betaFeedback

WorksPhysics and Astronomy

olLOSC: Unified and efficient density functional approximation to correct delocalization error in molecules and periodic materials

Abstract

Density functional theory (DFT) is the most promising method for calculating quantum properties of molecules and materials at moderate and large scales. However, commonly used density functional approximations (DFAs) have systematic delocalization error, as demonstrated by underestimated band gaps, over-delocalized charges, and energy level misalignment at interfaces, which limits its quantitative prediction. Extensive efforts, such as the GWGW approximation to many-body perturbation theory, system-specific tuning of DFA parameters, and correction functionals have been developed to address delocalization error. However, an accurate, efficient, and unified solution to describe total energy, charge density and band structure for both finite systems and materials is still not available. Building on the linear-response localized orbital scaling correction (lrLOSC), we introduce olLOSC: a localized orbital scaling correction with curvature calculated by orbital-free electronic linear response. olLOSC has comparable accuracy to lrLOSC, but is much more computationally efficient. olLOSC corrects delocalization error - especially underestimated gaps, but also the total energy - both in molecules and in materials with small and moderate band gaps, within the same orbital-free approximation. Critically, with a a unified approximation, olLOSC opens the path for robust and efficient DFT applications across molecules, materials, and interfaces.

main.pdf

Significance statement

Key properties of metals in biology and for developing advanced technologies like catalysts and solar energy converters depend on the quantum-mechanical

Introduction

Density functional theory (DFT) is the “workhorse of quantum chemistry and materials science” [1] and probably the method used most often to predict quantum properties of electronic systems. Among its greatest appeals is its reasonable accuracy at calculating ground-state properties, especially the total energy [2], of molecules and materials.

Kohn–Sham (KS) DFT [3] additionally provides single-particle orbital energies, whose meaning has long been a subject of study. Janak [4] showed that orbital energies computed by a density functional approximation (DFA) are equal to the derivative of the ground-state energy with respect to the orbital occupations, as long as the exchange-correlation functional ExcE_{\mathrm{xc}} is continuous in the density ρ(r)\rho(\mathbf{r}). However, this does not connect directly to physical observables. Two relationships connecting observables and orbital energies have been established rigorously. First, the ionization potential theorem [5, 6] demonstrates that, for the exact (local) potential, the highest occupied KS orbital (HOMO) energy is the negative of the first ionization energy (valence band edge) II. Second, the ground state chemical potential theorem [7, 8] shows that, in a ground state KS calculation with ExcE_{\mathrm{xc}} continuous in ρ\rho or a generalized KS Kohn–Sham (GKS) calculation with ExcE_{\mathrm{xc}} continuous in the KS density matrix γ(r,r′)\gamma(\mathbf{r},\mathbf{r}^{\prime}), the energy of the highest occupied (lowest unoccupied) molecular orbital is the chemical potential μ\mu of electron removal (addition). The PPLB linearity condition [5] proves that μ\mu computed by the exact functional are derivatives of the total energy with respect to the number of electrons; thus, for an N0N_0-electron system,

∂E∂N∣N0−=−Iand∂E∂N∣N0+=−A.\left.\frac{\partial E}{\partial N}\right|_{N_0^-} = -I \quad\text{and} \quad\left.\frac{\partial E}{\partial N}\right|_{N_0^+} = -A.

The frontier orbital energies thus correspond—in both KS and GKS calculations—to experimentally observable ionization potentials and electron affinities (in materials, band edge energies and band gaps). The ground state chemical potential established for the first time the physical meaning of the lowest unoccupied molecular orbital (LUMO) and conduction band minimum (CBM) orbital energy in ground-state (G)KS calculations [7]. DFT calculations thus provide estimates of key properties like the fundamental gap, defined as the difference between the ionization potential II and the electron affinity AA,

Eg=I−A≈{εLUMODFT−εHOMODFTin molecules,εCBMDFT−εVBMDFTin materials,(1)E_g = I - A \approx\begin{cases} \varepsilon_{\mathrm{LUMO}}^{\mathrm{DFT}}-\varepsilon_{\mathrm{HOMO}}^{\mathrm{DFT}} & \text{in molecules},\\ \varepsilon_{\mathrm{CBM}}^{\mathrm{DFT}}-\varepsilon_{\mathrm{VBM}}^{\mathrm{DFT}} & \text{in materials}, \end{cases} \tag*{(1)}

as long as ExcE_{\mathrm{xc}} is continuous in ρ(r)\rho(\mathbf{r}) or γ(r,r′)\gamma(\mathbf{r},\mathbf{r}'). In molecules, the DFT-computed gap is the difference between the LUMO and HOMO energies; in materials, the difference is between the analogous conduction band minimum (CBM) and valence band maximum (VBM) energies. If the exact density functional were known and used, this result would be exact [4, 5, 7]. Since it is not, (1) is only approximate, and traditional approaches to its approximation are quite poor. Common density functional approximations (DFAs) like LDA [3, 9, 10] and PBE [11] systematically underestimate the fundamental gap by as much as 40% [12]. This problem, as well the underestimation of chemical reaction energy barriers [13, 14], unphysically delocalized charge densities [15], and DFAs’ qualitative failure to describe molecular dissociation [16, 17], are now known to stem from delocalization error [15, 18, 19].

The manifestation of delocalization error depends on the size of the system [20, 21]. In small molecules near their equilibrium bond lengths, it arises from the failure of the energy E(N)E(N) of a DFA, as a function of the number of electrons, to obey the PPLB linearity condition [5, 22, 23]

E(N+δ)=(1−δ)E(N)+δE(N+1),0≤δ<1.(2)E(N+\delta) = (1-\delta)E(N) + \delta E(N+1), \quad0 \leq\delta< 1. \tag*{(2)}

In the exact functional, E(N)E(N) is piecewise linear, with discontinuities in the derivative at integer NN. Because the derivatives of EE with respect to the Kohn–Sham orbital occupations nnσn_{n\sigma} are the orbital energies [4], the derivative discontinuities give rise to the expression for the fundamental gap (1). E(N)E(N) is convex instead of piecewise linear, however, in almost all DFAs used in practice. This convexity softens the derivative discontinuity and, ultimately, causes the litany of problems due to delocalization error. (Note that Hartree–Fock theory suffers instead from localization error because its E(N)E(N) curve is concave [18]. The success of hybrid functionals, which mix a fraction of Hartree–Fock exchange with a DFA counterpart, is partially due to this cancellation of errors.)

Small molecules exhibit a pronounced curvature in E(N)E(N), but its prominence decreases for larger systems. In the bulk limit, translational symmetry forces E(N)E(N) to be piecewise linear, but the derivative discontinuities (hence band gaps) are still underestimated. In fact, the total energy E(N±1)E(N \pm1) becomes inaccurate in materials [18]. Because of these contrasting behaviors, it is challenging to correct delocalization error in both molecules and materials with the same approximation.

Kohn–Sham DFT’s underestimation of the (band) gap was perhaps the earliest hallmark of delocalization error to be recognized [12, 15], so there are several fairly mature theories that improve gaps and band structure. However, a method that corrects delocalization error in all scenarios must not only correct the band structure; it must also be able to change the total energy of the system in order to describe the dissociation limit correctly and to correct the charge density to describe interfacial systems [24, 25].

Methods to correct delocalization error

Range-separated hybrid functionals can mitigate delocalization error in molecules or materials, but they do so in completely different ways. In molecules, Coulomb-attenuated functionals [26] use Hartree–Fock (exact) exchange at long range; they cancel delocalization error imperfectly, but improve molecular fundamental gap calculations significantly [27, 28]. In bulk systems, however, the 1/r1/r asymptotic decay of Hartree–Fock exchange is qualitatively wrong. Long-range correlations decay exponentially in materials with a band gap [29, 30, 31], and exchange is further scaled by the inverse macroscopic dielectric constant 1/ϵ∞1/\epsilon_{\infty} [32]. (More generally, it is scaled by the inverse of the microscopic dielectric function, ϵ−1(r,r′)\epsilon^{-1}(\mathbf{r},\mathbf{r}^{\prime}); this approach is used in the GWGW method [33, 34].) Range-separated hybrids for materials therefore use Hartree–Fock exchange only for short-range interactions, with a fixed screening to describe long-range interactions [35, 36, 37]. The correct mixture of DFT and Hartree–Fock exchange also depends on the particular system. The optimally tuned range-separated hybrid functional [38, 39] incorporates both system-dependent exchange and ϵ∞\epsilon_{\infty} to correct delocalization error. With its recent extension to materials [40], it is effective at restoring the linearity condition, but the optimal tuning depends not only on the system but on its geometry; it is therefore difficult to obtain consistent energetics of (for instance) the reactants and products a chemical reaction, or of an interface and its separate components. Optimally tuned range-separated hybrid functionals are used primarily to correct band gaps, not total energies.

Koopmans-compliant functionals handily correct quasiparticle spectra in molecules [41, 42]. So does its periodic extension, the Koopmans spectral, or Koopmans-compliant Wannier (KCW) functional [43, 44]. KCW is currently implemented non-self-consistently [45], but because it corrects the Hamiltonian it could be applied self-consistently to provide a density correction, following a similar prescription as [46]. The Wannier–Koopmans method, which shares a similar philosophy, is similarly effective at improving band gap prediction [47–51]. However, Koopmans-compliant corrections do not correct the total energy of insulators or typical molecules.

The GWGW method [33, 34] improves gaps and band structures significantly, especially when coupled with some degree of self-consistency [52–55]. Furthermore, it offers an implicit correction to the density. However, its description of the total energy is suspect [56] unless computed fully self-consistently, which is generally prohibitive in computational cost. This scales naively with the fourth power of the number NeN_e of electrons; even after approximations to reduce the scaling [57], the prefactor is characteristically large, so fully self-consistent GWGW calculations are rare. Keeping delocalization error corrections at DFT cost, O(Ne3)O(N_e^3), with reasonable practical time cost, is thus desirable.

Both the Perdew–Zunger self-interaction correction (PZ-SIC) [58] and its modern descendant, the Fermi–Löwdin orbital self-interaction correction (FLO-SIC) correct delocalization error partially, focusing on its single-electron manifestation. Unlike its earlier predecessor, FLOSIC’s energy correction is invariant under unitary transformation of the occupied orbitals and size-consistent [59]. In addition to energy corrections, FLOSIC can be applied self-consistently [60] and yields a corrected density [61]. However, the quality of its correction to the orbital energies is less clear. FLOSIC provides improvement over DFAs for molecular ionization potentials [62, 63] and, when maximally localized Wannier functions [64, 65] are used to construct the Fermi orbitals, modest improvement in band gaps of semiconductors [66]. However, the improvements are somewhat inconsistent, and worsen when the DFA is a gradient-corrected functional.

The localized orbital scaling method (LOSC), introduced by [25], is a delocalization error correction that offers a correction to the total energy and electron density as well as to band gaps. It does this through the construction of localized orbitalets (called dually localized Wannier functions in materials [67]) that encode both spatial and spectral localization in a balanced way [25] [68]. The original implementation of LOSC performs well for molecular systems, but breaks down in bulk materials because it lacks a description of dielectric screening [69]. More recently, we developed linear-response (lr)LOSC, adding dielectric screening from linear-response theory [21] to LOSC. lrLOSC yields semiconductor and insulator band structures [70], molecular ionization potentials and electron affinities [71], and core-level binding energies [72] of comparable accuracy to (and in some cases better than) GWGW methods. Nevertheless, the linear-response kernel is expensive to compute, compared with the conventional DFT calculations, preventing lrLOSC from being broadly applicable to complex bulk systems and interfaces of interest.

In this work, we present olLOSC, a computationally efficient approximation to lrLOSC with an orbital-free linear-response curvature. olLOSC is shown as a unified functional approximation to correct delocalization error in both molecules and semiconductors. Looking ahead, we seek an approximation that makes DFT reliably accurate for computing the electronic structure of interfaces, which combine challenging features of both molecules and materials. olLOSC, applicable to both finite and bulk systems—each a limiting case of the more general interface—is an important step toward this goal.

Theory

All methods in the LOSC family correct delocalization error by applying a quadratic-to-linear correction functional to restore (2) approximately [21, 25] [73, 74]. The most advanced form is the LOSC correction

ΔELOSC=∑σ∑ij12λijσ(δij−λijσ)κijσ,(3)\Delta E_{\mathrm{LOSC}} = \sum_{\sigma}\sum_{ij} \frac{1}{2}\lambda_{ij\sigma}(\delta_{ij}-\lambda_{ij\sigma})\kappa_{ij\sigma}, \tag*{(3)}

where the matrix of local occupations λijσ\lambda_{ij\sigma} is the one-electron density matrix of the noninteracting reference system in a basis of localized orbitalets ∣ϕiσ⟩\lvert\phi_{i\sigma}\rangle [25],

λijσ=⟨ϕiσ∣ρ∣ϕjσ⟩.(4)\lambda_{ij\sigma} = \langle\phi_{i\sigma}|\rho|\phi_{j\sigma}\rangle. \tag*{(4)}

The curvature κijσ\kappa_{ij\sigma} measures the magnitude of the correction for each pair of orbitalets. δij\delta_{ij} is the Kronecker delta function, equal to 1 when i=ji=j and to 0 otherwise. ΔELOSC\Delta E_{\mathrm{LOSC}} is quadratic in λijσ\lambda_{ij\sigma}, which addresses the major contribution to delocalization error [75].1

The form of ΔELOSC\Delta E_{\mathrm{LOSC}} shares similarities to those of other well-known post-DFT methods, such as the rotationally invariant formulation [76] of DFT+UU [77] and the Koopmans-compliant functionals [41] [78]. The curvature κijσ\kappa_{ij\sigma} is the second derivative of the total energy with respect to the orbital occupation numbers [21] [79], expressed in the basis of localized orbitalets. It bears strong similarities to linear-response DFT+UU [80] and to the Koopmans-compliant Wannier functional [43, 44] [45]. Neither of the latter energy corrections, however, can be applied to molecules and materials in all situations because their respective localized orbitals have fixed occupations, either 1 or 0. Studies on homolytic dissociation demonstrate that the total energy from DFAs is typically accurate near equilibrium, but yields a qualitatively incorrect dissociation limit [15]. DFT+UU, based on local (often atomic d) orbital occupations, may provide a large correction to the energy even at molecular equilibrium geometries. On the other hand, the Koopmans-compliant energy correction is based on the occupations of Kohn–Sham orbitals or maximally localized Wannier functions, so it can offer no total energy correction in insulating systems, even stretched molecules. In particular, they fail to predict the correct dissociation of molecular ions such as H2+_2^+. A unified correction to delocalization error therefore requires dynamic localization, which is LOSC’s first key feature.

Localization and orbitalets

LOSC’s localized orbitals mix the occupied and virtual manifolds to balance localization in the spatial and energy domains. The resulting orbitals, known as orbitalets for finite systems [25] [68] and dually localized Wannier functions (DLWFs) for periodic systems [67] are obtained from the canonical (Kohn–Sham) orbitals ∣ψnσ⟩|\psi_{n\sigma}\rangle by unitary transformation,

∣ϕiσ⟩=∑nUniσ∣ψnσ⟩,(5)|\phi_{i\sigma}\rangle= \sum_n U^\sigma_{ni}|\psi_{n\sigma}\rangle, \tag*{(5)}

with UσU^\sigma chosen to minimize a cost function FF. Initially defined as a penalty based on the Kohn–Sham eigenvalues [25], FF was reformulated in [68] to take the form

Fσ=∑i[(1−γ)Δriσ2+CγΔhiσ2];0≤γ≤1.(6)F^\sigma= \sum_i \left[(1-\gamma)\Delta r_{i\sigma}^{2} + C\gamma\Delta h_{i\sigma}^{2}\right];\quad0 \leq\gamma\leq1. \tag*{(6)}

LOSC for materials is implemented in periodic boundary conditions and supports nontrivial sampling of the Brillouin zone; for simplicity, we omit the associated index kk in this text. We additionally require translational symmetry of orbitals in materials, making them (generalized) Wannier functions [64, 65, 67].

FσF^\sigma is thus a convex sum of the spatial variance Δriσ2=⟨ϕiσ∣r∣ϕiσ⟩2−⟨ϕiσ∣r2∣ϕiσ⟩\Delta r_{i\sigma}^{2} = \langle\phi_{i\sigma}|\mathbf{r}|\phi_{i\sigma}\rangle^{2} - \langle\phi_{i\sigma}|r^{2}|\phi_{i\sigma}\rangle of the orbitalets ∣ϕiσ⟩|\phi_{i\sigma}\rangle and their variance in energy Δhiσ2=⟨ϕiσ∣h∣ϕiσ⟩2−⟨ϕiσ∣h2∣ϕiσ⟩\Delta h_{i\sigma}^{2} = \langle\phi_{i\sigma}|h|\phi_{i\sigma}\rangle^{2} - \langle\phi_{i\sigma}|h^{2}|\phi_{i\sigma}\rangle [68, 81]. The constant CC ensures that FσF^\sigma is dimensionally consistent.2

The key difference between orbitalets (DLWFs) and all previous localization approaches is that they are not limited to mixing orbitals with the same occupation; that is, they allow mixing between occupied (valence) and virtual (conduction) orbitals. Allowing this mixing means that the diagonal occupations λiiσ\lambda_{ii\sigma} are not constrained to integral values, even in insulators. It is these noninteger local occupations that allow LOSC to correct the total energy, necessary for correcting delocalization error in molecular dissociation.

Molecular orbitalets are dynamic: different molecular environments yield qualitatively different orbitalets when FσF^\sigma is minimized. This is illustrated dramatically by stretched diatomic molecules. Near equilibrium, the relatively large gap between Kohn–Sham eigenvalues forces the orbitalets to approximate the canonical orbitals closely; the local and canonical occupations are essentially the same—integers—and LOSC yields almost no correction to the total energy. If the molecule is stretched, however, the canonical orbitals become increasingly delocalized, and the energy gap between them narrows. The coupled change in spatial and spectral characteristics drives a transition that localizes the orbitalets on their respective atoms. Their occupations λiiσ\lambda_{ii\sigma} become fractional, and LOSC provides a substantial correction to the total energy that improves the description of dissociation greatly [25].

Linear-response curvature

The curvature κ\kappa measures the deviation from linearity of the DFA energy with respect to orbital occupation numbers. Yang and coworkers derived this quantity to second order with respect to the Kohn–Sham (canonical) occupations nnσ=⟨ψnσ∣ρ∣ψnσ⟩n_{n\sigma} = \langle\psi_{n\sigma}|\rho|\psi_{n\sigma}\rangle [79], finding that

∂2E∂nnσ∂nmτ=⟨ψnσ∗ψmσ∣[fHxcστ+∑μνfHxcσμχμνfHxcντ]∣ψmτ∗ψmτ⟩.(7)\frac{\partial^{2} E}{\partial n_{n\sigma}\partial n_{m\tau}} = \langle\psi_{n\sigma}^{*}\psi_{m\sigma}| \left[ f_{\mathrm{Hxc}}^{\sigma\tau} + \sum_{\mu\nu} f_{\mathrm{Hxc}}^{\sigma\mu}\chi^{\mu\nu}f_{\mathrm{Hxc}}^{\nu\tau} \right] |\psi_{m\tau}^{*}\psi_{m\tau}\rangle. \tag*{(7)}

Here, fHxcστ=∂2EHxc/∂ρσ∂ρτf_{\mathrm{Hxc}}^{\sigma\tau} = \partial^{2}E_{\mathrm{Hxc}}/\partial\rho^{\sigma}\partial\rho^{\tau} is the Hartree–exchange-correlation (Hxc) kernel, and χμν=δρν/δvμ\chi^{\mu\nu} = \delta\rho^{\nu}/\delta v^{\mu} is the (static) linear response function of the density to an external perturbing potential. For details of multiple derivations of ∂2E/∂nnσ∂nmτ\partial^{2}E/\partial n_{n\sigma}\partial n_{m\tau}, see the Supporting Information of [21].

However, as mentioned above, local orbitals with fractional occupations are required to correct the total energy, necessary to correct delocalization error size-consistently. The linear-response (lr)LOSC curvature ansatz is (7) expressed in the orbitalet basis; for collinear spins, this yields [70, 72]

κijσ=⟨ρiσ∣fHxcσσ+∑νμfHxcσνχνμfHxcμσ∣ρjσ⟩.(8)\kappa_{ij\sigma}=\langle\rho_{i\sigma}\left|f_{\mathrm{Hxc}}^{\sigma\sigma}+\sum_{\nu\mu}f_{\mathrm{Hxc}}^{\sigma\nu}\chi^{\nu\mu}f_{\mathrm{Hxc}}^{\mu\sigma}\right|\rho_{j\sigma}\rangle. \tag*{(8)}

The linear response function χ\chi can be written in terms of the noninteracting linear response function χs\chi_s with the Dyson equation

χστ=χsστ+χsσνfHxcνμχμτ,(9)\chi^{\sigma\tau}=\chi_s^{\sigma\tau}+\chi_s^{\sigma\nu}f_{\mathrm{Hxc}}^{\nu\mu}\chi^{\mu\tau}, \tag*{(9)}

where repeated spins (and their corresponding spatial variables) are integrated over. In Kohn–Sham calculations, χs\chi_s takes the well-known form

χsστ(r,r′)=δρσ(r)δvKSτ(r′)=δστ×∑iaψiσ∗(r)ψaσ(r)ψaσ∗(r′)ψiσ(r′)εiσ−εaσ,(10)\chi_s^{\sigma\tau}(\mathbf{r},\mathbf{r}')=\frac{\delta\rho^\sigma(\mathbf{r})}{\delta v_{\mathrm{KS}}^\tau(\mathbf{r}')}=\delta_{\sigma\tau}\times\sum_{ia}\frac{\psi_{i\sigma}^{*}(\mathbf{r})\psi_{a\sigma}(\mathbf{r})\psi_{a\sigma}^{*}(\mathbf{r}')\psi_{i\sigma}(\mathbf{r}')}{\varepsilon_{i\sigma}-\varepsilon_{a\sigma}}, \tag*{(10)}

where vKSτ=vHxcτ+vτv_{\mathrm{KS}}^\tau=v_{\mathrm{Hxc}}^\tau+v^\tau is the effective potential, which includes both Hxc and external terms. In the second equality, ii indexes occupied and aa virtual Kohn–Sham orbitals.

Eq. (10) suggests a self-consistent method for calculating κijσ\kappa_{ij\sigma}, but since χ\chi depends on two spatial coordinates and χs\chi_s includes a sum over many unoccupied orbitals, such a method promises to be computationally expensive. It turns out that either self-consistency or the sum over unoccupied states can be avoided. We can rewrite the Dyson equation for χ\chi (omitting spin indices) as

χ−1=χs−1−fHxc.(11)\chi^{-1}=\chi_s^{-1}-f_{\mathrm{Hxc}}. \tag*{(11)}

If χs\chi_s can be inverted numerically—using, for example, a discretization enabled by the resolution-of-the-identity (RI) method [82]—then χ\chi can be computed noniteratively. This is the method used in lrLOSC for molecules [71, 72]. On the other hand, the Sternheimer equation [83] can be solved in the manner of density functional perturbation theory, bypassing the sum over virtual orbitals; this method is used in lrLOSC for materials [70].3

Computing κijσ\kappa_{ij\sigma} with linear response yields a very accurate correction to delocalization error for molecules, semiconductors, and insulators. lrLOSC predicts band gaps ranging from less than 0.5 eV to more than 15 eV with accuracy comparable to or better than self-consistent GW [70]; the fundamental gaps and core-level binding energies of molecules are of similar accuracy [72]. However, even with the speedup afforded by the Sternheimer equation, calculating curvature in materials with linear response remains computationally costly, limiting the application of lrLOSC to crystals with relatively small unit cells compared to systems of practical interest for commonly used DFT approximations.

Orbital-free linear-response curvature

There is another way to think about χ\chi and χs\chi_s that enables a faster calculation with only a modest loss in accuracy: the orbital-free approach [84]. The total energy E[ρ]E[\rho] obeys the stationarity condition

δEδρσ(r)=δFHKδρσ(r)+vσ(r)=μσ(12)\frac{\delta E}{\delta\rho^\sigma(\mathbf{r})}=\frac{\delta F_{\mathrm{HK}}}{\delta\rho^\sigma(\mathbf{r})}+v^\sigma(\mathbf{r})=\mu^\sigma \tag*{(12)}

where FHK[ρ]=Ts[ρ]+J[ρ]+Exc[ρ]F_{\mathrm{HK}}[\rho]=T_s[\rho]+J[\rho]+E_{\mathrm{xc}}[\rho] is the Hohenberg–Kohn universal functional, vσv^\sigma is the external potential, and μσ\mu^\sigma is the spin-σ\sigma chemical potential. Taking a second variation yields

δ2FHKδρσ(r)δρτ(r′)+δvσ(r)δρτ(r′)=δμσδρτ(r′)(13)\frac{\delta^2 F_{\mathrm{HK}}}{\delta\rho^\sigma(\mathbf{r})\delta\rho^\tau(\mathbf{r}')}+\frac{\delta v^\sigma(\mathbf{r})}{\delta\rho^\tau(\mathbf{r}')}=\frac{\delta\mu^\sigma}{\delta\rho^\tau(\mathbf{r}')} \tag*{(13)}

and rearranging gives the Euler–Lagrange equation

∑τ∫dr′[δ2FHKδρσ(r)δρτ(r′)]δρτ(r′)=δμσ−δvσ(r)(14)\sum_\tau\int d\mathbf{r}'\left[\frac{\delta^2 F_{\mathrm{HK}}}{\delta\rho^\sigma(\mathbf{r})\delta\rho^\tau(\mathbf{r}')}\right]\delta\rho^\tau(\mathbf{r}')=\delta\mu^\sigma-\delta v^\sigma(\mathbf{r}) \tag*{(14)}

for the perturbing potential δvσ(r)\delta v^\sigma(\mathbf{r}) and its linear response ∣δρτ⟩\lvert\delta\rho^\tau\rangle [84]. The Lagrange multiplier δμσ\delta\mu^\sigma ensures the conservation of charge (equivalently, electron number). We quickly obtain an expression for the linear-response function:

χστ(r,r′)=[δvσ(r)δρτ(r′)]−1=[δμσ−δ2FHKδρσ(r)δρτ(r′)]−1(15)\chi^{\sigma\tau}(\mathbf{r},\mathbf{r}')=\left[\frac{\delta v^\sigma(\mathbf{r})}{\delta\rho^\tau(\mathbf{r}')}\right]^{-1} = \left[\delta\mu^\sigma-\frac{\delta^2F_{\mathrm{HK}}}{\delta\rho^\sigma(\mathbf{r})\delta\rho^\tau(\mathbf{r}')}\right]^{-1} \tag*{(15)}

Finally, in the Kohn–Sham auxiliary system, there is no Hartree–exchange-correlation energy, so FHKF_{\mathrm{HK}} is exactly the noninteracting kinetic energy TsT_s. Thus

χsστ(r,r′)=[δμσ−δ2Tsδρσ(r)δρτ(r′)]−1=[δμσ−fTsστ(r,r′)]−1(16)\chi_s^{\sigma\tau}(\mathbf{r},\mathbf{r}')=\left[\delta\mu^\sigma-\frac{\delta^2T_s}{\delta\rho^\sigma(\mathbf{r})\delta\rho^\tau(\mathbf{r}')}\right]^{-1} =\left[\delta\mu^\sigma-f_{T_s}^{\sigma\tau}(\mathbf{r},\mathbf{r}')\right]^{-1} \tag*{(16)}

If we had the exact orbital-free kinetic energy functional Ts[ρσ]T_s[\rho^\sigma], (9) and (16) would lead to a χστ\chi^{\sigma\tau} identical to the coupled-perturbed Kohn–Sham equations. This relationship suggests a computationally simplification; we approximate fTsf_{T_s} by the kernel of an orbital-free kinetic energy functional. In this work, we choose the Thomas–Fermi functional [85, 86] with the von Weizsäcker correction [87], obtaining

fTFvWστ(r,r′)=fTFστ(r,r′)+λfvWστ(r,r′)(17)f_{\mathrm{TFvW}}^{\sigma\tau}(\mathbf{r},\mathbf{r}')=f_{\mathrm{TF}}^{\sigma\tau}(\mathbf{r},\mathbf{r}')+\lambda f_{\mathrm{vW}}^{\sigma\tau}(\mathbf{r},\mathbf{r}') \tag*{(17)}

where fTFστ=δ2TTF/δρσδρτf_{\mathrm{TF}}^{\sigma\tau}=\delta^2T_{\mathrm{TF}}/\delta\rho^\sigma\delta\rho^\tau is the Thomas–Fermi kernel, fvWf_{\mathrm{vW}} the von Weizsäcker kernel, and λ≥0\lambda\geq0 controls the amount of von Weizsäcker correction. (The analytic form of fTFvWf_{\mathrm{TFvW}} can be found in the Supporting Information.) Like its Kohn–Sham counterpart, fTFvWf_{\mathrm{TFvW}} is diagonal in spin, with fTFvWστ=fTFvWσσδστf_{\mathrm{TFvW}}^{\sigma\tau}=f_{\mathrm{TFvW}}^{\sigma\sigma}\delta_{\sigma\tau}. The Thomas–Fermi kernel is local in space, while fvWf_{\mathrm{vW}} includes derivatives of the Dirac delta distribution; however, the action of fTFvWf_{\mathrm{TFvW}} on ∣δρi⟩τ|\delta\rho_i\rangle_\tau can still be evaluated locally. We thus obtain an orbital-free approximation

χTFvWστ(r,r′)=[δμσ−fTFvWστ(r,r′)]−1(18)\chi_{\mathrm{TFvW}}^{\sigma\tau}(\mathbf{r},\mathbf{r}')=\left[\delta^{\mu\sigma}-f_{\mathrm{TFvW}}^{\sigma\tau}(\mathbf{r},\mathbf{r}')\right]^{-1} \tag*{(18)}

to χs\chi_s. Replacing χs\chi_s by χTFvW\chi_{\mathrm{TFvW}}4 in (9) yields the orbital-free approximation χol\chi_{\mathrm{ol}} to the interacting response function; solving for χol\chi_{\mathrm{ol}} directly or iteratively, we obtain the orbital-free linear-response (ol)LOSC curvature κijσol\kappa_{ij\sigma}^{\mathrm{ol}}.

The lrLOSC and olLOSC curvature expressions contain two components: a bare response to fHxcσσf_{\mathrm{Hxc}}^{\sigma\sigma}, indicated by the first summand in (8), and a response screened by χ\chi or χol\chi_{\mathrm{ol}} from the second summand. It is this screening, or orbital relaxation, that olLOSC approximates with an orbital-free kinetic energy kernel. Because screening is a collective, many-electron effect, using fTFvWf_{\mathrm{TFvW}} in olLOSC should capture the effect well; choosing a better Ts[ρ]T_s[\rho] would likely improve accuracy.

The partial random phase approximation

We also make a partial random phase approximation (RPA) in olLOSC, replacing the Hxc kernel fHxcμνf_{\mathrm{Hxc}}^{\mu\nu} by the Hartree kernel fHμν(r,r′)=1/∣r−r′∣f_{\mathrm{H}}^{\mu\nu}(\mathbf{r},\mathbf{r}')=1/\lvert\mathbf{r}-\mathbf{r}'\rvert in (9). There are two reasons for doing so. First, including the exchange-correlation kernel in χ\chi induces major numerical instabilities when computing κol\kappa^{\mathrm{ol}} in materials, while χolRPA\chi_{\mathrm{olRPA}} can be obtained by inverting a positive semidefinite matrix. Second, numerical tests on molecules actually demonstrate better fundamental gaps from χolRPA\chi_{\mathrm{olRPA}} than from χol\chi_{\mathrm{ol}}. See the Supporting Information for details. This approximation is akin to the screening described in the GWGW approximation [33, 34]: the operator WW accounts only for Coulombic screening, while exchange-correlation effects are contained in the vertex function (and therefore usually neglected). Note that exchange-correlation interactions are excluded only from χ\chi, not entirely removed from the olLOSC curvature; fxcf_{\mathrm{xc}} is still present in both terms of (8).

Hamiltonian and spectrum

We derive the LOSC correction to the DFA Hamiltonian from (3). At each spin component [25] [46],

Δhσ=δΔEδρσ=∑ij[∂ΔE∂λijσδλijσδρσ+∂ΔE∂λijσ∗δλijσ∗δρσ]=∑ijκijσ(12δij−Re⁡λijσ)∣ϕiσ⟩⟨ϕjσ∣.(19)\Delta h^\sigma= \frac{\delta\Delta E}{\delta\rho^\sigma} = \sum_{ij} \left[ \frac{\partial\Delta E}{\partial\lambda_{ij\sigma}} \frac{\delta\lambda_{ij\sigma}}{\delta\rho^\sigma} + \frac{\partial\Delta E}{\partial\lambda^*_{ij\sigma}} \frac{\delta\lambda^*_{ij\sigma}}{\delta\rho^\sigma} \right] = \sum_{ij} \kappa_{ij\sigma} \left( \frac{1}{2}\delta_{ij} - \operatorname{Re}\lambda_{ij\sigma} \right) \lvert\phi_{i\sigma} \rangle\langle\phi_{j\sigma} \rvert. \tag*{(19)}

Diagonalizing (hσ+Δhσ)(h^\sigma+ \Delta h^\sigma) yields corrected canonical orbitals ∣ψ~nσ⟩\lvert\widetilde{\psi}_{n\sigma} \rangle and orbital energies ε~nσ\widetilde{\varepsilon}_{n\sigma}. These eigenvalues are interpreted as quasiparticle energies that produce the LOSC-corrected band structure, including the fundamental gap [7] [88, 89, 90].

Results and Discussion

There are two parameters in olLOSC: the balance between spatial and energy localization γ\gamma and the fraction λ\lambda of von Weizsäcker kinetic energy. We found that γ=0.30\gamma= 0.30 and λ=0.75\lambda= 0.75 gave the best results when comparing between molecules and materials. Increasing the fraction λ\lambda of von Weizsäcker kinetic energy tends to increase the fundamental gap slightly (decreases the screened response ⟨δρ∣V⟩\langle\delta\rho\vert V \rangle), while increasing the spatial delocalization γ\gamma has the opposite effect. We note that most previous implementations of the LOSC method use γ=0.47714\gamma= 0.47714, but in some cases this decreased spatial localization leads the orbitalets of materials to become trapped in local minima (see the Supporting Information for details). All results in this section use (γ,λ)=(0.30,0.75)(\gamma, \lambda) = (0.30, 0.75). Underlying density functional calculations are performed with the PBE functional [11].

We tested the molecular implementation of olLOSC on the fundamental gaps of small and large molecules and on the ionization potentials of a set of polymers. The small molecules are a 17-member subset of the G2/97 test set [91] such that εPBELUMO≤−2\varepsilon_{\mathrm{PBE}}^{\mathrm{LUMO}} \leq-2 eV. The reference values for small molecules are fundamental gaps calculated by CCSD(T) in the Supporting Information of [68]. Large molecules are organic acceptors with reference CCSD(T) gaps from [92]. The polymers we study are polyacetylene, H(C2_2H2_2)n_nH, where 1≤n≤101 \leq n \leq10; reference ionization potentials are computed with RASPT2 in [93].

The bulk implementation was tested on a set of thirteen semiconductors and large-gap insulators, with lattice constants given by the experimental values found in [37]. The reference values are electronic gaps—experimental gaps adjusted for lattice-coupling effects via zero-point renormalization [94–96];5 they range from less than 1 to more than 21 eV.

As is clear from Fig. 1a, olLOSC provides substantial improvement in molecular ionization potentials and fundamental gaps compared with PBE. With the chosen parameters, we see little systematic error and a greatly improved absolute error. For example, the ionization potential of polyacetylene oligomers (triangles)

DFA and olLOSC fundamental gaps in molecules and materials

Figure 1. DFA (blue) and olLOSC (orange) fundamental gaps (ionization potentials for polyacetylenes), in (a) molecules and polymers, (b) materials. Reference values are from (a) CCSD(T) and RASPT2, (b) experiments, corrected for zero-point renormalization. The orbitalet space/energy mixing parameter γ=0.30\gamma= 0.30, and the kinetic energy functional is TF + 75% vW (λ=0.75\lambda= 0.75).

MSE (eV)MAE (eV)MARE (%)
Molecules
Small0.5280.5625.48
Large−0.2510.3524.18
All0.0800.4414.73
H(C2_2H2_2)n_nHPolymers (IP only)
(1 ≤n≤\le n \le 10)0.0640.1251.52
Materials
Gap ≤8\le8 eV−0.1180.1605.77
Gap > 8 eV−2.7422.74220.6
All−1.3291.35212.6

Table 1. Relative errors from olLOSC (γ=0.30\gamma= 0.30, λ=0.75\lambda= 0.75) of gaps (molecules, materials) and ionization potentials (polymers). Materials’ experimental gaps are renormalized for zero-point energy. Mean signed errors (MSE) and mean absolute errors (MAE) are in eV; mean absolute relative errors (MARE) are percentages.

are both more accurate and better follow the true trendline as their length is increased. Both small and large molecules have far better band gaps. We note a slight overestimation in the gap of molecules with especially large band gaps, in contrast to the situation for bulk insulators (see below).

The band gaps of semiconductors and moderately gapped insulators, up to about 8 eV, are similarly improved (Fig. 1b). For large-gapped insulators (in this dataset, LiCl, NaF, LiF, Ar, and Ne), band gaps are systematically underestimated, although the olLOSC predictions are still better than those of PBE.

The behavior of olLOSC in large-gapped systems differs between molecules and materials. In molecules, we see a small overestimation of the gap; in materials, the opposite effect is observed. Thus, we cannot tune the olLOSC parameters to eliminate the inaccuracy; at best, we can achieve a compromise. We attribute the difference to the limited accuracy of the relatively simple Thomas–Fermi–von Weizsäcker kinetic energy functional. Investigating more sophisticated functionals, such as those discussed in [97], is a promising avenue for future work. However, the TFvW functional is sufficiently accurate in the regime of interest for interfacial systems: inorganic and organic molecules, especially those with modest gaps, and semiconductors. We also expect that olLOSC with fTFvWf_{\mathrm{TFvW}} will perform adequately when extended to metals, which must account for perturbations to the Fermi energy [98, section II.C.4].

Conclusion

olLOSC corrects delocalization error consistently across system sizes, from small molecules to bulk semiconductors and insulators (with small and moderate band gaps). Besides this, the LOSC family has three crucial ingredients, each necessary for accurately calculating interfaces with DFT: correcting the band structure or quasiparticle energies [69, 70, 72]; modifying the total energy size-consistently [25] [68]; and correcting the charge density [46]. It is equally important to offer corrections to the total energy to describe chemical reactions occurring at interfaces. More fundamentally, approximate density functional calculations with delocalization error are not size-consistent [25], an error that could be exacerbated in interfaces. Finally, changes to the charge density are necessary to capture the charge transfer that can occur across an interface [99–103]. To the best of our knowledge, LOSC is the only method that can satisfy all three requirements in a unified approximation. In future work, we will explore the impact of other kinetic energy functionals [97] and add the long-wavelength correction to the Fermi energy in metals. Finally, we will implement the procedure of [46] in olLOSC to obtain a self-consistently optimized density. We expect that olLOSC will develop into a robust and efficient DFT approach for modeling interfaces between molecules and surfaces.

Materials and methods

Curvature implementation: Molecules

In molecular systems, we discretize χTFvW\chi_{\mathrm{TFvW}} and χol\chi_{\mathrm{ol}} in a low-rank function space inspired by the resolution-of-the-identity (RI) approximation [82, 104–106]; in the resulting auxiliary basis, we can invert them directly. We call our auxiliary, atom-centered basis functions Pσ(r)P_\sigma(\mathbf{r}). In the PσP_\sigma basis, the matrix elements of the orbital-free kinetic energy kernel are

fTFvWPσ,Qτ=[χTFvWPσ,Qτ]−1=⟨Pσ∣fTFvWστ∣Qτ⟩.(20)f_{\mathrm{TFvW}}^{P_\sigma,Q_\tau}=\left[\chi_{\mathrm{TFvW}}^{P_\sigma,Q_\tau}\right]^{-1}=\langle P_\sigma|f_{\mathrm{TFvW}}^{\sigma\tau}|Q_\tau\rangle. \tag*{(20)}

The Hartree kernel is similarly

fHPσ,Qτ=⟨Pσ∣fHστ∣Qτ⟩=∬dr dr′Pσ(r)Qτ(r′)∣r−r′∣;(21)f_{\mathrm{H}}^{P_\sigma,Q_\tau}=\langle P_\sigma|f_{\mathrm{H}}^{\sigma\tau}|Q_\tau\rangle=\iint\mathrm{d}r\,\mathrm{d}r'\frac{P_\sigma(r)Q_\tau(r')}{|r-r'|}; \tag*{(21)}

these matrix elements are computed readily with the RI-V machinery of [82]. Applying this discretization and the partial RPA to solve (11), we obtain the orbital-free linear-response function

χolPσ,Qτ=(fHPσ,Qτ+fTFvWPσ,Qτ)−1.(22)\chi_{\mathrm{ol}}^{P_\sigma,Q_\tau}=\left(f_{\mathrm{H}}^{P_\sigma,Q_\tau}+f_{\mathrm{TFvW}}^{P_\sigma,Q_\tau}\right)^{-1}. \tag*{(22)}

As mentioned before, we must impose charge conservation in χ\chi. However, given χolPσ,Qτ\chi_{\mathrm{ol}}^{P_\sigma,Q_\tau}, we can conserve charge without computing the chemical potential perturbation δμσ\delta\mu^\sigma directly. The total charge of Pσ(r)P_\sigma(r) is

dPσ=∫dr Pσ(r);(23)d_{P_\sigma}=\int\mathrm{d}r\,P_\sigma(r); \tag*{(23)}

then, following [84], the charge-conserving linear-response function χˉ\bar{\chi} has matrix elements

⟨Pσ∣χˉ∣Qτ⟩=χˉolPσ,Qτ=∑RSχolPσ,RτdRτdSσχolSσ,Qτ∑RSdRσχolRσ,SτdSτ−χolPσ,Qτ.(24)\langle P_\sigma|\bar{\chi}|Q_\tau\rangle=\bar{\chi}_{\mathrm{ol}}^{P_\sigma,Q_\tau}=\frac{\sum_{RS}\chi_{\mathrm{ol}}^{P_\sigma,R_\tau}d_{R_\tau}d_{S_\sigma}\chi_{\mathrm{ol}}^{S_\sigma,Q_\tau}}{\sum_{RS}d_{R_\sigma}\chi_{\mathrm{ol}}^{R_\sigma,S_\tau}d_{S_\tau}}-\chi_{\mathrm{ol}}^{P_\sigma,Q_\tau}. \tag*{(24)}

The olLOSC curvature is thus computed as

κijσol=⟨ρiσ∣fHxcσσ∣ρjσ⟩+∑Pμ,QνfHxciσ,PμχˉolPμ,QνfHxcjσ,Qν,(25)\kappa_{ij\sigma}^{\mathrm{ol}}=\langle\rho_{i\sigma}|f_{\mathrm{Hxc}}^{\sigma\sigma}|\rho_{j\sigma}\rangle+\sum_{P_\mu,Q_\nu}f_{\mathrm{Hxc}}^{i\sigma,P_\mu}\bar{\chi}_{\mathrm{ol}}^{P_\mu,Q_\nu}f_{\mathrm{Hxc}}^{j\sigma,Q_\nu}, \tag*{(25)}

where fHxciσ,Pμ=⟨ρiσ∣fHxcσμ∣Pμ⟩f_{\mathrm{Hxc}}^{i\sigma,P_\mu}=\langle\rho_{i\sigma}|f_{\mathrm{Hxc}}^{\sigma\mu}|P_\mu\rangle is a Hxc kernel matrix element between the orbitalet and auxiliary bases.

Curvature implementation: Materials

Following [44] [80], we decompose

κijσol=⟨ρiσ∣fHxc∣ρjσ⟩+∑μν⟨ρiσ∣fHxcσμχolμνfHxcνσ∣ρjσ⟩=⟨ρiσ∣Vjσ⟩+∑ν⟨δρiν∣Vjν⟩,(26)\begin{aligned} \kappa_{ij\sigma}^{\mathrm{ol}}&=\langle\rho_{i\sigma}|f_{\mathrm{Hxc}}|\rho_{j\sigma}\rangle+\sum_{\mu\nu}\langle\rho_{i\sigma}|f_{\mathrm{Hxc}}^{\sigma\mu}\chi_{\mathrm{ol}}^{\mu\nu}f_{\mathrm{Hxc}}^{\nu\sigma}|\rho_{j\sigma}\rangle\\ &=\langle\rho_{i\sigma}|V_{j\sigma}\rangle+\sum_\nu\langle\delta\rho_{i\nu}|V_{j\nu}\rangle, \tag*{(26)} \end{aligned}

where the potential ∣Vjν⟩\lvert V_{j\nu}\rangle and screened density response ∣δρiν⟩\lvert\delta\rho_{i\nu}\rangle are given by

Vjν(r)=∫dr′ fHxcνσ(r,r′)ρjσ(r′),V_{j\nu}(\mathbf{r})=\int d\mathbf{r}'\,f_{\mathrm{Hxc}}^{\nu\sigma}(\mathbf{r},\mathbf{r}')\rho_{j\sigma}(\mathbf{r}'),
δρiν(r)=∑μ∬dr′ dr′′ χolμν(r,r′)fHxcμσ(r′,r′′)ρiσ(r′′).(27)\delta\rho_{i\nu}(\mathbf{r})=\sum_{\mu}\iint d\mathbf{r}'\,d\mathbf{r}''\,\chi_{\mathrm{ol}}^{\mu\nu}(\mathbf{r},\mathbf{r}')f_{\mathrm{Hxc}}^{\mu\sigma}(\mathbf{r}',\mathbf{r}'')\rho_{i\sigma}(\mathbf{r}''). \tag*{(27)}

We compute ∣ρiσ⟩\lvert\rho_{i\sigma}\rangle and ∣Vjν⟩\lvert V_{j\nu}\rangle directly. But where ∣δρiν⟩\lvert\delta\rho_{i\nu}\rangle is computed in KCW [44] and lrLOSC via perturbations to the occupied Kohn–Sham orbitals, we obtain it more directly in olLOSC. We identify the perturbation δρν\delta\rho^\nu of (14) with ∣δρiν⟩\lvert\delta\rho_{i\nu}\rangle, yielding the linear system

\begin{pmatrix} f_{\mathrm{TFvW}}^{\alpha\alpha}+f_{\mathrm{H}}^{\alpha\alpha} & f_{\mathrm{H}}^{\alpha\beta} \\ f_{\mathrm{H}}^{\beta\alpha} & f_{\mathrm{TFvW}}^{\beta\beta}+f_{\mathrm{H}}^{\beta\beta} \end{pmatrix} \begin{pmatrix} \lvert\delta\rho_{i\alpha}\rangle\\ \lvert\delta\rho_{i\beta}\rangle \end{pmatrix} = \begin{pmatrix} \delta\mu^\alpha-\lvert\delta v^\alpha\rangle\\ \delta\mu^\beta-\lvert\delta v^\beta\rangle \end{pmatrix}. \tag*{(28)}\end{pmatrix}

(Without the partial RPA, we would replace fHf_{\mathrm{H}} by fHxcf_{\mathrm{Hxc}}.) Choosing δρν=∣δρiν⟩\delta\rho^\nu=\lvert\delta\rho_{i\nu}\rangle also fixes the perturbation to the external potential, so that δvτ=∣Viτ⟩\delta v^\tau=\lvert V_{i\tau}\rangle.

The matrix on the left-hand side of (28) is nonlocal in either real or reciprocal space, so it is too large even to store directly. We can, however, compute its action on a vector and use an iterative solver to obtain ∣δρiα⟩\lvert\delta\rho_{i\alpha}\rangle and ∣δρiβ⟩\lvert\delta\rho_{i\beta}\rangle simultaneously. Under the partial RPA, the matrix is positive semidefinite, so we solve it with the modified conjugate gradient algorithm of [107]. Instead of directly computing δμν\delta\mu^\nu, which changes each iteration, we impose charge conservation as a constraint via a projected preconditioner [107]. Without the partial RPA, a linear solver for indefinite matrices such as MINRES-QLP [108–110] would be required.

Computational details

olLOSC for molecules is implemented in the in-house code QM4DQM^{4}D, which uses Gaussian-type orbitals. Small and large molecules are computed with correlation-consistent triple-zeta Dunning basis sets augmented with diffuse functions (aug-cc-pVTZ) [111, 112]; the polymers use the same basis sets without augmentation (cc-pVTZ). We use (aug)-cc-pVTZ-RIFIT as the auxiliary basis for χol\chi_{\mathrm{ol}}.

olLOSC for materials is implemented as a module in a locally maintained fork [113] of the open-source Quantum ESPRESSO package [114, 115], version 7.2. The bulk orbitalets are computed with a fork [116] of wannier90 [117–119] version 3.1.0. Quantum ESPRESSO’s PWscf code is used for the underlying DFT calculations, which use the PBE functional [11] and optimized Vanderbilt norm-conserving pseudopotentials with scalar relativistic corrections [120] downloaded from the ABINIT PseudoDojo (http://www.pseudo-dojo.org/). We use a kinetic energy cutoff of 75 Ry; unless otherwise specified, we sample the Brillouin zone with a 6×6×66\times6\times6 Monkhorst–Pack (uniform) grid [121] centered at the origin Γ\Gamma of reciprocal space. The macroscopic dielectric constant ϵ∞\epsilon_\infty, needed for the Gygi–Baldereschi correction [122] to the Coulomb singularity, is computed by density functional perturbation theory [98] in the Phonon module of Quantum ESPRESSO, with the same parameters. The eigenvalues of the olLOSC Hamiltonian, used for corrected band structures, are computed along the paths specified by [123] for the systems’ respective lattices.

Computational complexity

Computing the curvature—in particular, the screened response—is the most computationally demanding part of a LOSC calculation. Thus, we restrict our complexity analysis to this bottleneck step.

In materials, the monochromatic decomposition means that the density response {∣δρiτ⟩}τ\{\lvert\delta\rho_{i\tau}\rangle\}_{\tau}, which encodes the screening, must be computed Nw×NkN_w \times N_k times, where NwN_w is the number of DLWFs and NkN_k is the number of k-points sampled in the Brillouin zone. Each calculation of ∣δρiτ⟩\lvert\delta\rho_{i\tau}\rangle requires a matrix inversion, computed by the projected preconditioned conjugate gradient [107]; each iteration scales linearly in the number of plane waves NGN_G (the length of ∣δρiτ⟩\lvert\delta\rho_{i\tau}\rangle) [110]. Each iteration also requires several conversions between real and reciprocal space (because the kinetic energy kernel is semilocal in r former, while the Coulomb kernel is local in G) via the fast Fourier transform, which scales as NGlog⁡NGN_G \log N_G. Thus, olLOSC for materials scales as O(NwNkNGlog⁡NG)O(N_w N_k N_G \log N_G). This saves a factor of NkN_k relative to lrLOSC [70]; even with the Sternheimer equation bypassing direct computation of χs\chi_s, constructing ∣δρiτ⟩\lvert\delta\rho_{i\tau}\rangle in lrLOSC requires coupled pairs of k-points.

In molecules, the interacting linear response function χ\chi is computed directly; the bottleneck step is inverting χ\chi and χol\chi_{\mathrm{ol}}. Within the orbital-free-approximation, all involved matrices are evaluated in a Naux×NauxN_{\mathrm{aux}} \times N_{\mathrm{aux}} basis of auxiliary functions, where NauxN_{\mathrm{aux}} is proportional to the number of orbitalets NwN_w. We invert χ\chi with an LU decomposition, so olLOSC for molecules scales as O(Nw3)O(N_w^3). In addition to the evaluation of the linear response function, the numerical integration of the exchange-correlation kernel can be a potential time-determining step because of its relatively large pre-factor. In olLOSC, this integral is constructed as shown in Equation 25; it scales as O(Nw2Ngrid)O(N_w^2 N_{\mathrm{grid}}), where NgridN_{\mathrm{grid}} is the number of grid points. lrLOSC, by comparison, requires the inversion of a rank-4 tensor instead of a rank-2 tensor because χs\chi_s (unlike χTFvW\chi_{\mathrm{TFvW}}) depends on the difference between occupied and virtual orbital energies.

Data Availability

The Python (molecules) and R (materials) analysis and figure generation scripts, as well as the underlying data, are included in the Supporting Information. olLOSC for materials is implemented as a fork of Quantum ESPRESSO and is available at [113].

Acknowledgements

We are grateful to Yang Shen for discussion about the positive semidefiniteness of ∣δρiν⟩\lvert\delta\rho_{i\nu}\rangle under the partial RPA. We acknowledge financial support from the National Science Foundation (Grant No. CHE-2154831) and the National Institutes of Health (Grant No. R35-GM158181).

Supporting Information for olLOSC: Unified and efficient density functional approximation to correct delocalization error in molecules and periodic materials

Yichen Fan; Jacob Z. Williams; Weitao Yang

supp.pdf

Additional notation for materials

Computing LOSC in periodic boundary conditions requires additional considerations (and more notation). We sample the complex Kohn–Sham orbitals at NkN_k different points k\mathbf{k} in the irreducible Brillouin zone. This yields an additional index kk to the canonical orbitals. They also obey Bloch’s theorem [1],

ψknσ(r)=eik⋅ruknσ(r),(29)\psi_{kn\sigma}(\mathbf{r})=e^{i\mathbf{k}\cdot\mathbf{r}}u_{kn\sigma}(\mathbf{r}), \tag*{(29)}

where uknσ(r)u_{kn\sigma}(\mathbf{r}) shares the translational symmetry of the underlying lattice.

The bulk orbitalets must also obey lattice’s translational symmetry. This is shown mathematically by localizing with a different unitary matrix UkσU^{k\sigma} at each k\mathbf{k}-point; the orbitalets are given by

∣ϕkiσ⟩=∑nUinkσ∣ψknσ⟩,(30)\lvert\phi_{ki\sigma}\rangle=\sum_n U_{in}^{k\sigma}\lvert\psi_{kn\sigma}\rangle, \tag*{(30)}

the orbitals and do not mix at different k\mathbf{k}. To obtain a more natural representation in real space, we transform to the (generalized) Wannier representation [2], yielding dually localized Wannier functions (DLWFs) [3]

∣wRiσ⟩=1Nk∑ke−ik⋅R∣ϕiσ⟩.(31)\lvert w_{Ri\sigma}\rangle=\frac{1}{N_k}\sum_k e^{-i\mathbf{k}\cdot\mathbf{R}}\lvert\phi_{i\sigma}\rangle. \tag*{(31)}

By the same token, maximally localized Wannier functions [4] are not an exact analogue of the molecular Foster–Boys orbitals [5]; loosening the translational invariance constraint could, in principle, yield localized orbitals with smaller spatial variance. (Of course, the periodic structure of Wannier functions is a major part of their usefulness and appeal.) Note that, in molecules and materials sampled only at Γ\Gamma, Nk=1N_k=1 and k=0\mathbf{k}=0. This renders the Fourier transformation of (3) trivial, and ∣ϕ0iσ⟩=∣ϕiσ⟩\lvert\phi_{0i\sigma}\rangle=\lvert\phi_{i\sigma}\rangle.

The monochromatic decomposition

It is well known [6, 7] that linear operators OO that are periodic on a supercell consisting of NkN_k uniformly sampled primitive unit cells—or, equivalently, on NkN_k k\mathbf{k}-points sampled in the Brillouin zone—can be decomposed into monochromatic components, each periodic on the primitive cell:

O(r,r′)=∑qeiq⋅rOq(r,r′)e−iq⋅r′.(32)O(\mathbf{r},\mathbf{r}')=\sum_{\mathbf{q}}e^{i\mathbf{q}\cdot\mathbf{r}}O^{\mathbf{q}}(\mathbf{r},\mathbf{r}')e^{-i\mathbf{q}\cdot\mathbf{r}'}. \tag*{(32)}

(For the weights of each OqO^{\mathbf{q}} to be equal, the Brillouin zone must be sampled uniformly [8], and the origin Γ=0\Gamma= 0 of reciprocal space must be one of the k-points.)

This decomposition applies to one-index quantities such as Bloch orbitals (including their perturbations, useful for density functional perturbation theory); indeed, it is nothing more than a statement of Bloch’s theorem [1]. It also applies to Wannier functions, with

wRiσ(r)=1Nk∑qeiq⋅rwRiσq(r);(33)w_{\mathbf{R}i\sigma}(\mathbf{r}) = \frac{1}{N_k} \sum_{\mathbf{q}} e^{i\mathbf{q}\cdot\mathbf{r}} w_{\mathbf{R}i\sigma}^{\mathbf{q}}(\mathbf{r}); \tag*{(33)}

the factor of 1/Nk1/N_k is a normalization convention.

More recently, Colonna et al. found in [9] that the density of a generalized Wannier function (such as a LOSC DLWF) also has this property:

ρRiσ(r)=∣wRiσ(r)∣2=1Nk∑qeiq⋅rρRiσq(r),(34)\rho_{\mathbf{R}i\sigma}(\mathbf{r}) = |w_{\mathbf{R}i\sigma}(\mathbf{r})|^2 = \frac{1}{N_k} \sum_{\mathbf{q}} e^{i\mathbf{q}\cdot\mathbf{r}} \rho_{\mathbf{R}i\sigma}^{\mathbf{q}}(\mathbf{r}), \tag*{(34)}

where

wRiσq(r)=1Nke−iq⋅R∑kϕkiσ∗(r)ϕ(k+q)iσ(r).(35)w_{\mathbf{R}i\sigma}^{\mathbf{q}}(\mathbf{r}) = \frac{1}{N_k} e^{-i\mathbf{q}\cdot\mathbf{R}} \sum_{\mathbf{k}} \phi_{\mathbf{k}i\sigma}^{*}(\mathbf{r})\phi_{(\mathbf{k}+\mathbf{q})i\sigma}(\mathbf{r}). \tag*{(35)}

Observe also that

∣ρRiσ⟩=1Nk∑qe−iq⋅R∣ρ0iσq⟩.(36)|\rho_{\mathbf{R}i\sigma}\rangle= \frac{1}{N_k} \sum_{\mathbf{q}} e^{-i\mathbf{q}\cdot\mathbf{R}}|\rho_{0i\sigma}^{\mathbf{q}}\rangle. \tag*{(36)}

Thus, we only compute quantities for the home unit cell R=0\mathbf{R} = 0 explicitly; for the rest, we require only the phase factors e−iq⋅Re^{-i\mathbf{q}\cdot\mathbf{R}}. Because the LOSC curvature operator is linear in the orbital et densities,

κRijσ=⟨ρ0iσ∣fHxcσσ+∑ντfHxcσνχντfHxcτσ∣ρRjσ⟩,(37)\kappa_{\mathbf{R}ij\sigma} = \langle\rho_{0i\sigma}|f_{\mathrm{Hxc}}^{\sigma\sigma} + \sum_{\nu\tau} f_{\mathrm{Hxc}}^{\sigma\nu}\chi^{\nu\tau}f_{\mathrm{Hxc}}^{\tau\sigma}|\rho_{\mathbf{R}j\sigma}\rangle, \tag*{(37)}

we can apply the monochromatic decomposition wholesale, obtaining

κRijσ=∑qe−iq⋅R⟨ρ0iσq∣fHxcq;σσ+∑ντfHxcq;σνχq;ντfHxcq;τσ∣ρ0jσq⟩=∑qe−iq⋅Rκ0ijσq.(38)\begin{aligned} \kappa_{\mathbf{R}ij\sigma} &= \sum_{\mathbf{q}} e^{-i\mathbf{q}\cdot\mathbf{R}}\langle\rho_{0i\sigma}^{\mathbf{q}}|f_{\mathrm{Hxc}}^{\mathbf{q};\sigma\sigma} + \sum_{\nu\tau} f_{\mathrm{Hxc}}^{\mathbf{q};\sigma\nu}\chi^{\mathbf{q};\nu\tau}f_{\mathrm{Hxc}}^{\mathbf{q};\tau\sigma}|\rho_{0j\sigma}^{\mathbf{q}}\rangle\\ &= \sum_{\mathbf{q}} e^{-i\mathbf{q}\cdot\mathbf{R}}\kappa_{0ij\sigma}^{\mathbf{q}}. \tag*{(38)} \end{aligned}

The monochromatic decomposition κRijσ=∑qe−iq⋅Rκ0ijσq\kappa_{\mathbf{R}ij\sigma} = \sum_{\mathbf{q}} e^{-i\mathbf{q}\cdot\mathbf{R}}\kappa_{0ij\sigma}^{\mathbf{q}} provides a substantial improvement to computational scaling. Naïvely, quantities like κ\kappa and ∣ρRiσ⟩|\rho_{\mathbf{R}i\sigma}\rangle are periodic on the supercell real- and reciprocal-space grids, which have NkNGN_kN_G elements (where NGN_G is the size of the primitive cell grid). Matrix operations on this space scale as O(Nk3NG3)O(N_k^3N_G^3). The monochromatic decomposition allows us to work with quantities periodic on the primitive cell, with matrix operations scaling as O(NG3)O(N_G^3). Each supercell quantity decomposes into Nk=NqN_k = N_q primitive-cell quantities, which means that computing each curvature element κ0ijσq\kappa_{0ij\sigma}^{\mathbf{q}} scales as O(NkNG3)O(N_kN_G^3) overall (up to logarithmic factors due to the fast Fourier transform), saving a factor of Nk2N_k^2. In lrLOSC, the density functional perturbation theory approach [6] uses the Sternheimer equation [10] to compute χντfHxcτσ∣ρ0jσq⟩\chi^{\nu\tau}f_{\mathrm{Hxc}}^{\tau\sigma}|\rho_{0j\sigma}^{\mathbf{q}}\rangle via the first-order variations ∣δψ(k+q)nσ⟩|\delta\psi_{(\mathbf{k}+\mathbf{q})n\sigma}\rangle of the Kohn–Sham orbitals ∣ψknσ⟩|\psi_{\mathbf{k}n\sigma}\rangle; the reciprocal-space points k\mathbf{k} and k+q\mathbf{k}+\mathbf{q} are paired, so computing κ0ijσq\kappa_{0ij\sigma}^{\mathbf{q}} in lrLOSC requires O(Nk2NG3)O(N_k^2N_G^3) effort, saving only one factor of NkN_k. For more details on the monochromatic decomposition, see [9] and the Supplemental Material of [11].

Implementation of the olLOSC curvature

The Thomas–Fermi–von Weizsäcker kernel

The Thomas–Fermi kinetic energy is

TTF[ρ]=cF∫dr ρ(r)5/3,(39)T_{\mathrm{TF}}[\rho] = c_F \int d\mathbf{r}\,\rho(\mathbf{r})^{5/3}, \tag*{(39)}

where cF=(3/10)(3π2)2/3c_{\mathrm{F}}=(3/10)(3\pi^2)^{2/3} is the Fermi constant. The variation in TTFT_{\mathrm{TF}} with respect to ρσ\rho^\sigma is

δTTFδρσ(r)=53cFρσ(r)2/3,(40)\frac{\delta T_{\mathrm{TF}}}{\delta\rho^\sigma(\mathbf{r})}=\frac{5}{3}c_{\mathrm{F}}\rho^\sigma(\mathbf{r})^{2/3}, \tag*{(40)}

and the kernel is

δ2TTFδρσ(r)δρτ(r′)=22/310cF9ρσ(r)−1/3δστδ(r−r′),(41)\frac{\delta^2 T_{\mathrm{TF}}}{\delta\rho^\sigma(\mathbf{r})\delta\rho^\tau(\mathbf{r}')}=2^{2/3}\frac{10c_{\mathrm{F}}}{9}\rho^\sigma(\mathbf{r})^{-1/3}\delta_{\sigma\tau}\delta(\mathbf{r}-\mathbf{r}'), \tag*{(41)}

the factor of 22/32^{2/3} is omitted for a spinless calculation [12]. The expression for the von Weizsäcker kernel is known, e.g. Eq. (A10) of [13] and Eq. (97) of [14], although the derivation is very long. In Rydberg units, (ℏ=1\hbar=1, but the electron mass m=1/2m=1/2), it is

δ2TvWδρσ(r)δρσ(r′)=14δvσ(r)δρσ(r′)=12[∇ρσ(r′)⋅∇δ(r−r′)ρσ(r′)+∇2ρσ(r′)ρσ(r′)2δ(r−r′)−∇ρσ(r′)⋅∇ρσ(r′)ρσ(r′)3δ(r−r′)−∇2δ(r−r′)ρσ(r′)].(42)\begin{aligned} \frac{\delta^2 T_{\mathrm{vW}}}{\delta\rho_\sigma(\mathbf{r})\delta\rho_\sigma(\mathbf{r}')}=\frac{1}{4}\frac{\delta v_\sigma(\mathbf{r})}{\delta\rho_\sigma(\mathbf{r}')} \\ &=\frac{1}{2}\left[\frac{\nabla\rho_\sigma(\mathbf{r}')\cdot\nabla\delta(\mathbf{r}-\mathbf{r}')}{\rho_\sigma(\mathbf{r}')}+\frac{\nabla^2\rho_\sigma(\mathbf{r}')}{\rho_\sigma(\mathbf{r}')^2}\delta(\mathbf{r}-\mathbf{r}')-\frac{\nabla\rho_\sigma(\mathbf{r}')\cdot\nabla\rho_\sigma(\mathbf{r}')}{\rho_\sigma(\mathbf{r}')^3}\delta(\mathbf{r}-\mathbf{r}')-\frac{\nabla^2\delta(\mathbf{r}-\mathbf{r}')}{\rho_\sigma(\mathbf{r}')}\right]. \tag*{(42)} \end{aligned}

Together, we have

fTFvWστ(r,r′)=fTFvWσ(r,r′)δστ=[fTFσ(r)δ(r−r′)+λfvWσ(r,r′)]δστ,(43)f_{\mathrm{TFvW}}^{\sigma\tau}(\mathbf{r},\mathbf{r}')=f_{\mathrm{TFvW}}^{\sigma}(\mathbf{r},\mathbf{r}')\delta_{\sigma\tau}=[f_{\mathrm{TF}}^\sigma(\mathbf{r})\delta(\mathbf{r}-\mathbf{r}')+\lambda f_{\mathrm{vW}}^\sigma(\mathbf{r},\mathbf{r}')]\delta_{\sigma\tau}, \tag*{(43)}

where λ≥0\lambda\ge0 is the amount of von Weizsäcker correction. Note that fvWf_{\mathrm{vW}} is semilocal; its action on ∣δρ⟩|\delta\rho\rangle at r\mathbf{r} depends on ∣δρ⟩|\delta\rho\rangle, ∇∣δρ⟩\nabla|\delta\rho\rangle, and ∇2∣δρ⟩\nabla^2|\delta\rho\rangle, but still only requires them to be evaluated at r\mathbf{r}.

olLOSC in molecules

Molecules have few enough basis functions that we can almost invert χ\chi in (9) directly. The last step is provided by the resolution-of-the-identity (RI) approximation [15, 16] with auxiliary basis functions. Within the RI approximation, the four-center Coulomb integral

(ij∣kl)=∬dr dr′ψi∗(r)ψj(r)ψk∗(r′)ψl(r′)∣r−r′∣,(44)(ij|kl)=\iint d\mathbf{r}\,d\mathbf{r}'\frac{\psi_i^*(\mathbf{r})\psi_j(\mathbf{r})\psi_k^*(\mathbf{r}')\psi_l(\mathbf{r}')}{|\mathbf{r}-\mathbf{r}'|}, \tag*{(44)}

can be evaluated in terms of the products of three-center and two-center integrals:

(ij∣kl)≈∑μνCijμVμν−1Cklν.(45)(ij|kl)\approx\sum_{\mu\nu}C_{ij}^{\mu}V_{\mu\nu}^{-1}C_{kl}^{\nu}. \tag*{(45)}

Here, we define

Vμν=∬drdr′Pμ(r)Pν(r′)∣r−r′∣,(46)V_{\mu\nu}=\iint d\mathbf{r}d\mathbf{r}'\frac{P_\mu(\mathbf{r})P_\nu(\mathbf{r}')}{|\mathbf{r}-\mathbf{r}'|}, \tag*{(46)}

and

Cijμ=∑ν(ij∣ν)Vμν−1;(47)C_{ij}^{\mu}=\sum_{\nu}(ij|\nu)V_{\mu\nu}^{-1}; \tag*{(47)}

here,

(ij∣ν)=∬dr dr′ψi∗(r)ψj(r)Pν(r′)∣r−r′∣.(48)(ij|\nu)=\iint d\mathbf{r}\,d\mathbf{r}'\frac{\psi_i^*(\mathbf{r})\psi_j(\mathbf{r})P_\nu(\mathbf{r}')}{|\mathbf{r}-\mathbf{r}'|}. \tag*{(48)}

μ,ν=1,2,…Naux\mu,\nu=1,2,\ldots N_{\mathrm{aux}} label the auxiliary basis functions PμP_\mu.

We also use the same set of auxiliary basis functions to discretize the Thomas–Fermi–von Weizsäcker (TFvW) kinetic energy kernel introduced in Equation 14, following the discussion in York and Yang [17]. In our molecular implementation, we evaluate the TFvW kernel in the auxiliary basis and store the resulting coefficients, which are

fμν,σTFvW=∬drdr′Pμ(r)δ2(TTF+λTvW)δρσ(r)δρσ(r′)Pν(r′).(49)f_{\mu\nu,\sigma}^{\mathrm{TFvW}}=\iint d\mathbf{r}d\mathbf{r}'P_\mu(\mathbf{r})\frac{\delta^2(T_{\mathrm{TF}}+\lambda T_{\mathrm{vW}})}{\delta\rho_\sigma(\mathbf{r})\delta\rho_\sigma(\mathbf{r}')}P_\nu(\mathbf{r}'). \tag*{(49)}

In the random phase approximation, the inverse of the many-body static linear response function χ\chi is

χστ−1(r,r′)=δVσ(r)δρτ(r′)=δ(Tσδστ+VHσ)δρτ(r′).(50)\chi^{-1}_{\sigma\tau}(\mathbf{r},\mathbf{r}')=\frac{\delta V^{\sigma}(\mathbf{r})}{\delta\rho^{\tau}(\mathbf{r}')}=\frac{\delta\left(T^{\sigma}\delta_{\sigma\tau}+V_{\mathrm{H}}^{\sigma}\right)}{\delta\rho^{\tau}(\mathbf{r}')} . \tag*{(50)}

We can evaluate the matrix elements of χ−1\chi^{-1} as the sum of the kinetic kernel (21) and Hartree kernel (18). We also implement a beyond-RPA olLOSC scheme, which adds VxcσV_{\mathrm{xc}}^{\sigma} to the right-hand side of (22). For results on this, see 4.3.

olLOSC in materials

To compute the screened density response ∣δρ0iτ⟩\lvert\delta\rho_{0i\tau}\rangle in materials, we solve the coupled linear equations

\begin{pmatrix} f_{\mathrm{T}}^{\alpha\alpha}+f_{\mathrm{H}}^{\alpha\alpha} & f_{\mathrm{T}}^{\alpha\beta}+f_{\mathrm{H}}^{\alpha\beta} \\ f_{\mathrm{T}}^{\beta\alpha}+f_{\mathrm{H}}^{\beta\alpha} & f_{\mathrm{T}}^{\beta\beta}+f_{\mathrm{H}}^{\beta\beta} \end{pmatrix} \begin{pmatrix} \lvert\delta\rho_{0i\alpha}\rangle\\ \lvert\delta\rho_{0i\beta}\rangle \end{pmatrix} = \begin{pmatrix} \delta\mu^{\alpha}-\lvert\delta V_{\mathrm{R}i\alpha}\rangle\\ \delta\mu^{\beta}-\lvert\delta V_{\mathrm{R}i\beta}\rangle \end{pmatrix}. \tag*{(51)}\end{pmatrix}

Because the Thomas–Fermi–von Weizsäcker kernel is diagonal in spin, this reduces to

\begin{pmatrix} f_{\mathrm{TFvW}}^{\alpha\alpha}+f_{\mathrm{H}}^{\alpha\alpha} & f_{\mathrm{H}}^{\alpha\beta} \\ f_{\mathrm{H}}^{\beta\alpha} & f_{\mathrm{TFvW}}^{\beta\beta}+f_{\mathrm{H}}^{\beta\beta} \end{pmatrix} \begin{pmatrix} \lvert\delta\rho_{0i\alpha}\rangle\\ \lvert\delta\rho_{0i\beta}\rangle \end{pmatrix} = \begin{pmatrix} \delta\mu^{\alpha}-\lvert\delta V_{0i\alpha}\rangle\\ \delta\mu^{\beta}-\lvert\delta V_{0i\beta}\rangle \end{pmatrix}. \tag*{(52)}\end{pmatrix}

Formally, we solve (24) by inverting the left-hand matrix MolM_{\mathrm{ol}}. However, it is nonlocal on either the real-space or reciprocal-space grid, and the number of grid points (plane waves) can exceed 10610^{6}. We therefore use a Krylov subspace algorithm to invert it iteratively. Under the partial RPA, MolM_{\mathrm{ol}} is Hermitian and positive semidefinite; as we show in 2.4, ⟨x∣Mol∣x⟩≥0\langle x\lvert M_{\mathrm{ol}}\rvert x\rangle\ge0 for any vector ∣x⟩=(∣x⟩α ∣x⟩β)T\lvert x\rangle=(\lvert x\rangle^{\alpha}\ \lvert x\rangle^{\beta})^{\mathrm{T}}, so we solve for ∣δρ0iσ⟩\lvert\delta\rho_{0i\sigma}\rangle with the conjugate gradient algorithm.

Instead of having to recalculate the charge conservation terms δμσ\delta\mu^{\sigma} every iteration, we account for them implicitly by constraining the optimization. Including δμσ\delta\mu^{\sigma} on the right side of (23) is equivalent to requiring

∫dr δρ0iα(r)=∫dr δρ0iβ(r)=0.(53)\int d\mathbf{r}\,\delta\rho_{0i\alpha}(\mathbf{r})=\int d\mathbf{r}\,\delta\rho_{0i\beta}(\mathbf{r})=0. \tag*{(53)}

We implement the projected preconditioning method of [18], using a preconditioner PP that preserves the constraint:

P=(−χTFvW;HEGα(G)00−χTFvW;HEGβ(G)),(54)P=\begin{pmatrix} -\chi_{\mathrm{TFvW;HEG}}^{\alpha}(\mathbf{G}) & 0 \\ 0 & -\chi_{\mathrm{TFvW;HEG}}^{\beta}(\mathbf{G}) \end{pmatrix}, \tag*{(54)}

where

χTFvW;HEGσ(G)=−[fTFvW;HEGσ]−1(G)=−kFmeπ2(1+3λη2).(55)\chi_{\mathrm{TFvW;HEG}}^{\sigma}(\mathbf{G})=-\left[f_{\mathrm{TFvW;HEG}}^{\sigma}\right]^{-1}(\mathbf{G})=-\frac{k_{\mathrm{F}}}{m_{\mathrm{e}}\pi^{2}(1+3\lambda\eta^{2})}. \tag*{(55)}

is the Thomas–Fermi–von Weizsäcker kernel for the homogeneous electron gas. kF=(3π2ρˉσ)1/3k_{\mathrm{F}}=(3\pi^{2}\bar{\rho}^{\sigma})^{1/3} is the Fermi wavevector given the average density of spin σ\sigma in one unit cell,

ρˉσ=∫UCdr ρσ(r);\bar{\rho}^{\sigma}=\int_{\mathrm{UC}}d\mathbf{r}\,\rho^{\sigma}(\mathbf{r});

λ\lambda is the fraction of von Weizsäcker exchange in fTFvW=fTF+λfvWf_{\mathrm{TFvW}}=f_{\mathrm{TF}}+\lambda f_{\mathrm{vW}}; mem_{\mathrm{e}} is the mass of the electron, unity in atomic units but 1/2 in Rydberg units, favored by Quantum ESPRESSO; and η=∣G∣/kF\eta=\lvert\mathbf{G}\rvert/k_{\mathrm{F}} is the reduced wavevector magnitude. Note that fTFvWf_{\mathrm{TFvW}} (and χTFvW\chi_{\mathrm{TFvW}} are local in both spin and reciprocal space, with

fTFvWσν(G,G′)=fTFvWσ(G)δσνδGG′.f_{\mathrm{TFvW}}^{\sigma\nu}(\mathbf{G},\mathbf{G}')=f_{\mathrm{TFvW}}^{\sigma}(\mathbf{G})\delta_{\sigma\nu}\delta_{\mathbf{G}\mathbf{G}'}.

MolM_{\mathrm{ol}} is positive semidefinite under the partial RPA

We decompose MolM_{\mathrm{ol}} into three components:

Mol=(fTFα00fTFβ)+(fvWα00fvWβ)+(fHααfHαβfHβαfHββ)=:M1+M2+M3.(56)M_{\mathrm{ol}}= \begin{pmatrix} f_{\mathrm{TF}}^{\alpha} & 0 \\ 0 & f_{\mathrm{TF}}^{\beta} \end{pmatrix} + \begin{pmatrix} f_{\mathrm{vW}}^{\alpha} & 0 \\ 0 & f_{\mathrm{vW}}^{\beta} \end{pmatrix} + \begin{pmatrix} f_{\mathrm{H}}^{\alpha\alpha} & f_{\mathrm{H}}^{\alpha\beta} \\ f_{\mathrm{H}}^{\beta\alpha} & f_{\mathrm{H}}^{\beta\beta} \end{pmatrix} =:M_{1}+M_{2}+M_{3}. \tag*{(56)}

By linearity, if M1M_1, M2M_2, and M3M_3 are each positive semidefinite, then

⟨x∣Mol∣x⟩=⟨x∣M1∣x⟩+⟨x∣M2∣x⟩+⟨x∣M3∣x⟩≥0,(57)\langle x|M_{\mathrm{ol}}|x\rangle= \langle x|M_1|x\rangle+ \langle x|M_2|x\rangle+ \langle x|M_3|x\rangle\ge0, \tag*{(57)}

and so is MolM_{\mathrm{ol}}.

The Thomas–Fermi kernel

This is the simplest to show. The Thomas–Fermi kernel is

fTFσ(r)=δ2TTFδρσ(r)δρτ(r′)=10cF9[ρσ(r)]−1/3δστδ(r−r′),(58)f_{\mathrm{TF}}^\sigma(\mathbf{r}) = \frac{\delta^2 T_{\mathrm{TF}}}{\delta\rho_\sigma(\mathbf{r})\delta\rho_\tau(\mathbf{r}')} = \frac{10c_F}{9}[\rho_\sigma(\mathbf{r})]^{-1/3}\delta_{\sigma\tau}\delta(\mathbf{r}-\mathbf{r}'), \tag*{(58)}

where the constant cF=(3/10)(3nsπ2)2/3>0c_F = (3/10)(3n_s\pi^2)^{2/3} > 0. (The number of spin components is nsn_s, here ns=2n_s = 2; and in Rydberg units, where ℏ=1\hbar= 1 but the electron mass me=1/2m_e = 1/2, cFc_F must be multiplied by 2.)

Given any nonzero two-spin vector xx,

x†M1x=(⟨xα∣⟨xβ∣)(fTFα00fTFβ)(∣xα⟩∣xβ⟩)=⟨xα∣fTFα∣xα⟩+⟨xβ∣fTFβ∣xβ⟩=∬dr dr′ xα∗(r)fTFα(r,r′)xα(r′)+∬dr dr′ xβ∗(r)fTFβ(r,r′)xβ(r′)=10cF9∫dr(∣xα(r)∣2[ρα(r)]−1/3+∣xβ(r)∣2[ρβ(r)]−1/3)≥0,(59)\begin{aligned} x^\dagger M_1 x = \begin{pmatrix} \langle x_\alpha| & \langle x_\beta| \end{pmatrix} \begin{pmatrix} f_{\mathrm{TF}}^\alpha& 0 \\ 0 & f_{\mathrm{TF}}^\beta \end{pmatrix} \begin{pmatrix} |x_\alpha\rangle\\ |x_\beta\rangle \end{pmatrix} \\ &= \langle x_\alpha|f_{\mathrm{TF}}^\alpha|x_\alpha\rangle+ \langle x_\beta|f_{\mathrm{TF}}^\beta|x_\beta\rangle\\ &= \iint d\mathbf{r}\,d\mathbf{r}'\,x_\alpha^*(\mathbf{r})f_{\mathrm{TF}}^\alpha(\mathbf{r},\mathbf{r}')x_\alpha(\mathbf{r}') + \iint d\mathbf{r}\,d\mathbf{r}'\,x_\beta^*(\mathbf{r})f_{\mathrm{TF}}^\beta(\mathbf{r},\mathbf{r}')x_\beta(\mathbf{r}') \\ &= \frac{10c_F}{9}\int d\mathbf{r}\left(|x_\alpha(\mathbf{r})|^2[\rho_\alpha(\mathbf{r})]^{-1/3} + |x_\beta(\mathbf{r})|^2[\rho_\beta(\mathbf{r})]^{-1/3}\right) \ge0, \tag*{(59)} \end{aligned}

since ρσ(r)≥0\rho^\sigma(\mathbf{r}) \ge0.

The von Weizsäcker kernel

Like the Thomas–Fermi kernel, the vW kernel is diagonal in spin, so x†M2x=⟨xα∣fvWα∣xα⟩+⟨xβ∣fvWβ∣xβ⟩x^\dagger M_2x = \langle x_\alpha|f_{\mathrm{vW}}^\alpha|x_\alpha\rangle+ \langle x_\beta|f_{\mathrm{vW}}^\beta|x_\beta\rangle. It suffices to consider ⟨xα∣fvWα∣xα⟩\langle x_\alpha|f_{\mathrm{vW}}^\alpha|x_\alpha\rangle; exactly the same argument will apply to the spin-β\beta term.

Observe that

TvWα[ρ]=λ8∫dr ∣∇ρα(r)∣2ρα(r)=λ8∫dr ∇ρα(r)⋅∇ρα(r)ρα(r)=λ8∫dr γαα(r)ρα(r)=λ8∫dr f(r,ρα,γαα),(60)\begin{aligned} T_{\mathrm{vW}}^\alpha[\rho] = \frac{\lambda}{8}\int d\mathbf{r}\,\frac{|\nabla\rho_\alpha(\mathbf{r})|^2}{\rho_\alpha(\mathbf{r})} = \frac{\lambda}{8}\int d\mathbf{r}\,\frac{\nabla\rho_\alpha(\mathbf{r})\cdot\nabla\rho_\alpha(\mathbf{r})}{\rho_\alpha(\mathbf{r})} = \frac{\lambda}{8}\int d\mathbf{r}\,\frac{\gamma_{\alpha\alpha}(\mathbf{r})}{\rho_\alpha(\mathbf{r})} \\ &= \frac{\lambda}{8}\int d\mathbf{r}\,f(\mathbf{r},\rho_\alpha,\gamma_{\alpha\alpha}), \tag*{(60)} \end{aligned}

where the reduced gradient γαβ=∇ρα⋅∇ρβ\gamma_{\alpha\beta} = \nabla\rho_\alpha\cdot\nabla\rho_\beta. Without loss of generality, we ignore the constant λ/8\lambda/8.

The derivatives of ff with respect to γαα\gamma_{\alpha\alpha} and ρα\rho_\alpha are [19]

∂f∂ρα=−γααρα2;∂2f∂ρα2=+2γααρα3;∂f∂γαα=1ρα;∂2f∂γαα2=0;\frac{\partial f}{\partial\rho_\alpha} = -\frac{\gamma_{\alpha\alpha}}{\rho_\alpha^2};\qquad \frac{\partial^2 f}{\partial\rho_\alpha^2} = +\frac{2\gamma_{\alpha\alpha}}{\rho_\alpha^3};\qquad \frac{\partial f}{\partial\gamma_{\alpha\alpha}} = \frac{1}{\rho_\alpha};\qquad \frac{\partial^2 f}{\partial\gamma_{\alpha\alpha}^2} = 0;
∂2f∂ρα∂γαα=−1ρα2.(61)\frac{\partial^2 f}{\partial\rho_\alpha\partial\gamma_{\alpha\alpha}} = -\frac{1}{\rho_\alpha^2}. \tag*{(61)}

Omitting the dependence on r\mathbf{r},

⟨xα∣fvWα∣xα⟩=2∫dr (∇xα∗⋅∇xα)∂f∂γαα+∫dr xα∗∂2f∂ρα2xα+2∫dr [(∇ρα⋅∇xα∗)xα+xα∗(∇ρ⋅∇xα)]∂2f∂ρα∂γαα+4∫dr (∇ρα⋅∇xα∗)∂2f∂γαα2(∇ρα⋅∇xα)=2∫dr ∣∇xα∣2ρα−2∫dr xα∗(∇ρα⋅∇xα)+(∇ρα⋅∇xα∗)xαρα2+2∫dr γαα∣xα∣2ρα3.(62)\begin{aligned} \langle x_\alpha|f_{\mathrm{vW}}^\alpha|x_\alpha\rangle ={}&2\int d\mathbf{r}\,(\nabla x_\alpha^* \cdot\nabla x_\alpha)\frac{\partial f}{\partial\gamma_{\alpha\alpha}}\\ &+\int d\mathbf{r}\,x_\alpha^*\frac{\partial^2 f}{\partial\rho_\alpha^2}x_\alpha\\ &+2\int d\mathbf{r}\,[(\nabla\rho_\alpha\cdot\nabla x_\alpha^*)x_\alpha+x_\alpha^*(\nabla\rho\cdot\nabla x_\alpha)]\frac{\partial^2 f}{\partial\rho_\alpha\partial\gamma_{\alpha\alpha}}\\ &+4\int d\mathbf{r}\,(\nabla\rho_\alpha\cdot\nabla x_\alpha^*)\frac{\partial^2 f}{\partial\gamma_{\alpha\alpha}^2}(\nabla\rho_\alpha\cdot\nabla x_\alpha)\\ ={}&2\int d\mathbf{r}\,\frac{|\nabla x_\alpha|^2}{\rho_\alpha} -2\int d\mathbf{r}\,\frac{x_\alpha^*(\nabla\rho_\alpha\cdot\nabla x_\alpha)+(\nabla\rho_\alpha\cdot\nabla x_\alpha^*)x_\alpha}{\rho_\alpha^2}\\ &+2\int d\mathbf{r}\,\frac{\gamma_{\alpha\alpha}|x_\alpha|^2}{\rho_\alpha^3}. \tag*{(62)} \end{aligned}

Multiplying by powers of ρα\rho_\alpha to obtain a common denominator, we have

2∫dr 1ρα3[ρα2(∇xα∗⋅∇xα)−ραxα∗(∇ρα⋅∇xα)−ραxα(∇ρα⋅∇xα∗)+(∇ρα⋅∇ρα)∣xα∣2]=2∫dr (ρα∇xα−xα∇ρα)∗×(ρα∇xα−xα∇ρα)ρα3=2∫dr ∣ρα∇xα−xα∇ρα∣2ρα3;(63)\begin{aligned} 2\int d\mathbf{r}\,\frac{1}{\rho_\alpha^3}\big[&\rho_\alpha^2(\nabla x_\alpha^*\cdot\nabla x_\alpha)-\rho_\alpha x_\alpha^*(\nabla\rho_\alpha\cdot\nabla x_\alpha)\\ &-\rho_\alpha x_\alpha(\nabla\rho_\alpha\cdot\nabla x_\alpha^*)+(\nabla\rho_\alpha\cdot\nabla\rho_\alpha)|x_\alpha|^2\big]\\ ={}&2\int d\mathbf{r}\,\frac{(\rho_\alpha\nabla x_\alpha-x_\alpha\nabla\rho_\alpha)^*\times(\rho_\alpha\nabla x_\alpha-x_\alpha\nabla\rho_\alpha)}{\rho_\alpha^3}\\ ={}&2\int d\mathbf{r}\,\frac{|\rho_\alpha\nabla x_\alpha-x_\alpha\nabla\rho_\alpha|^2}{\rho_\alpha^3}; \tag*{(63)} \end{aligned}

since ρα≥0\rho_\alpha\geq0, ⟨xα∣fvWα∣xα⟩≥0\langle x_\alpha|f_{\mathrm{vW}}^\alpha|x_\alpha\rangle\geq0 (as long as the denominator does not diverge). Replacing the spin index by β\beta means that ⟨xβ∣fvWβ∣xβ⟩≥0\langle x_\beta|f_{\mathrm{vW}}^\beta|x_\beta\rangle\geq0, whence x†M2x≥0x^\dagger M_2x\geq0.

The Hartree kernel

We work in reciprocal space, where fHqf_{\mathrm{H}}^{\mathbf{q}} is local, and note that positive semidefiniteness is invariant under the choice of basis. By the shift theorem of Fourier transforms,

fHq(G)=F[fHq(r)]=F[eiq⋅∣r′−r∣∣r′−r∣]=4π∣G+q∣2>0.(64)f_{\mathrm{H}}^{\mathbf{q}}(\mathbf{G})=\mathcal{F}[f_{\mathrm{H}}^{\mathbf{q}}(\mathbf{r})]=\mathcal{F}\left[\frac{e^{i\mathbf{q}\cdot|\mathbf{r}'-\mathbf{r}|}}{|\mathbf{r}'-\mathbf{r}|}\right]=\frac{4\pi}{|\mathbf{G}+\mathbf{q}|^2}>0. \tag*{(64)}

(The divergence at ∣G+q∣=0|\mathbf{G}+\mathbf{q}|=0 can be handled in various ways; it’s enough to know that fH(∣G+q∣=0)≥0f_{\mathrm{H}}(|\mathbf{G}+\mathbf{q}|=0)\geq0.)

The expectation value of any two-spin vector xx is

(⟨xα∣⟨xβ∣)(fHqfHqfHqfHq)(∣xα⟩∣xβ⟩)=⟨xα∣fHq∣xα⟩+⟨xα∣fHq∣xβ⟩+⟨xβ∣fHq∣xα⟩+⟨xβ∣fHq∣xβ⟩=∑G4π∣q+G∣2×(xα∗(G)xα(G)+xα∗(G)xβ(G)+xβ∗(G)xα(G)+xβ∗(G)xβ(G)).(65)\begin{aligned} \left(\begin{array}{cc}\langle x_\alpha|&\langle x_\beta|\end{array}\right) \left(\begin{array}{cc}f_{\mathrm{H}}^{\mathbf{q}}&f_{\mathrm{H}}^{\mathbf{q}}\\f_{\mathrm{H}}^{\mathbf{q}}&f_{\mathrm{H}}^{\mathbf{q}}\end{array}\right) \binom{|x_\alpha\rangle}{|x_\beta\rangle} ={}&\langle x_\alpha|f_{\mathrm{H}}^{\mathbf{q}}|x_\alpha\rangle+\langle x_\alpha|f_{\mathrm{H}}^{\mathbf{q}}|x_\beta\rangle+\langle x_\beta|f_{\mathrm{H}}^{\mathbf{q}}|x_\alpha\rangle+\langle x_\beta|f_{\mathrm{H}}^{\mathbf{q}}|x_\beta\rangle\\ ={}&\sum_{\mathbf{G}}\frac{4\pi}{|\mathbf{q}+\mathbf{G}|^2}\times\big(x_\alpha^*(\mathbf{G})x_\alpha(\mathbf{G})+x_\alpha^*(\mathbf{G})x_\beta(\mathbf{G})\\ &\qquad+x_\beta^*(\mathbf{G})x_\alpha(\mathbf{G})+x_\beta^*(\mathbf{G})x_\beta(\mathbf{G})\big). \tag*{(65)} \end{aligned}

We simplify the parts depending on xσx_\sigma:

xα∗(G)xα(G)+xα∗(G)xβ(G)+xβ∗(G)xα(G)+xβ∗(G)xβ(G)=[xα(G)+xβ(G)]∗[xα(G)+xβ(G)],(66)x_\alpha^*(\mathbf{G})x_\alpha(\mathbf{G})+x_\alpha^*(\mathbf{G})x_\beta(\mathbf{G})+x_\beta^*(\mathbf{G})x_\alpha(\mathbf{G})+x_\beta^*(\mathbf{G})x_\beta(\mathbf{G})=[x_\alpha(\mathbf{G})+x_\beta(\mathbf{G})]^*[x_\alpha(\mathbf{G})+x_\beta(\mathbf{G})], \tag*{(66)}

so that the total expectation value is the Hartree self-energy of the vector ∣xα+xβ⟩|x_\alpha+x_\beta\rangle. Since the Hartree kernel is nonnegative (manifest in reciprocal space),

⟨xα+xβ∣fHq∣xα+xβ⟩≥0,(67)\langle x_\alpha+x_\beta|f_{\mathrm{H}}^{q}|x_\alpha+x_\beta\rangle\ge0, \tag*{(67)}

with equality only if ∣xβ⟩=−∣xα⟩|x_\beta\rangle=-|x_\alpha\rangle.

The LOSC Hamiltonian

Using the chain rule from the generalized complex (or Wirtinger) derivative [20], the LOSC Hamiltonian is given by

Δh=δΔEδρ=∑σΔhσ,(68)\Delta h=\frac{\delta\Delta E}{\delta\rho}=\sum_\sigma\Delta h^\sigma, \tag*{(68)}

with

Δhσ=∑Rij[∂ΔE∂λRijσδλRijσδρ+∂ΔE∂λRijσ∗δλRijσ∗δρ].(69)\Delta h^\sigma=\sum_{Rij}\left[\frac{\partial\Delta E}{\partial\lambda_{Rij\sigma}}\frac{\delta\lambda_{Rij\sigma}}{\delta\rho}+\frac{\partial\Delta E}{\partial\lambda_{Rij\sigma}^{*}}\frac{\delta\lambda_{Rij\sigma}^{*}}{\delta\rho}\right]. \tag*{(69)}

We can split the correction by spins:

ΔE=∑σ∑RijλRijσ∗(δRijσ−λRijσ)κRijσ.(70)\Delta E=\sum_\sigma\sum_{Rij}\lambda_{Rij\sigma}^{*}(\delta_{Rij\sigma}-\lambda_{Rij\sigma})\kappa_{Rij\sigma}. \tag*{(70)}

Recall that λRijσ=⟨w0i∣ρ∣wRj⟩\lambda_{Rij\sigma}=\langle w_{0i}|\rho|w_{Rj}\rangle. We are also assuming that κRijσ\kappa_{Rij\sigma} is fixed; this is the frozen-orbitalet approximation for SCF calculation. The orbitalets can be relaxed in the so-called macro-SCF loop, although this has only minor effects [21].

An elementary result from the calculus of variations gives us

δλRijσδρ=δ⟨w0i∣ρ∣wRj⟩δρ=∣wRj⟩⟨w0i∣;(71)\frac{\delta\lambda_{Rij\sigma}}{\delta\rho}=\frac{\delta\langle w_{0i}|\rho|w_{Rj}\rangle}{\delta\rho}=|w_{Rj}\rangle\langle w_{0i}|; \tag*{(71)}

similarly, by the Hermiticity of λ\lambda,

δλRijσ∗δρ=δ⟨wRj∣ρ∣w0i⟩δρ=∣w0i⟩⟨wRj∣.(72)\frac{\delta\lambda_{Rij\sigma}^{*}}{\delta\rho}=\frac{\delta\langle w_{Rj}|\rho|w_{0i}\rangle}{\delta\rho}=|w_{0i}\rangle\langle w_{Rj}|. \tag*{(72)}

Next, we observe that

∂ΔE∂λRijσ=∂[12λRijσ∗(δRijσ−λRijσ)κRijσ]∂λRijσ=−12λRijσ∗κRijσ,(73)\frac{\partial\Delta E}{\partial\lambda_{Rij\sigma}}=\frac{\partial\left[\frac{1}{2}\lambda_{Rij\sigma}^{*}(\delta_{Rij\sigma}-\lambda_{Rij\sigma})\kappa_{Rij\sigma}\right]}{\partial\lambda_{Rij\sigma}}=-\frac{1}{2}\lambda_{Rij\sigma}^{*}\kappa_{Rij\sigma}, \tag*{(73)}

and similarly

∂ΔE∂λRijσ∗=∂[12λRijσ∗(δRijσ−λRijσ)κRijσ]∂λRijσ∗=−12(δRijσ−λRijσ)κRijσ.(74)\frac{\partial\Delta E}{\partial\lambda_{Rij\sigma}^{*}}=\frac{\partial\left[\frac{1}{2}\lambda_{Rij\sigma}^{*}(\delta_{Rij\sigma}-\lambda_{Rij\sigma})\kappa_{Rij\sigma}\right]}{\partial\lambda_{Rij\sigma}^{*}}=-\frac{1}{2}(\delta_{Rij\sigma}-\lambda_{Rij\sigma})\kappa_{Rij\sigma}. \tag*{(74)}

Thus

Δhσ=∑Rij[−12λRijσ∗κRijσ∣wRj⟩⟨w0i∣+12(δRijσ−λRijσ)κRijσ∣w0i⟩⟨wRj∣].(75)\Delta h^\sigma=\sum_{Rij}\left[-\frac{1}{2}\lambda_{Rij\sigma}^{*}\kappa_{Rij\sigma}|w_{Rj}\rangle\langle w_{0i}| +\frac{1}{2}(\delta_{Rij\sigma}-\lambda_{Rij\sigma})\kappa_{Rij\sigma}|w_{0i}\rangle\langle w_{Rj}|\right]. \tag*{(75)}

To simplify this, we separate diagonal and off-diagonal terms, with

Δhσ=∑i12κii0[(1−λii0)∣w0i⟩⟨w0i∣−λii0∣w0i⟩⟨w0i∣]−∑Rij≠0i12κRijσ[λRijσ∣w0i⟩⟨wRj∣+λRijσ∗∣wRj⟩⟨w0i∣].(76)\Delta h^\sigma=\sum_i\frac{1}{2}\kappa_{ii0}\left[(1-\lambda_{ii0})|w_{0i}\rangle\langle w_{0i}|-\lambda_{ii0}|w_{0i}\rangle\langle w_{0i}|\right] -\sum_{Rij\ne0i}\frac{1}{2}\kappa_{Rij\sigma}\left[\lambda_{Rij\sigma}|w_{0i}\rangle\langle w_{Rj}|+\lambda_{Rij\sigma}^{*}|w_{Rj}\rangle\langle w_{0i}|\right]. \tag*{(76)}

Since the diagonal elements of λ\lambda and κ\kappa are real, the first summand simplifies to

12κii0[(1−λii0)∣w0i⟩⟨w0i∣−λii0∣w0i⟩⟨w0i∣]=κii0(12−λii0)∣w0i⟩⟨w0i∣,(77)\frac{1}{2}\kappa_{ii0}\left[(1-\lambda_{ii0})|w_{0i}\rangle\langle w_{0i}|-\lambda_{ii0}|w_{0i}\rangle\langle w_{0i}|\right] = \kappa_{ii0}\left(\frac{1}{2}-\lambda_{ii0}\right)|w_{0i}\rangle\langle w_{0i}|, \tag*{(77)}

and the second to

∑Rj≠0i12κRijσ[λRijσ∣w0i⟩⟨wRj∣+λRijσ∗∣wRj⟩⟨w0i∣]=∑Rj≠0iκRijσλRijσ+λRijσ∗2∣w0i⟩⟨wRj∣=∑Rj≠0iκRijσRe⁡λRijσ∣w0i⟩⟨wRj∣.(78)\sum_{\mathbf{R}j\ne0i}\frac{1}{2}\kappa_{\mathbf{R}ij\sigma}\left[\lambda_{\mathbf{R}ij\sigma}|w_{0i}\rangle\langle w_{\mathbf{R}j}|+\lambda_{\mathbf{R}ij\sigma}^{*}|w_{\mathbf{R}j}\rangle\langle w_{0i}|\right] = \sum_{\mathbf{R}j\ne0i}\kappa_{\mathbf{R}ij\sigma}\frac{\lambda_{\mathbf{R}ij\sigma}+\lambda_{\mathbf{R}ij\sigma}^{*}}{2}|w_{0i}\rangle\langle w_{\mathbf{R}j}| = \sum_{\mathbf{R}j\ne0i}\kappa_{\mathbf{R}ij\sigma}\operatorname{Re}\lambda_{\mathbf{R}ij\sigma}|w_{0i}\rangle\langle w_{\mathbf{R}j}|. \tag*{(78)}

(In the above, we used the fact that (∣w0i⟩⟨wRj∣)†=∣wRj⟩⟨w0i∣(|w_{0i}\rangle\langle w_{\mathbf{R}j}|)^\dagger=|w_{\mathbf{R}j}\rangle\langle w_{0i}|, with (λ†)Rijσ=(λT)Rijσ∗=λji(−R)σ∗=λRijσ(\lambda^\dagger)_{\mathbf{R}ij\sigma}=(\lambda^{\mathrm{T}})_{\mathbf{R}ij\sigma}^{*}=\lambda_{ji(-\mathbf{R})\sigma}^{*}=\lambda_{\mathbf{R}ij\sigma}. Observe that the coefficients for each outer product of DLWFs come in complex conjugate pairs.)

Combining, we have

Δh=∑σΔhσ=∑σ∑RijκRijσ(12δRijσ−Re⁡λRijσ)∣w0i⟩⟨wRj∣=∑σ∑Rij12κRijσ(12δRijσ−λRijσ)∣w0i⟩⟨wRj∣+h.c.(79)\begin{aligned} \Delta h=\sum_{\sigma}\Delta h^\sigma=\sum_{\sigma}\sum_{\mathbf{R}ij}\kappa_{\mathbf{R}ij\sigma}\left(\frac{1}{2}\delta_{\mathbf{R}ij\sigma}-\operatorname{Re}\lambda_{\mathbf{R}ij\sigma}\right)|w_{0i}\rangle\langle w_{\mathbf{R}j}| \\ &=\sum_{\sigma}\sum_{\mathbf{R}ij}\frac{1}{2}\kappa_{\mathbf{R}ij\sigma}\left(\frac{1}{2}\delta_{\mathbf{R}ij\sigma}-\lambda_{\mathbf{R}ij\sigma}\right)|w_{0i}\rangle\langle w_{\mathbf{R}j}|+h.c. \tag*{(79)} \end{aligned}

In finite molecules, the Hamiltonian is

Δh=∑σΔhσ=∑σ(∑i(12−λiiσ)κiiσ∣ϕiσ⟩⟨ϕiσ∣−∑i≠jκijσ∣ϕiσ⟩⟨ϕjσ∣),(80)\Delta h=\sum_{\sigma}\Delta h^\sigma=\sum_{\sigma}\left(\sum_i\left(\frac{1}{2}-\lambda_{ii}^{\sigma}\right)\kappa_{ii}^{\sigma}|\phi_i^\sigma\rangle\langle\phi_i^\sigma|-\sum_{i\ne j}\kappa_{ij}^{\sigma}|\phi_i^\sigma\rangle\langle\phi_j^\sigma|\right), \tag*{(80)}

with ∣ϕi⟩|\phi_i\rangle as the localized molecular orbitals.

Choice of parameters

There are two parameters that can be tuned in olLOSC: the space/energy mixing 0≤γ≤10\le\gamma\le1 in the orbitalets, and the fraction λ≥0\lambda\ge0 of von Weizsäcker kinetic energy added to the Thomas–Fermi kernel.

In materials

The results of our parameter sweep are shown in Fig. 1. Since olLOSC is most accurate for systems with smaller gaps (≤8\le8 eV), we chose parameters based on their performance in the left column of the figure. It was clear that some nonzero fraction of von Weizsäcker kinetic energy is required: for all values of the DLWF localization parameter γ\gamma, there was a large negative mean signed error (c, g), indicating systematic underestimation of the band gap.

If γ=0.47714\gamma=0.47714, the value used in previous works on LOSC [21, 22, 23, 24], then 100% von Weizsäcker correction is preferred (λ=1\lambda=1). However, γ=0.30\gamma=0.30 yields improved performance in some metrics: γ=0.30\gamma=0.30, λ=1\lambda=1 has almost zero mean signed percent error (c), while γ=0.30\gamma=0.30, λ=0.75\lambda=0.75 has minimal mean signed error (eV; g).

We chose γ=0.30\gamma=0.30 for this work to reduce the appearance of local minima in the DLWF minimization (see below), and λ=0.75\lambda=0.75 to minimize the systematic error.

Convergence in Brillouin zone sampling

We investigate the convergence with respect to the k-mesh in Figure 2. (Recall that olLOSC requires a uniform k-sampling centered at Γ\Gamma for its monochromatic decomposition.) In general, the change in olLOSC gap is fairly small as the k-sampling is increased from 6×6×66 \times6 \times6 to 8×8×88 \times8 \times8. Noticeable differences appear in Si, C, BP, and SiC because their conduction band minimum is not at a high-symmetry k-point; thus, the olLOSC gap change is matched by a similar one in the PBE calculation.

Four-panel scatter plots showing olLOSC convergence with respect to k-points

Figure 2. (a) γ=0.3\gamma= 0.3, λ=0.75\lambda= 0.75 (all systems). (b) γ=0.3\gamma= 0.3, λ=0.75\lambda= 0.75 (gap ≤8\le8 eV). (c) γ=0.30\gamma= 0.30, λ=1\lambda= 1 (all systems). (d) γ=0.30\gamma= 0.30, λ=1\lambda= 1 (gap ≤8\le8 eV). Figure 2: olLOSC convergence with respect to k-points for materials at the two best parameter choices. Left column: All systems. Right column: Small-gapped systems (≤8\le8 eV). Note that NaF and Ne were only computed with a 6×6×66 \times6 \times6 k-mesh.

Large-gapped systems tend to have a slight reduction in the band gap as the k-sampling is increased separate from the DFA. This behavior generally coincides with differences in the ‘degeneracy pattern’ of the DLWFs, as noted in [3]; their convergence is often more difficult in larger-gapped systems. As olLOSC is most applicable to materials with smaller gaps, we did not investigate further. The large deviation in AlP when γ=0.47714\gamma= 0.47714 is also due to a qualitative difference in the DLWFs (see § 16.2 below); nevertheless, it is partially to avoid this possibility that we choose γ=0.30\gamma= 0.30 in the main text.

In molecules

The results of our parameter sweep in molecular systems are shown in Figure 3 and Figure 4; λ\lambda is one of 50%, 75%, and 100%, and γ∈{0.2,0.3,0.47714}\gamma\in\{0.2, 0.3, 0.47714\}. Size-consistency is of particular importance in molecular calculations. Small molecules are generally spatially localized but delocalized in energy, have the opposite problem. We therefore partitioned the molecular test based on system size. The small-molecule subset contains the molecules with no more than five atoms; the large-molecule subset contains the rest.

Relative errors (%) by LOSC $\gamma$ and the vW $\lambda$ in the molecules tested

Figure 3. Relative errors (%) by LOSC γ\gamma and the vW λ\lambda in the molecules tested. (a, d): Small molecules. (b, e): Large molecules. (c, f): All molecular systems.

Bar charts of unscaled errors by LOSC $\gamma$ and vW $\lambda$

Figure 4. Unscaled errors (eV) by LOSC γ\gamma and the vW λ\lambda in the molecules tested. (a, d): Small molecules. (b, e): Large molecules. (c, f): All molecular systems.

From the standpoint of mean signed error, the molecular datasets allow greater flexibility with respect to λ\lambda. For all three choices of λ\lambda, the fundamental gaps predicted by olLOSC with tf+λvWt_f+\lambda v_\mathrm{W} screening exhibit a mean absolute error of roughly 0.4 eV. Calculations on materials favor larger λ\lambda values, as can be seen in Table 7 (γ=0.3\gamma= 0.3, λ=0.75\lambda= 0.75) and 10 (γ=0.47714\gamma= 0.47714, λ=1.0\lambda= 1.0). Data for the rest of the parameter space can be found in the SI Dataset [25].

SystemEPBEE_{\mathrm{PBE}} (Eh)ΔE\Delta E (mEh)HOMO (eV) PBEHOMO (eV) olLOSCLUMO (eV) PBELUMO (eV) olLOSCGap (eV) PBEGap (eV) olLOSCGap (eV) Ref.olLOSC gap error eVolLOSC gap error %
Acridine-555.067-0.312-5.426-7.966-2.976-0.8732.4497.0937.394-0.301-4.069
Anthracene-539.0211.462-4.979-7.422-2.728-0.6832.2516.7397.214-0.475-6.579
Azulene-385.4788.534-4.970-7.499-2.881-0.5312.0896.9687.003-0.035-0.493
Benzonitrile-324.201-0.087-6.852-9.924-2.5210.0694.3319.99210.168-0.176-1.728
Benzoquinone-381.1601.487-6.348-10.259-4.566-1.6691.7818.5908.714-0.124-1.428
Dichlone-1453.559-0.958-6.580-9.899-4.385-2.0692.1957.8298.066-0.237-2.934
Dinitrobenzonitrile-733.009-0.448-7.725-11.315-4.583-2.0063.1429.3099.408-0.099-1.053
Fumaronitrile-262.854-0.312-7.851-11.310-4.094-1.1833.75710.12810.507-0.379-3.611
Maleicanhydride-379.039-0.029-7.129-11.342-4.199-1.1762.93010.16710.308-0.141-1.372
NDCA-686.0673.383-6.468-9.023-3.658-1.3942.8107.6297.918-0.289-3.655
Naphthalenedione-534.6821.114-6.151-9.659-4.171-1.6711.9797.9888.403-0.415-4.935
Nitrobenzene-436.435-0.581-6.816-10.760-3.485-0.5363.33210.2249.6970.5275.432
Nitrobenzonitrile-528.6080.651-7.301-11.042-4.136-1.4513.1659.5909.3480.2422.591
Phenazine-571.1060.464-5.735-8.959-3.404-1.1802.3317.7797.3870.3925.306
Phthalicanhydride-532.560-0.336-6.900-10.579-3.601-0.9863.2999.5939.700-0.107-1.107
Phthalimide-512.703-0.964-6.350-9.972-3.299-0.7413.0519.2319.482-0.251-2.650
TCNE-447.1733.614-8.597-11.754-5.780-3.2442.8178.5108.954-0.444-4.954
TCNQ-678.02216.349-7.050-9.377-5.514-3.6661.5365.7116.262-0.551-8.804
Tetrachloro-isobenzofuranedione-2370.3003.504-7.106-9.777-4.062-1.7213.0448.0578.430-0.373-4.427
Tetrachlorobenzoquinone-2218.9123.724-6.983-9.787-5.010-2.6041.9737.1847.762-0.578-7.451
Tetrafluorobenzenedicarbonitrile-813.1191.210-7.444-10.070-4.133-1.7033.3118.3679.151-0.784-8.569
Tetrafluorobenzoquinone-777.9132.135-7.374-10.343-5.127-2.2612.2488.0828.861-0.779-8.791
mDCBN-416.3740.572-7.393-10.240-3.288-0.8124.1059.4289.832-0.404-4.112

Table 7. Large organic molecules olLOSC calculation results with λ=0.75\lambda= 0.75 and γ=0.30\gamma= 0.30. (See b and e in Figure 3 and Figure 4) Reference gap (Ref.) is CCSD-T benchmark from Ref. [43].

olLOSC beyond the random phase approximation

As we mentioned earlier, the RPA is an approximation; at first glance, it should appear that we should avoid it if possible. In the subsequent discourse, a comparative analysis will be conducted between the beyond RPA olLOSC scheme and the RPA-based olLOSC scheme. Owing to certain numerical instability, wherein the preconditioner employed in bulk calculations may lose its positive-definiteness when the potential term includes exchange and correlation effects, the beyond RPA olLOSC protocol demonstrates robustness exclusively in molecular systems. In Figure 5 and Figure 6, parameter scanning is performed analogously to the conventional RPA-based olLOSC. For two parameter sets with particular interest, {γ=0.3,λ=0.75}\{\gamma= 0.3, \lambda= 0.75\} and {γ=0.47714,λ=1}\{\gamma= 0.47714, \lambda= 1\}, we also provide their numerical results in Table 12. The beyond-RPA olLOSC routine yields superior results for small-sized molecules, yet underperforms in large-sized molecules, with comparable outcomes observed for both sets of parameters. Based on this comparison and on the basis of self-consistence and robustness, it is preferred to adapt the RPA-based olLOSC as a conventional olLOSC routine.

Relative errors by LOSC $\gamma$ and vW $\lambda$ for the tested molecules

Figure 5. of-LOSC beyond RPA Relative errors (%) by LOSC γ\gamma and the vW λ\lambda in the molecules tested. (a, d): Small molecules. (b, e): Large molecules. (c, f): All molecular systems.

Bar charts of mean signed and absolute errors for small, large, and all molecular systems

Figure 6. of-LOSC beyond RPA unscaled errors (eV) by LOSC γ\gamma and the vW λ\lambda in the molecules tested. (a, d): Small molecules. (b, e): Large molecules. (c, f): All molecular systems.

γλisRPAMAE (eV)MSE (eV)MAREMSRE
Small Molecules (17 molecules)
0.300.75TRUE0.5620.5285.479%5.077%
0.300.75FALSE0.3760.2773.673%2.526%
0.477141.0TRUE0.7570.7387.446%7.231%
0.477141.0FALSE0.5880.5415.737%5.219%
Large Molecules (23 molecules)
0.300.75TRUE0.352-0.2514.176%-3.017%
0.300.75FALSE0.491-0.4575.779%-5.386%
0.477141.0TRUE0.393-0.2494.761%-3.071%
0.477141.0FALSE0.486-0.4135.791%-4.946%
All Molecules (40 molecules)
0.300.75TRUE0.4420.0804.730%0.423%
0.300.75FALSE0.443-0.1454.884%-2.024%
0.477141.0TRUE0.5480.1715.902%1.307%
0.477141.0FALSE0.529-0.0085.768%-0.626%

Table 12. Comparison of the default olLOSC routine (within random phase approximation) with the olLOSC routine beyond random phase approximation across various categories of molecular systems. Also see Figure 3–Figure 6 for paramter set other than {γ=0.3\gamma= 0.3, λ=0.75\lambda= 0.75} and {γ=0.477114\gamma= 0.477114, λ=1.0\lambda= 1.0}. Raw data and visualization scripts are in the SI dataset.

Total energy
Materials

The materials we tested are in their equilibrium geometry, so we expect only a small correction from olLOSC to the total energy; and that is what we see. For the production dataset seen in the main text with (γ=0.30\gamma= 0.30, λ=0.75\lambda= 0.75) (Fig. 1; Table 2), the maximum energy correction is 3.60×10−33.60 \times10^{-3} Ry, for silicon; it also has the largest relative energy change, −0.0213%-0.0213\%.

Systemk-meshTotal energy (Ry) EPBEE_{\mathrm{PBE}}ΔELOSC\Delta E_{\mathrm{LOSC}}VBM (eV) PBEVBM (eV) olLOSCCBM (eV) PBECBM (eV) olLOSCGap (eV) PBEGap (eV) olLOSCRef.olLOSC gap error eVolLOSC gap error %
AlP06−18.589.75×10−49.75 \times10^{-4}5.0214.3596.5976.7091.5762.3502.606−0.256−9.808
Ar06−45.184.48×10−54.48 \times10^{-5}−3.948−5.7564.7555.0668.70310.82214.200−3.378−23.788
BN06−26.812.88×10−52.88 \times10^{-5}11.1469.77215.67716.4274.5316.6556.6220.0330.500
BP10−19.798.28×10−48.28 \times10^{-4}9.2408.57010.49210.9321.2522.3622.501−0.139−5.570
C08−24.086.85×10−56.85 \times10^{-5}13.35512.82017.54418.2354.1895.4155.803−0.388−6.693
Ge06−357.282.13×10−32.13 \times10^{-3}9.5619.0949.5939.8550.0320.7610.773−0.012−1.604
LiCl06−47.475.99×10−55.99 \times10^{-5}1.6670.4227.9967.8426.3297.4209.836−2.416−24.564
LiF06−63.982.11×10−52.11 \times10^{-5}0.970−1.70210.15610.2419.18711.94315.431−3.488−22.605
MgO06−152.108.69×10−58.69 \times10^{-5}8.0996.32612.90113.3664.8037.0418.503−1.462−17.200
NaF06−140.977.98×10−57.98 \times10^{-5}0.999−1.6397.3898.0026.3909.64112.199−2.558−20.971
Ne06−66.712.25×10−52.25 \times10^{-5}−9.071−12.7162.5455.83311.61618.54921.700−3.151−14.521
Si10−16.922.52×10−32.52 \times10^{-3}6.2345.6926.8127.0340.5781.3421.2280.1149.267
SiC08−20.544.84×10−44.84 \times10^{-4}9.4968.63510.85711.0511.3612.4162.595−0.179−6.913

Table 2. The dataset seen in Fig. 1 (b) of the main text: the olLOSC parameters are γ=0.30\gamma= 0.30, λ=0.75\lambda= 0.75 We use a 6×6×66 \times6 \times6 Monkhorst–Pack kk-mesh, except in cases where the material’s conduction band minimum does not lie on a high-symmetry kk-point (as evidenced by the PBE CBM changing with kk). Thus, we use 8×8×88 \times8 \times8 kk-points for C and SiC, and 10×10×1010 \times10 \times10 kk-points for BP and Si.

Molecules

Total energies in the underlying DFA calculations for molecules are not size-consistent [22], so a large ΔELOSC\Delta E_\mathrm{LOSC} is expected for large molecules. Table 7 and 8 include EDFAE_\mathrm{DFA} and ΔELOSC\Delta E_\mathrm{LOSC} for the large molecules we tested. We observed olLOSC energy corrections up to a few mEh_\mathrm{h}. The largest is for TCNQ (tetracyanoquinodimethane), which is a highly conjugated molecule (thus, significant delocalization error is expected). It has ΔELOSC=\Delta E_\mathrm{LOSC} = 16.349 mEhmE_{\mathrm{h}} when γ=0.30\gamma= 0.30 and λ=0.75\lambda= 0.75. Note that increased spatial localization (a smaller γ\gamma) yields a larger ΔE\Delta E because non-integer local occupations λRijσ\lambda_{Rij\sigma} are penalized less. Thus, when γ=0.47714\gamma= 0.47714 and λ=1\lambda= 1, ΔELOSC\Delta E_{\mathrm{LOSC}} for TCNQ is 11.383 mEhmE_{\mathrm{h}}. The reaction barrier height is heavily based on total energy calculations. Table 13 includes mean signed error (MSE) and mean absolute error (MAE) for reaction barrier heights using the HTBH38 and NHTBH38 datasets [26, 27], where reference reaction barrier heights are considered to be the ’REF1’ (exclude relativistic effect) within these datasets. A γ\gamma value of 0.47714 generally replicates the results of PBE reaction barrier heights, whereas a γ\gamma value of 0.3 yields the lowest mean absolute error.

γ\gammaλ\lambdaNHTBH38HTBH38All reactions
MAEMSEMAEMSEMAEMSE
PBEN/A9.915-8.1419.424-9.4249.669-8.782
0.20.757.888-2.7219.873-9.8738.881-6.297
0.30.757.012-4.93710.116-10.1168.564-7.527
0.477140.759.249-7.4389.788-9.7889.518-8.613

Table 13. Mean signed error (MSE) and mean absolute error (MAE) for reaction barrier heights (in kcal/mol). NHTBH38 is a 38-reaction dataset of non-hydrogen transfer reaction barrier heights, while HTBH38 contains 38 hydrogen transfer barrier heights. We take REF1 (which excludes relativistic effect) from [26, 27] as a reference. The first row gives the reaction barrier calculated by the parent DFT method (PBE); the subsequent rows detail the olLOSC reaction barrier calculations using 75% von Weizsäcker kinetic energy (λ=0.75\lambda= 0.75) at different localization parameters γ\gamma.

Local minima in dually localized Wannier functions

DLWFs [3] are computed variationally, minimizing for each spin component the cost function

Fσ[γ]=(1−γ)∑i[⟨w0iσ∣r∣w0iσ⟩2−⟨w0iσ∣r2∣w0iσ⟩]+Cγ∑i[⟨w0iσ∣h∣w0iσ⟩2−⟨w0iσ∣h2∣w0iσ⟩]=(1−γ)∑iΔr0iσ2+Cγ∑iΔh0iσ2(81)\begin{aligned} F^\sigma[\gamma] = (1-\gamma)\sum_i \left[\langle w_{0i\sigma}|r|w_{0i\sigma}\rangle^2-\langle w_{0i\sigma}|r^2|w_{0i\sigma}\rangle\right] + C\gamma\sum_i \left[\langle w_{0i\sigma}|h|w_{0i\sigma}\rangle^2-\langle w_{0i\sigma}|h^2|w_{0i\sigma}\rangle\right] \\ &= (1-\gamma)\sum_i \Delta r_{0i\sigma}^2+C\gamma\sum_i \Delta h_{0i\sigma}^2 \tag*{(81)} \end{aligned}

where 0≤γ≤10 \le\gamma\le1 and C=1C = 1. (Note that wRiσ(r)=w0iσ(r−R)w_{R i\sigma}(r) = w_{0i\sigma}(r-R), so we do not need to sum over the unit cell index RR; that would only multiply Fσ[γ]F^\sigma[\gamma] by a constant, which does not affect the optimization.)

We minimize FσF^\sigma by modifying the method used for maximally localized Wannier functions (MLWFs) [4]. The difference is in the cost function, which includes the Hamiltonian, as first suggested by [28]. The key difference is that DLWFs are constructed from both valence and conduction bands. We first disentangle the conduction’ manifold [29], obtaining optimally smooth pseudo-bands that span the virtual subspace. (If the disentangled bands are analytic, the MLWFs they specify decay exponentially [30].) Once the bands are chosen, we formulate Fσ[γ]F^\sigma[\gamma] as a function of the elements of the localization unitaries UinσU^{\sigma}_{in}, which localize the disentangled bands:

ϕkiσ(r)=∑nUinσψknσ(r).(82)\phi_{ki\sigma}(r)=\sum_n U^{\sigma}_{in}\psi_{kn\sigma}(r). \tag*{(82)}

Then we minimize FσF^\sigma with the conjugate gradient method.

It was noted as early as [4] that, depending on the quality of the initial guess, the MLWF localization scheme can become trapped in local minima of the cost function. More than two decades of work has rendered MLWFs extremely robust to local minima, to the point that there are automated and high-throughput workflows for Wannierization [31]. This scheme scheme relies on the selected columns of the density matrix (SCDM) algorithm [32, 33] for its initial guess. SCDM produces orthonormal, localized Wannier-like functions deterministically by exploiting the exponential decay of the one-electron reduced density matrix [34],

γσ(r,r′)∼e−∣r−r′∣.(83)\gamma_\sigma(r,r') \sim e^{-|r-r'|}. \tag*{(83)}

We, too, use SCDM to generate initial guesses for DLWFs, but our minimization is still sensitive to local minima. This is due largely to the disentangled conduction bands, which have large energy variance Δh0iσ2\Delta h_{0i\sigma}^2. We avoid this as best we can by minimizing FσF^\sigma several times, over a variety of conjugate gradient step sizes, and choosing the smallest cost function value overall.

Mahler et al. also observed that the qualitative nature of the DLWFs changes as a function of γ\gamma (see Fig. 1 of [3]). At γ=0\gamma= 0 (the MLWF limit), all the WFs are degenerate; they are also typically equivalent up to symmetry operations in the primitive cell. At small positive γ\gamma, the valence and conduction DLWFs split off from one another. They have different shapes and different average energies ⟨h⟩0iσ\langle h\rangle_{0i\sigma}. Splitting of degenerate subsets continues as γ\gamma is increased further.

We see similar patterns in the DLWFs for these systems, but the degeneracy pattern varies somewhat with the kk-sampling—especially for our largest γ\gamma, 0.47714. As a particularly noticeable example, observe the large difference (0.442 eV) in the olLOSC gap (γ=0.47714\gamma= 0.47714, λ=1\lambda= 1) of BP between k = 6 and k = 8 (Figure 2(d) and Table 3). We attribute this difference to the DLWFs. Table 6 demonstrates how BP has two ‘classes’ of occupied DLWFs (with local occupations λ0iiσ>0.99\lambda_{0ii\sigma} > 0.99) when computed with the 6 × 6 × 6 k-mesh: one has ⟨h⟩0iσ≈−3.1 eV\langle h\rangle_{0i\sigma} \approx-3.1\,\mathrm{eV}, and three are nearly degenerate, with ⟨h⟩0iσ≈4.7 eV\langle h\rangle_{0i\sigma} \approx4.7\,\mathrm{eV}. In the 8×8×88 \times8 \times8 kk-sampled calculation, the lowest-energy DLWF is qualitatively very similar, but the other three occupied DLWFs have split into two sub-bands. One has ⟨h⟩0iσ≈3.0 eV\langle h\rangle_{0i\sigma} \approx3.0\,\mathrm{eV}, and two are nearly degenerate, with ⟨h⟩0iσ≈5.5 eV\langle h\rangle_{0i\sigma} \approx5.5\,\mathrm{eV}. They also have qualitatively different spatial spreads, as seen by the values of ⟨r2⟩0iσ\langle r^2\rangle_{0i\sigma} in Table 6.

Table (continued in Table 4): Choosing the γ\gamma and λ\lambda parameters in bulk systems (see Fig. 1). The k-mesh is as in Table 2. Reference gap (Ref.) is experimental − ZPR from Table 1.

SystemDLWF γ\gammavW λ\lambdaVBM PBE (eV)VBM olLOSC (eV)CBM PBE (eV)CBM olLOSC (eV)Gap PBE (eV)Gap olLOSC (eV)Ref.eV%
AlP0.200.005.0214.4166.5976.4821.5762.0662.606−0.540−20.733
AlP0.200.755.0214.3496.5976.7101.5762.3612.606−0.245−9.405
AlP0.201.005.0214.3356.5976.7321.5762.3972.606−0.209−8.035
AlP0.300.005.0214.4256.5976.4891.5762.0642.606−0.542−20.814
AlP0.300.755.0214.3596.5976.7091.5762.3502.606−0.256−9.808
AlP0.301.005.0214.3456.5976.7301.5762.3852.606−0.221−8.480
AlP0.480.005.0214.4386.5976.5011.5762.0632.606−0.543−20.837
AlP0.480.755.0214.3736.5976.7131.5762.3402.606−0.266−10.207
AlP0.481.005.0214.3606.5976.7331.5762.3732.606−0.232−8.922
Ar0.200.00−3.948−9.2364.755−9.2828.703−0.04714.200−14.247−100.329
Ar0.200.75−3.948−5.7614.7554.5828.70310.34314.200−3.857−27.162
Ar0.201.00−3.948−5.7784.7554.6458.70310.42314.200−3.777−26.599
Ar0.300.00−3.948−6.8304.755−7.7308.703−0.90014.200−15.100−106.339
Ar0.300.75−3.948−5.7564.7555.0668.70310.82214.200−3.378−23.788
Ar0.301.00−3.948−5.7744.7555.1198.70310.89314.200−3.307−23.291
Ar0.480.00−3.948−5.6014.755−6.4698.703−0.86814.200−15.068−106.115
Ar0.480.75−3.948−5.7524.7555.1098.70310.86114.200−3.339−23.513
Ar0.481.00−3.948−5.7694.7555.1588.70310.92814.200−3.272−23.044
BN0.200.0011.1469.86015.67716.5484.5316.6896.6220.0671.010
BN0.200.7511.1469.76415.67716.7454.5316.9816.6220.3595.423
BN0.201.0011.1469.74315.67716.7814.5317.0376.6220.4156.275
BN0.300.0011.1469.86715.67716.2504.5316.3836.622−0.239−3.617
BN0.300.7511.1469.77215.67716.4274.5316.6556.6220.0330.500
BN0.301.0011.1469.75215.67716.4584.5316.7066.6220.0841.272
BN0.480.0011.14610.37215.67716.3114.5315.9396.622−0.683−10.307
BN0.480.7511.14610.33315.67716.4004.5316.0676.622−0.555−8.380
BN0.481.0011.14610.32515.67716.4184.5316.0936.622−0.529−7.982
BP0.200.009.2408.62910.49210.8501.2522.2212.501−0.280−11.188
BP0.200.759.2408.55610.49210.9351.2522.3782.501−0.123−4.910
BP0.201.009.2408.54010.49210.9491.2522.4092.501−0.092−3.675
BP0.300.009.2408.64110.49210.8481.2522.2082.501−0.293−11.731
BP0.300.759.2408.57010.49210.9321.2522.3622.501−0.139−5.570
BP0.301.009.2408.55410.49210.9461.2522.3922.501−0.109−4.362
BP0.480.009.2408.59110.49210.8191.2522.2272.501−0.274−10.944
BP0.480.759.2408.57610.49210.8271.2522.2512.501−0.250−9.996
C0.200.0013.35512.39017.54418.5044.1896.1145.8030.3115.352
C0.200.7513.35512.30317.54418.6384.1896.3355.8030.5329.164
C0.201.0013.35512.28317.54418.6644.1896.3815.8030.5789.962
C0.300.0013.35512.85217.54418.1584.1895.3075.803−0.496−8.554
C0.300.7513.35512.82017.54418.2354.1895.4155.803−0.388−6.693
C0.301.0013.35512.81417.54418.2494.1895.4355.803−0.368−6.333
C0.480.0013.35512.93317.54418.0344.1895.1015.803−0.702−12.104
C0.480.7513.35512.91117.54418.0864.1895.1755.803−0.628−10.815
C0.481.0013.35512.90617.54418.0964.1895.1905.803−0.613−10.569
Ge0.200.009.5619.1929.5939.7480.0320.5550.773−0.218−28.150
Ge0.200.759.5619.0609.5939.8550.0320.7950.7730.0222.885
Ge0.201.009.5619.0319.5939.8740.0320.8430.7730.0709.056
Ge0.300.009.5619.2169.5939.7470.0320.5300.773−0.243−31.384
Ge0.300.759.5619.0949.5939.8550.0320.7610.773−0.012−1.604
Ge0.301.009.5619.0679.5939.8730.0320.8060.7730.0334.295
Ge0.480.009.5619.3599.5939.7350.0320.3770.773−0.396−51.294
Ge0.480.759.5619.2889.5939.8420.0320.5530.773−0.220−28.396
Ge0.481.009.5619.2739.5939.8600.0320.5870.773−0.186−24.036
Systemγ\gammaλ\lambdavWv_{\mathrm{W}}VBM PBE (eV)VBM olLOSC (eV)CBM PBE (eV)CBM olLOSC (eV)Gap PBE (eV)Gap olLOSC (eV)Ref.olLOSC gap error (eV)olLOSC gap error (%)
LiCl0.200.001.6670.5387.9964.9906.3294.4529.836−5.384−54.737
LiCl0.200.751.6670.4207.9967.7496.3297.3299.836−2.507−25.491
LiCl0.201.001.6670.4047.9967.8006.3297.3969.836−2.441−24.812
LiCl0.300.001.6670.5407.9965.5226.3294.9829.836−4.854−49.349
LiCl0.300.751.6670.4227.9967.8426.3297.4209.836−2.416−24.564
LiCl0.301.001.6670.4067.9967.8866.3297.4809.836−2.356−23.952
LiCl0.480.001.6670.5457.9965.9596.3295.4149.836−4.422−44.957
LiCl0.480.751.6670.4277.9967.9406.3297.5139.836−2.323−23.618
LiCl0.481.001.6670.4117.9967.9786.3297.5679.836−2.269−23.072
LiF0.200.000.970−1.54310.1567.6229.1879.16515.431−6.265−40.603
LiF0.200.750.970−1.71110.15610.1679.18711.87815.431−3.553−23.024
LiF0.201.000.970−1.73810.15610.2259.18711.96315.431−3.468−22.472
LiF0.300.000.970−1.53410.1567.9119.1879.44515.431−5.986−38.795
LiF0.300.750.970−1.70210.15610.2419.18711.94315.431−3.488−22.605
LiF0.301.000.970−1.72910.15610.2959.18712.02415.431−3.407−22.079
LiF0.480.000.970−1.52610.1568.2579.1879.78315.431−5.648−36.605
LiF0.480.750.970−1.69510.15610.3309.18712.02515.431−3.406−22.076
LiF0.481.000.970−1.72210.15610.3799.18712.10015.431−3.331−21.584
MgO0.200.008.0996.44612.90112.5084.8036.0618.503−2.442−28.716
MgO0.200.758.0996.31212.90113.2464.8036.9348.503−1.569−18.453
MgO0.201.008.0996.28812.90113.2824.8036.9948.503−1.509−17.748
MgO0.300.008.0996.46112.90112.9054.8036.4448.503−2.059−24.209
MgO0.300.758.0996.32612.90113.3664.8037.0418.503−1.462−17.200
MgO0.301.008.0996.30212.90113.3944.8037.0928.503−1.411−16.592
MgO0.480.008.0996.47012.90113.0904.8036.6208.503−1.883−22.146
MgO0.480.758.0996.33512.90113.4094.8037.0748.503−1.429−16.811
MgO0.481.008.0996.31112.90113.4284.8037.1178.503−1.386−16.304
NaF0.200.000.999−1.4857.3896.3966.3907.88112.199−4.318−35.396
NaF0.200.750.999−1.6527.3897.9996.3909.65212.199−2.547−20.883
NaF0.201.000.999−1.6797.3898.0356.3909.71412.199−2.485−20.370
NaF0.300.000.999−1.4707.3896.5406.3908.01012.199−4.189−34.337
NaF0.300.750.999−1.6397.3898.0026.3909.64112.199−2.558−20.971
NaF0.301.000.999−1.6667.3898.0356.3909.70012.199−2.499−20.482
NaF0.480.000.999−1.4587.3896.7436.3908.20112.199−3.998−32.772
NaF0.480.750.999−1.6277.3898.0056.3909.63212.199−2.567−21.039
NaF0.481.000.999−1.6547.3898.0346.3909.68812.199−2.511−20.586
Ne0.200.00−9.071−104.2152.545−109.06211.616−4.84721.700−26.547−122.336
Ne0.200.75−9.071−12.7212.5456.01611.61618.73721.700−2.963−13.654
Ne0.201.00−9.071−12.7562.5456.17711.61618.93321.700−2.767−12.749
Ne0.300.00−9.071−109.3892.545−96.80811.61612.58121.700−9.119−42.021
Ne0.300.75−9.071−12.7162.5455.83311.61618.54921.700−3.151−14.521
Ne0.301.00−9.071−12.7512.5455.98011.61618.73221.700−2.968−13.680
Ne0.480.00−9.071−74.7442.545−36.43411.61638.31021.70016.61076.544
Ne0.480.75−9.071−12.7122.5455.61911.61618.33021.700−3.370−15.529
Ne0.481.00−9.071−12.7472.5455.75111.61618.49921.700−3.201−14.753
Si0.200.006.2345.6916.8126.9730.5781.2821.2280.0544.406
Si0.200.756.2345.6186.8127.0340.5781.4161.2280.18815.309
Si0.201.006.2345.6016.8127.0420.5781.4411.2280.21317.386
Si0.300.006.2345.7556.8126.9730.5781.2181.228−0.010−0.790
Si0.300.756.2345.6926.8127.0340.5781.3421.2280.1149.267
Si0.301.006.2345.6776.8127.0420.5781.3651.2280.13711.132
Si0.480.756.2345.7356.8127.0330.5781.2981.2280.0705.692
Si0.481.006.2345.7226.8127.0410.5781.3191.2280.0917.427
SiC0.200.009.4968.69010.85710.9611.3612.2712.595−0.324−12.470
SiC0.200.759.4968.61710.85711.1841.3612.5672.595−0.028−1.094
SiC0.201.009.4968.60110.85711.2161.3612.6152.5950.0200.759
SiC0.300.009.4968.70710.85710.9021.3612.1952.595−0.400−15.399
SiC0.300.759.4968.63510.85711.0511.3612.4162.595−0.179−6.913
SiC0.301.009.4968.61910.85711.0691.3612.4502.595−0.145−5.580
SiC0.480.009.4968.86310.85710.8941.3612.0312.595−0.564−21.738
SiC0.480.759.4968.81710.85711.0081.3612.1912.595−0.404−15.584
SiC0.481.009.4968.80710.85711.0211.3612.2142.595−0.381−14.678

Table 3. This table (continued in Table 4): Choosing the γ\gamma and λ\lambda parameters in bulk systems (see Fig. 1). The k-mesh is as in Table 2. Reference gap (Ref.) is experimental − ZPR from Table 1.

Local occupation λ0iiσ\lambda_{0ii\sigma}Average energy ⟨h⟩0iσ\langle h\rangle_{0i\sigma} (eV)Spatial spread ⟨r2⟩0iσ\langle r^2\rangle_{0i\sigma} (a02a_0^2)
k=6×6×6k = 6 \times6 \times6k=8×8×8k = 8 \times8 \times8k=6×6×6k = 6 \times6 \times6k=8×8×8k = 8 \times8 \times8k=6×6×6k = 6 \times6 \times6k=8×8×8k = 8 \times8 \times8
1.00001.0000−3.075373−3.0650851.8270051.840974
0.99950.99974.7028813.0412652.3133043.948855
0.99950.99884.7031435.5117132.3132113.489266
0.99950.99884.7033875.5327602.3131903.553196
0.00080.001414.52651514.5831344.2156744.342439
0.00040.000815.86612215.7541404.3087664.357879
0.00020.000416.85164016.7353313.5980733.649210
0.00000.000118.79131718.5443734.5293174.627405

Table 6. The DLWFs (γ=0.47714\gamma= 0.47714) for BP, arranged in order of increasing energy. The local occupations are the diagonal elements of the density matrix ρ\rho in the DLWF basis, λ0iiσ=⟨w0iσ∣ρ∣w0iσ⟩\lambda_{0ii\sigma} = \langle w_{0i\sigma}|\rho|w_{0i\sigma}\rangle; they are real, with 0≤λ0iiσ≤10 \le\lambda_{0ii\sigma} \le1, so they measure DLWF occupation. Observe the threefold near-degeneracy of the occupied DLWFs in the 6×6×66 \times6 \times6 kk-sampled calculation and how it differs from the occupied submanifold in the 8×8×88 \times8 \times8 calculation.

Differences in spatial localization change, especially, the Coulomb interaction ⟨ρ0iσ∣f^Hσσ′∣ρRjσ⟩\langle\rho_{0i\sigma}|\hat{f}^{\sigma\sigma'}_{\mathrm{H}}|\rho_{Rj\sigma}\rangle in the curvature, leading to a different correction to the band gap. In particular, observe from Figure 2 (d) that the 6×6×66 \times6 \times6 calculation has a substantially higher gap than its 8×8×88 \times8 \times8 kk-sampled counterpart, corresponding to the increased spatial localization (smaller ⟨r2⟩\langle r^2\rangle) of its occupied DLWFs.

We are currently seeking an alternative to the DLWF minimization that avoids this weakness. Since there is no physical principle that specifies what γ\gamma should be, we believe that a deterministic algorithm—yielding Wannier functions approximately localized in both space and energy—will improve the robustness of olLOSC.

Figure 1 (PDF p. 37). (a) Mean absolute error (%) for small-gap (≤8\le8 eV) systems. (b) Mean absolute error (%) for all systems. (c) Mean signed error (%) for small-gap (≤8\le8 eV) systems. (d) Mean signed error (%) for all systems. (e) Mean absolute error (eV) for small-gap (≤8\le8 eV) systems. (f) Mean absolute error (eV) for all systems. (g) Mean signed error (eV) for small-gap (≤8\le8 eV) systems. (h) Mean signed error (eV) for all systems. Figure 1: Errors by DLWF mixing γ\gamma and the vW λ\lambda in the materials tested. (a–d): Relative errors (%); (e–h): unscaled errors (eV). (a, c, e, g): Small-gapped systems (gaps ≤8\le8 eV), for which olLOSC is most accurate; (b, d, f, h): All systems. (a, b, e, f): Absolute values; (c, d, g, h): Signed errors.

SystemLatticeaa (Å)Gap (eV)ZPR (eV)Gap − ZPR (eV)
CA43.575.48−0.32 [40]5.80
SiA45.431.17−0.06 [40]1.23
SiCB34.362.42−0.17 [40]2.59
GeA45.660.74−0.03 [41]0.77
MgOB14.217.97−0.53 [40]7.75
LiFB14.0214.20−1.23 [40]15.43
LiClB15.139.40−0.44 [42]9.84
NaFB14.6211.50−0.70 [42]12.20
BNB33.626.22−0.40 [40]6.62
BPB34.542.40−0.10 [42]2.50
AlPB35.462.51−0.10 [40]2.61
NeA14.4321.70n/a21.70
ArA15.2614.20n/a14.20

Table 1 (PDF p. 44). The bulk systems studied. The lattice is defined by the Strukturbericht symbol (A4: diamond; B3: zincblende; B1: halite; B3: wurtzite; A1: simple cubic). aa is the experimental lattice parameter, in angstrom; these are taken from [35] when available, [36] otherwise. The gap is experimental, taken from [37] (LiF, LiCl, Ne, Ar); [38] (NaF); [39 p. 184] (MgO); and [35] and containing references (the rest). ZPR is zero-point renormalization, computed in the indicated reference for each system.

Systemk-meshγ\gammaλ\lambdaEPBEE_{\mathrm{PBE}}ΔEolLOSC\Delta E_{\mathrm{olLOSC}}VBM PBE (eV)VBM olLOSC (eV)CBM PBE (eV)CBM olLOSC (eV)Gap PBE (eV)Gap olLOSC (eV)Ref.olLOSC gap error eV%
AlP060.300.75−18.581.64×10−31.64 \times10^{-3}5.0214.1506.5976.8161.5762.6662.6060.0602.306
AlP080.300.75−18.582.13×10−32.13 \times10^{-3}5.0164.1726.5956.7921.5792.6202.6060.0140.553
AlP060.301.00−18.581.81×10−31.81 \times10^{-3}5.0214.0996.5976.8561.5762.7572.6060.1515.810
Ar060.300.75−45.185.44×10−55.44 \times10^{-5}−3.948−5.9094.7555.0798.70310.98814.200−3.212−22.620
Ar080.300.75−45.185.07×10−55.07 \times10^{-5}−3.948−5.7594.7554.7608.70310.51914.200−3.681−25.923
Ar060.301.00−45.186.65×10−56.65 \times10^{-5}−3.948−6.0344.7555.1548.70311.18814.200−3.012−21.213
BN060.300.75−26.812.97×10−52.97 \times10^{-5}11.1469.71615.67716.4954.5316.7786.6220.1562.362
BN080.300.75−26.814.12×10−54.12 \times10^{-5}11.1439.80015.67616.7754.5336.9756.6220.3535.331
BN060.301.00−26.813.24×10−53.24 \times10^{-5}11.1469.64115.67716.5604.5316.9206.6220.2984.494
BP060.300.75−19.797.10×10−47.10 \times10^{-4}9.2468.29010.59111.2681.3452.9782.5010.47719.060
BP080.300.75−19.791.03×10−31.03 \times10^{-3}9.2418.31610.52211.1581.2812.8422.5010.34113.643
BP060.301.00−19.797.85×10−47.85 \times10^{-4}9.2468.21510.59111.3321.3453.1172.5010.61624.630
C060.300.75−24.073.92×10−53.92 \times10^{-5}13.35812.78517.56318.2894.2055.5055.803−0.299−5.144
C080.300.75−24.086.78×10−56.78 \times10^{-5}13.35512.82517.54418.2244.1895.3995.803−0.404−6.962
C060.301.00−24.074.04×10−54.04 \times10^{-5}13.35812.77817.56318.3044.2055.5265.803−0.277−4.770
Ge060.300.75−357.283.11×10−33.11 \times10^{-3}9.5618.9329.5939.9750.0321.0430.7730.27034.955
Ge080.300.75−357.284.39×10−34.39 \times10^{-3}9.5548.9339.5859.9510.0311.0180.7730.24531.759
Ge060.301.00−357.283.53×10−33.53 \times10^{-3}9.5618.8669.59310.0320.0321.1650.7730.39250.776
LiCl060.300.75−47.479.32×10−59.32 \times10^{-5}1.6670.2287.9967.8926.3297.6649.836−2.172−22.085
LiCl080.300.75−47.472.26×10−52.26 \times10^{-5}1.6670.3037.9967.4276.3297.1249.836−2.712−27.573
LiCl060.301.00−47.471.23×10−41.23 \times10^{-4}1.6670.1097.9967.9886.3297.8799.836−1.957−19.899
LiF060.300.75−63.982.10×10−52.10 \times10^{-5}0.970−1.70210.15610.2419.18711.94315.431−3.488−22.604
LiF080.300.75−63.982.45×10−52.45 \times10^{-5}0.970−1.53910.15610.0889.18711.62615.431−3.805−24.658
LiF060.301.00−63.982.23×10−52.23 \times10^{-5}0.970−1.72810.15610.2969.18712.02415.431−3.407−22.077
MgO060.300.75−152.101.02×10−41.02 \times10^{-4}8.0996.42612.90113.3624.8036.9368.503−1.567−19.257
MgO080.300.75−152.108.94×10−58.94 \times10^{-5}8.0996.44212.90113.3894.8036.9478.503−1.556−18.304
MgO060.301.00−152.107.97×10−57.97 \times10^{-5}8.0996.30212.90113.3694.8037.0678.503−1.436−16.897
Na060.300.75−140.172.19×10−52.19 \times10^{-5}−9.071−12.6382.5455.79611.61618.43421.700−3.266−15.050
Na080.300.75−140.172.35×10−52.35 \times10^{-5}−9.071−12.4832.4175.38411.48817.86721.700−3.833−17.664
Na060.301.00−140.172.20×10−52.20 \times10^{-5}−9.071−12.6802.5455.79611.61618.47621.700−3.224−14.853
Si060.300.75−16.922.35×10−32.35 \times10^{-3}6.2355.4926.8477.1720.6121.6801.2280.45236.816
Si080.300.75−16.922.13×10−32.13 \times10^{-3}6.2355.5446.8477.1720.6121.7281.2280.50040.713
Si060.301.00−16.922.20×10−32.20 \times10^{-3}6.2355.4416.8477.3130.6121.8721.2280.64452.443
SiC060.300.75−20.545.10×10−45.10 \times10^{-4}9.4638.40810.85711.1231.3612.6812.5950.0863.341
SiC080.300.75−20.546.69×10−46.69 \times10^{-4}9.4638.43910.85711.1231.3612.6842.5950.0893.434
SiC060.301.00−20.545.54×10−45.54 \times10^{-4}9.5038.35610.85911.1661.3562.8102.5950.2158.270

Table 5. kk-point convergence of bulk systems. VBM: valence band maximum; CBM: conduction band minimum. Reference values are the electronic gaps from Table 1. Gap error is EolLOSCg−ErefgE_{\mathrm{olLOSC}}^{g} - E_{\mathrm{ref}}^{g}.

SystemEPBEE_{\mathrm{PBE}} (Eh)ΔE\Delta E (mEh)HOMO PBEHOMO olLOSCLUMO PBELUMO olLOSCGap PBEGap olLOSCGap Ref.olLOSC gap error eVolLOSC gap error %
Acridine-555.067-0.935-5.426-7.809-2.976-0.8302.4496.9797.394-0.415-5.618
Anthracene-539.0210.266-4.979-7.378-2.728-0.6622.2516.7167.214-0.498-6.907
Azulene-385.4782.300-4.970-7.582-2.881-0.4322.0897.1497.0030.1462.089
Benzonitrile-324.201-0.216-6.852-9.848-2.5210.1294.3319.97710.168-0.191-1.874
Benzoquinone-381.160-0.022-6.348-10.256-4.566-1.6741.7818.5828.714-0.132-1.517
Dichlone-1453.559-1.344-6.580-9.914-4.385-2.0942.1957.8198.066-0.247-3.056
Dinitrobenzonitrile-733.0090.171-7.725-11.345-4.583-1.9243.1429.4219.4080.0130.139
Fumaronitrile-262.854-0.331-7.851-11.256-4.094-1.1133.75710.14310.507-0.364-3.469
Maleicanhydride-379.0390.071-7.129-11.342-4.199-1.0832.93010.25910.308-0.049-0.477
NDCA-686.0670.346-6.468-9.019-3.658-1.3412.8107.6787.918-0.240-3.036
Naphthalenedione-534.6820.036-6.151-9.668-4.171-1.6761.9797.9928.403-0.411-4.888
Nitrobenzene-436.435-0.010-6.816-10.809-3.485-0.4623.33210.3479.6970.6506.704
Nitrobenzonitrile-528.608-0.013-7.301-11.082-4.136-1.3853.1659.6969.3480.3483.725
Phenazine-571.1060.090-5.735-8.998-3.404-1.1102.3317.8877.3870.5006.773
Phthalicanhydride-532.560-0.428-6.900-10.598-3.601-0.9453.2999.6539.700-0.047-0.485
Phthalimide-512.703-0.751-6.350-9.990-3.299-0.6743.0519.3159.482-0.167-1.757
TCNE-447.1731.231-8.597-11.816-5.780-3.1652.8178.6518.954-0.303-3.383
TCNQ-678.02211.383-7.050-9.031-5.514-3.5871.5365.4446.262-0.818-13.070
Tetrachloro-isobenzofuranedione-2370.3001.541-7.106-9.671-4.062-1.7633.0447.9088.430-0.522-6.191
Tetrachlorobenzoquinone-2218.9120.299-6.983-9.490-5.010-2.6311.9736.8597.762-0.903-11.637
Tetrafluorobenzenedicarbonitrile-813.1190.201-7.444-9.995-4.133-1.6523.3118.3439.151-0.808-8.829
Tetrafluorobenzoquinone-777.9130.071-7.374-10.361-5.127-2.3482.2488.0138.861-0.848-9.571
mDCBN-416.374-0.150-7.393-10.172-3.288-0.7644.1059.4099.832-0.423-4.307

Table 8. Large organic molecules olLOSC calculation results with λ=1.0\lambda= 1.0 and γ=0.47714\gamma= 0.47714. (See b and e in Figure 3 and Figure 4) Reference gap (Ref.) is CCSD-T benchmark from Ref[43].

SystemTotal energy EPBEE_{\mathrm{PBE}} (EhE_{\mathrm{h}})ΔE\Delta E (mEhE_{\mathrm{h}})HOMO (eV) PBEHOMO (eV) olLOSCLUMO (eV) PBELUMO (eV) olLOSCGap (eV) PBEGap (eV) olLOSCGap (eV) Ref.olLOSC gap error (eV)olLOSC gap error (%)
CF2-237.570.03-7.36-12.22-3.730.683.6312.9012.680.221.71
CH3O-114.960.05-6.06-11.18-5.34-1.010.7310.179.510.666.93
CN-92.641.12-9.38-14.47-7.67-3.641.7110.8310.290.545.28
Cl2-920.04-0.00-7.42-11.73-4.66-0.922.7610.8110.610.201.89
H2C-39.110.01-5.73-10.93-3.32-0.462.4110.479.570.909.40
H2CS-437.270.02-5.54-9.41-3.64-0.111.909.299.160.131.44
H2N-55.830.03-7.27-12.90-4.54-0.282.7312.6211.371.2510.96
H2P-342.350.03-6.05-10.11-4.20-1.201.858.918.560.354.09
H2Si-290.460.07-5.86-9.43-4.06-1.241.808.198.48-0.29-3.42
H3C-39.790.00-5.42-10.44-2.950.242.4810.689.840.848.52
H3Si-291.080.13-5.36-9.12-3.55-1.121.827.997.930.060.79
HCO-113.780.14-5.00-9.66-3.300.491.7010.159.440.717.53
HN-55.170.03-7.91-14.23-4.320.023.6014.2613.151.118.40
HO-75.680.01-7.36-14.01-6.44-1.140.9212.8711.241.6314.51
HP-341.730.03-6.19-10.30-3.97-1.052.219.259.160.090.96
HS-398.590.02-6.22-10.53-5.79-2.240.438.298.050.243.01
S2-796.07-0.01-5.91-9.56-4.61-1.301.308.267.920.344.29

Table 9. Small molecules olLOSC calculation results with λ=0.75\lambda= 0.75 and γ=0.3\gamma= 0.3. (See a and d in Figure 3 and Figure 4) Reference gap (Ref.) is CCSD-T benchmark from [43].

SystemTotal energy EPBEE_{\mathrm{PBE}} (EhE_{\mathrm{h}})ΔE\Delta E (mEh\mathrm{m}E_{\mathrm{h}})HOMO (eV) PBEHOMO (eV) olLOSCLUMO (eV) PBELUMO (eV) olLOSCGap (eV) PBEGap (eV) olLOSCGap (eV) Ref.olLOSC gap error (eV)olLOSC gap error (%)
CF2-237.570.00-7.36-12.30-3.730.803.6313.1112.680.433.37
CH3O-114.960.01-6.06-11.26-5.34-0.900.7310.359.510.848.84
CN-92.640.12-9.38-14.61-7.67-3.491.7111.1210.290.838.05
Cl2-920.04-0.00-7.42-11.41-4.66-0.832.7610.5910.61-0.02-0.20
H2C-39.110.00-5.73-11.06-3.32-0.322.4110.749.571.1712.22
H2CS-437.270.00-5.54-9.49-3.64-0.021.909.469.160.303.32
H2N-55.830.01-7.27-13.05-4.54-0.132.7312.9211.371.5513.64
H2P-342.350.01-6.05-10.21-4.20-1.101.859.108.560.546.36
H2Si-290.460.02-5.86-9.51-4.06-1.161.808.348.48-0.14-1.63
H3C-39.790.00-5.42-10.57-2.950.362.4810.939.841.0911.05
H3Si-291.080.03-5.36-9.19-3.55-1.011.828.187.930.253.19
HCO-113.780.04-5.00-9.78-3.300.641.7010.429.440.9810.36
HN-55.170.01-7.91-14.41-4.320.173.6014.5913.151.4410.91
HO-75.680.00-7.36-14.21-6.44-0.950.9213.2611.242.0217.97
HP-341.730.01-6.19-10.40-3.97-0.952.219.459.160.293.16
HS-398.590.00-6.22-10.65-5.79-2.130.438.538.050.485.94
S2-796.07-0.00-5.91-9.64-4.61-1.221.308.437.920.516.39

Table 10. Small molecules olLOSC calculation results with λ=1.0\lambda= 1.0 and γ=0.47714\gamma= 0.47714. (See a and d in Figure 3 and Figure 4) Reference gap (Ref.) is CCSD-T benchmark from Ref[43].

H(C₂H₂)ₙHLOSC γvWλEPBEE_{\mathrm{PBE}} (EhE_{\mathrm{h}})ΔE (mEhE_{\mathrm{h}})PBEolLOSCRef.eV%
Total energyHOMO (eV)olLOSC error
10.3000.75-78.5010.000-6.722-10.870-10.481-0.3893.712
10.4771.00-78.5010.000-6.722-10.975-10.481-0.4944.712
20.3000.75-155.8310.165-5.819-9.306-9.176-0.1301.415
20.4771.00-155.8310.038-5.819-9.385-9.176-0.2092.282
30.3000.75-233.1621.026-5.366-8.390-8.177-0.2132.599
30.4771.00-233.1620.498-5.366-8.415-8.177-0.2382.915
40.3000.75-310.4942.130-5.094-7.832-7.691-0.1411.835
40.4771.00-310.4941.173-5.094-7.877-7.691-0.1862.414
50.3000.75-387.8274.028-4.913-7.372-7.329-0.0430.581
50.4771.00-387.8272.149-4.913-7.314-7.3290.015-0.202
60.3000.75-465.1604.596-4.783-7.076-7.044-0.0320.451
60.4771.00-465.1602.587-4.783-7.028-7.0440.016-0.232
70.3000.75-542.4936.656-4.685-6.795-6.8520.057-0.836
70.4771.00-542.4933.667-4.685-6.764-6.8520.088-1.285
80.3000.75-619.8267.781-4.609-6.631-6.6620.031-0.471
80.4771.00-619.8264.656-4.609-6.585-6.6620.077-1.158
90.3000.75-697.1599.879-4.548-6.440-6.5550.115-1.751
90.4771.00-697.1595.750-4.548-6.400-6.5550.155-2.362
100.3000.75-774.49310.759-4.499-6.313-6.4100.097-1.515
100.4771.00-774.4936.569-4.499-6.263-6.4100.147-2.294

Table 11. All trans-polyacetylene (H(C₂H₂)ₙH olLOSC HOMO energies with {λ=0.75\lambda= 0.75 and γ=0.30\gamma= 0.30} and {λ=1.00\lambda= 1.00 and γ=0.47714\gamma= 0.47714} (See a, b, e, f in Figure 7) Reference energies is RASPT2 calculated negative IP energies from Ref[44].

Bar charts of olLOSC errors in ionization energy for all-trans polyacetylene

Figure 7. olLOSC errors in ionization energy for all-trans polyacetylene with different LOSC γ\gamma and vW λ\lambda. (a,b): Percentage error for olLOSC within RPA; (c,d): Percentage error for olLOSC beyond RPA; (e,f): Unscaled error for olLOSC within RPA; (g,h): Unscaled error for olLOSC beyond RPA.

MoleculeElectronsBasis functionsolLOSC wall time (s)
H(HC=CH)1HH(\mathrm{HC} = \mathrm{CH})_{1}H1613015.15
H(HC=CH)2HH(\mathrm{HC} = \mathrm{CH})_{2}H3023034.17
H(HC=CH)3HH(\mathrm{HC} = \mathrm{CH})_{3}H4433065.85
H(HC=CH)4HH(\mathrm{HC} = \mathrm{CH})_{4}H58430116.92
H(HC=CH)5HH(\mathrm{HC} = \mathrm{CH})_{5}H72530188.40
H(HC=CH)6HH(\mathrm{HC} = \mathrm{CH})_{6}H86630291.36
H(HC=CH)7HH(\mathrm{HC} = \mathrm{CH})_{7}H100730436.64
H(HC=CH)8HH(\mathrm{HC} = \mathrm{CH})_{8}H114830638.80
H(HC=CH)9HH(\mathrm{HC} = \mathrm{CH})_{9}H128930868.04
H(HC=CH)10HH(\mathrm{HC} = \mathrm{CH})_{10}H14210301209.56

Table 14. Wall times in seconds for the olLOSC κ\kappa evaluation for polyacetylene using the cc-pVTZ basis set and cc-pVTZ-RIFIT as the auxiliary basis. The olLOSC method is implemented in QM4^{4}D and executed on a 32-core Intel(R) Xeon(R) Gold 6226R CPU @ 2.90 GHz, with a NVIDIA A100 GPU employed to accelerate the numerical integration. From the bolded wall times, numerical fitting indicates that the computational cost scales as O(Nelectrons2.49)O(N_{\mathrm{electrons}}^{2.49}) with R2=0.9952R^{2} = 0.9952 or, alternatively, as O(MCGTOs2.56)O(M_{\mathrm{CGTOs}}^{2.56}) with R2=0.9958R^{2} = 0.9958.

References

  1. [1]F. Bloch, “Über die Quantenmechanik der Elektronen in Kristallgittern”, Z. Physik 52, 555–600 (1929).DOI
  2. [2]G. H. Wannier, “The Structure of Electronic Excitation Levels in Insulating Crystals”, Phys. Rev. 52, 191–197 (1937).DOI
  3. [3]A. Mahler, J. Z. Williams, N. Q. Su, and W. Yang, Wannier Functions Dually Localized in Space and Energy, Jan. 2022.DOI
  4. [4]N. Marzari and D. Vanderbilt, “Maximally localized generalized Wannier functions for composite energy bands”, Phys. Rev. B 56, 12847–12865 (1997).DOI
  5. [5]J. M. Foster and S. F. Boys, “Canonical Configurational Interaction Procedure”, Rev. Mod. Phys. 32, 300–302 (1960).
  6. [6]S. Baroni, S. de Gironcoli, A. Dal Corso, and P. Giannozzi, “Phonons and related crystal properties from density-functional perturbation theory”, Rev. Mod. Phys. 73, 515–562 (2001).DOI
  7. [7]I. Timrov, N. Marzari, and M. Cococcioni, “Hubbard parameters from density-functional perturbation theory”, Phys. Rev. B 98, 085127 (2018).DOI
  8. [8]H. J. Monkhorst and J. D. Pack, “Special points for Brillouin-zone integrations”, Phys. Rev. B 13, 5188–5192 (1976).DOI
  9. [9]N. Colonna, R. De Gennaro, E. Linscott, and N. Marzari, “Koopmans Spectral Functionals in Periodic Boundary Conditions”, J. Chem. Theory Comput. 18, 5435–5448 (2022).arxiv.org/abs/2202.08155
  10. [10]R. M. Sternheimer, “Electronic Polarizabilities of Ions from the Hartree-Fock Wave Functions”, Phys. Rev. 96, 951–968 (1954).DOI
  11. [11]J. Z. Williams and W. Yang, Correcting Delocalization Error in Materials with Localized Orbitals and Linear-Response Screening, July 2024.DOI
  12. [12]G. L. Oliver and J. P. Perdew, “Spin-density gradient expansion for the kinetic energy”, Phys. Rev. A 20, 397–403 (1979).
  13. [13]P. K. Chattaraj, D. R. Roy, P. Geerlings, and M. Torrent-Sucarrat, “Local hardness: a critical account”, Theor Chem Account 118, 923–930 (2007).DOI
  14. [14]F. Della Sala, “Orbital-free methods for plasmonics: Linear response”, J. Chem. Phys. 157, 104101 (2022).DOI
  15. [15]O. Vahtras, J. Almlöf, and M. W. Feyereisen, “Integral approximations for LCAO-SCF calculations”, Chem. Phys. Lett. 213, 514–518 (1993).DOI
  16. [16]X. Ren, P. Rinke, V. Blum, J. Wieferink, A. Tkatchenko, A. Sanfilippo, K. Reuter, and M. Scheffler, “Resolution-of-identity approach to Hartree–Fock, hybrid density functionals, RPA, MP2 and GW with numeric atom-centered orbital basis functions”, New J. Phys. 14, 053020 (2012).DOI
  17. [17]D. M. York and W. Yang, “A chemical potential equalization method for molecular simulations”, J. Chem. Phys. 104, 159–172 (1996).DOI
  18. [18]N. I. M. Gould, M. E. Hribar, and J. Nocedal, “On the Solution of Equality Constrained Quadratic Programming Problems Arising in Optimization”, SIAM J. Sci. Comput. 23, 1376–1395 (2001).DOI
  19. [19]S. Hirata and M. Head-Gordon, “Time-dependent density functional theory for radicals An improved description of excited states with substantial double excitation character”, Chem. Phys. Lett. 302, 375–382 (1999).
  20. [20]W. Wirtinger, “Zur formalen Theorie der Funktionen von mehr komplexen Veränderlichen”, Math. Ann. 97, 357–375 (1927).DOI
  21. [21]Y. Mei, Z. Chen, and W. Yang, “Self-Consistent Calculation of the Localized Orbital Scaling Correction for Correct Electron Densities and Energy-Level Alignments in Density Functional Theory”, J. Phys. Chem. Lett. 11, 10269–10277 (2020).DOI
  22. [22]C. Li, X. Zheng, N. Q. Su, and W. Yang, “Localized orbital scaling correction for systematic elimination of delocalization error in density functional approximations”, Nat. Sci. Rev. 5, 203–215 (2018).DOI
  23. [23]N. Q. Su, A. Mahler, and W. Yang, “Preserving Symmetry and Degeneracy in the Localized Orbital Scaling Correction Approach”, J. Phys. Chem. Lett. 11, 1528–1535 (2020).
  24. [24]A. Mahler, J. Williams, N. Q. Su, and W. Yang, “Localized orbital scaling correction for periodic systems”, Phys. Rev. B 106, 035147 (2022).DOI
  25. [25]Y. Fan, J. Z. Williams, and W. Yang, Data and scripts from: ollosc: unified and efficient density functional approximation to correct delocalization error in molecules and periodic materials [data set], Duke Research Data Repository, 2026.DOI
  26. [26]J. Zheng, Y. Zhao, and D. G. Truhlar, “The dbh24/08 database and its use to assess electronic structure model chemistries for chemical reaction barrier heights”, Journal of Chemical Theory and Computation 5, 808–821 (2009).
  27. [27]R. Peverati and D. G. Truhlar, “Quest for a universal density functional: the accuracy of density functionals across a broad spectrum of databases in chemistry and physics”, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 372, 10.1098/RSTA.2012.0476 (2014).
  28. [28]F. Gygi, J.-L. Fattebert, and E. Schwegler, “Computation of Maximally Localized Wannier Functions using a simultaneous diagonalization algorithm”, Computer Physics Communications 155, 1–6 (2003).DOI
  29. [29]I. Souza, N. Marzari, and D. Vanderbilt, “Maximally localized Wannier functions for entangled energy bands”, Phys. Rev. B 65, 035109 (2001).DOI
  30. [30]D. Monaco, G. Panati, A. Pisante, and S. Teufel, “Optimal Decay of Wannier functions in Chern and Quantum Hall Insulators”, Commun. Math. Phys. 359, 61–100 (2018).DOI
  31. [31]V. Vitale, G. Pizzi, A. Marrazzo, J. R. Yates, N. Marzari, and A. A. Mostofi, “Automated high-throughput Wannierization”, npj Comput Mater 6, 1–18 (2020).
  32. [32]A. Damle, L. Lin, and L. Ying, “Compressed Representation of Kohn–Sham Orbitals via Selected Columns of the Density Matrix”, J. Chem. Theory Comput. 11, 1463–1469 (2015).DOI
  33. [33]A. Damle, L. Lin, and L. Ying, “SCDM-k: Localized orbitals for solids via selected columns of the density matrix”, Journal of Computational Physics 334, 1–15 (2017).arxiv.org/abs/1507.03354
  34. [34]E. Prodan and W. Kohn, “Nearsightedness of electronic matter”, PNAS 102, 11635–11638 (2005).pubmed.ncbi.nlm.nih.gov/16087868
  35. [35]J. Heyd, J. E. Peralta, G. E. Scuseria, and R. L. Martin, “Energy band gaps and lattice parameters evaluated with the Heyd-Scuseria-Ernzerhof screened hybrid functional”, J. Chem. Phys. 123, 174101 (2005).pubmed.ncbi.nlm.nih.gov/16375511
  36. [36]R. W. G. Wyckoff, Crystal Structures, 2nd (Wiley, New York, 1973).
  37. [37]F. Tran and P. Blaha, “Accurate Band Gaps of Semiconductors and Insulators with a Semilocal Exchange-Correlation Potential”, Phys. Rev. Lett. 102, 226401 (2009).pubmed.ncbi.nlm.nih.gov/19658882
  38. [38]R. T. Poole, J. G. Jenkin, J. Liesegang, and R. C. G. Leckey, “Electronic band structure of the alkali halides. I. Experimental parameters”, Phys. Rev. B 11, 5179–5189 (1975).
  39. [39]O. Madelung, Semiconductors: Data Handbook (Springer Berlin Heidelberg, Berlin, Heidelberg, 2004).DOI
  40. [40]M. Engel, H. Miranda, L. Chaput, A. Togo, C. Verdi, M. Marsman, and G. Kresse, “Zero-point renormalization of the band gap of semiconductors and insulators using the projector augmented wave method”, Phys. Rev. B 106, 094316 (2022).DOI
  41. [41]A. Miglio, V. Brousseau-Couture, E. Godbout, G. Antonius, Y.-H. Chan, S. G. Louie, M. Côté, M. Giantomassi, and X. Gonze, “Predominance of non-adiabatic effects in zero-point renormalization of the electronic band gap”, npj Comput Mater 6, 1–8 (2020).DOI
  42. [42]H. Shang, J. Zhao, and J. Yang, “Assessment of the Mass Factor for the Electron–Phonon Coupling in Solids”, J. Phys. Chem. C 125, 6479–6485 (2021).DOI
  43. [43]R. M. Richard, M. S. Marshall, O. Dolgounitcheva, J. V. Ortiz, J. L. Brédas, N. Marom, and C. D. Sherrill, “Accurate ionization potentials and electron affinities of acceptor molecules i. reference data at the ccsd(t) complete basis set limit”, Journal of Chemical Theory and Computation 12, 595–604 (2016).DOI
  44. [44]A. R. M. Shahi, C. J. Cramer, and L. Gagliardi, “Second-order perturbation theory with complete and restricted active space reference functions applied to oligomeric unsaturated hydrocarbons”, Physical Chemistry Chemical Physics 11, 10964–10972 (2009).DOI

Paper details

vlnc:W_k6ae2225
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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