betaFeedback

WorksPhysics and Astronomy

G0W0G_0W_0@HF and BSE methods in periodic systems from Hartree-Fock theory: gaussian orbital and density fitting approach

Abstract

The GWGW method for calculating quasi-particle energies of solids commonly begin from a DFT Hamiltonian and Kohn-Sham orbitals in a plane wave basis. Screening of the coulomb interaction is implemented using the inverse dielectric function in the random phase approximation (RPA). We present G0W0G_0W_0 calculations which begin from the Hartree-Fock method in a basis of gaussian orbitals. The screened coulomb interaction, WW, is obtained using a WW = vv + vΠvvΠv approach without invoking a plasmon pole approximation. The polarizability, ΠΠ, in WW is treated at the RPA level. RPA polarizabilities require solution of Bethe-Salpeter equations (BSE) for each unique Q\textbf{Q} point. A strategy for obtaining self-energies which are converged with respect to number of virtual states is employed in which G0W0G_0W_0 yields the majority of the self-energy and the remaining part from high energy virtual levels is evaluated at second-order. The methods are evaluated by applying them to elemental semiconductors (C, Si) and oxides (MgO and anatase and rutile TiO2_2). Common errors of HF theory applied to materials include overestimation of both the band gap and valence band widths. These are corrected in the approach employed here. Typically, the RPA screened interaction results in overestimation of band gaps while the G0W0G_0W_0 self-energy band width renormalization yields band widths for diamond and Si which are in good agreement with experiment. HF calculations are performed in gaussian orbital basis sets and G0W0G_0W_0 and BSE calculations are performed using density fitting with a coulomb metric.

Introduction

Ab-initio GWGW and Bethe-Salpeter equation[1] (BSE) methods have been applied to crystalline materials over the last 60 years[2, 3, 4]. This has mostly been achieved using plane wave (PW) basis sets and DFT hamiltonians[4, 5] and PW/DFT implementations of GWGW and BSE methods are available in a number of codes, including BerkeleyGW[4, 6], Yambo[7, 8], VASP[9, 10] and Abinit[11, 12]. Here we report GoWoG_{o}W_{o}/BSE calculations for tetrahedral semiconductors (Si and C) and wide gapped oxides (MgO and TiO2_2) in a gaussian orbital (GO) basis using a density fitting approach using the Exciton code for both single-particle and GoWoG_{o}W_{o}/BSE calculations[13, 14, 15]. In contrast to most previous applications of the GoWoG_{o}W_{o}/BSE method in solids, we use Hartree-Fock (HF) wavefunctions and single-particle energies in the unperturbed hamiltonian. The HF method is well known to overestimate band gaps and valence band widths in solids and we show that the GoWoG_{o}W_{o} self-energies that we employ are capable of renormalizing the bandwidths of Si and diamond, to yield agreement with experimental bandwidths, as well as yielding GoWoG_{o}W_{o} band structures in reasonable agreement band structures from PW GoWoG_{o}W_{o} calculations.

The QQ-dependent screened interaction, W(Q)W(Q), in both GWGW and BSE calculations in this work is obtained from the interacting polarizability in the random phase approximations (RPA). This approach has the advantage that it does not require a plasmon pole approximation to the inverse dielectric function, however, it does require diagonalization of large RPA hamiltonians at all unique QQ points in the Brillouin zone (BZ) while generating the BSE hamiltonian. This approach can also be applied to calculation of exciton dispersion.

Density fitting of products of wavefunction orbitals is a long established technique in finite[16, 17, 18, 19] and periodic systems and periodic systems[15, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30]. The GO/density fitting approach has the advantage of a much reduced number of integrals

required over approaches that do not factorize them. It lends itself to calculations involving other interacting particles such as positrons[31], since the products above can be calculated separately for electron or positron wavefunctions. It may be possible to treat systems with large unit cells and open volumes such as polymers, amorphous and crystalline organic materials or metal organic frameworks. Gaussian basis sets in quantum chemistry are commonly generated using variational principles for the ground state, but this leaves open the question of their suitability for conduction bands of solids. We address this question here by calculating the free electron band structures of the materials chosen for study.

The present work follows earlier work in which density fitting was applied to time-dependent Hartree-Fock (TDHF) calculations[15] where screening of the electron-hole attraction term in the TDHF hamiltonian was implemented using a simple scaling factor. It employs coulomb-weighted density fitting in which the metric is the coulomb potential, rather than the overlap of the auxiliary basis set. Using DFT wavefunctions and single-particle energies in the unperturbed hamiltonian, Rohlfing and coworkers showed that a GO basis could reproduce planewave/DFT quasiparticle (QP) energy corrections in tetrahedral semiconductors[32, 33] and oxides[34] as well as excitonic optical spectra in wide and narrow gapped materials[35]. More recently, Zhu and Chan[36] reported GoWoG_{o}W_{o}@PBE calculations of valence and core level ionization potentials[36], Zgid and coworkers reported finite temperature[37] and relativistic[38] self-consistent scGWscGW calculations and García-Blázquez and Palacios[39] reported BSE calculations, all of which were in a GO basis with density fitting. Correlated electron methods such as coupled cluster singles and doubles (CCSD) for periodic systems[40] have also been implemented in a GO basis.

The remainder of this paper is organized as follows: the BSE for finite wavevector excitations is introduced using a linear response formalism, followed by the bare and interacting polarizabilities and second order and G0W0G_{0}W_{0} self-energies. The following section describes how these are implemented in a GO basis using density fitting. Results of G0W0G_{0}W_{0} self-energy and BSE calculations are reported in the following sections for diamond and Si, MgO and anatase and rutile TiO2_2. The final section provides discussion and further analysis of these results and conclusions regarding the GO/density fitting approach to excitations in gapped materials.

Theory

The BSE formalism is commonly introduced beginning from the two-body Green’s function[4, 10, 41], which leads to the well-known form,

Π(Q,ϵ)=Π0(Q,ϵ)+Π0(Q,ϵ)K(Q,ϵ)Π(Q,ϵ),(1)\Pi(\mathbf{Q},\epsilon)=\Pi_{0}(\mathbf{Q},\epsilon)+\Pi_{0}(\mathbf{Q},\epsilon)K(\mathbf{Q},\epsilon)\Pi(\mathbf{Q},\epsilon), \tag*{(1)}

where Π(Q,ϵ)\Pi(\mathbf{Q},\epsilon) and Π0(Q,ϵ)\Pi_{0}(\mathbf{Q},\epsilon) are, respectively, the interacting and non-interacting polarizability and Q\mathbf{Q} and ϵ\epsilon are wave vector and energy. The choice of interaction kernel, K(Q,ϵ)K(\mathbf{Q},\epsilon), determines the level of approximation for Π(Q,ϵ)\Pi(\mathbf{Q},\epsilon). In an RPA calculation, K(Q)K(\mathbf{Q}) contains the static, bare coulomb interaction, v(Q)v(\mathbf{Q}), in ring diagrams; in a BSE calculation K(Q,ϵ)K(\mathbf{Q},\epsilon) contains v(Q)v(\mathbf{Q}) in ring diagrams as well as the energy-dependent screened electron-hole interaction, W(Q,ϵ)W(\mathbf{Q},\epsilon).

Figure 1 shows two indirect transitions in which electron and hole wavevectors for each electron-hole pair differ by Q\mathbf{Q} while the wavevector connecting each electron-hole pair is q\mathbf{q}. This is the meaning of q\mathbf{q} and Q\mathbf{Q} in this work.

Band energy versus wave vector diagram for indirect transitions

Figure 1. Band energy versus wave vector diagram for indirect transitions with change in wave vector Q\mathbf{Q} between states ψik\psi_{ik} and ψak+Q\psi_{ak+Q} and states ψjk+q\psi_{jk+q} and ψbk+q+Q\psi_{bk+q+Q}. These electron-hole pairs differ in wave vector q\mathbf{q}.

In a conventional G0W0G_{0}W_{0}/BSE calculation of the optical properties of a material, Π(Q,ϵ)\Pi(\mathbf{Q},\epsilon) is calculated for Q→0\mathbf{Q}\to0. Energy denominators in Π0(Q,ϵ)\Pi_{0}(\mathbf{Q},\epsilon) are single particle Kohn-Sham eigenvalues with G0W0G_{0}W_{0} self-energy corrections. The screened interaction, W0(Q,ϵ)W_{0}(\mathbf{Q},\epsilon), is calculated using the RPA inverse dielectric function,

W0(Q,ϵ)=ϵ−1,RPA(Q,ϵ)v(Q)(2)W_{0}(\mathbf{Q},\epsilon)=\epsilon^{-1,\mathrm{RPA}}(\mathbf{Q},\epsilon)v(\mathbf{Q}) \tag*{(2)}

where

ϵRPA(Q,ϵ)=1−v(Q)Π0(Q,ϵ)(3)\epsilon^{\mathrm{RPA}}(\mathbf{Q},\epsilon)=1-v(\mathbf{Q})\Pi_{0}(\mathbf{Q},\epsilon) \tag*{(3)}

The frequency dependent inverse dielectric function in Eq. (2) is needed for the G0W0G_{0}W_{0} self-energy, which is the convolution of G0G_{0} and W0W_{0}. Inversion of the dielectric matrix as a function of frequency can be avoided using a plasmon pole approximation, whereby the dielectric matrix is inverted at zero frequency and a frequency dependence ansatz is introduced, for example, by several pole frequencies whose strengths and positions are fitted using sum rules[5].

We adopt an approach in which the screened interaction is given by the equivalent form,

W0(Q,ϵ)=v(Q)+v(Q)Π(Q,ϵ)v(Q),(4)W_{0}(\mathbf{Q},\epsilon)=v(\mathbf{Q})+v(\mathbf{Q})\Pi(\mathbf{Q},\epsilon)v(\mathbf{Q}), \tag*{(4)}

where Π\Pi is the RPA dressed polarizability obtained by solving RPA equations at unique, finite wave vectors, Q\mathbf{Q}. This screened interaction is used in the final BSE that is solved and used to obtain the imaginary part of the optical dielectric function at Q→0\mathbf{Q}\to0. This approach requires no plasmon pole approximation since the frequency dependence of W0(Q,ϵ)W_{0}(\mathbf{Q},\epsilon) is obtained exactly (within the limitations of the RPA and basis sets used) when the self-energy is constructed from RPA eigenvectors and eigenvalues. A similar approach was applied recently to ionization and excitation energies in medium sized molecules[42, 43].

Linear response equation of motion

A linear response approach, which emphasizes the connection of the BSE to linear response theory has been reviewed by Drew and Head-Gordon[44]. It yields the BSE equations for finite wave vector Q\mathbf{Q} excitations in a transparent manner, beginning from the equation of motion for the density matrix,

∂P^∂t=[F^,P^].(5)\frac{\partial\hat{P}}{\partial t}=[\hat{F},\hat{P}]. \tag*{(5)}

This is solved in the basis of eigenfunctions of the Fock operator {i,a}\{i,a\}. The Fock operator, F^\hat{F}, and density matrix, P^\hat{P}, are split into zeroth (0) and first order (1) parts, so that an element of the density matrix is given by,

Ppq=Ppq(0)+Ppq(1).(6)P_{pq}=P_{pq}^{(0)}+P_{pq}^{(1)}. \tag*{(6)}

The zeroth order Fock matrix, representing the ground state, is diagonal with elements,

Fpq(0)=ϵpδpq.(7)F_{pq}^{(0)}=\epsilon_{p}\delta_{pq}. \tag*{(7)}

where ϵp\epsilon_{p} are occupied and virtual state HF eigenvalues, ϵi\epsilon_{i} and ϵa\epsilon_{a}. The zeroth order density matrix in the {i,a}\{i,a\} basis is diagonal with unit values for states which are occupied in the ground state,

Ppq(0)=δpqθ(ϵkF−ϵp),(8)P_{pq}^{(0)}=\delta_{pq}\theta(\epsilon_{k_F}-\epsilon_{p}), \tag*{(8)}

and zero otherwise. ϵkF\epsilon_{k_F} is the Fermi energy. The electron-electron part of the zeroth order Fock matrix is,

Fpk+Qqk(0)=[2(ψpk+Q∗ψqk∗∣V∣ψrk+qψsk+q+Q)−(ψpk+Q∗ψsk+q+Q∗∣W∣ψrk+qψqk)]Psk+q+Q/rk+q,(9)F^{(0)}_{p k+Q q k} = \left[2(\psi^{*}_{p k+Q}\psi^{*}_{q k}|V|\psi_{r k+q}\psi_{s k+q+Q})-(\psi^{*}_{p k+Q}\psi^{*}_{s k+q+Q}|W|\psi_{r k+q}\psi_{q k})\right]P_{s k+q+Q/r k+q}, \tag*{(9)}

The static, screened interaction, WW, in the direct term in F(0)F^{(0)} is introduced here on the basis that screening of these terms is found in derivation of the BSE beginning from the two-body Green’s function and that this is usually approximated by the static, screened interaction which enables solution of the resulting equations as a generalized eigenvalue problem[4]. Two-electron integrals in the Fock matrix are given in chemists’ notation which lends itself to the density fitting approach outlined below. The first order Fock matrix is given by,

Fai(1)=∂Fai(0)∂PjbPjb(1)+∂Fai(0)∂PbjPbj(1),(10)F^{(1)}_{ai}=\frac{\partial F^{(0)}_{ai}}{\partial P_{jb}}P^{(1)}_{jb}+\frac{\partial F^{(0)}_{ai}}{\partial P_{bj}}P^{(1)}_{bj}, \tag*{(10)}

where the first order density matrix is,

Pbj(1)(t)=12[Xbje−iϵt+Yjb∗e+iϵt].(11)P^{(1)}_{bj}(t)=\frac{1}{2}\left[X_{bj}e^{-i\epsilon t}+Y^{*}_{jb}e^{+i\epsilon t}\right]. \tag*{(11)}

XX and YY are constant amplitudes to be determined from the equation of motion and Pbj(1)(t)=Pjb∗(1)(t)P^{(1)}_{bj}(t)=P^{*(1)}_{jb}(t). The system is assumed to be initially in its ground state, so that Pii(0)=1P^{(0)}_{ii}=1 and Paa(0)P^{(0)}_{aa} is zero, and to remain predominantly in its ground state, so that Paa(1)P^{(1)}_{aa} remains negligible. Off-diagonal elements of the commutators at first order are given by,

(F(0)P(1)−P(1)F(0))ai=ϵaPai(1)−Pai(1)ϵi,(12)\left(F^{(0)}P^{(1)}-P^{(1)}F^{(0)}\right)_{ai}=\epsilon_aP^{(1)}_{ai}-P^{(1)}_{ai}\epsilon_i, \tag*{(12)}

and

(F(1)P(0)−P(0)F(1))ai=Fai(1)Pii(0)−Paa(0)Fai(1).(13)\left(F^{(1)}P^{(0)}-P^{(0)}F^{(1)}\right)_{ai}=F^{(1)}_{ai}P^{(0)}_{ii}-P^{(0)}_{aa}F^{(1)}_{ai}. \tag*{(13)}

We now consider in more detail the equation of motion for excitations at wave vector QQ in a crystal. A complex conjugate pair of off-diagonal, first order density matrix elements is,

Pjk+q bk+q+Q(1)(t)=12[Xjk+q bk+q+Qe−iϵt+Yjk+q bk+q+Q∗e+iϵt](14)P^{(1)}_{j k+q\,b k+q+Q}(t)=\frac{1}{2}\left[X_{j k+q\,b k+q+Q}e^{-i\epsilon t}+Y^{*}_{j k+q\,b k+q+Q}e^{+i\epsilon t}\right] \tag*{(14)}

and

Pbk+q+Q jk+q(1)(t)=12[Xbk+q+Q jk+qe+iϵt+Ybk+q+Q jk+qe−iϵt].(15)P^{(1)}_{b k+q+Q\,j k+q}(t)=\frac{1}{2}\left[X_{b k+q+Q\,j k+q}e^{+i\epsilon t}+Y_{b k+q+Q\,j k+q}e^{-i\epsilon t}\right]. \tag*{(15)}

Inserting the density matrix into (12) yields,

Fak+Q,ak+Q(0)Pak+Q,ik(1)(t)−Pak+Q,ik(1)(t)Fik,ik(0)=(ϵak+Q−ϵik)Pak+Q,ik(1)(t),(16)F^{(0)}_{a k+Q,a k+Q}P^{(1)}_{a k+Q,i k}(t)-P^{(1)}_{a k+Q,i k}(t)F^{(0)}_{i k,i k}=(\epsilon_{a k+Q}-\epsilon_{i k})P^{(1)}_{a k+Q,i k}(t), \tag*{(16)}

(13) becomes,

Fak+Q,ik(1)Pik,ik(0)=[(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q)−(ψak+Q∗ψbk+q+Q∣W∣ψjk+q∗ψik)]Pjk+q+Q bk+q(1)+[(ψak+Q∗ψik∣V∣ψbk+q∗ψjk+q+Q)−(ψak+Q∗ψjk+q+Q∣W∣ψbk+q∗ψik)]Pjk+q+Q b∗k+q(1).(17)\begin{aligned} F^{(1)}_{a k+Q,i k}P^{(0)}_{i k,i k}=\left[(\psi^{*}_{a k+Q}\psi_{i k}|V|\psi^{*}_{j k+q}\psi_{b k+q+Q})-(\psi^{*}_{a k+Q}\psi_{b k+q+Q}|W|\psi^{*}_{j k+q}\psi_{i k})\right]P^{(1)}_{j k+q+Q\,b k+q} \\ &+\left[(\psi^{*}_{a k+Q}\psi_{i k}|V|\psi^{*}_{b k+q}\psi_{j k+q+Q})-(\psi^{*}_{a k+Q}\psi_{j k+q+Q}|W|\psi^{*}_{b k+q}\psi_{i k})\right]P^{(1)}_{j k+q+Q\,b^{*} k+q}. \tag*{(17)} \end{aligned}

Inserting (16) and 17 into (5) and equating coefficients of e−iϵte^{-i\epsilon t} with space variation eiQ⋅re^{i\mathbf{Q}\cdot\mathbf{r}} yields,

(ϵak+Q−ϵik−ϵ)Xik,a∗k+Q+[(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q)−(ψak+Q∗ψbk+q+Q∣W∣ψjk+q∗ψik)]Xjk+q b∗k+q+Q+[(ψak+Q∗ψik∣V∣ψbk+q∗ψjk+q+Q)−(ψak+Q∗ψjk+q+Q∣W∣ψbk+q∗ψik)]Yj∗k+q+Q bk+q∗.(18)\begin{aligned} (\epsilon_{a k+Q}-\epsilon_{i k}-\epsilon)X_{i k,a^{*} k+Q} +\left[(\psi^{*}_{a k+Q}\psi_{i k}|V|\psi^{*}_{j k+q}\psi_{b k+q+Q})-(\psi^{*}_{a k+Q}\psi_{b k+q+Q}|W|\psi^{*}_{j k+q}\psi_{i k})\right]X_{j k+q\,b^{*} k+q+Q} \\ &+\left[(\psi^{*}_{a k+Q}\psi_{i k}|V|\psi^{*}_{b k+q}\psi_{j k+q+Q})-(\psi^{*}_{a k+Q}\psi_{j k+q+Q}|W|\psi^{*}_{b k+q}\psi_{i k})\right]Y^{*}_{j^{*} k+q+Q\,b k+q}. \tag*{(18)} \end{aligned}

A similar analysis of the iki k, ak+Qa k+Q element of the density matrix in (5) yields,

(ϵak+Q−ϵik+ϵ)Yak+Q,ik+[(ψik+Q∗ψak∣V∣ψjk+q∗ψbk+q+Q)−(ψik+Q∗ψbk+q+Q∣W∣ψjk+q∗ψak)]Xjk+q b∗k+q+Q+[(ψik+Q∗ψak∣V∣ψbk+q∗ψjk+q+Q)−(ψik+Q∗ψjk+q+Q∣W∣ψbk+q∗ψak)]Yj∗k+q+Q bk+q.(19)\begin{aligned} (\epsilon_{a k+Q}-\epsilon_{i k}+\epsilon)Y_{a k+Q,i k} +\left[(\psi^{*}_{i k+Q}\psi_{a k}|V|\psi^{*}_{j k+q}\psi_{b k+q+Q})-(\psi^{*}_{i k+Q}\psi_{b k+q+Q}|W|\psi^{*}_{j k+q}\psi_{a k})\right]X_{j k+q\,b^{*} k+q+Q} \\ &+\left[(\psi^{*}_{i k+Q}\psi_{a k}|V|\psi^{*}_{b k+q}\psi_{j k+q+Q})-(\psi^{*}_{i k+Q}\psi_{j k+q+Q}|W|\psi^{*}_{b k+q}\psi_{a k})\right]Y_{j^{*} k+q+Q\,b k+q}. \tag*{(19)} \end{aligned}

(19) and 20 and their complex conjugates can be written in the form,

AX+BY=+ΩXB∗X+A∗Y=−ΩYA∗X∗+B∗Y∗=−ΩX∗BX∗+AY∗=+ΩY∗(20)\begin{aligned} AX+BY&=+\Omega X\\ B^{*}X+A^{*}Y&=-\Omega Y\\ A^{*}X^{*}+B^{*}Y^{*}&=-\Omega X^{*}\\ BX^{*}+AY^{*}&=+\Omega Y^{*} \tag*{(20)} \end{aligned}

The correspondence between two-electron matrix elements in Eq. (19) and 20 and the AA and BB matrix blocks in Eq. (20) is given in Table I as well as further Tamm-Dancoff (TDA), RPA and TDHF approximations. Feynman diagram representations of the matrix elements in Eq. (19) and 20 are shown in Figure 2 and Figure 3. The system of coupled equations in Eq. (20) can be represented as the BSE,

Feynman diagrams contributing to the exchange parts of Fock matrix elements

Figure 2. Feynman diagrams which contribute to the exchange parts of Fock matrix elements Fak+Q,ikF_{ak+Q,ik} and Fik+Q,akF_{ik+Q,ak}. (a) AX, (b) BY (c) AY and (d) BX. Momentum transfer is indicated by Q.

Feynman diagrams contributing to the direct parts of Fock matrix elements

Figure 3. Feynman diagrams which contribute to the direct parts of Fock matrix elements Fak+Q,ikF_{ak+Q,ik} and Fik+Q,akF_{ik+Q,ak}. (a) AX, (b) BY (c) AY and (d) BX. Momentum transfer is indicated by q.

(ABB∗A∗)(XY∗YX∗)=(Ω00−Ω)(XY∗YX∗)(21)\begin{pmatrix} A & B \\ B^{*} & A^{*} \end{pmatrix} \begin{pmatrix} X & Y^{*} \\ Y & X^{*} \end{pmatrix} = \begin{pmatrix} \Omega& 0 \\ 0 & -\Omega\end{pmatrix} \begin{pmatrix} X & Y^{*} \\ Y & X^{*} \tag*{(21)} \end{pmatrix}

Eq. (21) is a generalized eigenvalue problem in which eigenvalues occur in positive and negative pairs, which we denote as Ω±S\Omega_{\pm}^{S}, with Ω+S\Omega_{+}^{S} denoting the positive eigenvalue of the pair. Components of the SthS^{th} eigenvector at wave vector Q are denoted, Xikak+QX_{ikak+Q} and Ykiak+QY_{kiak+Q}, (Eq. (21)). The corresponding eigenvalue is denoted, ΩS(Q)\Omega^{S}(Q).

Off-diagonal BB sub-blocks of the BSE matrix, which couple the resonant, AA, and anti-resonant, A∗A^{*}, sub-blocks of the matrix, typically have a limited effect on optical absorption

properties predicted by the BSE[10]. When this is the case, YY components of the eigenvectors in Eq. (21) are small compared to XX components. The Tamm-Dancoff approximation (TDA) to the BSE omits the BB sub-blocks in Eq. (21) and requires instead solution of the ordinary eigenvalue problems,

AX=XΩ(22)AX=X\Omega \tag*{(22)}
A∗X∗=−X∗Ω.A^{*}X^{*}=-X^{*}\Omega.

The TDA is used from here on in this work.

Bare and dressed polarizabilities

Resonant and anti-resonant parts of the non-interacting polarizability, Π0(Q,ϵ)\Pi^{0}(Q,\epsilon), are usually obtained as products of HF electron and hole Green’s functions. They can also be obtained as the inverse of the BSE matrix in Eq. (21). For the case where AA and BB blocks are omitted from Eq. (21) so that only virtual and occupied energy eigenvalue differences remain on the diagonal of the BSE matrix, X\mathbf{X} is the unit matrix and Y\mathbf{Y} is zero. Resonant and anti-resonant parts of the non-interacting HF polarizability, Π0(Q,ϵ)\Pi^{0}(\mathbf{Q},\epsilon), are then given by,

iΠRes.0(Q,ϵ)=∑i,a,kψik(x1)ψak+Q∗(x1)ψik∗(x2)ψak+Q(x2)ϵ−(ϵak+Q−ϵik)+iδ(23)i\Pi_{\mathrm{Res.}}^{0}(\mathbf{Q},\epsilon)=\sum_{i,a,\mathbf{k}}\frac{\psi_{i\mathbf{k}}(\mathbf{x}_{1})\psi^{*}_{a\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{1})\psi^{*}_{i\mathbf{k}}(\mathbf{x}_{2})\psi_{a\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{2})}{\epsilon-(\epsilon_{a\mathbf{k}+\mathbf{Q}}-\epsilon_{i\mathbf{k}})+i\delta} \tag*{(23)}

and,

(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q∗)(24)\left(\psi^{*}_{a\mathbf{k}+\mathbf{Q}}\psi_{i\mathbf{k}}|V|\psi^{*}_{j\mathbf{k}+\mathbf{q}}\psi^{*}_{b\mathbf{k}+\mathbf{q}+\mathbf{Q}}\right) \tag*{(24)}

and using time-reversal symmetry, ψak∗=ψa,−k\psi^{*}_{a\mathbf{k}}=\psi_{a,-\mathbf{k}}, yields,

(ψa,−(k+Q)∗ψi,−k∣V∣ψj,−(k+q)∗ψb,−(k+q+Q))(25)\left(\psi^{*}_{a,-(\mathbf{k}+\mathbf{Q})}\psi_{i,-\mathbf{k}}|V|\psi^{*}_{j,-(\mathbf{k}+\mathbf{q})}\psi_{b,-(\mathbf{k}+\mathbf{q}+\mathbf{Q})}\right) \tag*{(25)}

Adding a reciprocal lattice vector, G\mathbf{G}, to bring each wave vector into the first BZ yields,

(ψak′′∗ψik′′+Q∣V∣ψjk′′+q′′+Q∗ψbk′′+q′′)(26)\left(\psi^{*}_{a\mathbf{k}^{\prime\prime}}\psi_{i\mathbf{k}^{\prime\prime}+\mathbf{Q}}|V|\psi^{*}_{j\mathbf{k}^{\prime\prime}+\mathbf{q}^{\prime\prime}+\mathbf{Q}}\psi_{b\mathbf{k}^{\prime\prime}+\mathbf{q}^{\prime\prime}}\right) \tag*{(26)}
k′′=−(k+Q)+Gq′′=−q(27)\begin{aligned} \mathbf{k}^{\prime\prime}&=-(\mathbf{k}+\mathbf{Q})+\mathbf{G}\\ \mathbf{q}^{\prime\prime}&=-\mathbf{q} \tag*{(27)} \end{aligned}

i.e. the RPA-TDA matrix A∗\mathbf{A}^{*} has eigenvectors Xik+Qak∗X^{*}_{i\mathbf{k}+\mathbf{Q}a\mathbf{k}} which are complex conjugates of eigenvectors of the RPA-TDA matrix A\mathbf{A} with the same eigenvalues. Wave vectors k\mathbf{k} in Xikak+QX_{i\mathbf{k}a\mathbf{k}+\mathbf{Q}} are reordered in Xik+Qak∗X^{*}_{i\mathbf{k}+\mathbf{Q}a\mathbf{k}}. Later, where the polarizability is used in the GWGW self-energy, these eigenvectors are contracted with wave function products summed over all wave vectors and thus this reordering is immaterial. The original RPA-TDA anti-resonant polarizability,

iΠAnti−Res.(Q,ϵ)=−∑i,a,k,Sψak(x1)ψik+Q∗(x1)Xik+QakSXik+QakS∗ψak∗(x2)ψik+Q(x2)ϵ+ΩS(Q)−iδ(28)i\Pi_{\mathrm{Anti-Res.}}(\mathbf{Q},\epsilon)=-\sum_{i,a,\mathbf{k},S}\frac{\psi_{a\mathbf{k}}(\mathbf{x}_{1})\psi^{*}_{i\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{1})X^{S}_{i\mathbf{k}+\mathbf{Q}a\mathbf{k}}X^{S*}_{i\mathbf{k}+\mathbf{Q}a\mathbf{k}}\psi^{*}_{a\mathbf{k}}(\mathbf{x}_{2})\psi_{i\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{2})}{\epsilon+\Omega^{S}(\mathbf{Q})-i\delta} \tag*{(28)}

can therefore be obtained as,

iΠAnti−Res.(Q,ϵ)=−∑i,a,k,Sψik(x1)ψak+Q∗(x1)Xikak+QSXikak+QS∗ψik∗(x2)ψak+Q(x2)ϵ+ΩS(Q)−iδ(29)i\Pi_{\mathrm{Anti-Res.}}(\mathbf{Q},\epsilon)=-\sum_{i,a,\mathbf{k},S}\frac{\psi_{i\mathbf{k}}(\mathbf{x}_{1})\psi^{*}_{a\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{1})X^{S}_{i\mathbf{k}a\mathbf{k}+\mathbf{Q}}X^{S*}_{i\mathbf{k}a\mathbf{k}+\mathbf{Q}}\psi^{*}_{i\mathbf{k}}(\mathbf{x}_{2})\psi_{a\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{2})}{\epsilon+\Omega^{S}(\mathbf{Q})-i\delta} \tag*{(29)}

Self-energy

The G0W0G_{0}W_{0} self-energy is the convolution of the non-interacting Green’s function, G0(r,r′,ϵ)G_{0}(\mathbf{r},\mathbf{r}^{\prime},\epsilon), with a screened interaction (Eq. 4),

iΠAnti−Res.0(Q,ϵ)=−∑i,a,kψak(x1)ψik+Q∗(x1)ψak∗(x2)ψik+Q(x2)ϵ+(ϵak−ϵik+Q)−iδ(30)i\Pi_{\mathrm{Anti-Res.}}^{0}(\mathbf{Q},\epsilon)=-\sum_{i,a,\mathbf{k}}\frac{\psi_{a\mathbf{k}}(\mathbf{x}_{1})\psi^{*}_{i\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{1})\psi^{*}_{a\mathbf{k}}(\mathbf{x}_{2})\psi_{i\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{2})}{\epsilon+(\epsilon_{a\mathbf{k}}-\epsilon_{i\mathbf{k}+\mathbf{Q}})-i\delta} \tag*{(30)}

Similarly, the RPA-TDA polarizability, Π(Q,ϵ)\Pi(\mathbf{Q},\epsilon), can be expressed in terms of eigenvectors, X\mathbf{X} and RPA-TDA eigenvalues, ΩS(Q)\Omega^{S}(\mathbf{Q}), which are obtained by solving the BSE equation using RPA-TDA A\mathbf{A} matrix elements in Table 1. Resonant and anti-resonant parts of the polarizability are obtained as inverses of the two subsystems in (22). The resonant part is obtained from the first of these equations and is given by,

MethodAai,bjA_{ai,bj}Aai,bj∗A^{*}_{ai,bj}Bai,bjB_{ai,bj}Bai,bj∗B^{*}_{ai,bj}
TDHF2(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi^{*}_{jk+q}\psi_{bk+q+Q})−(ψak+Q∗ψbk+q+Q∣V∣ψjk+q∗ψik)-(\psi^{*}_{ak+Q}\psi_{bk+q+Q}|V|\psi^{*}_{jk+q}\psi_{ik})2(ψak+Q∗ψik∣V∣ψbk+qψjk+q+Q∗)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi_{bk+q}\psi^{*}_{jk+q+Q})−(ψak+Q∗ψjk+q+Q∗∣V∣ψbk+qψik)-(\psi^{*}_{ak+Q}\psi^{*}_{jk+q+Q}|V|\psi_{bk+q}\psi_{ik})
BSE2(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi^{*}_{jk+q}\psi_{bk+q+Q})−(ψak+Q∗ψbk+q+Q∣W∣ψjk+q∗ψik)-(\psi^{*}_{ak+Q}\psi_{bk+q+Q}|W|\psi^{*}_{jk+q}\psi_{ik})2(ψak+Q∗ψik∣V∣ψbk+qψjk+q+Q∗)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi_{bk+q}\psi^{*}_{jk+q+Q})−(ψak+Q∗ψjk+q+Q∗∣W∣ψbk+qψik)-(\psi^{*}_{ak+Q}\psi^{*}_{jk+q+Q}|W|\psi_{bk+q}\psi_{ik})
RPA2(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi^{*}_{jk+q}\psi_{bk+q+Q})2(ψak+Q∗ψik∣V∣ψbk+qψjk+q+Q∗)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi_{bk+q}\psi^{*}_{jk+q+Q})
TDHF-TDA2(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi^{*}_{jk+q}\psi_{bk+q+Q})−(ψak+Q∗ψbk+q+Q∣V∣ψjk+q∗ψik)-(\psi^{*}_{ak+Q}\psi_{bk+q+Q}|V|\psi^{*}_{jk+q}\psi_{ik})
BSE-TDA2(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi^{*}_{jk+q}\psi_{bk+q+Q})−(ψak+Q∗ψbk+q+Q∣W∣ψjk+q∗ψik)-(\psi^{*}_{ak+Q}\psi_{bk+q+Q}|W|\psi^{*}_{jk+q}\psi_{ik})
RPA-TDA2(ψak+Q∗ψik∣V∣ψjk+q∗ψbk+q+Q)2(\psi^{*}_{ak+Q}\psi_{ik}|V|\psi^{*}_{jk+q}\psi_{bk+q+Q})

Table 1. Two-electron matrix elements in AA and BB blocks of the BSE matrix (Eq. (21)) used for calculation of the RPA-TDA polarizability, Π\Pi, and for BSE-TDA excited state calculations. Factors of 2 appear from spin summation. The diagonals of the AA and A∗A^{*} blocks also contain eigenvalue differences, (ϵak+Q−ϵik)δak+Q,bk+Qδik,jk(\epsilon_{ak+Q}-\epsilon_{ik})\delta_{ak+Q,bk+Q}\delta_{ik,jk}. For RPA-TDA polarizability calculations these are HF single-particle eigenvalues and for BSE-TDA they are HF eigenvalues with GoWoG_{o}W_{o}@HF self-energy corrections.

iΠRes.(Q,ϵ)=∑i,a,k,Sψik(x1)ψak+Q∗(x1)Xikak+QSXikak+QS∗ψik∗(x2)ψak+Q(x2)ϵ−ΩS(Q)+iδ(31)i\Pi_{\mathrm{Res.}}(\mathbf{Q},\epsilon)=\sum_{i,a,\mathbf{k},S}\frac{\psi_{i\mathbf{k}}(\mathbf{x}_{1})\psi^{*}_{a\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{1})X^{S}_{i\mathbf{k}a\mathbf{k}+\mathbf{Q}}X^{S*}_{i\mathbf{k}a\mathbf{k}+\mathbf{Q}}\psi^{*}_{i\mathbf{k}}(\mathbf{x}_{2})\psi_{a\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{2})}{\epsilon-\Omega^{S}(\mathbf{Q})+i\delta} \tag*{(31)}

The RPA-TDA A\mathbf{A} block matrix elements for the resonant polarizability are shown in Figure 2(a). The anti-resonant part can be obtained by setting up an additional RPA-TDA matrix corresponding to A∗\mathbf{A}^{*} (Figure 2(c)) and diagonalizing again. However, it is possible to obtain the non-resonant part using eigenvalues and eigenvectors from the same A\mathbf{A} matrix as the resonant part. Taking the complex conjugates of each wave function in the matrix element,

where,

k′′=−(k+Q)+G(32)\mathbf{k}^{\prime\prime}=-(\mathbf{k}+\mathbf{Q})+\mathbf{G} \tag*{(32)}
q′′=−q(33)\mathbf{q}^{\prime\prime}=-\mathbf{q} \tag*{(33)}

i.e. the RPA-TDA matrix A∗\mathbf{A}^{*} has eigenvectors Xik+Qak∗X^{*}_{i\mathbf{k}+\mathbf{Q}a\mathbf{k}} which are complex conjugates of eigenvectors of the RPA-TDA matrix A\mathbf{A} with the same eigenvalues. Wave vectors k\mathbf{k} in Xikak+QX_{i\mathbf{k}a\mathbf{k}+\mathbf{Q}} are reordered in Xik+Qak∗X^{*}_{i\mathbf{k}+\mathbf{Q}a\mathbf{k}}. Later, where the polarizability is used in the GWGW self-energy, these eigenvectors are contracted with wave function products summed over all wave vectors and thus this reordering is immaterial. The original RPA-TDA anti-resonant polarizability,

iΠAnti−Res.(Q,ϵ)=−∑i,a,k,Sψak(x1)ψik+Q∗(x1)Xik+QakSXik+QakS∗ψak∗(x2)ψik+Q(x2)ϵ+ΩS(Q)−iδ(34)i\Pi_{\mathrm{Anti-Res.}}(\mathbf{Q},\epsilon)=-\sum_{i,a,\mathbf{k},S}\frac{\psi_{a\mathbf{k}}(\mathbf{x}_{1})\psi^{*}_{i\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{1})X^{S}_{i\mathbf{k}+\mathbf{Q}a\mathbf{k}}X^{S*}_{i\mathbf{k}+\mathbf{Q}a\mathbf{k}}\psi^{*}_{a\mathbf{k}}(\mathbf{x}_{2})\psi_{i\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{2})}{\epsilon+\Omega^{S}(\mathbf{Q})-i\delta} \tag*{(34)}

can therefore be obtained as,

iΠAnti−Res.(Q,ϵ)=−∑i,a,k,Sψik(x1)ψak+Q∗(x1)Xikak+QSXikak+QS∗ψik∗(x2)ψak+Q(x2)ϵ+ΩS(Q)−iδ(35)i\Pi_{\mathrm{Anti-Res.}}(\mathbf{Q},\epsilon)=-\sum_{i,a,\mathbf{k},S}\frac{\psi_{i\mathbf{k}}(\mathbf{x}_{1})\psi^{*}_{a\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{1})X^{S}_{i\mathbf{k}a\mathbf{k}+\mathbf{Q}}X^{S*}_{i\mathbf{k}a\mathbf{k}+\mathbf{Q}}\psi^{*}_{i\mathbf{k}}(\mathbf{x}_{2})\psi_{a\mathbf{k}+\mathbf{Q}}(\mathbf{x}_{2})}{\epsilon+\Omega^{S}(\mathbf{Q})-i\delta} \tag*{(35)}
ΣGW(r,r′,ϵ)=∫−∞+∞dϵ′2πiG0(r,r′,ϵ−ϵ′)W0(r,r′,ϵ′).(36)\Sigma^{GW}(\mathbf{r},\mathbf{r}',\epsilon)=\int_{-\infty}^{+\infty}\frac{d\epsilon'}{2\pi i}G_{0}(\mathbf{r},\mathbf{r}',\epsilon-\epsilon')W_{0}(\mathbf{r},\mathbf{r}',\epsilon'). \tag*{(36)}

The bare coulomb interaction, v(Q)v(\mathbf{Q}), in Eq. (4) contributes the static, HF exchange part of the self-energy, which is included in the HF SCF calculation in this work. The second order self-energy, Σ(2)(r,r′,ϵ)\Sigma^{(2)}(\mathbf{r},\mathbf{r}',\epsilon), is obtained using the non-interacting polarizability (Eq. (30) and 25) in Eq. (4). Diagonal matrix elements of the second order self-energy, Σ(2)(Q,ϵ)\Sigma^{(2)}(\mathbf{Q},\epsilon) (Figure 4), are given by,

Diagrams for matrix elements of the second order self-energy

Figure 4. Diagrams for matrix elements of the second order self-energy ⟨ψnk|Σ(2)(Q,ϵ)|ψnk⟩\left\langle\psi_{nk}\middle|\Sigma^{(2)}(\mathbf{Q},\epsilon)\middle|\psi_{nk}\right\rangle.

⟨ψnk|Σ(2)(ϵ)|ψnk⟩=∑j,b,j′,b′,k′,Q(ψnkψbk−Q∗|ψj′k′ψb′k′+Q∗)(ψj′k′∗ψb′k′+Q|ψbk−Qψnk∗)ϵ−ϵbk−Q−(ϵb′k′+Q−ϵj′k′)+(ψnkψjk−Q∗|ψj′k′ψb′k′+Q∗)(ψj′k′∗ψb′k′+Q|ψjk−Qψnk∗)ϵ−ϵjk−Q−(ϵj′k′+Q−ϵb′k′)(37)\left\langle\psi_{nk}\middle|\Sigma^{(2)}(\epsilon)\middle|\psi_{nk}\right\rangle=\sum_{j,b,j',b',k',\mathbf{Q}}\frac{\left(\psi_{nk}\psi^{*}_{bk-\mathbf{Q}}\middle|\psi_{j'k'}\psi^{*}_{b'k'+\mathbf{Q}}\right)\left(\psi^{*}_{j'k'}\psi_{b'k'+\mathbf{Q}}\middle|\psi_{bk-\mathbf{Q}}\psi^{*}_{nk}\right)}{\epsilon-\epsilon_{bk-\mathbf{Q}}-\left(\epsilon_{b'k'+\mathbf{Q}}-\epsilon_{j'k'}\right)}+\frac{\left(\psi_{nk}\psi^{*}_{jk-\mathbf{Q}}\middle|\psi_{j'k'}\psi^{*}_{b'k'+\mathbf{Q}}\right)\left(\psi^{*}_{j'k'}\psi_{b'k'+\mathbf{Q}}\middle|\psi_{jk-\mathbf{Q}}\psi^{*}_{nk}\right)}{\epsilon-\epsilon_{jk-\mathbf{Q}}-\left(\epsilon_{j'k'+\mathbf{Q}}-\epsilon_{b'k'}\right)} \tag*{(37)}

The GoWoG_{o}W_{o} self-energy is obtained using the interacting polarizability (Eq. 30 and 31) in Eq. (4). Diagonal matrix elements of ΣGW(r,r′,ϵ)\Sigma^{GW}(\mathbf{r},\mathbf{r}',\epsilon) are given by,

⟨ψnk|ΣGW(ϵ)|ψnk⟩=∑j,b,j′,b′,k′,Q,S(ψnkψbk−Q∗|ψj′k′ψb′k′+Q∗)Xj′k′/b′k′+QSXj′′k′′/b′′k′′+QS†(ψj′k′∗ψb′k′+Q|ψbk−Qψnk∗)ϵ−ϵbk−Q−ΩS(Q)+(ψnkψjk−Q∗|ψb′k′ψj′k′+Q∗)Xj′k′/b′k′+QSXj′′k′′/b′′k′′+QS†(ψb′k′∗ψj′k′+Q|ψjk−Qψnk∗)ϵ−ϵjk−Q+ΩS(Q)(38)\left\langle\psi_{nk}\middle|\Sigma^{GW}(\epsilon)\middle|\psi_{nk}\right\rangle=\sum_{j,b,j',b',k',\mathbf{Q},S}\frac{\left(\psi_{nk}\psi^{*}_{bk-\mathbf{Q}}\middle|\psi_{j'k'}\psi^{*}_{b'k'+\mathbf{Q}}\right)X^{S}_{j'k'/b'k'+\mathbf{Q}}X^{S\dagger}_{j''k''/b''k''+\mathbf{Q}}\left(\psi^{*}_{j'k'}\psi_{b'k'+\mathbf{Q}}\middle|\psi_{bk-\mathbf{Q}}\psi^{*}_{nk}\right)}{\epsilon-\epsilon_{bk-\mathbf{Q}}-\Omega^{S}(\mathbf{Q})}+\frac{\left(\psi_{nk}\psi^{*}_{jk-\mathbf{Q}}\middle|\psi_{b'k'}\psi^{*}_{j'k'+\mathbf{Q}}\right)X^{S}_{j'k'/b'k'+\mathbf{Q}}X^{S\dagger}_{j''k''/b''k''+\mathbf{Q}}\left(\psi^{*}_{b'k'}\psi_{j'k'+\mathbf{Q}}\middle|\psi_{jk-\mathbf{Q}}\psi^{*}_{nk}\right)}{\epsilon-\epsilon_{jk-\mathbf{Q}}+\Omega^{S}(\mathbf{Q})} \tag*{(38)}

Various self-energy approximations are possible within this approach by solving particular approximations to Eq. (21) and using the resulting eigenvectors and excitation energies to construct the self-energy. The computational expense of the RPA-TDA calculations used to construct the self-energy depends on the number of occupied and virtual states used. Most of the self-energy for valence excitations is obtained from transitions between states close to the Fermi level. Here we use Eq. (38) to recover most of the self-energy and Eq. (37) for the remainder. The latter does not require a RPA-TDA calculation and is therefore less expensive to evaluate. For example, for diamond, four occupied states and 10 virtual states were included in the GoWoG_{o}W_{o} self-energy and transitions between these occupied states and an additional 20 states were included in the second order self-energy. The RPA-TDA matrix for an 8×8×88 \times8 \times8 Monkhorst-Pack (MP) mesh with these

occupied and virtual states is of dimension 20480 and for a 10×10×1010 \times10 \times10 mesh it is of dimension 40000.

Density fitting

Density fitting is used to project wavefunction products,

ψak∗(r)ψbk+q(r),\psi^{*}_{ak}(\mathbf{r})\psi_{bk+\mathbf{q}}(\mathbf{r}),
ψak∗(r)ψik+q(r),\psi^{*}_{ak}(\mathbf{r})\psi_{ik+\mathbf{q}}(\mathbf{r}),
ψik∗(r)ψak+q(r),\psi^{*}_{ik}(\mathbf{r})\psi_{ak+\mathbf{q}}(\mathbf{r}),
ψik∗(r)ψjk+q(r),\psi^{*}_{ik}(\mathbf{r})\psi_{jk+\mathbf{q}}(\mathbf{r}),

i.e. virtual-virtual, virtual-occupied, occupied-virtual and occupied-occupied products at wave vectors k\mathbf{k} and k+q\mathbf{k}+\mathbf{q}, onto an auxiliary Gaussian orbital basis. Bloch functions with lattice translation symmetry are constructed from crystal orbitals (CO),

ψik(r)=cikmϕm(r−m−A)eik⋅A,(39)\psi_{ik}(\mathbf{r})=c_{ikm}\phi_{m}(\mathbf{r}-\mathbf{m}-\mathbf{A})e^{i\mathbf{k}\cdot\mathbf{A}}, \tag*{(39)}

which are linear combinations of basis functions, ϕm(r−m−A)\phi_{m}(\mathbf{r}-\mathbf{m}-\mathbf{A}), at site, m\mathbf{m}, in the unit cell with lattice translation vector, A\mathbf{A}, and expansion coefficient, cikmc_{ikm}, for the ithi^{\mathrm{th}} occupied state at wavevector, k\mathbf{k}. Wavefunction product densities (Figure 5) are expressed in auxiliary basis functions, χβ+q(r)\chi_{\beta}^{+\mathbf{q}}(\mathbf{r}), as15^{15},

Charge density at point, r, arising from wavefunction product

Figure 5. Charge density at point, r, arising from wavefunction product ψik∗(r)ψak+q(r)\psi^{*}_{ik}(\mathbf{r})\psi_{ak+\mathbf{q}}(\mathbf{r}).

ϕik∗(r)ψak+q(r)=(ϕik∗ψak+q|χα−q)[Vαβ−q]−1χβ+q(r).(40)\phi_{i\mathbf{k}}^{*}(\mathbf{r})\psi_{a\mathbf{k}+\mathbf{q}}(\mathbf{r})=\left(\phi_{i\mathbf{k}}^{*}\psi_{a\mathbf{k}+\mathbf{q}}\middle|\chi_{\alpha}^{-\mathbf{q}}\right)\left[V_{\alpha\beta}^{-\mathbf{q}}\right]^{-1}\chi_{\beta}^{+\mathbf{q}}(\mathbf{r}). \tag*{(40)}

The auxiliary basis CO, χα+q(r)\chi_{\alpha}^{+\mathbf{q}}(\mathbf{r}), is a lattice sum of auxiliary basis functions on site, n, in unit cell, A,

χα+q(r)=χα(r−n−A)eiq⋅A.(41)\chi_{\alpha}^{+\mathbf{q}}(\mathbf{r})=\chi_{\alpha}(\mathbf{r}-\mathbf{n}-\mathbf{A})e^{i\mathbf{q}\cdot\mathbf{A}}. \tag*{(41)}

(ψik∗ψak+q|χα−q)\left(\psi_{i\mathbf{k}}^{*}\psi_{a\mathbf{k}+\mathbf{q}}\middle|\chi_{\alpha}^{-\mathbf{q}}\right) is the three-center integral,

Vmnβk,q=∑B,C∫drdr′ϕm∗(r)ϕn(r−C)χβ∗(r′−B)∣r−r′−B∣ei(k+q)⋅Ce−iq⋅B,(42)V_{mn\beta}^{\mathbf{k},\mathbf{q}}=\sum_{\mathbf{B},\mathbf{C}}\int d\mathbf{r}d\mathbf{r}'\frac{\phi_{m}^{*}(\mathbf{r})\phi_{n}(\mathbf{r}-\mathbf{C})\chi_{\beta}^{*}(\mathbf{r}'-\mathbf{B})}{\lvert\mathbf{r}-\mathbf{r}'-\mathbf{B}\rvert}e^{i(\mathbf{k}+\mathbf{q})\cdot\mathbf{C}}e^{-i\mathbf{q}\cdot\mathbf{B}}, \tag*{(42)}

contracted with expansion coefficients of the wavefunctions, ψik∗(r)\psi_{i\mathbf{k}}^{*}(\mathbf{r}) and ψak+q(r)\psi_{a\mathbf{k}+\mathbf{q}}(\mathbf{r}), and [Vαβ−q]−1\left[V_{\alpha\beta}^{-\mathbf{q}}\right]^{-1} is the inverse of the matrix,

Vαβ−q=∑A∫drdr′χα(r)χβ∗(r′)∣r−r′−A∣e−iq⋅A.(43)V_{\alpha\beta}^{-\mathbf{q}}=\sum_{\mathbf{A}}\int d\mathbf{r}d\mathbf{r}'\frac{\chi_{\alpha}(\mathbf{r})\chi_{\beta}^{*}(\mathbf{r}')}{\lvert\mathbf{r}-\mathbf{r}'-\mathbf{A}\rvert}e^{-i\mathbf{q}\cdot\mathbf{A}}. \tag*{(43)}

Both are integrals over the Ewald potential,

∑Aeiq⋅A∣r−r′−A∣=∑G4πΩe−∣q+G∣24γ2∣q+G∣2ei(q+G)⋅(r−r′)+∑Aerfc⁡(γ1/2∣r−r′−A∣)∣r−r′−A∣eiq⋅A.(44)\sum_{\mathbf{A}}\frac{e^{i\mathbf{q}\cdot\mathbf{A}}}{\lvert\mathbf{r}-\mathbf{r}'-\mathbf{A}\rvert}=\sum_{\mathbf{G}}\frac{4\pi}{\Omega}\frac{e^{-\frac{\lvert\mathbf{q}+\mathbf{G}\rvert^{2}}{4\gamma^{2}}}}{\lvert\mathbf{q}+\mathbf{G}\rvert^{2}}e^{i(\mathbf{q}+\mathbf{G})\cdot(\mathbf{r}-\mathbf{r}')}+\sum_{\mathbf{A}}\frac{\operatorname{erfc}(\gamma^{1/2}\lvert\mathbf{r}-\mathbf{r}'-\mathbf{A}\rvert)}{\lvert\mathbf{r}-\mathbf{r}'-\mathbf{A}\rvert}e^{i\mathbf{q}\cdot\mathbf{A}}. \tag*{(44)}

Combining two factors of density products in Eq. (40) and the Ewald potential, results in the following expression15^{15} for the leftmost, unscreened two-electron integral in Figure 2(a),

(ψak+Q∗(r)ψik(r)|ψjk+q(r′)ψbk+q+Q(r′))=(ψik∗ψak+Q|χα−Q)[Vαβ−Q]−1(ψjk+q∗ψbk+q+Q|χβQ).(45)\left(\psi_{a\mathbf{k}+\mathbf{Q}}^{*}(\mathbf{r})\psi_{i\mathbf{k}}(\mathbf{r})\middle|\psi_{j\mathbf{k}+\mathbf{q}}(\mathbf{r}')\psi_{b\mathbf{k}+\mathbf{q}+\mathbf{Q}}(\mathbf{r}')\right)=\left(\psi_{i\mathbf{k}}^{*}\psi_{a\mathbf{k}+\mathbf{Q}}\middle|\chi_{\alpha}^{-\mathbf{Q}}\right)\left[V_{\alpha\beta}^{-\mathbf{Q}}\right]^{-1}\left(\psi_{j\mathbf{k}+\mathbf{q}}^{*}\psi_{b\mathbf{k}+\mathbf{q}+\mathbf{Q}}\middle|\chi_{\beta}^{\mathbf{Q}}\right). \tag*{(45)}

Further details of the density fitting procedure are given in Ref. [15].

The G0W0G_{0}W_{0} self-energy in Eq. (38) contains two two-electron integrals contracted with X amplitudes from Q\mathbf{Q}-dependent RPA-TDA calculations and a denominator consisting of single-particle energies, ϵjk−Q\epsilon_{j\mathbf{k}-\mathbf{Q}} or ϵbk−Q\epsilon_{b\mathbf{k}-\mathbf{Q}}, and RPA-TDA energies, ΩS(Q)\Omega^{S}(\mathbf{Q}). The numerator of the first term on the rhsrhs of Eq. (38) translated into the two- and three-center integrals of Eq. 30 and Eq. 31 and RPA-TDA eigenvectors, Xj′k′b′k′+QSX_{j'k'b'k'+Q}^{S}, becomes,

(ψnkψbk−Q∗|χα−Q)[VαβQ]−1∗(χβ+Q|ψj′k′∗ψb′k′+Q)Xj′k′b′k′+QSXj′′k′′b′′k′′+QS∗(ψj′k′∗ψb′k′+Q|χδ−Q)[Vδγ−Q]−1∗(χγ+Q|ψbk−Qψnk∗)(46)\left(\psi_{n\mathbf{k}}\psi_{b\mathbf{k}-\mathbf{Q}}^{*}\middle|\chi_{\alpha}^{-\mathbf{Q}}\right)\left[V_{\alpha\beta}^{\mathbf{Q}}\right]^{-1*}\left(\chi_{\beta}^{+\mathbf{Q}}\middle|\psi_{j'\mathbf{k}'}^{*}\psi_{b'\mathbf{k}'+\mathbf{Q}}\right)X_{j'k'b'k'+Q}^{S}X_{j''k''b''k''+Q}^{S*}\left(\psi_{j'\mathbf{k}'}^{*}\psi_{b'\mathbf{k}'+\mathbf{Q}}\middle|\chi_{\delta}^{-\mathbf{Q}}\right)\left[V_{\delta\gamma}^{-\mathbf{Q}}\right]^{-1*}\left(\chi_{\gamma}^{+\mathbf{Q}}\middle|\psi_{b\mathbf{k}-\mathbf{Q}}\psi_{n\mathbf{k}}^{*}\right) \tag*{(46)}

While this expression might seem cumbersome, it is relatively straightforward to assemble. First, the RPA eigenvectors are contracted with three-center integrals to form products,

XS,β=(χβ+Q|ψj′k′∗ψb′k′+Q)Xj′k′b′k′+QS(47)X_{S,\beta}=\left(\chi_{\beta}^{+\mathbf{Q}}\middle|\psi_{j'\mathbf{k}'}^{*}\psi_{b'\mathbf{k}'+\mathbf{Q}}\right)X_{j'k'b'k'+Q}^{S} \tag*{(47)}

Second, these products are contracted with,

(ψnkψbk−Q∗|χα−Q)[Vαβ−Q]−1∗(48)\left(\psi_{n\mathbf{k}}\psi_{b\mathbf{k}-\mathbf{Q}}^{*}\middle|\chi_{\alpha}^{-\mathbf{Q}}\right)\left[V_{\alpha\beta}^{-\mathbf{Q}}\right]^{-1*} \tag*{(48)}

to yield the factor on the lhslhs of Eq. (46). For diagonal elements of the self-energy operator, the factor on the rhsrhs is simply the complex conjugate of the lhslhs factor. In practice, parallel calculation of three-center integrals is distributed over auxiliary basis set centers, α\alpha, and RPA-TDA eigenvectors, Xj′k′b′k′+QSX_{j'k'b'k'+\mathbf{Q}}^{S}, are distributed over cores in a block-cyclic format following diagonalization using the ELPA package^{45,46}. The first step is completed by circulating RPA-TDA eigenvectors and multiplying them by the three-center integrals on each core. This is followed by contraction over auxiliary basis function index, α\alpha, to yield the left half of the numerator in Eq. (46). The self-energy matrix

Screened Interaction

The self-energy matrix is assembled using an MPI-reduce call to sum self-energy contributions over auxiliary basis function sites.

The bare Coulomb interaction in this work is represented by the Coulomb matrix in Eq. (43), which is in the auxiliary basis. The auxiliary basis is larger than the wavefunction basis and is of order 100 functions per atom. A matrix representing a unit cell with 100 atoms is therefore of dimension around 10,000. Inverses of the Coulomb matrix arise in representation of wavefunction product densities in Eq. (40); a matrix of this dimension is relatively inexpensive to invert.

Screened interactions in BSE matrix element calculations are approximated using a static, inverse dielectric matrix4^{4}. In this work, a static approximation to the screened interaction is obtained by omitting the energy dependence of the dressed polarizability in Eq. 30 and Eq. 31 and combining resonant and anti-resonant terms into a single term. Products of densities in the polarizability are expressed as combinations of three-center integrals and inverted Coulomb matrices (Eq. (40)). The matrix representation of the static, dressed polarizability is,

i[Παβ−q]−1=4[Vαβq−1]∗(χβ+q∣ψj′k′∗ψb′k′+q)Xj′k′b′k′+qS′Xj′′k′′b′′k′′+qS′′(ψj′k′∗ψb′k′+q∣χδ−q)[Vδγ−q]∗]ΩS(q).(49)i[\Pi_{\alpha\beta}^{-q}]^{-1}=4\frac{[V_{\alpha\beta}^{q-1}]^*(\chi_{\beta}^{+q}|\psi_{j'k'}^*\psi_{b'k'+q})X_{j'k'b'k'+q}^{S'}X_{j''k''b''k''+q}^{S''}(\psi_{j'k'}^*\psi_{b'k'+q}|\chi_{\delta}^{-q})[V_{\delta\gamma}^{-q}]^*]}{\Omega^S(q)}. \tag*{(49)}

where the factor of 4 arises from combining the resonant and anti-resonant terms in the polarizability and summing over spin. The screening part of W0(Q,ϵ)W_0(\mathbf{Q},\epsilon) is obtained by multiplying this polarizability on the left and right by the density products ψak+Q∗(r)ψbk+q+Q(r)\psi_{ak+Q}^*(\mathbf{r})\psi_{bk+q+Q}(\mathbf{r}) and ψj′k′+q∗(r′)ψik(r′)\psi_{j'k'+q}^*(\mathbf{r}')\psi_{ik}(\mathbf{r}') and integrating over the CO auxiliary basis functions which transform Eq. (49) into its coordinate representation to give,

(ψbk+q+Q∗ψak+Q∣χα−q)∗[iΠαβ−q]−1(χβ−q∣ψik∗ψjk+q).\left(\psi_{bk+q+Q}^*\psi_{ak+Q}\left|\chi_{\alpha}^{-q}\right)^*\left[i\Pi_{\alpha\beta}^{-q}\right]^{-1}\left(\chi_{\beta}^{-q}\right|\psi_{ik}^*\psi_{jk+q}\right).

The total screened interaction (Eq. (4)) is a sum of the bare Coulomb interaction plus this screening part with matrix representation,

[Wαβ−q]−1=[Vαβ−q]−1+[iΠαβ−q]−1(50)\left[W_{\alpha\beta}^{-q}\right]^{-1}=\left[V_{\alpha\beta}^{-q}\right]^{-1}+\left[i\Pi_{\alpha\beta}^{-q}\right]^{-1} \tag*{(50)}

Once this has been obtained, a range of screened interaction matrix elements (Table I) can be obtained by multiplying the appropriate three-center integrals on the right and left. For example, diagram (a) in Figure 2 is given by,

(ψak+Q∗(r)ψbk+q+Q(r)|W|ψjk+q∗(r′)ψik(r′))=(ψbk+q+Q∗ψak+Q|χα−q)∗[Wαβ−q]−1(χβ−q|ψik∗ψjk+q).(51)\left(\psi_{ak+Q}^*(\mathbf{r})\psi_{bk+q+Q}(\mathbf{r})\middle|W\middle|\psi_{jk+q}^*(\mathbf{r}')\psi_{ik}(\mathbf{r}')\right)=\left(\psi_{bk+q+Q}^*\psi_{ak+Q}\middle|\chi_{\alpha}^{-q}\right)^*\left[W_{\alpha\beta}^{-q}\right]^{-1}\left(\chi_{\beta}^{-q}\middle|\psi_{ik}^*\psi_{jk+q}\right). \tag*{(51)}

Small QQ limits

The leading terms in the contribution to the self-energy and statically screened electron-hole interaction in the limit as Q→0\mathbf{Q}\to0 in Eq. (38) and Eq. (49) arise from forward scattering of the electron or hole in the two diagrams at the top of Figure 4 and diagrams (a) and (c) in Figure 3. The numerator of the first term in the G0W0G_0W_0 self-energy in Eq. (38) was translated into its density-fitted expression in Eq. (46). This is divergent as Q→0\mathbf{Q}\to0 for the G=0G=0 term in the Ewald potential in Eq. (44). In order to isolate the Q2Q^2 divergence, the first term on the rhs of Eq. (38) is rewritten in its Fourier representation as,

2 4πQ2∣(ψnk|e−iQ⋅r|ψbk−Q∗)∣24πQ2∣Xj′k′b′k′+QS(ψj′k′|eiQ⋅r|ψb′k′+Q)∣2ϵ−ϵbk−Q−ΩS(Q)(52)\frac{2\,4\pi}{Q^2}\left|\left(\psi_{nk}\middle|e^{-i\mathbf{Q}\cdot\mathbf{r}}\middle|\psi_{bk-Q}^*\right)\right|^2\frac{4\pi}{Q^2}\frac{\left|X_{j'k'b'k'+Q}^{S}\left(\psi_{j'k'}\middle|e^{i\mathbf{Q}\cdot\mathbf{r}}\middle|\psi_{b'k'+Q}\right)\right|^2}{\epsilon-\epsilon_{bk-Q}-\Omega^S(\mathbf{Q})} \tag*{(52)}

where repeated indices are summed over and a factor of 2 is included for spin summation in the polarizability. The leading contributions to the first factor come from ψnk=ψbk\psi_{nk}=\psi_{bk}; the matrix element of eiQ⋅re^{i\mathbf{Q}\cdot\mathbf{r}} approaches unity as Q→0\mathbf{Q}\to0. Matrix elements with other wavefunctions, ψnk≠ψbk\psi_{nk}\ne\psi_{bk}, at that order vanish by orthogonality. Leading contributions to the second factor are found at first order in eiQ⋅re^{i\mathbf{Q}\cdot\mathbf{r}}: iQ⋅(ψj′k′∣r′∣ψb′k′+Q)i\mathbf{Q}\cdot(\psi_{j'k'}|\mathbf{r}'|\psi_{b'k'+Q}). This is evaluated using[47],

i⟨0∣p∣S⟩=⟨0∣[r,HRPA]∣S⟩=ΩS(Q)⟨0∣r∣S⟩,(53)i\langle0|\mathbf{p}|S\rangle=\langle0|[\mathbf{r},H^{RPA}]|S\rangle=\Omega^S(\mathbf{Q})\langle0|\mathbf{r}|S\rangle, \tag*{(53)}

where 0 and SS are the ground and SSth RPA excited states, respectively, and HRPAH^{RPA} is the RPA Hamiltonian. It is well

known that the non-local HF exchange operator in the single-particle Fock operator does not commute with the position operator[48] leading to an extra term in the commutator above when the Hamiltonian is the Fock operator. However, the RPA Hamiltonian commutes with the position operator[47], hence the commutator of the exchange and position operators does not appear here. In our notation the contraction of the RPA eigenvector for state SS, Xj′k′b′k′+QSX_{j'k'b'k'+Q}^{S}, with the position matrix element in Eq. (53) becomes,

(ψj′k′∣r′∣ψb′k′+Q)Xj′k′b′k′+QS=i(ψj′k′∣p∣ψb′k′+Q)Xj′k′b′k′+QSΩS(Q)(54)\left(\psi_{j'k'}|\mathbf{r}'|\psi_{b'k'+Q}\right)X_{j'k'b'k'+Q}^{S}=i\frac{\left(\psi_{j'k'}|\mathbf{p}|\psi_{b'k'+Q}\right)X_{j'k'b'k'+Q}^{S}}{\Omega^S(\mathbf{Q})} \tag*{(54)}

The contribution to the self-energy of virtual state ψnk=ψbk−Q\psi_{nk}=\psi_{bk-Q} with energy ϵ=ϵbk−Q\epsilon=\epsilon_{bk-Q} in the limit of small Q\mathbf{Q} is

−1NQNk′16π2∣Q^⋅(ψj′k′∣p∣ψb′k′+Q∗)Xj′k′b′k′+QS∣2Q2ΩS(Q)3.(55)-\frac{1}{N_QN_{k'}}\frac{16\pi^2\left|\hat{\mathbf{Q}}\cdot\left(\psi_{j'k'}|\mathbf{p}|\psi_{b'k'+Q}^*\right)X_{j'k'b'k'+Q}^{S}\right|^2}{Q^2\Omega^S(\mathbf{Q})^3}. \tag*{(55)}

where Q^\hat{\mathbf{Q}} is a unit vector parallel to Q\mathbf{Q} and NQN_Q and Nk′N_{k'} are the number of Q\mathbf{Q} and k′\mathbf{k}' wave vectors in the Brillouin zone. There is a similar, positive term corresponding to the second term on the right in Eq. (38) for valence states. The Q2Q^2 divergence is treated using the method of Gygi and Baldereschi[49] for handling this divergence in exchange energies of cubic solids; otherwise the extension of this method to general lattices by Carrier, Rohra and Görling[50] is used. The form given in Eq. (55) is used for the non-analytic, divergent part as Q→0\mathbf{Q}\to0 and the non-divergent, density-fitted form in Eq. (38) is used for finite Q. The main uncertainty in this approach to the small Q limit is the equivalence of momentum and position matrix elements implied by Eq. (53) in an incomplete basis. Similar divergences in the screened electron-hole attraction (Eq. (51)) are treated in the same way.

Computational Methods

Basis sets in this work were all electron G0 basis sets adapted from the def2-TZVP basis sets of Weigend and Ahlrichs[51]. Gaussian orbital basis sets are commonly generated using the variational principle, i.e. minimization of total energy of the system in its ground state. However, response properties require conduction bands to be well represented as well as occupied states. The completeness of the basis sets used in the kinetic energy range to 100 eV were tested by diagonalizing the kinetic energy operator and generating free-electron band structures for the materials studied. These are shown in the supporting information. This was done for the def2-TZVP basis sets and for augmented basis sets with modified exponents and additional high angular momentum basis functions. Diffuse basis functions (orbital exponent α≤0.1\alpha\le0.1) cause linear dependence problems and are not essential to construct free-electron band structures up to 100 eV.

For MgO, the def2-TZVP Mg(O) basis sets have 4s3p2d(5s3p2d1f) functions (with diffuse functions excluded). Deviations from the free-electron band structure for MgO are of order several eV for kinetic energies as low as 20 eV (Fig. S1). However, augmenting this to 6s4p2d1f1g for both Mg and O results in a free-electron band structure which shows deviations from the free electron band structure around 60 eV, but generally is much improved over the def2-TZVP basis for energies of 20 to 100 eV.

For diamond, the def2-TZVP C 11s5p2d1f(5s3p2d1f) basis set was replaced by the 11s6p2d2f(5s4p2d2f) basis given in SI Table S1. The def2-TZVP basis set core functions were retained and diffuse s and p functions were replaced by less diffuse functions. The free electron band structure is reproduced well by this basis up to 120 eV (SI Fig. S1).

For Si, the 13s9p1d(6s5p1d) basis from Heyd and coworkers[52] was replaced by the 13s9p3d1f(6s5p3d1f) basis given in SI Table S2. Changes to the original basis were to add d and f functions.

For anatase and rutile TiO2_2 the def2-TZVP 17s11p7d1f(6s4p4d1f) basis was replaced by the 16s11p7d2f(5s4p3d2f) basis given in SI tbl:S3. Diffuse s, p and d functions were replaced by less diffuse functions. The free electron band structure for the TiO2_2 phases is reproduced well up to about 60 eV. The O basis for the TiO2_2 phases differed from the O basis for MgO in the number and exponents of d and f functions.

Auxiliary basis sets used for density fitting were the def2-TZVP-RIFIT sets[53]. Basis functions of h or higher angular momentum were omitted. Modified wavefunction basis sets are listed in Supporting Information Tables S1 to S5.

Results

The materials chosen for study here are those used in a previous study of the performance of TDHF in wide gap materials with a scaled electron-hole attraction[15], namely diamond C, MgO and anatase and rutile TiO2_2. Here we also consider Si as an example of a narrow, indirect gap semiconductor with a valence band of sp3sp^3 character. HF theory overestimates band widths of materials where the bottom and top of the band are of different orbital character. For example, the bottom and top of the valence band in diamond and Si are of ss and pp character, respectively, and HF theory overestimates the band width in diamond C and Si by about 25%. Below we show that G0W0G_0W_0@HF corrects this and predicts band widths in agreement with experiment. In materials such as the oxides studied here the O 2s and 2p bands are distinct. In this case the HF O 2p bands are nearly indistinguishable from the G0W0G_0W_0@HF bands once both band structures have their VBM aligned. The O 2s levels shift upwards relative to the valence pp bands on going from HF to G0W0G_0W_0@HF. There is a large renormalisation of the HF band gap in all these materials. G0W0G_0W_0@HF band gaps are larger than G0W0G_0W_0@LDA gaps and, typically, experimental gaps lie between the theoretical predictions.

In a conventional PW G0W0G_0W_0 calculation wave functions are expanded in PW basis sets. PW cutoff energies of 230 and 680 eV for wave functions in diamond and Si were used in early G0W0G_0W_0@PW/LDA calculations[3] and found to be converged. RPA dielectric matrices for diamond and Si with dimension 220 x 220 and 140 x 140 G vectors were used in the same work. Shell closings around 220 G vectors in diamond and 140 vectors in Si correspond to free electron energies of 32 and 10 eV, respectively. Thus RPA excitations which are likely to be important in screening are relatively low in energy. Tests of virtual state cutoff in RPA screening in molecular tetracene (C. H. Patterson unpublished) showed that a virtual state cutoff energy of 60 eV results in a difference in low energy excited states of less than 10 meV in a cc-pVTZ basis, compared to inclusion of all virtual states, and that convergence is essentially reached by including virtual states up to 50 eV. GO basis sets used in this work in GoWoG_oW_o@HF calculations for diamond reproduce the free electron dispersion relation accurately to about 120 eV, when electron-electron and electron-nuclear interactions are turned off (SI Fig. S3). GO in which wave functions are expanded have radial decay exponents which correspond to electron kinetic energies well above 120 eV. Similarly, the PW cutoff energies used for wave functions in diamond or Si are well above this energy. However, in practice in both PW and GO calculations cutoff energies for the screening response do not need to exceed 100 eV.

Table II shows the number of occupied and virtual states included in the GoWoG_oW_o@HF (Eq. (38)) and second-order (Eq. (37)) self-energies. For diamond, a virtual state cutoff energy of 39 eV is reached with just 10 conduction bands in the RPA calculation used in the self-energy and screened interaction (Eq. (22)). A further 24 conduction bands used in the second-order self-energy results in a cutoff energy of 110 eV. For Si, 14 and 32 virtual states in each self-energy result in lower cutoffs of 34 eV in the RPA calculation and 58 eV in the second order self-energy, but well above the cutoff energy of 10 eV implied by 140 G vectors in the PW RPA matrix. For MgO, 14 and 24 virtual states in the respective self-energies resulted in cutoffs of 56 and 107 eV.

For the two TiO2_2 phases, the cutoff in the RPA calculation is 30 eV for an RPA matrix size of 51840 with 40 (rutile) and 30 (anatase) unique Q points at which the RPA matrix must be diagonalized. To achieve a cutoff of 50 eV in the RPA screening calculations would require 44 virtual states for rutile TiO2_2 and 48 virtual states for anatase TiO2_2 and RPA matrix dimensions of 114048 and 124416, which would need to be diagonalized at 40 and 30 unique Q points for a 6×\times6×\times6 MP mesh. This is beyond our computational resources.

GW Band Structure

Diamond

The band structure of diamond has been investigated by photoemission[54, 55, 56], optical absorption[57] and electron energy loss spectroscopy[58]. The GoWoG_oW_o band structure of diamond has been calculated starting from LDA[32, 56, 59, 60, 61] in GO[32], LAPW[61] and PW[59, 60] basis sets. Valence and conduction band energies for diamond at high symmetry points of the BZ in these calculations are compared to our HF starting point calculations in a GO basis in Table III. The valence band width of diamond has been reported to be between 21 eV[54] and 24.2 eV[62], with most experimental estimates lying between 23 and 24 eV[56].

The GO HF band structure for diamond in Table III and Figure 6 shows a valence band width of 29.07 eV, in good agreement with previously reported HF values of 28.67 eV[64] or 29.43 eV[65]. When GoWoG_oW_o self-energy corrections are made, the valence band width reduces to 23.94 eV, in good agreement with the range of measured values from photoemission[55, 56, 66]. Direct and indirect HF band gaps, 14.71 and 12.62 eV, are in good agreement with previously reported values of 14.7 and 12.6 eV[67]. The GoWoG_oW_o@HF direct gap is 7.95 eV, 0.6 eV greater than the experimental value of Roberts and Walker[57], namely 7.3 eV. A previous GO GoWoG_oW_o@LDA calculation[32] predicted a direct gap about 0.3 eV greater than the experimental value. GoWoG_oW_o@LDA valence band widths are 1-2 eV less than GoWoG_oW_o@HF values. The GoWoG_oW_o@HF indirect gap in diamond is 5.75 eV, 0.3 eV greater than the experimental value of 5.48 eV[63].

Band structures of Si and C in the diamond structure

Figure 6. Band structures of Si and C in the diamond structure from HF and GoWoG_oW_o calculations. HF band structures are shown as solid blue lines. GoWoG_oW_o self-energies were obtained in an 8x8x8 MP mesh at points shown.

Edmonds and coworkers[56] reported a constant initial state (CIS) photoemission study of the (100) surface of H-terminated diamond and gave precise values of valence band energies at the X point. They found the lower and upper VB at -14.01 and -6.97 eV (Table III), which compares to -14.68 and -7.11 eV from GoWoG_oW_o@HF and -13.42 and -6.74 eV from GoWoG_oW_o@LDA with a PW basis. Table III also compares positions of higher CB at the Γ and X points and these are in reasonable agreement with GoWoG_oW_o@HF and GoWoG_oW_o@LDA calculations, although it should be noted that experimental error bars are several eV wide[56].

Silicon

HF and GoWoG_oW_o@HF band structures of Si are shown in Figure 6 and energies of valence and conduction band states at high symmetry points are compared to experimental values and previous GoWoG_oW_o@LDA calculations in Table IV. The HF valence bandwidth of Si in this work is 17.00 eV, considerably larger than the experimental value of 12.5 ±\pm 0.6 eV[63]. The bottom of the VB has 3s character and the top has 3p character; the GoWoG_oW_o@HF self-energy for the bottom of the

StateHFa^{a}G0W0G_{0}W_{0}@HFa^{a}G0W0G_{0}W_{0}@LDA b^{b}G0W0G_{0}W_{0}@LDA c^{c}G0W0G_{0}W_{0}@LDA d^{d}G0W0G_{0}W_{0}@LDA e^{e}G0W0G_{0}W_{0}@LDA f^{f}Exptf^{f}Expt
Γc\Gamma_{c}38.7230.2928.5031.1
Γc\Gamma_{c}37.3930.0028.1028.6
Γc\Gamma_{c}29.6721.22
Γ2c\Gamma_{2c}24.7116.1914.5414.4115.3±0.5g\pm0.5^{g}
Γ15c\Gamma_{15c}14.687.957.637.387.447.437.3h^{h}
Γ25v\Gamma_{25v}0.000.000.000.000.000.000.00.00
Γ1v\Gamma_{1v}-29.08-23.94-22.88-22.09-22.0-24.0 ±0.5\pm0.5-23.0±0.2i\pm0.2^{i}
L2c′L'_{2c}26.9718.5818.1417.4720±1.5j\pm1.5^{j}
L15cL_{15c}18.3011.1510.2310.4410.38
Lv′L'_{v}-3.78-3.13-2.98-3.01-2.94
L2v′L'_{2v}-17.94-15.15-14.27-13.97-12.8±0.3j\pm0.3^{j}
L3vL_{3v}-21.89-18.02-16.95-16.58-15.2±0.3j\pm0.3^{j}
XcX_{c}50.4842.7539.9057.2
XcX_{c}44.0936.0532.7736.2
XcX_{c}36.1326.9524.4623.9
XcX_{c}29.8320.8519.5023.9
XcX_{c}13.316.306.306.236.26
X4v′X'_{4v}-8.38-7.11-6.69-6.72-6.58-6.74-6.97 ±0.075\pm0.075
X1v′X'_{1v}-17.78-14.68-13.80-13.42-14.01 ±0.075\pm0.075
KcK_{c}12.297.497.22
K1v′K'_{1v}-9.29-6.05-5.74
WcW_{c}20.5013.0212.43
WvW_{v}-10.84-9.36-8.85
EgE_{g}12.635.755.675.755.635.48j^{j}

Table III. Valence and conduction band energies for diamond at high symmetry points in the BZ from HF, G0W0G_{0}W_{0}@HF and G0W0G_{0}W_{0}@LDA and photoemission experiments. All values are relative to the VBM at the Γ\Gamma point. All G0W0G_{0}W_{0} energies, with the exception of this work, used a LDA starting point and GO, LAPW or PW basis sets. This work used a HF starting point and a GO basis set.

band is larger than the top, resulting in a narrowing of the valence bandwidth to 13.03 eV. The predicted VB width lies within the experimental uncertainty range and compares to a G0W0G_{0}W_{0}@LDA value of 12.04 eV{^4} (Table IV). Duchemin and Gygi{^74} reported the HF band structure for Si using a PW basis. They obtained conduction band energies of 9.12, 7.92 and 6.62 eV at the Γ\Gamma, LL and XX points of the Brillouin zone, and an indirect gap of 6.43 eV which compares to 8.41, 7.81 and 6.37 eV in this work and 6.21 eV for the indirect gap (Table IV) in this work using an all-electron GO basis.

StateHFGoWoG_oW_o@HFGoWoG_oW_o@LDAa^{a}GoWoG_oW_o@LDAb^{b}Exptc^{c}
Γ2c′\Gamma'_{2c}10.635.194.274.084.2
Γ15c\Gamma_{15c}8.413.533.303.353.35d^{d}
Γ25v′\Gamma'_{25v}0.000.000.000.00
Γ1v\Gamma_{1v}-17.00-13.0312.0412.5±0.612.5\pm0.6
L2c′L'_{2c}9.664.544.114.244.15±0.14.15\pm0.1
L15cL_{15c}7.812.842.302.272.1e,2.4±0.15f2.1^{e},2.4\pm0.15^{f}
L3v′L'_{3v}-1.94-1.51-1.19-1.271.2±0.2,1.5g1.2\pm0.2,1.5^{g}
L2v′L'_{2v}-9.67-7.73-7.186.7±0.26.7\pm0.2
L1vL_{1v}-14.01-10.87-9.799.3±0.49.3\pm0.4
X1cX_{1c}6.371.471.44
X4v′X'_{4v}-4.29-3.48-2.993.3±0.2h3.3\pm0.2^{h}
X2v′X'_{2v}-11.38-8.96
K1cK_{1c}6.591.59
K4v′K'_{4v}-4.03-3.36
W1cW_{1c}10.485.13
W4v′W'_{4v}-5.23-4.43
EgE_g6.211.381.241.291.17

Table IV. HF, GoWoG_oW_o@HF and GoWoG_oW_o@LDA conduction and valence band energies for Si at high symmetry points relative to the VBM at the Γ\Gamma point.

The G0W0G_{0}W_{0}@DFT band structure of Si is well established{^4,33,68}. These data and photoemission and optical transition measurements allow a detailed comparison of the G0W0G_{0}W_{0}@HF calculations. A comparison of critical point transition energies from theory and experiment is given in Table V. G0W0G_{0}W_{0}@LDA calculations agree with ellipsometry values to within 0.1 eV while G0W0G_{0}W_{0}@HF values overestimate transition energies by between 0.2 eV (E0E_{0}) and 0.9 eV (E1E_{1}).

TransitionGoWoG_oW_o@HFGoWoG_oW_o@LDAa^{a}GoWoG_oW_o@LDAb^{b}Exptc^{c}
Γ25v→Γ15c\Gamma_{25v}\rightarrow\Gamma_{15c} E0E_03.533.303.353.35
L3v→L1cL_{3v}\rightarrow L_{1c} E1E_14.353.543.493.46
L3v→L3cL_{3v}\rightarrow L_{3c} E1′E'_16.055.305.515.4
X4v→X1cX_{4v}\rightarrow X_{1c} E2E_24.954.43-4.32

Table V. Critical point transition energies for Si from GoWoG_oW_o@HF, GoWoG_oW_o@LDA calculations and ellipsometry.

Rock Salt MgO

The G0W0G_{0}W_{0}@HF band structure of MgO calculated is shown in Figure 7 and conduction and valence band energies are compared to results of a previous G0W0G_{0}W_{0}@LDA calculation and the experimental gap in Table VI. Rö ssler and Walker{^75} reported an exciton binding energy of 0.08 eV with an exciton dissociation limit of 7.77 eV from optical reflectance measurements and Walker and coworkers reported exciton fine structure between 7.67 and 7.83 eV{^76}. Schönberger and Aryasetiawan{^77} reported a G0W0G_{0}W_{0}@LDA band gap of 7.7 eV using an LMTO-ASA method. More recently, band gaps of 7.90 eV{^78} (PW/LDA), 7.63 eV{^61} (FLAPW/LDA), 7.32 eV{^79} (numerical atomic orbitals/PBE), 8.53 eV{^80} (FLAPW/HSE06) and 9.97 eV{^81} quasiparticle self-consistent GWGW (QSGW) have been reported. The G0W0G_{0}W_{0}@HF gap value from this work is 9.13 eV. Antonius and coworkers{^82} reported zero point renormalization reduction in the band gaps of diamond and MgO by over 0.3 eV in diamond and nearly 0.3 eV in MgO.

Figure 7. Band structures of rock salt MgO from HF and GoWoG_oW_o calculations. HF band structures are shown as solid blue lines. GoWoG_oW_o self-energies were obtained in an 8x8x8 MP mesh at points shown.

StateHFGoWoG_oW_o@HFa^{a}GoWoG_oW_o@LDAb^{b}Exptc^{c}
Γc\Gamma_c15.979.137.327.77(8.1)d^{d}
Γv\Gamma_v0.000.000.00
LcL_c18.7811.7410.46
LvL_v-0.95-0.98-0.77
XcX_c19.8312.7511.43
XvX_v-1.89-1.98-1.56

Table VI. HF, GoWoG_oW_o@HF and GoWoG_oW_o@LDA lowest conduction band and highest valence band energies for MgO relative to the VBM at the Γ\Gamma point.

a^{a}[68]

b^{b}[4]

c^{c}[63] except where noted

d^{d}[69]

e^{e}[70]

f^{f}[71]

g^{g}[72]

h^{h}[73]

a^{a}[68]

b^{b}[4]

c^{c}[69]

First principles calculations omit screening by polar phonons, which is, of course, affects experiental measurements.

The value of 9.13 eV for the band gap of MgO from GoWoG_oW_o@HF is therefore about 1.1 eV above the experimental gap of 7.77 eV, taking these effects into account. It is 0.6 eV higher than the FLAPW/HSE06 value80^{80} mentioned above. The HF and GoWoG_oW_o@HF band structures for MgO in Figure 7 show that the O 2pp valence band widths are not strongly affected by self energy corrections, while the O 2ss band shifts upwards relative to the VBM by 4.0 eV.

Anatase and Rutile TiO2_2

Angle resolved photoelectron spectroscopy (ARPES) measurements on anatase TiO2_2 at 20 K show that the VBM

a^{a}This work GO

b^{b}[61]

c^{c}[76]

d^{d}Estimate without electron-phonon interaction [82]

a^{a}This work GO

b^{b}[84]

b^{b}[83]

occurs at the XX point of the BZ and that it lies approximately 0.5 eV above the highest band at the Γ\Gamma point83^{83}. The same work found a QP band gap at Γ\Gamma of 3.97 eV83^{83} using nn-type samples with O vacancies. HF and GoWoG_oW_o@HF band energies for anatase are compared to GoWoG_oW_o@LDA, GoWoG_oW_o@PBE values and ARPES data in Table VII and the band structures are shown in Figure 8.

StateHFGoWoG_oW_o@HFa^{a}GoWoG_oW_o@LDAb^{b}GoWoG_oW_o@PBEc^{c}Exptc^{c}
Γc\Gamma_c12.804.833.563.463.47
Γv\Gamma_v-0.27-0.29-0.58-0.46-0.50
XcX_c14.306.024.89--
XvX_v0.000.00-0.060.000.00

Table VII. HF, GoWoG_oW_o@HF, GoWoG_oW_o@LDA and GoWoG_oW_o@PBE lowest conduction band and highest valence band energies for anatase TiO2_2 relative to the VBM at the XX point.

Figure 8. Band structures of anatase TiO₂ from HF and GoWoG_oW_o@HF calculations. The HF band structure is shown as solid blue lines. GoWoG_oW_o@HF self-energies were obtained in a 6x6x6 MP mesh at points shown.

Kang and Hybertsen84^{84} found the VBM close to the XX point along the Γ−X\Gamma-X direction, the CBM at Γ\Gamma and an indirect gap of 3.56 eV and direct gaps at Γ\Gamma and XX of 4.14 and 4.95 eV, respectively. Baldini and coworkers found the VBM at XX and the CBM at Γ83\Gamma^{83} with an indirect gap of 3.46 eV, in good agreement with ARPES data in the same work.

Low temperature, high resolution absorption measurements on rutile TiO₂ show the presence of a dark exciton at 3.03 eV, which has been used to infer a direct QP gap at Γ for rutile[85]. Kang and Hybertsen84 found an indirect gap of 3.34 eV in rutile, with the VBM at the R point. However, the highest VB at Γ is just 40 meV below that at the R point. More recent calculations by Baldini and coworkers86 found a direct gap of 3.34 eV. Similar values for the rutile QP gap in the range 3.30[87] to 3.59 eV[88] have also been reported. The direct gap at Γ from GoWoG_oW_o@HF is 3.85 eV and is less than the indirect gap values of 3.88 and 3.94 eV at the M and R points, respectively. The GoWoG_oW_o@HF band structure is shown in Figure 9 and band edge energies are compared to GoWoG_oW_o@LDA values in Table VIII.

Figure 9. Band structures of rutile TiO₂ from HF and GoWoG_oW_o calculations. HF band structures are shown as solid blue lines. GoWoG_oW_o self-energies were obtained in a 6x6x6 MP mesh at points shown.

StateHFGoWoG_oW_o@HFᵃGoWoG_oW_o@LDAᵇ
Γc\Gamma_c12.413.853.38
Γv\Gamma_v0.000.000.00
RcR_c12.544.003.34
RvR_v-1.38-1.49-1.12
McM_c12.343.893.40
MvM_v-1.16-1.48-1.15

Table VIII. HF, GoWoG_oW_o@HF and GoWoG_oW_o@PW/LDA lowest conduction band and highest valence band energies for rutile TiO₂ relative to the VBM at the Γ point. ᵃ This work GO. ᵇ Ref. [84]

The direct gap at Γ from GoWoG_oW_o@HF for anatase is 5.12 eV, larger than the experimental value of 3.97 eV[83]. Thus the gaps in anatase and rutile TiO₂ exceed experimental or GoWoG_oW_o@LDA values by 0.5 eV or more. Band ranges used in the GoWoG_oW_o@HF self-energy calculation for both anatase and rutile TiO₂ (Table II) were 12 valence bands and 20 conduction bands and a cutoff of 30 eV. This band range requires diagonalization of RPA matrices of dimension 51840 at the unique Q points in the 6x6x6 MP mesh (30 for anatase and 40 for rutile). This cutoff energy is much lower than the smaller diamond, Si and MgO unit cells owing to the larger numbers of bands at low energies. A higher cutoff may therefore increase the self-energy magnitudes and reduce the valence-conduction band gap.

MaterialOccupiedVirtualsa^aVirtualsb^bMP2-netRPAc^c
C410 (39)24 (110)10x10x1040000
Si414 (34)32 (58)10x10x1056000
MgO414 (56)24 (107)10x10x1056000
Anatase TiO2_21220 (30)2 (33)6x6x651840
Rutile TiO2_21220 (30)2 (33)6x6x651840

Table II. Occupied and virtual state ranges and energy cutoff in eV used in G0W0G_0W_0@HF and BSE@HF calculations and MP meshes and dimensions for RPA screening calculations.

Dielectric functions from GoWoG_oW_o@HF/BSE

We present dielectric functions from GoWoG_oW_o@HF/BSE-TDA calculations in this Section. In the previous Section, nxnxn MP meshes with n=8n = 8 or n=6n = 6 were used for self-energy corrections to band structures. Points in the BZ in the Exciton code must be integers modulo nn. BZ high symmetry points for the lattices used contain reciprocal lattice vector fractions such as G/4 or G/3. Consequently, to construct straight lines

in reciprocal space between high symmetry points (apart from the Γ point), MP mesh dimensions for cubic diamond, Si or MgO must be multiples of 4 and those for anatase and rutile TiO₂ must be multiples of 3. For BSE-TDA calculations of dielectric functions this restriction is lifted as all k-points are used.

BSE-TDA calculations reported in this Section used 10x10x10 or 11x11x11 MP meshes centered on the Γ point. Symmetric meshes permit symmetry to be used to reduce the number of Q points at which the RPA-TDA Hamiltonian (Eq. (22)) must be diagonalized in order to construct the dressed polarizability on which the screened interaction depends. Symmetry equivalence of k-points means limited sampling of the BZ, however. Dielectric functions presented demonstrate that the method produces dielectric spectra which agree with experiment reasonably well. Methods of improving convergence of the approach are discussed later in Section VII.

A study of oscillator strengths in molecules[89], which compared oscillator strengths from a range of DFT exchange-correlation functionals using Kohn-Sham orbitals and RPA and CIS methods using HF orbitals to EOM-CCSD, found that oscillator strengths from RPA and CIS methods typically exceeded those from the reference method while those from DFT underestimate them. The difference depended on the bonding type - for alkenes, CIS overestimates them by a factor of 1.45 on average, while LDA underestimates by a factor of 0.86. For azabenzenes, CIS overestimates them by a factor of 2.22 on average, while LDA is almost in agreement with EOM-CCSD (0.99). The CIS method is BSE-TDA with no screening of the electron-hole interaction and is the method most similar to BSE-TDA of the methods used in Ref. [89]. We find similar overestimates of oscillator strength in the crystal systems studied here using HF wave functions. For diamond and Si there is reasonable agreement with experimental oscillator strengths, while for the oxides chosen for study, there is significant overestimation of oscillator strengths.

Diamond

The dielectric function of diamond from GoWoG_{o}W_{o}@HF/BSE is compared to experimental data in Figure 10(a). It was obtained by averaging spectra with n=10n = 10 and n=11n = 11 to improve BZ sampling and by smoothing with a spline fit. The self-energy calculation used four valence and 10 conduction bands in the GoWoG_{o}W_{o}@HF self-energy (Eq. (38)) plus a further 20 conduction bands in the second order self-energy (Eq. (37)). The BSE-TDA calculation used the same bands as the GoWoG_{o}W_{o}@HF self-energy. The experimental lineshape and intensity is recovered well except for overemphasis of a peak around 15 eV. Peak positions in the BSE-TDA and experiment are in good agreement (11.8 eV in both). Compared to previous calculations of the dielectric function of diamond in a GO basis11, which used the HF band structure and a TDHF Hamiltonian with a uniformly reduced electron-hole attraction, the width and position of the absorption peak from GoWoG_{o}W_{o}@HF/BSE is in much better agreement with experiment then before.

Figure 10. Dielectric functions of diamond (left panel) and Si (right panel) from BSE calculations which used GoWoG_{o}W_{o}@HF self-energies as input to the BSE-TDA calculation. Experimental data redrawn from Ref. [90] is shown in solid black lines.

Silicon

The dielectric function of Si from GoWoG_{o}W_{o}@HF/BSE is compared to experimental data in Figure 10(b). It was obtained from a single calculation with n=10n = 10. The self-energy calculation used four valence and 14 conduction bands in the GoWoG_{o}W_{o}@HF self-energy (Eq. (38)) plus a further 32 conduction bands in the second order self-energy (Eq. (37)). The BSE-TDA calculation used the same bands as the GoWoG_{o}W_{o}@HF self-energy. Peak positions in the dielectric function are in good agreement with experimental values (3.45, 4.24 and 5.43 eV versus 3.44, 4.29 and 5.30 eV in experiment), however the intensity of the first peak (which is enhanced by electron-hole attraction) is about a factor of two too low and the weak shoulder at 5.30 eV in experiment is a strong peak in the calculation.

Rock Salt MgO

The BSE-TDA dielectric function of MgO is shown in Figure 11 along with the experimental dielectric function from Ref. [75]. A single n=10n = 10 calculation with four valence bands, 17 conduction bands in the GoWoG_{o}W_{o}@HF self-energy (Eq. (38)) plus a further 32 conduction bands in the second order self-energy (Eq. (37)).

Figure 11. Dielectric functions of MgO from BSE-TDA calculations which used GoWoG_{o}W_{o}@HF self-energies as input to the BSE-TDA Hamiltonian. Experimental data redrawn from Ref. [75] is shown in solid black lines.

Peak positions in the BSE-TDA dielectric function (8.09, 10.50, 12.64 and 16.93 eV) are in reasonable agreement with experimental values (7.6, 10.8, 13.3 and 16.9 eV). Unlike diamond and silicon, peak intensities from BSE-TDA are greater than experiment. Similar overestimation in the dielectric function intensities are also found in anatase and rutile TiO2_2. The tetrahedral semiconductors considered here have delocalized, nearly free electron wave functions while the polar, ionic oxides have more localized occupied HF wave functions.

The experimental optical gap of MgO is 7.83 eV12. The experimental exciton binding energy is 0.08 eV13. FLAPW/HSE06 calculations14 predict an exciton binding energy of 0.44 eV and 0.60 eV using PBEsol and HSE06 DFT starting points in the GoWoG_{o}W_{o}/BSE calculations. Here we find an exciton binding energy of 1 eV. Recent calculations of the renormalization of exciton energies in MgO15 by polar phonons report a redshift of around 0.05 eV for MgO, which is much less than the differences between experimental and first principles values.

Anatase and Rutile TiO2_2

The dielectric functions of anatase and rutile TiO2_2 from G0W0G_0W_0@HF/BSE are compared to experimental data in Figure 12. They were obtained using an MP mesh with n=6n=6. Self-energy and BSE screening calculations for both phases used 12 valence and 20 conduction bands (Table II). Hosaka and coworkers obtained the dielectric function of anatase TiO2_2 from optical reflection data[92] and Baldini and coworkers[83] reported the dielectric function of anatase TiO2_2 from spectroscopic ellipsometry (SE) and compared it to results of G0W0G_0W_0@PBE/BSE-TDA calculations (Table IX).

Figure 12. Dielectric functions of anatase (upper panel) and rutile (lower panel) TiO2_2 from G0W0G_0W_0@HF/BSE-TDA calculations self-energies as input to the BSE-TDA calculation. BSE spectra were scaled by a factor of 0.6. Experimental data for anatase TiO2_2 are redrawn from Refs. [92] (Expt I, optical reflectance, black line) and [83] (Expt II, SE, blue line). Experimental data for rutile TiO2_2 are redrawn from Refs. [93] (Expt III, optical reflectance, black line) and [86] (Expt IV, SE, blue line).

FieldOpt. Refl.a^{a}Spec. Ellips.b^{b}BSE-TDAb^{b}BSE-TDAc^{c}
Anatase
E∥\parallela3.903.793.764.01
4.674.614.814.78
E∥\parallelc4.274.134.284.56
5.014.964.94
8.088.24
Rutile
E∥\parallela4.053.933.993.71
4.514.574.32
5.435.425.375.12
E∥\parallelc4.204.154.243.80

Table IX. Peak positions in dielectric functions of anatase and rutile TiO2_2 from optical reflectivity, spectroscopic ellipsometry and G0W0G_0W_0@HF/BSE-TDA and G0W0G_0W_0@PBE/BSE-TDA calculations.

G0W0G_0W_0@HF/BSE-TDA calculations reproduce the experimental peak positions for anatase TiO2_2 mostly within 0.1 to 0.2 eV. For E∥\parallela, peaks occur at 4.01 and 4.61 eV (compared to 3.90/3.79 eV and 4.67/4.61 eV from optical reflectance/SE data). For E∥\parallelc they occur at 4.56 and 4.94 eV (compared to 4.27/4.13 eV and 5.01/4.96 eV from experiment). As noted in the introduction to § 8 spectral intensities predicted by G0W0G_0W_0@HF/BSE for oxides exceed experimental intensities. To facilitate comparison of predicted and measured spectra, intensities have been scaled by a factor of 0.6. G0W0G_0W_0@PBE/BSE-TDA calculations in Ref. [83] predict peaks at 3.76 and 4.81 eV for E∥\parallela and at 4.28 eV for E∥\parallelc, which are close to peak positions in SE experiments at 3.79, 4.61 and 4.13 eV (Table IX).

Tiwald and Schubert obtained the dielectric function of rutile TiO2_2 from optical reflectivity[93] and Baldini and coworkers obtained it from spectroscopic ellipsometry[86]. For E∥\parallela, SE shows a peak at 3.93 eV, a shoulder at 4.51 eV and a further peak at 5.42 eV (Table IX). Optical reflectivity shows a peak at 4.05 eV and a second peak at 5.43 eV. G0W0G_0W_0@PBE/BSE-TDA calculations by Baldini and coworkers[86] find peaks to match each of these features at 3.99, 4.57 and 5.37 eV.

G0W0G_0W_0@HF/BSE-TDA calculations reproduce the experimental peaks but are shifted to lower energy by

0.2 to 0.4 eV. For E∥\parallela peaks occur at 3.71, 4.32 eV and 5.12 eV (0.22, 0.19 and 0.30 eV below the SE values) and for E∥\parallelc the main peak occurs at 3.80 eV (0.35 eV below the SE value). The cause of this shift is believed to be the small QQ approximation to the screened electron-hole attraction described in Section IIF. Bare electron-hole attraction terms in the small QQ limit for anatase and rutile TiO2_2 with an n=6n=6 MP mesh are 1.56 and 0.94 eV, respectively. When screened these become 1.03 and 0.65 eV, corresponding to long wavelength dielectric functions of 1.45 and 1.50, well below a value of 5 which might be expected in a moderate gap energy oxide. If a value of 5 were used instead, the screened small QQ limit electron-hole attraction terms would become 0.31 eV and 0.19 eV. The larger value for rutile TiO2_2 may explain the tendency for peaks to be shifted to lower energy in that case. This point is discussed further in Section VII.

Discussion

Most GW/BSE calculations for periodic, solid materials are performed using codes which employ PW basis sets. These codes have been referenced throughout this work and our GO basis results have been compared to them. GO basis GW and BSE we have presented a GW/BSE method.

Early G0W0G_0W_0@LDA calculations by Rohlfing and coworkers[32] used a model dielectric function to treat the band structures of five tetrahedral semiconductors. Later these authors adopted the dielectric band structure plasmon pole model[94] for the dielectric function in GoWo@LDAG_{o}W_{o}@\mathrm{LDA} calculations applied to bulk Si and the Si(001) surface[33]. Rohlfing and Louie[35] then applied this GoWo@LDAG_{o}W_{o}@\mathrm{LDA} method to GoWo@LDAG_{o}W_{o}@\mathrm{LDA}/BSE-TDA calculations applied to inert gas solids, tetrahedral semiconductors and wide gap insulators. Galamić-Mulaomerović and Patterson used the dielectric band structure model for the dielectric function to calculate the band structures[95] and excitonic absorption spectra[96] of solid Ne and Ar. In contrast to the current work, that work used the bare coulomb potential in its plane wave representation and used wave functions and single-particle band energies from the Crystal code[97] Here it is represented in the auxiliary CO basis (Eq. (41)).

Recently we applied the density-fitting approach used here to perform TDHF calculations with a scaled (rather than screened) electron-hole interaction[15] applied to the same materials as in this work (except Si). García-Blázquez and Palacio[39] reported a GO density-fitting approach which used wave functions and single-particle band energies from the Crystal code[98] and using a ’scissors shift’ to align hybrid DFT band gaps with QP gaps.

The current work establishes a density fitting, GO method for GoWo@HFG_{o}W_{o}@\mathrm{HF}/BSE-TDA calculations without invoking a plasmon pole approximation. To generate the self-energy and screened interaction matrices, diagonalization of a matrix which is of the same size as the final BSE-TDA matrix, must be diagonalized at each unique QQ point in the BZ. As the system size grows the number of valence and especially conduction bands must be increased so that the method is limited by the size of matrices that can be diagonalized in a reasonable amount of time. However, increased availability of GPU hardware combined with the ELPA code[99] will increase significantly the sizes of matrices which can be diagonalized at a reasonable computational cost.

The symmetric MP mesh that is used in the current work results in limited sampling of the BZ, especially affecting optical spectra. Meshes of the sizes that have been used for BSE-TDA calculations on MgO in this work result in well converged spectra[80] when an off-centered mesh is used. Employing symmetry in reciprocal space reduces the amount of time required for Ewald potential integral generation (Eq. (44)). The Ewald potential has absolutely convergent real and reciprocal space terms. The wave vector dependence of the real space term is a simple multiplication of an integral by a phase factor containing qq. On the other hand the wave vector dependence of the reciprocal space term cannot be separated in this way. An approach to calculating density fitted matrix elements at all k points rather than symmetry unique points would be to use only the real space part of the Ewald potential for calculation of fitting coefficients. With some well chosen value of γ\gamma this would limit the number of three-center integrals to be calculated, real space integrals could be stored in memory and Fourier transformed by multiplying in the qq-dependent phase factor as needed. Matrices appearing as [VQ−1]αβ∗[V_{Q}^{-1}]_{\alpha\beta}^{*} would then become triple products of matrices of this size with the inverse of the real space part of the Ewald potential on either side and the full (uninverted) Coulomb potential in the middle.

The formalism outlined in Section II A differs from the usual field theoretic way of introducing the BSE[1] and makes clear the connection of the method linear response theory with the additional provision of screening of the Coulomb potential. The approach outlined here in which RPA calculations are performed for finite Q\mathbf{Q} vectors could easily be adapted to permit exciton dispersion calculations.

This formalism was applied recently to ionization energies and optical excitations in molecules[42, 43]. As in this case, the starting point for the calculations was HFT[100, 101, 102, 103] rather than DFT. In our previous work the mean signed error (MSE) for GoWo@HFG_{o}W_{o}@\mathrm{HF} in π\pi-bonded molecules was 0.3 to 0.4 eV, i.e. the ionization energies were overestimated by 0.3 to 0.4 eV by GoWo@HFG_{o}W_{o}@\mathrm{HF}. For N localized lone pairs states the overestimate was 0.8 to 1.1 eV. Electron affinities were not calculated, however, these overestimates contribute to overestimation of the QP gap when starting GoWoG_{o}W_{o} from HFT. In this work, the overestimate of the GoWo@HFG_{o}W_{o}@\mathrm{HF} QP gap for diamond is 0.27 eV, for bulk Si it is 0.19 eV, for MgO it is 1.36 eV (1.0 eV allowing for polar phonon screening). These values seem seem reasonable compared to those obtained for a range of molecules.

Acknowledgments

This work was supported by Science Foundation Ireland under grant number 19/FFP/6582, SOOMAT. Calculations were performed on the Boyle cluster, maintained by the Trinity Centre for High Performance Computing and funded by Science Foundation Ireland.

Data Availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

References

  1. [1]H. A. Bethe and E. E. Salpeter, Phys. Rev. 82, 309 (1951).
  2. [2]L. Hedin, Phys. Rev. 139, A796 (1965).
  3. [3]M. S. Hybertsen and S. G. Louie, Phys. Rev. Lett. 55, 1418 (1985).
  4. [4]M. S. Hybertsen and S. G. Louie, Phys. Rev. B 34, 5390 (1986).
  5. [5]M. S. Hybertsen and S. G. Louie, Phys. Rev. B 35, 5585 (1987).
  6. [6]J. Deslippe, G. Samsonidze, D. A. Strubbe, M. Jain, M. L. Cohen, and S. G. Louie, Comput. Phys. Commun. 183, 1269 (2012).
  7. [7]A. Marini, C. Hogan, M. Grüning, and D. Varsano, Comput. Phys. Commun. 180, 1392 (2009).
  8. [8]D. Sangalli, A. Ferretti, H. Miranda, C. Attaccalite, I. Marri, E. Cannuccia, P. M. Melo, M. Marsili, F. Paleari, A. Marrazzo, G. Prandini, P. Bonfa, M. O. Atambo, F. Affinito, M. Palummo, A. M. Sanchez, C. Hogan, M. Grüning, D. Varsano, and A. Marini, J. Phys. Cond. Matt. 31, 325902 (2019).
  9. [9]M. Shishkin, M. Marsman, and G. Kresse, Phys. Rev. Lett. 99, 246403 (2007).
  10. [10]T. Sander, E. Maggio, and G. Kresse, Phys. Rev. B 92, 045209 (2015).
  11. [11]F. Bruneval, N. Vast, and L. Reining, Phys. Rev. B 74, 045102 (2006).
  12. [12]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. Cavig nac, J.-B. Charraud, W. Chen, M. Côté, S. Cott enier, J. Denier, G. Geneste, P. Ghosez, M. Gian tomassi, 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. L. Marques, A. Martin, C. Martins, H. P. C. 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. V. Troeye, M. J. Verstraete, G. Zérah, and J. W. Zwan ziger, Comput. Phys. Commun. 248, 107042 (2020).
  13. [13]C. H. Patterson, Mol. Phys. 108, 3181 (2010).
  14. [14]C. H. Patterson, Phys. Rev. Mater. 3, 043804 (2019).
  15. [15]C. H. Patterson, J. Chem. Phys. 153, 064107 (2020).
  16. [16]J. L. Whitten, J. Chem. Phys. 58, 4496 (1973).
  17. [17]B. I. Dunlap, J. W. D. Connolly, and J. R. Sabin, J. Chem. Phys. 71, 3396 (1979).
  18. [18]J. W. Mintmire, J. R. Sabin, and S. B. Trickey, Phys. Rev. B 26, 1743 (1982).
  19. [19]S. Reine, E. Teilgren, A. Krapp, T. Kjaergaard, T. Helgaker, B. Jansik, S. Høst, and P. Salek, J. Chem. Phys. 129, 104101 (2008).
  20. [20]L. Maschio, D. Usvyat, F. R. Manby, S. Casassa, C. Pisani, and M. Schütz, Phys. Rev. B 76, 075101 (2007).
  21. [21]D. Usvyat, L. Maschio, F. R. Manby, S. Casassa, M. Schütz, and C. Pisani, Phys. Rev. B 76, 075102 (2007).
  22. [22]M. Miliko, J. Noga, and S. Varga, Int. J. Quantum Chem. 107, 2158 (2007).
  23. [23]S. Varga, Int. J. Quantum Chem. 108, 1518 (2008).
  24. [24]A. M. Burow, M. Sierka, and F. Mohamed, J. Chem. Phys. 131, 214101 (2009).
  25. [25]M. Katouda and S. Nagase, J. Chem. Phys. 133, 184103 (2010).
  26. [26]M. Lorenz, L. Maschio, M. Schütz, and D. Usvyat, J. Chem. Phys. 137, 204119 (2012).
  27. [27]M. D. Ben, J. Hutter, and J. VandeVondele, J. Chem. Theory Comput. 9, 2654 (2013).
  28. [28]M. Francini, P. H. T. Philipsen, E. van Lenthe, and L. Visscher, J. Chem. Theory Comput. 10, 1994 (2014).
  29. [29]W. G. Sun, J. J. Wang, C. Lu, X. X. Xia, X. Y. Kuang, and A. Hermann, Inorg. Chem. 56, 1241 (2017).
  30. [30]X. Wang, C. A. Lewis, and E. F. Valeev, “Efficient evaluation of exact exchange for periodic systems via concentric atomic density fitting,” (2020).DOI
  31. [31]J. Hofierka, B. Cunningham, C. M. Rawlins, C. H. Patterson, and D. G. Green, Nature 606, 688 (2022).
  32. [32]M. Rohlfing, P. Krüger, and J. Pollmann, Phys. Rev. B 48, 17791 (1993).
  33. [33]M. Rohlfing, P. Krüger, and J. Pollmann, Phys. Rev. B 52, 1905 (1995).
  34. [34]M. Rohlfing, P. Krüger, and J. Pollmann, Phys. Rev. B 57, 6485 (1998).
  35. [35]M. Rohlfing and S. G. Louie, Phys. Rev. B 62, 4927 (2000).
  36. [36]T. Zhu and G. K. L. Chan, J. Chem. Theory Comput. 17, 727 (2021).
  37. [37]C.-N. Yeh, S. Iskakov, D. Zgid, and E. Gull, Phys. Rev. B 106, 235104 (2022).
  38. [38]V. Abraham, G. Harsha, and D. Zgid, J. Chem. Theory Comput. 20, 4579 (2024).
  39. [39]M. A. García-Blázquez and J. J. Palacios, Phys. Rev. Res. 7, 013156 (2025).
  40. [40]H.-Z. Ye and T. C. Berkelbach, J. Chem. Theory Comput. 20, 8948 (2024).
  41. [41]G. Strinati, Phys. Rev. B 29, 5718 (1988).
  42. [42]C. H. Patterson, J. Chem. Theory Comput. 20, 7479 (2024).
  43. [43]D. Waide and C. H. Patterson, J. Chem. Phys. 161, 224108 (2024).
  44. [44]A. Drew and M. Head-Gordon, Chem. Rev. 105, 4009 (2005).
  45. [47]D. L. Lin, Phys. Rev. A 16, 600 (1977).
  46. [48]A. Starace, Phys. Rev. A 3, 1242 (1971).
  47. [49]M. Guzzi, F. Pio, G. Spinolo, A. Vedda, C. B. Azzoni, and A. Paleari, Phys. Rev. B 4405, 34 (1986).
  48. [50]P. Carrier, S. Rohra, and A. Görling, Phys. Rev. B 75, 205126 (2007).
  49. [51]F. Weigend and R. Ahlrichs, Phys. Chem. Chem. Phys. 7, 3297 (2005).
  50. [52]J. Heyd, J. E. Peralta, G. E. Scuseria, and R. L. Martin, J. Chem. Phys. 123, 174101 (2005).
  51. [53]F. Weigend, M. Häser, H. Patzelt, and R. Ahlrichs, Chem. Phys. Lett. 294, 143 (1998).
  52. [54]F. J. Himpsel, J. F. Vanderveen, and D. E. Eastman, Phys. Rev. B 22, 1967 (1980).
  53. [55]I. Jiménez, L. J. Terminello, D. G. J. Sutherland, J. A. Carlisle, E. L. Shirley, and F. J. Himpsel, Phys. Rev. B 56, 7215 (1997).
  54. [56]M. T. Edmonds, A. Tadich, M. Wanke, K. M. O’Donnell, Y. Smets, K. J. Rietwyk, J. D. Riley, C. I. Pakes, and L. Ley, Phys. Rev. B 87, 085123 (2013).
  55. [57]R. A. Roberts and W. C. Walker, Phys. Rev. 161, 730 (1967).
  56. [58]S. Korneychuk, G. Guzzinati, and J. Verbeeck, Phys. Stat. Solidi A 215, 1800318 (2018).
  57. [59]H. Löfås, A. Grigoriev, J. Isberg, and R. Ahuja, AIP Adv. 1, 032139 (2011).
  58. [60]S.-P. Gao, Phys. Stat. Sol. B 252, 235 (2015).
  59. [61]D. Nabok, A. Gulans, and C. Draxl, Phys. Rev. B 94, 035118 (2016).
  60. [62]F. R. McFeely, S. P. Kowalczyk, L. Ley, R. G. Cavell, R. A. Pollak, and D. A. Shirley, Phys. Rev. B 9, 5268 (1974).
  61. [63]Numerical Data and Functional Relationships in Science and Technology, Landolt-Börnstein, New Series group III, Vol. 17a, Ed. K. H. Hellwege, O. Madelung and H. Weiss Springer, New York, (1982)).
  62. [64]A. S. Barnard, S. P. Russo, and I. K. Snook, Phil. Mag. 82, 1767 (2002).
  63. [65]A. Stoyanova, A. O. Mitrushchenkov, L. Ho zoi, H. Stoll, and P. Fulde, Phys. Rev. B 89, 235121 (2014).
  64. [66]T. Yokoya, T. Nakamura, T. Matsushita, T. Muro, H. Okazaki, M. Arita, K. Shimada, H. Namatame, M. Taniguchi, Y. Takano, M. Nagao, T. Takenouchi, H. Kawarada, and T. Oguchi, Sci. Technol. Adv. Mater. 7, S12 (2006).
  65. [67]T. Shimazaki and Y. Asai, Chem. Phys. Lett. 466, 91 (2008).
  66. [68]R. Godby, M. Schlüter, and L. J. Sham, Phys. Rev. B 37, 10159 (1988).
  67. [69]P. Lautenschlager, M. Garriga, L. Viña, and M. Cardona, Phys. Rev. B 36, 4821 (1987).
  68. [70]R. Hulthén and N. G. Nilsson, Sol. State Commun. 18, 1341 (1976).
  69. [71]D. Straub, L. Ley, and F. J. Himpsel, Phys. Rev. Lett. 54, 142 (1985).
  70. [72]F. J. Himpsel, P. Heimann, and D. E. Eastman, Phys. Rev. B 24, 2003 (1981).
  71. [73]A. L. Wachs, T. Miller, T. C. Hsieh, A. P. Shapiro, and T.-C. Chiang, Phys. Rev. B 32, 2326 (1985).
  72. [75]D. M. Roessler and W. C. Walker, Phys. Rev. 159, 733 (1967).
  73. [76]R. C. Whited, C. J. Flaten, and W. C. Walker, Sol. State Commun. 13, 1903 (1973).
  74. [80]V. Begum, M. E. Gruner, C. Vorwerk, C. Draxl, and R. Pentcheva, Phys. Rev. B 103, 195128 (2021).
  75. [82]G. Antonius, S. Poncé, E. Lantagne-Hurtubise, G. Auclair, X. Gonze, and M. Côté, Phys. Rev. B 92, 085137 (2015).
  76. [83]E. Baldini, L. Chiodo, A. Dominguez, M. Palummo, S. Moser, M. Yazdi-Rizi, G. Auböck, B. P. P. Mallett, H. Berger, A. Magrez, C. Bernhard, M. Grioni, A. Rubio, and M. Chergui, Nature Commun. 8, 1 (2018).
  77. [84]W. Kang and M. S. Hybertsen, Phys. Rev. B 82, 085203 (2010).
  78. [85]J. Pascual, J. Camassel, and H. Mathieu, Phys. Rev. Lett. 39, 1490 (1977).
  79. [86]E. Baldini, A. Dominguez, L. Chiodo, E. Sheveleva, M. Yazdi-Rizi, C. Bernhard, A. Rubio, and M. Chergui, Phys. Rev. B 96, 041204(R) (2017).
  80. [87]M. Zhang, S. Ono, and K. Ohno, Phys. Rev. B 92, 035205 (2015).
  81. [88]L. Chiodo, J. M. Garcia-Lastra, A. Iacomino, S. Ossicini, J. Zhao, H. Petek, and A. Rubio, Phys. Rev. B 82, 045207 (2010).
  82. [89]M. Caricato, G. W. Trucks, M. J. Frisch, and K. B. Wiberg, J. Chem. Theory Comput. 7, 456 (2011).
  83. [90]E. D. Palik, ed., Handbook of Optical Constants of Solids (Academic Press, New York, 1997).
  84. [92]N. Hosaka, T. Sekiya, C. Satoko, and S. Kurita, J. Phys. Soc. Jpn. 66, 877 (1997).
  85. [93]T. E. Tiwald and M. Schubert, Proc. SPIE 4103, 19 (2000).
  86. [94]A. Baldereschi and E. Tosatti, Solid State Commun. 34, 1 (1980).
  87. [95]S. Galamić-Mulaomerović and C. H. Patterson, Phys. Rev. B 71, 195103 (2005).
  88. [96]S. Galamić-Mulaomerović and C. H. Patterson, Phys. Rev. B 72, 35127 (2005).
  89. [97]V. R. Saunders, R. Dovesi, C. Roetti, M. Causà, R. Orlando, C. M. Zicovich-Wilson, N. M. Harrison, K. Doll, B. Civalleri, I. Bush, P. D’Arco, and M. Llunell, Crystal03 User’s Manual, University of Torino, Torino, 2003. (www.crystal.unito.it).
  90. [98]A. Erba, J. K. Desmarais, S. Casassa, B. Civalleri, L. Donà, I. J. Bush, B. Searle, L. Maschio, L.-E. Daga, A. Cossard, C. Ribaldone, E. Ascrizzi, N. L. Marana, J.-P. Flament, and B. Kirtman, J. Chem. Theory Comput. 19, 6891 (2023).
  91. [99]P. Karpov, A. Marek, T. Melson, A. Pöppl, V. W. Yu, B. Hourahine, A. Garcia, W. Dawson, Y. Yao, W. Huhn, J. Moussa, S. Hall, R. Maurer, U. Herath, K. Lion, S. Kokott, and V. Blum, “Solders for large-scale electronic structure theory: Elpa and elsi,” arXiv:2502.02460.arxiv.org/abs/2502.02460
  92. [100]N. Marom, F. Caruso, X. Ren, O. T. Hofmann, T. Körzdörfer, J. R. Chelikowsky, A. Rubio, M. Scheffler, and P. Rinke, Phys. Rev. B 86, 245127 (2012).
  93. [101]F. Bruneval and M. A. L. Marques, J. Chem. Theory Comput. 9, 324 (2013).
  94. [102]J. W. Knight, X. Wang, L. Gallandi, O. Dolgounitcheva, X. Ren, J. V. Ortiz, P. Rinke, T. Körzdörfer, and N. Marom, J. Chem. Theory Comput. 12, 615 (2016).
  95. [103]F. Bruneval and A. Forster, J. Chem. Theory Comput. 20, 3218 (2024).

Paper details

vlnc:W_fchg9pp4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

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