betaFeedback

WorksMathematics

Fourier Transform-Based Computation of Two-Electron Four-Center Coulomb Repulsion Integrals

Abstract

We present a Fourier-based method for evaluating two-electron four-center Coulomb repulsion integrals over Slater-type orbitals. Spherical Fourier-Bessel coefficients of the convolution factors are precomputed and reused. Since the spherical part of the integral can be evaluated directly, there remains only a one-dimensional quadrature problem in the radial direction. For the ten reference integrals reported here, the largest absolute discrepancy is 1.7×10−91.7\times10^{-9}. A preliminary resident-cache benchmark computes approximately 4.2×1054.2\times10^5 unique integrals per second on an NVIDIA V100 at the contraction stage.

Introduction

Fast ab initio calculations of molecular properties require efficient evaluation of electron repulsion integrals (ERIs). In orbital-based methods, molecular orbitals are expressed as linear combinations of basis functions [1, 2]. Each ERI describes the Coulomb interaction between two orbital-pair densities and involves four basis functions. Slater-type orbitals (STOs) [3] are a possible choice. They give rise to accurate computations using fewer basis functions compared to competing approaches, such as Gaussian-type orbitals, but the resulting multicenter integrals are known to be difficult to evaluate.

The ERI can be viewed as a convolution of two functions, each of which is the product of two STOs, mediated by the Coulomb potential as a kernel. It is therefore natural to replace the six-dimensional integral with the three-dimensional integration of the Fourier transforms of the two factors against the Fourier transform of the potential. A primary computational challenge is to represent these transforms compactly and reuse them across the many integrals sharing the same orbital pairs.

Partial-wave analysis has long been central to this approach. Trivedi and Steinborn derived a two-center transform of reduced Bessel functions explicitly suited to partial-wave decomposition and proposed reusing previously computed transforms [4]. Grotendorst and Steinborn subsequently developed a general multicenter treatment using BB functions, with a three-dimensional numerical integration remaining for two-electron integrals [5]. Direct STO formulations include the shift-operator construction of Niehaus, López, and Rico [6] and the one-dimensional, non-oscillatory hypergeometric representation of Vuković and Dmitrović [7]. These developments provide analytic representations of pair transforms and multicenter integrals.

Our focus is on the organization of reusable local pair data and angular contractions for the evaluation of large batches of ERIs on GPUs. We developed a localized, reusable partial-wave formulation. A pair-specific Becke partition suppresses the off-center nuclear structure, allowing each component to be expanded about its own center. The center phases are kept outside these expansions and analytically integrated during contraction. All local harmonic channels are stored at each radial momentum node, and the resulting finite angular couplings are evaluated in GPU batches.

The remainder of the paper is organized as follows. § 2 presents the Fourier formulation of the ERIs. In § 3 we construct the localized pair representation and its reusable coefficients. In § 4 the analytic angular contraction is derived. In § 5 we describe the cache layout and computational organization. Numerical results are presented in § 6, followed by brief concluding remarks in § 7.

Fourier formulation

Let {ϕμ}μ=1N\{\phi_\mu\}_{\mu=1}^{N} be real, normalized STOs,

ϕμ(r)=Nμrμnμ−1e−ζμrμYℓμmμR(rμrμ),\phi_\mu(\mathbf r)=\mathcal{N}_\mu r_\mu^{n_\mu-1}e^{-\zeta_\mu r_\mu}Y^{\mathrm{R}}_{\ell_\mu m_\mu}\left(\frac{\mathbf r_\mu}{r_\mu}\right),

where rμ=r−Aμ\mathbf r_\mu=\mathbf r-\mathbf A_\mu, rμ=∣rμ∣r_\mu=\lvert\mathbf r_\mu\rvert, nμn_\mu is the principal index, ζμ>0\zeta_\mu>0, and YℓmRY^{\mathrm{R}}_{\ell m} denotes an orthonormal basis of real spherical harmonics. For p=(μ,ν)p=(\mu,\nu) with μ≤ν\mu\le\nu, set ρp=ϕμϕν\rho_p=\phi_\mu\phi_\nu. There are M=N(N+1)/2M=N(N+1)/2 such pairs. The ERI associated with p=(μ,ν)p=(\mu,\nu) and q=(λ,σ)q=(\lambda,\sigma) is

Cpq=(μν∣λσ)=∬R3×R3ρp(r)ρq(r′)∣r−r′∣ dr dr′.C_{pq}=(\mu\nu\mid\lambda\sigma)=\iint_{\mathbb{R}^{3}\times\mathbb{R}^{3}}\frac{\rho_p(\mathbf r)\rho_q(\mathbf r')}{\lvert\mathbf r-\mathbf r'\rvert}\,\mathrm{d}\mathbf r\,\mathrm{d}\mathbf r'.

Denoting the Fourier transform by

Fp(k)=∫R3ρp(r)e−ik⋅r dr,F_p(\mathbf{k}) = \int_{\mathbb{R}^3} \rho_p(\mathbf{r})e^{-i\mathbf{k}\cdot\mathbf{r}}\,d\mathbf{r},

the convolution theorem and Parseval’s identity give

Cpq=1(2π)3∫R34π∣k∣2Fp(k)‾Fq(k) dk.C_{pq} = \frac{1}{(2\pi)^3}\int_{\mathbb{R}^3}\frac{4\pi}{|\mathbf{k}|^2}\overline{F_p(\mathbf{k})}F_q(\mathbf{k})\,d\mathbf{k}.

Writing k=kω\mathbf{k}=k\boldsymbol{\omega}, with ω∈S2\boldsymbol{\omega}\in S^2 and k≥0k\geq0, we have the volume element dk=k2 dk dωd\mathbf{k}=k^2\,dk\,d\boldsymbol{\omega}. The spherical Jacobian cancels the k−2k^{-2} Coulomb multiplier, giving

Cpq=12π2∫0∞∫S2Fp(kω)‾Fq(kω) dω dk.(1)C_{pq}=\frac{1}{2\pi^2}\int_0^\infty\int_{S^2}\overline{F_p(k\boldsymbol{\omega})}F_q(k\boldsymbol{\omega})\,d\boldsymbol{\omega}\,dk. \tag*{(1)}

A common-origin expansion Fp(kω)=∑ℓmfp,ℓm(k)Yℓm(ω)F_p(k\boldsymbol{\omega})=\sum_{\ell m}f_{p,\ell m}(k)Y_{\ell m}(\boldsymbol{\omega}) consequently gives

Cpq=12π2∫0∞∑ℓmfp,ℓm(k)‾fq,ℓm(k) dk.C_{pq}=\frac{1}{2\pi^2}\int_0^\infty\sum_{\ell m}\overline{f_{p,\ell m}(k)}f_{q,\ell m}(k)\,dk.

Unfortunately, using this will result in a larger cache size, described in § 5, because a translated spherical density has the transform g^(k)e−ikω⋅X\widehat{g}(k)e^{-i k\boldsymbol{\omega}\cdot\mathbf{X}}, whose Rayleigh coefficients contain jℓ(k∣X∣)j_\ell(k|\mathbf{X}|) through degrees of order k∣X∣k|\mathbf{X}|. For a molecule of radius RsysR_{\mathrm{sys}} relative to the common origin, the global cutoff can therefore grow on the scale of kRsyskR_{\mathrm{sys}}. We instead retain translations explicitly and expand only locally-centered amplitudes.

Localized representation

Consider a pair with distinct centers A\mathbf{A} and B\mathbf{B}. An expansion about A\mathbf{A} encounters the displaced nuclear cusp on shells passing through B\mathbf{B}, where a radial feature becomes a localized angular feature. We use a two-center Becke partition [8] to suppress that feature in the A\mathbf{A}-centered component:

ρp=ρpA+ρpB,ρpA=wAρp,ρpB=wBρp.\rho_p=\rho_p^A+\rho_p^B,\qquad\rho_p^A=w_A\rho_p,\qquad\rho_p^B=w_B\rho_p.

Let dAB=∣A−B∣d_{AB}=|\mathbf{A}-\mathbf{B}| and define

t(r)=∣r−A∣−∣r−B∣dAB,s(t)=32t−12t3,t(\mathbf{r})=\frac{|\mathbf{r}-\mathbf{A}|-|\mathbf{r}-\mathbf{B}|}{d_{AB}},\qquad s(t)=\frac{3}{2}t-\frac{1}{2}t^3,
wA(r)=12[1−s(s(s(t(r))))],wB(r)=1−wA(r).w_A(\mathbf{r})=\frac{1}{2}\left[1-s\left(s\left(s(t(\mathbf{r}))\right)\right)\right],\qquad w_B(\mathbf{r})=1-w_A(\mathbf{r}).

The partition is exact before truncation. More importantly, endpoint flatness gives wA=O((rB/dAB)8)w_A=O((r_B/d_{AB})^8) near B\mathbf{B}, where rB=∣r−B∣r_B=|\mathbf{r}-\mathbf{B}|, and conversely for wBw_B near A\mathbf{A}. Each component therefore suppresses the other center’s cusp to high order while retaining its own in local radial coordinates. Same-center pairs require no split.

Let C(p)C(p) denote the resulting one or two component centers. For X∈C(p)X \in C(p), write r=X+rΩ\mathbf{r}=\mathbf{X}+r\mathbf{\Omega} and project the component onto real spherical harmonics:

ρp,ℓmX(r)=∫S2ρpX(X+rΩ)YℓmR(Ω) dΩ.\rho_{p,\ell m}^{X}(r)=\int_{S^2}\rho_p^{X}(\mathbf{X}+r\mathbf{\Omega})Y_{\ell m}^{\mathrm{R}}(\mathbf{\Omega})\,d\mathbf{\Omega}.

The reusable radial coefficients are

Rp,ℓmX(k)=∫0∞r2jℓ(kr)ρp,ℓmX(r) dr(2)R_{p,\ell m}^{X}(k)=\int_0^\infty r^2j_\ell(kr)\rho_{p,\ell m}^{X}(r)\,dr \tag*{(2)}

where jℓj_\ell is the spherical Bessel function of the first kind. These coefficients are real and are evaluated numerically for 0≤ℓ≤L0\leq\ell\leq L, −ℓ≤m≤ℓ-\ell\leq m\leq\ell, where L=Lmax⁡L=L_{\max}.

The Fourier transform of a degree-ℓ\ell spherical-harmonic component has the same angular harmonic, multiplied by 4π(−i)ℓ4\pi(-i)^\ell and its radial spherical-Bessel transform. Applying this identity to each component and retaining degrees through LL defines

Fp(L)(kω)=4π∑X∈C(p)e−ikω⋅XGpX(k,ω),(3)F_p^{(L)}(k\boldsymbol{\omega})=4\pi\sum_{X\in C(p)}e^{-i k\boldsymbol{\omega}\cdot\mathbf{X}}G_p^X(k,\boldsymbol{\omega}), \tag*{(3)}
GpX(k,ω)=∑ℓ=0L∑m=−ℓℓ(−i)ℓRp,ℓmX(k)YℓmR(ω).G_p^X(k,\boldsymbol{\omega})=\sum_{\ell=0}^{L}\sum_{m=-\ell}^{\ell}(-i)^\ell R_{p,\ell m}^{X}(k)Y_{\ell m}^{\mathrm{R}}(\boldsymbol{\omega}).

Only GpXG_p^X is represented by a finite angular expansion: the translation phase is neither expanded nor sampled. Localization reduces the degree needed for off-center structure.

Analytic angular contraction

Our goal is to evaluate the angular integral in (1) directly from the cached coefficients, leaving a one-dimensional momentum quadrature. Substituting the local representation (3) reduces this task to component-center pairs. For each pair (X,Y)(X,Y), write DXY=X−Y\mathbf{D}_{XY}=\mathbf{X}-\mathbf{Y} and DXY=∣DXY∣D_{XY}=\lvert\mathbf{D}_{XY}\rvert. The required contribution is

ApX,qY(k)=∫S2eikω⋅DXYGpX(k,ω)‾GqY(k,ω) dω.A_{pX,qY}(k)=\int_{S^2}e^{i k\boldsymbol{\omega}\cdot\mathbf{D}_{XY}}\overline{G_p^X(k,\boldsymbol{\omega})}G_q^Y(k,\boldsymbol{\omega})\,d\boldsymbol{\omega}.

Summing up these contributions and multiplying by (4π)2(4\pi)^2 gives the angular integral of Fp(L)Fq(L)F_p^{(L)}F_q^{(L)}.

For the derivation, we express the local amplitudes in complex spherical harmonics:

GpX(k,ω)=∑ℓ=0L∑m=−ℓℓup,ℓmX(k)Yℓm(ω),up,ℓmX(k)=(−i)ℓcp,ℓmX(k),G_p^X(k,\boldsymbol{\omega})=\sum_{\ell=0}^{L}\sum_{m=-\ell}^{\ell}u_{p,\ell m}^{X}(k)Y_{\ell m}(\boldsymbol{\omega}),\qquad u_{p,\ell m}^{X}(k)=(-i)^\ell c_{p,\ell m}^{X}(k),

where cp,ℓmXc_{p,\ell m}^{X} is the unitary conversion of the cached real coefficients Rp,ℓmXR_{p,\ell m}^{X}.

For distinct centers, we rotate the coordinates to align DXY\mathbf{D}_{XY} with the positive zz-axis. Let QQ denote this rotation and η=Qω\boldsymbol{\eta}=Q\boldsymbol{\omega} the rotated direction. Both amplitudes are expressed in this same frame using the Wigner rotation matrices:

Yℓs(QTη)=∑m=−ℓℓSms(ℓ)(Q)Yℓm(η).Y_{\ell s}(Q^{T}\boldsymbol{\eta})=\sum_{m=-\ell}^{\ell}S^{(\ell)}_{ms}(Q)Y_{\ell m}(\boldsymbol{\eta}).
u~p,ℓmX(k)=∑s=−ℓℓSms(ℓ)(Q)up,ℓsX(k).\widetilde{u}^{X}_{p,\ell m}(k)=\sum_{s=-\ell}^{\ell}S^{(\ell)}_{ms}(Q)u^{X}_{p,\ell s}(k).

In this frame, the phase becomes eizte^{izt}, with z=kDXYz=kD_{XY} and tt the cosine of the polar angle. Azimuthal integration retains equal magnetic orders, giving the translation blocks

Tℓℓ′m(z)=2πNℓaNℓ′a∫−11Pℓa(t)Pℓ′a(t)eizt dt,a=∣m∣,T^{m}_{\ell\ell'}(z)=2\pi N_{\ell a}N_{\ell'a}\int_{-1}^{1}P_{\ell}^{a}(t)P_{\ell'}^{a}(t)e^{izt}\,\mathrm{d}t,\qquad a=|m|,

where Nℓa=(2ℓ+1)(ℓ−a)!/[4π(ℓ+a)!]N_{\ell a}=\sqrt{(2\ell+1)(\ell-a)!/[4\pi(\ell+a)!]} is the spherical-harmonic normalization. For each mm, the degrees ℓ,ℓ′=∣m∣,…,L\ell,\ell'=|m|,\ldots,L form a block of order L−∣m∣+1L-|m|+1.

To evaluate these blocks, we precompute the normalized Legendre-product expansions

NℓaNℓ′aPℓa(t)Pℓ′a(t)=∑n=0ℓ+ℓ′bn,ℓℓ′aPn(t).N_{\ell a}N_{\ell'a}P_{\ell}^{a}(t)P_{\ell'}^{a}(t)=\sum_{n=0}^{\ell+\ell'}b^{a}_{n,\ell\ell'}P_n(t).

Combining this expansion with ∫−11Pn(t)eizt dt=2injn(z)\int_{-1}^{1}P_n(t)e^{izt}\,\mathrm{d}t=2i^n j_n(z) gives

Tℓℓ′m(z)=4π∑n=0ℓ+ℓ′bn,ℓℓ′∣m∣injn(z).T^{m}_{\ell\ell'}(z)=4\pi\sum_{n=0}^{\ell+\ell'}b^{|m|}_{n,\ell\ell'}i^n j_n(z).

The coefficients depend only on angular indices. The selection rules restrict their support to ∣ℓ−ℓ′∣≤n≤ℓ+ℓ′|\ell-\ell'|\le n\le\ell+\ell' with ℓ+ℓ′+n\ell+\ell'+n even. Thus Bessel orders through 2L2L suffice for every value of kDXYkD_{XY}.

Collect the rotated coefficients of each magnetic (u~p,ℓmX(k))ℓ=∣m∣L(\widetilde{u}^{X}_{p,\ell m}(k))_{\ell=|m|}^{L} order into u~p,mX(k)\widetilde{\mathbf{u}}^{X}_{p,m}(k). The component contribution is then

ApX,qY(k)=∑m=−LLu~p,mX†(k)Tm(kDXY)u~q,mY(k).\mathcal{A}_{pX,qY}(k)=\sum_{m=-L}^{L}\widetilde{\mathbf{u}}^{X\dagger}_{p,m}(k)T^{m}(kD_{XY})\widetilde{\mathbf{u}}^{Y}_{q,m}(k).

For coincident centers, the original frame gives Tm(0)=IT^{m}(0)=I, the identity, and the contraction reduces to an ordinary coefficient inner product.

Finally, using quadrature nodes and weights {ki,vi}i=1Nk\{k_i,v_i\}_{i=1}^{N_k} on [0,Kmax⁡][0,K_{\max}], we obtain

C^pq=8∑i=1Nkvi∑X∈C(p)∑Y∈C(q)Re⁡ApX,qY(ki).\widehat{C}_{pq}=8\sum_{i=1}^{N_k}v_i\sum_{X\in\mathcal{C}(p)}\sum_{Y\in\mathcal{C}(q)}\operatorname{Re}\mathcal{A}_{pX,qY}(k_i).

The factor 8=(4π)2/(2π2)8=(4\pi)^2/(2\pi^2) combines the Fourier and Coulomb prefactors. The angular integration is exact for the retained harmonic representation.

Implementation and storage

To generate the coefficients in (2), the angular projections are evaluated using 48 Gauss–Legendre nodes in u=cos⁡θu=\cos\theta and 96 uniformly spaced azimuthal nodes. We note that Lebedev quadrature is more efficient, as we observed in [9], but we choose to use Gauss–Legendre for simplicity in terms of implementation; this part of the computation entails a relatively insignificant computational effort anyway.

The radial integrals use 24 panels with 16 Gauss–Legendre nodes per panel. The grid is shared within each center pairs for all orbitals:

Rp,ℓmX(ki)≈∑a=124×16wa(r)ra2jℓ(kira)⏟reused across orbital pairsρp,ℓmX(ra).R^{X}_{p,\ell m}(k_i) \approx\sum_{a=1}^{24\times16} \underbrace{w_a^{(r)}r_a^2j_\ell(k_i r_a)}_{\text{reused across orbital pairs}}\rho^{X}_{p,\ell m}(r_a).

We also enforced a panel boundary at r=dABr=d_{AB} to ensure each side is approximated separately, as discussed in § 3. We determine the choices of numerical parameters empirically and note that they can be easily adjusted to fit the application needs. For each center pair, we choose the radial cutoff to be rmax⁡=dAB+τ/βmin⁡r_{\max}=d_{AB}+\tau/\beta_{\min}, with βmin⁡\beta_{\min} being the smallest sum ζμ+ζν\zeta_\mu+\zeta_\nu among its orbital pairs. For p=(μ,ν)p=(\mu,\nu), the STO decay and ∣jℓ(kr)∣≤1\lvert j_\ell(kr)\rvert\le1 give the following bound on the tail:

∣∫rmax⁡∞r2jℓ(kr)ρp,ℓmX(r) dr∣≤Cp(1+τ)nμ+nνe−τ.\left|\int_{r_{\max}}^\infty r^2j_\ell(kr)\rho^{X}_{p,\ell m}(r)\,\mathrm{d}r\right| \le C_p(1+\tau)^{n_\mu+n_\nu}e^{-\tau}.

where the constant CpC_p depends on the orbital parameters and geometry, but is independent of kk and τ\tau. We choose τ=30\tau=30 for the computation demonstrated in § 6.

The persistent cache stores Rp,ℓmX(ki)R^{X}_{p,\ell m}(k_i) as real double-precision values in blocks indexed by component, together with pair and center maps. Every block retains all H=(L+1)2H=(L+1)^2 channels at every momentum node; complex coefficients and rotations are formed for contraction. With Ncomp=∑p∣C(p)∣N_{\mathrm{comp}}=\sum_p|\mathcal{C}(p)| and M≤Ncomp≤2MM\le N_{\mathrm{comp}}\le2M, the coefficient storage in bytes is

Spair=8NcompNk(L+1)2=O(N2Nk(L+1)2).(4)S_{\mathrm{pair}}=8N_{\mathrm{comp}}N_k(L+1)^2=O(N^2N_k(L+1)^2). \tag*{(4)}

Thus the cache is quadratic in the number of basis functions at fixed NkN_k and LL. Keeping translations outside the cache avoids a factor set by Kmax⁡RsysK_{\max}R_{\mathrm{sys}} in the stored angular degree. The dense layout also avoids irregular per-pair channel lists. Device residency is nonetheless limited by this quadratic cache and the additional contraction workspaces.

Contractions are grouped by component-center pair so that rotations and translation matrices are reused, with spherical-Bessel values shared at repeated distances. The inner exact angular contraction exposes batched matrix operations rather than repeated angular quadratures. For a single component-pair contribution at one momentum node, the direct bilinear form involves

∑m=−LL(L−∣m∣+1)2=(L+1)(2L2+4L+3)3.\sum_{m=-L}^{L}(L-\lvert m\rvert+1)^2=\frac{(L+1)(2L^2+4L+3)}{3}.

matrix entries. Thus, once rotations and coupling matrices are available, this gives a direct evaluation cost of O(Nk(L+1)3)\mathcal{O}(N_k(L+1)^3) per component-pair contribution, with further reuse available across a batch. Each ERI accumulates at most four component contributions.

Pair preprocessing is amortized over the integrals that share each pair. Explicitly forming all unique ERIs produces M(M+1)/2M(M+1)/2 output values, which scales as O(N4)\mathcal{O}(N^4).

Results

Table 1 reports the four-center test cases from [9]. The reference values are obtained from the FORTRAN code described in [10]. The four orbitals in each entry are associated, in order, with centers c1c_1, c2c_2, c3c_3, and c4c_4 given in the caption. The comparison uses the orbital labeling of the cited test cases.

Integral (Φ1Φ2∣Φ3Φ4)(\Phi_1\Phi_2 \mid\Phi_3\Phi_4)ResultRef ResultAbs. Error
(1s 1s∣1s 1s)(1s\ 1s \mid1s\ 1s)0.31118378400.3111837840<1.0×10−10< 1.0 \times10^{-10}
(1s 1s∣2s 1s)(1s\ 1s \mid2s\ 1s)0.26229857720.2622985772<1.0×10−10< 1.0 \times10^{-10}
(2s 2s∣2s 2s)(2s\ 2s \mid2s\ 2s)0.28490439590.28490439601.0×10−101.0 \times10^{-10}
(1s 1s∣2s 2p−1)(1s\ 1s \mid2s\ 2p_{-1})−0.1226144752-0.1226144752−0.1226144752-0.1226144752<1.0×10−10< 1.0 \times10^{-10}
(1s 2s∣2s 2p−1)(1s\ 2s \mid2s\ 2p_{-1})−0.1006677390-0.1006677390−0.1006677390-0.1006677390<1.0×10−10< 1.0 \times10^{-10}
(2s 2p−1∣1s 1s)(2s\ 2p_{-1} \mid1s\ 1s)−0.0038018631-0.0038018631−0.0038018631-0.0038018631<1.0×10−10< 1.0 \times10^{-10}
(3d−1 3d−1∣3d−1 3d−1)(3d_{-1}\ 3d_{-1} \mid3d_{-1}\ 3d_{-1})0.17401228790.17401228961.7×10−91.7 \times10^{-9}
(3d+2 3d+1∣3d+2 3d+1)(3d_{+2}\ 3d_{+1} \mid3d_{+2}\ 3d_{+1})0.00269661720.0026966172<1.0×10−10< 1.0 \times10^{-10}
(3d+1 3d+1∣3d+2 3d+2)(3d_{+1}\ 3d_{+1} \mid3d_{+2}\ 3d_{+2})0.15312601480.15312601631.5×10−91.5 \times10^{-9}
(3d+1 3d+1∣3d+1 3d+1)(3d_{+1}\ 3d_{+1} \mid3d_{+1}\ 3d_{+1})0.17401228790.17401228961.7×10−91.7 \times10^{-9}

Table 1. Four-center STO electron-repulsion integrals for centers c1=(0,0,0)c_1 = (0, 0, 0), c2=(1,0,0)c_2 = (1, 0, 0), c3=(2,0,0)c_3 = (2, 0, 0), and c4=(3,0,0)c_4 = (3, 0, 0) with Lmax⁡=22L_{\max} = 22, Kmax⁡=40K_{\max} = 40, and a momentum quadrature of 24 panels with 16 Gauss–Legendre nodes per panel.

The examples include ss-orbital products, mixed ss–pp products, and several dd-orbital combinations. The reported absolute discrepancies are at or below 10−1010^{-10} for seven entries and between 1.5×10−91.5 \times10^{-9} and 1.7×10−91.7 \times10^{-9} for the remaining three. The largest reported discrepancy is 1.7×10−91.7 \times10^{-9}, occurring in two of the dd-orbital examples. The signed mixed-orbital values and the different magnetic orders provide further evidence of the validity of the angular factors in addition to the radial dependence.

Figure 6 shows the end-to-end runtime for various tests, which is the wall clock time to compute all ERIs of a molecule. This includes pair cache generation, IO time, memory transfer, and additional computational overhead. The time taken for the cache build depends on the geometry of the molecule and the number of basis functions. If the total memory of the workspace used by the contraction stage exceeds the GPU memory, then the workload must be split and streamed in smaller parts, resulting in more time spent on memory transfer. Overall, the diagram shows that we sustain about 40,000 ERIs per second end-to-end.

Runtime breakdown per workload

Figure 1. Total runtime breakdown per workload using the parameters of Table 1.

A contraction benchmark reaches approximately 4.2×1054.2 \times10^{5} unique ERIs per second on an NVIDIA V100. This is a contraction-stage rate, not an end-to-end timing including cache generation. The throughput measures the reuse of precomputed pair coefficients during ERI assembly. Together with the storage formula (4), it characterizes the contraction rate and coefficient storage of the implementation.

Conclusions

The localized Fourier formulation for evaluating four-center Coulomb repulsion integrals over STOs that we have presented separates pair preprocessing from ERI contraction. A two-center Becke partition produces locally centered components, and their radial spherical-Bessel coefficients provide reusable data for all integrals involving the corresponding orbital pairs.

We take advantage of the attractive computational features of our GPU environment, and the results are promising. The small absolute discrepancies and the large number of unique ERIs that are computed per second suggest that further study of the accuracy–storage–throughput balance through systematic refinement of the local representation and momentum quadrature may be beneficial.

References

  1. [1]A. Szabo, N. S. Ostlund, Modern Quantum Chemistry: Introduction to Advanced Electronic Structure Theory, 1st Edition, Dover Publications, Mineola, New York, 1996.
  2. [2]D. B. Cook, Handbook of Computational Quantum Chemistry, Dover Books on Chemistry, Dover Publications, Mineola, N.Y, 2005.
  3. [3]J. C. Slater, The Self Consistent Field and the Structure of Atoms, Physical Review 32 (3) (1928) 339–348. doi:10.1103/PhysRev.32.339.DOI
  4. [4]H. P. Trivedi, E. O. Steinborn, Fourier transform of a two-center product of exponential-type orbitals: Application to one- and two-electron multicenter integrals, Physical Review A 27 (2) (1983) 670–679. doi:10.1103/PhysRevA.27.670.DOI
  5. [5]J. Grotendorst, E. O. Steinborn, Numerical evaluation of molecular one- and two-electron multicenter integrals with exponential-type orbitals via the fourier-transform method, Physical Review A 38 (8) (1988) 3857–3876. doi:10.1103/PhysRevA.38.3857.DOI
  6. [6]T. A. Niehaus, R. López, J. F. Rico, Efficient evaluation of the fourier transform over products of slater-type orbitals on different centers, Journal of Physics A: Mathematical and Theoretical 41 (48) (2008) 485205.
  7. [7]T. Vuković, S. Dmitrović, Extremely compact formulas for the fourier transform of a product of two-centre slater-type orbitals, Journal of Physics A: Mathematical and Theoretical 43 (45) (2010) 455208.DOI
  8. [8]A. D. Becke, A multicenter numerical integration scheme for polyatomic molecules, The Journal of Chemical Physics 88 (4) (1988) 2547–2553.DOI
  9. [9]A. Kaur, W. Chen, G. K. Luhana, L. Silberman, M. Ewert, A. Bünger, U. Segal, C. Greif, Fast and accurate gpu-accelerated computation of two-electron four-center coulomb repulsion integrals, Journal of Computational Chemistry 47 (20) (2026) e70472. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/jcc.70472, doi:https://doi.org/10.1002/jcc.70472. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/jcc.70472
  10. [10]L. Berlu, H. Safouhi, Multicentre two-electron Coulomb and exchange integrals over Slater functions evaluated using a generalized algorithm based on nonlinear transformations, Journal of Physics A: Mathematical and General 37 (10) (2004) 3393–3410. doi:10.1088/0305-4470/37/10/006.DOI

Paper details

vlnc:W_byzn9r6g
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents