An overview of the EPH code¶
This page provides a quick introduction to the new electronphonon driver integrated with the ABINIT executable. We discuss important technical details related to the implementation and the associated input variables. The drawbacks/advantages with respect to the implementation available in ANADDB are also discussed.
Why a new EPH code?¶
First of all, let’s try to answer the question: why did you decide to implement a new code for electronphonon calculations?
As you may know, eph calculations have been available through the ANADDB executable for a long time. This ANADDBbased implementation is essentially a postprocessing of the eph matrix elements computed by the DFPT code. This approach presents advantages as well as drawbacks. On the one hand, most of the work required to compute eph matrix elements is implemented directly by the DFPT routines. This means that eph calculations with advanced features such as PAW, SOC, noncollinear magnetism etc are readily available in the ANADDB version once support in the DFPT part is implemented.
On the other hand, this postprocessing approach implies that the number of \kk/\qqpoints in the eph matrix elements is automatically fixed at the level of the DFPT run. In other words, if you want to compute phononlimited mobilities with e.g. a 90×90×90 \kk and \qqmesh, you need to perform DFPT calculations with the same sampling thus rendering the computation almost impossible. In principle, it is possible to use tricks such as a linear interpolation of the eph matrix elements to densify the sampling, but in order to get a decent interpolation one usually needs initial BZ meshes that are significantly denser than the ones needed to converge the DFPT part alone.
As a matter of fact, electrons, phonons and eph properties present completely different convergence rates. In silicon, for instance, a 9×9×9 mesh both for phonons and electrons is enough to converge the electron density and the vibrational spectrum [Petretto2018]. On the contrary, the phononlimited mobility of Si requires a 45×45×45 \kkgrid and a 90×90×90 \qqgrid to reach a 5% relative error [Brunin2020b]. Roughly speaking, an explicit computation of phonons in Si with a 90×90×90 \qqmesh requires around 20000 × 3 × natom DFPT calculations so you can easily get an idea of the cost of a fully abinitio evaluation of the eph matrix elements for all these \qqpoints.
The EPH code bypasses this bottleneck by interpolating the DFPT potentials in \qqspace while Bloch states are computed nonselfconsistently on arbitrarily dense \kkmeshes. As a net result, the three problems (electrons, phonons and electronphonon) are now partly decoupled and can be converged separately. Keep in mind, however, that the fact that one can easily densify the \qqsampling in the EPH code does not mean that one can use underconverged values for the groundstate (GS) and DFPT parts. Indeed, the quality of the interpolation depends on the initial \kk and \qqmeshes. The takehome message is that one should always converge carefully both electronic and vibrational properties before moving to EPH computations.
For further information about the difference between EPH and ANADDB, see also [Gonze2019]. Further details about the EPH implementation are available in [Brunin2020b].
EPH workflow¶
A typical EPH workflow with arrows denoting dependencies between the different steps is schematically represented in the figure below:
The brown boxes are standard DFPT calculations done with relatively coarse \kk and \qqmeshes (for instance, 9×9×9 in Si). Each DFPT run produces a (partial) DDB file with a portion of the full dynamical matrix as well as POT files with the firstorder derivative of the KS potential (referred to as the DFPT potential below). The partial POT files are merged with the mrgdv utility to produce a single DVDB file (Derivatives of V(\rr) DataBase). As usual, the partial DDB files are merged with mrgddb (see the second tutorial on DFPT).
The EPH driver (blue box) receives in input the total DDB and the DVDB as well as a GS WFK file that is usually produced with a different \kkmesh (in some cases, even with a different number of bands when highenergy empty states are needed). These ingredients are then used to compute (interpolate) eph matrix elements and the associated physical properties.
The EPH calculation is activated by using optdriver = 7 while eph_task defines the physical properties to be computed. To read the external files, one specifies the filepath with the three variables: getwfk_filepath, getddb_filepath and getdvdb_filepath.
Internally, the code starts by reading the DDB file to construct the interatomic force constants (IFCs) in the realspace supercell (\RRspace). Then, the EPH code computes phonon bands and DOS. Finally a specialized routine is invoked depending on the value of eph_task.
The following physical properties can be computed:

Imaginary part of phe selfenergy in metals (eph_task 1) that gives access to:
 Phonon linewidths induced by eph coupling
 Eliashberg function
 Superconducting properties within the isotropic MigdalEliashberg formalism

Real and imaginary parts of the eph selfenergy (eph_task 4) that gives access to:
 Zeropoint renormalization of the band gap
 QP corrections due to eph scattering as a function of T
 Spectral function A(\ww) and Eliashberg functions

Imaginary part of the eph selfenergy at the KS energy (eph_task 4) that gives access to:
 Phononlimited carrier mobility, electrical conductivity and Seebeck coefficient
 Phononlimited carrier mean free path and relaxation times (or scattering rates)
 All the calculations above can be done as a function of temperature and doping, for nonpolar and polar materials.
At the time of writing (November 11, 2020), the following features are not yet supported by EPH:
 PAW calculations
 Spinorbit coupling
 Noncollinear magnetism (nspinor = 2 and nspden = 4)
 Nonlocal part of the pseudopotential applied with useylm = 1
In this introduction, we focus on the parts that are common to the different subdrivers i.e.:
 Computation of vibrational properties via Fourier interpolation of the dynamical matrix.
 Fourier interpolation of the DFPT potentials.
The use of the different subdrivers is discussed in more detail in the specialized lessons:
Phonon bands and DOS with EPH¶
Since phonon frequencies and displacements are needed for eph calculations, it is not surprising that some of the ANADDB features related to the treatment of the dynamical matrix are integrated in the EPH code as well. In many cases, the variables names are the same in EPH and ANADDB especially for important variables such as dipdip, asr, and chneut. There are however important differences with respect to the ANADDB input file. More specifically, in EPH the name of the DDB file is specified by getddb_filepath whereas the DFPT \qqmesh associated to the DDB file is given by ddb_ngqpt. These two variables are mandatory when performing EPH calculations.
Important
In Abinit9 the default values of dipdip, asr, and chneut have been changed. In particular, the acoustic sum rule and the charge neutrality of the Born effecive charges are enforced by default. It is responsability of the user to check whether the breaking of these sum rules (always present due to numerical inaccuracies) is reasonable. By the same token, make sure that no vibrational instabilty is present before embarking on big EPH calculations. If the spectrum presents instabilities around \Gamma due to a Fourier interpolation done with coarse \qqsampling, you may try to use rifcsph.
Variables for phonon DOS¶
By default, the EPH code computes the phonon DOS and the atomprojected PHDOS by interpolating the IFCs on the dense \qqmesh specified by ph_ngqpt. The default \qqgrid is 20×20×20. You may want to increase this value for more accurate results. The step of the (linear) frequency mesh is governed by ph_wstep. The linear tetrahedron method by [Bloechl1994] is used by default. The Gaussian method can be activated via ph_intmeth with ph_smear defining the Gaussian smearing (in Hartree units by default). The final results are stored in the PHDOS.nc file (same format at the one produced by ANADDB). The computation of the PHDOS can be disabled by setting prtphdos = 0 to make the calculation a bit faster.
To summarize: to compute the PHDOS with the tetrahedron method and a 40x40x40 \Gammacentered \qqmesh, one should use
prtphdos 1 # 0 to disable this part. ph_ngqpt 40 40 40
To visualize the results with AbiPy , execute
abiopen.py out_PHDOS.nc e
Variables for phonon band structures¶
The computation of the phonon band structure is activated automatically provided the input file defines the highsymmetry \qqpath in terms of ph_nqpath vertices listed in the ph_qpath array. The ph_ndivsm variable defines the number of divisions used to sample the smallest segment of the path so that the number of points in each segment is proportional to the length of the segment. The computation of the phonon band structure can be deactivated by setting prtphbands = 0. The final results are stored in the PHBST.nc file (same format at the one produced by ANADDB).
For instance, the section for the phonon band structure could look like
prtphbands 1 ph_nqpath 5 ph_qpath 0.0 0.0 0.0 # Gamma 1/2 0.0 1/2 # X 1/2 1/4 3/4 # W 3/8 3/8 3/4 # K 0.0 0.0 0.0 # Gamma
To obtain the list of highsymmetry qpoints, one can use the abistruct script provided by AbiPy :
abistruct.py kpath in_DDB
and the phonon bands can be visualized using
abiopen.py out_PHBST.nc e
Electronphonon matrix elements¶
The eph matrix elements \gkq are defined by
where \psi_{n\kk} is the KS Bloch state and \Delta_\qnu V^\KS is the firstorder variation of the selfconsistent KS potential induced by the phonon mode \qnu. The scattering potential can be expressed as:
where \Delta_{\qq\nu} v^\KS(\rr) is a lattice periodic function [Giustino2017]. Note that ABINIT computes the response to the atomic perturbation defined by the three variables qpt, rfdir and rfatpol when rfphon is set to 1. The connection between the phonon representation and the atomic perturbation employed by the DFPT code is given by:
where {e}_{\kappa\alpha,\nu}(\qq) is the \alphath Cartesian component of the phonon eigenvector for the atom \kappa in the unit cell, M_\kappa its atomic mass and \partial_{\kappa\alpha,\qq} v^\KS(\rr) is the firstorder derivative of the KS potential that can be obtained with DFPT by solving selfconsistently a system of Sternheimer equations for a given (\kappa\alpha, \qq) perturbation [Gonze1997] [Baroni2001].
The DVDB file stores \partial_{\kappa\alpha,\qq} v^\KS(\rr) for all the \qqpoints in the IBZ and all the irreducible atomic perturbations. More rigorously, we should say that the DVDB file stores the local part of the DFPT potential (variation of the Hartree + XC + local part of the pseudo) but this is a rather technical point discussed in more detail in [Brunin2020b] that is not relevant for the present discussion so we do not elaborate more on this.
Important
The number of irreducible atomic perturbations depends on the \qqpoint and the symmetries of the system. Actually only a particular subset of the point group of the unperturbed crystal can be exploited at the DFPT level. Fortunately, you don’t have to worry about these technical details as symmetries are fully supported in EPH. Just run the DFPT calculation for the irreducible perturbations in the IBZ as usual. Also the computation of the WFK can be limited to the IBZ. The EPH code will employ symmetries to reconstruct the different quantities at runtime.
Fourier interpolation of the DFPT potentials¶
The EPH code employs the Fourier interpolation proposed in [Eiguren2008] to obtain the scattering potentials at arbitrary \qqpoints. In this interpolation technique, one uses the DFPT potential stored in the DVDB file to compute the \qq \rightarrow \RR Fourier transform
where the sum is over the N_\qq BZ \qqpoints belonging to the abinitio ddb_ngqpt grid. Once W_{\kappa\alpha}(\rr,\RR) is known, it is possible to interpolate the potential at an arbitrary point \tilde{\qq} using the inverse transform
where the sum is over the lattice vectors inside the Bornvon Karman supercell. The algorithm used to define the \RR points of the supercell with the corresponding weights is specified by dvdb_rspace_cell.
The accuracy of the interpolation depends on the localization of W_{\kappa\alpha} in \RRspace. This means that the Bornvon Karman supercell corresponding to the ddb_ngqpt grid should be large enough to capture the spatial decay of W_{\kappa\alpha}(\rr,\RR) as a function of \RR. As a consequence, ddb_ngqpt should be subject to convergence studies.
In metals, W_{\kappa\alpha} is expected to be shortranged provided one ignores possible Kohn anomalies. On the contrary, a special numerical treatment is needed in semiconductors and insulators due to the presence of longranged (LR) dipolar and quadrupolar fields in \RRspace. These LR terms determine a nonanalytic behaviour of the scattering potentials in the longwavelength limit \qq \rightarrow 0 [Vogl1976]. To handle the LR part, the EPH code uses an approach that is similar in spirit to the one employed for the Fourier interpolation of the dynamical matrix [Gonze1997].
The idea is relatively simple. One subtracts the LR part from the DFPT potentials before computing Eq. \eqref{eq:dfpt_pot_realspace} thus making the realspace representation amenable to Fourier interpolation. The nonanalytical part is then restored back to Eq. \eqref{eq:dfpt_pot_interpolation} when interpolating the potential at \tilde{\qq}.
In polar materials, the leading term is given by the dipolar field [Verdi2015], [Sjakste2015]:
where {\bm{\tau}}_\kappa is the atom position, \Omega is the volume of the unit cell, \bm{Z}^* and {\bm{\varepsilon}}^\infty are the Born effective charge tensor and the dielectric tensor, respectively, and summation over the Cartesian directions \beta is implied. This term is present in polar materials (nonzero \bm{Z}^*) and diverges as 1/q for \qq \rightarrow 0 hence the singularity is integrable in 3D systems. \bm{Z}^* and {\bm{\varepsilon}}^\infty are read automatically from the DDB file (if present) so we strongly recommend to compute these quantities with DFPT in order to prepare an EPH calculation in semiconductors.
In nonpolar materials, the Born effective charges are zero but the scattering potentials are still nonanalytic due to presence of jump discontinuities. As discussed in [Brunin2020], the nonanalytic behaviour can be fully captured by using the more general expression:
Important
The computation of the dynamical quadrupoles tensor within the DFPT framework will be made available in a future release, together with a specific tutorial. Once it is computed and stored in the DDB, the EPH code reads it automatically and uses it for the LR model.
In the implementation, each Fourier component is multiplied by the Gaussian filter e^{\frac{\qG^2}{4\alpha}} in order to damp the short range components that are not relevant for the definition of the LR model. The \alpha parameter can be specified via the dvdb_qdamp input variable. The default value worked well in the majority of the systems investigated so far yet this parameter should be subject to convergence tests. For testing purposes, it is possible to deactivate the treatment of the LR part by setting dvdb_add_lr to 0.
On the importance of the initial DFPT mesh¶
At this point it is worth commenting about the importance of the initial DFPT \qqmesh. The Fourier interpolation implicitly assumes that the signal in \RRspace decays quickly hence the quality of the interpolated phonon frequencies and of the interpolated DFPT potentials between the abinitio points depends on the spacing of the initial \qqmesh that in turns defines the size of the BornvonKarman supercell. In other words, the denser the DFPT mesh, the larger the realspace supercell and the better the interpolation, especially for \qqpoints close to \Gamma.
From a more practical point of view, this implies that one should always monitor the convergence of the physical properties with respect to the initial DFPT \qqmesh. The LR model implemented in ABINIT facilitates the convergence as the nonanalytic behaviour for \qq \rightarrow 0 is properly described yet the Fourier interpolation can introduce oscillations between the abinitio \qqpoints and these oscillations may affect the quality of the physical results [Brunin2020].
Tricks to accelerate the computation and reduce the memory requirements¶
Each subdriver implements specialized techniques to accelerate the calculation and reduce the memory requirements. Here we focus on the techniques that are common to the different EPH subdrivers. Additional tricks specific to the particular value of eph_task are discussed in more detail in the associated lessons.
First of all, note that the memory requirements for the W_{\kappa\alpha}(\rr,\RR) array scales as %nfft × product(ddb_ngqpt). This quantity should be multiplied by 3 * natom if each MPI process stores all the perturbations in memory. The MPI parallelism over perturbations (see eph_np_pqbks) allows one to decrease this prefactor from 3 × natom to 3 × natom / nprocs_pert.
Also, the total number of \rrpoints (%nfft) plays an important role both at the level of memory as well as the level of the walltime. To optimize this part, one can decrease the value of boxcutmin to a value smaller than 2 e.g. 1.5 or the more aggressive 1.1. Note that one is not obliged to run the GS/DFPT part with the same boxcutmin. The EPH code will automatically interpolate the DFPT potentials if the input FFT mesh computed from ecut and boxcutmin differs from the one found in the DVDB file. Just to clarify, you can change the value of boxcutmin in the EPH part but not ecut.
A significant fraction of the walltime in EPH is spent for performing the FFTs required to apply the firstorder Hamiltonian H^1. The use of single precision in the FFT routines allows one to decrease the computational cost without loosing precision. This trick is activated by setting mixprec = 1. Note that this feature is available only if the code is linked with FFTW3 or intelMKL.
Important
The boxcutmin and mixprec tricks are not activated by default because users are supposed to perform preliminary tests to make sure the quality of the results is not affected by these options.
By default, the EPH code stores the KS wavefunctions in a single precision array although the majority of the calculations are done with double precision arithmetic (except for the FFT when mixprec is used). The precision used for the internal buffer can be specified at configure time with:
enable_gw_dpc=“yes” # Store wavefunctions in double precision buffers.
The default value if “no” i.e. single precision and we suggest not to change this option unless you have a good reason to do so.
We terminate the discussion with another trick that is not directly related to the EPH code but to the DFPT computation. Since the EPH code does not need the first order change of the wavefunctions (1WF files) we suggest to avoid the output of these files by using prtwf = 1 in the DFPT part as these files are quite large and the overall space on disk scales as nqpt × 3 × natom. When prtwf is set to 1, the DFPT code writes the 1WF file only if the DFPT SCF cycle is not converged so that one can still restart from these wavefunctions if needed (restarting a DFPT run from the first order wavefunctions is more effective than restarting from the first order density).
Starfunction interpolation of the KS eigenvalues¶
As mentioned in the introduction, in the EPH implementation Bloch states are computed nonselfconsistently on arbitrarily dense \kkmeshes without having to resort to interpolation schemes. The advantage of such approach is that calculations can be easily automated. The drawback is that the computational cost of the NSCF quickly increases with the density of the \kkmesh and nband hence for “big calculations” the cost of the NSCF part may be even greater than the eph computation itself.
There are however physical properties whose computation does not require the knowledge of the KS states for each \kkpoint in the IBZ. For instance, the computation of mobilities in semiconductors require the knowledge of the KS states whose energy is slight above (below) the CMB (VBM), let’s say ~0.2 eV. In metals, only states close to the Fermi level are needed to compute superconducting properties with the standard formalism. In other words, several eph calculations in which delta functions are involved require extremely dense BZ meshes to converge but as a matter of fact only a relatively small fraction of the BZ compatible with energy and crystallinemomentum conservation is nededed.
At this point a question naturally arises: can we avoid the NSCF computation of \kkpoints that are supposed to give negligible contribution to the final physical results? The answer is yes provided we are able to predict in some easy way and with reasonable accuracy the KS eigenvalues \ee_\nk without actually solving the KS equations.
The aproach used in the EPH code is based on the starfunction interpolation by ShanklandKoellingWood (SKW) with the improvements described in [Pickett1988]. In this method, the singleparticle energies are expressed in terms of the (symmetrized) Fourier sum
where the star function, S_\RR(\kk), is defined by
\RR is a lattice vector and the sum is over the N rotations of the crystallographic point group. By construction, the expansion in Eq.\eqref{eq:skw_expansion} fulfills the basic symmetry properties of the singleparticle energies:
In principle, the expansion coefficients in Eq.\eqref{eq:skw_expansion} can be uniquely determined by using a number of star functions equal to the number of ab initio \kkpoints but this usually leads to sharp oscillations between the input eigenvalues. To avoid this problem, one uses more star functions than ab initio \kkpoints and constrains the fit so that the interpolant function passes through the input energies and a roughness function is minimized [Pickett1988].
This einterp variable activates the interpolation of the electronic eigenvalues. The user can specify the number of star functions per ab initio \kkpoint and an optional Fourier filtering as proposed in [Uehara2000].
Without entering into details that will be discussed in the other specialized lessons, one can use the SKW algorithm to find the relevant \kkpoints, perform an abinitio NSCF run for these wavevectors only in order to produce a WFK file that can be used by the EPH code. The entire procedure is performed in an automatic way inside ABINIT but before running big EPH calculations, we strongly recommend to check whether the SKW interpolation gives reasonable results. A typical test would be:

Compute a WFK file on a reasonably dense IBZ (let’s call the file IBZ_WFK)

Perform a NSCF band structure calculation along a highsymmetry \kkpath covering the most important regions of the BZ (e.g. band edges in semiconductors). This step produces another WFK file that will be used as reference to check the interpolation (let’s call it KPATH_WKF)

Use the abitk executable in src/98_main to interpolate the KS energies using the eigenvalues in IBZ_WFK and compare the results with the abinitio band structure stored in KPATH_WFK.
The syntax is:
abitk skw_compare IBZ_WFK KPATH_WFK [lpratio 5] [rcut 0.0] [rsigma 0]
where optional arguments are placed between square brackets and the default value is indicated. If you prefer, it is possible to replace WFK files with GSR.nc file as in
abitk skw_compare IBZ_GSR.nc KPATH_GSR.nc
as only KS energies are needed for the SKW interpolation.
To compare the bands with AbiPy, use the abicomp script with the ebands
command:
abicomp.py ebands abinitio_EBANDS.nc skw_EBANDS.nc p combiplot
If the fit is not satisfactory, you may want to try one of the following options (in order of importance):
 Increase the abinitio mesh in IBZ_WFK.
 Increase the value of lpratio
 Play with rcut and rsigma to damp the oscillations in the interpolant
Note that it is sometimes difficult to get completely rid of spurious oscillations or artifacts in the SKW interpolation especially in the presence of degeneracies or band crossing/anticrossing, Remember, however, that achieving perfect agreement between the SKW interpolation and the abinitio results is not crucial since the SKW bands are only used to find those \kkpoints that are sufficiently close to the band edges (Fermi level). All these wavevectors will be recomputed afterwards with KSDFT and possible oscillations or artifacts will disappear in the abinitio results.
In a nutshell, you need to make sure that the SKW bands are reasonably close to the abinitio results especially in the region around the band edges for semiconductors or around the Fermi level for metals. Small deviations between SKW and abinitio bands can always be accounted for by increasing the value of sigma_erange used for generating the KERANGE.nc file.
Examples of input files to compute WFK files with the KERANGE are given in the last section of the mobility tutorial.