on parallel scaling benchmarks. In the ‘Discussion’ section we discuss possible future directions for the EPW software project.
Results
In this section, we provide an overview of additional or expanded capabilities of the EPW code. In particular, we discuss how EPW calculates carrier transport properties via the ab initio Boltzmann transport equation (aiBTE), including both carrierphonon and carrierimpurity scattering; how the superconducting critical temperature and superconducting gap are computed via the solution of the Eliashberg equations; how we solve the ab initio polaron equations to investigate small and large polarons without using large supercells; and how we perform calculations of optical absorption spectra in indirectgap semiconductors by including phononassisted optical transitions. We also describe an alternative to Wannier interpolation to study electron–phonon interactions, based on the special displacement method. Of these features, the superconducting module has been enhanced and expanded with respect to the previous release in 2016^{18}; all the other modules described in the following have been developed ex novo since the 2016 release. The notations and conventions employed below are described in detail in the ‘Methods’ section. In the same section, we provide an overview of the Wannier–Fourier interpolation of electron–phonon matrix elements.
Phononlimited carrier transport using the ab initio Boltzmann transport equation
Background and formalism
The calculation of the electronic transport properties of metals and semiconductors is conveniently dealt with by the ab initio Boltzmann transport equation (aiBTE). The Boltzmann equation describes the nonequilibrium distribution function of electrons and holes in the presence of external electric or magnetic fields^{2}. It carries strong predictive power as recently demonstrated for many common semiconductors^{22}. Although the Boltzmann formalism is usually derived within the semiclassical approximation, the theory can rigorously be derived from a nonequilibrium manybody Green’s function formalism and is understood as the quasiparticle approximation to the Kadanoff–Baym theory^{23,24,25,26}.
The EPW code implements the linearized Boltzmann transport equation, which describes the firstorder response of the distribution function to external fields and is suitable for investigating the conductivity of metals and the lowfield mobility of semiconductors. In this section, we focus on the case of phononlimited transport properties. The extension to include charged defects is discussed in the next section.
We consider homogeneous extended solids held at uniform temperature, and carrier scattering by electron–phonon interactions only, for now. Within the aiBTE, the linear response \({\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}\) of the carrier distribution function to an external electric field E is given by^{22,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41}
In this expression, f_{nk} is the electron distribution function, n denotes a Kohn–Sham band index, and k is the electron wavevector. \({\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}\) is a shorthand notation for (∂f_{nk}/∂E_{β})∣_{E=0}, v_{nkα} = ℏ^{−1}∂ε_{nk}/∂k_{α} is the intraband velocity matrix element for the Kohn–Sham eigenvalue ε_{nk}, and δ denotes the Dirac delta function. g_{mnν}(k, q) is the electron–phonon matrix element describing the scattering from the state nk to the state mk + q via the phonon of branch ν and wavevector q (see the “Methods” 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 \({f}_{n{{{\bf{k}}}}}^{0}\) and n_{qν}, respectively. The quantity τ_{nk} in Eq. (1) is the carrier relaxation time, and is obtained from Fermi’s golden rule:
The electrical conductivity tensor σ_{αβ}, which corresponds to the variation of the current density with respect to the electric field, is computed from \({\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}\) using:
and the drift mobility is obtained from this expression upon dividing by the carrier density n_{c}:
These equations are valid for metals and for electrons or holes in semiconductors taken separately. A common approximation to the aiBTE is the relaxation time approximation, which consists of neglecting the second term on the righthand side of Eq. (1). In this case the mobility takes the simpler form:
We refer to this approximation as the “selfenergy relaxation time approximation” (SERTA) since the transport lifetime in Eq. (2) corresponds to the quasiparticle lifetime, which is proportional to the imaginary part of the electronphonon selfenergy^{1,22,24}.
The electrical conductivity and the drift mobility described by Eq. (1) can be measured by timeofflight measurements or by THz photoconductivity measurements. In the case of Hall and De Haas–Van Alphen measurements, an additional magnetic field B is applied, and the resulting Lorentz force must be taken into account in the aiBTE. In these cases, EPW solves the following aiBTE equation^{22,24,32,42}:
The variation \({\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}({{{\bf{B}}}})\) is computed for magnetic fields sufficiently small that their effect on the electronic and vibrational properties can be ignored; accordingly, the Kohn–Sham energies, phonons, and their couplings in the above equation are all evaluated for B = 0. The application of a magnetic field in the direction \(\hat{{{{\bf{B}}}}}\) results in an orthogonal flow of the charge carriers which can be described by the linear response of the mobility to the field:
From this expression, we obtain the Hall mobility in terms of the drift mobility as follows:
where we have introduced the dimensionless Hall factor as^{22}
Also in this case, one can simplify the solution of Eq. (6) by ignoring the second term on the righthand side. This choice leads to the SERTA approximation in the presence of a magnetic field.
Computational considerations
EPW implements an iterative solver to obtain \({\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}\) from Eq. (1), with the possibility of Broyden mixing^{43} to accelerate the convergence. The accurate evaluation of Eqs. (3) and (4) is computationally demanding because it requires the knowledge of \({\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}\) for a dense set of kpoints in an energy window of the order of a few tens of meV around the Fermi energy. In particular, since the scattering integral on the righthand side of Eq. (1) couples linear response coefficients at k and k + q points, the kgrid and the qgrid must be commensurate. As this term is not evaluated in the SERTA approximation, SERTA calculations can instead be performed using incommensurate grids, including for example random and quasirandom sampling^{31}. It should be noted, however, that the SERTA approximation tends to underestimate transport coefficients by up to 50% as compared to the aiBTE^{22}.
Regardless of the approximation chosen to calculate transport coefficients, EPW exploits crystal symmetry operations to eliminate symmetryequivalent wavevectors in the solution of Eq. (1); accordingly, \({\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}\) is evaluated within the irreducible wedge of the Brillouin zone. Furthermore, the solution of Eq. (1) is restricted to the set of wavevectors for which both the initial and the final electronic state lie within a userdefined energy window around reference energy. In particular, the wavevectors k and q are retained only if there exists a pair of bands m and n such that ∣ε_{nk}−E^{ref}∣ < Δ and ∣ε_{mk+q}−E^{ref}∣ < Δ, where E^{ref} is reference energy and Δ is the width of the window. In the case of metals, the reference energy is set to the Fermi energy. In the case of semiconductors, the Fermi energy is calculated from the userspecified temperature and carrier concentration, using the bisection method.
Magnetotransport calculations using Eq. (6) are harder to converge than calculations without magnetic fields. The iterative solution is initialized by using the result of Eq. (1) as a first approximation for \({\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}({{{\bf{B}}}})\), and the quantity \({\nabla }_{{{{\bf{k}}}}}{\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}({{{\bf{B}}}})\) appearing on the lefthand side of Eq. (6) is evaluated via finite differences. The correctness of the solution is tested by checking for the conservation of the carrier density, which corresponds to the condition \({\sum }_{n{{{\bf{k}}}}}{\partial }_{{E}_{\beta }}{f}_{n{{{\bf{k}}}}}({{{\bf{B}}}})=0\).
To ensure numerically accurate evaluation of transport coefficients with EPW, it is important to verify that results be converged with respect to (i) the density of the Brillouin zone grids employed in Eqs. (1) and (6); (ii) the density of the coarse grids used for the Wannier–Fourier interpolation of the DFT and DFPT data^{22}; (iii) the Gaussian smearing employed to compute the Dirac delta functions appearing in Eq. (1); and (iv) to check for the sensitivity of results to lattice parameters, exchange and correlation functionals, and pseudopotentials. To facilitate the convergence with respect to the Gaussian smearing and the kpoint sampling, EPW offers the possibility of using the adaptive smearing method of ref. ^{44}.
Application example
To demonstrate the implementation of the transport module in EPW, we investigate the electron and hole mobilities of cubic boron nitride, as well as the associated Hall factors. We use the relativistic Perdew–Burke–Ernzerhof (PBE) parametrization^{45} of the generalized gradient approximation to DFT. The pseudopotentials are normconserving, generated using the ONCVPSP code^{46}, and optimized via the PseudoDojo initiative^{47}. We consider room temperature and low carrier concentrations of 10^{13} cm^{−3}. cBN is a polar wide gap semiconductor with isotropic Born effective charges \({Z}_{{{{\rm{B}}}}}^{* }=1.91\,e\) and \({Z}_{{{{\rm{N}}}}}^{* }={Z}_{{{{\rm{B}}}}}^{* }\); isotropic dynamical quadrupoles Q_{κ,αβγ} = Q_{κ}∣ϵ_{αβγ}∣ where ϵ_{αβγ} is the Levi–Civita tensor and Q_{κ} an atomdependent scalar with values Q_{B} = 3.46ea_{0} and Q_{N} = −0.63ea_{0}, where a_{0} is the Bohr radius; and an isotropic highfrequency relative dielectric constant ϵ^{∞} = 4.54^{22}. As will be discussed in Fig. 9, longrange electron–phonon interactions are important in the case of cBN. Wannier functions are calculated separately for the valence and the conduction band manifold in order to reduce the computational cost.
Figure 1a shows the convergence of calculated mobilities with the coarse grid of kpoints used in the Wannierization procedure. For this test, we employ a coarse qpoint grid with half the number of the kpoints; and we employ identical fine grids with 50^{3}k and qpoints for the solution of Eq. (1); we use an energy window Δ = 0.3 eV around the reference energy E^{ref} set to the band edge, and we employ adaptive smearing. This panel shows that the mobilities are converged with an accuracy of 1% when coarse grids with 18^{3} and 14^{3} points are employed for electrons and holes, respectively. Figure 1b shows the convergence of calculated mobilities with respect to the fine grids. We see that both electron and hole mobilities are almost converged for grids consisting of 250^{3} points. Generally speaking, the fine grid required to achieve converged results depends on the effective masses, with smaller masses requiring denser grids. As an example, accurate calculations of the electron mobility of GaAs require homogeneous grids of order 400^{3} points, while in the case of AlAs or GaP 80^{3} points are sufficient^{22} (see Table 1).
In Fig. 1c we show convergence curves for the mobility evaluated in the SERTA approximation. In this case, the convergence with the density of points in the fine grids is much slower than for complete aiBTE calculations. More importantly, SERTA mobilities can differ by up to a factor of two from the corresponding aiBTE results. The large difference between full aiBTE calculations and SERTA seems to be a common trend in polar materials^{22}. For the same k and qpoint grids, the computational saving afforded by the SERTA approximation is minimal; therefore, we recommend using SERTA only when the use of commensurate grids is too demanding, for example in the case of systems with many atoms in the unit cell. Figure 1d, e show the computed Hall factors for electrons and holes in cBN, respectively. The Hall factor in the SERTA approximation appears to be close to the full aiBTE result, which is expected since it is defined as the ratio of two mobilities (see Eq. (9))^{48}.
In Fig. 1(f) we analyze the role of the Gaussian smearing parameter in the calculations. To this end, we compute the electron mobility of cBN as a function of grid size, for varying smearing parameter. It is seen that the size of the grid necessary for convergence increases when decreasing the smearing in the few meV’s range. On the other hand, the use of adaptive smearing affords fast convergence even for smaller grid sizes. Based on this comparison, adaptive smearing is the computationally most convenient strategy.
Defectlimited carrier transport using the ab initio Boltzmann transport equation
Background and formalism
Electron–phonon interactions are the dominant scattering mechanism in highpurity single crystals with low defect concentration, typically up to 10^{15} cm^{−3} ^{49}. In tetrahedral semiconductors and a host of other materials, charged defects are ubiquitous since dopant elements are employed to introduce free carriers^{50}. Donors release electrons to the conduction band, thus becoming positively charged defects; similarly, acceptors release holes into the valence band, and become negatively charged defects. In either case, the impurity generates a longranged Coulomb potential that scatters charge carriers. This scattering mechanism tends to dominate over electron–phonon processes at high doping concentrations.
There exist popular semiempirical relations to estimate the effect of ionized impurity scattering on carrier transport in semiconductors, such as the Brooks–Herring^{51} and Conwell–Weisskopf^{52} formulas. However, these expressions rely on simplified parabolic band models and do not carry predictive power in the case of materials with multiband or multivalley band extrema. Firstprinciples calculations offer a modern alternative to these earlier approaches, and have achieved considerable success in recent years^{53,54,55,56}. The EPW code implements a module for charged defect scattering based on a randomized distribution of point charges. This approach is described in detail in ref. ^{57} and summarized below.
For convenience we rewrite the aiBTE from Eq. (1) in the more compact form:
where the quantity \({\tau }_{n{{{\bf{k}}}}\to m{{{\bf{k}}}}+{{{\bf{q}}}}}^{1}\) is the partial scattering rate from the Kohn–Sham state nk to the state mk + q. In the case of electron–phonon scattering, \({\tau }_{n{{{\bf{k}}}}\to m{{{\bf{k}}}}+{{{\bf{q}}}}}^{1}\) is obtained from Eq. (2) by removing the summation over m and the integral over the Brillouin zone, i.e. \(\mathop{\sum}\nolimits_{m}\int d{\bf{q}}/\Omega_{\rm{BZ}}\). When both electron–phonon scattering and charged impurity scattering are taken into account, the partial scattering rate is written as the sum of the individual partial rates
where the superscripts refer to carrierphonon (ph) and carrierimpurity (imp) scattering, respectively. In EPW, the carrierimpurity partial scattering rate \(1/{\tau }_{n{{{\bf{k}}}}\to m{{{\bf{k}}}}+{{{\bf{q}}}}}^{\,{{\rm{imp}}}\,}\) is calculated under the following simplifying approximations: (i) each impurity is described by an idealized point charge, embedded in the dielectric continuum of the host material; (ii) the scattering rate from each impurity is evaluated within the first Born approximation; (iii) defects are sufficiently diluted that the scattering rates from different impurities are additive; (iv) impurities are randomly distributed, and this random distribution is formally taken into account by using the Kohn and Luttinger ensemble average^{58}. Within these approximations, the charged impurity scattering rate reads^{57}:
where N_{imp} is the number of impurities per crystal unit cell (dimensionless), Ze is the charge of each impurity, and \({{{{\boldsymbol{\epsilon }}}}}_{}^{0}\) is the static relative dielectric constant tensor. The scattering rate in Eq. (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 of charged defects in supercells is that it can be used systematically without requiring detailed knowledge of the defect physics and energetics in each material.
Computational considerations
In the EPW code, the scattering rate given in Eq. (12) is added to the electron–phonon scattering rate using Eqs. (11) and (10), and the aiBTE is solved as already described for the phononlimited mobility. The computational overhead as compared to phonononly calculations is minimal.
In Eq. (12), the sum over Gvectors is handled via a Gaussian filter, as described in the “Methods” section; the reader is referred to that section for more details. The overlap integrals between the initial and final Kohn–Sham states appearing in Eq. (12), 〈ψ_{mk+q}∣e^{i(q+G)⋅r}∣ψ_{nk}〉, are evaluated by means of the unitary Wannier function matrices described in the Methods.
The integral over the scattering wavevectors q in Eq. (10) of the impurity scattering rates given in Eq. (12) contains a singular ∣q∣^{−4} term that is not integrable. This issue is resolved by introducing the screening of the defect potential by free carriers released upon ionization. To this end, we replace \({{{{\boldsymbol{\epsilon }}}}}_{}^{0}\) in Eq. (12) with the total dielectric function:
where \({{{{\boldsymbol{I}}}}}_{}^{}\) denotes the identity matrix, and q_{TF} is the Thomas–Fermi wavenumber obtained from the longwavelength limit of the Lindhard function^{49,56,59}:
Temperature enters this expression via the equilibrium Fermi–Dirac distribution of the electrons or holes, \({f}_{n{{{\bf{k}}}}}^{0}\).
The concentration of charged defects is an external input parameter in these calculations. In the case of ionized impurities in semiconductors, it is also possible to use a simple estimate for the fraction f_{ii} of ionized impurities at a given temperature by using the impurity energy level in the gap ε_{d}^{49,56,60}.
where f is the Fermi–Dirac distribution and n_{d} is the total concentration of defects that may thermally ionize.
Owing to the ∣q∣^{−4} scaling of the ionizedimpurity scattering matrix elements, achieving convergence in the presence of impurities is slightly more demanding than for calculations of phononlimited mobility. For example, in recent calculations of carrier mobilities in Si, GaP, and SiC, fine grids with 100^{3}, 100^{3}, and 180^{3} points were necessary, respectively^{57} (see Table 1).
Application example
To illustrate the impurity scattering capability of EPW, we calculate the mobility of electrons and holes in silicon as a function of both temperature and dopant concentration. We consider coarse Brillouin zone grids with 12^{3}kpoints and 6^{3}qpoints, and fine grids with 100^{3} points. We use the PBE functional^{45} and ONCV pseudopotentials^{46,61}, and we include spin–orbit coupling for the valence bands. We account for quadrupole corrections using the procedure described in the Methods section.
Figure 2a shows the calculated electron mobility of silicon as a function of temperature. When considering phonon scattering only, our calculations are in very good with measurements on highpurity silicon^{62}. Upon introducing ionized impurity scattering, the electron mobility at 100 K decreases from 11,813 to 4725 cm^{2} V^{−1} s^{−1} for a dopant concentration of 1.75 × 10^{16} cm^{−3}, and to 1769 cm^{2} V^{−1} s^{−1} for a dopant concentration of 1.3 × 10^{17} cm^{−3}. These calculations agree well with experimental data^{63}. At higher temperatures, the reduction in the mobility due to impurity scattering is less significant since phonons provide the dominant scattering mechanism. The hole mobility in Fig. 2b exhibits a similar trend. Upon introducing ionized impurity scattering, the hole mobility at 100 K decreases from 8877 to 2884 cm^{2} V^{−1} s^{−1} for a dopant concentration of 2.4 × 10^{16} cm^{−3}, and to 1056 cm^{2} V^{−1} s^{−1} for a dopant concentration of 2.0 × 10^{17} cm^{−3}. Also in this case, the agreement with experimental data is very good^{63}.
Figure 2c shows the calculated electron mobility of silicon at 300 K, as a function of ionized impurity concentration. Up to a dopant concentration of 10^{16} cm^{−3}, phonon scattering dominates and the mobility is relatively insensitive to impurity scattering. Beyond this concentration, the mobility decreases sharply and approximately as the inverse of the impurity density, in line with Eq. (12). The calculations agree well with experiments up to an impurity concentration around 10^{18} cm^{−3}^{64}. Beyond this concentration, it is expected that additional mechanisms such as twoimpurity scattering and plasmon scattering^{65} will further reduce the mobility. The hole mobility, which is shown in Fig. 2d, follows a similar trend. The calculated hole mobility at low doping, 603 cm^{2} V^{−1} s^{−1}, slightly overestimates the experimental range 450–500 cm^{2} V^{−1} s^{−1}; this effect can be traced to the underestimation of the heavy hole mass by DFT^{31}. Upon increasing the impurity concentration, the mobility decreases following the same trend as for the electrons. These calculations agree well with experiments^{64,66} and with previous firstprinciples calculations^{56}.
Phononmediated superconductivity using the ab initio Eliashberg theory
Background and formalism
Firstprinciples calculations of phononmediated superconductors are primarily based on three approaches, namely semiempirical methods based on the McMillan formula^{67}, the ab initio Eliashberg theory^{68,69}, and the densityfunctional theory for superconductors^{70,71,72,73}. The EPW code implements the former two approaches. In this section, we briefly review methods based on the McMillan formula, and then we describe the Eliashberg formalism and its implementation.
The superconducting critical temperature T_{c} can be estimated using the semiempirical McMillan equation in Allen–Dynes form^{74}:
where k_{B} is Boltzmann’s constant, μ^{*} is the semiempirical Coulomb pseudopotential, \({\omega }_{\log }\) is the logarithmic average of the phonon frequencies, and λ is the electron–phonon coupling constant, as defined in ref. ^{1}. The parameter μ^{*} is largely material independent, for a large class of superconductors a value in the 0.10.15 range being typically required to fit the experimental T_{c}^{74,75,76}. Note that for very strong coupling (λ > 1.5), in ref. ^{74} two correction prefactors f_{1} and f_{2} are introduced in Eq. (16) to better estimate the critical temperature in this regime.
In addition, the code allows the user to calculate the critical temperature using a more recent prescription based on a machine learning approach^{77}:
where the correction factors f_{ω} and f_{μ} are given by
In these expressions, \({\bar{\omega }}_{2}\) is the square root of the second moment of the normalized weight function g(ω) = 2α^{2}F(ω)/(ωλ), with α^{2}F being the Eliashberg spectral function, as defined in ref. ^{1}. Equations (16) and (17) are useful for preliminary calculations but do not carry the predictive power of the ab initio Eliashberg theory.
The Eliashberg theory^{68,69} describes the superconducting phase transition by means of finitetemperature Green’s functions. In this theory, superconducting pairing arises from an attractive electron–electron interaction mediated by phonons, which is partly compensated by the interelectron Coulomb repulsion. The Eliashberg theory can be formulated as a Dyson equation for a generalized 2 × 2 matrix Green’s function via the NambuGor’kov formalism^{78,79}. The offdiagonal elements of this matrix describe Cooperpair amplitudes in the superconducting state and are related to the superconducting gap function. These elements become nonzero below the critical temperature, marking the transition to the superconducting state. It is standard practice to expand the pairing selfenergy using Pauli matrices \({\hat{\tau }}_{i}\) (i = 0, …, 3) as follows^{75,80,81,82,83,84,85,86}:
where iω_{j} = i(2j + 1)πT is the fermionic Matsubara frequency with j being an integer, T is the absolute temperature, Z_{nk}(iω_{j}) is the mass renormalization function, χ_{nk}(iω_{j}) is the energy shift, and ϕ_{nk}(iω_{j}) is the order parameter. This selfenergy is expressed in terms of the electron Green’s function using the Migdal approximation^{87} for the electron–phonon contribution, and the GW approximation for the electron–electron contribution^{88,89}. Using Eq. (19) inside the Dyson equation for the electron Green’s function yields a set of coupled equations for Z_{nk}, χ_{nk}, and ϕ_{nk}:
having defined the auxiliary function:
In Eqs. (20)–(23), N(ε_{F}) is the density of states (DOS) per spin at the Fermi level, μ_{F} is the chemical potential, and the quantities V_{nk,mk+q} denote the matrix elements of the screened Coulomb interaction W between electron pairs, as given in refs. ^{90,91}. The anisotropic electron–phonon coupling parameters \(\lambda (n{{{\bf{k}}}},m{{{\bf{k}}}}+{{{\bf{q}}}},{\omega }_{j}{\omega }_{{j}^{{\prime} }})\) appearing in these equations are calculated as
Equations (20)–(23) are supplemented by a statement of particle number conservation, which determines the chemical potential^{84}:
where N_{e} is the number of electrons per unit cell. Equations (20)–(25) are referred to as the anisotropic fullbandwidth (FBW) Eliashberg equations^{92} since they explicitly take into account scattering processes involving electrons with energies and momenta that are not restricted to the vicinity of the Fermi surface. To find the temperaturedependent superconducting gap, in EPW these equations are solved iteratively for different temperatures; the highest temperature for which a nontrivial solution exists (ϕ_{nk} ≠ 0) is the superconducting critical temperature T_{c}. Below this temperature, the superconducting gap Δ_{nk} is given by
The numerical solution of Eqs. (20)–(25) is computationally demanding. A common simplification of these equations consists of restricting the energy range close to the Fermi level^{75,80,81,82,83,84,85,86,93}. In this approach, it is assumed that the DOS within this energy window is constant. It can be shown that, within these approximations, the energy shift χ_{nk} vanishes and the requirement in Eq. (25) is automatically satisfied. As a result, only two equations for Z_{nk} and ϕ_{nk} need to be solved selfconsistently:
These equations are referred to as the anisotropic Fermi surface restricted (FSR) Eliashberg equations^{18,85}.
To extract physical quantities of interest, such as the tunneling density of states and the heat capacity, the gap function on the real frequency axis is required^{75,81,82}. In the EPW code, the continuation of Δ_{nk}(iω_{j}) from the imaginary to the real axis is performed either using Padé approximants^{94,95}, or using the exact analytic continuation of ref. ^{96}, as described in ref. ^{85}.
Accounting for the Coulomb repulsion in the Eliashberg equations requires the evaluation of the matrix elements V_{nk,mk+q}. While this is feasible in principle, it is common practice to replace the product N(ε_{F})V_{nk,mk+q} with the semiempirical Morel–Anderson pseudopotential \({\mu }_{{{{\rm{c}}}}}^{* }\)^{97}. In this approach, the value of \({\mu }_{{{{\rm{c}}}}}^{* }\) is linked to the cutoff frequency ω_{c} for the Matsubara summation on the imaginary axis, with ω_{c} typically set to ten times the maximum phonon frequency ω_{ph}. Note that when converting between the μ^{*} in Allen–Dynes formula in Eq. (16) and \({\mu }_{{{{\rm{c}}}}}^{* }\), a ratio ω_{c}/ω_{ph} of ten is often used^{74}. In EPW, \({\mu }_{{{{\rm{c}}}}}^{* }\) is specified by the user as an external parameter, and in many applications, values in the range \({\mu }_{{{{\rm{c}}}}}^{* }=0.1\)0.2 yield reasonable agreement with experiments. More recently, firstprinciples calculations of \({\mu }_{{{{\rm{c}}}}}^{* }\) as a Fermi surface average of V_{nk,mk+q} have been used in conjunction with EPW^{91,98,99,100}.
In the case of simple superconductors which do not exhibit significant anisotropy, it may be a good approximation to neglect the band and momentum dependence of the superconducting gap. In these cases, instead of Eqs. (20)–(25), it is possible to solve a simplified version of the equations where all properties are averaged over the Fermi surface^{76,101,102,103}. The relations are referred to as the isotropic Eliashberg equations:
where the counterpart of Eq. (23) is
In Eqs. (29)–(31), \(\lambda ({\omega }_{j}{\omega }_{{j}^{{\prime} }})\) is a momentumaveraged version of Eq. (24):
Similar to the anisotropic case, Eqs. (29)–(31) can also be reduced to a set of two equations for Z(iω_{j}) and ϕ(iω_{j}), under the same conditions as discussed for Eqs. (27) and (28).
Computational considerations
Solving Eqs. (20)–(25) or the simplified version Eqs. (27) and (28) is computationally demanding because a fine sampling of electronphonon processes near the Fermi surface is required. In addition, due to the implicit form of these equations, the k and qpoint grids must be commensurate. To accelerate calculations, EPW exploits crystal symmetries so that the gap function and related quantities are only computed for kpoints in the irreducible wedge of the Brillouin zone, and only electronic states within a small energy window centered at the Fermi energy are considered. Numerical tests show that converged results are obtained when this energy window is of the order of a few times the maximum phonon energy. In general, grids with an average point density Δk ~ 0.02 Å^{−1} and Δq ~ 0.04 Å^{−1} are sufficient to converge the superconducting gap^{85,91,104}, but denser meshes may be required for superconductors with complex Fermi surfaces. As an example, 120^{2} and 40^{3}kpoint grids and 60^{2} and 20^{3}qpoint grids were used to describe the twogap structure of bilayer C_{6}CaC_{6} and the singlegap structure of bulk CaC_{6}, respectively^{91} (see Table 1).
One further complication in the solution of the Eliashberg equations is that all quantities depend on Matsubara frequencies. The Matsubara frequencies are proportional to the absolute temperature, therefore superconductors with low T_{c} require a larger number of frequencies within the same energy range. While it is common to set the Matsubara frequency cutoff to ten times the maximum phonon frequency^{85,86,101,105,106}, we recommend always performing convergence tests with respect to this parameter. In EPW, the number of Matsubara frequencies can be reduced using a sparse sampling scheme whereby only a subset of frequencies is retained with a density that decreases logarithmically with the Matsubara index. The grid density is controlled by an adjustable parameter; using the default setting for this parameter, EPW yields a ~30% reduction of Matsubara frequencies, while keeping all the lowest ~40% of frequencies. Numerical tests show that this approach preserves the accuracy of more expensive fullgrid calculations.
FBW Eliashberg calculations are computationally more demanding than FSR calculations. In particular, to determine the chemical potential from Eq. (25) one needs a Matsubara frequency cutoff of at least a few times the Kohn–Sham energy window, leading to a considerable increase in computational cost. To circumvent this issue, the default behavior of EPW is to keep the chemical potential fixed at the Fermi level unless it is instructed by the user to solve Eq. (25). In this case, careful convergence tests with respect to the Matsubara frequency cutoff are warranted.
Application example
To demonstrate the implementation of the superconducting module in EPW, we investigate the superconducting properties of hexagonal niobium diselenide (2HNbSe_{2}). 2HNbSe_{2} exhibits a superconducting phase transition and a charge density wave instability below 7.2 and 33 K, respectively; both phases are driven by a strong momentumdependent electronphonon coupling^{107,108,109,110,111,112}. For brevity, in this section, we focus on superconductivity and ignore the lattice instability.
We employ the PBE exchange and correlation functional^{45} and the optB86bvdW van der Waals functional^{113,114,115,116}, as well as ONCV pseudopotentials^{46,61} including semicore electrons for Nb. We use a plane waves kinetic energy cutoff of 80 Ry, Γcentered 24 × 24 × 12 kpoint and 6 × 6 × 4 qpoint coarse grids, and Methfessel–Paxton smearing^{117} of 0.025 Ry. The optimized lattice parameters are a = 3.46 Å and c = 12.57 Å, in good agreement with the experimental data a = 3.43 Å and c = 12.55 Å^{108}.
The Eliashberg equations are solved on a uniform 60 × 60 × 40 kpoint and 30 × 30 × 20 qpoint fine grids. We consider energy 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 \({\mu }_{{{{\rm{c}}}}}^{* }=0.2\). The Dirac deltas in the FSR approach are replaced by Gaussians of width 50 meV.
Figure 3a shows the band structure and DOS of 2HNbSe_{2}. Three bands cross the Fermi level. This system exhibits soft phonons along the ΓM direction^{109,111,112}, which are stabilized by anharmonic effects^{109}. Here, for simplicity, we avoid soft phonons by slightly increasing the electronic smearing to 0.025 Ry in phonon calculations, as shown in Fig. 3b. The same panel also shows the distribution of the electron–phonon coupling strength as described by the Eliashberg spectral function α^{2}F. Figure 3c shows the anisotropic gap function colorcoded on the Fermi surface of 2HNbSe_{2}. We find a highly anisotropic twogap structure: the highgap region corresponds to the Fermi arcs around the K point, and the lowgap region corresponds to the Γcentered pockets. These results are similar to previous calculations for the related compound NbS_{2}^{98}. Figure 3d shows the energy distribution of the superconducting gap function Δ_{nk} for various temperatures. The gap is seen to close gradually with increasing temperature, vanishing at the critical temperature T_{c} = 19 K. The calculated critical temperature overestimates the experimental value of 7.2 K, consistent with prior calculations for bulk and monolayer transition metal dichalcogenides^{98,109,112,118,119,120}. The calculations in Fig. 3d are performed within the FBW Eliashberg approach, with the chemical potential set to the Fermi energy. Calculated gap values for energy windows of 0.4 and 0.8 eV are very similar, therefore in the following we only consider the former. Figure 3e shows the sensitivity of the superconducting gap to the chemical potential: by requiring particle number conservation via Eq. (25), the chemical potential shifts by 25 meV with respect to the Fermi energy, and the critical temperature slightly increases, by <1 K. In Fig. 3f we perform the same test, but this time with a wider energy window, obtaining similar results.
In Fig. 3g we also compare the FBW Eliashberg approach with the FSR approach, which is approximately 30% faster for the system under consideration. The critical temperatures calculated from these approaches agree within 0.5 K. This finding relates to the DOS of 2HNbSe_{2} being slowly varying within a few hundred meV’s around the Fermi energy. Larger differences are expected for systems with rapidly varying DOS, such as for example the highT_{c} hydride superconductors^{105}.
When using a logarithmic Matsubara frequency grid, the results are essentially unaffected as shown in Fig. 3h. At the same time, the sparse Matsubara grids afford a significant computational saving of almost a factor of two.
Polarons and electron selftrapping
Background and formalism
An important manifestation of electronphonon interactions is the formation of polarons. The polaron is a quasiparticle consisting of an electron or a hole dressed by a distortion of the crystal lattice. The lattice distortion can be interpreted as a cloud of virtual phonons accompanying the electron or hole. In the presence of strong electron–phonon coupling, the electron or hole can become trapped in the lattice distortion that it produced; this configuration is referred to as a selftrapped polaron^{121,122,123}.
In the context of firstprinciples calculations, the formation of polarons has generally been investigated by means of DFT calculations of an excess charge added to a large supercell. When the calculation is initialized with a localized distortion of the crystal lattice, it may be energetically favorable for the electron or hole to localize around this distortion, leading to the formation of a polaron^{124,125,126,127,128,129,130}. This approach faces two challenges: first, the existence of polaronic states and their formation energy are very sensitive to the choice of the exchange and correlation functional; second, the size of the supercells required to obtain converged energies and wavefunctions may be prohibitive.
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 displacements associated with the polaron^{131,132,133,134}. The founding principle of this methodology is that the DFT formation energy of the polaron can be expressed as a selfinteractionfree functional of the polaron wavefunction ψ(r) and the atomic displacements in the polaronic state Δτ_{καp}^{132} (all symbols are defined in the ‘Methods’ section):
where \({\hat{H}}_{{{{\rm{KS}}}}}^{0}\) and \({C}_{\kappa \alpha p,{\kappa }^{{\prime} }{\alpha }^{{\prime} }{p}^{{\prime} }}^{0}\) are the Kohn–Sham Hamiltonian and the matrix of interatomic force constants in the groundstate structure without the polaron, and the integrals are over the Born–von Kármán (BvK) supercell. Variational minimization of this energy functional leads to the coupled system of equations:
where \({V}_{{{{\rm{KS}}}}}^{0}\) is the Kohn–Sham potential in the groundstate structure without polaron, and ε represents the quasiparticle excitation energy of the polaron^{133,134}. To avoid performing calculations in large supercells, in EPW Eqs. (36) and (37) are rewritten more conveniently in terms of Kohn–Sham states, phonons, and electron–phonon matrix elements evaluated in the primitive unit cell of the crystal. To this aim, the wavefunction is expanded on the basis of Kohn–Sham states:
and the atomic displacements are expanded on the basis of lattice vibrational eigenmodes:
With these definitions, Eqs. (36) and (37) become a nonlinear system of equations for the polaron coefficients A_{nk} and B_{qν}:
These relations are referred to as the ab initio polaron equations. The polaron formation energy in Eq. (35) can be expressed in terms of the polaron coefficients A_{nk} and B_{qν} as^{132}
where ε_{CBM} is the Kohn–Sham eigenvalue of the conduction band minimum. This expression holds for electron polarons; in the case of hole polarons, ε_{CBM} is replaced by the eigenvalue of the valence band maximum, and the first term on the righthand side of Eq. (42) acquires a minus sign. Equations (38), (39), and (42) provide the realspace wavefunction of the polaron, the accompanying atomic displacements, and its formation energy.
Computational considerations
The solution of Eqs. (40) and (41) require the knowledge of the KohnSham energies ε_{nk}, the phonon frequencies ω_{qν}, and the electronphonon matrix elements g_{mnν}(k, q). All of these quantities are calculated by EPW using Wannier–Fourier interpolation as described in the Methods. The solution of the coupled system of equations is performed using an iterative procedure as follows. First, A_{nk} is initialized in order to compute B_{qν} via Eq. (41). Then, Eq. (40) is solved by constructing and diagonalizing the effective Hamiltonian
The set of solution coefficients A_{nk} is used again in Eq. (41), and the process is repeated until convergence is achieved.
The initialization of the coefficients A_{nk} is achieved by using a Gaussian function in reciprocal space centered around the band extremum. Alternatively, a polaron solution from a previous calculation can be used. In the construction of the effective Hamiltonian matrix in Eq. (43), we set the gauge condition \({e}_{\kappa \alpha ,\nu }({{{\bf{q}}}})={e}_{\kappa \alpha ,\nu }^{* }({{{\bf{q}}}})\)^{135}. This condition ensures that \({B}_{{{{\bf{q}}}}\nu }={B}_{{{{\bf{q}}}}\nu }^{* }\), so only half of the B_{qν} coefficients need to be evaluated at each iteration. In addition, for parallel execution, we set a global gauge for all interpolated Kohn–Sham wavefunctions and vibrational eigenmodes. The diagonalization of the effective Hamiltonian is performed via the Davidson method^{136} as implemented in Quantum ESPRESSO. The convergence of the iterative procedure is tested by evaluating the atomic displacements in real space via Eq. (39) and comparing these realvalued vectors between successive iterations. In the case of polaron calculations, choosing a fine Brillouin zone grid with N × N × N points implies that one is studying polarons in a periodic BvK supercell consisting of N × N × N unit cells. The formation energy of an isolated polaron is obtained by taking the limit of N → ∞. For large enough N, we observe that the formation energy scales as 1/N, which is consistent with the Makov–Payne asymptotic expansion of the energy of a point charge in a periodic cell^{132,137}. Based on this observation, to obtain the energy of an isolated polaron, we perform a MakovPayne extrapolation to the limit of infinite supercell size. Generally speaking, we find that for small polarons the wavefunction shape is converged already for BvK supercells of the order of 5 × 5 × 5 unit cells, while for large polarons it is necessary to use supercells that are large enough to fully contain the polaron density. While there is no formal upper limit to this size, the largest BvK supercell sizes that we employed are up to 50 × 50 × 50 for the large electron polaron in NaCl (see Table 1).
To visualize the polaron, EPW expresses the wavefunction as a linear combination of maximally localized Wannier functions:
where the coefficients A_{mp} are obtained from the solution vectors A_{nk} via the Wannier–Fourier transformation:
and U_{mnk} is the unitary matrix that generates the smooth Bloch gauge [Eq. (64) of the “Methods” section]. The inversion of this last relation also allows one to interpolate the coefficients A_{nk} throughout the Brillouin zone, which is useful to visualize how specific bands contribute to the polaron wavefunction (Fig. 4).
Application example
To demonstrate the implementation of the polaron module in EPW, we investigate the electron and hole polarons in rocksalt NaCl, a prototypical polar insulator that hosts small hole polarons called V_{K} centers^{138,139}. We perform calculations using the PBE exchange and correlation functional^{45}, ONCV pseudopotentials^{47}, and a planewaves kinetic energy cutoff of 150 Ry. Groundstate and latticedynamical calculations are performed using a 12 × 12 × 12 uniform and unshifted Brillouin zone mesh for both k and qpoints. The Kohn–Sham states, phonons, and electronphonon matrix elements that are needed to solve Eqs. (40) and (41) are generated on Brillouin zone grids with up to 50 × 50 × 50 points by Wannier–Fourier interpolation. To this end, we use three Wannier functions to describe the Cl3p states in the valence band, and one Wannier function to describe a single conduction band formed by the Na3s states. The use of additional valence or conduction bands only brings negligible changes to the polaron formation energy. In order to obtain the polaron formation energies in the limit of infinite supercell size, we solve Eqs. (40) and (41) for N × N × N k and qpoint grids with increasing N, and we perform a linear extrapolation of the energy vs. 1/N curve. Figure 4a shows an isosurface of the calculated wavefunction of the electron polaron in NaCl. The formation energy of this polaron in the infinite supercell limit is 37 meV; with a radius of 147 Å, this polaron spans several tens of crystalline unit cells. The wavefunction is isotropic and can be described as a combination of Na3s orbitals modulated by an approximately Gaussian envelope.
In Fig. 4b we show the atomic displacements associated with the electron polaron of Fig. 4a. For clarity we only show the displacement of the Cl ions. The ions tend to move away from the center of the electron wavefunction, consistent with the fact that the electron polaron tends to repel anions. Figure 4c shows an isosurface of the calculated hole polaron of NaCl. The formation energy in the dilute limit is 820 meV, indicating a strongly bound polaron. Consistent with the large formation energy, we find a very small polaron of radius 6 Å, which is comparable to the lattice parameter of 5.69 Å. This polaron consists primarily of a single 3p orbital centered on Cl and oriented along the [100] direction. The accompanying atomic displacements are shown in the same panel. As expected from the negative charge of the polaron, the largest displacements are found for the Na cations which tend to move away from the polaron center.
In Fig. 4d we overlay the coefficients ∣A_{nk}∣^{2} with the band structure in order to determine which electronic states contribute to the electron (blue) and hole (orange) polaron. In the case of the electron polaron, only electrons near the bottom of the conduction band contribute; the narrow distribution of A_{nk} near k = 0 is consistent with the large spatial extent of the electron polaron in real space. Conversely, the hole polaron draws weight from the entire Brillouin zone, in line with the strong localization in real space. Similarly, in Fig. 4e we show the coefficients ∣B_{qν}∣^{2} on the phonon dispersion relations. The electron polaron is primarily driven by longwavelength longitudinaloptical phonons and longitudinal acoustic phonons, while the hole polaron is driven mostly by shortrange optical modes. These spectral weight distributions suggest that the electron polaron in NaCl should be classified as a Fröhlichtype^{140} polaron, while the hole polaron should be classified as a Holsteintype polaron^{141}.
Phononassisted indirect absorption
Background and formalism
The EPW code has the capability to compute, from first principles, optical absorption spectra in indirect band gap semiconductors by including phononassisted optical transitions within secondorder timedependent perturbation theory. Phononassisted transitions involve two virtual processes, namely the absorption of a photon, and the absorption or emission of a phonon, in either order. In this section, we outline the general formalism to describe these processes^{142,143}.
We consider a linearly polarized electromagnetic wave with vector potential of amplitude A_{0}, frequency ω, and polarization vector e. In the following equations, the longwavelength limit and the corresponding electric dipole approximation are understood. Secondorder timedependent perturbation theory states that the transition rate of an electron from an initial Kohn–Sham state nk to a final state mk + q involving a photon and a phonon qν is^{142}:
In this equation, β = +1 and −1 represent phonon emission and absorption processes, respectively. S_{1,mnν}(k, q) and S_{2,mnνβ}(k, q) are the transition amplitudes for the processes illustrated in Fig. 5a: S_{1,mnν}(k, q) refers to a process whereby the electron absorbs a photon, and then absorbs or emits a phonon; S_{2,mnνβ}(k, q) describes a process whereby phonon absorption/emission takes place, followed by the absorption of a photon. Explicit expressions for these amplitudes are:
where v_{mn} denotes velocity matrix elements between the Kohn–Sham states and the sum extends to all possible occupied and unoccupied states. The energy η is a small parameter to avoid singular denominators. From the transition rates W_{mnν}(k, q; ω), we obtain the imaginary part of the dielectric function by summing over all possible transitions in the Brillouin zone and in the band manifold^{142,143}:
where the factor of two results from the electron spin in the case of spinunpolarized systems; this factor is omitted in the case of calculations including spin–orbit coupling. The quantities P_{mnνβ}(k, q) contain the temperaturedependent Fermi–Dirac and Bose–Einstein distributions as follows:
In Eq. (49), the scalar quantity ϵ(ω) represents the projection of the dielectric tensor along the polarization direction e of the electric field. For easier comparison with experimental measurements, after we obtain the imaginary part of the dielectric function from EPW, we calculate the absorption coefficient using^{142}:
where n(ω) is the real part of the refractive index and c is the speed of light. n(ω) can be calculated from the standard relations between the dielectric function and the refractive index^{144}. In this case, the real part of the dielectric function can be obtained by performing a Kramers–Kronig transformation of Im[ϵ(ω)] and applying a rigid shift to match the value of ϵ(ω = 0) computed from DFPT using Quantum ESPRESSO. Alternatively, n(ω) can be taken from experiments^{143}.
Computational considerations
In the optics module of EPW, the imaginary part of the dielectric function is calculated using Eq. (49). The summation over virtual states in Eqs. (47) and (48) are restricted to the manifold of valence and conduction bands included in the Wannierization procedure. The Dirac delta functions appearing in Eq. (49) are replaced by Gaussian functions or Lorentzian functions with a finite broadening. The small parameter η in Eqs. (47) and (48) are used to avoid singular denominators which arise if, for a given photon energy ℏω, direct transitions are resonant with indirect transitions. These situations are encountered, for example, at the onset of direct transitions, i.e. when the photon energy matches the direct gap. In these situations, the dielectric function and the absorption spectrum become sensitive to the choice of η: too small an η leads to a divergence of the spectrum, and too large an η leads to an excessive broadening. To probe the sensitivity of the spectra to this parameter, the optics module of EPW calculates Im[ϵ(ω)] for a range of broadening parameters between 1 meV and 0.5 eV. Efforts are currently ongoing to eliminate these spurious singularities: we believe that they arise from an intrinsic limitation of secondorder perturbation theory when direct and indirect transitions are in resonance; a more general theory that correctly describes these resonances is under development^{145}.
In Eqs. (47) and (48), the velocity matrix elements are sensitive to the details of the electronic structure. For accurate calculations, it is preferable to include GW quasiparticle corrections to the Kohn–Sham eigenvalues^{89}. These corrections require a renormalization of the velocity matrix elements to preserve the fsum rule of optical transitions^{146,147}. In the EPW code, this renormalization is carried out by evaluating the velocity via finite differences. To this end, we consider quasiparticle energies and Kohn–Sham eigenvalues evaluated on kpoint grids slightly offset along the Cartesian directions. The renormalized velocity matrix elements are given by^{147}
where δq_{α} is a small wavevector along the Cartesian direction α, of magnitude 10^{−3} × 2π/a (a is the lattice parameter).
To achieve convergence in calculations of indirect phononassisted optical absorption spectra, Brillouin zone grids with 40 × 40 × 40 points have been employed in previous work on silicon^{143}, as well as in more recent work on silicon and diamond^{145} (see Table 1). In addition, convergence of the phononassisted optical spectra for cubic boron arsenide^{148} and cubic silicon carbide^{149} was achieved with Brillouin zone grids of 32 × 32 × 32. When freecarrier absorption is considered, denser grids are required for converged spectra. For example, a grid of 48 × 48 × 48 is needed for converged phononassisted freecarrier absorption spectra in silicon^{150}.
Application example
To demonstrate the implementation of the optics module in EPW, we examine phononassisted optical absorption in silicon. Goundstate DFT calculations are performed using the PBE exchange and correlation functional^{45}, ONCV pseudopotentials^{46,61}, a planewaves kinetic energy cutoff of 60 Ry, and a 6 × 6 × 6 Brillouin zone sampling of both the kpoint grid and the qpoint grid. The optimized DFT lattice parameter is a = 5.478 Å, similar to previous studies^{151,152,153}.
We evaluate quasiparticle corrections on a 6 × 6 × 6 kpoint grid within the GW method using the BerkeleyGW code^{154}. The GWcorrected band gap is 1.31 eV, which slightly overestimates the experimental value of 1.12 eV at room temperature^{155}, but agrees well with prior work using similar settings^{89}. The resulting quasiparticle band structure is shown in Fig. 5a. Our calculations agree well with experimental measurements at several critical points^{89,156,157,158}, shown by marks in the figure. In Fig. 5a we also show representative optical transitions leading to the amplitudes in Eqs. (47) and (48). The fundamental gap is indirect, between the top of the valence band at the Γ point and the bottom of the conduction band near the X point. To correctly describe optical transitions near the fundamental gap, it is essential to include phononassisted processes. Figure 5b shows calculated phonon dispersion relations, which agree well with prior calculations by us^{18} and other groups^{159}, as well as experimental data^{160,161} (solid symbols). Figure 5c shows the imaginary part of the dielectric function, as obtained from Eq. (49) using 32 × 32 × 32 fine k and qpoint grids. The temperature is set to 300 K, and the Dirac delta functions in Eq. (49) are approximated via Gaussians of width 50 meV. In this plot, a rigid shift of −0.19 eV is applied in order to match the GW band gap to the measured gap of silicon at room temperature (1.12 eV)^{155}. Figure 5d shows the related optical absorption coefficient from Eq. (51). This calculation requires the real part of the frequencydependent refractive index, for which we used experimentally measured values from ref. ^{162} for simplicity.
Our calculated spectra agree very well with experiments in the photon energy range between the indirect gap and the direct gap at 3.3 eV. Beyond the lineshape, the magnitude also agrees well with the experiment over several orders of magnitude, with the theory underestimating the experimental data by 50% at most. This residual underestimation might be related to the rigid shift of the band gap, which we did not include in the velocity renormalization expressed by Eq. (52), and to the fact that electron–phonon matrix elements are slightly too weak in DFT as a result of the band gap problem^{1,163}.
Special displacement method
Background and formalism
The EPW package contains a set of standalone modules, the ZG toolkit, for calculating finitetemperature properties including quantum zeropoint effects via the special displacement method (SDM)^{164,165,166}. The special displacement method is a supercellbased approach that is distinct from and complementary to the Wannier–Fourier interpolation method employed in all previous sections. The founding principle of this method is that the effects of electron–phonon couplings on the electronic and optical properties of extended solids can be captured by performing calculations for a large supercell where the atoms have been displaced away from their equilibrium crystallographic sites. The displacements are chosen in such a way that the corresponding atomic configuration represents the best singlepoint approximant to the quantum thermal distribution of the atomic coordinates^{165}.
Under the approximations of adiabatic Born–Oppenheimer decoupling and harmonic lattice, the quantum thermal average of an electronic or optical property described by the observable O can be written as^{165}
where q runs over the set of wavevectors in a uniform Brillouin zone grid which excludes timereversal invariant points and timereversal partners. In Eq. (53), z_{qν} denote normal mode coordinates with real part x_{qν} and imaginary part y_{qν}. \({u}_{{{{\bf{q}}}}\nu }^{2}=(\hslash /2{M}_{0}{\omega }_{{{{\bf{q}}}}\nu })(2{n}_{{{{\bf{q}}}}\nu }+1)\) is the meansquare displacement for the oscillator qν with Bose–Einstein occupation n_{qν}; O({τ}) is the property of interest, such as for example the Kohn–Sham eigenvalues, density of states, or optical absorption spectrum, calculated for the set of atomic coordinates {τ_{καp}}. The relation between these coordinates and the normal coordinates z_{qν}, which is required to carry out the integral in Eq. (53), is provided in ref. ^{165}.
In the special displacement method, the configurational average expressed by Eq. (53) is approximated by a single calculation for an optimum configuration:
where \({\tau }_{\kappa \alpha p}^{0}\) represent atomic coordinates in the DFT groundstate at zero temperature, and the optimum ‘ZG’ displacement is given by^{165}
As in Eq. (53), the summation is restricted to qpoints from a uniform Brillouin zone grid which are not timereversal invariant and are not timereversal partners. This partitioning is described in Appendix B of ref. ^{1}. The quantities S_{qν} appearing in Eq. (55) are signs (±1) determined by the ZG module so as to guarantee that the resulting displacements make \(O(\{{\tau }_{\kappa \alpha p}^{0}+{{\Delta }}{\tau }_{\kappa \alpha p}^{{{{\rm{ZG}}}}}\})\) the best possible approximant to Eq. (53). In the thermodynamic limit of a large supercell, the ZG displacements reproduce the exact meansquare anisotropic displacement tensors, which are given by^{167}
In the same limit, a single evaluation of the property \(O(\{{\tau }_{\kappa \alpha p}^{0}+{{\Delta }}{\tau }_{\kappa \alpha p}^{{{{\rm{ZG}}}}}\})\) tends to the exact thermal average in Eq. (53). In the case of nonperiodic systems, such as for example nanocrystals and quantum dots, these equations are replaced by their Γpoint only versions^{164,168,169}.
The special displacement method is similar in spirit to computing thermodynamic averages using pathintegral molecular dynamics^{170}, but it differs insofar as a single calculation is required to evaluate the average instead of many molecular dynamics snapshots. In this method, electron–phonon couplings are included nonperturbatively through the changes of the Kohn–Sham energies and wavefunctions caused by the ZG displacements. The method can be applied to compute any property that can be expressed by means of a Fermi Golden Rule, such as for example temperaturedependent band structures, density of states, and optical spectra. The main limitation of this approach as compared to the Wannier–Fourier interpolation method employed in the preceding sections is that, being an adiabatic theory, fine spectral features on the scale of the phonon energy are averaged out. Conversely, its main advantage is that it is easy to use as it requires a single DFT calculation.
The Kohn–Sham energy eigenvalues generated by the special displacement method capture temperature renormalization and quantum zeropoint effects at the same level as the adiabatic Allen–Heine theory of temperaturedependent band structures^{171}. The calculation of the imaginary part of the dielectric function including temperature effects, zeropoint corrections, and phononassisted indirect processes, is performed by evaluating:
where Ω_{sc} and Ω_{BZ,sc} are the volumes of the supercell and the supercell Brillouin zone, respectively, the summations over v, c refer to valence and conduction states, and K is a wavevector of the supercell Brillouin zone. This expression contains both direct and indirect optical transitions: the indirect transitions are hidden in the dependence of the wavefunctions ψ_{vK} and ψ_{cK} on the ZG displacement. Similarly, this expression includes temperature renormalization via the dependence of the energies ε_{vK} and ε_{cK} on the ZG displacements^{164}.
Computational considerations
Calculations using the special displacement method proceed as follows. First, one performs phonon calculations for the crystalline unit cell (not the supercell) using standard DFPT on a coarse uniform Brillouin zone grid. From this calculation, the phonon frequencies and eigenmodes are computed on a finer Brillouinzone grid with N_{1} × N_{2} × N_{3}qpoints using standard interpolation of the force constant matrix^{8}. Then the ZG displacement within a supercell consisting of N_{1} × N_{2} × N_{3} unit cells is evaluated at the temperature T via Eq. (55). The desired property is finally computed with this supercell, with the atoms displaced according to \(\Delta {\tau }_{\kappa \alpha p}^{{{{\rm{ZG}}}}}\). The special displacements are generated by the ZG module; the procedure is computationally inexpensive and is performed serially.
The ZG toolkit provides several codes to analyze the results of supercell calculations performed with special displacements, for example, Brillouinzone unfolding of temperaturedependent band structures from the supercell to the unit cell, the density of states, electronic spectra, and vibrational spectra.
Band unfolding is performed using the procedure outlined in ref. ^{172}, which consists of determining the spectral function in the primitive Brillouin zone by projecting the wavefunctions of the supercell into the wavefunctions of the unit cell. In practice, the spectral density for the wavevector k at the energy ε is obtained as
where the spectral weights P_{mK,k} are given by
Here, c_{mK} denotes planewaves coefficients of supercell wavefunctions, and G_{sc} denotes reciprocal lattice vectors of the supercell. Equation (59) refers to normconserving pseudopotential implementations. The ZG module includes additional terms that are required in this expression when using ultrasoft and PAW (projector augmentedwave) pseudopotentials^{165,169}. A similar unfolding strategy is employed to analyze lattice dynamics, as well as Xray and neutron diffuse scattering intensities accounting for multiphonon interactions^{173}.
The supercell size required to achieve converged results in calculations employing the special displacement method depends on the target property. For example, to compute optical spectra over a broad photon energy range, supercells of size 4 × 4 × 4 unit cells are typically sufficient. However, to obtain accurate phononinduced band gap renormalization, say within a 10 meV accuracy, supercells as large as 10 × 10 × 10 may be required^{165} (see Table 1).
Application example
To demonstrate the implementation of the ZG module, we investigate the temperaturedependent band structure renormalization and phononassisted optical absorption spectra of silicon and BaSnO_{3}. We perform calculations using the local density approximation (LDA) for the exchange and correlation^{174,175} and ONCV pseudopotentials^{46,61}. We use a planewaves kinetic energy cutoff of 40 Ry for silicon and 120 Ry for BaSnO_{3}, and 6 × 6 × 6 uniform kpoint grids for groundstate calculations. With these settings, we obtain indirect and direct band gaps of 0.49 and 2.56 eV for silicon, respectively, and indirect and direct band gaps of 1.06 and 1.55 eV for BaSnO_{3}, respectively. We perform calculations of dynamical matrices on 4 × 4 × 4 qpoint grids in both cases and use the ZG module to generate displacements in 3 × 3 × 3 supercells. The signs S_{qν} appearing in Eq. (55) are determined by minimizing the error descriptor in Eq. (54) of ref. ^{165} with a dimensionless threshold η = 0.1, after enforcing a smooth Berry connection between vibrational eigenmodes across the Brillouin zone^{165}. Kohn–Sham energies in the structures with ZG displacements are calculated using a 12 × 12 × 12 uniform Kpoint grid in the supercell Brillouin zone, and 108 unoccupied states. For calculations of dielectric functions, we use up to 200 randomly generated Kpoints in the supercell Brillouin zone, as well as 27 and 135 conduction bands for silicon and BaSnO_{3}, respectively. All Dirac deltas are replaced by Gaussians of width 30 meV.
Figure 6a shows the joint density of states (JDOS) of silicon at 0 K (red) and 300 K (blue), as calculated using the special displacement method. For comparison, the JDOS computed for the DFT ground state structure is also shown in green. We see that, upon including electronphonon interactions via the special displacements, even at 0 K the JDOS is redshifted with respect to the DFT ground state. This is a manifestation of the zeropoint band gap renormalization^{176}. Upon increasing temperature, the onset of the JDOS further redshifts. This temperatureinduced band gap narrowing is referred to as the Varshni effect^{177}. The horizontal offset between the JDOS for the DFT ground state structure and the JDOS computed with special displacements at 0 K yields a zeropoint renormalization of 50 meV, in good agreement with prior work^{178,179}. More accurate values can be calculated by increasing the supercell size^{164,165}. The band gap renormalization can also be computed without using the JDOS; to this end, one needs to evaluate the Kohn–Sham eigenvalues at the supercell Kpoints that unfold onto the wavevectors of the band extrema in the Brillouin zone of the unit cell.
Figure 6b and c report color maps of the electronic spectral functions of silicon and BaSnO_{3} at 0 K calculated using the special displacement method. For comparison, the band structures in the DFT ground state are overlaid with these color maps. These maps can directly be compared to angleresolved photoelectron spectroscopy data. From these images, we identify numerically the quasiparticle band structures by extracting the spectral peaks. In the case of silicon, the valence band maximum blueshifts by 32 meV with respect to groundstate DFT, and the conduction band redshifts by 18 meV. The resulting gap renormalization of 50 meV is in agreement with the value determined in Fig. 6a via the JDOS. In the case of BaSnO_{3}, Fig. 6 shows a zeropoint renormalization of the band gap of 10 meV; however, we emphasize that this value is not fully converged, and larger supercells, as well as corrections for Fröhlich couplings^{180}, are necessary to obtain accurate data.
Figure 6d reports a convergence test of the imaginary part of the dielectric function of silicon, as computed with the special displacement method. In this calculation, we keep the supercell fixed, and we increase the number of random Kpoints in the supercell Brillouin zone. It is seen that the dielectric function converges relatively rapidly with the number of points and full convergence is achieved with 200 points when the Dirac delta functions in Eq. (57) are replaced by Gaussians with 30 meV width. Clearly, a larger smearing would require fewer Kpoints.
In Fig. 6e we compare the imaginary part of the dielectric function of silicon computed in the DFT ground state structure (red) with the special displacement method (blue). We see that the special displacements correctly capture phononassisted indirect optical transitions in the energy range between the indirect gap and direct gap of silicon. This approach provides an alternative strategy for computing optical spectra to Eq. (49) implemented in the indirect optics module of EPW. In the special displacement method, both temperaturedependent band structure renormalization and phononassisted processes are included on the same footing. Figure 6f shows the imaginary part of the dielectric function of BaSnO_{3} at 0 K (blue) and 300 K (green), including phononassisted processes. For comparison, we also show the calculation using the DFT groundstate structure (red), which misses phononassisted transitions. The spectra compare well with prior work using the special displacement method^{181}.
Implementation and HPC benchmarks
In this section, we describe some of the implementation and parallel programming models of EPW. We first outline the computational workflow and basic capabilities of EPW. Then we describe our recent efforts to prepare EPW for exascale HPC systems with a focus on an explorative implementation of a highly scalable parallelization scheme, hybrid twolevel MPI, and OpenMP parallelization. We present benchmarking results which demonstrate that, with this improved parallelization strategy, EPW can run at nearly full scale on the preexascale HPC system Frontera at the Texas Advanced Computing Center (TACC). Finally, we describe the I/O strategy employed in this explorative implementation, which is based on XML (eXtensible Markup Language) and HDF5 file formats. We emphasize that EPW continues to evolve with changes and improvements, and the reader is referred to the EPW website for its most uptodate features and functionalities^{182}.
Computational workflow and basic capabilities
The EPW code is logically partitioned into two components. The first component serves as an interface to the PWscf and PHonon codes of Quantum ESPRESSO. The code imports groundstate charge density and wavefunctions with computational parameters from 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 evaluated on coarse k and qpoints grids, and the phonon wavevectors are restricted to the irreducible wedge of the Brillouin zone. Then the code calls Wannier90 in library mode to calculate the unitary rotation matrices U_{mnk} [Eq. (64) of the “Methods” section] needed to transform Bloch states into maximally localized Wannier functions. Subsequently, the code unfolds the irreducible wedge of the phonon Brillouin zone into the full zone, evaluates electron–phonon matrix elements on coarse k and qpoint grids, and it transforms the Hamiltonian, dynamical matrices, and electron–phonon matrix elements into the Wannier representation. At this point, the code can print relevant quantities for subsequent runs with epw.x, and it can continue to the interpolation part. The interpolation engine of EPW carries out Wannier interpolation of the Hamiltonian, dynamical matrices, and electron–phonon matrix elements on fine k and qpoint grids after reading these quantities in the Wannier representation, and then it calculates electron–phononrelated properties on these grids.
The current snapshot of the code supports spinunpolarized and nonmagnetic spin–orbit calculations; work is currently in progress to extend the code to the cases of spinpolarized and noncollinear magnetic cases. Regarding pseudopotentials (PPs), normconserving (NC) PPs are supported. Since both ultrasoft PPs and the PAW method violate fundamental symmetry relations that the electron–phonon matrix elements must fulfill^{183}, extending the current NCPP implementation to these cases poses some challenges. We will proceed to these extensions in future releases.
Parallelization
At the time of writing of this manuscript, we witness the launch of the first supercomputer entering the exascale computing era, namely Frontier at Oak Ridge National Laboratory; additional exascale systems are currently being deployed worldwide. Exascale HPC systems are characterized by manycore and heterogeneous architectures. Since 2020, we have made efforts to ready EPW for the exascale transition. In the following, we describe one strategy that we are currently exploring, a hybrid twolevel MPI/OpenMP parallelization scheme, which makes EPW work seamlessly and efficiently on manycore architectures.
Historically EPW has adopted a onelevel parallelization over the electron wavevectors (kpoint parallelization) via MPI. This strategy has three shortcomings, which lead to an early saturation of the speedup with the number of cores in largescale calculations: (i) There is an upper bound for the number of total MPI tasks which guarantees good scalability; this bound is determined by the number of kpoints. (ii) As the number of MPI tasks increases, the overhead cost associated with MPI communications, in particular collective communications, increases. (iii) Due to the internal buffers of MPI, the memory overhead also increases with the number of MPI tasks. To address these issues, we are exploring the extension of the onelevel parallelization over kpoints to the hybrid twolevel MPI and OpenMP parallelization over both k and qpoints.
In the hierarchical twolevel 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 qpoint pools, and each pool is further divided into q and kpoint pools, respectively. Each q or kpoint pool contains a set of MPI tasks, and OpenMP parallelization is employed within each MPI task. The parallelization option employed in the upperlevel pool (over kpoints or qpoints) is chosen so as to maximize the calculation efficiency.
As compared to the onelevel parallelization, the hierarchical twolevel 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 computational workload over a large number of cores.
To demonstrate the scaling behavior of EPW, we perform strongscaling tests for the twogap superconductor MgB_{2}. We evaluate electronphonon matrix elements on 150 × 150 × 150 k and qpoint grids (Fig. 7). These calculations are performed on the Frontera supercomputer at TACC. Frontera consists of 8368 nodes, each equipped with two Intel Xeon Platinum 8280 ("Cascade Lake”) processors with 28 cores per processor. To minimize statistical fluctuations in these benchmarks, we execute each calculation four times, and we average the resulting Wall times. Figure 7 shows that using this parallelization scheme, we achieve ~92% of the ideal speedup up to 448,000 cores in the evaluation of the electron–phonon matrix elements.
The hybrid twolevel MPI and OpenMP parallelization scheme is controlled by three parameters that can be optimized to achieve maximum parallel scaling efficiency: the number of upperlevel MPI tasks, the number of lowerlevel MPI tasks, and the number of OpenMP threads per MPI task. In the benchmarks shown in Fig. 7, we fix the number of kpoint pools to 250, by varying the number of qpoint pools depending on the number of cores. With this choice, we are able to achieve nearideal speedup on extreme scaling tests; we expect that a speedup even superior to 92% could be achieved upon further optimization of the parallelization parameters We also point out that, in largescale runs, I/O time dominates the total Wall time. To overcome this I/O bottleneck, we employ a lowI/O mode whereby memory is exploited as much as possible and minimum usage of storage media is made.
Our strategy for OpenMP parallelization is to use threaded versions of scientific libraries such as Intel MKL and Cray LibSci, rather than explicitly using OpenMP directives throughout the code. For the benchmarks reported in Fig. 7, we find that exceeding 7 OpenMP threads per MPI task worsens the performance. Therefore we use 4 OpenMP threads per MPI task, and we make all threads belonging to the same MPI task reside in the same nonuniform memory access (NUMA) domain.
Structured I/O
Recently we have been investigating the use of XML and the HDF5 data formats. Both formats can describe hierarchical data and are selfdescribing, flexible, and portable. For small to mediumsized data, the XML format is used; for large binary data, the HDF5 file format is used in parallel. In previous versions of EPW, several large arrays were read or written via perprocess I/O or parallel MPI I/O. The former quickly overwhelms the file systems in largescale calculations by generating a large number of files; the latter is not easily portable. Both issues can be addressed for largesize binary data by employing parallel I/O via parallel HDF5.
We note that parallel I/O performance heavily depends on the details of file systems and the parallelization settings. For instance, in the case of the LUSTRE file system, the number and size of stripes play an essential role in achieving high performance. In the case of the ROMIO implementation of MPIIO, identifying optimum parameters in largescale runs is nontrivial and requires systematic experimentation on HPC systems.
Discussion
Since the last technical paper in 2016^{18}, the EPW code has considerably expanded in scope, to the point of becoming a robust and efficient software platform for developing, testing, and deploying methods addressing electron–phonon physics and related materials properties. It is then natural to ask which components and functionalities are still in need of improvement, what are emerging directions in this area, and more generally what comes next for the EPW project.
On the functionality front, we expect to continue expanding the capabilities for transport, superconductivity, polarons, and optics described in the ‘Results’ section. For example, the transport module is based on the steadystate linearized Boltzmann transport equation; here, it would be desirable to generalize the methodology to highfield transport and to timedependent driving fields. Similarly, in the superconductivity module, the treatment of electron–electron repulsion could be improved by seamlessly integrating EPW with standard GW codes. The study of polaron physics from an ab initio manybody perspective is just beginning, therefore we can expect a number of developments in this area, from the study of polaron transport to their optical properties and their relation with the theory of band structure renormalization. Calculations of phononassisted optical transitions are currently restricted to optical absorption spectra, but the formalism can just as well be employed for investigating related phenomena such as for example Auger–Meitner recombination^{184}. Beyond these core modules of the EPW project, we anticipate growth in the calculation of electronic and optical properties using the special displacement method within the EPW/ZG code; for example, the current implementation focuses on harmonic systems, but generalizations to strongly anharmonic systems have recently been proposed^{185}.
On the accuracy front, we anticipate that future work will focus on improving the precision of the Wannier interpolation method that is the basis for EPW. For example, methods that take dimensionality into account in the interpolation of the electron–phonon matrix elements will be essential to perform predictive and reliable calculations of transport, optics, superconductivity, and polarons in 2D materials^{186,187}. Furthermore, it would be highly desirable to improve the predictive power of the DFPT electron–phonon matrix elements before even proceeding to Wannier–Fourier interpolation; a recent proposal to employ the GW method to calculate manybody corrections to the DFPT matrix elements offers a promising path toward this goal^{188}.
On the front of HPC, we envision continuing the current effort to enable EPW for exascale computing architectures. In addition to the hybrid MPI/OpenMP parallelization, it will be necessary to leverage diverse GPU architectures and to enable the code for largescale runs that will become possible with the exascale supercomputers that are being deployed worldwide.
Another important direction will be to enhance the interoperability of EPW with other major electronic structure software packages such as Abinit^{189}, VASP^{190}, Siesta^{191}, and GPAW^{192}. As several codes currently use the information generated by EPW for a variety of postprocessing tasks, it will make sense to develop standardized data structures that adhere to the FAIR data principles.
Regarding the programming model, transitioning toward objectoriented programming (OOP) will be useful due to the potential for high modularity, extensibility, and reusability, which will facilitate flexible and sustained software development^{193}. We expect that targeted use of OOP will greatly increase the efficiency and productivity in the development and maintenance of code^{194}.
Alongside these developments, we anticipate increased attention to the issue of automation and the capability of performing electron–phonon calculations at scale for highthroughput approaches, data science, and artificial intelligence/machine learning (AI/ML) applications.
In summary, in this manuscript, we provided a comprehensive update on the current status, functionalities, and performance of the EPW code, we offered an overview of possible future directions for this and similar codes. We hope that, beyond capturing a snapshot of the current status of the EPW project, the present manuscript will contribute to making advanced electron–phonon calculations more accessible and more widely used in computational materials discovery and design.
Methods
The key element of any calculation of electronphonon interactions and related materials properties is the electron–phonon matrix element. The EPW code employs physicsbased Wannier–Fourier interpolation to compute electronphonon matrix elements accurately and efficiently and uses these matrix elements as the starting point for advanced manybody calculations of electronic, optical, and transport properties of solids including temperature dependence and phononassisted quantum processes.
In the following we introduce the notations and conventions employed throughout the manuscript, we provide a brief overview of the general principles of Wannier–Fourier interpolation of electron–phonon matrix elements, and we outline the extension of this interpolation scheme to the case of polar materials. A more comprehensive description of this methodology including detailed derivations can be found in ref. ^{10}.
Notation and definitions
In the study of the electronic structure and lattice dynamics of crystalline solids, infinitely extended crystals are described by considering a periodic BvK supercell. This supercell consists of N_{uc} = N_{1} × N_{2} × N_{3} primitive unit cells of the crystal and is subject to periodic boundary conditions. The pth unit cell within the BvK supercell is identified by the direct lattice vector R_{p}, with p = 1, …, N_{uc}. The dual lattice of the direct lattice vectors R_{p} in reciprocal space defines a uniform grid of N_{uc} = N_{1} × N_{2} × N_{3} wavevectors in the Brillouin zone. Throughout the manuscript, we use k and q to denote electron and phonon wavevectors in the crystal Brillouin zone, respectively, and K to denote electron wavevectors in the Brillouin zone of the BvK supercell.
We use ψ_{nk} to indicate the eigenfunctions of the DFT KohnSham Hamiltonian \({\hat{H}}_{{{{\rm{KS}}}}}\), and ε_{nk} the corresponding eigenvalues. This wavefunction is periodic and normalized over the BvK supercell, and can be expressed in the Bloch form:
with u_{nk} being a latticeperiodic function normalized in the primitive unit cell of volume Ω. The position vector of the atomic nucleus κ belonging to the unit cell identified by R_{p} is \({{{{\boldsymbol{\tau }}}}}_{\kappa p}^{}={{{{\bf{R}}}}}_{p}+{{{{\boldsymbol{\tau }}}}}_{\kappa }^{}\), with \({{{{\boldsymbol{\tau }}}}}_{\kappa }^{}\) being the position vector within the primitive unit cell. The components of this vector along the Cartesian directions α = 1, 2, 3 are denoted by τ_{καp}.
Vibrational eigenmodes and eigenfrequencies are obtained by diagonalizing the dynamical matrix, which is computed via DFPT using Quantum ESPRESSO^{12}. In turn, the dynamical matrix is the Fourier transform of the interatomic force constants \({C}_{\kappa \alpha p,{\kappa }^{{\prime} }{\alpha }^{{\prime} }{p}^{{\prime} }}\), that is the Hessian of the DFT total energy in the atomic coordinates. The relation between interatomic force constants and dynamical matrix is^{135}
where M_{κ} is the mass of the κth nucleus. The eigenvalues and eigenvectors of the dynamical matrix are denoted by \({\omega }_{{{{\bf{q}}}}\nu }^{2}\) and e_{κα,ν}(q), respectively, and the index ν runs from 1 to 3M where M is the number of atoms in the primitive cell. ω_{qν} corresponds to the vibrational frequency, and e_{κα,ν}(q) is the normal mode of vibration or polarization vector.
The electron–phonon matrix element is defined by^{1,10}
where the integral is evaluated over the unit cell, and the latticeperiodic component of the variation of the Kohn–Sham potential is given by
In this expression, M_{0} is an arbitrary reference mass introduced for convenience. MLWFs are defined in terms of Bloch states as^{9}
where U_{nmk} is a unitary matrix in the indices m and n.
Throughout the manuscript, we used interchangeably the notation
to indicate summations or integrals over the Brillouin zone. The quantity Ω_{BZ} on the right indicates the Brillouin zone volume.
The unitary matrix U_{nmk} appearing in Eq. (64) is determined by requiring that the functions w_{mp}(r) be as localized as possible^{195}, i.e. MLWFs^{9}. The computation of these unitary matrices is performed by minimizing the spatial spread of the resulting Wannier functions^{9}. This procedure is notoriously challenging since the spread functional exhibits multiple local minima, especially in the case of entangled electron bands. In practice, the construction of optimal MLWFs often requires finetuning of several computational parameters, such as the inner and outer energy windows, and the projections used to initialize the minimization algorithm. Historically, this optimization has often been performed by using chemical intuition, and sometimes a trialanderror approach.
In an effort to enable more systematic construction of MLWFs for highthroughput calculations and datadriven materials research, several strategies have been developed, such as for example those of refs. ^{196,197}. These strategies fall into two categories, one relying on the projectability concept^{198}, and the other on the selectedcolumnsofthedensitymatrix (SCDM) method^{199,200}. The former approach relies on the projectability of Bloch states on pseudoatomic orbitals associated with the pseudopotentials; the projectability metric of each Bloch state is used to downselect the band manifold for subsequent minimization of the spread functional. The latter approach does not rely on the pseudoatomic orbitals or other choice of initial projections but generates an optimal set of states from the columns of the density matrix via QR decomposition with column pivoting.
Both approaches are very promising and bring Wannier functions one step closer to complete automation. Nevertheless, some challenges remain. For example, to automatically construct Wannier functions and achieve highquality Wannier interpolation, both strategies require the use of all the occupied valence bands (except for lowlying semicore states) as well as a few empty conduction bands for metallic or semimetallic systems. In systems with mediumsized to large unit cells, the number of bands thus identified may be relatively large, and this will make EPW calculations more demanding because the number of operations required to obtain electronphonon matrix elements scales with the square of the number of Wannier functions. Therefore, more work will be needed in the future to achieve automation whilst preserving calculation efficiency.
General principles of Wannier interpolation of electron–phonon matrix elements
The localized nature of MLWFs provides the basis for accurate and efficient interpolation of Kohn–Sham wavefunctions and energies^{201}. This is achieved by expressing the Kohn–Sham Hamiltonian in the Wannier representation, and noting that the Hamiltonian matrix elements decrease rapidly with the distance between the Wannier function centers. The Hamiltonian matrix elements in the Wannier representation are^{195}
Using Eq. (64), these matrix elements can be expressed in terms of the Kohn–Sham Hamiltonian in the Bloch representation \({H}_{{m}^{{\prime} }{n}^{{\prime} }}({{{\bf{k}}}})\):
where the matrix U_{mnk} is the same as in Eq. (64). Once H_{mn}(R_{p}) has been determined, Eq. (67) can be inverted to generate \({H}_{{m}^{{\prime} }{n}^{{\prime} }}({{{\bf{k}}}})\) anywhere in the Brillouin zone.
Similarly to the electronic case, the dynamical matrix can be expressed in the phonon Wannier representation as^{10}
where D_{μν}(q) is the matrix element of the dynamical matrix in the Bloch representation for phonons. Once \({D}_{\kappa \alpha ,{\kappa }^{{\prime} }{\alpha }^{{\prime} }}({{{{\bf{R}}}}}_{p})\) has been determined, this relation can be inverted to obtain the phonon eigenvectors and eigenvalues anywhere in the Brillouin zone; this is a standard procedure employed to compute phonon dispersion relations^{7,202}.
Equations (67) and (68) can be generalized to the case of electron–phonon matrix elements by considering Fourier transforms for both the Kohn–Sham states and for the vibrational eigenmodes. The resulting electron–phonon matrix elements in the Wannier representation are given by^{10}:
Once the \({g}_{mn\kappa \alpha }({{{{\bf{R}}}}}_{p},{{{{\bf{R}}}}}_{{p}^{{\prime} }})\) have been computed, this relation can be inverted to generate electronphonon matrix elements anywhere in the Brillouin zone.
Equations (67)–(69) constitute the backbone of the interpolation engine of EPW. In practice, the interpolation module of EPW reads in the DFT electron density, dynamical matrices, and variations of the Kohn–Sham potential evaluated by Quantum ESPRESSO on a coarse Brillouin zone grid; computes the Kohn–Sham wavefunctions and electron–phonon matrix elements on this grid; calls the Wannier90 code in library mode to obtain the Wannier matrices in Eq. (64); transforms Hamiltonian, dynamical matrix, and electron–phonon matrix elements in the Wannier representation; and interpolates all these quantities onto arbitrarily dense k and qpoint grids. A qualitative schematic of this process is shown in Fig. 8.
Treatment of longrange electron–phonon interactions
In semiconductors and insulators, the ionic displacements associated with a phonon can induce a variation of the Kohn–Sham potential that is longranged in nature. As a result, the electron–phonon matrix element g_{mnν}(k, q) associated with longitudinal–optical (LO) phonons becomes singular in the longwavelength limit q → 0. More specifically, in any material exhibiting nonvanishing Born effective charges, the matrix elements associated with LO phonons diverge as q/∣q∣^{2} at small q. In these cases, the assumption of locality that underpins the methodology described in the previous section ceases to hold, and the Wannier interpolation procedure needs to be modified to correctly capture the singularity.
The singularity in the matrix element can be dealt with by considering a multipole expansion of the Kohn–Sham potentials resulting from individual atomic displacements. The first order in this expansion is the dipole potential^{203}, which scales as q/∣q∣^{2} and is responsible for the wellknown Fröhlich electron–phonon interaction^{204}. The modification of the electron–phonon interpolation method to include longrange effects consists of separating shortrange and longrange parts in the matrix elements, as follows:
where the first term on the righthand side is the shortrange component, the second term is the dipole component, and the third term is the quadrupole component. Additional multipoles could be considered, but the dipole and quadrupole terms already lead to very accurate results. The dipole matrix element \({g}_{mn\nu }^{{{{\mathcal{L}}}},{\rm {D}}}({{{\bf{k}}}},{{{\bf{q}}}})\) was derived in refs. ^{205,206} and reads:
In this expression, ϵ_{0} is the vacuum permittivity, \({{{{\boldsymbol{\epsilon }}}}}_{}^{\infty }\) is the highfrequency dielectric tensor of the material, G represents a reciprocal lattice vector, \({{{\bf{{Z}}}_{\kappa }^{* }}}\) is the Born effective charge tensor of the atom κ, the bracket indicates the integral over the BvK supercell, and we use the notation a ⋅ B ⋅ c = ∑_{αβ}a_{α}B_{αβ}c_{β}. The unitary matrices U_{mnk} in Eq. (71) are the Wannier transformation obtained following the standard procedure^{207,208} and coming from the q + G → 0 limit of the overlap integrals between Kohn–Sham wavefunctions \(\langle {\psi }_{m{{{\bf{k}}}}+{{{\bf{q}}}}} {{\rm {e}}}^{i{{{\bf{q}}}}\cdot {{{\bf{r}}}}} {\psi }_{n{{{\bf{k}}}}}\rangle ={\sum }_{p}{U}_{mp{{{\bf{k}}}}+{{{\bf{q}}}}}{U}_{pn{{{\bf{k}}}}}^{{\dagger} }+{{{\mathcal{O}}}}({{{\bf{q}}}})\). This matrix element reduces to the standard Fröhlich interaction^{204} when one considers parabolic electron bands and a dispersionless LO mode in a cubic material^{186,205}. This term is of the order of ∣q∣^{−1}. The quadrupole contribution is the second term of the multipole expansion of the Kohn–Sham potential and is of the order of ∣q∣^{0}. The corresponding matrix element \({g}_{mn\nu }^{{{{\mathcal{L}}}},{\rm {Q}}}({{{\bf{k}}}},{{{\bf{q}}}})\) was derived in refs. ^{40,41,187,209,210} and reads:
In this expression, Q_{κ} is the dynamical quadrupole tensor^{211}, \({V}^{{{\rm{Hxc}}},{{{{\mathcal{E}}}}}_{\gamma }}({{{\bf{r}}}})\) is the selfconsistent potential induced by a uniform electric field \({{{{\mathcal{E}}}}}_{\gamma }\) along the Cartesian direction γ^{41,187}, \({A}_{sp{{{\bf{k}}}},\gamma }^{\,{{\rm{W}}}\,}=i\langle \partial {u}_{s{{{\bf{k}}}}}^{\,{{\rm{W}}}\,}/\partial {k}_{\gamma } {u}_{p{{{\bf{k}}}}}^{\,{{\rm{W}}}\,}\rangle\) is the Berry connection, and \({u}_{p{{{\bf{k}}}}}^{\,{{\mbox{W}}}\,}\) is the smooth Bloch basis^{187}. In EPW, the calculation of the Berry connection term is automatically activated for quadrupole calculations when the velocity matrix is computed in the Wannier representation, cf. Eq. (27) of ref. ^{22}. This term can be nonnegligible in particular for materials where the quadrupole tensor Q_{κ} vanishes by symmetry, such as SrO^{187}. The \({V}_{}^{\,{{\rm{Hxc}}}\,,{{{\boldsymbol{{{{\mathcal{E}}}}}}}}}\) term has been implemented in Abinit^{40,41,212} and EPW^{187,213} and has been tested for Si, GaP, GaAs, SrO, hBN monolayer, MoS_{2} monolayer, and InSe monolayer. In all these cases, it was found that the contribution from this term represents <0.1% of the total quadrupole correction^{41,187}. Since this term is very small and its implementation is rather cumbersome, it has not been included in public releases of these codes.
The calculation strategy employed by EPW is as follows. First, the complete matrix elements g_{mnν}(k, q) are evaluated on coarse k and q grids using DFPT. Second, the longrange contributions \({g}_{mn\nu }^{{{{\mathcal{L}}}},{\rm {D}}}({{{\bf{k}}}},{{{\bf{q}}}})\) and \({g}_{mn\nu }^{{{{\mathcal{L}}}},Q}({{{\bf{k}}}},{{{\bf{q}}}})\) are subtracted from the DFPT matrix elements using Eqs. (71) and (72), leaving out the shortrange component \({g}_{mn\nu }^{{{{\mathcal{S}}}}}({{{\bf{k}}}},{{{\bf{q}}}})\) on the coarse grids. Third, the standard Wannier electron–phonon interpolation of ref. ^{10} is applied to the shortrange component only. And fourth, the longrange contributions are added back using Eqs. (71) and (72) on the fine k and q grids.
The infinite sum over the G vectors in Eqs. (71) and (72) ensure the periodicity of the matrix elements in reciprocal space. In practical calculations, this sum might be computationally demanding to converge. In principle, periodicity could be enforced by including only one reciprocal lattice vector in the sum, namely \({\tilde{{{{\bf{G}}}}}}_{{{{\bf{q}}}}}\) such that \( {{{\bf{q}}}}+{\tilde{{{{\bf{G}}}}}}_{{{{\bf{q}}}}} =\mathop{\min }\nolimits_{{{{\bf{G}}}}} {{{\bf{q}}}}+{{{\bf{G}}}}\). However, this choice introduces derivative discontinuities in \({g}_{mn\nu }^{{{{\mathcal{L}}}},{\rm {D}}}({{{\bf{k}}}},{{{\bf{q}}}})\) at the Brillouin zone boundaries, which in turn cause spurious oscillations in the interpolation of \({g}_{mn\nu }^{{{{\mathcal{S}}}}}({{{\bf{k}}}},{{{\bf{q}}}})\). We illustrate this point in Fig. 9a–d, focusing on the longitudinaloptical (LO) phonon of cubic boron nitride (cBN).
An alternative strategy to avoid the sum over G is to cut off the interaction range in Eqs. (71) and (72) using a Gaussian filter, \(\exp [({{{\bf{q}}}}+{{{\bf{G}}}})\cdot {{{{\boldsymbol{\epsilon }}}}}^{\infty }\cdot ({{{\bf{q}}}}+{{{\bf{G}}}})/4\alpha ]\). This choice finds motivation in the Ewald summation method that is commonly employed to evaluate the nonanalytic contribution to the dynamical matrix in polar materials^{7,8,202}. It was employed in previous versions of EPW as well as in other codes^{13,40,206}, and it avoids derivative discontinuity. However, it does not preserve the periodicity of the matrix elements in reciprocal space, which is required by the periodic gauge. In EPW, to maintain the periodicity of the matrix elements, we sum over a shell of G vectors centered around \({\tilde{{{{\bf{G}}}}}}_{{{{\bf{q}}}}}\), and then apply the Gaussian filter. Figure 9c shows how this procedure yields periodic and smooth longrange matrix elements with the correct behavior next to the singularities, and enables accurate interpolation of the complete matrix elements, as shown in Fig. 9b.
The α parameter in the Gaussian filter is chosen in such a way as to ensure accurate interpolation (which requires large α) whilst using as few G vectors as possible in Eqs. (71) and (72) (which requires small α). In Fig. 9e we compare the interpolated matrix element g_{mnν}(k, q) to explicit DFPT calculations, for different values of α, with the reciprocal space summation restricted to those G vectors such that \(({{{\bf{q}}}}+{{{\bf{G}}}})\cdot {{{{\boldsymbol{\epsilon }}}}}_{}^{\infty }\cdot ({{{\bf{q}}}}+{{{\bf{G}}}}) < 56\alpha\). With this choice, only Gvectors yielding a value of the Gaussian filter larger than \(\exp (14)\simeq 1{0}^{6}\) are included in the sum. We find that when 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 that, for layered materials, more G vectors are automatically included in the sum along the perpendicular direction due to the anisotropy of the dielectric tensor. For completeness, in Fig. 9f we show how the number of required G vectors increases with α, and in Fig. 9g we show the interpolation error in the matrix elements as a function of α. We emphasize that, since EPW implements this Gaussian filter, caution should be used when extracting dipolar and quadrupolar matrix elements from EPW for separate postprocessing, because \({g}_{mn\nu }^{{{{\mathcal{L}}}},{\rm {D}}}({{{\bf{k}}}},{{{\bf{q}}}})\) and \({g}_{mn\nu }^{{{{\mathcal{L}}}},{\rm {Q}}}({{{\bf{k}}}},{{{\bf{q}}}})\) have the expected qdependence only at long wavelength.
Figure 9h, i illustrate the importance of describing dipole and quadrupole interactions using Eqs. (71) and (72), for the case of cBN. To facilitate the comparison between the interpolation results and explicit DFPT calculations for longwavelength acoustic modes, we use the following descriptor which removes the factor \({\omega }_{{{{\bf{q}}}}\nu }^{1/2}\) from the matrix element and averages over electronic degeneracies:
Here, the sum over bands is carried over the N_{W} states in the Wannier manifold and ρ is the mass density of the crystal. D_{ν}(k, q) has units of energy divided by length and can be thought of as a deformation potential of sort. In Fig. 9h, i, we choose the manifold composed by the top three valence bands of cBN, and set k = 0. We see that the dipole term is necessary to correctly describe the singular behavior of the LO phonon at long wavelength, and the quadrupole term is necessary to correctly describe the discontinuous behavior of the longitudinal acoustic (LA) phonons at a long wavelength [see expanded view in Fig. 9i]. The inclusion of both terms guarantees a highquality interpolation of the electronphonon matrix elements across the entire Brillouin zone. In particular, in this example, we find that the deformation potential calculated by including dipole and quadrupole terms agrees with direct DFPT calculations within 0.01 eV/Bohr (<1% of the maximum value) for the LA mode of cBN along the ΓL line near the zone center. In refs. ^{22,187} a number of additional comparisons between the present method and direct DFPT calculations are discussed for standard threedimensional (3D) and twodimensional (2D) semiconductors.
The EPW code also implements longrange corrections for the phonon dynamical matrix by computing dipole–dipole, dipole–quadrupole, and quadrupole–quadrupole terms as discussed in refs. ^{7,8,202,211}. Systematic tests of the interpolation procedure described in this section are provided in ref. ^{22}.
The expressions for the longrange dipole and quadrupole 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 years^{186,187,213,214,215,216}. In the first of such generalizations^{214}, the authors obtain the longwavelength limit of polar electron–phonon couplings in 2D by evaluating the average dielectric polarization of a 2D slab enclosing the layer and removing the interaction between periodic replicas via Coulomb truncation. In ref. ^{215}, the authors additionally consider the contributions to the polar matrix elements arising from outofplane dipoles and demonstrate an improvement in the interpolation accuracy. In addition to these studies, more recent work by some of us focused on including quadrupole corrections for 2D materials^{187} and on establishing the connection between polar matrix elements in 3D and 2D materials^{186}. In particular, in ref. ^{187}, the 2D longrange matrix elements to quadrupolar order are given by
where the 2D longrange scattering potential to first order in q is
where S is the unitcell area, and the dielectric functions appearing at the denominators are
In these expressions, \({{{{\boldsymbol{\alpha }}}}}_{}^{\parallel }\) and α^{⊥} are macroscopic polarizabilities in and outofplane, respectively, and \(f(q)=1\tanh (qL/2)\) is a rangeseparation function and the distance L is determined according to Eq. (63) of ref. ^{187}. The 2D extension of the overlap integral is provided in Eqs. (46) and (47) of ref. ^{187}. This formulation also uses Coulomb truncation and generalizes the previous approaches of refs. ^{214,215} to the case of quadrupole corrections. In the approach of Ref. ^{186}, the authors propose a unified formulation of polar electron–phonon interactions in 3D and 2D that reduces to the approach of refs. ^{206,217} in 3D, and to the approach of refs. ^{214,215} in 2D. The theory is developed along the same lines as in ref. ^{217}, by evaluating the potential of point dipoles resulting from the atomic displacements. The key difference is that, in this case, the dipoles belong to a supercell consisting of a dielectric stack where one component is the material under consideration, and the other is a homogeneous dielectric (or vacuum). By solving this electrostatic problem, the authors obtain the following matrix element:
where ϵ_{∞,env} is the dielectric constant of the homogeneous dielectric surrounding the 2D material, and K(q + G, τ_{κz}) is a kernel that depends on the thickness of the 2D layer and the surrounding medium, as well as their dielectric constants. The expression for K is given by Eq. (A1) of ref. ^{186}. The main practical difference between this formulation and other approaches is that it can be used without the need for Coulomb truncation.
Data availability
The data presented in each figure are available from the corresponding author upon reasonable request.
Code availability
The codes used for the current study, namely Quantum ESPRESSO, Wannier90, and EPW, are all opensource software distributed via their corresponding websites. In particular, the EPW code described in this manuscript can be downloaded from https://epwcode.org. The data presented in each figure are available from the corresponding author upon reasonable request.
References
Giustino, F. Electron–phonon interactions from first principles. Rev. Mod. Phys. 89, 015003 (2017).
Ziman, J. M. Electrons and Phonons (Oxford University Press, 1960).
Hohenberg, P. & Kohn, W. Inhomogeneous electron gas. Phys. Rev. 136, B864–B871 (1964).
Kohn, W. & Sham, L. J. Selfconsistent equations including exchange and correlation effects. Phys. Rev. 140, A1133–A1138 (1965).
Baroni, S., Giannozzi, P. & Testa, A. Green’sfunction approach to linear response in solids. Phys. Rev. Lett. 58, 1861–1864 (1987).
Savrasov, S. Y. Linear response calculations of lattice dynamics using muffintin basis sets. Phys. Rev. Lett. 69, 2819–2822 (1992).
Gonze, X. & Lee, C. Dynamical matrices, Born effective charges, dielectric permittivity tensors, and interatomic force constants from densityfunctional perturbation theory. Phys. Rev. B 55, 10355–10368 (1997).
Baroni, S., de Gironcoli, S., Corso, A. D. & Giannozzi, P. Phonons and related properties of extended systems from densityfunctional perturbation theory. Rev. Mod. Phys. 73, 515 (2001).
Marzari, N., Mostofi, A. A., Yates, J. R., Souza, I. & Vanderbilt, D. Maximally localized Wannier functions: theory and applications. Rev. Mod. Phys. 84, 1419–1475 (2012).
Giustino, F., Cohen, M. L. & Louie, S. G. Electron–phonon interaction using Wannier functions. Phys. Rev. B 76, 165108 (2007).
GNU General Public License v3, Free Software Foundation. Gnu general public license v3. https://www.gnu.org/licenses (2007).
Giannozzi, P. et al. Advanced capabilities for materials modelling with QUANTUM ESPRESSO. J. Phys.: Condens. Matter 29, 465901 (2017).
Zhou, J.J. et al. Perturbo: A software package for ab initio electron–phonon interactions, charge transport and ultrafast dynamics. Comput. Phys. Commun. 264, 107970 (2021).
Protik, N. H., Li, C., Pruneda, M., Broido, D. & Ordejón, P. The elphbolt ab initio solver for the coupled electron–phonon Boltzmann transport equations. npj Comput. Mater. 8, 28 (2022).
Cepellotti, A., Coulter, J., Johansson, A., Fedorova, N. S. & Kozinsky, B. Phoebe: a highperformance framework for solving phonon and electron Boltzmann transport equations. J. Phys. Mater. 5, 035003 (2022).
Marchese, G. et al. Epiq (2022) (last accessed 27 January 2023); https://theepiqteam.gitlab.io/epiqsite.
Noffsinger, J. et al. EPW: a program for calculating the electron–phonon coupling using maximally localized Wannier functions. Comput. Phys. Commun. 181, 2140–2148 (2010).
Poncé, S., Margine, E. R., Verdi, C. & Giustino, F. EPW: electron–phonon coupling, transport and superconducting properties using maximally localized Wannier functions. Comput. Phys. Commun. 209, 116–133 (2016).
Message Passing Interface Forum. MPI: A MessagePassing Interface Standard Version 5.0 (Message Passing Interface Forum, 2022).
Dagum, L. & Menon, R. Openmp: an industry standard API for sharedmemory programming. Comput. Sci. Eng. 5, 46–55 (1998).
The HDF Group. Hierarchical Data Format, version 5 (1997NNNN). https://www.hdfgroup.org/HDF5/.
Poncé, S. et al. Firstprinciples predictions of Hall and drift mobilities in semiconductors. Phys. Rev. Res. 3, 043022 (2021).
Kadanoff, L. P. & Baym, G. Quantum Statistical Mechanics: Green’s Function Methods in Equilibrium and Nonequilibrium Problems. Frontiers in Physics. A Lecture Note and Reprint Series (Benjamin, 1962).
Poncé, S., Li, W., Reichardt, S. & Giustino, F. Firstprinciples calculations of charge carrier mobility and conductivity in bulk semiconductors and twodimensional materials. Rep. Prog. Phys. 83, 036501 (2020).
Mahan, G. D. Quantum transport equation for electric and magnetic fields. Phys. Rep. 145, 251–318 (1987).
Macheda, F. Ab Initio Prediction of Thermomagnetic and Thermoelectric Transport Phenomena in 3d and 2d Materials. Ph.D. thesis, King’s College London (2021).
Restrepo, O. D., Varga, K. & Pantelides, S. T. Firstprinciples calculations of electron mobilities in silicon: phonon and Coulomb scattering. Appl. Phys. Lett. 94, 212103 (2009).
Li, W. Electrical transport limited by electron–phonon coupling from Boltzmann transport equation: an ab initio study of Si, Al, and MoS_{2}. Phys. Rev. B 92, 075405 (2015).
Fiorentini, M. & Bonini, N. Thermoelectric coefficients of ndoped silicon from first principles via the solution of the Boltzmann transport equation. Phys. Rev. B 94, 085204 (2016).
Zhou, J.J. & Bernardi, M. Ab initio electron mobility and polar phonon scattering in GaAs. Phys. Rev. B 94, 201201 (2016).
Poncé, S., Margine, E. R. & Giustino, F. Towards predictive manybody calculations of phononlimited carrier mobilities in semiconductors. Phys. Rev. B 97, 121201(R) (2018).
Macheda, F. & Bonini, N. Magnetotransport phenomena in pdoped diamond from first principles. Phys. Rev. B 98, 201201(R) (2018).
Sohier, T., Campi, D., Marzari, N. & Gibertini, M. Mobility of twodimensional materials from first principles in an accurate and automated framework. Phys. Rev. Mater. 2, 114010 (2018).
Ma, J., Nissimagoudar, A. S. & Li, W. Firstprinciples study of electron and hole mobilities of Si and GaAs. Phys. Rev. B 97, 045201 (2018).
Poncé, S., Schlipf, M. & Giustino, F. Origin of low carrier mobilities in halide perovskites. ACS Energy Lett. 4, 456–463 (2019).
Poncé, S., Jena, D. & Giustino, F. Hole mobility of strained GaN from first principles. Phys. Rev. B 100, 085204 (2019).
Poncé, S., Jena, D. & Giustino, F. Route to high hole mobility in GaN via reversal of crystalfield splitting. Phys. Rev. Lett. 123, 096602 (2019).
Lee, N.E., Zhou, J. J., Chen, H. Y. & Bernardi, M. Ab initio electrontwophonon scattering in GaAs from nexttoleading order perturbation theory. Nat. Commun. 11, 1607 (2020).
Poncé, S. & Giustino, F. Structural, electronic, elastic, power, and transport properties of βGa_{2}O_{3} from first principles. Phys. Rev. Res. 2, 033102 (2020).
Brunin, G. et al. Phononlimited electron mobility in Si, GaAs, and GaP with exact treatment of dynamical quadrupoles. Phys. Rev. B 102, 094308 (2020).
Brunin, G. et al. Electronphonon beyond Fröhlich: dynamical quadrupoles in polar and covalent solids. Phys. Rev. Lett. 125, 136601 (2020).
Macheda, F., Poncé, S., Giustino, F. & Bonini, N. Theory and computation of Hall scattering factor in graphene. Nano Lett. 20, 8861–8865 (2020).
Broyden, C. G. A class of methods for solving nonlinear simultaneous equations. Math. Comput. 19, 577 (1965).
Li, W., Carrete, J., Katcho, N. A. & Mingo, N. ShengBTE: a solver of the Boltzmann transport equation for phonons. Comput. Phys. Commun. 185, 1747–1758 (2014).
Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 77, 3865 (1996).
Hamann, D. R. Optimized normconserving Vanderbilt pseudopotentials. Phys. Rev. B 88, 085117 (2013).
van Setten, M. J. et al. The PseudoDojo: training and grading a 85 element optimized normconserving pseudopotential table. Comput. Phys. Commun. 226, 39–54 (2018).
Di Paola, C., Macheda, F., Laricchia, S., Weber, C. & Bonini, N. Firstprinciples study of electronic transport and structural properties of Cu_{12}Sb_{4}S_{13} in its hightemperature phase. Phys. Rev. Res. 2, 033055 (2020).
Ashcroft, N. W. & Mermin, N. D.Solid State Physics (Saunders College Publishing, 1976).
Lundstrom, M. Fundamentals of Carrier Transport 2nd edn (Cambridge University Press, 2000).
Brooks, H. Theory of the electrical properties of germanium and silicon. In Advances in Electronics and Electron Physics (ed., Marton, L.), Ch. 2, 85–182 (Academic Press, 1955).
Debye, P. P. & Conwell, E. M. Electrical properties of ntype germanium. Phys. Rev. 93, 693–706 (1954).
Restrepo, O. D., Varga, K. & Pantelides, S. T. Firstprinciples calculations of electron mobilities in silicon: phonon and Coulomb scattering. Appl. Phys. Lett. 94, 212103 (2009).
Graziosi, P., Kumarasinghe, C. & Neophytou, N. Material descriptors for the discovery of efficient thermoelectrics. ACS Appl. Energy Mater. 3, 5913–5926 (2020).
Ganose, A. M. et al. Efficient calculation of carrier scattering rates from first principles. Nat. Commun. 12, 2222 (2021).
Lu, I.T., Zhou, J.J., Park, J. & Bernardi, M. Firstprinciples ionizedimpurity scattering and charge transport in doped materials. Phys. Rev. Mater. 6, L010801 (2022).
Leveillee, J., Zhang, X., Kioupakis, E. & Giustino, F. Ab initio calculation of carrier mobility in semiconductors including ionizedimpurity scattering. Phys. Rev. B 107, 125207 (2023).
Kohn, W. & Luttinger, J. M. Quantum theory of electrical transport phenomena. Phys. Rev. 108, 590–611 (1957).
Lindhard, J. On the properties of a gas of charged particles. Dan. Mat. Fys. Medd. 28, 3–58 (1954).
Sanders, N. & Kioupakis, E. Phonon and defectlimited electron and hole mobility of diamond and cubic boron nitride: a critical comparison. Appl. Phys. Lett. 119, 062101 (2021).
Schlipf, M. & Gygi, F. Optimization algorithm for the generation of ONCV pseudopotentials. Comput. Phys. Commun. 196, 36 (2015).
Canali, C., Jacoboni, C., Nava, F., Ottaviani, G. & AlberigiQuaranta, A. Electron drift velocity in silicon. Phys. Rev. B 12, 2265–2284 (1975).
Morin, F. J. & Maita, J. P. Electrical properties of silicon containing arsenic and boron. Phys. Rev. 96, 28–35 (1954).
Jacoboni, C., Canali, C., Ottaviani, G. & Alberigi Quaranta, A. A review of some charge transport properties of silicon. Solid State Electron. 20, 77–89 (1977).
Caruso, F. & Giustino, F. Theory of electron–plasmon coupling in semiconductors. Phys. Rev. B 94, 115208 (2016).
Misiakos, K. & Tsamakis, D. Accurate measurements of the silicon intrinsic carrier density from 78 to 340 K. J. Appl. Phys. 74, 3293–3297 (1993).
McMillan, W. L. Transition temperature of strongcoupled superconductors. Phys. Rev. 167, 331 (1968).
Eliashberg, G. Interactions between electrons and lattice vibrations in a superconductor. Sov. Phys.JETP 11, 696 (1960).
Eliashberg, G. Temperature Green’s function for electrons in a superconductor. Sov. Phys.JETP 12, 1000 (1961).
Oliveira, L. N., Gross, E. K. U. & Kohn, W. Densityfunctional theory for superconductors. Phys. Rev. Lett. 60, 2430–2433 (1988).
Lüders, M. et al. Ab initio theory of superconductivity. I. Density functional formalism and approximate functionals. Phys. Rev. B 72, 024545 (2005).
Marques, M. A. L. et al. Ab initio theory of superconductivity. II. Application to elemental metals. Phys. Rev. B 72, 024546 (2005).
Sanna, A., Pellegrini, C. & Gross, E. K. U. Combining Eliashberg theory with density functional theory for the accurate prediction of superconducting transition temperatures and gap functions. Phys. Rev. Lett. 125, 057001 (2020).
Allen, P. B. & Dynes, R. C. Transition temperature of strongcoupled superconductors reanalyzed. Phys. Rev. B 12, 905–922 (1975).
Carbotte, J. P. Properties of bosonexchange superconductors. Rev. Mod. Phys. 62, 1027–1157 (1990).
Sanna, A. et al. Ab initio Eliashberg theory: making genuine predictions of superconducting features. J. Phys. Soc. Jpn. 87, 041012 (2018).
Xie, S. R. et al. Machine learning of superconducting critical temperature from Eliashberg theory. npj Comput. Mater. 8, 1–8 (2022).
Gor’kov, L. P. Microscopic derivation of the Ginzburg–Landau equations in the theory of superconductivity. Sov. Phys.JETP 7, 505 (1958).
Nambu, Y. Quasiparticles and gauge invariance in the theory of superconductivity. Phys. Rev. 117, 648–663 (1960).
Scalapino, D. J., Schrieffer, J. R. & Wilkins, J. W. Strongcoupling superconductivity. I. Phys. Rev. 148, 263–279 (1966).
Scalapino, D. J. The Electron–phonon Interaction and Strongcoupling Superconductors, Vol. 1 (Dekker, 1969).
Allen, P. B. & Mitrović, B. Theory of Superconducting T_{c}, Vol. 37 of Solid State Physics (Academic Press, 1983).
Choi, H. J., Cohen, M. L. & Louie, S. G. Anisotropic Eliashberg theory of MgB_{2}: T_{c}, isotope effects, superconducting energy gaps, quasiparticles, and specific heat. Phys. C: Supercond. 385, 66–74 (2003).
Marsiglio, F. & Carbotte, J. P. Electron–Phonon Superconductivity (Springer, 2008).
Margine, E. R. & Giustino, F. Anisotropic Migdal–Eliashberg theory using Wannier functions. Phys. Rev. B 87, 024505 (2013).
Marsiglio, F. Eliashberg theory: a short review. Ann. Phys. 417, 168102 (2020).
Migdal, A. Interaction between electrons and lattice vibrations in a normal metal. Sov. Phys.JETP 34, 996 (1958).
Hedin, L. New method for calculating the oneparticle Green’s function with application to the electrongas problem. Phys. Rev. 139, A796 (1965).
Hybertsen, M. S. & Louie, S. G. Electron correlation in semiconductors and insulators: Band gaps and quasiparticle energies. Phys. Rev. B 34, 5390 (1986).
Lee, K.H., Chang, K. J. & Cohen, M. L. Firstprinciples calculations of the coulomb pseudopotential μ^{*}: application to Al. Phys. Rev. B 52, 1425–1428 (1995).
Margine, E. R., Lambert, H. & Giustino, F. Electron–phonon interaction and pairing mechanism in superconducting Caintercalated bilayer graphene. Sci. Rep. 6, 1–6 (2016).
Aperis, A. & Oppeneer, P. M. Multiband fullbandwidth anisotropic Eliashberg theory of interfacial electron–phonon coupling and high T_{c} superconductivity in FeSe/SrTiO_{3}. Phys. Rev. B 97, 060501 (2018).
Allen, P. B. Fermisurface harmonics: a general method for nonspherical problems. application to Boltzmann and Eliashberg equations. Phys. Rev. B 13, 1416–1427 (1976).
Vidberg, H. & Serene, J. Solving the Eliashberg equations by means of Npoint Padé approximants. J. Low Temp. Phys. 29, 179–192 (1977).
Leavens, C. & Ritchie, D. Extension of the Npoint Padé approximants solution of the Eliashberg equations to T ~ T_{c}. Solid State Commun. 53, 137–142 (1985).
Marsiglio, F., Schossmann, M. & Carbotte, J. P. Iterative analytic continuation of the electron selfenergy to the real axis. Phys. Rev. B 37, 4965–4969 (1988).
Morel, P. & Anderson, P. W. Calculation of the superconducting state parameters with retarded electron–phonon interaction. Phys. Rev. 125, 1263–1271 (1962).
Heil, C. et al. Origin of superconductivity and latent charge density wave in NbS_{2}. Phys. Rev. Lett. 119, 087003 (2017).
Heil, C., di Cataldo, S., Bachelet, G. B. & Boeri, L. Superconductivity in sodalitelike yttrium hydride clathrates. Phys. Rev. B 99, 220502 (2019).
Di Cataldo, S., Heil, C., von der Linden, W. & Boeri, L. LaBH_{8}: towards highT_{c} lowpressure superconductivity in ternary superhydrides. Phys. Rev. B 104, L020511 (2021).
Pickett, W. E. Generalization of the theory of the electron–phonon interaction: thermodynamic formulation of superconducting and normalstate properties. Phys. Rev. B 26, 1186–1207 (1982).
Ummarino, G. A. C. Eliashberg Theory, Vol. 3 of Lecture Notes of the Autumn School Correlated Electrons (Forschungszentrum Jülich GmbH Institute for Advanced Simulation, 2013).
Davydov, A. et al. Ab initio theory of plasmonic superconductivity within the Eliashberg and densityfunctional formalisms. Phys. Rev. B 102, 214508 (2020).
Kafle, G. P. et al. Ab initio study of Li–Mg–B superconductors. Phys. Rev. Mater. 6, 084801 (2022).
Sano, W., Koretsune, T., Tadano, T., Akashi, R. & Arita, R. Effect of van Hove singularities on highT_{c} superconductivity in H_{3}S. Phys. Rev. B 93, 094525 (2016).
Schrodi, F., Oppeneer, P. M. & Aperis, A. Fullbandwidth Eliashberg theory of superconductivity beyond Migdal’s approximation. Phys. Rev. B 102, 024503 (2020).
Valla, T. et al. Quasiparticle spectra, chargedensity waves, superconductivity, and electronphonon coupling in 2HNbSe_{2}. Phys. Rev. Lett. 92, 086401 (2004).
Weber, F. et al. Extended phonon collapse and the origin of the chargedensity wave in 2HNbSe_{2}. Phys. Rev. Lett. 107, 107403 (2011).
Leroux, M. et al. Strong anharmonicity induces quantum melting of charge density wave in 2HNbSe_{2} under pressure. Phys. Rev. B 92, 140303(R) (2015).
Yokoya, T. et al. Fermi surface sheetdependent superconductivity in 2HNbSe_{2}. Science 294, 2518 (2001).
Anikin, A. et al. Ultrafast dynamics in the highsymmetry and in the charge density wave phase of 2HNbSe_{2}. Phys. Rev. B 102, 205139 (2020).
Sanna, A. et al. Realspace anisotropy of the superconducting gap in the chargedensity wave material 2HNbSe_{2}. npj Quantum Mater. 7, 6 (2022).
Klimeš, J., Bowler, D. R. & Michaelides, A. van der Waals density functionals applied to solids. Phys. Rev. B 83, 195131 (2011).
Klimeš, J., Bowler, D. R. & Michaelides, A. Chemical accuracy for the van der Waals density functional. J. Phys.: Condens. Matter 22, 022201 (2010).
Thonhauser, T. et al. Spin signature of nonlocal correlation binding in metalorganic frameworks. Phys. Rev. Lett. 115, 136402 (2015).
Thonhauser, T. et al. van der Waals density functional: selfconsistent potential and the nature of the van der Waals bond. Phys. Rev. B 76, 125112 (2007).
Methfessel, M. & Paxton, A. T. Highprecision sampling for Brillouinzone integration in metals. Phys. Rev. B 40, 3616 (1989).
Zheng, F. & Feng, J. Electronphonon coupling and the coexistence of superconductivity and chargedensity wave in monolayer NbSe_{2}. Phys. Rev. B 99, 161119 (2019).
Wickramaratne, D., Khmelevskyi, S., Agterberg, D. F. & Mazin, I. I. Ising superconductivity and magnetism in NbSe_{2}. Phys. Rev. X 10, 041003 (2020).
Das, S., Paudyal, H., Margine, E. R., Agterberg, D. F. & Mazin, I. I. Electron–phonon coupling and spin fluctuations in the Ising superconductor NbSe_{2}. npj Comput. Mater. 9, 66 (2023).
Alexandrov, A. S. & Devreese, J. T.Advances in Polaron Physics, Vol. 159 of Springer Series in SolidState Sciences (Springer, 2010).
Emin, D. Polarons (Cambridge University Press, 2012).
Franchini, C., Reticcioli, M., Setvin, M. & Diebold, U. Polarons in materials. Nat. Rev. Mater. 6, 560 (2021).
Deskins, N. A. & Dupuis, M. Electron transport via polaron hopping in bulk TiO_{2}: a density functional theory characterization. Phys. Rev. B 75, 195212 (2007).
Franchini, C., Kresse, G. & Podloucky, R. Polaronic hole trapping in doped BaBiO_{3}. Phys. Rev. Lett. 102, 256402 (2009).
Lany, S. & Zunger, A. Polaronic hole localization and multiple hole binding of acceptors in oxide widegap semiconductors. Phys. Rev. B 80, 085202 (2009).
Varley, J. B., Janotti, A., Franchini, C. & Van de Walle, C. G. Role of selftrapping in luminescence and ptype conductivity of widebandgap oxides. Phys. Rev. B 85, 081109(R) (2012).
Setvin, M. et al. Direct view at excess electrons in TiO_{2} rutile and anatase. Phys. Rev. Lett. 113, 086402 (2014).
Kokott, S., Levchenko, S. V., Rinke, P. & Scheffler, M. Firstprinciples supercell calculations of small polarons with proper account for longrange polarization effects. New J. Phys. 20, 033023 (2018).
Falletta, S. & Pasquarello, A. Polarons free from manybody selfinteraction in density functional theory. Phys. Rev. B 106, 125119 (2022).
Sio, W. H., Verdi, C., Poncé, S. & Giustino, F. Polarons from first principles, without supercells. Phys. Rev. Lett. 122, 246403 (2019).
Sio, W. H., Verdi, C., Poncé, S. & Giustino, F. Ab initio theory of polarons: Formalism and applications. Phys. Rev. B 99, 235139 (2019).
LafuenteBartolomé, J. et al. Unified approach to polarons and phononinduced band structure renormalization. Phys. Rev. Lett. 129, 076402 (2022).
LafuenteBartolomé, J. et al. Ab initio selfconsistent manybody theory of polarons at all couplings. Phys. Rev. B 106, 075119 (2022).
Maradudin, A. A. & Vosko, S. H. Symmetry properties of the normal vibrations of a crystal. Rev. Mod. Phys. 40, 1–37 (1968).
Davidson, E. R. The iterative calculation of a few of the lowest eigenvalues and corresponding eigenvectors of large realsymmetric matrices. J. Comput. Phys. 17, 87–94 (1975).
Makov, G. & Payne, M. C. Periodic boundary conditions in ab initio calculations. Phys. Rev. B 51, 4014–4022 (1995).
Castner, T. G. & Känzig, W. The electronic structure of Vcenters. J. Phys. Chem. Solids 3, 178–195 (1957).
Jette, A. N., Gilbert, T. L. & Das, T. P. Theory of the selftrapped hole in the alkali halides. Phys. Rev. 184, 884–894 (1969).
Fröhlich, H., Pelzer, H. & Zienau, S. XX. Properties of slow electrons in polar materials. Lond. Edinb. Dublin Philos. Mag. J. Sci. 41, 221–242 (1950).
Holstein, T. Studies of polaron motion: Part I. The molecularcrystal model. Ann. Phys. 8, 325–342 (1959).
Bassani, F. & Parravicini, G. P. Electronic States and Optical Transitions in Solids (Pergamon Press, 1975).
Noffsinger, J., Kioupakis, E., Van de Walle, C. G., Louie, S. G. & Cohen, M. L. Phononassisted optical absorption in silicon from first principles. Phys. Rev. Lett. 108, 167402 (2012).
Giustino, F. Materials Modelling Using Density Functional Theory: Properties and Predictions (Oxford University Press, 2014).
Tiwari, S. & Giustino, F. Abinitio manybody calculations of optical absorption spectra including both direct and indirect phononassisted processes (2023) (submitted).
Levine, Z. H. & Allan, D. C. Quasiparticle calculation of the dielectric response of silicon and germanium. Phys. Rev. B 43, 4187–4207 (1991).
Rohlfing, M. & Louie, S. G. Electronhole excitations and optical spectra from first principles. Phys. Rev. B 62, 4927–4944 (2000).
Song, B. et al. Optical properties of cubic boron arsenide. Appl. Phys. Lett. 116, 141903 (2020).
Zhang, X. & Kioupakis, E. Phononassisted optical absorption of SiC polytypes from first principles. Phys. Rev. B 107, 115207 (2023).
Zhang, X., Shi, G., Leveillee, J. A., Giustino, F. & Kioupakis, E. Ab initio theory of freecarrier absorption in semiconductors. Phys. Rev. B 106, 205203 (2022).
Mo, Y., Tang, H., Bansil, A. & Tao, J. Accurate lattice geometrical parameters and bulk moduli from a semilocal density functional. AIP Adv. 8, 095209 (2018).
Haas, P., Tran, F. & Blaha, P. Calculation of the lattice constant of solids with semilocal functionals. Phys. Rev. B 79, 085104 (2009).
Favot, F. & Dal Corso, A. Phonon dispersions: Performance of the generalized gradient approximation. Phys. Rev. B 60, 11427–11431 (1999).
Deslippe, J. et al. BerkeleyGW: a massively parallel computer package for the calculation of the quasiparticle and optical properties of materials and nanostructures. Comput. Phys. Commun. 183, 1269–1289 (2012).
Sze, S. M., Li, Y. & Ng, K. K. Physics and Properties of Semiconductors—A Review, Ch. 1, 5–75 (Wiley, 2006).
Wachs, A. L., Miller, T., Hsieh, T. C., Shapiro, A. P. & Chiang, T. C. Angleresolved photoemission studies of Ge(111)c(2 × 8), Ge(111)(1 × 1)H, Si(111)(7 × 87), and Si(100)(2 × 1). Phys. Rev. B 32, 2326–2333 (1985).
Straub, D., Ley, L. & Himpsel, F. J. Conductionband and surfacestate critical points in Si: an inversephotoemission study. Phys. Rev. Lett. 54, 142–145 (1985).
Madelung, O. Semiconductors: Group IV Elements and III–V Compounds (Springer, 1991).
Petretto, G. et al. Highthroughput densityfunctional perturbation theory phonons for inorganic materials. Sci. data 5, 1–12 (2018).
Dolling, G. Inelastic scattering of neutrons in solids and liquids. IAEA 2, 37 (1963).
Nilsson, G. & Nelin, G. Study of the homology between silicon and germanium by thermalneutron spectrometry. Phys. Rev. B 6, 3777–3786 (1972).
Schinke, C. et al. Uncertainty analysis for the coefficient of bandtoband absorption of crystalline silicon. AIP Adv. 5, 067168 (2015).
Antonius, G., Poncé, S., Boulanger, P., Côté, M. & Gonze, X. Manybody effects on the zeropoint renormalization of the band structure. Phys. Rev. Lett. 112, 215501 (2014).
Zacharias, M. & Giustino, F. Oneshot calculation of temperaturedependent optical spectra and phononinduced bandgap renormalization. Phys. Rev. B 94, 075125 (2016).
Zacharias, M. & Giustino, F. Theory of the special displacement method for electronic structure calculations at finite temperature. Phys. Rev. Res. 2, 013357 (2020).
Zacharias, M. et al. Multiphonon diffuse scattering in solids from first principles: Application to layered crystals and twodimensional materials. Phys. Rev. B 104, 205109 (2021).
Brüesch, P. Phonons, Theory and Experiments, Vol. 1 (Springer, 1982).
Zacharias, M. & Kelires, P. C. Temperature dependence of the optical properties of silicon nanocrystals. Phys. Rev. B 101, 245122 (2020).
Zacharias, M. & Kelires, P. C. Quantum confinement of electron–phonon coupling in graphene quantum dots. J. Phys. Chem. Lett. 12, 9940–9946 (2021).
Ramírez, R., Herrero, C. P. & Hernández, E. R. Pathintegral molecular dynamics simulation of diamond. Phys. Rev. B 73, 245202 (2006).
Allen, P. B. & Heine, V. Theory of the temperature dependence of electronic band structures. J. Phys. C: Solid State Phys. 9, 2305 (1976).
Popescu, V. & Zunger, A. Extracting E versus\(\overrightarrow{k}\) effective band structure from supercell calculations on alloys and impurities. Phys. Rev. B 85, 085201 (2012).
Zacharias, M. et al. Efficient firstprinciples methodology for the calculation of the allphonon inelastic scattering in solids. Phys. Rev. Lett. 127, 207401 (2021).
Perdew, J. P. & Zunger, A. Selfinteraction correction to densityfunctional approximations for manyelectron systems. Phys. Rev. B 23, 5048–5079 (1981).
Ceperley, D. M. & Alder, B. J. Ground state of the electron gas by a stochastic method. Phys. Rev. Lett. 45, 566–569 (1980).
Giustino, F., Louie, S. G. & Cohen, M. L. Electron–phonon renormalization of the direct band gap of diamond. Phys. Rev. Lett. 105, 265501 (2010).
Varshni, Y. P. Temperature dependence of the energy gap in semiconductors. Physica 34, 149–154 (1967).
Karsai, F., Engel, M., Kresse, G. & FlageLarsen, E. Electronphonon coupling in semiconductors within the GW approximation. N. J. Phys. 20, 123008 (2018).
Poncé, S. et al. Temperature dependence of the electronic structure of semiconductors and insulators. J. Chem. Phys. 143, 102813 (2015).
Nery, J. P. & Allen, P. B. Influence of Fröhlich polaron coupling on renormalized electron bands in polar semiconductors: results for zincblende GaN. Phys. Rev. B 94, 115135 (2016).
Kang, Y., Peelaers, H., Krishnaswamy, K. & Van de Walle, C. G. Firstprinciples study of direct and indirect optical absorption in BaSnO_{3}. Appl. Phys. Lett 112, 062106 (2018).
Lee, H. et al. EPW Project (2023) (last retrieved 02062023) https://epwcode.org.
Engel, M. et al. Zeropoint renormalization of the band gap of semiconductors and insulators using the projector augmented wave method. Phys. Rev. B 106, 094316 (2022).
Bushick, K. & Kioupakis, E. Phononassisted AugerMeitner recombination in silicon from first principles. Phys. Rev. Lett. 131, 076902 (2023).
Zacharias, M., Volonakis, G., Giustino, F. & Even, J. Anharmonic lattice dynamics via the special displacement method. Phys. Rev. B 108, 035155 (2023).
Sio, W. H. & Giustino, F. Unified ab initio description of fröhlich electronphonon interactions in twodimensional and threedimensional materials. Phys. Rev. B 105, 115414 (2022).
Poncé, S., Royo, M., Stengel, M., Marzari, N. & Gibertini, M. Longrange electrostatic contribution to electron–phonon couplings and mobilities of twodimensional and bulk materials. Phys. Rev. B 107, 155424 (2023).
Li, Z., Antonius, G., Wu, M., da Jornada, F. H. & Louie, S. G. Electron–phonon coupling from ab initio linearresponse theory within the GW method: correlationenhanced interactions and superconductivity in Ba_{1−x}K_{x}BiO_{3}. Phys. Rev. Lett. 122, 186402 (2019).
Gonze, X. et al. The Abinit project: Impact, environment and recent developments. Comput. Phys. Commun. 248, 107042 (2020).
Kresse, G. & Hafner, J. Ab initio molecular dynamics for liquid metals. Phys. Rev. B 47, 558–561 (1993).
Soler, J. M. et al. The SIESTA method for ab initio orderN materials simulation. J. Phys.: Condens. Matter 14, 2745 (2002).
Enkovaara, J. et al. Electronic structure calculations with GPAW: a realspace implementation of the projector augmentedwave method. J. Phys.: Condens. Matter 22, 253202 (2010).
Gamma, E., Helm, R., Johnson, R. & Vlissides, J. Design Patterns, Elements of Reusable ObjectOriented Software (AddisonWesley, 1995).
IsmailBeigi, S. & Arias, T. A. New algebraic formulation of density functional calculation. Comput. Phys. Commun. 128, 1–45 (2000).
Pizzi, G. et al. Wannier90 as a community code: new features and applications. J. Phys.: Condens. Matter. 32, 165902 (2019).
Vitale, V. et al. Automated highthroughput wannierisation. npj Comput. Mater. 6, 66 (2020).
Qiao, J., Pizzi, G. & Marzari, N. Projectability disentanglement for accurate and automated electronicstructure Hamiltonians. Preprint athttps://doi.org/10.48550/arXiv.2303.07877 (2023).
Agapito, L. A., Ferretti, A., Calzolari, A., Curtarolo, S. & Buongiorno Nardelli, M. Effective and accurate representation of extended Bloch states on finite Hilbert spaces. Phys. Rev. B 88, 165127 (2013).
Damle, A., Lin, L. & Ying, L. Compressed representation of Kohn–Sham orbitals via selected columns of the density matrix. J. Chem. Theory Comput. 11, 1463–1469 (2015).
Damle, A. & Lin, L. Disentanglement via entanglement: a unified method for wannier localization. Multiscale Model. Simul. 16, 1392–1410 (2018).
Yates, J. R., Wang, X., Vanderbilt, D. & Souza, I. Spectral and Fermi surface properties from Wannier interpolation. Phys. Rev. B 75, 195121 (2007).
Giannozzi, P., de Gironcoli, S., Pavone, P. & Baroni, S. Ab initio calculation of phonon dispersions in semiconductors. Phys. Rev. B 43, 7231–7242 (1991).
Vogl, P. Microscopic theory of electron–phonon interaction in insulators or semiconductors. Phys. Rev. B 13, 694–704 (1976).
Fröhlich, H. Electrons in lattice fields. Adv. Phys. 3, 325–361 (1954).
Verdi, C. & Giustino, F. Fröhlich electron–phonon vertex from first principles. Phys. Rev. Lett. 115, 176401 (2015).
Sjakste, J., Vast, N., Calandra, M. & Mauri, F. Wannier interpolation of the electronphonon matrix elements in polar semiconductors: polaroptical coupling in GaAs. Phys. Rev. B 92, 054307 (2015).
Marzari, N. & Vanderbilt, D. Maximally localized generalized Wannier functions for composite energy bands. Phys. Rev. B 56, 12847–12865 (1997).
Souza, I., Marzari, N. & Vanderbilt, D. Maximally localized Wannier functions for entangled energy bands. Phys. Rev. B 65, 035109 (2001).
Jhalani, V. A., Zhou, J.J., Park, J., Dreyer, C. E. & Bernardi, M. Piezoelectric electron–phonon interaction from ab initio dynamical quadrupoles: Impact on charge transport in wurtzite GaN. Phys. Rev. Lett. 125, 136602 (2020).
Park, J., Zhou, J.J., Jhalani, V. A., Dreyer, C. E. & Bernardi, M. Longrange quadrupole electron–phonon interaction from first principles. Phys. Rev. B 102, 125203 (2020).
Royo, M. & Stengel, M. Firstprinciples theory of spatial dispersion: dynamical quadrupoles and flexoelectricity. Phys. Rev. X 9, 021050 (2019).
Gonze, X. et al. Recent developments in the ABINIT software package. Comput. Phys. Commun. 205, 106–131 (2016).
Poncé, S., Royo, M., Gibertini, M., Marzari, N. & Stengel, M. Accurate prediction of Hall mobilities in twodimensional materials through gaugecovariant quadrupolar contributions. Phys. Rev. Lett. 130, 166301 (2023).
Sohier, T., Calandra, M. & Mauri, F. Twodimensional Fröhlich interaction in transitionmetal dichalcogenide monolayers: theoretical modeling and firstprinciples calculations. Phys. Rev. B 94, 085415 (2016).
Deng, T. et al. Ab initio dipolar electron–phonon interactions in twodimensional materials. Phys. Rev. B 103, 075410 (2021).
Sio, W. H. & Giustino, F. Polarons in twodimensional atomic crystals. Nat. Phys. 19, 629 (2023).
Verdi, C. & Giustino, F. Fröhlich electron–phonon vertex from first principles. Phys. Rev. Lett. 115, 176401 (2015).
Towns, J. et al. XSEDE: accelerating scientific discovery. Comput. Sci. Eng. 16, 62–74 (2014).
Ottaviani, G., Reggiani, L., Canali, C., Nava, F. & AlberigiQuaranta, A. Hole drift velocity in silicon. Phys. Rev. B 12, 3318–3329 (1975).
Kawamura, M. Fermisurfer: Fermisurface viewer providing multiple representation schemes. Comput. Phys. Commun. 239, 197–203 (2019).
Spicer, W. E. & Eden, R. C. Photoemission investigation of the band structure of semiconductors. In Proc. Ninth International Conference on the Physics of Semiconductors, Vol. 1 (ed. Ryvkin, S. M.) p. 65 (Nauka Publishers, Moscow, 1968).
Zucca, R. R. L. & Shen, Y. R. Wavelengthmodulation spectra of some semiconductors. Phys. Rev. B 1, 2668–2676 (1970).
Aspnes, D. E. & Studna, A. A. Dielectric functions and optical parameters of Si, Ge, GaP, GaAs, GaSb, InP, InAs, and InSb from 1.5 to 6.0 eV. Phys. Rev. B 27, 985–1009 (1983).
Tiedje, T., Yablonovitch, E., Cody, G. & Brooks, B. Limiting efficiency of silicon solar cells. IEEE Trans. Electron Devices 31, 711–716 (1984).
Green, M. A. & Keevers, M. J. Optical properties of intrinsic silicon at 300 K. Prog. Photovolt. 3, 189–192 (1995).
Kim, H. J. et al. Physical properties of transparent perovskite oxides (Ba,La)SnO_{3} with high electrical mobility at room temperature. Phys. Rev. B 86, 165205 (2012).
Lee, H. et al. EPW v6alpha (last accessed 5 November 2023). https://docs.epwcode.org/doc/DownloadAndInstall.html (2023).
Acknowledgements
This research is supported by: the Computational Materials Sciences Program funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award No. DESC0020129 (project coordination, scaleup, polaron module, transport module, optics module, special displacement module); the National Science Foundation, Office of Advanced Cyberinfrastructure and Division of Materials Research under Grants Nos. 2103991 and 2035518 (superconductivity module, interoperability); the NSF Characteristic Science Applications for the Leadership Class Computing Facility program under Grant No. 2139536 (preparation for LCCF); the Fond National de la Recherche Scientifique of Belgium (F.R.S.FNRS) and the European Union’s Horizon 2020 research and innovation program under grant agreements No. 881603Graphene Core3 (transport module); the NSF DMREF award 2119555 (quasidegenerate perturbation theory module). This research used resources of the National Energy Research Scientific Computing Center and the Argonne Leadership Computing Facility, which are DOE Office of Science User Facilities supported by the Office of Science of the U.S. Department of Energy, under Contracts No. DEAC0205CH11231 and DEAC0206CH11357, respectively. The authors acknowledge the Texas Advanced Computing Center (TACC) at The University of Texas at Austin for providing access to Frontera, Lonestar6, and Texascale Days, which have contributed to the research results reported within this paper (http://www.tacc.utexas.edu); the Extreme Science and Engineering Discovery Environment (XSEDE)^{218} which is supported by National Science Foundation grant number ACI1548562, and in particular Expanse at the San Diego Supercomputer Center through allocation TGDMR180071. S.P. acknowledges computational resources provided by the PRACE award granting access to Discoverer in SofiaTech, Bulgaria (OptoSpin project id. 2020225411), and by the Consortium des Équipements de Calcul Intensif (CÉCI), funded by the FRSFNRS under Grant No. 2.5020.11 and by the Walloon Region, as well as computational resources awarded on the Belgian share of the EuroHPC LUMI supercomputer. K.B. acknowledges the support of the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research, Department of Energy Computational Science Graduate Fellowship under Award Number DESC0020347.
The authors wish to thank Zhenbang Dai, Nikolaus Kandolf, VietAnh Ha, and Amanda Wang for their contributions to the EPW project that are not discussed in this manuscript; John Cazes and Hang Liu at TACC for their support with the Characteristic Science Applications project, Paolo Giannozzi for his support with Quantum ESPRESSO; and Stefano Baroni for fruitful discussions.
Author information
Authors and Affiliations
Contributions
H.L. coordinated the development of the code, developed hybrid parallelism, and performed scaling tests. S.P. codeveloped the transport module and maintained the test farm. K.B. codeveloped the indirect absorption module. S.H. developed the fullband anisotropic Eliashberg module. J.L.B. tested and expanded the polaron module. J.L. developed the ionizedimpurity scattering module. C.L. codeveloped the polaron module. J.M.L. developed the Wannierfunction perturbation theory module. F.M. codeveloped the transport module. H.M. developed the intermediate representation for the superconducting module. H.P. tested and expanded the anisotropic Eliashberg module. W.S. developed the initial version of the polaron module. S.T. developed the quasidegenerate perturbation theory module. M.Z. developed the special displacement module. X.Z. codeveloped the indirect absorption module. N.B. supervised the development of the transport module. E.K. supervised the development of the indirect absorption module. E.R.M. supervised the development of the superconductivity module. F.G. supervised the project. All authors contributed to the writing of the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
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.
Rights and permissions
Open Access This article is licensed under a Creative Commons 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 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 indicated otherwise in a credit line to the material. If material is not included in the 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 from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Lee, H., Poncé, S., Bushick, K. et al. Electron–phonon physics from first principles using the EPW code. npj Comput Mater 9, 156 (2023). https://doi.org/10.1038/s41524023011073
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41524023011073
This article is cited by

A momentumresolved view of polaron formation in materials
npj Computational Materials (2024)

Straininduced activation of chiralphonon emission in monolayer WS2
npj 2D Materials and Applications (2024)

Temperature and quantum anharmonic lattice effects on stability and superconductivity in lutetium trihydride
Nature Communications (2024)

Fullbandwidth anisotropic MigdalEliashberg theory and its application to superhydrides
Communications Physics (2024)

Accelerating the calculation of electron–phonon coupling strength with machine learning
Nature Computational Science (2024)