betaFeedback

WorksPhysics and Astronomy

Efficient Interpolation of the Bethe-Salpeter Hamiltonian at Arbitrary Momenta

Abstract

We develop an efficient reciprocal-space method to reconstruct the static Bethe-Salpeter (BSE) Hamiltonian at arbitrary crystal momenta from coarse first-principles grids. Bloch states are interpolated within an optimally smooth Wannier subspace, while the screened Coulomb interaction is reconstructed from the dielectric response. This decouples the electron-hole transition momenta and transferred momentum from the underlying ab initio sampling. We benchmark the method on monolayer hBN, monolayer MoS2_2, and bulk GaAs against explicit ab initio calculations, finding excellent agreement for dense-grid Bloch states, dielectric screening, and BSE optical spectra. Furthermore, our cost analysis estimates speedups of up to 22×22\times on the grids considered here. Arbitrary transferred-momentum sampling gives direct access to the long-wavelength exciton dispersion: in hBN and MoS2_2, we resolve the nonanalytic behaviour induced by long-range electron-hole exchange at momenta well below the spacing of practical uniform grids. For bulk GaAs, locally refined transition-momentum sampling yields a converged binding energy for the threefold-degenerate lowest exciton, in close agreement with experiment and reference calculations, using far fewer BSE points than conventional uniform meshes. As the present framework provides exciton energies and amplitudes at arbitrary crystal momenta, it opens the way to incorporating ab initio excitonic effects into nonlocal electron-phonon vertices and to extensions towards exciton-phonon interactions and nonequilibrium exciton dynamics.

Introduction

The Bethe–Salpeter equation (BSE) provides a standard first-principles description of neutral excitations in solids, including optical spectra, exciton binding energies, and correlated electron–hole states [1, 2, 3, 4]. Its predictive accuracy, however, comes with a substantial computational cost. The electron–hole Hamiltonian couples transitions throughout the Brillouin zone, and converged calculations often require dense momentum sampling, particularly for spatially extended excitons [5, 6]. The screened Coulomb interaction introduces a further sampling bottleneck, since its momentum dependence can vary rapidly in the long-wavelength limit, especially in low-dimensional systems [7, 8].

The need for flexible momentum sampling becomes even more apparent away from the optical limit. Standard first-principles implementations rely on regular electronic momentum meshes, so a finite-momentum BSE calculation can be performed directly only when the electronic states required for a given transferred momentum Q\mathbf{Q}, corresponding to the total crystal momentum of the electron–hole pair, are represented on those grids. An arbitrary Q\mathbf{Q} will generally not satisfy this condition, and accessing it may require a new, typically denser, underlying calculation. This becomes particularly restrictive close to Γ\Gamma in two-dimensional systems, where long-range electron–hole exchange gives rise to a nonanalytic exciton dispersion and distinct longitudinal and transverse branches [9, 10, 11]. Resolving this regime requires transferred momenta well below the spacing of typical ab initio meshes, as recently demonstrated by the direct observation of a linear exciton dispersion in monolayer hBN [12, 13].

Dense interpolation in reciprocal space is already widely used in first-principles calculations. Maximally localised Wannier functions [14, 15, 16] provide a smooth representation of Bloch states from which band structures can be reconstructed on dense momentum grids. The same framework was subsequently extended to electron–phonon matrix elements [17, 18], while more recent work has shown that reciprocal-space quantities can be interpolated directly within an optimally smooth Wannier subspace, without an intermediate transformation to real space [19]. These developments now underpin dense-grid first-principles calculations of electron–phonon coupling, transport, spectroscopy, and carrier dynamics [20, 21, 22, 23].

A common smooth electronic representation becomes especially important when several interactions are treated within the same computational framework. Transport and nonequilibrium calculations require electronic states and interaction vertices to be evaluated on compatible dense momentum grids. This is the case, for example, in coupled Boltzmann treatments including both electron–phonon and electron–electron scattering [24], and in real-time descriptions of carrier dynamics in which both scattering channels are treated simultaneously [25, 26, 27, 28, 29]. Beyond interpolation, a common Wannier representation allows matrix elements associated with different interactions to be combined consistently.

The same requirement arises when excitons are coupled to lattice degrees of freedom. First-principles exciton–phonon frameworks combine BSE amplitudes at different transferred momenta with electron–phonon matrix elements to describe scattering and relaxation [30, 31], phonon-assisted absorption and emission [32], dynamical spectral renormalisation and phonon sidebands [33], exciton lifetimes and optical linewidths [34], and phonondriven exciton dynamics and thermalisation in real time [35, 36]. The symmetry properties of finite-momentum excitons further constrain these processes through selection rules for exciton–phonon coupling [37]. Recent formulations have also addressed the consistent combination of excitonic and electron–phonon quantities within a common electronic representation [38, 39]. More broadly, the interplay between electronic correlations and lattice dynamics motivates a common treatment of the corresponding interaction vertices [40]. The ability to reconstruct the BSE interaction at arbitrary momenta within the same smooth electronic representation as other interaction vertices is therefore relevant not only for equilibrium excitonic properties, but also for exciton transport and nonequilibrium dynamics.

Making such calculations practical also requires reducing the cost of constructing and solving the underlying many-body problem. Compact representations of the polarisation based on localised orbitals and formulations avoiding explicit sums over empty states have been developed for GW calculations [41, 42]. Maximally localised Wannier functions have been used in large-scale GW–BSE calculations to reduce their formal scaling [43] and to reduce the cost of real-time HSEX calculations of excitonic and nonlinear optical response [44]. Locality has further been exploited directly at the excitonic level, yielding sparse real-space BSE Hamiltonians and efficient evaluations of the electron–hole interaction [45]. Lanczos-chain and response-based approaches avoid explicit diagonalisation of the full response problem [46, 47], while density-matrix [48, 49] and finite-field formulations [50] reduce or eliminate explicit summations over virtual states and dielectric matrices.

Dense momentum sampling nevertheless remains a distinct computational bottleneck in BSE calculations, and several strategies have been developed to alleviate it. Early reciprocal-space schemes interpolated the electron–hole kernel from a coarse to a finer momentum mesh [51], with later formulations evaluating the interpolated matrix elements on the fly to reduce memory and computational cost in iterative BSE solvers [52]. Double-grid approaches instead retain the expensive kernel on the coarse mesh and use a finer auxiliary grid for the single-particle contribution and optical response [53, 54]. Nonuniform samplings, including locally refined and hybrid meshes, concentrate momentum points in the regions most relevant to the exciton wave function [5, 6], while interpolation of the screened interaction has been used to treat its sharp momentum dependence near the long-wavelength limit in reduced-dimensional systems [55]. At a different level, maximally localised exciton Wannier functions [56], constructed from Bloch excitons, allow exciton energies and wave functions to be interpolated as functions of the transferred momentum Q\mathbf{Q}.

In this work, we develop a reciprocal-space framework in which both the Bloch states and screened interaction can be reconstructed independently at arbitrary momenta. The approach is implemented in the code EPWlo [23].

Bloch states are interpolated within an optimally smooth Wannier subspace [19], while the screened interaction is reconstructed from an interpolated dielectric response [55]. This decouples the BSE sampling from the underlying first-principles grids, allowing both the transition-momentum grid k\mathbf{k} and the transferred momentum Q\mathbf{Q} to be chosen independently without repeating the expensive ab initioab\ initio calculation. At the same time, all electronic quantities remain tied to a common smooth representation, allowing the resulting exciton wave functions to be combined consistently with other interaction vertices defined in the same electronic representation.

We benchmark the method for monolayer hBN and MoS2_2 and bulk GaAs, first at the level of the interpolated Bloch states and dielectric response and then through BSE optical spectra on dense target meshes. We also quantify the computational gain obtained by replacing dense first-principles sampling with interpolation. Finally, we demonstrate the flexibility of the target BSE sampling with two applications: resolving the long-wavelength exciton dispersions of hBN and MoS2_2 at arbitrary transferred momenta, and converging the binding energies of weakly bound GaAs excitons using locally refined transition-momentum meshes.

Theoretical framework

We reconstruct the static BSE Hamiltonian at arbitrary exciton momenta by interpolating two quantities: the reciprocal-space components of the Bloch states in an optimally smooth Wannier representation [19] and the static dielectric screening. The direct and exchange electron–hole kernels are then calculated and the BSE Hamiltonian solved.

Rydberg atomic units (a.u.) are used throughout unless otherwise stated,

ℏ=2me=e22=4πϵ0=1,(1)\hbar= 2m_e = \frac{e^2}{2} = 4\pi\epsilon_0 = 1, \tag*{(1)}

with mem_e the electron mass, ee the magnitude of the elementary charge, and ϵ0\epsilon_0 the vacuum permittivity.

Finite-momentum Bethe–Salpeter Hamiltonian

The Bloch functions are written as

ψkn(r,σ)=1Neik⋅rukn(r,σ),(2)\psi_{\mathbf{k}n}(\mathbf{r},\sigma) = \frac{1}{\sqrt{N}}e^{i\mathbf{k}\cdot\mathbf{r}}u_{\mathbf{k}n}(\mathbf{r},\sigma), \tag*{(2)}

where σ\sigma labels the spinor component and NN is the number of primitive cells in the Born–von Kármán (BvK) supercell. The Bloch states ψkn\psi_{\mathbf{k}n} are normalised over the BvK volume V=NΩ\mathcal{V}=N\Omega, with Ω\Omega the primitive-cell volume, while the periodic parts uknu_{\mathbf{k}n} are normalised over Ω\Omega. For a uniform BZ sampling, NN coincides with the number NkN_{\mathbf{k}} of crystal momenta in the full BZ.

A neutral electron–hole excitation with transferred momentum Q\mathbf{Q}, corresponding to the total crystal momentum of the electron–hole pair, is represented in the transition basis

∣mnk;Q⟩=c^k+Qn†c^km∣0⟩,(3)|mn\mathbf{k};\mathbf{Q}\rangle= \hat{c}^{\dagger}_{\mathbf{k}+\mathbf{Q}n}\hat{c}_{\mathbf{k}m}|0\rangle, \tag*{(3)}

where c^kn\hat{c}_{\mathbf{k}n} (c^kn†\hat{c}^{\dagger}_{\mathbf{k}n}) annihilates (creates) an electron in Bloch state nn at momentum k\mathbf{k}, with single-particle energy εkn\varepsilon_{\mathbf{k}n}. The momentum k\mathbf{k} therefore labels the electron–hole transition, with the hole at k\mathbf{k} and the electron at k+Q\mathbf{k}+\mathbf{Q}.

We consider insulating and semiconducting systems within the resonant Tamm–Dancoff approximation. The transition space is restricted to excitations from occupied valence states vv to unoccupied conduction states cc, corresponding to m=vm=v and n=cn=c in Eq. (3). Neglecting the coupling between resonant and anti-resonant sectors then yields a Hermitian eigenvalue problem [4, 57]. For a given transferred momentum Q\mathbf{Q}, the BSE Hamiltonian is

HBSE(Q)=H0(Q)+Kd(Q)+Kx(Q),(4)H^{\mathrm{BSE}}(\mathbf{Q}) = H^{0}(\mathbf{Q}) + K^{d}(\mathbf{Q}) + K^{x}(\mathbf{Q}), \tag*{(4)}

with independent-particle contribution

Hvck,v′c′k′0(Q)=[εk+Qc−εkv]δvv′δcc′δkk′,(5)H^{0}_{vc\mathbf{k},v'c'\mathbf{k}'}(\mathbf{Q}) = [\varepsilon_{\mathbf{k}+\mathbf{Q}c} - \varepsilon_{\mathbf{k}v}]\delta_{vv'}\delta_{cc'}\delta_{\mathbf{k}\mathbf{k}'}, \tag*{(5)}

and screened direct and bare exchange kernels KdK^{d} and KxK^{x}, respectively [3, 4]. The momenta k\mathbf{k} and k′\mathbf{k}' label the two electron–hole transitions coupled by a matrix element of the BSE Hamiltonian.

The interaction kernels require transition densities between Bloch states,

ρmnG(k,q)=∑σ∫Ωdr ψk+qm∗(r,σ)ei(q+G)⋅rψkn(r,σ),(6)\rho^{\mathbf{G}}_{mn}(\mathbf{k},\mathbf{q}) = \sum_{\sigma}\int_{\Omega} d\mathbf{r}\,\psi^{*}_{\mathbf{k}+\mathbf{q}m}(\mathbf{r},\sigma)e^{i(\mathbf{q}+\mathbf{G})\cdot\mathbf{r}}\psi_{\mathbf{k}n}(\mathbf{r},\sigma), \tag*{(6)}

or, equivalently, in terms of the cell-periodic parts,

ρmnG(k,q)=∑σ∫Ωdr uk+qm∗(r,σ)eiG⋅rukn(r,σ).(7)\rho^{\mathbf{G}}_{mn}(\mathbf{k},\mathbf{q}) = \sum_{\sigma}\int_{\Omega}d\mathbf{r}\,u^{*}_{\mathbf{k}+\mathbf{q}m}(\mathbf{r},\sigma)e^{i\mathbf{G}\cdot\mathbf{r}}u_{\mathbf{k}n}(\mathbf{r},\sigma). \tag*{(7)}

Here, k\mathbf{k} and q\mathbf{q} lie in the first Brillouin zone, while G\mathbf{G} is a reciprocal-lattice vector.

The statically screened direct kernel is

Kvck,v′c′k′d(Q)=−wkwk′Ω∑GG′ρcc′G(k′+Q,q)WGG′(q)[ρvv′G′(k′,q)]∗,(8)K^{d}_{vc\mathbf{k},v'c'\mathbf{k}'}(\mathbf{Q}) = -\frac{\sqrt{w_{\mathbf{k}}w_{\mathbf{k}'}}}{\Omega}\sum_{\mathbf{G}\mathbf{G}'}\rho^{\mathbf{G}}_{cc'}(\mathbf{k}'+\mathbf{Q},\mathbf{q})W_{\mathbf{G}\mathbf{G}'}(\mathbf{q})\left[\rho^{\mathbf{G}'}_{vv'}(\mathbf{k}',\mathbf{q})\right]^{*}, \tag*{(8)}

where wkw_{\mathbf{k}} and wk′w_{\mathbf{k}'} are the Brillouin-zone integration weights associated with the transition momenta k\mathbf{k} and k′\mathbf{k}'. For uniform sampling over the full Brillouin zone, wk=wk′=1/Nkw_{\mathbf{k}}=w_{\mathbf{k}'}=1/N_{\mathbf{k}}. Here, q=k−k′\mathbf{q}=\mathbf{k}-\mathbf{k}' is the momentum exchanged between the two electron–hole transitions coupled by the direct kernel. It is distinct from the transferred momentum Q\mathbf{Q}, which is the total crystal momentum of the electron–hole pair. The quantity WGG′(q)W_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) is the static screened interaction and is obtained from the inverse dielectric matrix ϵGG′−1(q)\epsilon^{-1}_{\mathbf{G}\mathbf{G}'}(\mathbf{q}). The dielectric response is evaluated from the independent-particle response within the random-phase approximation, as described in Appendix A.

The bare exchange kernel reads

Kvck,v′c′k′x(Q)=gswkwk′Ω∑GρcvG(k,Q)vG(Q)[ρc′v′G(k′,Q)]∗,(9)K^{x}_{vc\mathbf{k},v'c'\mathbf{k}'}(\mathbf{Q}) = \frac{g_{s}\sqrt{w_{\mathbf{k}}w_{\mathbf{k}'}}}{\Omega}\sum_{\mathbf{G}}\rho^{\mathbf{G}}_{cv}(\mathbf{k},\mathbf{Q})v_{\mathbf{G}}(\mathbf{Q})\left[\rho^{\mathbf{G}}_{c'v'}(\mathbf{k}',\mathbf{Q})\right]^{*}, \tag*{(9)}

where vG(Q)v_{\mathbf{G}}(\mathbf{Q}) is the Fourier transform of the bare Coulomb interaction, whose explicit form for different dimensionalities is given in Appendix B. Depending on the response function of interest, the full interaction vv in Eq. (9) can be replaced by the modified interaction vˉ\bar{v}, defined by setting its macroscopic component to zero, vˉG=0(Q)=0\bar{v}_{\mathbf{G}=0}(\mathbf{Q})=0 [4]. The full interaction gives the polarisation χ\chi, whereas vˉ\bar{v} is used to obtain the modified polarisation χˉ\bar{\chi} entering the macroscopic optical dielectric function. In the finite-Q\mathbf{Q} exciton calculations considered below, we retain the full interaction; in two-dimensional systems, its G=0\mathbf{G}=0 component gives rise to the non-analytic long-wavelength exciton dispersion [10, 11]. In spin-degenerate scalar or collinear singlet calculations, gs=2g_{s}=2, while gs=1g_{s}=1 when spin is treated explicitly in a non-collinear spinorial formulation. In the conventional spin-conserving BSE, the exchange term vanishes in the triplet channel [3].

In the following, we show how the BSE Hamiltonian can be constructed on an arbitrary transition-momentum grid and for an arbitrary transferred momentum Q\mathbf{Q}. The procedure is:

  1. For each pair of transition momenta k\mathbf{k} and k′\mathbf{k}', we interpolate the reciprocal-space components of the Bloch states at the required momenta k\mathbf{k}, k+Q\mathbf{k}+\mathbf{Q}, k′\mathbf{k}', and k′+Q\mathbf{k}'+\mathbf{Q} within the optimally smooth Wannier subspace [19], and construct the corresponding transition densities.

  1. For the resulting exchanged momentum q=k−k′\mathbf{q}=\mathbf{k}-\mathbf{k}', we interpolate the screening response and reconstruct ϵGG′−1(q)\epsilon^{-1}_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) and WGG′(q)W_{\mathbf{G}\mathbf{G}'}(\mathbf{q}). The bare Coulomb interaction is never interpolated and is evaluated directly at the required momentum.

  1. We construct the direct and exchange kernels and assemble the BSE Hamiltonian of Eq. (4).

  1. We solve the BSE Hamiltonian to obtain the desired excitonic properties at the chosen transferred momentum Q\mathbf{Q}.

We now describe these steps in detail.

Interpolation of the BSE Hamiltonian at arbitrary momenta

Wannier interpolation of transition densities

The direct and exchange kernels in Eqs. (8) and (9) require transition densities involving Bloch states at momenta that do not necessarily belong to the original first-principles mesh. We obtain these states by directly interpolating the reciprocal-space components of the Bloch wave functions within the optimally smooth subspace (OSS) [19].

The periodic part of a Bloch function is expanded in plane waves as

ukn(r,σ)=1Ω∑Gukn(G,σ)eiG⋅r,(10)u_{kn}(\mathbf{r},\sigma) = \frac{1}{\sqrt{\Omega}}\sum_{\mathbf{G}} u_{kn}(\mathbf{G},\sigma)e^{i\mathbf{G}\cdot\mathbf{r}}, \tag*{(10)}

where ukn(r,σ)u_{kn}(\mathbf{r},\sigma) and ukn(G,σ)u_{kn}(\mathbf{G},\sigma) denote its real- and reciprocal-space representations, respectively.

Direct interpolation of the coefficients ukn(G,σ)u_{kn}(\mathbf{G},\sigma) in the band-eigenstate representation is hindered by the absence of a smooth representation of the Bloch eigenstates obtained independently at each crystal momentum. Their phases are arbitrary and, within degenerate or entangled manifolds, the states may also undergo unitary rotations between neighbouring momenta. The smoothness of the Bloch functions in reciprocal space is directly related to the localisation of the corresponding Wannier functions in real space [58, 59, 60].

We therefore use maximally localised Wannier functions (MLWFs) [14, 15, 16] to construct a smooth electronic representation for interpolation. In the general case of an entangled band manifold, the transformation to the OSS reads

uknOSS(G,σ)=∑mukm(G,σ)[Udis(k)U(k)]mn,(11)u_{kn}^{\mathrm{OSS}}(\mathbf{G},\sigma) = \sum_m u_{km}(\mathbf{G},\sigma)\left[U^{\mathrm{dis}}(\mathbf{k})U(\mathbf{k})\right]_{mn}, \tag*{(11)}

where Udis(k)U^{\mathrm{dis}}(\mathbf{k}) defines the disentangled subspace and U(k)U(\mathbf{k}) is the unitary transformation defining the optimally smooth Wannier representation. For an isolated group of bands, Udis(k)U^{\mathrm{dis}}(\mathbf{k}) is not required. The resulting OSS plane-wave components vary smoothly with k\mathbf{k} and can therefore be interpolated directly in reciprocal space [19].

The transition densities must ultimately be constructed in the band-eigenstate representation. At each target momentum, the interpolated OSS states are therefore rotated back to the Hamiltonian eigenstates. The Wannier Hamiltonian is first reconstructed at k\mathbf{k} using either conventional real-space Fourier interpolation [16] or direct Wannier interpolation in the OSS [19], and then diagonalised,

∑mHnm(k)Vml(k)=Vnl(k)εkl,(12)\sum_m H_{nm}(\mathbf{k})V_{ml}(\mathbf{k}) = V_{nl}(\mathbf{k})\varepsilon_{kl}, \tag*{(12)}

where Hnm(k)H_{nm}(\mathbf{k}) is the interpolated Wannier Hamiltonian, εkl\varepsilon_{kl} are its band eigenvalues, and the columns of the unitary matrix V(k)V(\mathbf{k}) are the corresponding eigenvectors. The plane-wave components of the resulting band eigenstates are then

ukn(G,σ)=∑mukmOSS(G,σ)Vmn(k).(13)u_{kn}(\mathbf{G},\sigma) = \sum_m u_{km}^{\mathrm{OSS}}(\mathbf{G},\sigma)V_{mn}(\mathbf{k}). \tag*{(13)}

The BSE transition densities can thus be constructed directly from the interpolated band eigenstates. To evaluate a generic transition density ρmn(k,q)\rho_{mn}(\mathbf{k},\mathbf{q}), we define the final-state momentum

kf=k+q(14)\mathbf{k}_f = \mathbf{k} + \mathbf{q} \tag*{(14)}

and decompose it as

kf=k~f+Gkf,(15)\mathbf{k}_f = \widetilde{\mathbf{k}}_f + \mathbf{G}_{\mathbf{k}_f}, \tag*{(15)}

where k~f\widetilde{\mathbf{k}}_f is the representative of kf\mathbf{k}_f in the first BZ and Gkf\mathbf{G}_{\mathbf{k}_f} is the reciprocal-lattice vector required to fold kf\mathbf{k}_f into it. If kf\mathbf{k}_f already lies in the first BZ, then k~f=kf\widetilde{\mathbf{k}}_f = \mathbf{k}_f and Gkf=0\mathbf{G}_{\mathbf{k}_f} = 0. Substituting the plane-wave expansion into Eq. (7) gives

ρmnG(k,q)=∑G′,σ[u~kfm(G′+G+Gkf,σ)]∗ukn(G′,σ).(16)\rho_{mn}^{\mathbf{G}}(\mathbf{k},\mathbf{q}) = \sum_{\mathbf{G}',\sigma}\left[\widetilde{u}_{k_fm}(\mathbf{G}'+\mathbf{G}+\mathbf{G}_{\mathbf{k}_f},\sigma)\right]^*u_{kn}(\mathbf{G}',\sigma). \tag*{(16)}

This equation provides the transition densities entering the BSE kernels directly from the interpolated reciprocal-space Bloch states. Following the direct Wannier interpolation approach in the OSS [19], the reciprocal-space interpolation scheme itself can be chosen independently of the Wannier construction. Details of the scheme adopted here are given in Appendix C.

Interpolation of dielectric screening

The direct interpolation of the inverse dielectric function ϵ−1\epsilon^{-1} is inconvenient because its momentum dependence retains the rapid variation associated with the bare Coulomb interaction, particularly in the long-wavelength limit of reduced-dimensional systems. Following [55], we instead interpolate a function in which the explicit Coulomb dependence is removed. Defining

dGG′(q)=vG(q)vG′(q),(17)d_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) = \sqrt{v_{\mathbf{G}}(\mathbf{q})v_{\mathbf{G}'}(\mathbf{q})}, \tag*{(17)}

and

SGG′(q)=ϵGG′−1(q)−δGG′,(18)S_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) = \epsilon^{-1}_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) - \delta_{\mathbf{G}\mathbf{G}'}, \tag*{(18)}

we introduce

fGG′(q)=SGG′(q)dGG′(q)[SGG′(q)+1].(19)f_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) = \frac{S_{\mathbf{G}\mathbf{G}'}(\mathbf{q})}{d_{\mathbf{G}\mathbf{G}'}(\mathbf{q})\left[S_{\mathbf{G}\mathbf{G}'}(\mathbf{q})+1\right]}. \tag*{(19)}

All operations in Eq. (19) are elementwise; in particular, the scalar 11 in the denominator is not a Kronecker delta and no matrix inversion is implied.

The function fGG′(q)f_{\mathrm{GG}^{\prime}}(\mathbf{q}) has a much smoother momentum dependence than ϵ−1\epsilon^{-1}, since the singular or rapidly varying Coulomb factors have been removed. In the absence of local-field effects, its head reduces to

f00(q)=χ000(q),(20)f_{00}(\mathbf{q})=\chi^{00}_{0}(\mathbf{q}), \tag*{(20)}

while in the general case it also contains the effects of local-field coupling [55]. This construction is particularly advantageous for two-dimensional systems with a truncated Coulomb interaction and is also used for the non-head components in bulk systems. For three-dimensional bulk systems, however, we interpolate the head ϵ00−1(q)\epsilon^{-1}_{00}(\mathbf{q}) directly. This quantity is smooth near Γ\Gamma, whereas interpolating f00f_{00} would require its quadratic small-qq behaviour to be reproduced with high accuracy in order to cancel the 1/q21/q^{2} Coulomb divergence.

At an arbitrary target momentum q\mathbf{q}, the function fGG′(q)f_{\mathrm{GG}^{\prime}}(\mathbf{q}) is interpolated using the scheme described in Appendix C, except for the bulk head treated as described above. The Coulomb factors are then evaluated directly at the same momentum and the inverse dielectric matrix is reconstructed as

ϵGG′−1(q)=δGG′+fGG′(q)dGG′(q)1−fGG′(q)dGG′(q).(21)\epsilon^{-1}_{\mathrm{GG}^{\prime}}(\mathbf{q})=\delta_{\mathrm{GG}^{\prime}}+\frac{f_{\mathrm{GG}^{\prime}}(\mathbf{q})d_{\mathrm{GG}^{\prime}}(\mathbf{q})}{1-f_{\mathrm{GG}^{\prime}}(\mathbf{q})d_{\mathrm{GG}^{\prime}}(\mathbf{q})}. \tag*{(21)}

The screened interaction follows from

WGG′(q)=vG(q) ϵGG′−1(q) vG′(q).(22)W_{\mathrm{GG}^{\prime}}(\mathbf{q})=\sqrt{v_{\mathrm{G}}(\mathbf{q})}\,\epsilon^{-1}_{\mathrm{GG}^{\prime}}(\mathbf{q})\,\sqrt{v_{\mathrm{G}^{\prime}}(\mathbf{q})}. \tag*{(22)}

These reconstruction formulas are again understood elementwise. The bare Coulomb interaction is never interpolated, but evaluated analytically at each required momentum.

Reconstruction of the static BSE Hamiltonian at arbitrary momentum

The reconstruction of the BSE Hamiltonian from coarse first-principles inputs is summarised in Figure 1. Starting from the Bloch states and static dielectric response, the electronic states are transformed into an optimally smooth Wannier representation, while the auxiliary screening response fGG′(q)f_{\mathrm{GG}^{\prime}}(\mathbf{q}) is constructed from the dielectric matrix. Both quantities are then interpolated independently at the momenta required by the target BSE calculation.

Figure 1. Schematic representation of the reciprocal-space BSE interpolation scheme. The Bloch states are obtained from density-functional theory using pw.x from Quantum ESPRESSO [61, 62], while the static dielectric matrix is computed with our epsilon_lr.q.x implementation within Quantum ESPRESSO. The Bloch states are transformed into the optimally smooth Wannier subspace using Wannier90 [63], yielding the matrices U(k)U(\mathbf{k}) and Udis(k)U^{\mathrm{dis}}(\mathbf{k}). The smooth Bloch states and screening response fGG′(q)f_{\mathrm{GG}^{\prime}}(\mathbf{q}) are independently interpolated onto the target momenta. Reconstruction of the band eigenstates and screened interaction, construction of the BSE kernels, and solution of the BSE are performed within EPIq [23], yielding the exciton energies, amplitudes, and optical response.

The interpolated Bloch states are rotated back to the band-eigenstate representation using Eq. (13), from which the transition densities are constructed through Eq. (16). For each pair of transitions, the screened interaction is reconstructed at the exchanged momentum q=k−k′\mathbf{q}=\mathbf{k}-\mathbf{k}^{\prime}, while the bare Coulomb interaction is evaluated directly at Q\mathbf{Q}. These quantities determine the direct and exchange kernels of Eqs. (8) and (9).

The screened interaction entering the direct kernel can also be averaged over the reciprocal-space cell associated with q\mathbf{q} to improve convergence with momentum sampling [55]. The independent-particle contribution H0H^{0} is constructed from the chosen band energies,

which can be obtained from Wannier-interpolated Kohn–Sham bands, optionally supplemented by a scissor correction, or from interpolated quasiparticle energies. Combining these contributions yields the BSE Hamiltonian of Eq. (4), which can be constructed and solved for arbitrary transition momenta k\mathbf{k}, k′\mathbf{k}^{\prime} and transferred momentum Q\mathbf{Q}, independently of the underlying first-principles grids.

Solution of the Bethe–Salpeter equation

Once HBSE(Q)H^{\mathrm{BSE}}(\mathbf{Q}) has been constructed, we employ different iterative solvers depending on the quantity of interest, avoiding explicit diagonalisation of the full electron–hole Hamiltonian. Optical spectra are computed using the Lanczos–Haydock recursion [64, 65, 66]. Starting from the optical transition vector, the method constructs a reduced tridiagonal representation of the BSE Hamiltonian through successive matrix–vector products. The resulting continued fraction directly yields the frequency-dependent optical response and, in particular, the imaginary part of the dielectric function, ϵ2(Q,ω)=Im⁡ϵ(Q,ω)\epsilon_2(\mathbf{Q},\omega)=\operatorname{Im}\epsilon(\mathbf{Q},\omega), without requiring the full excitonic eigenspectrum.

When individual excitonic states are required, we instead employ Davidson iterative diagonalisation [67]. The reduced subspace is iteratively refined using the residuals of the approximate eigenvectors, yielding selected low-lying exciton energies Eλ(Q)E_\lambda(\mathbf{Q}) and amplitudes Avckλ(Q)A^\lambda_{vck}(\mathbf{Q}) without diagonalising the full BSE Hamiltonian.

Computational details

Density-functional theory calculations are performed with Quantum ESPRESSO [61, 62] within the local-density approximation (LDA) [68], using optimised norm-conserving Vanderbilt pseudopotentials [69]. Plane-wave kinetic-energy cutoffs of 80 Ry are used for monolayer hBN and MoS2_2, and 60 Ry for bulk GaAs. The periodic images of the monolayers are separated by 17 Å of vacuum and we employ a truncated Coulomb interaction [70]. Ground-state charge densities are computed on Γ\Gamma-centred Monkhorst–Pack meshes [71] of 18×18×118 \times18 \times1 for hBN and MoS2_2 and 12×12×1212 \times12 \times12 for GaAs. Spin–orbit coupling is included for MoS2_2.

Maximally localised Wannier functions are constructed with Wannier90 [63], retaining 10, 22, and 14 Wannier functions from manifolds of 20, 34, and 20 bands for hBN, MoS2_2, and GaAs, respectively. Several uniform kk meshes are used to assess the dependence of the interpolation accuracy on the source sampling. The reciprocal-space components of the Bloch states are extracted from the same states used for the Wannierisation through a modified version of the wfck2r.x utility of Quantum ESPRESSO. The quality of the Wannier representation is assessed in Appendix § 1.2.4, where the ab initio and Wannier-interpolated band structures are compared for the valence and conduction manifolds relevant to the BSE calculations. The initial orbital projections are also reported there.

The static RPA dielectric response used for the screening interpolation is computed with epsilon_q.x, a utility implemented by us within Quantum ESPRESSO. The independent-particle response includes 1000 empty states and uses block-size cutoffs of 20 Ry for hBN and MoS2_2 and 15 Ry for GaAs. The internal Brillouin-zone kk-point summation entering the independent-particle response, see Eq. (35), is converged on 48×48×148 \times48 \times1, 36×36×136 \times36 \times1, and 24×24×2424 \times24 \times24 meshes for hBN, MoS2_2, and GaAs, respectively. For the monolayers, the Coulomb interaction is truncated along the non-periodic direction. Several uniform qq meshes are used to assess the dependence of the screening interpolation on the source sampling.

The interpolated calculations are validated against explicit ab initio BSE calculations performed with YAMBO [72, 73]. Quasiparticle corrections are approximated by rigid scissor shifts of 2.8, 0.8, and 0.5 eV for hBN, MoS2_2, and GaAs, respectively. In the ab initio calculations, the scissor shifts are applied directly on the corresponding mesh, whereas in the interpolated calculations they are applied after Wannier interpolation onto the target mesh. The screened direct kernel is truncated at 10 Ry for hBN and MoS2_2 and 8 Ry for GaAs, while a 40 Ry cutoff is used for the bare exchange term in all cases. The transition space contains the two highest valence and two lowest conduction bands for hBN, four valence and four conduction bands for MoS2_2, and three valence and four conduction bands for GaAs. Optical spectra are evaluated with constant broadenings of 0.10, 0.075, and 0.15 eV for hBN, MoS2_2, and GaAs, respectively.

The only external codes required by the present implementation are Quantum ESPRESSO for the DFT Bloch states and Wannier90 for the Wannier representation. The dielectric response is computed with epsilon_q.x, while Bloch-state and screening interpolation, construction of the BSE kernels, and solution of the BSE are performed within EPIq [23]. No external many-body perturbation theory code is therefore required; YAMBO is used only to generate independent reference calculations for validation.

Benchmarks

Interpolation accuracy of Bloch states and dielectric screening

We first benchmark the Bloch-state and dielectric-screening interpolation separately before assessing their combined accuracy at the BSE level. Monolayer MoS2_2 is used as a representative case. Starting from a 12×12×112 \times12 \times1 source mesh, both quantities are interpolated on a 36×136 \times1 target grid and compared point by point with explicit first-principles calculations. This directly tests the interpolation at momenta absent from the source grid.

Figure 2 assesses the Bloch-state interpolation through the quantity

Figure 2. Interpolation benchmark for the Bloch-state plane-wave components of monolayer MoS2_2, expressed through the unitary-invariant quantity SG(k)S_{\mathbf{G}}(\mathbf{k}) defined in Eq. (23). (a) SG(k)S_{\mathbf{G}}(\mathbf{k}) for G=(0,0,0)\mathbf{G}=(0,0,0) from the explicit 36×36×136 \times36 \times1 calculation (left), interpolation from a 12×12×112 \times12 \times1 source mesh (centre), and absolute interpolation error (right). (b) SG(k)S_{\mathbf{G}}(\mathbf{k}) along the KK–Γ\Gamma–MM path for three representative reciprocal-lattice vectors. Red lines denote the interpolated values, open circles the explicit 36×36×136 \times36 \times1 reference, and blue diamonds the 12×12×112 \times12 \times1 source points. The lower panels show the absolute interpolation error at the reference momenta.

SG(k)=∑nσ∣ukn(G,σ)∣2,(23)S_G(\mathbf{k}) = \sum_{n\sigma} \left|u_{kn}(\mathbf{G},\sigma)\right|^2 , \tag*{(23)}

where the sum runs over all bands spanning the Wannier subspace and over spin. Since SG(k)S_G(\mathbf{k}) is invariant under unitary rotations within the subspace, it probes the plane-wave content of the interpolated Bloch manifold independently of the representation of the individual eigenstates.

For G=(0,0,0)\mathbf{G}=(0,0,0), the interpolated Brillouin-zone map is essentially indistinguishable from the explicit calculation, with a maximum relative error below 1% over the full target grid, as shown in Figure 2(a). Similarly, along the KK–Γ\Gamma–MM path, the interpolated curves closely reproduce the reference values for the representative reciprocal-lattice vectors shown in Figure 2(b). The interpolation therefore accurately reproduces the plane-wave content of the Bloch subspace throughout the Brillouin zone.

The dielectric response provides a more stringent test because of its rapid variation in the long-wavelength region. Figure 3 compares the reconstructed and explicit inverse dielectric matrix ϵGG′−1(q)\epsilon^{-1}_{\mathbf{G}\mathbf{G}'}(\mathbf{q}). The quantity directly interpolated is the screening response fGG′(q)f_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) introduced in Eq. (19), from which ϵGG′−1(q)\epsilon^{-1}_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) is reconstructed using Eq. (21). The head element, G=G′=(0,0,0)\mathbf{G}=\mathbf{G}'=(0,0,0), is accurately reproduced over most of the Brillouin zone, as shown in Figure 3(a), with similarly good agreement for representative diagonal and off-diagonal elements along KK–Γ\Gamma–MM in Figure 3(b).

Figure 3. Interpolation benchmark for the static dielectric response of monolayer MoS2_2. (a) Real part of the head element ϵGG′−1(q)\epsilon^{-1}_{\mathbf{G}\mathbf{G}'}(\mathbf{q}), with G=G′=(0,0,0)\mathbf{G}=\mathbf{G}'=(0,0,0), from the explicit 36×36×136 \times36 \times1 calculation (left), reconstructed from the screening response interpolated from a 12×12×112 \times12 \times1 source mesh (centre), and modulus of the interpolation error (right). (b) Representative elements of Re⁡ϵGG′−1(q)\operatorname{Re}\epsilon^{-1}_{\mathbf{G}\mathbf{G}'}(\mathbf{q}) along the KK–Γ\Gamma–MM path. Red lines denote the reconstructed values, open circles the explicit 36×36×136 \times36 \times1 reference, and blue diamonds the 12×12×112 \times12 \times1 source points. The lower panels show the modulus of the complex interpolation error at the reference momenta.

The largest deviations occur for the head element close to Γ\Gamma, where the dielectric response varies on a momentum scale only sparsely resolved by the 12×12×112 \times12 \times1 source mesh. Even there, the maximum relative error remains around 5% and decreases rapidly away from the long-wavelength limit. The screening reconstruction is therefore already accurate for this deliberately coarse source grid, while its convergence remains sensitive to the source sampling. Refining this mesh progressively resolves the sharp variation around Γ\Gamma and systematically improves the accuracy, as examined below.

Together, Figure 2 and Figure 3 show that Bloch states and dielectric screening can be accurately interpolated well beyond the original first-principles sampling, including at momenta absent from the source grids. This enables the BSE sampling to be decoupled from the underlying ab initio calculation. We next assess the combined accuracy at the level of the full electron–hole Hamiltonian.

Reconstructed BSE optical spectra

Having benchmarked the Bloch states and dielectric screening separately, we now assess their combined accuracy at the level of the BSE optical response. Figure 4 compares spectra obtained from the reconstructed BSE Hamiltonian with explicit ab initio calculations for monolayer hBN and MoS2_2 and bulk GaAs. The reference calculations use 48×48×148 \times48 \times1, 36×36×136 \times36 \times1, and 24×24×2424 \times24 \times24 meshes for hBN, MoS2_2, and GaAs, respectively, and the reconstructed spectra are evaluated on the same target grids. The Bloch states and dielectric screening entering the reconstructed Hamiltonian are instead obtained from substantially coarser source meshes.

Figure 4. BSE optical spectra obtained from the reciprocal-space interpolation scheme compared with explicit ab initio calculations for monolayer hBN and MoS2_2 and bulk GaAs. Black solid lines show the explicit Yambo reference spectra, calculated on 48×48×148 \times48 \times1, 36×36×136 \times36 \times1, and 24×24×2424 \times24 \times24 meshes, respectively. Blue dashed and red dash-dotted lines show reconstructed spectra; the legend indicates the source and target BSE meshes as source →\to target. (a) Convergence with respect to the source q mesh used for the dielectric screening, with the Bloch-state source mesh fixed at the coarser sampling. (b) Convergence with respect to the source k mesh used for the Bloch states, with the dielectric screening fixed at the finer q sampling. Insets magnify representative excitonic features where the residual differences are most visible.

To separate the sources of interpolation error, we first examine convergence with respect to the source q\mathbf{q} mesh used for the dielectric screening, as shown in Figure 4(a), while keeping the Bloch-state mesh fixed at the coarser sampling. We then refine the Bloch-state source k\mathbf{k} mesh in Figure 4(b), with the screening fixed at the finer q\mathbf{q} sampling.

For all three materials, the reconstructed spectra closely reproduce the explicit reference, including excitonic line shapes, peak positions, and oscillator strengths, already from the coarser source meshes. The remaining differences are mainly small energy shifts and localised changes in spectral intensity, highlighted by the insets, and decrease systematically upon refinement.

Convergence with respect to the screening mesh in Figure 4(a) is the more demanding of the two, consistent with the benchmark of Figure 3. In hBN, refining the source q mesh from 12×12×112 \times12 \times1 to 24×24×124 \times24 \times1 substantially reduces the shift of the dominant excitonic peak. The improvement is even stronger for MoS2_2, where the main excitonic structure becomes essentially indistinguishable from the explicit result upon refinement from 12×12×112 \times12 \times1 to 18×18×118 \times18 \times1. For GaAs, the spectrum is already well reproduced from the 8×8×88 \times8 \times8 source mesh, with only minor changes upon refinement to 12×12×1212 \times12 \times12. These trends reflect the need to resolve the momentum dependence of the dielectric response on the source grid, particularly in low-dimensional systems.

Figure 4(b) shows the convergence with respect to the Bloch-state k mesh, with the screening fixed at the finer sampling. For hBN, the two reconstructed spectra are nearly indistinguishable, indicating that the remaining discrepancy is dominated by the screening. In MoS2_2 and GaAs, refining the k mesh mainly corrects small localised intensity differences, while leaving the overall line shape and peak positions essentially unchanged. This is consistent with the sub-percent errors found in the unitary-invariant Bloch-state benchmark of Figure 2 and confirms that the interpolated plane-wave components accurately reproduce the transition densities entering the electron–hole kernels.

Because the residual differences are dominated by small energy shifts and localised redistributions of spectral weight, a pointwise relative error is not a particularly meaningful measure of spectral agreement. We therefore quantify the convergence using the first Wasserstein distance between the normalised spectra. Defining the cumulative spectral weight

F(ω)=∫−∞ωdω′ ϵ2(ω′)∫−∞+∞dω′ ϵ2(ω′).(24)F(\omega) = \frac{ \int_{-\infty}^{\omega} d\omega'\,\epsilon_2(\omega')}{ \int_{-\infty}^{+\infty} d\omega'\,\epsilon_2(\omega')}. \tag*{(24)}

the spectral distance is

W1=∫dω ∣F(ω)−Fref(ω)∣.(25)W_1 = \int d\omega\,\left|F(\omega)-F_{\mathrm{ref}}(\omega)\right|. \tag*{(25)}

For two identical line shapes rigidly displaced by ΔE\Delta E, this gives simply W1=∣ΔE∣W_1 = \lvert\Delta E\rvert. The values in Table I quantify the convergence visible in Figure 4. Refining the screening mesh reduces W1W_1 from

MaterialMesh typeSource-mesh refinementW1W_1 (meV)
hBNqq12×12×1→24×24×112 \times12 \times1 \rightarrow24 \times24 \times124.1→9.524.1 \rightarrow9.5
kk12×12×1→24×24×112 \times12 \times1 \rightarrow24 \times24 \times19.5→9.09.5 \rightarrow9.0
MoS2_2qq12×12×1→18×18×112 \times12 \times1 \rightarrow18 \times18 \times123.1→1.723.1 \rightarrow1.7
kk12×12×1→18×18×112 \times12 \times1 \rightarrow18 \times18 \times11.7→0.31.7 \rightarrow0.3
GaAsqq8×8×8→12×12×128 \times8 \times8 \rightarrow12 \times12 \times128.9→8.08.9 \rightarrow8.0
kk8×8×8→10×10×108 \times8 \times8 \rightarrow10 \times10 \times108.0→4.98.0 \rightarrow4.9

caption: Convergence of the reconstructed BSE optical spectra with respect to the source qq mesh used for the dielectric screening and the source kk mesh used for the Bloch states. During the qq-mesh convergence, the Bloch-state source mesh is kept fixed at the coarser sampling; during the subsequent kk-mesh convergence, the screening is kept fixed at the finer sampling. The arrows indicate refinement of the source mesh, while the target BSE meshes are those specified in Figure 4. W1W_1 is the first Wasserstein distance from the explicit ab initio BSE spectrum, defined in Eq. (25).

:::

24.1 to 9.5 meV in hBN and from 23.1 to 1.7 meV in MoS2_2, while in GaAs it decreases from 8.9 to 8.0 meV. Subsequent refinement of the Bloch-state mesh reduces the distance further to 9.0, 0.3, and 4.9 meV for hBN, MoS2_2, and GaAs, respectively. The relative importance of the two interpolated quantities varies among the materials, but in all cases the final spectra closely reproduce the explicit reference.

These results provide the central validation of the reconstruction scheme. Starting from coarse Bloch-state and screening meshes, the method reproduces the optical response of much denser explicit calculations without interpolating the final spectrum or excitonic eigenvalues. Instead, the BSE Hamiltonian itself is reconstructed from the interpolated Bloch states and screened interaction.

Results

Computational cost and scaling

The accuracy demonstrated above is accompanied by a substantial reduction in the cost of dense BSE calculations. Figure 5 compares the cost of obtaining BSE optical spectra for monolayer MoS2_2 and bulk GaAs from explicit calculations on the target mesh and from BSE Hamiltonians reconstructed using Bloch states and screening interpolated from coarser source meshes. Timings are reported in core hours and decomposed into the first-principles calculation of Bloch states and dielectric response, interpolation to the target momenta, construction of the transition densities and BSE kernel, and iterative solution of the BSE. Reference timings are measured on 36×36×136 \times36 \times1 for MoS2_2 and 24×24×2424 \times24 \times24 for GaAs; the remaining costs are obtained from the scaling relations discussed below.

Figure 5. Computational cost of explicit ab initio and reconstructed BSE calculations as a function of the target momentum mesh for monolayer MoS2_2 (top) and bulk GaAs (bottom). The cost is reported in core hours and decomposed into the first-principles calculation of the Bloch states and dielectric response, interpolation to the target momenta, construction of the transition densities and BSE kernel, and iterative solution of the BSE. Hatched bars denote explicit calculations and filled bars reconstructed calculations. In the latter, the Bloch states and dielectric response are computed only on fixed source meshes, 18218^2 for MoS2_2 and 838^3 for GaAs, and subsequently interpolated to the target sampling. Numbers above the bars give the ratio between the total explicit and reconstructed computational costs.

The first-principles contribution includes the non-self-consistent-field (NSCF) calculation of the Bloch states

and the dielectric response. In the explicit calculations, the NSCF kk mesh is refined together with the qq mesh required for screening. For fixed band and plane-wave cutoffs, the NSCF cost scales approximately with the number of Bloch momenta, NkN_k. The independent-particle polarizability of Eq. (35) must instead be evaluated for every momentum transfer qq, involving an internal kk-point summation. In our cost estimates, this internal sampling is kept fixed at 36×36×136 \times36 \times1 for MoS2_2 and 24×24×2424 \times24 \times24 for GaAs, as specified in Sec. IV. The dielectric-response cost therefore scales linearly with the number of sampled momentum transfers, NqN_q, giving

TFP∝Nk+Nq.(26)T_{\mathrm{FP}} \propto N_k + N_q. \tag*{(26)}

This is a conservative assumption: in conventional implementations such as Yambo, the internal k\mathbf{k} mesh entering χ0\chi^0 is typically refined together with the BSE sampling, increasing the cost of the explicit screening calculation. By contrast, in the reconstructed calculations, the Bloch states and dielectric response are computed only on the fixed 18×18×118 \times18 \times1 and 8×8×88 \times8 \times8 source meshes for MoS2_2 and GaAs, respectively, so their cost remains constant as the target mesh is refined. The speedups reported below therefore represent conservative estimates of the computational advantage of interpolation.

The interpolation cost also grows with the target sampling. Bloch states are reconstructed once at each required k\mathbf{k} point and stored, while the dielectric response is reconstructed independently at each required q\mathbf{q}. Hence,

Tinterp∝Nk+Nq.(27)T_{\mathrm{interp}} \propto N_k + N_q. \tag*{(27)}

The subsequent BSE-kernel construction has the same momentum scaling in the explicit and reconstructed calculations. On the uniform meshes considered here, the transition densities and interaction matrix elements must be evaluated over all (k,q)(\mathbf{k},\mathbf{q}) pairs. For fixed band and plane-wave cutoffs, this gives

Tkernel∝NkNq.(28)T_{\mathrm{kernel}} \propto N_k N_q. \tag*{(28)}

The iterative BSE solution is dominated by Hamiltonian applications over the electron–hole transition space and, for fixed numbers of valence and conduction bands, scales approximately as

TBSE∝Nk2.(29)T_{\mathrm{BSE}} \propto N_k^2. \tag*{(29)}

For the uniform samplings considered here, NqN_q grows together with NkN_k, so both kernel construction and BSE solution scale quadratically with the target mesh size. The resulting trend is shown in Figure 5. At the source meshes, the explicit and reconstructed calculations have comparable costs because they require the same coarse first-principles ingredients. The advantage grows rapidly as the target mesh is refined. For MoS2_2, the speed-up reaches approximately 22×22\times at 96×96×196 \times96 \times1. In the explicit calculation, the increasing cost is dominated by the dielectric response on the dense q\mathbf{q} mesh together with the NSCF calculation of the corresponding Bloch states. In the reconstructed calculation, both first-principles contributions remain fixed at the source-mesh cost, while the interpolation overhead remains comparatively small. For bulk GaAs, the speed-up reaches about 10×10\times at 28×28×2828 \times28 \times28. At dense samplings, the gain becomes smaller because construction of the transition densities

and BSE kernel accounts for an increasing fraction of the reconstructed calculation. This stage is common to both approaches and retains the same NkNqN_kN_q scaling. The interpolation scheme therefore does not alter the asymptotic cost of assembling the BSE Hamiltonian, but removes the need to recompute Bloch states and dielectric screening from first principles on each dense target mesh. The BSE Hamiltonian can therefore be constructed on dense target meshes from coarse first-principles inputs at substantially lower cost than the corresponding explicit ab initio calculation. Together with the spectral agreement of Figure 4, this makes flexible momentum sampling practical for the applications discussed below.

Finite-momentum excitons in the long-wavelength limit

A key advantage of the present interpolation scheme is that the transferred momentum Q\mathbf{Q} can be chosen independently of the underlying first-principles grids. Figure 6 shows the exciton dispersions of monolayer hBN and MoS2_2 along the full MM–Γ\Gamma–KK path, with a finely resolved region around Γ\Gamma. The exciton energies themselves are not interpolated: for each target Q\mathbf{Q}, the required Bloch states and screened interaction are reconstructed and a new BSE Hamiltonian is solved.

Finite-momentum exciton dispersion of monolayer hBN and MoS₂ along M–Γ–K

Figure 6. Finite-momentum exciton dispersion of monolayer hBN (top) and MoS₂ (bottom) along MM–Γ\Gamma–KK. Near Γ\Gamma, the lowest hBN doublet separates into transverse (blue) and longitudinal (red) branches, while the bright MoS₂ AA-exciton doublet separates into parabolic (blue) and light-like (red) branches. Other excitons are shown in grey. Insets resolve the long-wavelength region; open symbols denote calculated energies and dashed lines the corresponding fits. Dotted vertical lines mark the closest non-zero transferred momenta to Γ\Gamma accessible on uniform 48×48×148 \times48 \times1 (hBN) and 72×72×172 \times72 \times1 (MoS₂) meshes, highlighting the finer resolution enabled by the interpolation.

In hBN, the lowest exciton doublet at Γ\Gamma separates into a weakly dispersive transverse branch (T) and a longitudinal branch (L) with a nonanalytic dispersion induced by long-range electron–hole exchange [9, 10, 11]. In the long-wavelength limit,

ET(Q)=E0+ℏ2Q22MT,(30)E_{\mathrm{T}}(Q) = E_0 + \frac{\hbar^2 Q^2}{2M_{\mathrm{T}}}, \tag*{(30)}
EL(Q)−ET(Q)=2A∣Q∣,(31)E_{\mathrm{L}}(Q) - E_{\mathrm{T}}(Q) = 2A\lvert Q\rvert, \tag*{(31)}

where E0E_0 is the common energy at Γ\Gamma and MTM_{\mathrm{T}} is the transverse exciton mass. Taking advantage of the arbitrary Q\mathbf{Q} sampling, we fit only the first few non-zero momenta closest to Γ\Gamma, where this limit applies. We obtain A=4.26A = 4.26 and 4.144.14 eV Å along Γ\Gamma–MM and Γ\Gamma–KK, respectively, indicating a nearly isotropic linear splitting. The corresponding exchange velocities v=2A/ℏv = 2A/\hbar are 1.29×1061.29 \times10^6 and 1.26×1061.26 \times10^6 m s−1^{-1}, while the fitted transverse masses are MT=3.46 meM_{\mathrm{T}} = 3.46\,m_e and 2.20 me2.20\,m_e. The pronounced linear splitting is consistent with the strong long-range electron–hole exchange reported for monolayer hBN [12, 13].

In MoS2_2, spin–orbit coupling separates the low-energy excitons into unlike-spin dark states and a like-spin bright AA-exciton doublet. Exchange vanishes in the unlike-spin channel, while at finite QQ it splits the bright doublet into a parabolic (−-) and a light-like (++) branch [10]. Their long-wavelength dispersion

E−(Q)=E0+ℏ2Q22M−,(32)E_{-}(Q) = E_0 + \frac{\hbar^2 Q^2}{2M_{-}}, \tag*{(32)}
E+(Q)−E−(Q)=2A∣Q∣.(33)E_{+}(Q) - E_{-}(Q) = 2A\lvert Q\rvert. \tag*{(33)}

Using the two closest non-zero momenta, we obtain A=0.57A = 0.57 and 0.510.51 eV Å along Γ\Gamma–MM and Γ\Gamma–KK, corresponding to velocities of 1.75×1051.75 \times10^{5} and 1.55×1051.55 \times10^{5} m s−1^{-1}. The fitted linear coefficients are therefore comparable along the two directions. The corresponding parabolic masses are M−=1.32 meM_{-} = 1.32\,m_e and 1.26 me1.26\,m_e. Quantitative agreement with previous literature is not expected from the present scissor-corrected LDA electronic structure, but the weaker light-like splitting relative to hBN is consistent with the expected exchange-driven behaviour.

These fits probe transferred momenta well below the resolution of the corresponding uniform meshes. The dotted lines in the insets of Figure 6 mark the closest non-zero momenta to Γ\Gamma accessible on those meshes. For hBN, their moduli along Γ\Gamma–MM and Γ\Gamma–KK are ∣Q∣=0.0606|\mathbf{Q}| = 0.0606 and 0.10490.1049 Å−1^{-1} on a 48×48×148 \times48 \times1 mesh, whereas the fits extend only to ∣Q∣=0.0087|\mathbf{Q}| = 0.0087 and 0.01510.0151 Å−1^{-1}. For MoS₂, the corresponding values on a 72×72×172 \times72 \times1 mesh are 0.03190.0319 and 0.05520.0552 Å−1^{-1}, compared with 0.00460.0046 and 0.00800.0080 Å−1^{-1} in the fits. In both materials, the sampled momenta are therefore about seven times smaller in modulus. This resolution is essential because the linear form already begins to deviate from the calculated dispersion very close to Γ\Gamma. Reaching the same regime with uniform grids would require substantially denser first-principles calculations of both Bloch states and screening. Recent work has accessed it through an approximate long-wavelength treatment [74]; here, it is obtained directly by constructing and solving the BSE Hamiltonian at each Q\mathbf{Q}, without introducing an additional long-wavelength approximation.

This illustrates the main advantage of decoupling the BSE sampling from the source grids: transferred momenta can be placed directly where the exciton dispersion varies most rapidly, giving access to the long-wavelength regime without uniformly refining the Brillouin zone.

Convergence of exciton binding energies with locally refined sampling

The freedom to choose the transition-momentum sampling independently of the underlying first-principles grids allows BSE points to be concentrated where they are most relevant. This is particularly useful for weakly bound Wannier–Mott excitons [75], whose amplitudes are localised in reciprocal space around the band extrema. Uniform meshes can therefore converge slowly, whereas local refinement provides the required momentum resolution with far fewer points [5, 6, 51].

We illustrate this for bulk GaAs, where the lowest exciton wave functions at Q=0\mathbf{Q} = 0 are concentrated around Γ\Gamma. Alongside uniform meshes, we consider locally refined samplings obtained by retaining points within a sphere of radius 0.040.04 Å−1^{-1} around Γ\Gamma from progressively denser uniform meshes. The parent uniform mesh determines the local momentum resolution, while only the retained points enter the BSE transition space. In both cases, the Bloch states and screened interaction are reconstructed from the same coarse first-principles inputs. The binding energy of exciton λ\lambda is defined as

Ebλ=Eg−Eλ(Q=0),(34)E_b^\lambda= E_g - E_\lambda(\mathbf{Q} = 0), \tag*{(34)}

where EgE_g is the single-particle gap entering the BSE. Figure 7 shows the convergence of the binding energy of the threefold-degenerate lowest exciton. Uniform samsampling remains far from convergence even on a 48×48×4848 \times48 \times48 mesh, containing more than 10510^{5} transition momenta. In contrast, local refinement yields a converged binding energy of approximately 4.0 meV with only a few thousand points. This agrees closely with experimental values of 3.8–4.2 meV [76, 77, 78] and the Wannier–Mott model prediction of 4.1 meV [79].

Figure 7. Convergence of the binding energy of the threefold-degenerate lowest exciton in bulk GaAs with the number of BSE kk points. Red squares denote uniform sampling and blue circles locally refined sampling around Γ\Gamma. Labels on the uniform series indicate the mesh dimensions, while those in parentheses indicate the uniform meshes required to achieve comparable momentum resolution near Γ\Gamma.

Earlier first-principles GW–BSE calculations by Rohlfing and Louie obtained a binding energy of 4.0 meV using locally concentrated momentum sampling [51]. Our result agrees closely with this benchmark, while the present approach constructs the BSE kernels from independently interpolated Bloch states and reconstructed dielectric screening.

These results demonstrate that accurate exciton binding energies can be obtained with substantially fewer transition momenta by refining the BSE sampling locally, without increasing the underlying first-principles grids.

Conclusions

We have developed a reciprocal-space framework for reconstructing the static Bethe–Salpeter Hamiltonian at arbitrary crystal momenta from coarse first-principles grids. Bloch states are interpolated within an optimally smooth Wannier subspace, while the screened Coulomb interaction is reconstructed from an independently interpolated dielectric response. The resulting BSE Hamiltonian can be evaluated on momentum grids chosen independently of the underlying first-principles sampling.

Although we use static screening within the random-phase approximation here, the interpolation requires only the microscopic inverse dielectric matrix on a coarse qq mesh, independently of the approximation used to calculate it. The dielectric response can therefore be obtained from different many-body implementations, such as Yambo [72, 73], BerkeleyGW [80], or ABINIT [81]. Likewise, the single-particle energies can be obtained from various electronic-structure approaches, including semilocal and hybrid functionals, Koopmans-compliant functionals [82, 83], DFT+UU, and GWGW quasiparticle calculations. The use of a Wannier representation is not a major practical limitation, as automated procedures now allow Wannier Hamiltonians to be constructed with limited material-specific tuning, as demonstrated in high-throughput studies across broad classes of materials [84, 85]. These features make the method applicable to a broad range of materials, levels of theory, and electronic-structure codes.

Benchmarks against explicit first-principles calculations show excellent agreement for the interpolated Bloch-state plane-wave components and reconstructed dielectric screening. The resulting BSE optical spectra also closely reproduce the reference calculations for monolayer hBN, MoS2_2, and bulk GaAs. Our cost analysis estimates speedups of up to 22×22\times for MoS2_2 and 10×10\times for GaAs on the grids considered here. These estimates are conservative: the internal momentum grid used to evaluate the independent-particle response is kept fixed, whereas conventional BSE implementations often refine it together with the BSE sampling, further increasing the cost of dielectric screening.

The freedom to choose both the transition-momentum grid and the transferred momentum QQ also enables calculations tailored to different excitonic properties. For monolayer hBN and MoS2_2, we resolve the long-wavelength exciton dispersions at transferred momenta about seven times smaller than those accessible on the corresponding uniform meshes, recovering the nonanalytic splitting of low-energy excitons induced by long-range electron–hole exchange. Each point is obtained by reconstructing and solving a new BSE Hamiltonian. For bulk GaAs, locally refined transition-momentum sampling converges the binding energy of the threefold-degenerate lowest exciton to approximately 4.0 meV with only a few thousand BSE points, substantially outperforming uniform meshes. This agrees closely with experiment and previous first-principles calculations.

The ability to reconstruct excitonic states at arbitrary momenta within a common smooth electronic representation also provides a basis for combining them consistently with other interaction vertices. This opens the way to first-principles treatments of exciton–phonon coupling [38, 39], excitonic corrections to nonlocal electron–phonon vertices [40], and nonequilibrium exciton dynamics [36].

Acknowledgments

Funded by the European Union (ERC, DELIGHT, 101052708). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union nor the granting authority can be held responsible for them.

We are grateful for fruitful discussions with Dr. Iurii Timrov and Dr. Nicola Colonna.

Data Availability Statement

The data underlying the results presented in this work will be made publicly available upon publication.

Static dielectric screening

The screened interaction entering the direct kernel of Eq. (8) requires the static microscopic dielectric response. The independent-particle density-density response function can be written in reciprocal space as

χGG′0(q,ω)=1NkΩ∑knmfkn−fk+qmω+εkn−εk+qm+iηρmnG′(k,q)[ρmnG(k,q)]∗(35)\chi^{0}_{GG'}(\mathbf{q},\omega)=\frac{1}{N_k\Omega}\sum_{\mathbf{k}nm}\frac{f_{\mathbf{k}n}-f_{\mathbf{k}+\mathbf{q}m}}{\omega+\varepsilon_{\mathbf{k}n}-\varepsilon_{\mathbf{k}+\mathbf{q}m}+i\eta}\rho^{G'}_{mn}(\mathbf{k},\mathbf{q})\left[\rho^{G}_{mn}(\mathbf{k},\mathbf{q})\right]^{*} \tag*{(35)}

where fknf_{\mathbf{k}n} denotes the occupation of the state (k,n)(\mathbf{k},n) and the transition densities are defined in Eq. (7). (35) involves a sum over the internal k\mathbf{k}-point sampling, distinct from the q\mathbf{q} mesh on which the dielectric response is evaluated.

We denote matrices in reciprocal-lattice-vector space by a hat, such that for a generic quantity AGG′A_{GG'},

[A^]GG′=AGG′.(36)\left[\hat{A}\right]_{GG'}=A_{GG'}. \tag*{(36)}

In particular, the bare Coulomb interaction is represented by the diagonal matrix

[v^(q)]GG′=vG(q)δGG′.(37)\left[\hat{v}(\mathbf{q})\right]_{GG'}=v_G(\mathbf{q})\delta_{GG'}. \tag*{(37)}

Unless otherwise stated, products and inverses of hatted quantities in the following equations are understood as matrix products and matrix inverses in reciprocal-lattice-vector space. Starting from χ^0\hat{\chi}^{0}, the interacting density response satisfies the Dyson equation [4, 86, 87]

χ^(q,ω)=χ^0(q,ω)+χ^0(q,ω)[v^(q)+f^xc(q,ω)]χ^(q,ω),(38)\hat{\chi}(\mathbf{q},\omega)=\hat{\chi}^{0}(\mathbf{q},\omega)+\hat{\chi}^{0}(\mathbf{q},\omega)\left[\hat{v}(\mathbf{q})+\hat{f}_{\mathrm{xc}}(\mathbf{q},\omega)\right]\hat{\chi}(\mathbf{q},\omega), \tag*{(38)}

where f^xc\hat{f}_{\mathrm{xc}} is the exchange–correlation kernel. Within the random-phase approximation, adopted throughout this work, f^xc=0\hat{f}_{\mathrm{xc}}=0, yielding

χ^RPA(q,ω)=[1^−χ^0(q,ω)v^(q)]−1χ^0(q,ω).(39)\hat{\chi}^{\mathrm{RPA}}(\mathbf{q},\omega)=\left[\hat{1}-\hat{\chi}^{0}(\mathbf{q},\omega)\hat{v}(\mathbf{q})\right]^{-1}\hat{\chi}^{0}(\mathbf{q},\omega). \tag*{(39)}

The symmetrised inverse microscopic dielectric matrix is then

ϵ^−1(q,ω)=1^+v^1/2(q)χ^(q,ω)v^1/2(q),(40)\hat{\epsilon}^{-1}(\mathbf{q},\omega)=\hat{1}+\hat{v}^{1/2}(\mathbf{q})\hat{\chi}(\mathbf{q},\omega)\hat{v}^{1/2}(\mathbf{q}), \tag*{(40)}

where v^1/2\hat{v}^{1/2} is the diagonal matrix with elements vG(q)\sqrt{v_G(\mathbf{q})}. The screened Coulomb interaction follows as

W^(q,ω)=v^1/2(q)ϵ^−1(q,ω)v^1/2(q),(41)\hat{W}(\mathbf{q},\omega)=\hat{v}^{1/2}(\mathbf{q})\hat{\epsilon}^{-1}(\mathbf{q},\omega)\hat{v}^{1/2}(\mathbf{q}), \tag*{(41)}

or, equivalently,

W^(q,ω)=v^(q)+v^(q)χ^(q,ω)v^(q).(42)\hat{W}(\mathbf{q},\omega)=\hat{v}(\mathbf{q})+\hat{v}(\mathbf{q})\hat{\chi}(\mathbf{q},\omega)\hat{v}(\mathbf{q}). \tag*{(42)}

The matrix elements of W^\hat{W} are the quantities WGG′W_{GG'} entering Eq. (8). As customary in the static BSE approximation [3, 4], the screening is evaluated at zero frequency,

W^(q)≡W^(q,ω=0).(43)\hat{W}(\mathbf{q})\equiv\hat{W}(\mathbf{q},\omega=0). \tag*{(43)}

Bare Coulomb interaction

The bare Coulomb interaction is evaluated analytically at the required momentum. For three-dimensional periodic systems, its reciprocal-space form is

vG3D(q)=8π∣q+G∣2.(44)v^{3\mathrm{D}}_{G}(\mathbf{q})=\frac{8\pi}{\lvert\mathbf{q}+\mathbf{G}\rvert^{2}}. \tag*{(44)}

For two-dimensional systems in a repeated-slab geometry, we instead employ a Coulomb interaction truncated along the non-periodic direction [88, 89]. Defining

K=q+G(45)\mathbf{K}=\mathbf{q}+\mathbf{G} \tag*{(45)}

and its components perpendicular and parallel to the layer as

K⊥=K⋅n^,K∥=∣K∣2−K⊥2,(46)K_{\perp}=\mathbf{K}\cdot\hat{\mathbf{n}},\qquad K_{\parallel}=\sqrt{\lvert\mathbf{K}\rvert^{2}-K_{\perp}^{2}}, \tag*{(46)}

where n^\hat{\mathbf{n}} is the slab normal, the truncated interaction is

vGslab(q)=8π∣K∣2[1−e−K∥Lcos⁡(K⊥L)].(47)v^{\mathrm{slab}}_{G}(\mathbf{q})=\frac{8\pi}{\lvert\mathbf{K}\rvert^{2}}\left[1-e^{-K_{\parallel}L}\cos(K_{\perp}L)\right]. \tag*{(47)}

Here, LL is the truncation length, equal to half the supercell extent along the non-periodic direction for the symmetric slab cutoff used here.

Reciprocal-covariant RBF–TPS interpolation

Following the direct Wannier interpolation approach in the optimally smooth subspace (OSS) of Ref. [19], the reciprocal-space interpolation scheme can be chosen independently of the Wannier construction. Here, we interpolate the smooth OSS wave-function components uknOSS(G,σ)u^{\mathrm{OSS}}_{\mathbf{k}n}(\mathbf{G},\sigma) and screening response fGG′(q)f_{GG'}(\mathbf{q}) using local radial basis functions (RBFs) with the thin-plate-spline (TPS) kernel of Refs. [90, 91]. This avoids an intermediate transformation to real space and does not require a rectangular tensor-product grid, making the method well suited to non-orthogonal reciprocal lattices and arbitrary sampling within the first Brillouin zone.

The reciprocal-space quantities considered here are not scalar periodic functions of crystal momentum. A reciprocal-lattice translation G′′\mathbf{G}'' of the momentum is accompanied by a corresponding shift of the Fourier indices,

uk+G′′nOSS(G,σ)=uknOSS(G+G′′,σ),(48)u^{\mathrm{OSS}}_{\mathbf{k}+\mathbf{G}''n}(\mathbf{G},\sigma)=u^{\mathrm{OSS}}_{\mathbf{k}n}(\mathbf{G}+\mathbf{G}'',\sigma), \tag*{(48)}
fGG′(q+G′′)=fG+G′′,G′+G′′(q).(49)f_{\mathbf{G}\mathbf{G}'}(\mathbf{q}+\mathbf{G}'')=f_{\mathbf{G}+\mathbf{G}'',\mathbf{G}'+\mathbf{G}''}(\mathbf{q}). \tag*{(49)}

These covariance relations are used to augment the source sampling with neighbouring reciprocal-lattice images. We consider translations

G′′=∑μ∈Amμbμ,mμ∈{−1,0,1},(50)\mathbf{G}''=\sum_{\mu\in\mathcal{A}}m_\mu\mathbf{b}_\mu,\qquad m_\mu\in\{-1,0,1\}, \tag*{(50)}

where bμ\mathbf{b}_\mu are reciprocal-lattice vectors and A\mathcal{A} contains the sampled reciprocal directions.

Only images compatible with the retained plane-wave basis are included. For the wave functions, the shifted component G+G′′\mathbf{G}+\mathbf{G}'' must be available, while for the screening response both G+G′′\mathbf{G}+\mathbf{G}'' and G′+G′′\mathbf{G}'+\mathbf{G}'' must be retained. The interpolation neighbourhood may therefore depend on the Fourier component being considered.

For a target momentum x\mathbf{x} in the first Brillouin zone, representing either a Bloch momentum k\mathbf{k} or a screening momentum q\mathbf{q}, we first fix the Fourier component to be interpolated: G\mathbf{G} for the wave functions or (G,G′)(\mathbf{G},\mathbf{G}') for the screening response. The source momenta xi\mathbf{x}_i are supplemented by their admissible images xi+G′′\mathbf{x}_i+\mathbf{G}'', from which a local set is selected according to the Cartesian reciprocal-space distance from the target. Labelling each source image by a=(i,G′′)a=(i,\mathbf{G}''), we define

Δxa=xi+G′′−x(51)\Delta\mathbf{x}_a=\mathbf{x}_i+\mathbf{G}''-\mathbf{x} \tag*{(51)}

and the scaled local coordinates

ya=Δxas,s=max⁡μ∣Δxaμ∣.(52)\mathbf{y}_a=\frac{\Delta\mathbf{x}_a}{s},\qquad s=\max_\mu\lvert\Delta x_{a\mu}\rvert. \tag*{(52)}

We use the RBF kernel

ϕ(r)={r2ln⁡r,r>0,0,r=0,(53)\phi(r)= \begin{cases} r^2\ln r, & r>0,\\ 0, & r=0, \end{cases} \tag*{(53)}

together with a quadratic polynomial augmentation,

p(y)=(1,y1,…,yd,{yμyν}1≤μ≤ν≤d)T,(54)\mathbf{p}(\mathbf{y})=\left(1,y_1,\ldots,y_d,\{y_\mu y_\nu\}_{1\leq\mu\leq\nu\leq d}\right)^{\mathrm{T}}, \tag*{(54)}

where d=2d=2 or 33 is the dimensionality of the reciprocal-space sampling. For a local set of NlocN_{\mathrm{loc}} source images, the RBF matrix is

Φab=ϕ(∥ya−yb∥).(55)\Phi_{ab}=\phi(\lVert\mathbf{y}_a-\mathbf{y}_b\rVert). \tag*{(55)}

Figure 8. Comparison between the DFT band structures (blue solid lines) and their Wannier-interpolated counterparts (red dashed lines) for (a) monolayer hBN, (b) monolayer MoS2_2, and (c) bulk GaAs. The bands shown include the valence and conduction states relevant to the BSE calculations. The Wannier interpolation closely reproduces the DFT dispersions along the Γ\Gamma–MM–KK–Γ\Gamma path for hBN and MoS2_2 and the LL–Γ\Gamma–XX–LL path for GaAs. In each case, the centre of the band gap is set to zero energy.

The interpolation weights h\mathbf{h} and the auxiliary coefficients λ\boldsymbol{\lambda} associated with the polynomial constraints are obtained from

\begin{pmatrix} \Phi& P\\ P^{\mathrm{T}} & 0 \end{pmatrix} \begin{pmatrix} \mathbf{h}\\ \boldsymbol{\lambda} \end{pmatrix} = \begin{pmatrix} \boldsymbol{\phi}_*\\ \mathbf{p}(\mathbf{0}) \end{pmatrix}, \tag*{(56)}\end{pmatrix}

where

Paℓ=pℓ(ya),(ϕ∗)a=ϕ(∥ya∥).(57)P_{a\ell}=p_{\ell}(y_a), \qquad(\phi^*)_a=\phi(\lVert y_a\rVert). \tag*{(57)}

The coefficient hah_a is the interpolation weight associated with source image aa, whereas λ\boldsymbol{\lambda} enforces the polynomial-reproduction constraints and does not enter directly in the interpolation of the physical quantity. The quadratic augmentation ensures exact reproduction of polynomials up to second order.

Once the geometric weights are known, they are applied to the reciprocal-space quantities using the covariance relations of Eqs. (48) and (49). For a fixed wave-function component,

uxnOSS(G,σ)=∑(i,G′′)∈SG(x)hiG′′(x)ukiOSS(G+G′′,σ),(58)u^{\mathrm{OSS}}_{x_n}(\mathbf{G},\sigma)=\sum_{(i,\mathbf{G}'')\in\mathcal{S}_{\mathbf{G}}(\mathbf{x})}h_{i\mathbf{G}''}(\mathbf{x})u^{\mathrm{OSS}}_{k_i}(\mathbf{G}+\mathbf{G}'',\sigma), \tag*{(58)}

while for a fixed screening component,

fGG′(x)=∑(i,G′′)∈SGG′(x)hiG′′(x)fG+G′′,G′+G′′(qi).(59)f_{\mathbf{G}\mathbf{G}'}(\mathbf{x})=\sum_{(i,\mathbf{G}'')\in\mathcal{S}_{\mathbf{G}\mathbf{G}'}(\mathbf{x})}h_{i\mathbf{G}''}(\mathbf{x})f_{\mathbf{G}+\mathbf{G}'',\mathbf{G}'+\mathbf{G}''}(\mathbf{q}_i). \tag*{(59)}

Here, ki\mathbf{k}_i and qi\mathbf{q}_i denote the source momenta for the wave-function and screening interpolations, respectively, while G′′\mathbf{G}'' identifies the reciprocal-lattice translation of each source image. The sets SG\mathcal{S}_{\mathbf{G}} and SGG′\mathcal{S}_{\mathbf{G}\mathbf{G}'} contain only images for which the required shifted Fourier components are available. Since the weights depend only on the geometry of the corresponding local set, the same interpolation is applied independently to the real and imaginary parts of complex quantities.

Wannier-interpolated band structures

Figure 8 compares the DFT and Wannier-interpolated band structures for the three materials considered in this work, including the valence and conduction manifolds relevant to the BSE calculations. For hBN, the initial projections consist of B and N sp2sp^2 and pzp_z states together with two interstitial ss-like functions. For MoS2_2, Mo dd and S pp states are used, while for GaAs the projections include Ga and As ss and pp states together with dxyd_{xy}, dxzd_{xz}, and dyzd_{yz} states on both atomic species.

In all cases, the Wannier-interpolated bands closely reproduce the DFT dispersions along the selected high-symmetry paths, including the band-edge regions relevant to the electron–hole transition space. This confirms the accuracy of the Wannier representation used to reconstruct the electronic states entering the BSE at arbitrary kk points.

References

  1. [1]G. Strinati, Application of the green’s functions method to the study of the optical properties of semiconductors, La Rivista del Nuovo Cimento (1978-1999) 11, 1 (1988).
  2. [2]S. Albrecht, L. Reining, R. Del Sole, and G. Onida, Ab initio calculation of excitonic effects in the optical spectra of semiconductors, Phys. Rev. Lett. 80, 4510 (1998).
  3. [3]M. Rohlfling and S. G. Louie, Electron-hole excitations and optical spectra from first principles, Phys. Rev. B 62, 4927 (2000).
  4. [4]G. Onida, L. Reining, and A. Rubio, Electronic excitations: density-functional versus many-body green’s-function approaches, Rev. Mod. Phys. 74, 601 (2002).
  5. [5]F. Fuchs, C. Rödl, A. Schleife, and F. Bechstedt, Efficient O(N²) approach to solve the bethe-salpeter equation for excitonic bound states, Phys. Rev. B 78, 085103 (2008).
  6. [6]A. M. Alvertis, A. Champagne, M. Del Ben, F. H. da Jornada, D. Y. Qiu, M. R. Filip, and J. B. Neaton, Importance of nonuniform brillouin zone sampling for ab initio bethe-salpeter equation calculations of exciton binding energies in crystalline solids, Phys. Rev. B 108, 235117 (2023).
  7. [7]F. Hüser, T. Olsen, and K. S. Thygesen, How dielectric screening in two-dimensional crystals affects the convergence of excited-state calculations: Monolayer MoS₂, Phys. Rev. B 88, 245309 (2013).
  8. [8]D. Y. Qiu, F. H. da Jornada, and S. G. Louie, Screening and many-body effects in two-dimensional crystals: Monolayer MoS₂, Phys. Rev. B 93, 235435 (2016).
  9. [9]L. C. Andreani and F. Bassani, Exchange interaction and polariton effects in quantum-well excitons, Phys. Rev. B 41, 7536 (1990).
  10. [10]D. Y. Qiu, T. Cao, and S. G. Louie, Nonanalyticity, valley quantum phases, and lightlike exciton dispersion in monolayer transition metal dichalcogenides: Theory and first-principles calculations, Phys. Rev. Lett. 115, 176801 (2015).
  11. [11]P. Cudazzo, L. Sponza, C. Giorgetti, L. Reining, F. Sottile, and M. Gatti, Exciton band structure in two-dimensional materials, Phys. Rev. Lett. 116, 066803 (2016).
  12. [12]L. Y. Liu, S. Y. Woo, J. Wu, B. Hou, C. Su, and D. Y. Qiu, Direct observation of massless excitons and linear exciton dispersion, Nature Physics 22, 521 (2026).
  13. [13]J. Hong, A. Guandalini, W. Wu, H. Sun, F. W. Chen, C. Ma, K. Suenaga, T. Pichler, and F. Mauri, Exciton dispersion fine structure and deep ultraviolet optical conductivity of freestanding two-dimensional h-BN, Nature Communications 17, 7399 (2026).
  14. [14]N. Marzari and D. Vanderbilt, Maximally localized generalized wannier functions for composite energy bands, Phys. Rev. B 56, 12847 (1997).
  15. [15]I. Souza, N. Marzari, and D. Vanderbilt, Maximally localized wannier functions for entangled energy bands, Phys. Rev. B 65, 035109 (2001).
  16. [16]N. Marzari, A. A. Mostofi, J. R. Yates, I. Souza, and D. Vanderbilt, Maximally localized wannier functions: Theory and applications, Rev. Mod. Phys. 84, 1419 (2012).
  17. [17]F. Giustino, M. L. Cohen, and S. G. Louie, Electron-phonon interaction using wannier functions, Phys. Rev. B 76, 165108 (2007).
  18. [18]M. Calandra, G. Profeta, and F. Mauri, Adiabatic and nonadiabatic phonon dispersion in a wannier function approach, Phys. Rev. B 82, 165111 (2010).
  19. [19]G. Volpato, S. Mocatti, G. Marini, and M. Calandra, Wannier interpolation of reciprocal-space periodic and nonperiodic matrix elements in the optimally smooth subspace, Phys. Rev. B 111, 035105 (2025).DOI
  20. [20]H. Lee, S. Poncé, K. Bushick, S. Hajinazar, J. Lafuente-Bartolome, J. Leveillee, C. Lian, J.-M. Lihm, F. Macheda, H. Mori, H. Paudyal, W. H. Sio, S. Tiwari, M. Zacharias, X. Zhang, N. Bonini, E. Kioupakis, E. R. Margine, and F. Giustino, Electron–phonon physics from first principles using the epw code, npj Computational Materials 9, 156 (2023).arxiv.org/abs/2302.08085
  21. [21]J.-J. Zhou, J. Park, I.-T. Lu, I. Maliiyov, X. Tong, and M. Bernardi, Perturbo: A software package for ab initio electron–phonon interactions, charge transport and ultrafast dynamics, Computer Physics Communications 264, 107970 (2021).DOI
  22. [22]A. Cepellotti, J. Coulter, A. Johansson, N. S. Fedorova, and B. Kozinsky, Phoebe: A high-performance framework for solving phonon and electron boltzmann transport equations, Journal of Physics: Materials 5, 035003 (2022).DOI
  23. [23]G. Marini, G. Marchese, G. Profeta, J. Sjåkste, F. Macheda, N. Vast, F. Mauri, and M. Calandra, epiq: An open-source software for the calculation of electron-phonon interaction related properties, Computer Physics Communications 295, 108950 (2024).DOI
  24. [24]F. Macheda and T. Sohier, Coupled dynamical boltzmann transport equations with long-range electron-phonon and electron-electron interactions in two-dimensional materials, Phys. Rev. B 114, 245419 (2026).
  25. [25]A. Marini, Competition between the electronic and phonon-mediated scattering channels in the out-of-equilibrium carrier dynamics of semiconductors: an ab-initio approach, Journal of Physics: Conference Series 427, 012003 (2013).DOI
  26. [26]D. Sangalli and A. Marini, Ultra-fast carriers relaxation in bulk silicon following photo-excitation with a short and polarized laser pulse, Europhysics Letters 110, 47004 (2015).arxiv.org/abs/1409.1706
  27. [27]E. Perfetto and G. Stefanucci, Real-time gw-ehrenfest-fan-migdal method for nonequilibrium 2d materials, Nano Letters 23, 7029 (2023).arxiv.org/abs/2305.07458
  28. [28]G. Stefanucci and E. Perfetto, Semiconductor electron-phonon equations: A rung above boltzmann in the many-body ladder, SciPost Phys. 16, 073 (2024).arxiv.org/abs/2311.03980
  29. [29]S. Mocatti, G. Marini, G. Volpato, P. Cudazzo, and M. Calandra, Nonequilibrium photocarrier and phonon dynamics from first principles: a unified treatment of carrier-carrier, carrier-phonon, and phonon-phonon scattering, npj Computational Materials 12, 260 (2026).DOI
  30. [30]H.-Y. Chen, D. Sangalli, and M. Bernardi, Exciton-phonon interaction and relaxation times from first principles, Phys. Rev. Lett. 125, 107401 (2020).DOI
  31. [31]G. Antonius and S. G. Louie, Theory of exciton-phonon coupling, Phys. Rev. B 105, 085111 (2022).DOI
  32. [32]F. Paleari, H. P. C. Miranda, A. Molina-Sánchez, and L. Wirtz, Exciton-phonon coupling in the ultraviolet absorption and emission spectra of bulk hexagonal boron nitride, Phys. Rev. Lett. 122, 187401 (2019).DOI
  33. [33]P. Cudazzo, First-principles description of the exciton-phonon interaction: A cumulant approach, Phys. Rev. B 102, 045136 (2020).DOI
  34. [34]Y.-h. Chan, J. B. Haber, M. H. Naik, J. B. Neaton, D. Y. Qiu, F. H. da Jornada, and S. G. Louie, Exciton lifetime and optical line width profile via exciton–phonon interactions: Theory and first-principles calculations for monolayer mos2, Nano Letters 23, 3971 (2023).DOI
  35. [35]G. Stefanucci and E. Perfetto, Excitonic bloch equations from first principles, SciPost Phys. 18, 009 (2025).DOI
  36. [36]C. Betancur, G. Stefanucci, and E. Perfetto, Excitonic effects in the photocarrier dynamics of two-dimensional materials, Nano Letters 26, 9455 (2026).DOI
  37. [37]M. Nalbatohuda, D. Sangalli, F. Paleari, S. Reichardt, and L. Wirtz, Symmetries of excitons, Phys. Rev. B 113, 205130 (2026).
  38. [38]A. M. Alvertis, J. B. Haber, Z. Li, C. J. N. Coveney, S. G. Louie, M. R. Filip, and J. B. Neaton, Phonon screening and dissociation of excitons at finite temperatures from first principles, Proceedings of the National Academy of Sciences 121, e2403434121 (2024).DOI
  39. [39]G. Marini, M. Calandra, and P. Cudazzo, Optical absorption and photoluminescence of single-layer boron nitride from a first-principles cumulant approach, Nano Letters 24, 6017 (2024).DOI
  40. [40]A. Guandalini, F. Macheda, G. Caldarelli, and F. Mauri, Excitonic effects in phonons: Reshaping the graphene kohn anomalies and lifetimes, Phys. Rev. Lett. 135, 076401 (2025).
  41. [41]P. Umari, G. Stenuit, and S. Baroni, Optimal representation of the polarization propagator for large-scale gw calculations, Phys. Rev. B 79, 201104(R) (2009).DOI
  42. [42]P. Umari, G. Stenuit, and S. Baroni, GW quasiparticle spectra from occupied states only, Phys. Rev. B 81, 115104 (2010).DOI
  43. [43]M. Marsili, E. Mosconi, F. De Angelis, and P. Umari, Large-scale gw-bse calculations with N3 scaling: Excitonic effects in dye-sensitized solar cells, Phys. Rev. B 95, 075415 (2017).DOI
  44. [44]G. Cistaro, M. Sá, D. Sangalli, A. Picón, and N. Colonna, Efficient nonequilibrium electron dynamics from first-principles: Leveraging koopmans spectral functionals and wannier localization, Journal of Chemical Theory and Computation (2026).arxiv.org/abs/2608.16226
  45. [45]K. Merkel and F. Ortmann, Linear scaling approach for optical excitations using maximally localized wannier functions, Journal of Physics: Materials 7, 015001 (2023).arxiv.org/abs/2309.06834
  46. [46]D. Rocca, R. Gebauer, Y. Saad, and S. Baroni, Turbo charging time-dependent density-functional theory with lanczos chains, The Journal of Chemical Physics 128, 154105 (2008).DOI
  47. [47]S. Baroni, R. Gebauer, O. Baris Malcioglu, Y. Saad, P. Umari, and J. Xian, Harnessing molecular excited states with lanczos chains, Journal of Physics: Condensed Matter 22, 074204 (2010).DOI
  48. [48]D. Rocca, D. Lu, and G. Galli, Ab initio calculations of optical absorption spectra: Solution of the bethe-salpeter equation within density matrix perturbation theory, The Journal of Chemical Physics 133, 164109 (2010).
  49. [49]D. Rocca, Y. Ping, R. Gebauer, and G. Galli, Solution of the bethe-salpeter equation without empty electronic states: Application to the absorption spectra of bulk systems, Phys. Rev. B 85, 045116 (2012).DOI
  50. [50]N. L. Nguyen, H. Ma, M. Govoni, F. Gygi, and G. Galli, Finite-field approach to solving the bethe-salpeter equation, Phys. Rev. Lett. 122, 237402 (2019).DOI
  51. [51]M. Rohlfing and S. G. Louie, Electron-hole excitations in semiconductors and insulators, Phys. Rev. Lett. 81, 2312 (1998).
  52. [52]Y. Gillet, M. Giantomassi, and X. Gonze, Efficient on-the-fly interpolation technique for bethe-salpeter calculations of optical spectra, Computer Physics Communications 203, 83 (2016).
  53. [53]D. Kammerlander, S. Botti, M. A. L. Marques, A. Marini, and C. Attaccalite, Speeding up the solution of the bethe-salpeter equation by a double-grid method and wannier interpolation, Phys. Rev. B 86, 125203 (2012).DOI
  54. [54]I. M. Alliati, D. Sangalli, and M. Grüning, Double k-grid method for solving the bethe-salpeter equation via lanczos approaches, Frontiers in Chemistry Volume 9 - 2021 (2022).
  55. [55]A. Guandalini, P. D’Amico, A. Ferretti, and D. Varsano, Efficient gw calculations in two dimensional materials through a stochastic integration of the screened potential, npj Computational Materials 9, 44 (2023).arxiv.org/abs/2205.11946
  56. [56]J. B. Haber, D. Y. Qiu, F. H. da Jornada, and J. B. Neaton, Maximally localized exciton and wannier functions for solids, Phys. Rev. B 108, 125118 (2023).arxiv.org/abs/2308.03012
  57. [57]T. Sander, E. Maggio, and G. Kresse, Beyond the tamm-dancoff approximation for extended systems using exact diagonalization, Phys. Rev. B 92, 045209 (2015).
  58. [58]W. Kohn, Analytic properties of bloch waves and wannier functions, Phys. Rev. 115, 809 (1959).DOI
  59. [59]J. D. Cloizeaux, Energy bands and projection operators in a crystal: Analytic and asymptotic properties, Phys. Rev. 135, A685 (1964).DOI
  60. [60]C. Brouder, G. Panati, M. Calandra, C. Mourougane, and N. Marzari, Exponential localization of wannier functions in insulators, Phys. Rev. Lett. 98, 046402 (2007).pubmed.ncbi.nlm.nih.gov/17358792
  61. [61]P. Giannozzi, S. Baroni, N. Bonini, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, G. L. Chiarotti, M. Cococcioni, I. Dabo, A. Dal Corso, S. de Gironcoli, S. F. Fabris, G. Fratesi, R. Gebauer, U. Gerstmann, C. Gougoussis, A. Kokalj, M. Lazzeri, L. Martin-Samos, N. Marzari, F. Mauri, R. Mazzarello, S. Paolini, A. Pasquarello, L. Paulatto, C. Sbraccia, S. Scandolo, G. Sclauzero, A. P. Seit­sonen, A. Smogunov, P. Umari, and R. M. Wentzcovitch, Quantum espresso: a modular and open-source software project for quantum simulations of materials, Journal of Physics: Condensed Matter 21, 395502 (2009).DOI
  62. [62]P. Giannozzi, O. Andreussi, T. Brumme, O. Bunau, M. Buongiorno Nardelli, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, M. Cococcioni, N. Colonna, I. Carnimeo, A. Dal Corso, S. de Gironcoli, P. Delugas, R. A. DiStasio, A. Ferretti, A. Floris, G. Fratesi, G. Fugallo, R. Gebauer, U. Gerstmann, F. Giustino, T. Gorni, G. Jia, M. Kawamura, H.-Y. Ko, A. Kokalj, E. Küçükbenli, M. Lazzeri, M. Marsili, N. Marzari, F. Mauri, N. L. Nguyen, H.-V. Nguyen, A. Otero-de-la-Roza, L. Paulatto, S. Poncé, D. Rocca, R. Sabatini, B. Santra, M. Schliph, A. P. Seit­sonen, A. Smogunov, I. Timrov, T. Thonhauser, P. Umari, N. Vast, X. Wu, and S. Baroni, Advanced capabilities for materials modelling with quantum espresso, Journal of Physics: Condensed Matter 29, 465901 (2017).
  63. [63]G. Pizzi, V. Vitale, R. Arita, S. Blügel, F. Freimuth, G. Géranton, M. Gibertini, D. Gresch, C. Johnson, T. Koretsune, J. Ibáñez-Azpiroz, H. Lee, J.-M. Lihm, D. Marchand, A. Marrazzo, Y. Mokrousov, J. I. Mustafa, Y. Nohara, Y. Nomura, L. Paulatto, S. Poncé, T. Ponweiser, J. Qiao, F. Thöle, S. S. Tsirkin, M. Wierzbowska, N. Marzari, D. Vanderbilt, I. Souza, A. A. Mostofi, and J. R. Yates, Wannier90 as a community code: new features and applications, Journal of Physics: Condensed Matter 32, 165902 (2020).
  64. [64]C. Lanczos, An iteration method for the solution of the eigenvalue problem of linear differential and integral operators, Journal of Research of the National Bureau of Standards 45, 255 (1950).DOI
  65. [65]R. Haydock, The recursive solution of the Schrödinger equation, in Solid State Physics, Vol. 35, edited by H. Ehrenreich, F. Seitz, and D. Turnbull (Academic Press, New York, 1980) pp. 215–298.
  66. [66]L. X. Benedict, E. L. Shirley, and R. B. Bohn, Optical absorption of insulators and the electron-hole interaction: An ab initio calculation, Phys. Rev. Lett. 80, 4514 (1998).
  67. [67]E. R. Davidson, The iterative calculation of a few of the lowest eigenvalues and corresponding eigenvectors of large real-symmetric matrices, Journal of Computational Physics 17, 87 (1975).DOI
  68. [68]W. Kohn and L. J. Sham, Self-consistent equations including exchange and correlation effects, Phys. Rev. 140, A1133 (1965).DOI
  69. [69]D. R. Hamann, Optimized norm-conserving vanderbilt pseudopotentials, Phys. Rev. B 88, 085117 (2013).DOI
  70. [70]T. Sohier, M. Calandra, and F. Mauri, Density functional perturbation theory for gated two-dimensional heterostructures: Theoretical developments and application to flexural phonons in graphene, Phys. Rev. B 96, 075448 (2017).DOI
  71. [71]H. J. Monkhorst and J. D. Pack, Special points for brillouin-zone integrations, Phys. Rev. B 13, 5188 (1976).DOI
  72. [72]A. Marini, C. Hogan, M. Grüning, and D. Varsano, yambo: An ab initio tool for excited state calculations, Computer Physics Communications 180, 1392 (2009).DOI
  73. [73]D. Sangalli, A. Ferretti, H. Miranda, C. Attaccalite, I. Marri, E. Cannuccia, P. Melo, M. Mar­ s ili, F. Paleari, A. Marrazzo, G. Prandini, P. Bonfà, M. O. Atambo, F. Affinito, M. Palummo, A. Molina-Sánchez, C. Hogan, M. Grüning, D. Varsano, and A. Marini, Many-body perturbation theory calculations using the yambo code, Journal of Physics: Condensed Matter 31, 325902 (2019).
  74. [74]A. Guandalini, A. Ferretti, D. Varsano, P. Barone, and F. Mauri, Unified description of exciton, phonon and plasmon dispersions in 2d materials from an optical conductivity approximation (2026), arXiv:2609.29789 [cond-mat.mtrl-sci].arxiv.org/abs/2609.29789
  75. [75]L. J. Sham and T. M. Rice, Many-particle derivation of the effective-mass equation for the wannier exciton, Phys. Rev. 144, 708 (1966).DOI
  76. [76]D. Bimberg and W. Schairer, Nonhydrogenic exciton and energy gap of gaas, Phys. Rev. Lett. 28, 442 (1972).
  77. [77]D. D. Sell, Resolved free-exciton transitions in the optical-absorption spectrum of gaas, Phys. Rev. B 6, 3750 (1972).DOI
  78. [78]S. B. Nam, D. C. Reynolds, C. W. Litton, R. J. Almassy, T. C. Collins, and C. M. Wolfe, Free-exciton energy spectrum in gaas, Phys. Rev. B 13, 761 (1976).DOI
  79. [79]A. D’Andrea and R. Del Sole, Wannier-mott excitons in semi-infinite crystals: Wave functions and normal-incidence reflectivity, Phys. Rev. B 25, 3714 (1982).DOI
  80. [80]J. Deslippe, G. Samsonidze, D. A. Strubbe, M. Jain, M. L. Cohen, and S. G. Louie, Berkeleygw: A massively parallel computer package for the calculation of the quasi- particle and optical properties of materials and nanostructures, Computer Physics Communications 183, 1269 (2012).
  81. [81]X. Gonze, B. Amadon, G. Antonius, F. Arnardi, L. Baguet, J.-M. Beuken, J. Bieder, F. Bottin, J. Bouchet, E. Bousquet, N. Brouwer, F. Bruneval, G. Brunin, T. Cavignac, J.-B. Charraud, W. Chen, M. Côté, S. Cottenier, J. Denier, G. Geneste, P. Ghosez, M. Giantomassi, Y. Gillet, O. Gingras, D. R. Hamann, G. Hautier, X. He, N. Helbig, N. Holzwarth, Y. Jia, F. Jollet, W. Lafargue-Dit-Hauret, K. Lejaeghere, M. A. Marques, A. Martin, C. Martins, H. P. Miranda, F. Naccarato, K. Persson, G. Petretto, V. Planes, Y. Pouillon, S. Prokhorenko, F. Ricci, G.-M. Rignanese, A. H. Romero, M. M. Schmitt, M. Torrent, M. J. van Setten, B. Van Troeye, M. J. Verstraete, G. Zérah, and J. W. Zwanziger, The abinitoproject: Impact, environment and recent developments, Computer Physics Communications 248, 107042 (2020).DOI
  82. [82]N. L. Nguyen, N. Colonna, A. Ferretti, and N. Marzari, Koopmans-compliant spectral functionals for extended systems, Phys. Rev. X 8, 021051 (2018).DOI
  83. [83]J. D. Elliott, N. Colonna, M. Marsili, N. Marzari, and P. Umari, Koopmans meets the β-sel⟭ter: Excitonic optical spectra without gw, Journal of Chemical Theory and Computation 15, 3710 (2019).
  84. [84]V. Vitale, G. Pizzi, A. Marrazzo, J. R. Yates, N. Marzari, and A. A. Mostofi, Automated high-throughput wannierisation, npj Computational Materials 6, 66 (2020).
  85. [85]J. Qiao, G. Pizzi, and N. Marzari, Prizeability disentanglement for accurate and automated electronic-structure hamiltonians, npj Computational Materials 9, 208 (2023).DOI
  86. [86]S. L. Adler, Quantum theory of the dielectric constant in real solids, Phys. Rev. 126, 413 (1962).DOI
  87. [87]N. Wiser, Dielectric constant with local field effects included, Phys. Rev. 129, 62 (1963).DOI
  88. [88]C. A. Rozzi, D. Varsano, A. Marini, E. K. U. Gross, and A. Rubio, Exact coulomb cutoff technique for supercell calculations, Phys. Rev. B 73, 205119 (2006).DOI
  89. [89]S. Ismail-Beigi, Truncation of periodic image interactions for confined systems, Phys. Rev. B 73, 233103 (2006).
  90. [90]J. Duchon, Splines minimizing rotation-invariant seminorms in sobolev spaces, in Constructive Theory of Functions of Several Variables, edited by W. Schemp and K. Zeller (Springer Berlin Heidelberg, Berlin, Heidelberg, 1977) pp. 85–100.
  91. [91]C. A. Micchelli, Interpolation of scattered data: Distance matrices and conditionally positive definite functions, Constructive Approximation 2, 11 (1986).

Paper details

vlnc:W_t38wqa43
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents