betaFeedback

WorksPhysics and Astronomy

Accelerating real-time hybrid-functional calculations with learned ACE and symmetry-based FFT reuse

Abstract

Real-time hybrid-functional calculations require repeated reconstruction of nonlocal exchange, whose cost grows rapidly with Brillouin-zone sampling. We develop two methods to improve the efficiency of time evolution and spatial processing. The first learns the time evolution of adaptively compressed exchange (ACE) factors and predicts their updates, reducing the frequency of explicit reconstruction. The second uses crystal symmetry to reuse exchange fast Fourier transforms (FFTs) among equivalent real-space points. This symmetry-based FFT reuse (SBFR) reduces the number of FFTs while retaining the full k\bm k-point sum. The methods can be used with a variety of hybrid functionals, and we use HSE06 in this work. Relative to an ACE reference using full-row mesh FFTs, learned prediction provides more than fivefold acceleration, SBFR more than fourteenfold, and their combination more than sixty-threefold. The combined method shows efficient parallel scaling. After learning, propagation takes less than 0.5 s per step. The reduced computational cost allows us to study nonlinear and nonequilibrium electron dynamics with hybrid functionals and to examine the role of excitonic effects in semiconductors and insulators.

Introduction

Real-time time-dependent density-functional theory (RT-TDDFT) describes electronic response to external fields [1, 2, 3, 4]. Nonlinear and nonequilibrium phenomena in solids [5] include laser-induced optical breakdown [6], coherent phonon generation [7], pump–probe response [8], the dynamical Franz–Keldysh effect [9], and high-harmonic generation [10, 11]. Their accurate description requires reliable band gaps and electron–hole interactions, which semilocal approximations often inadequately capture [12]. Hybrid functionals such as HSE06 improve this description through nonlocal exchange [13, 14, 15] and have been applied to nonlinear excitation, optical spectra, and excitonic response [16, 17, 18, 19].

Hybrid functionals such as HSE and PBE0 are substantially more expensive than semilocal functionals such as PBE because of nonlocal exchange [20]. This cost is a major obstacle in real-time calculations. Nonlocal exchange couples each target orbital to occupied orbitals throughout the Brillouin zone. As these orbitals evolve, exchange must be updated over thousands of time steps [21], and finer kk-point sampling further increases the work. This cost makes systematic calculations over laser intensity, frequency, and polarization difficult, even for a small unit cell. Reducing the exchange cost is therefore essential for studying nonlinear dynamics routinely with hybrid functionals.

Adaptively compressed exchange (ACE) reduces repeated exchange-operator application costs [20]; combined with parallel-transport integration, it enables efficient hybrid RT-TDDFT, including finite-temperature simulations [21, 22]. Density fitting, orbital locality, and low-rank representations further reduce exchange work [23, 24, 25, 26], with Fourier convolution [27] and tensor hypercontraction [28] addressing kk-point sampling. Despite these advances, explicit exchange evaluation during ACE reconstruction remains costly. In this work, we make hybrid-functional propagation more practical by reducing both the reconstruction frequency in time and the FFT work in space.

The first method accelerates time evolution by learning the changes in ACE factors and reducing the reconstruction frequency. Previous learning approaches model molecular time-dependent Hartree–Fock potentials [29] or electron-density propagators [30]. We fit the coefficients to gauge-aligned factor increments during propagation and use them to predict the factors between explicit corrections. The orbitals, density, and semilocal potentials are updated at every time step.

The second method improves spatial processing by reducing the number of fast Fourier transforms (FFTs) in each reconstruction. We achieve this through symmetry-based FFT reuse (SBFR). Representative-orbital transformations [31], symmetry-reduced exchange-matrix construction with localized resolution of identity [32], and symmetry-adapted density fitting for many-body calculations [33] provide related precedents. We reuse mesh FFTs of symmetry-equivalent density-matrix rows directly on the real-space grid, without fitting orbital products or reducing the exchange sum over k′k'. Spatial-pair distribution parallelizes the remaining FFTs beyond the number of independent rows.

We implement both methods in SALMON [34] and demonstrate them with HSE06. Section II presents the methods, Sec. III defines the validation measures, and Sec. IV reports silicon benchmarks. Section V summarizes the findings and outlook.

Theory and methods

Screened exchange in real time

Atomic units are used unless stated otherwise. HSE replaces a fraction of short-range Perdew–Burke–Ernzerhof (PBE) exchange by Hartree–Fock exchange, retaining PBE correlation and long-range exchange [35]. Its exchange-correlation energy is

ExcHSE=aExHF,SR(ωSR)+(1−a)ExPBE,SR(ωSR)+ExPBE,LR(ωSR)+EcPBE,(1)E_{\mathrm{xc}}^{\mathrm{HSE}}=aE_{\mathrm{x}}^{\mathrm{HF,SR}}(\omega_{\mathrm{SR}})+(1-a)E_{\mathrm{x}}^{\mathrm{PBE,SR}}(\omega_{\mathrm{SR}})+E_{\mathrm{x}}^{\mathrm{PBE,LR}}(\omega_{\mathrm{SR}})+E_{\mathrm{c}}^{\mathrm{PBE}}, \tag*{(1)}

where HSE06 uses a=1/4a=1/4 and ωSR=0.11\omega_{\mathrm{SR}}=0.11 [14]. For each spin channel, define the one-particle density matrix

γσ(r,r′;t)=∑nkwkfnkσψnkσ(r,t)ψnkσ∗(r′,t).(2)\gamma_{\sigma}(\mathbf{r},\mathbf{r}';t)=\sum_{nk}w_kf_{nk\sigma}\psi_{nk\sigma}(\mathbf{r},t)\psi_{nk\sigma}^{*}(\mathbf{r}',t). \tag*{(2)}

Here fnkσf_{nk\sigma} is the occupation per spin, and wkw_k is the normalized Brillouin-zone weight. The screened exchange action is

[Vx,σSRψnkσ](r,t)=−∑R∫Ωd3r′ K0(r−r′−R)γσ(r,r′+R;t)ψnkσ(r′+R,t).(3)[V_{\mathrm{x},\sigma}^{\mathrm{SR}}\psi_{nk\sigma}](\mathbf{r},t)=-\sum_{\mathbf{R}}\int_{\Omega}d^{3}r'\,K_{0}(\mathbf{r}-\mathbf{r}'-\mathbf{R})\gamma_{\sigma}(\mathbf{r},\mathbf{r}'+\mathbf{R};t)\psi_{nk\sigma}(\mathbf{r}'+\mathbf{R},t). \tag*{(3)}

Here Ω\Omega denotes the real-space simulation box and, when used as a scalar, its volume. The vectors R\mathbf{R} are lattice translations and K0(r)=erfc⁡(ωSR∣r∣)/∣r∣K_{0}(r)=\operatorname{erfc}(\omega_{\mathrm{SR}}|r|)/|r|. Translated orbitals carry their Bloch phases. For an occupied orbital ψn′k′σ\psi_{n'k'\sigma} and target orbital ψnkσ\psi_{nk\sigma}, the pair density is ψn′k′σ∗ψnkσ\psi_{n'k'\sigma}^{*}\psi_{nk\sigma}. Explicit exchange evaluates the corresponding convolution and multiplies its result by ψn′k′σ\psi_{n'k'\sigma}. The source sum is weighted by the occupations.

For occupied-space reconstruction, direct exchange requires convolutions over occupied-band pairs and source and target wave vectors. We use a uniform real-space grid with NrN_r points; the resulting costs are summarized in Table I.

Real-time propagation follows the self-consistent time-dependent Kohn–Sham equation [1, 36, 37],

i∂tψnk=[hsl[n(t),A(t)]+aVxSR[γ(t)]]ψnk,(4)i\partial_t\psi_{nk}=[h_{\mathrm{sl}}[n(t),\mathbf{A}(t)]+aV_{\mathrm{x}}^{\mathrm{SR}}[\gamma(t)]]\psi_{nk}, \tag*{(4)}

where hslh_{\mathrm{sl}} includes kinetic, ionic, Hartree, and the appropriate semilocal HSE terms.

ACE construction and application

To reduce the cost of repeated exchange applications during propagation, we construct ACE separately at each target kk point from its NoccN_{\mathrm{occ}} occupied orbitals. Wave-vector and spin labels are suppressed in this subsection. The matrix Φ\Phi collects these orbitals, and W=VxSRΦW=V_{\mathrm{x}}^{\mathrm{SR}}\Phi contains their explicit exchange images. The overlap matrix is defined by spatial integration,

Mij=∫Ωϕi∗(r)Wj(r) d3r.M_{ij}=\int_{\Omega}\phi_i^{*}(\mathbf{r})W_j(\mathbf{r})\,d^{3}r.

In operator notation this gives

M=Φ†W,−M=LL†,(5)M=\Phi^{\dagger}W,\qquad-M=LL^{\dagger}, \tag*{(5)}

with LL lower triangular. For a nonsingular negative-definite overlap matrix, define

Ξ=WL−†,VxACE=−ΞΞ†.(6)\Xi=WL^{-\dagger},\qquad V_{\mathrm{x}}^{\mathrm{ACE}}=-\Xi\Xi^{\dagger}. \tag*{(6)}

Here L−†=(L†)−1L^{-\dagger}=(L^{\dagger})^{-1} is evaluated by triangular solves. Since M−1=−L−†L−1M^{-1}=-L^{-\dagger}L^{-1},

VxACE=WM−1W†,VxACEΦ=W.(7)V_{\mathrm{x}}^{\mathrm{ACE}}=WM^{-1}W^{\dagger},\qquad V_{\mathrm{x}}^{\mathrm{ACE}}\Phi=W. \tag*{(7)}

Thus ACE reproduces explicit exchange on the construction subspace.

For any target ϕ\phi, the ACE action is

VxACEϕ=−∑μ=1Noccξμ(r)∫Ωξμ∗(r′)ϕ(r′) d3r′,(8)V_{\mathrm{x}}^{\mathrm{ACE}}\phi=-\sum_{\mu=1}^{N_{\mathrm{occ}}}\xi_{\mu}(\mathbf{r})\int_{\Omega}\xi_{\mu}^{*}(\mathbf{r}')\phi(\mathbf{r}')\,d^{3}r', \tag*{(8)}

where ξμ(r)\xi_{\mu}(\mathbf{r}) is the μ\muth column of Ξ\Xi. ACE replaces pairwise FFTs with two matrix products over orbitals and linear combinations. The factors are reused between explicit reconstructions.

ACE agrees closely with direct screened exchange in the laser-driven Si benchmark (Appendix E).

Temporal acceleration with learned ACE

ACE makes exchange application inexpensive, but reconstructing its factors at every time step remains costly. To reduce this work, we predict the factors from their recent evolution and retain periodic explicit corrections. The density and semilocal potentials are updated at every step. Related history-based methods reduce self-consistency work in polarizable molecular dynamics [38] and expensive force evaluations in ACE-based Born–Oppenheimer molecular dynamics [39].

Unitary column rotations leave the ACE operator unchanged but can introduce artificial variations in its factor history. Before fitting or prediction, we align successive explicit factors by a complex extension of orthogonal Procrustes alignment [40]; see Appendix C.

The aligned explicit factor at endpoint mm is Ξm\Xi_m, with endpoints Ξm−KΔt\Xi_{m-K\Delta t}. Within an interval, s=ℓ/Ks=\ell/K for ℓ=1,…,K−1\ell=1,\ldots,K-1. To learn the factor increments, we retain three explicit factors and form

D1=Ξm−Ξm−1,D2=Ξm−1−Ξm−2.(9)D_1=\Xi_m-\Xi_{m-1},\qquad D_2=\Xi_{m-1}-\Xi_{m-2}. \tag*{(9)}

For factor functions, we define the real-valued inner product

(A,B)=ℜ∑μ∫ΩAμ∗(r)Bμ(r) d3r.(A,B)=\Re\sum_{\mu}\int_{\Omega}A_{\mu}^{*}(\mathbf{r})B_{\mu}(\mathbf{r})\,d^{3}r.

Here μ\mu labels factor columns and ∥A∥F2=(A,A)\lVert A\rVert_{F}^{2}=(A,A). For a later explicit endpoint, the target increment is Y=Y= Ξm+1−Ξm\Xi_{m+1}-\Xi_m. The two real coefficients c=(c1,c2)\mathbf{c}=(c_1,c_2) are determined by cumulative regularized least squares,

min⁡c1,c2∑m∥Ym−c1D1,m−c2D2,m∥F2+λ(c12+c22).(10)\min_{c_1,c_2}\sum_m\left\lVert Y_m-c_1D_{1,m}-c_2D_{2,m}\right\rVert_F^2+\lambda(c_1^2+c_2^2). \tag*{(10)}

To stabilize the fit for nearly dependent increments, we set λ=10−8∑m(∥D1,m∥F2+∥D2,m∥F2)\lambda=10^{-8}\sum_m(\lVert D_{1,m}\rVert_F^2+\lVert D_{2,m}\rVert_F^2). The fit determines two real coefficients for each target-k\mathbf{k} history.

The learned predictor adds a correction to linear extrapolation, giving

Ξ^m,ℓ=Ξm+[s+s2(c1−1)]D1+s2c2D2.(11)\widehat{\Xi}_{m,\ell}=\Xi_m+\left[s+s^2(c_1-1)\right]D_1 +s^2c_2D_2. \tag*{(11)}

It reproduces the learned endpoint increment at s=1s=1 and reduces to linear extrapolation when c=(1,0)\mathbf{c}=(1,0). At ℓ=K\ell=K, explicit exchange supplies the next factor and updates the fit. For either predictor, the exchange operator is constructed as −Ξ^Ξ^†-\widehat{\Xi}\widehat{\Xi}^{\dagger}, preserving its algebraic sign and Hermiticity.

During warmup, ACE is reconstructed at every time step, while training increments are accumulated at intervals of KK steps. We set Nhi=3N_{\mathrm{hi}}=3 history factors and Nteach=8N_{\mathrm{teach}}=8 training increments. These determine the warmup length automatically, giving 80 steps at K=8K=8 (Appendix C).

Spatial acceleration with SBFR

Crystal symmetry makes many real-space contributions to exchange equivalent. To avoid repeating the corresponding FFTs, we first perform the occupied-band sum and express exchange through the density matrix resolved by k′\mathbf{k}'. For a mesh with reference point k0\mathbf{k}_0, we use the phase convention

vnk′(r)=e−ik0⋅rψnk′(r)=ei(k′−k0)⋅runk′(r),(12)v_{n\mathbf{k}'}(\mathbf{r})=e^{-i\mathbf{k}_0\cdot\mathbf{r}}\psi_{n\mathbf{k}'}(\mathbf{r}) =e^{i(\mathbf{k}'-\mathbf{k}_0)\cdot\mathbf{r}}u_{n\mathbf{k}'}(\mathbf{r}), \tag*{(12)}
γ‾k′(r,r′)=∑n∈occvnk′(r)vnk′∗(r′).(13)\overline{\gamma}_{\mathbf{k}'}(\mathbf{r},\mathbf{r}')=\sum_{n\in\mathrm{occ}}v_{n\mathbf{k}'}(\mathbf{r})v_{n\mathbf{k}'}^*(\mathbf{r}'). \tag*{(13)}

Here ψnk′=eik′⋅runk′\psi_{n\mathbf{k}'}=e^{i\mathbf{k}'\cdot\mathbf{r}}u_{n\mathbf{k}'} is a Bloch orbital and k0\mathbf{k}_0 is the first point of the full uniform mesh. Unit occupations per spin and equal weights 1/Nk1/N_k are assumed; spin and time labels are suppressed. For each coordinate pair, the mesh transforms are

γ^R(r,r′)=∑k′∈Mke−i(k′−k0)⋅Rγ‾k′(r,r′),(14)\widehat{\gamma}_{\mathbf{R}}(\mathbf{r},\mathbf{r}')= \sum_{\mathbf{k}'\in\mathcal{M}_k}e^{-i(\mathbf{k}'-\mathbf{k}_0)\cdot\mathbf{R}}\overline{\gamma}_{\mathbf{k}'}(\mathbf{r},\mathbf{r}'), \tag*{(14)}
Bk(r,r′)=∑R∈Bei(k−k0)⋅RKB(r−r′−R)γ^R(r,r′).(15)B_{\mathbf{k}}(\mathbf{r},\mathbf{r}')= \sum_{\mathbf{R}\in\mathcal{B}}e^{i(\mathbf{k}-\mathbf{k}_0)\cdot\mathbf{R}} \mathcal{K}_{\mathcal{B}}(\mathbf{r}-\mathbf{r}'-\mathbf{R})\widehat{\gamma}_{\mathbf{R}}(\mathbf{r},\mathbf{r}'). \tag*{(15)}

where Mk\mathcal{M}_k is the complete NkN_k-point mesh, B\mathcal{B} is its dual set of lattice translations in the Born–von Karman supercell, and KB\mathcal{K}_{\mathcal{B}} is the screened kernel periodic on that supercell. These transforms act along the k′\mathbf{k}' and R\mathbf{R} indices, with both spatial coordinates held fixed. The exchange action is

[VkSRunk](r)=−e−i(k−k0)⋅rNk∫ΩBk(r,r′)vnk(r′) d3r′.(16)[V_{\mathbf{k}}^{\mathrm{SR}}u_{n\mathbf{k}}](\mathbf{r})= -\frac{e^{-i(\mathbf{k}-\mathbf{k}_0)\cdot\mathbf{r}}}{N_k} \int_{\Omega}B_{\mathbf{k}}(\mathbf{r},\mathbf{r}')v_{n\mathbf{k}}(\mathbf{r}')\,\mathrm{d}^3r'. \tag*{(16)}

Appendix B derives these expressions and their discrete normalization.

The real-space representation is particularly useful for this procedure because symmetry-related grid points are connected by coordinate permutations and phase factors. We can therefore reuse equivalent density-matrix rows and distribute the remaining FFTs over spatial pairs independently of orbital ownership.

Crystal symmetry reduces orbital propagation to representative k\mathbf{k} points. At each k′\mathbf{k}', we copy the orbitals from its corresponding representative point by permuting the real-space grid according to the symmetry operation and multiplying by the associated Bloch phase factors. We form γ‾k′\overline{\gamma}_{\mathbf{k}'} from these orbitals, averaging the projectors when several symmetry operations reach the same k′\mathbf{k}'. SBFR additionally reuses FFT results among symmetry-equivalent real-space rows: Eqs. (14) and (15) are evaluated only at inequivalent first coordinates r\mathbf{r}, while r′\mathbf{r}' spans the full cell. The remaining rows of BkB_{\mathbf{k}} are restored by symmetry for contraction in Eq. (16). Thus representative-point reduction decreases the number of independently propagated orbitals, whereas SBFR decreases the number of mesh FFTs; each FFT and the exchange sum retain all NkN_k points.

Symmetry reconstruction transforms the wave vector and both spatial coordinates together, including the Bloch phases associated with cell wrapping (Appendix B). Retained operations must preserve the occupied projector, discrete kernel, and external field. For the Si tests, we use 32 space-group operations preserving the zz-directed field, including fractional translations.

Choosing rs\mathbf{r}_s from the representative first-coordinate points reduces the number of coordinate pairs requiring mesh FFTs from

Nr2⟶NaNr,Nk′=Nk in both routes,(17)N_r^2\longrightarrow N_aN_r,\qquad N_k'=N_k\ \text{in both routes}, \tag*{(17)}

where NaN_a is the number of symmetry-inequivalent real-space grid points. For the 12312^3 grid used here, Nr=1728N_r=1728 and Na=75N_a=75, reducing the number of pairs requiring FFTs by a factor of 23. We distribute the NaNrN_aN_r pairs over MPI ranks in batches. Each task performs a full NkN_k-point transform, symmetry restoration, and partial target contraction, avoiding storage of the full two-coordinate density matrix. The SBFR Off reference uses the same representative orbitals and mesh transforms but evaluates all first-coordinate rows. Timings include projector generation, FFTs, and target contraction; detailed costs are given in Appendix D.

Summary of computational costs

tbl:I summarizes the computational work. Here NoccN_{\mathrm{occ}} is the occupied-orbital count per kk point, NrN_{r} the grid size, NaN_{a} the number of symmetry-inequivalent grid points, and PkP_{k} the MPI rank count used for kk-point and spatial-pair distributions. The common costs are Capply=O(NkNocc2Nr)C_{\mathrm{apply}}=O(N_{k}N_{\mathrm{occ}}^{2}N_{r}) for ACE application and Cpredict=O(NkNoccNr)C_{\mathrm{predict}}=O(N_{k}N_{\mathrm{occ}}N_{r}) for prediction at fixed history length. The reconstruction costs CbuildglobalC_{\mathrm{build}}^{\mathrm{global}} and CbuildsymC_{\mathrm{build}}^{\mathrm{sym}} refer to conventional exchange and SBFR, including factor construction. In the performance benchmarks, CbuildglobalC_{\mathrm{build}}^{\mathrm{global}} uses the full-row mesh-FFT reference. Symmetry reduces mesh-FFT work by Nr/NaN_{r}/N_{a}, while prediction reduces reconstruction frequency by KK. The per-rank estimate assumes balanced spatial-pair tasks with Pk≲NaNrP_{k}\lesssim N_{a}N_{r}. At fixed cell and Pk∝NkP_{k}\propto N_{k}, the mesh-FFT contribution per rank grows as O(log⁡Nk)O(\log N_{k}). Communication, padding, and the other reconstruction costs are discussed in Appendix D.

Validation and performance measures

We compare learned prediction with conventional ACE and SBFR On with SBFR Off, keeping all other calculation settings fixed.

The absorbed energy per cell is evaluated from the matter-current density as

ΔEabs(t)=−Ω∫0tEtot(t′)⋅J(t′) dt′.(18)\Delta E_{\mathrm{abs}}(t)=-\Omega\int_{0}^{t}E_{\mathrm{tot}}(t')\cdot J(t')\,\mathrm{d}t'. \tag*{(18)}

Here EtotE_{\mathrm{tot}} is the total electric field and JJ is the matter-current density output by SALMON.

The HHG spectrum is evaluated as

S(ω)=ω2∣∫0Tw(t)Jz(t)eiωt dt∣2,(19)S(\omega)=\omega^{2}\left|\int_{0}^{T}w(t)J_{z}(t)e^{i\omega t}\,\mathrm{d}t\right|^{2}, \tag*{(19)}

using a common window w(t)=sin⁡2(πt/T)w(t)=\sin^{2}(\pi t/T) over the trajectory duration TT.

We quantify the current error over the complete trajectory by

ϵJ=[∫0T∣Jz(t)−Jzref(t)∣2 dt∫0T∣Jzref(t)∣2 dt]1/2.(20)\epsilon_{J}=\left[\frac{\int_{0}^{T}\left|J_{z}(t)-J_{z}^{\mathrm{ref}}(t)\right|^{2}\,\mathrm{d}t}{\int_{0}^{T}\left|J_{z}^{\mathrm{ref}}(t)\right|^{2}\,\mathrm{d}t}\right]^{1/2}. \tag*{(20)}

The energy error is the relative difference of final absorbed energies.

For transverse impulse-response calculations [3], we obtain the conductivity and dielectric function from

σzz(ω)=1Iz∫0Twimp(t)Jz(t)eiωt dt,(21)\sigma_{zz}(\omega)=\frac{1}{I_{z}}\int_{0}^{T}w_{\mathrm{imp}}(t)J_{z}(t)e^{i\omega t}\,\mathrm{d}t, \tag*{(21)}
ϵzz(ω)=1+4πiωσzz(ω),(22)\epsilon_{zz}(\omega)=1+\frac{4\pi i}{\omega}\sigma_{zz}(\omega), \tag*{(22)}

where IzI_{z} is the applied electric-field impulse amplitude in the SALMON convention and wimp(t)=1−3(t/T)2+2(t/T)3w_{\mathrm{imp}}(t)=1-3(t/T)^{2}+2(t/T)^{3}. We evaluate the transform by Δt∑j\Delta t\sum_{j} using the recorded currents at tj=jΔtt_{j}=j\Delta t.

Performance is reported as propagation time per step, with learning and post-learning intervals separated. Step timings average the maximum duration across MPI ranks at each step and exclude checkpoint and timer-record output. Each configuration is timed once.

Results

We use an eight-atom conventional Si cell with 32 valence electrons. We set the cell side length to 10.26 a.u., the real-space grid to 12312^{3} points, and the ground-state orbital count to 32. We use a norm-conserving Si pseudopotential generated in-house with the Perdew–Zunger local-density parametrization [41], treating four valence electrons per atom. We set uniform kk meshes shifted by half a mesh spacing in each reciprocal direction. Real-time calculations propagate the 16 occupied orbitals using fourth-order Taylor propagation with predictor–corrector updates and Δt=0.16\Delta t=0.16 a.u. The screened kernel is untruncated.

Numerical assessment of predictive ACE

Reducing the reconstruction frequency is useful only if prediction preserves the laser-driven response. To determine the correction interval, we compare learned ACE with conventional ACE on a 434^3 kk mesh over a 15-fs trajectory. We set the laser photon energy to 1.55 eV and the peak intensity to 101210^{12} W cm−2^{-2}. We use a zz-polarized vector potential with a cos⁡2\cos^2 envelope of 10-fs full support. We vary the explicit correction interval over K=2K = 2, 4, 8, 12, and 16.

Figure 1 compares the induced current and absorbed energy; the conventional-ACE reference absorbs 0.38 eV per cell. At K=4K = 4 and K=8K = 8, prediction also reproduces the conventional-ACE high-harmonic spectrum through the ninth harmonic (Figure 2).

Figure 1. Laser-driven (a) matter-current density and (b) absorbed energy per conventional Si cell for HSE06 with Nk=43N_k = 4^3. Learned ACE with correction intervals K=2K = 2, 4, 8, 12, and 16 is compared with ACE without learning. The dotted line marks the pulse end.

Figure 2. HHG spectra from conventional ACE and learned ACE with correction intervals K=2K = 2, 4, 8, 12, and 16 for the laser parameters in Sec. IV A. The horizontal axis is the harmonic order, with ℏω0=1.55\hbar\omega_0 = 1.55 eV. All spectra use the same 15-fs window and are normalized by the maximum conventional-ACE intensity over the displayed range.

Figure 3 quantifies the current and absorbed-energy errors. At K=8K = 8, these errors are 3.1% and 0.90%, whereas they increase to 16% and 7.1% at K=16K = 16. The lower reconstruction cost at K=16K = 16 thus comes at the expense of response accuracy. We therefore set K=8K = 8 for the following calculations. At this interval, learning reduces the current error from 11% for linear extrapolation to 3.1% at essentially the same cost (Appendix § 1.12).

Figure 3. Relative errors of the current and absorbed energy versus explicit correction interval KK for the calculations in Figure 1. Current errors use the full-record L2L^2 norm. Absorbed-energy errors use the final field-current integral. The reference is ACE without temporal learning. Lines guide the eye.

SBFR performance

To separate the gains from temporal prediction and spatial FFT reuse, we compare the two methods individually and in combination.

We conduct the benchmarks on Fugaku. Each node has one Fujitsu A64FX CPU with 48 compute cores and 32 GiB of high-bandwidth HBM2 memory, providing a peak memory bandwidth of 1024 GB/s. Nodes are connected by Tofu Interconnect D in a six-dimensional mesh/torus topology [42].

Exchange tests confirm that SBFR preserves numerical precision across shifted and unshifted meshes, fractional translations, MPI layouts, and FFT array layouts. For timing, we compare conventional ACE, learned ACE, and learned ACE with SBFR at Nk=43N_k = 4^3, using full-row mesh FFTs for both SBFR Off cases. We assign the 12 representative kk points to 12 MPI ranks on three nodes and set 12 OpenMP threads per rank, giving one representative point per rank. We retain crystal symmetry and all 434^3 source k′k' points in every case. We apply a zz-directed impulse of 10−410^{-4} a.u. and propagate for 160 steps (0.62 fs). We average the step times over steps 1–80 during learning and 81–160 after learning, with 70 prediction steps and 10 explicit corrections in the latter interval.

Table 2 shows the two contributions to the acceleration. During learning, explicit exchange is evaluated at every step, and SBFR alone reduces the step time by a factor of 15. After learning, prediction reduces the reconstruction frequency and gives a 5.3-fold acceleration without SBFR. Combining the two methods gives a 63-fold acceleration relative to conventional ACE.

MethodSBFRDuring learning (s/step)SpeedupAfter learning (s/step)Speedup
Conventional ACEOff36.9091.0036.8891.00
Learned ACEOff36.6801.016.9785.29
Learned ACEOn2.53914.540.58263.39

Table 2. Fugaku performance for Nk=43N_k = 4^3. SBFR Off and SBFR On use full-row and SBFR mesh FFTs, respectively, with identical orbital symmetry reduction. Speedups are relative to conventional ACE.

SBFR reduces explicit correction steps from 55 to 3.8 s, while prediction steps remain near 0.13 s. It therefore accelerates the costly steps that temporal prediction cannot eliminate.

Both scaling tests use this combined method with the same impulse and timing intervals, propagating symmetry-inequivalent representative kk points while retaining the full mesh in the exchange sum. We set four MPI ranks per node and 12 OpenMP threads per rank for both tests.

Strong scaling

To examine whether a fixed problem can be solved efficiently with more MPI resources, we first study strong scaling. We set Nk=163N_k = 16^3 with 576 representative points and use 16, 36, 72, and 144 nodes, corresponding to 64, 144, 288, and 576 MPI ranks, respectively. Figure 4 shows that, from 16 to 144 nodes, the learning time decreases from 19 to 1.7 s/step, while the post-learning time decreases from 4.4 to 0.45 s/step, giving speedups of 11 and 9.8, respectively, for a ninefold increase in node count. Orbitals are distributed over representative kk points, whereas exchange FFTs are distributed over spatial-grid pairs (r,r′)(r,r'). Increasing the number of ranks reduces the number of spatial pairs held by each rank. The resulting smaller FFT working set may fit more effectively in cache and explain the slightly superlinear speedups.

Figure 4. Fugaku strong scaling for a conventional Si cell at Nk=163N_k = 16^3, using learned ACE and SBFR with spatial-pair distribution. Dashed lines indicate inverse-node-count references anchored to the two 16-node timings.

Weak scaling

We next examine whether the step time can be maintained as the kk mesh grows. For this weak-scaling test, we increase the MPI resources to keep one symmetry-inequivalent representative kk point per rank. For Nk=43N_k = 4^3, 838^3, 16316^3, and 32332^3, symmetry gives 12, 80, 576, and 4352 representative points on 3, 20, 144, and 1088 nodes, respectively.

Figure 5(a) compares the learning and post-learning step times with constant-time reference lines. The post-learning cost remains below 0.80 s/step over the full series. It decreases from 0.58 to 0.35 s/step between 434^3 and 838^3, then rises to 0.45 and 0.77 s/step at 16316^3 and 32332^3. The initial decrease may likewise reflect improved cache utilization, as discussed for strong scaling. The learning curve follows the same trend. From 838^3 to 32332^3, a 64-fold increase in the full kk mesh increases the post-learning time by a factor of 2.2.

Figure 5. Fugaku weak scaling and optical response for a conventional Si cell. (a) Mean time per step during learning and after learning versus the full mesh size NkN_k, with one representative kk point per MPI rank. Dashed lines indicate constant-time references at Nk=43N_k = 4^3. (b) Imaginary dielectric response for Nk=43N_k = 4^3, 838^3, 16316^3, and 32332^3 from 20-fs impulse-response calculations.

We can understand this increase by separating prediction steps from explicit corrections. Prediction steps cost 0.13–0.15 s across the series, whereas explicit corrections increase from 1.9 s at 838^3 to 5.1 s at 32332^3. Although the number of propagated representative points per rank is fixed, each pair FFT retains the growing full source mesh. Reconstruction every eight steps limits this contribution to the average cost (Appendix A).

To determine the sampling needed for the linear response, we compare the dielectric spectra in Figure 5(b). The structure of the coarse-mesh spectrum becomes smoother as the number of points increases. The Nk=163N_k = 16^3 and 32332^3 spectra are close to each other, indicating that kk-point convergence is nearly reached on this real-space grid.

Laser-driven response with kk-point sampling

To determine whether the same sampling also describes nonlinear excitation, we compare the laser-driven responses at Nk=43N_k = 4^3, 838^3, 16316^3, and 32332^3 in Figure 6, using the pulse of Sec. IV A and the same learning schedule. The Nk=323N_k = 32^3 calculation over 15 fs required 3.2×1033.2 \times10^3 s (about 54 min), including initialization, on 1088 Fugaku nodes. The pronounced post-pulse current oscillations of the coarsest mesh are suppressed with finer sampling. The final absorbed energies are 0.38, 0.33, 0.28, and 0.28 eV per cell, respectively; the change from 16316^3 to 32332^3 is only 0.085%. The current waveforms also nearly converge at Nk=163N_k = 16^3, with a full-record relative difference of 6.1% from 32332^3. Above the seventh harmonic, differences between the two finest meshes are still visible, although the qualitative spectral features are unchanged. Under these conditions, 16316^3 sampling describes the main response. Finer sampling is needed for quantitative comparison of the higher harmonics.

Figure 6. Laser-driven Si response for Nk=43N_k = 4^3, 838^3, 16316^3, and 32332^3: (a) matter-current density, (b) absorbed energy per cell, and (c) HHG spectra. The laser parameters are those of Sec. IV A. Dotted vertical lines mark the pulse end.

Conclusions

We have accelerated real-time hybrid-functional calculations in two directions: time evolution with learned ACE and spatial processing with SBFR. The two reductions can be combined because they act on the frequency and cost of exchange reconstruction, respectively.

For silicon, we found a 5.3-fold acceleration from learned prediction after learning and a 15-fold acceleration from SBFR during learning, using full-row mesh-FFT ACE as the reference. Combining the two methods gives a 63-fold acceleration after learning. For the tested laser pulse, K=8K = 8 balances reconstruction frequency and response accuracy. Spatial-pair parallelism enables strong scaling of the remaining FFT work. The post-learning cost reaches 0.45 s/step at Nk=163N_k = 16^3 on 144 Fugaku nodes. It remains below 0.80 s/step through Nk=323N_k = 32^3 with one representative point per rank. With these step times, we can perform repeated HSE06 calculations to examine kk-point convergence and the dependence of electron dynamics on laser parameters.

The methods also apply to PBE0 and other hybrid functionals. Their improved description of the band gap and electron–hole interaction is important for nonlinear electron dynamics in semiconductors and insulators. The reduced cost makes it practical to examine how excitonic effects appear under strong laser excitation. By combining the present methods with SALMON’s multiscale framework [34, 43], we can further treat electron dynamics and light propagation simultaneously and approach experimental irradiation conditions more accurately.

Acknowledgments

This work was supported by the MEXT Quantum Leap Flagship Program (MEXT Q-LEAP), Grant No. JPMXS0118067246; JSPS KAKENHI, Grant No. 24K23024; and the MEXT Next-Generation Computational Science Global Reach & Innovation Program (GRIP), Japan Grant Number JPMXP1920260201. This work used computational resources of the supercomputer Fugaku provided by the RIKEN Center for Computational Science (Project ID: hp260497). The research ideas, computational approach, and analysis procedures were specified by the author. OpenAI Codex (GPT-6.1 Sol) was used under the author’s direction to assist with code development, numerical postprocessing, and manuscript preparation. The author reviewed and verified the resulting code, analyses, and manuscript.

Data availability

The implementation described in this article is planned for inclusion in a future release of SALMON. The input files will be made publicly available on GitHub. The numerical data supporting the findings are available from the author upon reasonable request.

Detailed timings for weak scaling

Table III gives the step times underlying Figure 5(a), including the prediction and explicit-correction contributions.

Density-matrix formulation of mesh exchange

Write a Bloch orbital as ψnk(r)=eik⋅runk(r)\psi_{nk}(\mathbf{r}) = e^{i\mathbf{k}\cdot\mathbf{r}}u_{nk}(\mathbf{r}), where unku_{nk} is cell periodic. The periodic pair density and its spatial Fourier transform are

ρ~n′nk′k(r′)=un′k′∗(r′)unk(r′).(23)\tilde{\rho}_{n'n}^{k'k}(\mathbf{r}') = u_{n'k'}^*(\mathbf{r}')u_{nk}(\mathbf{r}'). \tag*{(23)}
ρ~n′nk′k(G)=∫Ωe−iG⋅r′ρ~n′nk′k(r′) d3r′.(24)\tilde{\rho}_{n'n}^{k'k}(\mathbf{G}) = \int_{\Omega} e^{-i\mathbf{G}\cdot\mathbf{r}'}\tilde{\rho}_{n'n}^{k'k}(\mathbf{r}')\,\mathrm{d}^3r'. \tag*{(24)}

Here G\mathbf{G} is a reciprocal lattice vector. The usual exchange convolution multiplies this transform by K~SR(G+q)\tilde{K}_{\mathrm{SR}}(\mathbf{G}+\mathbf{q}) and transforms back to real space, where q=k−k′\mathbf{q}=\mathbf{k}-\mathbf{k}' is the momentum transfer. The screened multiplier is

K~SR(Q)=4πQ2[1−e−Q2/(4ωSR2)],(25)\tilde{K}_{\mathrm{SR}}(\mathbf{Q}) = \frac{4\pi}{Q^2}\left[1-e^{-Q^2/(4\omega_{\mathrm{SR}}^2)}\right], \tag*{(25)}

with limit π/ωSR2\pi/\omega_{\mathrm{SR}}^2 at Q=0Q=0. For an unscreened hybrid such as PBE0, the multiplier becomes 4π/Q24\pi/Q^2, and its Q=0Q=0 Coulomb singularity requires a separate treatment; the factor prediction and SBFR construction otherwise remain applicable. Here Q=G+q\mathbf{Q}=\mathbf{G}+\mathbf{q}.

To change the order of evaluation, use Bloch periodicity in Eq. (3),

ψn′k′∗(r′+R)ψnk(r′+R)=ei(k−k′)⋅Rψn′k′∗(r′)ψnk(r′).(26)\psi_{n'k'}^*(\mathbf{r}'+\mathbf{R})\psi_{nk}(\mathbf{r}'+\mathbf{R}) = e^{i(\mathbf{k}-\mathbf{k}')\cdot\mathbf{R}}\psi_{n'k'}^*(\mathbf{r}')\psi_{nk}(\mathbf{r}'). \tag*{(26)}

The occupied-orbital sum can then be performed before the sums over k′\mathbf{k}' and lattice translations. The density matrix contains products of the same occupied orbital at two positions, whereas the pair density contains an occupied orbital and the target orbital at the same position. Their relation is

∑n′∈occun′k′(r)ρ~n′nk′k(r′)=[∑n′∈occun′k′(r)un′k′∗(r′)]unk(r′).(27)\sum_{n'\in\mathrm{occ}}u_{n'k'}(\mathbf{r})\tilde{\rho}_{n'n}^{k'k}(\mathbf{r}') = \left[\sum_{n'\in\mathrm{occ}}u_{n'k'}(\mathbf{r})u_{n'k'}^*(\mathbf{r}')\right]u_{nk}(\mathbf{r}'). \tag*{(27)}
kk meshNodesMPI ranksDuring learningPost-learningPrediction onlyExplicit correction
434^{3}3122.5390.5820.1273.763
838^{3}20801.2720.3450.1291.858
16316^{3}1445761.7280.4480.1482.546
32332^{3}108843523.4240.7670.1505.086

Table III. Step times for the size-scaling series in Figure 5(a). All times are in seconds per step. The learning and post-learning columns average steps 1–80 and 81–160, respectively. Prediction-only and explicit-correction times are averaged separately within the post-learning interval.

The target orbital thus factors out of the occupied-band sum.

We use the uniform mesh

k=k0+∑α=13jαMαbα,0≤jα<Mα,(28)\mathbf{k} = \mathbf{k}_{0} + \sum_{\alpha=1}^{3} \frac{j_{\alpha}}{M_{\alpha}}\mathbf{b}_{\alpha}, \qquad0 \le j_{\alpha} < M_{\alpha}, \tag*{(28)}

where bα\mathbf{b}_{\alpha} are reciprocal lattice vectors and Nk=M1M2M3N_{k}=M_{1}M_{2}M_{3}. The origin k0\mathbf{k}_{0} may describe an unshifted or shifted mesh. Its dual translations are R=∑αlαaα\mathbf{R}=\sum_{\alpha}l_{\alpha}\mathbf{a}_{\alpha}, with 0≤lα<Mα0 \le l_{\alpha}<M_{\alpha} and aα⋅bβ=2πδαβ\mathbf{a}_{\alpha}\cdot\mathbf{b}_{\beta}=2\pi\delta_{\alpha\beta}. These translations form the set B\mathcal{B} in the Born–von Karman supercell. The sum over k′\mathbf{k}' runs over the full mesh Mk\mathcal{M}_{k}.

With the rephased orbitals and occupied projector defined in Sec. II D, the relation to the density matrix in Eq. (2) is

γˉk′(r,r′)=e−ik0⋅(r−r′)∑n∈occψnk′(r)ψnk′∗(r′),(29)\bar{\gamma}_{\mathbf{k}'}(\mathbf{r},\mathbf{r}')=e^{-i\mathbf{k}_{0}\cdot(\mathbf{r}-\mathbf{r}')}\sum_{n\in\mathrm{occ}}\psi_{n\mathbf{k}'}(\mathbf{r})\psi_{n\mathbf{k}'}^{*}(\mathbf{r}'), \tag*{(29)}
γ(r,r′)=eik0⋅(r−r′)Nk∑k′γˉk′(r,r′).(30)\gamma(\mathbf{r},\mathbf{r}')=\frac{e^{i\mathbf{k}_{0}\cdot(\mathbf{r}-\mathbf{r}')}}{N_{k}}\sum_{\mathbf{k}'}\bar{\gamma}_{\mathbf{k}'}(\mathbf{r},\mathbf{r}'). \tag*{(30)}

Thus γˉk′\bar{\gamma}_{\mathbf{k}'} is the density matrix resolved by k′\mathbf{k}', with a common phase removed to accommodate a shifted mesh. Let KB\mathcal{K}_{B} be the screened kernel made periodic on the Born–von Karman supercell. Folding the lattice translations into B\mathcal{B} gives

[VχSRunk](r)=−e−i(k−k0)⋅rNk∫Ωd3r′ vnk(r′)∑R∈BKB(r−r′−R)ei(k−k0)⋅R[V_{\chi}^{\mathrm{SR}}u_{n\mathbf{k}}](\mathbf{r})=-\frac{e^{-i(\mathbf{k}-\mathbf{k}_{0})\cdot\mathbf{r}}}{N_{k}}\int_{\Omega}d^{3}r'\,v_{n\mathbf{k}}(\mathbf{r}')\sum_{\mathbf{R}\in\mathcal{B}}\mathcal{K}_{B}(\mathbf{r}-\mathbf{r}'-\mathbf{R})e^{i(\mathbf{k}-\mathbf{k}_{0})\cdot\mathbf{R}}
×∑k′∈Mke−i(k′−k0)⋅Rγˉk′(r,r′).(31)\qquad{}\times\sum_{\mathbf{k}'\in\mathcal{M}_{k}}e^{-i(\mathbf{k}'-\mathbf{k}_{0})\cdot\mathbf{R}}\bar{\gamma}_{\mathbf{k}'}(\mathbf{r},\mathbf{r}'). \tag*{(31)}

The innermost sum is a discrete Fourier transform because

(k′−k0)⋅R=2π∑αjαlαMα.(\mathbf{k}'-\mathbf{k}_{0})\cdot\mathbf{R}=2\pi\sum_{\alpha}\frac{j_{\alpha}l_{\alpha}}{M_{\alpha}}.

For each fixed coordinate pair, Eqs. (14) and (15) evaluate the two mesh transforms with kernel multiplication between them. The remaining spatial integration gives Eq. (16), equivalent to Eq. (31) with the full k′\mathbf{k}' sum.

The spatial Fourier transform is performed when preparing KB\mathcal{K}_{B} from Eq. (25). On the NrNkN_{r}N_{k}-point supercell grid, the implementation stores the quadrature-weighted kernel

K(d)=1NrNkRe⁡∑Q∈QBvKK~SR(Q)eiQ⋅d,(32)\mathcal{K}(\mathbf{d})=\frac{1}{N_{r}N_{k}}\operatorname{Re}\sum_{\mathbf{Q}\in\mathcal{Q}_{\mathrm{BvK}}}\widetilde{\mathcal{K}}_{\mathrm{SR}}(\mathbf{Q})e^{i\mathbf{Q}\cdot\mathbf{d}}, \tag*{(32)}

where QBvK\mathcal{Q}_{\mathrm{BvK}} is the supercell reciprocal grid. The real part imposes the discrete even-kernel convention at Nyquist modes. Using this stored kernel in Eq. (15) includes the spatial quadrature in the final contraction. The mesh FFTs use the unnormalized FFTW convention, leaving the exchange weight 1/Nk1/N_{k} explicit in Eq. (16).

The reconstruction must transform the wave vector and both spatial coordinates together. For a space-group operation s={S∣t}s=\{S|\mathbf{t}\}, define cell coordinates by

Srs+t=r+Ls(rs),(33)S\mathbf{r}_{s}+\mathbf{t}=\mathbf{r}+\mathbf{L}_{s}(\mathbf{r}_{s}), \tag*{(33)}

where SS is the orthogonal matrix for the rotational part (including improper rotations), t\mathbf{t} is the translation vector, Ls\mathbf{L}_{s} is a lattice translation, and ks\mathbf{k}_{s} is the mesh representative of S−1kS^{-1}\mathbf{k} modulo a reciprocal lattice vector. Covariance of the occupied projector gives

Bk(r,r′)=eik⋅[Ls(rs′)−Ls(rs)]eik0⋅[(rs−rs′)−(r−r′)] Bks(rs,rs′).(34)B_{\mathbf{k}}(\mathbf{r},\mathbf{r}')=e^{i\mathbf{k}\cdot[\mathbf{L}_{s}(\mathbf{r}'_{s})-\mathbf{L}_{s}(\mathbf{r}_{s})]}e^{i\mathbf{k}_{0}\cdot[(\mathbf{r}_{s}-\mathbf{r}'_{s})-(\mathbf{r}-\mathbf{r}')]}\,B_{\mathbf{k}_{s}}(\mathbf{r}_{s},\mathbf{r}'_{s}). \tag*{(34)}

The first phase accounts for cell wrapping. The second comes from rephasing the Bloch orbitals and vanishes when k0=0\mathbf{k}_{0}=0. No individual-band phase choice is required because common orbital phases cancel in the projector.

Factor alignment and learning schedule

For an unaligned factor Ξraw\Xi^{\mathrm{raw}} and the latest accepted factor Ξref\Xi^{\mathrm{ref}}, the column rotation solves

Q⋆=arg min⁡Q†Q=I∥ΞrawQ−Ξref∥F.(35)Q_{\star}=\underset{Q^{\dagger}Q=I}{\operatorname{arg\,min}}\left\lVert\Xi^{\mathrm{raw}}Q-\Xi^{\mathrm{ref}}\right\rVert_{F}. \tag*{(35)}

The factor norms are invariant under QQ, so minimizing this distance is equivalent to maximizing Re⁡Tr⁡[Q†M]\operatorname{Re}\operatorname{Tr}[Q^{\dagger}M], where M=(Ξraw)†ΞrefM=(\Xi^{\mathrm{raw}})^{\dagger}\Xi^{\mathrm{ref}}. If M=UΣV†M=U\Sigma V^{\dagger} is its singular-value decomposition, Q⋆=UV†Q_{\star}=UV^{\dagger} and the aligned factor is Ξ=ΞrawQ⋆\Xi= \Xi^{\mathrm{raw}}Q_{\star}. Since Q⋆Q⋆†=IQ_{\star}Q_{\star}^{\dagger}=I, this transformation leaves −ΞΞ†-\Xi\Xi^{\dagger} unchanged. Each explicit factor is aligned to the already aligned reference, so the stored factors and their increments share a common column gauge.

The cumulative least-squares fit requires only a real symmetric 2×22 \times2 matrix and a two-component vector per target kk history. For accepted training samples, we accumulate

Gab=∑m(Da,m,Db,m),(36)G_{ab}=\sum_m(D_{a,m},D_{b,m}), \tag*{(36)}
ba=∑m(Da,m,Ym),a,b=1,2.(37)b_a=\sum_m(D_{a,m},Y_m),\qquad a,b=1,2. \tag*{(37)}

Differentiating the regularized objective gives

\begin{pmatrix} G_{11}+\lambda& G_{12}\\ G_{12} & G_{22}+\lambda \end{pmatrix} \begin{pmatrix} c_1\\ c_2 \end{pmatrix} = \begin{pmatrix} b_1\\ b_2 \end{pmatrix}, \tag*{(38)}\end{pmatrix}

where λ=10−8(G11+G22)\lambda=10^{-8}(G_{11}+G_{22}). Only explicit endpoints supply training targets YmY_m. The quadratic interpolation in ss retains the latest increment as its initial slope and approaches the fitted increment at the next correction endpoint.

The predicted ACE factor is held fixed during the internal predictor–corrector stages of each time step.

Starting from one explicit factor at the initial time, Nhi−1N_{\mathrm{hi}}-1 correction intervals are needed to populate the factor history, followed by NteachN_{\mathrm{teach}} intervals to collect the training samples. The default number of warmup time steps is therefore

Nwarm=(Nhi−1+Nteach)K,(39)N_{\mathrm{warm}}=(N_{\mathrm{hi}}-1+N_{\mathrm{teach}})K, \tag*{(39)}

where NhiN_{\mathrm{hi}} is the number of stored factors, NteachN_{\mathrm{teach}} is the minimum number of explicit target increments YmY_m required for fitting before prediction starts, and KK is the number of time steps between explicit corrections. Each training sample comprises (D1,m,D2,m,Ym)(D_{1,m},D_{2,m},Y_m). For the history and training counts used here, Nwarm=10KN_{\mathrm{warm}}=10K. Linear extrapolation uses the same prediction-start schedule.

Detailed computational and memory costs

The mesh-FFT estimates in Table 1 include transforms and kernel multiplication. Projector generation, symmetry restoration, target contraction, and ACE factor construction contribute to the full reconstruction cost CbuildsymC_{\mathrm{build}}^{\mathrm{sym}}. ACE application and prediction use the symmetry-inequivalent target points, whereas each reconstructed exchange action retains the full source mesh.

Method or operationComputational work
Direct exchange (spatial FFT)O(Nocc2Nr2Nrlog⁡Nr)O(N_{\mathrm{occ}}^{2}N_{r}^{2}N_{r}\log N_{r})
Full-row mesh FFTs (Off reference)O(Nr2Nklog⁡Nk)O(N_{r}^{2}N_{k}\log N_{k})
SBFR mesh FFTsO(NaNrNklog⁡Nk)O(N_{a}N_{r}N_{k}\log N_{k})
Distributed mesh FFTs (per rank)O(NaNrNklog⁡Nk/Pk)O(N_{a}N_{r}N_{k}\log N_{k}/P_{k})
ACECapply+CbuildglobalC_{\mathrm{apply}}+C_{\mathrm{build}}^{\mathrm{global}}
Learned ACECapply+Cpredict+Cbuildglobal/KC_{\mathrm{apply}}+C_{\mathrm{predict}}+C_{\mathrm{build}}^{\mathrm{global}}/K
SBFR + ACECapply+CbuildsymC_{\mathrm{apply}}+C_{\mathrm{build}}^{\mathrm{sym}}
SBFR + learned ACECapply+Cpredict+Cbuildsym/KC_{\mathrm{apply}}+C_{\mathrm{predict}}+C_{\mathrm{build}}^{\mathrm{sym}}/K

Table 1. Leading computational work, excluding MPI padding and communication. Costs are totals unless marked per rank. ACE combinations give post-learning averages; the mesh-FFT rows cover transforms and kernel multiplication only.

At fixed NkN_k, grid spacing, and density, NrN_r and NoccN_{\mathrm{occ}} grow in proportion to system size NN. For fixed Na/NrN_a/N_r, the SBFR mesh-FFT work is O(N2)O(N^2), but orbital contractions and ACE factor construction retain an O(N3)O(N^3) cost, including O(NkNocc3)O(N_kN_{\mathrm{occ}}^3) dense orbital-space operations. At fixed KK, learned prediction reduces reconstruction frequency without changing its system-size exponent. Alignment and coefficient fitting are included in the explicit-update cost for learned ACE.

For the CPU distribution over spatial coordinate pairs, a pass assigns at most cc pairs (ra,r′)(r_a,r') to each of the PkP_k ranks. The number of passes and allocated pair slots are

qc=⌈NaNrPkc⌉,Npairpad=Pkqc.(40)q_c=\left\lceil\frac{N_aN_r}{P_kc}\right\rceil,\qquad N_{\mathrm{pair}}^{\mathrm{pad}}=P_kq_c. \tag*{(40)}

Each active rank executes forward and backward mesh FFTs for cc arrays of length NkN_k per pass. Ranks without a valid pair skip the FFTs but participate in the transposes. The executed-work upper bounds are

WFFTrank=O(qccNklog⁡Nk),(41)W_{\mathrm{FFT}}^{\mathrm{rank}}=O(q_c cN_k\log N_k), \tag*{(41)}
WFFTall=O(NpairpadNklog⁡Nk).(42)W_{\mathrm{FFT}}^{\mathrm{all}}=O(N_{\mathrm{pair}}^{\mathrm{pad}}N_k\log N_k). \tag*{(42)}

Ignoring batch padding, balanced assignment of the independent coordinate pairs gives the maximum local FFT work

WFFTrank,max=O[max⁡(1,NaNrPk)Nklog⁡Nk].(43)W_{\mathrm{FFT}}^{\mathrm{rank,max}}=O\left[\max\left(1,\frac{N_aN_r}{P_k}\right)N_k\log N_k\right]. \tag*{(43)}

Thus, spatial-pair parallelism reduces the local FFT work up to Pk∼NaNrP_k\sim N_aN_r, where each active rank handles one pair.

The batch size cc balances the coordinate pairs across ranks subject to a workspace bound. For the present 12312^3 grid, Na=75N_a=75 gives 129600 independent pairs. Negligible padding recovers the useful-work estimate O(NaNrNklog⁡Nk)O(N_aN_rN_k\log N_k).

Each pass uses two fixed-count all-to-all transposes. If nk,max⁡n_{k,\max} is the largest number of full-mesh points assigned to one rank, the per-rank payload over one reconstruction is

Vtransposerank=O(qcPknk,max⁡)(44)V_{\mathrm{transpose}}^{\mathrm{rank}}=O(q_cP_kn_{k,\max}) \tag*{(44)}

in complex elements, excluding collective latency and protocol overhead. For balanced full-mesh ownership, nk,max⁡=O(Nk/Pk)n_{k,\max}=O(N_k/P_k) and the payload is O(qcNk)O(q_cN_k) per rank. In the balanced regime qc=O(NaNr/Pk)q_c=O(N_aN_r/P_k), the local FFT work and payload are O(NaNrNklog⁡Nk/Pk)O(N_aN_rN_k\log N_k/P_k) and O(NaNrNk/Pk)O(N_aN_rN_k/P_k), respectively. Communication time also depends on the collective algorithm and network mapping.

ACE factors are distributed over local real-space rows. Their overlaps require spatial reductions, while local matrix products and grid operations use OpenMP. For complex double precision, the three stored histories require

Bhist=3×16Nr,localNoccNk,local bytes.(45)B_{\mathrm{hist}}=3\times16N_{r,\mathrm{local}}N_{\mathrm{occ}}N_{k,\mathrm{local}}\ \text{bytes}. \tag*{(45)}

Peak memory additionally includes the current and predicted factors, orbitals, and workspaces.

Comparison of ACE and direct exchange

To check the accuracy of ACE before introducing temporal prediction, we compare it with direct screened exchange under the laser-driven conditions of Sec. IV A. Both calculations use identical initial orbitals, grid, kk mesh, field, and time integrator.

Figure 7 compares the current and absorbed energy. The full-record relative current difference is 0.15%. The final absorbed-energy difference is 0.010%. Figure 8 shows that the HHG spectra also closely agree, using the same 15-fs window and spectral definition in Sec. III.

Figure 7. Comparison of direct screened exchange and ACE without temporal prediction for (a) the charge-current density and (b) the absorbed energy per Si cell. The laser photon energy is 1.55 eV, the peak intensity is 101210^{12} W cm−2^{-2}, and the pulse has 10-fs full support. The dotted line marks the pulse end.

Figure 8. HHG spectra for direct screened exchange and ACE without temporal prediction. Both use the same 15-fs window and are normalized by the maximum direct-exchange intensity over the displayed range.

Comparison with linear extrapolation

The simplest predictor extends the change between the two latest explicit factors,

Ξm,ℓlin=Ξm+s(Ξm−Ξm−1),(46)\Xi^{\mathrm{lin}}_{m,\ell} = \Xi_m + s\left(\Xi_m-\Xi_{m-1}\right), \tag*{(46)}

using the endpoint and interval notation of Sec. II C. This predictor retains gauge alignment and periodic explicit corrections but requires no fitted coefficients.

PredictorCurrent error (%)Energy error (%)
Linear extrapolation10.680.55
Learned prediction3.120.90

Table IV. Errors of linear extrapolation and learned prediction at K=8K = 8. The current error is integrated over the full trajectory, and the absorbed-energy error is the relative difference at the final time.

To determine whether fitting the coefficients improves prediction, we compare linear extrapolation and learned ACE against conventional ACE under identical conditions (Sec. IV A).

For the tested correction intervals K=2K = 2, 4, 8, and 12, linear extrapolation gives current errors approximately 2.5–3.4 times those of learned prediction. At K=8K = 8, both final absorbed-energy errors are below 1%, with linear extrapolation giving the smaller error. The ratio of RT computation times for linear extrapolation to learned prediction is approximately 0.99. Learned prediction therefore improves current fidelity at essentially the same cost.

Choice of time propagator

Parallel-transport implicit propagators allow larger time steps but require iterative solution of the time-dependent equations [21, 22]. In preliminary tests of our implementation, repeated ACE reconstruction and Hamiltonian applications within these iterations offset the benefit of the larger time steps. We therefore use fourth-order Taylor propagation with predictor–corrector updates. Learned ACE reduces reconstruction frequency directly: prediction steps require only factor combinations with two fitted coefficients, without an iterative exchange solve. Gauge alignment is performed only at explicit history updates.

References

References

  1. [1]E. Runge and E. K. U. Gross, Density-functional theory for time-dependent systems, Phys. Rev. Lett. 52, 597 (1984).DOI
  2. [2]K. Yabana and G. F. Bertsch, Time-dependent local-density approximation in real time, Phys. Rev. B 54, 4484 (1996).
  3. [3]G. F. Bertsch, J.-I. Iwata, A. Rubio, and K. Yabana, Real-space, real-time method for the dielectric function, Phys. Rev. B 62, 7998 (2000).DOI
  4. [4]S. Meng and E. Kaxiras, Real-time, local basis-set implementation of time-dependent density functional theory for excited state dynamics simulations, J. Chem. Phys. 129, 054110 (2008).DOI
  5. [5]J. Xu, T. E. Carney, R. Zhou, C. Shepard, and Y. Kanai, Real-time time-dependent density functional theory for simulating nonequilibrium electron dynamics, J. Am. Chem. Soc. 146, 5011 (2024).DOI
  6. [6]T. Otobe, M. Yamagiwa, J.-I. Iwata, K. Yabana, T. Nakatsukasa, and G. F. Bertsch, First-principles electron dynamics simulation for optical breakdown of dielectrics under an intense laser field, Phys. Rev. B 77, 165104 (2008).DOI
  7. [7]Y. Shinohara, K. Yabana, Y. Kawashita, J.-I. Iwata, T. Otobe, and G. F. Bertsch, Coherent phonon generation in time-dependent density functional theory, Phys. Rev. B 82, 155110 (2010).DOI
  8. [8]S. A. Sato, K. Yabana, Y. Shinohara, T. Otobe, and G. F. Bertsch, Numerical pump-probe experiments of laser-excited silicon in nonequilibrium phase, Physical Review B 89, 064304 (2014).DOI
  9. [9]T. Otobe, Y. Shinohara, S. A. Sato, and K. Yabana, Femtosecond time-resolved dynamical Franz–Keldysh effect, Phys. Rev. B 93, 045124 (2016).DOI
  10. [10]S. Ghimire, A. D. DiChiara, E. Sistrunk, P. Agostini, L. F. DiMauro, and D. A. Reis, Observation of high-order harmonic generation in a bulk crystal, Nat. Phys. 7, 138 (2011).DOI
  11. [11]N. Tancogne-Dejean, O. D. Mücke, F. X. Kärtner, and A. Rubio, Ellipticity dependence of high-harmonic generation in solids originating from coupled intraband and interband dynamics, Nat. Commun. 8, 745 (2017).
  12. [12]L. Reining, V. Olevano, A. Rubio, and G. Onida, Excitonic effects in solids described by time-dependent density-functional theory, Phys. Rev. Lett. 88, 066404 (2002).
  13. [13]J. Heyd, G. E. Scuseria, and M. Ernzerhof, Hybrid functionals based on a screened coulomb potential, J. Chem. Phys. 118, 8207 (2003).DOI
  14. [14]A. V. Krukau, O. A. Vydrov, A. F. Izmaylov, and G. E. Scuseria, Influence of the exchange screening parameter on the performance of screened hybrid functionals, J. Chem. Phys. 125, 224106 (2006).pubmed.ncbi.nlm.nih.gov/17176133
  15. [15]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
  16. [16]S. A. Sato, Y. Taniguchi, Y. Shinohara, and K. Yabana, Nonlinear electronic excitations in crystalline solids using meta-generalized gradient approximation and hybrid functional in time-dependent density functional theory, J. Chem. Phys. 143, 224116 (2015).DOI
  17. [17]T. Ketolainen, N. Macháčová, and F. Karlický, Optical gaps and excitonic properties of 2D materials by hybrid time-dependent density functional theory: Evidences for monolayers and prospects for van der waals heterostructures, J. Chem. Theory Comput. 16, 5876 (2020).DOI
  18. [18]C. D. Pemmaraju, Valence and core excitons in solids from velocity-gauge real-time TDDFT with range-separated hybrid functionals: An LCAO approach, Com- put. Condens. Matter 18, e00348 (2019).
  19. [19]J. Xu, C. Shepard, and Y. Kanai, All-electron RT-TDDFT with exact exchange based on numeric atom-centered orbitals for condensed matter (2026), arXiv:2610.12896 [physics.comp-ph].arxiv.org/abs/2610.12896
  20. [20]L. Lin, Adaptively compressed exchange operator, J. Chem. Theory Comput. 12, 2242 (2016).DOI
  21. [21]W. Jia and L. Lin, Fast real-time time-dependent hybrid functional calculations with the parallel transport gauge and the adaptively compressed exchange formulation, Comput. Phys. Commun. 240, 21 (2019).DOI
  22. [22]R. Liu, Z. Guo, Q. Sha, T. Zhao, H. Li, W. Hu, L. Liu, G. Tan, and W. Jia, Large scale finite-temperature real-time time dependent density functional theory calculation with hybrid functional on ARM and GPU systems, in 2025 IEEE International Parallel and Distributed Processing Symposium (IPDPS) (IEEE, 2025) pp. 801–812.arxiv.org/abs/2501.03061
  23. [23]J. Lu and L. Ying, Compression of the electron repulsion integral tensor in tensor hypercontraction format with cubic scaling cost, J. Comput. Phys. 302, 329 (2015).arxiv.org/abs/1410.7757
  24. [24]W. Hu, L. Lin, and C. Yang, Interpolative separable density fitting decomposition for accelerating hybrid density functional calculations with applications to defects in silicon, J. Chem. Theory Comput. 13, 5420 (2017), arXiv:1707.09141.arxiv.org/abs/1707.09141
  25. [25]H.-Y. Ko, M. F. Calegari Andrade, Z. M. Sparrow, J.-a. Zhang, and R. A. DiStasio, Jr., High-throughput condensed-phase hybrid density functional theory for large-scale finite-gap systems: The SeA approach, J. Chem. Theory Comput. 19, 4182 (2023).DOI
  26. [26]J. Li, L. Wan, S. Jiao, W. Hu, and J. Yang, Low-rank approximations to accelerate hybrid functional enabled real-time time-dependent density functional theory within plane waves, Electron. Struct. 5, 014008 (2023).DOI
  27. [27]K. Wu, X. Qin, W. Hu, and J. Yang, Low-rank approximations accelerated plane-wave hybrid functional calculations with k-point sampling, J. Chem. Theory Comput. 18, 206 (2022).DOI
  28. [28]A. Rettig, J. Lee, and M. Head-Gordon, Even faster exact exchange for solids via tensor hypercontraction, J. Chem. Theory Comput. 19, 5773 (2023).DOI
  29. [29]H. S. Bhat, P. Gupta, and C. M. Isborn, Scalable learning of potentials to predict time-dependent Hartree–Fock dynamics, APL Mach. Learn. 2, 046112 (2024).arxiv.org/abs/2408.04765
  30. [30]K. Shah and A. Cangi, Machine learning time propagators for time-dependent density functional theory simulations, Mach. Learn.: Sci. Technol. 7, 035019 (2026).
  31. [31]VASP developers, ISYM, https://vasp.at/wiki/index.php/ISYM (2026), vASP Wiki, accessed October 20, 2026.
  32. [32]Y. Cao, M.-Y. Zhang, P. Lin, M. Chen, and X. Ren, Applying space-group symmetry to speed up hybrid-functional calculations within the framework of numerical atomic orbitals, J. Chem. Theory Comput. 21, 8086 (2025).arxiv.org/abs/2504.02422
  33. [33]C.-N. Yeh and M. A. Morales, Low-scaling algorithms for GW and constrained random phase approximation using symmetry-adapted interpolative separable density fitting, J. Chem. Theory Comput. 20, 3184 (2024).arxiv.org/abs/2401.12308
  34. [34]M. Noda, S. A. Sato, Y. Hirokawa, M. Uemoto, T. Takeuchi, S. Yamada, A. Yamada, Y. Shinohara, M. Yamaguchi, K. Iida, I. Floss, T. Otobe, K.-M. Lee, K. Ishimura, T. Boku, G. F. Bertsch, K. Nobusada, and K. Yabana, SAMMON: Scalable ab-initio light–matter simulator for optics and nanoscience, Comput. Phys. Commun. 235, 356 (2019).
  35. [35]J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized gradient approximation made simple, Phys. Rev. Lett. 77, 3865 (1996).pubmed.ncbi.nlm.nih.gov/10062328
  36. [36]O. Sugino and Y. Miyamoto, Density-functional approach to electron dynamics: Stable simulation under a self-consistent field, Phys. Rev. B 59, 2579 (1999).DOI
  37. [37]A. Castro, M. A. L. Marques, and A. Rubio, Propagators for the time-dependent Kohn–Sham equations, J. Chem. Phys. 121, 3425 (2004).
  38. [38]J. Kolafa, Time-reversible always stable predictor-corrector method for molecular dynamics of polarizable molecules, J. Comput. Chem. 25, 335 (2004).pubmed.ncbi.nlm.nih.gov/14696069
  39. [39]S. Mandal and N. N. Nair, Speeding-up ab initio molecular dynamics with hybrid functionals using adaptively compressed exchange operator based multiple timestepping, J. Chem. Phys. 151, 151102 (2019).DOI
  40. [40]P. H. Schönemann, A generalized solution of the orthogonal Procrustes problem, Psychometrika 31, 1 (1966).DOI
  41. [41]J. P. Perdew and A. Zunger, Self-interaction correction to density-functional approximations for many-electron systems, Physical Review B 23, 5048 (1981).DOI
  42. [42]RIKEN Center for Computational Science, About Fugaku, https://www.r-ccs.riken.jp/en/fugaku/about/, accessed October 20, 2026.
  43. [43]K. Yabana, T. Sugi­yama, Y. Shinohara, T. Otobe, and G. F. Bertsch, Time-dependent density functional theory for strong electromagnetic fields in crystalline solids, Phys. Rev. B 85, 045134 (2012).DOI

Paper details

vlnc:W_7ntfx5m6
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 2

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

  • Physics and Astronomy

OpenAI Codex (GPT-6.1 Sol)

View statement
Looking for related papers…
Contents