Boosting the efficiency of ab initio electron-phonon coupling calculations through dual interpolation

Page created by Leonard Love
 
CONTINUE READING
Boosting the efficiency of ab initio electron-phonon coupling calculations through dual interpolation
Page

                                                         Boosting the efficiency of ab initio electron-phonon coupling calculations through dual
                                                                                                 interpolation
                                                                   Anderson S. Chaves,1, 2 Alex Antonelli,2 Daniel T. Larson,3 and Efthimios Kaxiras3, 1
                                                                                       1 John
                                                                                            A. Paulson School of Engineering and Applied Sciences,
                                                                                      Harvard University, Cambridge, Massachusetts, 02138, USA
                                                                        2 Gleb Wataghin Institute of Physics and Center for Computing in Engineering & Sciences,
                                                                                    University of Campinas, PO Box 13083-859, Campinas, SP, Brazil
                                                                           3 Department of Physics, Harvard University, Cambridge, Massachusetts, 02138, USA
                                                                                                          (Dated: July 14, 2020)
                                                                     The coupling between electrons and phonons in solids plays a central role in describing many
arXiv:2006.16954v2 [cond-mat.mtrl-sci] 10 Jul 2020

                                                                  phenomena, including superconductivity and thermoelecric transport. Calculations of this coupling
                                                                  are exceedingly demanding as they necessitate integrations over both the electron and phonon
                                                                  momenta, both of which span the Brillouin zone of the crystal, independently. We present here an ab
                                                                  initio method for efficiently calculating electron-phonon mediated transport properties by dramatically
                                                                  accelerating the computation of the double integrals with a dual interpolation technique that combines
                                                                  maximally localized Wannier functions with symmetry-adapted plane waves. The performance gain
                                                                  in relation to the current state-of-the-art Wannier-Fourier interpolation is approximately 2ns × M,
                                                                  where ns is the number of crystal symmetry operations and M, a number in the range 5 − 60, governs
                                                                  the expansion in star functions. We demonstrate with several examples how our method performs
                                                                  some ab initio calculations involving electron-phonon interactions.

                                                                  PACS numbers: 71.15.Nc,36.40.-c,72.80.Ga

                                                        The electron-boson coupling is ubiquitous in phys-             phonon momenta, both spanning the entire Brillouin
                                                     ical phenomena through the whole spectrum of the                  Zone (BZ), independently. This double integration re-
                                                     physics of solids. In particular, electron-phonon (el-ph)         quires very fine sampling of the electron and phonon
                                                     coupling plays a fundamental role in the renormaliza-             wavevectors to achieve numerical convergence, which
                                                     tion of electronic and vibrational energy scales, thus            represents the bulk of the computational burden. The
                                                     determining the coupling itself, with important con-              application of crystal symmetry properties for the full
                                                     sequences for transport properties[1–4]. Conventional             integration of TP, which depend directly on the el-ph
                                                     superconductivity is a case in point, where the interac-          matrix elements, is not allowed. Even if the wavevector
                                                     tions between electrons and phonons give rise to Cooper           k in the el-ph matrix element lies within the symmetry-
                                                     pairing.[5] Other examples include the temperature de-            reduced portion (irreducible wedge) of the BZ, the trans-
                                                     pendence of electronic conductivity and thermoelectric            ferred momenta k + q spread out in the whole zone
                                                     transport properties,[6, 7] as well as phonon-assisted            because q belongs to an uniform mesh. Dense sampling
                                                     optical absorption in indirect-gap semiconductors.[8] In-         of the BZ is prohibitive, the reason being the connec-
                                                     terest in thermoelectrics has increased rapidly in recent         tion between transferred momenta with equally dense
                                                     years, partly due to the expectation of discovering higher        k-point meshes.
                                                     figure-of-merit materials, boosted by the nanotechnol-               Specialized numerical techniques have been devel-
                                                     ogy revolution. [9] A major goal of theory has been               oped to address this problem. One attempt to simplify
                                                     to predict thermoelectric transport properties directly           the brute-force integration is based on pre-screening of
                                                     from atomistic-scale calculations without any adjustable          subsets of the reciprocal space, such as relevant con-
                                                     or empirical parameters,[4, 6, 7] particularly combining          duction pockets within the neighborhood of band ex-
                                                     density functional theory (DFT) and many-body pertur-             trema that significantly contribute to the integral. Al-
                                                     bation theory. Despite great advances, such calculations          ternatively, interpolation schemes, such as linear[12] or
                                                     remain very demanding and still pose a challenge, even            Wannier-based[1, 13] ones, have been developed to im-
                                                     for simple crystalline bulk systems.                              prove convergence. In particular, the interpolation of
                                                        A well-established approach, using a first-principles          the el-ph matrix elements on the basis of Wannier func-
                                                     description of el-ph coupling, relies on solving the el-          tions introduced by Giustino, Cohen, and Louie,[13] has
                                                     ph matrix elements through density-functional pertur-             proven very successful in calculating properties with
                                                     bation theory (DFPT)[11]. The el-ph matrix elements,              more favorable scaling than using directly the DFPT ap-
                                                     g(k, k + q) = (hk + q| δq,β V KS |ki)uc , correspond to the       proach. In this Letter, we present a novel method for
                                                     electronic scattering calculated from the variations of           the computation of el-ph mediated TP which uses two
                                                     the Kohn-Sham (KS) potential due to phonon pertur-                interpolations: the first one is the usual Wannier-Fourier
                                                     bations with wavevector q and branch index β within               (W-F) interpolation, followed by a second one based
                                                     the unit cell (uc). To obtain transport properties (TP),          on symmetry-adapted plane-waves (PW). Our method
                                                     | g(k, k + q)|2 must be integrated over the electron and          leads to an efficient sampling of extremely fine, homo-
Boosting the efficiency of ab initio electron-phonon coupling calculations through dual interpolation
2

                                     8
                                   10                                                                                                                                       280
                                                                           (a)                                                                                                                              (b)                     3
                                                                                                    2.4

                                                                                                          Lorenz function, Λ (10-8WΩ/K2)

                                                                                                                                                                                                                                          Thermal conductivity, κ (W/mK)
                                                                                         Λn
Electrical Conductivity, σ (S/m)

                                                                                                                                           Seebeck Coefficient, S (µV/K)
                                                                                                                                                                            140                                            Sp
                                   107                                                                                                                                                                                              2.4
                                                                                                    2.2
                                                                                                                                                                               0
                                                                                         Λp
                                                                                                                                                                                                                                    1.8
                                   106                                                              2                                                                       -140                                           Sn
                                                                                         σn                                                                                                                                         1.2
                                     5                                                              1.8                                                                     -280                                      κn
                                   10
                                                                                         σp
                                                                                                                                                                                                                                    0.6
                                                                                                                                                                            -420                                      κp
                                     4
                                                                                                    1.6
                                   10
                                                                                                                                                                                                                                    0
                                         0.25          0.5             1             2        4                                                                                    0.25       0.5       1         2             4
                                                 Carrier Concentration (1020 cm-3)                                                                                                        Carrier Concentration (1020 cm-3)

Figure 1. Thermoelectric transport properties TP for p- and n-type doped Si polycrystals: (a) Electrical conductivity, σ, (red)
compared with experimental values[10] (filled black symbols) and Lorenz function, Λ, (blue); (b) Seebeck coefficient, S, (red)
compared with experimental data[10] (filled black symbols) and thermal conductivity due to the carriers, κ, (blue) calculated
from the relaxation times due to scatterings by el-ph coupling and ionized impurities (see text for details).

geneous k and q grids, with a significant decrease in                                                                                                           Wigner-Seitz (WS) supercell with Born-von-Kármán
computational cost compared to W-F calculations with                                                                                                            (BvK) periodic boundary conditions, Uk0 (uq0 ) is a di-
a single interpolation.                                                                                                                                         agonalizer matrix over k0 (q0 ) indices from Wannier to
   To illustrate the capability of our method, we con-                                                                                                          Bloch representations for electrons (phonons) and the
sidered realistic properties of solids (see SM for more                                                                                                         el-ph matrix elements in the Wannier representation are
details). Fig. 1 shows the calculated thermoelectric TP                                                                                                         given by
for Si polycrystals using our dual interpolation method
                                                                                                                                                                                            1
to calculate the relaxation time within the Boltzmann                                                                                                                      g(Re , R p ) =
                                                                                                                                                                                            Np   ∑ e−i(k·Re +q·R p ) U†k+q g(k, q)Uk u−q 1 ,
transport equation (BTE). We calculated the electrical                                                                                                                                           k,q
conductivity, σ, Seebeck coefficient, S, Lorenz function,                                                                                                                                                                  (2)
Λ, and thermal conductivity due to the carriers, κ, as                                                                                                          Uk is a unitary matrix corresponding to the rotation
functions of carrier concentration. Our results for n                                                                                                           of the corresponding electronic states from Bloch to
and p-doped Si polycrystals agree reasonably well with                                                                                                          Wannier representations within the gauge of maximally
available experimental TP. For these calculations we also                                                                                                       localized Wannier functions (MLWF),[14] and uq is a uni-
included the scattering by ionized impurities, within the                                                                                                       tary rotation matrix from Bloch to MLWF for phonons.
Brooks-Herring theory (see SM), considering in all cal-                                                                                                         The strength of the W-F interpolation method is the fact
culations a fixed unitary ratio between impurities and                                                                                                          that one only needs to perform calculations over the
carrier concentrations, which are the only input param-                                                                                                         initial coarse k, q meshes, and then can use Eq. (1) to
eter along with the crystal structure. Fig. 2 shows the                                                                                                         determine g(k0 , q0 ) on finer k0 , q0 meshes. For this, we
results of phonon-assisted optical absorption for Si at                                                                                                         neglect the matrix elements outside the WS supercell
296K and 78K as a function of the photon energies. Our                                                                                                          generated from the initial coarse BZ mesh.
calculations, based on our dual interpolation method                                                                                                               The accuracy of W-F calculations strongly depends
and the theory developed by Hall, Bardeen and Blatt (see                                                                                                        on the spatial localization of g(Re , R p ) within Eq. (1). A
[8]), are in good agreement with experimental results.                                                                                                          more detailed analysis suggests g(Re , R p ) should decay
  Before presenting our method we review the basic                                                                                                              in the variable Re at least with the rapidity of MLWFs.
concept and analyze the advantages and drawbacks of                                                                                                             For Re = 0, g(0, R p ) decays with R p due to the screened
W-F interpolation. Using W-F interpolation, the el-ph                                                                                                           Coulomb interaction of the dipole potential generated by
matrix elements g(k, q) can be calculated on coarse k, q                                                                                                        atomic displacement. Thus the localization of g(Re , R p )
meshes and then interpolated onto much finer k0 , q0                                                                                                            depends strongly on the dielectric properties of the sys-
meshes through simple matrix multiplication.[13] The                                                                                                            tem. In particular, Friedel oscilations[16] (|R p |−3 ) and
matrix elements on the fine mesh are given by                                                                                                                   quadrupole behavior[17] (|R p |−4 ) are intimately related
                                                                                                                                                                to the screening properties of metals and nonpolar semi-
                                                1                  0             0
g(k0 , q0 ) =
                                                Ne    ∑        ei(k ·Re +q ·R p ) Uk0 +q0 g(Re , R p )U†k0 uq0 ,                                                conductors, respectively.
                                                     Re ,R p                                                                                                       Despite the advantages of the W-F interpolation
                                                   (1)                                                                                                          and its more favorable scaling, there are still some
where Re and R p are primitive lattice vectors of the                                                                                                           drawbacks. The method is computationally intensive
Boosting the efficiency of ab initio electron-phonon coupling calculations through dual interpolation
3

                                   105                                                                           interpolation over the whole k-space on a finer grid.
                                                                                   296K                          Such an interpolation problem may present severe dif-
                                   104                                                                           ficulties if one uses an inadequate basis set. We show
Absorption Coefficient, α (cm-1)

                                                                                                      78K        below that such an interpolation over the entire BZ, com-
                                        3
                                   10                                                                            bined with periodic or other boundary conditions, can
                                                                                                                 be properly constructed from a basis set possessing the
                                   102                                                                           appropriate crystal symmetry.
                                        1
                                   10

                                   100
                                                                                                                    As the full symmetry of the crystal’s reciprocal space
                                                                                                                 is contained in the function f (k̄l ), it is natural to use
                                   10-1                                                                          symmetry-adapted PW or star functions, Υm (k0 ), as a
                                                                                                                 basis set to Fourier expand f [19]:
                                   10-2
                                            1        1.25   1.5     1.75   2    2.25       2.5        2.75   3                                      M
                                                                  Photon Energy (eV)                                                f˜(k0 ) =     ∑     am Υm (k0 ) ,                 (3)
                                                                                                                                                 m =1
Figure 2.   Phonon-assisted optical absorption, α, of Si at
296K and 78K as a function of photon energies. Our results
                                                                                                                 where Υm (k0 ) = n1s ∑{υ} exp(i (υRm ) · k0 ) , with the
(blue and magenta) are in good agreement with corresponding
experimental results (black)[15].                                                                                sum running over all ns point group symmetry oper-
                                                                                                                 ations {υ} on the direct lattice translations, Rm . Star
                                                                                                                 functions obey orthogonality relations involving BZ
                                                                                                                 summations[19], are totally symmetric under all point-
when many k/q points are needed to achieve con-
                                                                                                                 group operations, and are ordered such that the mag-
verged values for TP. The main computational oper-
                                                                                                                 nitude of Rm is nondecreasing as m increases, defining
ations are the simple matrix multiplications shown in
                                                                                                                 each star function to a given shell of lattice vectors. By
Eq. 1, with a computational complexity of O(n3 ) for                                                             taking into account the symmetry, it is expected that
classical computation, where n is the matrix size in-                                                            this expansion would converge much faster than using
                                                                                       0         q0
volved. For final dense grids with N kf (N f ) k0 (q0 )                                                          a regular Fourier expansion. Following the approach
points, the number of floating-point operations reach                                                            proposed by Shankland-Koelling-Wood[20, 21], we take
                                     0          q0                                                               the number of star functions in the expansion, M, to be
≈ N kf × N f n3 . To reduce the computational cost of                                                            greater than the number of data points (M > nk̄ ). We
Wannier-based calculations, some strategies have been                                                            then require the fit function, f˜, to pass through the data
adopted, including double grid schemes[6] or (quasi)                                                             points and use the extra freedom from additional basis
Monte-Carlo (MC) integrations.[18] In the former, only                                                           functions to minimize a spline-like roughness functional
the bandstructure and phonon dispersion are calcu-                                                               in order to suppress oscillations between data points,
lated over an ultrafine grid (N f × N f × N f ), while el-ph                                                     resulting in a well behaved function throughout the BZ.
matrix elements are computed over a moderate grid
(Nel − ph × Nel − ph × Nel − ph , with Nel − ph = s × N f and
s = 1/2, 1/3) and extrapolated to the ultrafine grid as-                                                           We adopt the spline-like roughness functional defined
suming that the el-ph coupling function is smooth. The                                                           by Pickett, Krakauer and Allen [22],
drawbacks here consist of the extrapolation which is
often fraught with risk, and the very modest gain fac-                                                                                          M
tor of the method. On the other hand, by using (quasi)                                                                                 Π=       ∑ | a m |2 ρ ( R m )                  (4)
MC integration, one has to test very dense sets of ran-                                                                                     m =2
dom (or quasi-random) k or q-points, which is a serious                                                                                                   2 2
drawback.
                                                                                                                                                
                                                                                                                                                     Rm
                                                                                                                 with ρ( Rm ) =        1 − c1       Rmin           + c2 ( RRmin
                                                                                                                                                                             m 6
                                                                                                                                                                                ) , where
   Our method proceeds as follows. For clarity, we de-
scribe the procedure for doing the partial q0 integration                                                        Rm = |Rm |, Rmin is the magnitude of the smallest
first, but one can easily switch the order of integration.                                                       nonzero lattice vector, and c1 = c2 = 3/4. Such a func-
We begin by computing g(k, q) over coarse k and q                                                                tional is more physically appealing than the original
meshes. Next, using W-F interpolation, we determine                                                              functional proposed by Shankland-Koelling-Wood, in
g over a finer q0 mesh and perform the partial integra-                                                          the sense that departures of f˜ is minimized from its
tion at each of the nk̄ irreducible k-points, k̄l , corre-                                                       mean value, a1 , instead of zero. The main problem in
sponding to a moderate regular         k-mesh (kr ). Thus we                                                     the expansion by star functions in Eq. (3) is the determi-
obtain a function f (k̄l ) ∝ BZ | g(k̄l + q0 , k̄l )|2 dq0 con-
                                  R
                                                                                                                 nation of the Fourier coefficients, am . Thus, a Lagrange
taining first-principles el-ph coupling properties defined                                                       multiplier method can be used toward this goal, once
at selected high-symmetry points in the corresponding                                                            the problem has been reduced to minimizing Π sub-
k-space. Given f (k̄l ), the next goal is to find a smooth                                                       ject to the constraints, f˜(k̄l ) = f (k̄l ), in relation to am .
4

Consequently, the result of this minimization is                                  following point group operations of crystal symmetry.
                                                                                  The corresponding translation vectors can be given as
                     n −1
     (
       ρ( Rm )−1 ∑l =k̄ 1 λ∗l Υ∗m (k̄l ) − Υ∗m (k̄nk̄ ) , m > 1,                  R = u1 a1 + u2 a2 + u3 a3 , in which a1 , a2 , a3 are related
                                                      
am =                  M                                                           to the crystal’s direct primitive vectors. Such points
       f (k̄nk̄ ) − ∑m =2 am Υm ( k̄nk̄ ),                m = 1,
                                                                                  are generated     inside a sphere with a radius defined as
                                                            (5)                     0
                                                                                  R = 3 · nk̄ · ns · M · Ω/4π, in which Ω is the volume
                                                                                         p3
in which the Lagrange multipliers, λ∗l , can be evaluated
from                                                                              of the unit cell. Consequently, R0 determines the full ex-
                                                                                  tension of the real space and can be properly increased,
                                                nk̄ −1                            for example, by increasing M, the number of star func-
                    f (k̄ p ) − f (k̄nk̄ ) =     ∑       H pl λ∗l ,         (6)   tions per k-point. In order to capture crystal anisotropy,
                                                 l =1                             the extension of the real space can be determined for
                                                                                  each crystal direction, defining spheres for each crys-
with                                                                              tallographic axis with  √ the maximum radius given by
           M  Υm (k̄ p ) − Υm (k̄nk̄ ) Υ∗m (k̄l ) − Υ∗m (k̄nk̄ )
                                                                              Rmax (t) = I NT ( R0 · bt · bt ) + 1, where bt are the re-
 H pl = ∑                                                           ,             spective reciprocal primitive vectors, with t = {1, 2, 3},
        m =2
                                  ρ( Rm )                                         and I NT ( x ) takes the largest integer number that does
                                                                  (7)             not exceed the magnitude of x.
a positive-definite symmetric matrix that can be deter-                               The star functions are ordered in such a way that the
mined once for a given crystal problem and can be easily                          magnitude of Rm is nondecreasing as m increases. Thus,
crafted numerically.                                                              a 3D array containing all vectors are sorted considering
   Once the Fourier coefficients are determined, a repre-                         their concentric radius, r, from the sphere center defined
sentation of f˜ is generated, which can be written more                           for each axis, and provided that all vectors, R, have
clearly as a linear mapping of the W-F data,                                      different star functions, m. The magnitude of each Rm
                                                                                  vector is defined through the metric tensor formalism.
                         nk̄ −1
                                                                                  For all Rm in the Bravais lattice, the reciprocal lattice
           f˜(k0 ) =      ∑       J (k̄l , k0 )[ f (k̄l ) − f (k̄nk̄ )] ,   (8)   is characterized by a set of wavevectors k, such that,
                          l =1
                                                                                  e2πik·Rm = 1. Given Rm and k in the same direction,
where J is the interpolation formula given by                                     the magnitude of the vector k in the reciprocal space
                                                                                  is given by |k| = (k1 u1 + k2 u2 + k3 u3 )/r = (nint (1) +
                     nk̄ −1 M
                                  [Υ∗m (k̄ p ) − Υ∗m (k̄nk̄ )]Υm (k0 )            nint (2) + nint (3))/r, where nint (t) = 1, 2, ..., k max (t) are
   J (k̄l , k0 ) =    ∑ ∑                      ρ( Rm )H pl
                                                                       ,    (9)   integer numbers with k max (t) = 2Rmax (t) + 1. To de-
                      p =1 m                                                      termine all k vectors from Rm , a 3D FFT is performed.
                                                                                  In practice, k max (t) defines the number of data points
which transforms one k-mesh into another one, that is,                            on each dimension and should be carefully taken as
k̄l → k0 . This is the main result of our approach, which                         the product of small primes in order to improve the
allows great computational savings by transforming the                            efficiency of FFT.
W-F data obtained over the k-mesh of irreducible points                               In fact, the FFT computational complexity is
(k̄l ) into a homogeneous dense grid (k0 ) that is larger                         O( N log N ), where N corresponds to the number of
than the regular grid (kr ) that generates such irreducible                       data points related to the product of FFT dimensions,
points. One important point to stress is that J does                                                                                             3
                                                                                  namely
                                                                                  √          N√ = k max√  (1) × k max (2) × k max (3) ≈ 8R0 ·
not depend on data, but it is completely defined by the
                                                                                     b1 · b1 · b2 · b2 · b3 · b3 ≈ 6/π (nk̄ × ns × M ). Con-
lattice, namely the set of irreducible sampling (k̄l ), the
                                                                                  sequently, the number of floating-point operations by
number of star functions (M), and the form of spline-                                                             q0
like roughness functional (Π). In practice, in order to                           using our approach is ≈ N f × nk̄ n3 + 6/π (nk̄ × ns ×
get a denser mesh, we rely on a Fast Fourier Transform                            M) ln (6/π (nk̄ × ns × M)). The first term comes from
(FFT) from the real space to the reciprocal space in order                        the first W-F interpolation by using nk̄ irreducible points,
to compute the expansion given in Eq. (3). We take                                while the second one comes from the symmetry-adapted
advantage of the BvK periodic boundary conditions to                              PW interpolation. The gain in performance by using our
increase the real space by the expansion factor, M, as                            method in comparison with single W-F calculation, to
will be explained below, to get proportionally a new                              get approximately the same final homogeneous grid,
homogeneous k0 -mesh finer than the original one. As a                            can be given by ≈ 2(ns × M), assuming N kf ≈ N
                                                                                                                                      0

result we get the full integration over very fine k0 and                                   0       q0
q0 meshes in order to calculate transport properties, that                        and N kf = N f . Clearly, high symmetry systems al-
is, TP ∝ ∑k0 f˜(k0 ) .                                                            low greater computational savings, however the factor
   Our implementation for the second interpolation is                             M, typically ranging from 5 − 60 enables remarkably
based on modifications and adaptations of some subrou-                            significant performance gain even for low symmetry
tines of the BoltzTraP[26] code. Lattice points and their                         systems.
respective star functions are generated in the real space                           In order to test our implementation, we carried out TP
5

Figure 3. (a) Scattering rates for Si at 300 K calculated with the dual interpolation method (green dots), DFT with linear
interpolation (black[23] and light blue[4] lines and black squares[12]), W-F interpolation using EPW (purple triangles[24] and
light orange circles), and tight-binding calculations (red[25] line). The W-F calculations used (30)3 k/(60)3 q ((100)3 k/(40)3 q)
meshes in calculations represented by the purple triangles (orange circles). (b) Walltime required to perform calculations for the
electron self-energy due to el-ph coupling in the Fan-Migdal approximation, for different grid sizes. Empty magenta triangles
correspond to the direct calculation using EPW over homogeneous grids. Empty (filled) green squares, empty (filled) red circles
and empty (filled) blue diamonds correspond to the calculations using T-EPW (only second PW interpolation), starting from 256,
1661 and 5216 irreducible k points, respectively (see text for details).

calculations for silicon. We computed the imaginary part           on W-F interpolations (EPW) grows almost exponen-
of the electron self-energy in the Fan-Migdal approxima-           tially with increasing k0 /q0 density. By applying our
tion, Im Σ, which gives the relaxation time due to e-ph            method a drastic reduction in the computational time
scattering, and consequently, thermoelectric TP using              is obtained, which is generally greater than two orders
the BTE, as shown in Fig. 1. Additionally, we also stud-           of magnitude, as can be observed from the curves for
ied phonon-assisted optical absorption for Si, as shown            the total time of T-EPW. These curves also demonstrate
in Fig. 2. More details about these calculations can               a roughly exponential growth with k0 /q0 density, which
be found in the Supplemental Material (SM). Since the              is due to increasing the grid size in the first W-F in-
first step in computing the double BZ integrals is based           terpolation, independent of the number of initial irre-
on a W-F interpolation, our implementation has been                ducible points k̄l . In these test calculations we consid-
built on top of the Electron-Phonon Wannier (EPW) [27]             ered nk̄ = 256, 1661, 5216, leading to regular meshes of
code, which is contained in the Quantum Espresso pack-             kr = (20)3 , (40)3 , (50)3 . The plateaus in the computa-
age [28]. We have modified the EPW code in order                   tional time for T-EPW are due to increasing the value of
to include the second PW interpolation, as described               M from 5 to 60, while keeping fixed the grid of first W-F
above, which we call Turbo-EPW (T-EPW). In Fig. 3(a)               interpolation ((100)3 q0 points). As shown in Fig. 3(b),
we show Im Σ for Si at 300 K calculated using T-EPW                using PW interpolation one can achieve much denser
in comparison with other approaches, namely, previous              grids by increasing the value of M with negligible in-
DFT with linear interpolation[4, 12, 23], W-F interpola-           crease in computational time. As shown in the SM, the
tion with EPW[24], and tight-binding calculations[25].             accompanying error due to the increase of M to generate
Our results, calculated using (100)3 k0 /(100)3 q0 grids,          denser grids diminishes by increasing nk̄ ; a solution that
are in good agreement with other W-F calculations us-              can also be used to minimize errors from possible kink
ing the EPW code directly on (30)3 k0 /(60)3 q0 [24] and           structures derived from band crossings, which leads to
(100)3 k0 /(40)3 q0 grids, but with significantly reduced          a Gibbs ringing in Fourier series analysis.
computational time.                                                   In summary, our method can be used to calculate effi-
  In order to estimate the performance gain of our ap-             ciently el-ph-based TP. The computational performance
proach, in Fig. 3(b) we show the computational time                gain is remarkable, being ≈ 2(ns × M) faster than state-
required to finalize the calculation of Im Σ for differ-           of-the-art EPW calculations without loosing accuracy.
ent k0 /q0 grids, all using the same computational hard-           It should be emphasized that this novel approach can
ware. The time required for calculations based only                also be used as an efficient and stable numerical tool in
6

order to calculate ubiquitous double BZ integrals, and
potentially extending to many further applications, for
instance phonon-assisted nonlinear optical properties,
superconducting critical temperature and its related ther-
modynamic properties and electron-plasmon coupling
TP from first-principles. Moreover, this method may
allow previously impractical calculations and can serve
as a starting point to explore the effects of the vertex
corrections to the Migdal approximation as well as to ad-
dress the e-ph coupling in complex systems with many
atoms in the unit cell. This last capability would be
useful for the discovery of efficient materials for energy
applications, such as high-performance thermoelectrics.

      Appendix A    Details of the calculations for Si

   First, we compute the self-consistent potential and       Figure 4. Convergence analysis of the scattering rate of Si at
Kohn-Sham states on a 12 × 12 × 12 Monkhorst-Pack            300K due to el-ph coupling in the Fan-Migdal approximation
                                                             as a function of the electron energy. The calculation has been
k-point grid using DFT and lattice-dynamical properties
                                                             performed on different q meshes, namely, (20)3 (green crosses),
with DFPT[11] on a 3 × 3 × 3 q-point grid, as imple-
                                                             (60)3 (blue squares), (100)3 (red dots) q-meshes in the first W-
mented in the Quantum Espresso distribution[28] us-
                                                             F interpolation, while keeping fixed the number of irreducible
ing the Perdew-Burke-Ernzerhoff exchange-correlation         k points in the second plane-waves interpolation, leading to
functional[29].    We used a full-relativistic norm-         (100)3 k-mesh.
conserving optimized Vanderbilt pseudopotential[30].
The unit cell consists of Si in the diamond structure
with an experimental lattice parameter of 5.43 Å. The
e-ph matrix elements are first computed on coarse grids,     k-mesh, is within ≈ ±6%. For the remaining, the differ-
then they are determined in the significantly finer          ence between the data points is about the same, within
grids using both Wannier-Fourier (W-F) interpolation         ≈ ±3%.
only, through EPW code and our dual interpolation
method, Turbo-EPW (T-EPW). Maximally localized Wan-
nier functions[14] for the wannierization procedure are         Appendix C          Calculation of electron self-energy and
obtained from Wannier90.[31] Thus, Bloch-to-Wannier                                       thermoelectric properties
rotation matrices and then Wannier-to-Bloch diagonal-
izer matrices are used to interpolate el-ph matrix ele-
                                                               The expression for the imaginary part of electronic
ments.
                                                             self-energy due to el-ph coupling in the Fan-Migdal
                                                             approximation can be derived from quantum field
                                                             theory[1] and it is expressed as
           Appendix B    Convergence analysis
                                                                                                            dq
                                                                                                        Z
                                                                                 Σ00n,k (ω, T ) = π ∑           |g    (k, q)|2
   Fig. 4 shows how the scattering rate aproaches con-                                            m,β BZ
                                                                                                            Ω BZ mn,β
vergence by increasing the mesh size of W-F interpo-                  "
lation, from (20)3 to (100)3 q points, while keeping              ×
                                                                                               
                                                                              nqβ ( T ) + f mk+q δ(ω − (emk+q − e F ) + ωqβ )
the number of irreducible points fixed at 5216 k points.
The second interpolation by star functions leads into                                                                      #
a converged grid with (100)3 k points. Fig. 5 shows           +[nqβ ( T ) + 1 − f mk+q ]δ(ω − (emk+q − eF ) − ωqβ ) ,
the difference in scattering rates calculated by different
approaches, namely, different parameters in the sec-                                                                       (10)
ond plane-waves interpolation (PWI), over equivalent
meshes. The analysis shows that the accompanying er-         where nqβ ( T ) and f mk+q are the Bose-Einstein and the
ror due to the increase in M decreases by enlarging the      Fermi-Dirac distributions, Ω BZ is the BZ volume, m
number of irreducible points, nk̄ (see main text). More-     and n are the corresponding electronic states, while
over, the difference between scattering rates calculated     β represents the phonon branch, emk+q are the elec-
over (100)3 q/ 1661 k points expanded by using M = 5         tronic eigenenergies of the state mk + q and ωqβ are the
to reach (100)3 k-mesh and data from (100)3 q/ 256           corresponding eigenfrequencies with wavevector q and
k points expanded by using M = 30 to reach (100)3            phonon branch β. Basically, from the first W-F interpola-
7

                                                                 perimental conditions of zero temperature gradient
                                                                 (∇ T = 0) and zero electric current, the kinetic co-
                                                                 efficient tensors can be identified with the electrical
                                                                 conductivity tensor, σ = Λ(0) , the Seebeck coeffi-
                                                                 cient tensor, S = (eT )−1 Λ(1) /Λ(0) , and the charge
                                                                 carrier contribution  to thermal conductivity   tensor,
                                                                                               −1
                                                                                                          
                                                                 κ e = ( e 2 T ) −1 Λ (1) · Λ (0) · Λ (1) − Λ (2) . Conse-
                                                                 quently, once we have τn,k from Eq. (11), we can com-
                                                                 pute the electrical conductivity, the Seebeck coefficient,
                                                                 Lorenz function and the charge carrier contribution
                                                                 to thermal conductivity from the solution of Eq. (13)
                                                                 and Eq. (12). Note that the both bandstructure and
                                                                 phonon dispersion have also been interpolated by the
                                                                 method presented in the main text over the same grid
                                                                 as electron self-energy. Indeed, we have implemented
                                                                 these equations on top of the BoltzTraP code,[26] from
                                                                 which we can obtain these transport properties directly
Figure 5. Difference between scattering rates computed by
                                                                 from first principles.
using different parameters in the second plane-waves interpo-
lation, over equivalent meshes. Green crosses correspond to
the difference between scattering rates calculated over (100)3
                                                                       Appendix D     Scattering by ionized impurities
q/ 1661 k points expanded by using M = 5 to reach (100)3 k-
mesh and data from (100)3 q/ 256 k points expanded by using
M = 30 to reach (100)3 k-mesh. Blue squares correspond to           For the calculation of thermoelectric transport proper-
the difference between scattering rates calculated over (100)3   ties of n- and p-type Si polycrystals, we have also consid-
q/ 5216 k points expanded by using M = 10 to reach (180)3        ered the scattering by ionized impurities. Such scattering
k-mesh and data from (100)3 q/ 1661 k points expanded by
                                                                 has been treated theoretically by Brooks and Herring
using M = 30 to reach (180)3 k-mesh. Red dots correspond to
                                                                 (B-H)[32, 33] by considering a screened Coulomb po-
the difference between scattering rates calculated over (100)3
                                                                 tential, the Born approximation for the evaluation of
q/ 5216 k points expanded by using M = 8 to reach (160)3
                                                                 transition probabilities and neglecting perturbation ef-
k-mesh and data from (100)3 q/ 1661 k points expanded by
                                                                 fects of the impurities on the electron energy levels and
                                                                 wave functions. In the B-H theory the electron is scat-
using M = 25 to reach (160)3 k-mesh.
                                                                 tered independently by dilute concentrations of ionized
                                                                 centers randomly distributed in the semiconductor.
                                                                    The per-unit-time transition probability for the scatter-
tion we can get f (k̄) = Σ00n,k̄ (ω, T ), over the irreducible   ing of charge carriers by ionized impurities can be given
points, which will be interpolated throughout the whole          in the plane-wave approximation as
BZ by star functions, resulting in Σ00n,k0 (ω, T ) over denser
k0 grids.                                                                      2π Ni
                                                                   W (k|k0 ) =
   Σ00 is directly related to the scattering rate, that is,                     h̄ V
inversely proportional to the relaxation time                      Z                            2
                                                                     U (r) exp i (k − k0 ; r) dr δ(e(k0 ) − e(k)) ,
                                                                                           
                                                                                                                         (14)
                   1
                        = 2Σ00n,k (ω = 0, T ) ,          (11)
                 τn,k                                               where U (r) is the scattering potential and Ni is the ion-
                                                                    ized impurity concentration.
which enters in kinetic transport equations. Indeed, the
kinetic coefficient tensors can be expressed through                   A long-range Coulomb field, U (r ) = eφ(r ) = ±e2 /ζr,
                                                                    with potential φ at a point r of the crystal is created
                                                             !      by the presence of positive (donor) or negative (accep-
                                          ∂ f (0) (µ; e, T )
                  Z
  (α)           2
Λ (µ; T ) = e        Ξ(e, µ, T )(e − µ) −
                                       α
                                                               de , tor) impurity ions, within a medium with dieletric con-
                                                  ∂e                stant ζ. The straightforward application of this field in
                                                           (12)     Eq. (14) leads into a logarithmic divergence, and hence, a
where µ is the chemical potential and Ξ(e, µ, T ) is the            screened Coulomb potential has to be considered. From
transport distribution kernel given by                              the B-H theory the potential can be expressed in a more
                                                                    rigorous form as φ(r ) = ±e/ζr (exp{(−r/r0 )}), where
                                                         dk
               Z
 Ξ(e, µ, T ) = ∑ vn,k ⊗ vn,k τn,k (µ, T )δ(e − en,k ) 3 ,           r0 is the radius of ion field screening defined by
                   n                                    8π
                                                           (13)                             4πe2       ∂ f0
                                                                                                  Z
                                                                                  −2
                                                                                r0 (k) =            −       g(e)de ,      (15)
with vn,k being the electron velocity. From both ex-                                         ζ0       ∂e(k)
8

where f (0) (e) is the equilibrium electron distribution                   with
function, ζ 0 is the static dielectric constant, and g(e)
is the density of states, calculated numerically on an
                                                                                                       vim (k) gmj,β (k, q)
energy grid with spacing de sampled over Nk k-points                               S1 (kq) =   ∑ em,k − eik − h̄ω + iΓm,k     ,       (20)
                                                                                               m
                               dk        1            δ(e − en,k )
           Z
  g(e) =       ∑ δ(e − en,k ) 8π3   =
                                        ΩNk     ∑         de
                                                                   ,
               n                                n,k
                                                                                                   gim,β (k, q)vmj (k + q)
                                                                  (16)      S2 (kq) =   ∑ em,k+q − eik ± h̄ωβq + iΓm,k+q          ,   (21)
where Ω is the volume of the unit cell.                                                  m
  Within the relaxation time approximation for the Boltz-
mann transport equations, the relaxation time for the                      and
scattering of the charge carriers by ionized impurities
can be expressed as                                                                                  1 1
                                                                                     P = (n βq +      ± )( f − f j,k+q ) .            (22)
                                                                                                     2 2 ik
                           h̄ζ 0 2       ∂e(k)
           τimp (k) =                 k2                         (17)
                         4
                      2πe Ni Fimp (k)     ∂k                                In these equations, ω is the photon frequency, c is the
                                                                         speed of light, nr is the refractive index (for silicon we
where                                                                    used nr = 3.4), and λ is the photon polarization. S1 and
                                                                         S2 are the two possible ways for the indirect absorption
                                             η
                Fimp (k) = ln(1 + η ) −         ,                (18)    process, while P is related to the carrier and phonon
                                            1+η                          statistics. For calculations of Si, the DFT band gap and
                                                                         all conduction bands have been shifted up by 0.7eV to
is the screening function with η = (2kr0 )2 . Here, we                   simulate experimental gap. Our calculations have been
                   ∂e(k)
interpolated ∂k within Eq. (17) by using the plane-                      carried out over (60)3 q/(40)3 k meshes. It took ≈ 36
waves interpolation (see the main text) and used Math-                   minutes to perform the calculation by using our dual
iessen’s rule to consider both the scattering by ionized                 interpolation method on 8 CPU cores at the Odyssey
impurities and phonons.                                                  cluster (Harvard University).

                                                                                               Acknowledgments
   Appendix E        Calculation of phonon-assisted optical
                                   absorption
                                                                           The authors thank the Harvard FAS Research Comput-
   To calculate the phonon-assisted absorption coeffi-                   ing facility and the Brazilian CCJDR-IFGW-UNICAMP
cient, we use the Fermi’s golden rule expression[8, 34]:                 for computational resources. A.S.C. and A.A. grate-
                                                                         fully acknowledge financial support from the Brazil-
                 4π 2 e2 1 1                                             ian agency FAPESP under Grants No.2015/26434-2,
    α(ω ) = 2
                ωcnr (ω ) Ω Nk Nq       ∑     |λ · (S1 + S2 )|2 (19)     No.2016/23891-6, No.2017/26105-4, No.2018/01274-0
                                      βijkq                              and No.2019/26088-8. A.S.C. also acknowledges the
                     × Pδ(e j,k+q − ei,k − h̄ω − ±h̄ω βq ) ,             kind hospitality of SEAS-Harvard University.

 [1] F. Giustino, Reviews of Modern Physics 89, 015003 (2017).            [9] A. I. Boukai, Y. Bunimovich, J. Tahir-Kheli, J.-K. Yu, W. A.
 [2] A. Marini, Physical Review Letters 101, 106405 (2008).                   Goddard III, and J. R. Heath, in Materials For Sustainable
 [3] C.-H. Park, F. Giustino, M. L. Cohen, and S. G. Louie,                   Energy: A Collection of Peer-Reviewed Research and Review
     Physical Review Letters 99, 086804 (2007).                               Articles from Nature Publishing Group (World Scientific,
 [4] O. Restrepo, K. Varga, and S. Pantelides, Applied Physics                2011) pp. 116–119.
     Letters 94, 212103 (2009).                                          [10] M. Strasser, R. Aigner, C. Lauterbach, T. Sturm, M. Fra-
 [5] E. Margine and F. Giustino, Physical Review B 90, 014518                 nosch, and G. Wachutka, Sensors and Actuators A: Phys-
     (2014).                                                                  ical 114, 362 (2004).
 [6] M. Fiorentini and N. Bonini, Physical Review B 94, 085204           [11] S. Baroni, S. De Gironcoli, A. Dal Corso, and P. Giannozzi,
     (2016).                                                                  Reviews of Modern Physics 73, 515 (2001).
 [7] Z. Wang, S. Wang, S. Obukhov, N. Vast, J. Sjakste,                  [12] W. Li, Physical Review B 92, 075405 (2015).
     V. Tyuterev, and N. Mingo, Physical Review B 83, 205208             [13] F. Giustino, M. L. Cohen, and S. G. Louie, Physical Review
     (2011).                                                                  B 76, 165108 (2007).
 [8] J. Noffsinger, E. Kioupakis, C. G. Van de Walle, S. G.              [14] N. Marzari and D. Vanderbilt, Physical Review B 56, 12847
     Louie, and M. L. Cohen, Physical Review Letters 108,                     (1997).
     167402 (2012).
9

[15] R. Braunstein, A. R. Moore, and F. Herman, Physical         [25] D. Rideau, W. Zhang, Y. Niquet, C. Delerue, C. Tavernier,
     Review 109, 695 (1958).                                          and H. Jaouen, in 2011 International Conference on Simula-
[16] A. L. Fetter and J. D. Walecka, Quantum theory of many-          tion of Semiconductor Processes and Devices (SISPAD) (IEEE,
     particle systems (Courier Corporation, 2012).                    New York, 2011) pp. 47–50.
[17] R. M. Pick, M. H. Cohen, and R. M. Martin, Physical         [26] G. K. Madsen and D. J. Singh, Computer Physics Com-
     Review B 1, 910 (1970).                                          munications 175, 67 (2006).
[18] S. Poncé, G. Antonius, P. Boulanger, E. Cannuccia,          [27] S. Poncé, E. R. Margine, C. Verdi, and F. Giustino, Com-
     A. Marini, M. Côté, and X. Gonze, Computational Mate-            puter Physics Communications 209, 116 (2016).
     rials Science 83, 341 (2014).                               [28] P. Giannozzi et al., Journal of Physics: Condensed Matter
[19] D. Chadi and M. L. Cohen, Physical Review B 8, 5747              21, 395502 (2009).
     (1973).                                                     [29] J. P. Perdew, K. Burke, and M. Ernzerhof, Physical Review
[20] D. G. Shankland, in Computational Methods in Band Theory         Letters 77, 3865 (1996).
     (Plenum, New York, 1971) p. 362.                            [30] P. Scherpelz, M. Govoni, I. Hamada, and G. Galli, Journal
[21] D. Koelling and J. Wood, Journal of Computational                of chemical theory and computation 12, 3523 (2016).
     Physics 67, 253 (1986).                                     [31] A. A. Mostofi, J. R. Yates, Y.-S. Lee, I. Souza, D. Vanderbilt,
[22] W. E. Pickett, H. Krakauer, and P. B. Allen, Physical            and N. Marzari, Computer physics communications 178,
     Review B 38, 2721 (1988).                                        685 (2008).
[23] Y. Sun, S. Boggs, and R. Ramprasad, Applied Physics         [32] H. Brooks, in Advances in electronics and electron physics,
     Letters 101, 132906 (2012).                                      Vol. 7 (Elsevier, 1955) pp. 85–182.
[24] B. Qiu, Z. Tian, A. Vallabhaneni, B. Liao, J. M. Mendoza,   [33] D. Chattopadhyay and H. Queisser, Reviews of Modern
     O. D. Restrepo, X. Ruan, and G. Chen, EPL (Europhysics           Physics 53, 745 (1981).
     Letters) 109, 57006 (2015), arXiv:1409.4862 (2014).         [34] F. Bassani and G. P. Parravicini, Electronic States and Optical
                                                                      Transitions in Solids (Pergamon press, New York, 1975).
You can also read