Long-range (Fröhlich and quadrupole) electron-phonon coupling: plan#

Status (Oct 25 2026): implemented (module elphon/long_range.py, hooks below, unit tests tests/unit/elphon/test_long_range.py); the SiC parity with EPW and the integration asset are in progress (see “Validation” at the end). Without a model the behaviour is unchanged: check_long_range_terms (elphon/do_pao_eph.py) refuses polar materials (|Z*| > 0.1) unless allow_missing_long_range=True and warns for non-polar insulators (quadrupole term missing). With long_range=LongRangeModel(...) the drivers subtract and restore the model; the Born charges and ε∞ come from either coupling source (load_epw_coupling, CoarseVertexSetup.long_range_tensors()), and LongRangeModel.from_coupling(coupling_dir, save_dir, quadrupoles=...) builds the model from the ph.x outputs and the nscf save.

Why the interpolation needs it#

The coupling of a polar crystal has a part that diverges as \(1/|q|\) (Fröhlich), and the dynamical matrix has a non-analytic dipole-dipole part (LO-TO splitting). Neither is a smooth function of q, so the Fourier interpolation from the coarse grid to the dense one cannot represent them: the dense \(g(k,q)\) near Γ is wrong by the whole long-range part and the LO branch collapses onto the TO. The standard cure (Verdi and Giustino, PRL 115, 176401 (2015); Sjakste et al., PRB 92, 054307 (2015); EPW lpolar) is subtract-and-restore: remove an analytic long-range model on the coarse grid, interpolate the short-range remainder, add the model back at every dense q. The quadrupole extension (Brunin et al., PRL 125, 136601 (2020); Poncé et al., PRR 3, 043022 (2021)) is the next order in q and is what non-polar semiconductors such as Si need for the small-q behaviour.

EPW’s implementation (QE 7.5 / EPW 6, EPW/src/longrange.f90)#

Vertex, rgd_blk_epw (3D branch), in the Wannier gauge, Rydberg units (\(e^2 = 2\)):

\[g^{\mathcal L}_{mn,\kappa\alpha}(k,q) = i\,\frac{4\pi e^2}{\Omega}\sum_{G\ne -q} \frac{e^{-(q+G)^2/4\alpha}}{(q+G)\cdot\varepsilon^\infty\cdot(q+G)}\, e^{-i(q+G)\cdot\tau_\kappa}\Big[(q+G)\cdot Z^*_\kappa\Big]_\alpha\,\delta_{mn} \;+\;\text{quadrupole: } -\tfrac{i}{2}(q+G)\cdot Q^\kappa_\alpha\cdot(q+G)\]
  • alph = 1.0 in units of \((2\pi/a)^2\) (metric), gmax = 14: the G box is the set with \((q+G)^2/4\alpha < 14\) (relative damping \(e^{-14}\)), built around the smallest \(|q+G|\) (find_min_qG).

  • The overlap \(\langle u_{m,k+q}|u_{n,k}\rangle\) is taken as \([U_{k+q}U_k^\dagger]_{mn}\), which is the identity in the Wannier gauge: EPW applies coeff_delta(imode) times \(\delta_{mn}\) (no Wannier-centre phase). A position-operator correction coeff_r (from the Wannier position matrix crrw/rrf) is used only in the 2D quadrupole branch.

  • zeu(ipol, i, na): first index field, second displacement, as PAOFLOW’s born_charges (nat, 3, 3).

  • Subtracted once per coarse (k, q) in the Wannier gauge, after the rotation by the U matrices and before the \(k \to R_e\) transform (bloch2wannier.f90); added back at every fine q in the Wannier gauge after the \(R_e \to k'\) transform and before the rotation to the fine Bloch states (wannier2bloch.f90).

  • Phonons, rgd_blk: the dipole-dipole term of the dynamical matrix with the same Ewald sum (dipole, dipole-quadrupole and quadrupole-quadrupole pieces), including the q = 0 self term subtracted from the diagonal (the ASR of the long-range part); subtracted from the coarse \(D(q)\) before \(q \to R_p\), added at the fine q. matdyn.x does the same (PHonon/PH/rigid.f90).

Mapping onto PAOFLOW#

The PAO gauge \(g^{\rm PAO}_{ij}(k,q) = \sum_{mn} A_{mi}(k+q)\,d_{mn}(k,q)\,A^*_{nj}(k)\) (cell-phase Bloch sums of atom-centred orbitals) plays the role of the Wannier gauge, so the long-range vertex is k-independent and diagonal in the orbitals: \(g^{\mathcal L}_{ij,\kappa\alpha}(q) = \delta_{ij}\, c_{\kappa\alpha}(q)\), one G-sum per q. Optional refinement: keep the orbital-centre phase \(e^{i(q+G)\cdot\tau_i}\) (exact for point orbitals, convention='centres'); the default mirrors EPW (convention='epw') because the convention must be the same on both grids and parity needs EPW’s. The quadrupole term enters the same coefficient. Metals keep \(Z^* = 0\) and are unaffected; free-carrier screening of the Fröhlich term in doped semiconductors (EPW lscreen) is out of scope.

New module elphon/long_range.py (~400 lines)#

Function

Role

LongRangeTensors

ε∞ (3,3), Z* (nat,3,3), optional Q (nat,3,3,3), τ (alat), alat, Ω, at/bg; from_dfpt(setup), from_epb(epw), read_quadrupole_fmt(path) (EPW’s quadrupole.fmt)

ewald_g_box(q_cart, bg, alpha=1.0, gmax=14.0)

the (q+G) list of EPW’s box, excluding q+G = 0

long_range_mode_coefficients(q_cryst, tensors, alpha, gmax, quadrupole)

\(c_{\kappa\alpha}(q)\), complex (3 nat,), Ry/bohr, Cartesian displacement index

long_range_vertex_pao(q_cryst, tensors, nawf, orbital_positions=None, convention='epw')

\((nawf, nawf, 3nat)\) to subtract/add

long_range_dynamical_matrix(q_cryst, tensors, alpha, gmax)

rgd_blk incl. the self term, \((3nat, 3nat)\), Ry mass-weighted after division by \(\sqrt{M_\kappa M_{\kappa'}}\)

Hooks (numerics unchanged when long_range=None)#

  1. Coarse subtraction in prepare_dense_vertex (do_pao_eph_dense_q.py): the per-q PAO-gauge vertex \(g_q(k)\) \((nk, nawf, nawf, 3nat)\) exists for both sources just before pao_vertex_to_cells ('dfpt': pao_vertex_k / unfolded_pao_vertices members; 'epw': inside vertex_from_epw before vertex_pao_R, which needs a long_range argument). Subtract \(g^{\mathcal L}(q)\) from every k. With unfold='irreducible' subtract at each star member after the rotation (the model is evaluated at the member’s q directly, so covariance is automatic).

  2. Dense restoration in band_vertex (elph_bloch.py): new argument long_range \((nawf, nawf, 3nat)\) added to gk after the Wigner-Seitz phases and before the mode and band rotations. Three callers compute it once per dense q: lambda_q_dense_ws_fast (Eliashberg, transport, pair coupling), carrier_scattering_dense_q (mobility/SERTA/MRTA), phonon_assisted_absorption_dense_q.

  3. Phonons: phonon_interp_from_dyn (dense drivers) and phonon_interpolator (phonon/qe_force_constants.py, q2r_matdyn) take long_range=tensors; subtract \(D^{\mathcal L}(q)\) from each coarse \(D(q)\) before \(q \to R_p\), add it in phonon_at_q before diagonalising. The is_polar warning of q2r_matdyn goes away. At exactly Γ the non-analytic limit is direction-dependent; dense grids only hit Γ itself (zeroed anyway) and paths may pass a direction.

  4. API: drivers gain long_range='dipole' | 'dipole+quadrupole' | None (default 'dipole' when \(|Z^*| > 0.1\), else None), quadrupoles= (array or quadrupole.fmt path), ewald_alpha=1.0; check_long_range_terms becomes the dispatcher (raise only when the tensors are missing, i.e. ph.x without epsil). The vertex cache key gets the long-range settings.

  5. Cost: one Ewald sum per q (~\(10^3\) G vectors, microseconds) on both grids; the subtraction is a broadcast over k. Memory unchanged.

Tests#

Unit (synthetic, fast, tests/unit/elphon/test_long_range.py):

  1. Ewald box: the sum is converged (α = 1 vs α = 0.5 agree to \(10^{-8}\) after the complementary real-space part is absent, i.e. the G box alone is converged at gmax = 14).

  2. Subtract-then-add identity on a coarse grid for a random vertex: prepare_dense_vertex with long_range followed by band_vertex at the coarse q returns the original \(g\) to \(10^{-12}\) (both conventions).

  3. Fröhlich limit of a cubic diatomic rigid-ion model (scalar \(Z^*\), ε∞, masses): along \(q \to 0\), \(|g^{\mathcal L}_{\rm LO}(q)| \to \frac{4\pi e^2}{\Omega}\frac{Z^*}{\varepsilon^\infty |q|} \frac{1}{\sqrt{2\mu\,\omega_{\rm LO}}}\) (with the mode factor), the TO modes get zero, and the LO-TO splitting of long_range_dynamical_matrix equals \(\omega_{\rm LO}^2-\omega_{\rm TO}^2 = 4\pi e^2 Z^{*2}/(\Omega\mu\varepsilon^\infty)\) (Lyddane-Sachs-Teller for the rigid-ion model).

  4. Symmetry covariance: \(g^{\mathcal L}(Sq)\) equals rotate_pao_vertex applied to \(g^{\mathcal L}(q)\) (consistency with the irreducible-q unfolding), and \(D^{\mathcal L}\) is Hermitian with the self term satisfying the ASR at q → 0.

  5. Quadrupole term: with \(Z^* = 0\) and a model \(Q\), \(|g| \propto |q|^0\) at small q and the dipole-quadrupole cross term of \(D^{\mathcal L}\) matches a finite-difference derivative of the dipole term.

Parity with EPW (local scripts, then an integration asset):

  1. 3C-SiC, the EPW test case test-suite/epw_polar (2 atoms, 20 Ry, scf 6³, ph.x 3³ with epsil, nscf 3³ with 8 bands, lpolar): run EPW with prtgkk = .true. on a short (k, q) list (filkf/filqf) through Γ, and compare \(\sum_{\rm degenerate}|g_{mn\nu}(k,q)|^2\) with PAOFLOW’s band_vertex on the same list from the minimal PAO basis (Si 3s3p, C 2s2p, 8 orbitals: the 4 valence bands are reproduced exactly). Expect agreement to the interpolation accuracy of the two bases (≈ 1 % away from the coarse q) and to \(10^{-6}\) at coarse q; the \(1/|q|\) behaviour near Γ is the decisive check. Asset ≈ 10 MB (sic/ next to si/ in the elphon tarball), CI test test_long_range_parity.py.

  2. Phonons: LO-TO splitting of SiC at Γ-approach and the whole dispersion against matdyn.x with the same Z*/ε∞ (q2r_matdyn path), to \(10^{-3}\) THz.

  3. GaN mobility (test-suite/epw_mob_polar, LDA, 2³/2³ coarse): room-temperature electron mobility of the PAOFLOW carrier driver vs EPW’s epw2.in/epw3.in output; literature values (Poncé et al., PRB 100, 085204 (2019)) as the physical reference. Not a CI test (minutes to hours).

  4. Si with quadrupoles (test-suite/epw_qdpt, quadrupole.fmt from Brunin et al.): mobility and the phonon-assisted absorption threshold change by a few per cent; compare with the Si example numbers.

Order of work#

  1. Module + unit tests 1-5 (2 days). 2. Hooks 1-3 and the dispatcher (2 days). 3. SiC parity and the asset (2 days, needs the EPW build already used for the DFPT port). 4. Quadrupole input and the Si check (1 day). 5. Docs: this note, the e-ph guide’s “Long-range” bullet, the Si example README (quadrupoles), check_long_range_terms docstring.

Validation (Oct 25-26 2026)#

Local scripts in Workspace/PAOFLOW/validation/dfpt_vertex_parity/sic{,6,40}/ (QE 7.5 / EPW 6.0, PAOFLOW PBE pseudopotentials, 3C-SiC at a = 8.237 bohr: ε∞ = 7.51, Z* = +2.65 (Si) / −2.83 (C)).

  • EPW’s prtgkk output needs a wide fsthick (with 10 eV the deep valence band is dropped from the printed block and the sums are not comparable); epw2_fs100.out is the reference.

  • At a coarse point (k = Γ, q = (1/3, 1/3, 0)): the DFPT band vertex of the 8-band and 4-band nscf runs and EPW’s .epb agree to \(4\times10^{-15}\); PAOFLOW’s dense path with the model reproduces them to 0.4 % (standard basis, 26 orbitals; 6 % with the minimal basis: projectability), EPW’s fine-grid values too.

  • Away from the coarse points on the 3³/3³ grids the two interpolations differ by up to 20 % for the LO mode near Γ and up to 60 % for the weaker modes: the short-range remainder is interpolated with different localised bases (four sp³ Wannier functions vs the PAO set), and the PAO long-range term carries the physical valence-subspace overlap \(\sum_{mn}|\langle u_{k+q}|u_k\rangle|^2/4\) (0.93 at \(q = 1/12\), 0.42 at \(q = 1/3\) along (1,1,0)) where the four-band Wannier model uses the identity. Both are consistent subtract-and-restore schemes; the convergence with the coarse grid (6³/6³ run) is the meaningful comparison, see below.

  • Convergence with the coarse grid (6³ k / 6³ q, same fine points, standard basis): PAOFLOW and EPW agree to 1-5 % at every (k, q, mode group), worst 12 % (the LA mode at |q| = 1/12 along (1,0,0)); the LO mode near Γ, dominated by the Fröhlich term, agrees to 2-5 %. The two subtract-and-restore schemes therefore converge to the same answer, and the residual on 3³ grids is the short-range interpolation, not the long-range model.

  • Integration test tests/integration/elphon/test_long_range_sic.py (3³ asset at 40 Ry): coarse-point identity < 2 %, LO coupling grows by > 2.5× when |q| halves only with the model and exceeds the plain interpolation by > 3× near Γ, EPW agreement < 2 % at coarse points and < 70 % elsewhere, LO-TO splitting within 3 % of the rigid-ion value from ε∞ and Z*.

  • q2r_matdyn (Oct 26 2026): long_range='auto' builds the model from the dyn files (LongRangeModel.from_dynamical_matrices, zasr='simple', QE’s Ewald parameter \(\alpha = (a/2\pi)^2\) in \((2\pi/a)^2\) units, which rigid.f90 uses with G in \(2\pi/a\); EPW uses 1), subtracts its force constants before the q → R transform, writes the T flag with α, ε∞ and Z* into the flfrc file, restores the term at every q and takes the direction-dependent limit at Γ along a path (gamma_limit_points, as matdyn.x). SiC 3³: matdyn.x reproduced to 0.01 cm⁻¹ (optical branches 1e-4), LO-TO at Γ 790/952 cm⁻¹ (tests/unit/phonon/test_qe_force_constants.py, data tests/unit/phonon/data/sic_dyn).

  • Quadrupoles from finite-q DFPT (elphon/quadrupoles.py, Oct 26 2026): instead of ABINIT’s long-wave DFPT, the multipole expansion of the induced electron density of single-q ph.x runs (fildrho): $\varepsilon_L(\hat q),[-\Omega\langle\tilde n_{\kappa\beta}\rangle e^{iq\cdot\tau_\kappa}

    • iq_\beta Z^{\rm ion}\kappa] = -iq\cdot Z^* - \tfrac12 q\cdot Q^\kappa\beta\cdot q + O(q^3)\(. The screening of a finite-q DFPT run acts on the total (ion + electron) charge, so the point ion that moves rigidly with the atom is added before dividing out \)\varepsilon_L\(; the imaginary part then reproduces the Born charges of the Γ run (and fixes the sign convention of the file), the real part at \)q^2\( is the quadrupole, free of the \)q^2\( dependence of \)\varepsilon_L\( and of the ionic term (imaginary, orders \)q^3\( and \)q\(). Six directions × two magnitudes (`single_q_phonon_inputs`), least squares on the symmetric quadratic forms, Richardson extrapolation in \)|q|\(, `write_quadrupole_fmt` for EPW's format and `LongRangeModel(quadrupoles=...)`. Driver: `coupling/quadrupoles.py` of every example (`PAOFLOW.elphon.workflows.quadrupoles`). Validation on Si (`Workspace/PAOFLOW/production/Si_mobility_quadrupoles/quadrupoles_si`: scf 12³ k, 38 Ry, PBE `Si.upf`, Γ run with `epsil` ε∞ = 13.11, twelve single-q runs with |q| = 0.02 and 0.04 (2π/a), 45 s each): the only non-zero components are \)Q_{x,yz}\( and its permutations, \)Q^{\rm Si_1}{x,yz} = -14.37\( (|q| = 0.02), \)-14.21\( (0.04), extrapolated \)-14.42\( e bohr, atom 2 opposite in sign, against 13.66 e bohr of Brunin et al., PRL 125, 136601 (2020) (ABINIT long-wave DFPT, different pseudopotential and cutoff); the fitted Born charges agree with the Γ run to 0.03 e (Z* ≈ 0 for Si, the residual is the \)q^3\( term of the \)|q| = 0.04\( samples amplified by ε∞). A charge-density file convention would have given \)Z^*{\rm el} - Z_{\rm ion}\( with the wrong sign and is detected by `calibrate_sign`. On the polar 3C-SiC (`examples/elphon/SiC/coupling/quadrupoles.py`, 6³ k, 60 Ry, PBE, |q| = 0.02 and 0.04, 10 s per run): \)Q^{\rm Si}{x,yz} = +7.49\( and \)Q^{\rm C}{x,yz} = -3.01\( e bohr (Si at the origin), against 7.41 and −2.63 fitted from small-q DFPT by Poncé et al., Phys. Rev. Research 3, 043022 (2021), Table IV (Abinit LDA: 7.03 / −1.42); the fitted Born charges reproduce the Γ-point \)Z^* = +2.65 / -2.83$ to 0.05 e, which checks the screening bookkeeping and the relative sign of the dipole and quadrupole terms (EPW’s coeff_delta = facq (i q·Z* + ½ q·Q·q)). Not done: convergence of Q with the k grid and cutoff.