In this paper
AbstractIntroductionMethodology and designLocal potential across the periodic tableNonlocal channelsResultsLearned potentials of elements.ConclusionsMethodsHamiltonian and basisSpecie potentialsNonlocal projectorsValence electron condition, data and splitsLoss, eigensolver and trainingEvaluationData and code availabilityAcknowledgmentsCompeting interestsSupplementary InformationThe network, the Hamiltonian and the training lossAccuracy and band structures of held-out structuresAtomic construction of channelsCost against cell sizeComparison with alternative modelsLimitations of the current EpNet modelReferencesWorksPhysics and Astronomy
Learning Electronic Band Structures Across the s, p, d Blocks with an Empirical Pseudopotential Network
Abstract
Electronic band structures govern how crystals conduct electrons and absorb light. Density-functional theory (DFT) computes them through costly self-consistent iterations, while existing machine-learning models either predict isolated scalar properties or output Hamiltonians that still require DFT solvers. Here we present EpNet, a learned empirical pseudopotential for crystals spanning 63 elements across the s, p and d blocks. An equivariant neural network builds its Hamiltonian from physics-motivated local potentials and nonlocal projectors. Trained solely on band energies, EpNet predicts complete band structures directly from atomic coordinates. On unseen chemical systems, its median errors are \winAll\ eV for band energies and \gapAll\ eV for band gaps. On 600,000 structures from an independent database, its band-gap errors are overall less than half those of existing models, at about 1% of DFT cost. EpNet thus enables fast screening of electronic properties for materials discovery and electronic-structure simulations of large systems.
Introduction
The electronic band structure of a crystal is critical for technological applications because it determines the band gap, carrier masses, Fermi surface, and optical response, serving as the starting point for transport, defect, and electron–phonon calculations in semiconducting and superconducting devices. Density-functional theory (DFT) [1, 2] is the standard method for computing band structures and has been widely adopted to construct online databases providing public access to precomputed band gaps for millions of structures [3, 4, 5, 6, 7, 8, 9].
However, modern materials design demands rapid screening across vast chemical spaces consisting of several million to billion candidate structures, where standard DFT calculations become a severe bottleneck. Furthermore, DFT is computationally prohibitive for large-scale applications such as disordered systems, finite-temperature dynamics, or multi-body interactions like electron–phonon coupling.
Machine learning has emerged as a powerful tool to accelerate electronic structure calculations [10]. Direct property-prediction models use equivariant graph neural networks to map crystal structures to target quantities like band gaps or densities of states [11, 12, 13, 14, 15], reaching mean gap errors as low as 0.16 eV [15, 16]. However, because they lack an explicit Hamiltonian formulation, predicted properties need not be mutually consistent, and derived features like carrier effective masses or gap nature cannot be extracted without training on separate labels. To preserve the operator structure, machine-learned DFT Hamiltonians have been developed to predict converged Kohn–Sham matrices using localized atomic orbitals [17, 18, 19, 20, 21, 22], with universal implementations spanning the periodic table at meV-level matrix-element errors [23, 24, 25]. Nevertheless, these models remain bound to specific DFT codes and basis sets. Extending them to higher-level electronic structure methods requires generating new matrix training sets [26], plane-wave representations disrupt network locality [27], and actual band-energy errors are often 10 to 50 times larger than the reported matrix-element errors [22]. Models trained on computed properties rather than on matrices avoid the need for higher-level matrices. For molecules, an equivariant network that predicts an effective electronic Hamiltonian, trained on CCSD(T) data, proves to be more accurate than hybrid density functionals for a range of properties [28].
On the other hand, semi-empirical tight-binding models eliminate the need for a DFT solver [29, 30, 31], but their minimal basis set degrades conduction-band quality, leading to average band-gap errors of 0.7–0.8 eV [32]. Machine-learned tight-binding variants resolve this accuracy gap but remain limited to system-specific parameterizations [21].
The empirical pseudopotential method (EPM) [33, 34, 35] combines several of these strengths. Assembled from atomic form factors in a plane-wave basis, an EPM Hamiltonian yields complete band structures at any -point through a single matrix diagonalization while bypassing self-consistent iterations in common DFT solvers. Although traditional EPM suffered from poor transferability due to material-by-material fitting, machine-learning models have recently revived the framework by predicting screened local potentials from DFT potentials [36] or band energies [37, 38, 39]. However, these models remain limited to single elements or narrow semiconductor families. Scaling EPM across the periodic table presents a fundamental physical challenge: while main-group and states are extended and well-described by weak, smooth pseudopotentials, transition-metal states form narrow resonances governed by localized orbitals [40, 41, 42]. Historically, empirical approaches captured states only through strongly attractive terms fitted to individual metals [43], or by importing nonlocal operators and overlap matrices directly from ab initio pseudopotentials [44, 45, 46, 47].
To address this challenge, we present EpNet, a unified learned EPM framework for crystals spanning 63 elements across the , , and blocks. EpNet assigns each atom a local potential derived from its chemical identity and local environment, alongside nonlocal channels matched to its valence states. Parameterized by 78,964 weights and trained end-to-end through the eigensolver on band energies alone, EpNet predicts complete band structures directly from crystal coordinates. On held-out chemical systems, it achieves median errors of 0.061 eV for band energies and 0.102 eV for band gaps. Evaluated on 600,000 completely unseen structures from the Alexandria database, EpNet achieves a median band-gap error of 0.11 eV—overall less than half that of existing models—while reducing computational cost by two to three orders of magnitude relative to plane-wave DFT. To our knowledge, EpNet is the first learned EPM model capable of predicting complete band structures across more than 60 elements from crystal geometry alone. In the following sections, we detail the architecture and design of EpNet, evaluate its performance across benchmark datasets, discuss its downstream applications, and present our conclusions.
Methodology and design
In the classical EPM, a crystal structure is mapped to a plane-wave Hamiltonian matrix (Figure 1a). For a unit cell of volume with atomic positions , the Hamiltonian matrix elements in the plane-wave basis are given by
Figure 1. Figure 1. Design of EpNet. a, An equivariant network maps element features and atomic environments to atom-centered operators, which are assembled into a plane-wave Hamiltonian and overlap , followed by a diagonalization. Training propagates the error in the PBE band energies of about 12,000 Materials Project crystals through the eigensolver. b, Local potential: starts from a screened empty core () or the atom’s own potential (). c, Nonlocal projectors. d, Nonlocal projectors, built from the scalar-relativistic all-electron atom on its own local potential. e, The 63 elements, shaded by the number of training structures that contain each; gray, outside the scope. f, Training band error by element.
where . The overlap matrix is the identity matrix, except in cells that contain -block atoms, as described below.
Local potential across the periodic table
To make the EPM transferable across chemical environments, we express the local potential of atom as summarized in Figure 1b:
This formulation comprises three main terms. First, the atomic specie potential defines an isotropic base potential parameterized by elemental features (e.g., period, group, electronegativity, and covalent radius). For elements, it starts from a Thomas–Fermi-screened empty-core potential[48]; for elements, it starts from the neutral atom’s self-consistent potential. Second, the environmental correction accounts for local coordination and chemical bonding. An equivariant neural network with a 6 Å cutoff predicts invariant, vector, and rank-2 tensor features for each atom, from which four multipole readouts (, each with components) are constructed. A router mixes these readouts with per-atom weights (a softmax, ) computed from the local valence-electron density and the electronegativity contrast relative to neighboring atoms. This term injects smooth, anisotropic environment effects into the potential, ensuring that the predicted band energies remain smoothly differentiable with respect to atomic displacements. Third, a Gaussian charge distribution on each atom (constrained to maintain cell neutrality) captures long-range electrostatic differences between cations and anions, ensuring accurate band gaps for ionic compounds.
In addition, two design choices ensure that behaves as a true pseudopotential rather than a basis-set fit: (i) the plane-wave kinetic energy cutoff is element-dependent (12 Ry by default, 16 Ry for B and C, and 20 Ry for N, O, F, and -block elements) to accommodate compact orbitals, and (ii) band energies are explicitly constrained to remain convergent as the basis size is enlarged during the training stage.
Nonlocal channels
The nonlocal potential distinguishes between extended and localized valence states. For atoms, the nonlocal operator (Figure 1c) consists of Kleinman–Bylander projectors[49] built from two Gaussian functions per angular momentum channel at radii set by the element’s valence orbitals, while the specie-level strengths come from the hypernetwork, with environment-dependent corrections per atom. Each -block element instead carries a channel constructed directly from its free atom (Figure 1d) to capture the localized states. From the scalar-relativistic atom[50], we take the level and orbital , and replace the orbital inside a core radius by a smooth, nodeless pseudo-orbital . Crucially, is not renormalized. Beyond it is identical to the all-electron orbital, so its tail, which overlaps with neighboring atoms and therefore sets the -band width, is exact. Inside the smooth holds less norm than the true orbital, and this missing norm is restored by an overlap operator . The local potential is the neutral atom’s own beyond a radius , stored as a fixed table . On this potential, the channel is an ultrasoft pseudopotential[51, 52] with two reference energies, and , with overlap . The projectors are dual to the pseudo-orbitals at the two energies (), and the strengths are
where are the all-electron radial orbitals at the two reference energies (), the brackets are radial integrals over the core region, so that and are matrices of numbers, and restores the norm and overlaps that the pseudo-orbitals miss inside (Supplementary equations (10)–(15)). By construction, the pseudo-atom reproduces the all-electron level and orbital exactly at both reference energies and interpolates the scattering between them. On the other hand, the and channels of -block elements each use a single reference energy, their own atomic level ( or ). Because these projectors are recomputed from the stored atomic states at every evaluation, the reference channels remain exact as the learned network parameters evolve.
In a crystal environment, the channel changes through two learned corrections: (i) a shift of the level, one learned value per element plus one per atom from its environment, introduced as , which shifts the atomic level by to first order without changing the projectors; and (ii) a learned, bounded radial stretch of the pseudo-orbitals, , one per element, for the seven elements where it remains numerically well conditioned (Supplementary § 13). In addition, the augmentation term is added to the projector strengths, , in which is the crystal’s local potential at the atom, minus the atom’s own, averaged over the augmentation charge, so that the level follows the local electrostatic potential of the surrounding crystal[51].
Model training. We considered a total of 12,136 PBE[53] band structures calculated along high-symmetry paths from the Materials Project[3] for model training. The dataset comprises non-magnetic crystal structures containing up to 24 atomic sites without corrections. In total, the dataset covers 63 elements, ranging from 31 compounds containing Mn to 2,674 compounds containing O (Figure 1e). Unlike previous approaches[16], our held-out validation and test sets are defined by chemical system and structure type rather than through random splitting. Entire chemical systems (all structures sharing a given set of elements) and distinct structure types are assigned to the held-out sets via hashing prior to training. This strategy prevents polymorphs or close structural analogues of test materials from appearing in the training set, avoiding the artificially inflated predictive performance on novel chemical domains typical of random splits[54]. The held-out sets are excluded from all training stages, and no held-out chemical system appears in the training split.
EpNet was trained end-to-end using 78,964 parameters according to the architecture detailed in Figure S1. The loss function combines the band-energy error within an energy window around the band gap or Fermi level, a dedicated gap penalty, and a weakly weighted regularization term across all occupied bands. For gapped compounds featuring open-shell -block elements, the loss function also incorporates target labels for the unoccupied manifold. Training follows a curriculum learning strategy[55], from simpler to harder systems, with each stage initialized from the weights of the preceding one. First, -block compounds are trained to establish the universal local potential. Next, small unit cells representing distinct -block subgroups are introduced sequentially ( elements, early -block, elements, and open-shell elements). Finally, the dataset is expanded to include cells from 8 sites to 24 sites via two steps. The resulting median training errors range between 0.03 eV and 0.17 eV across the elements, as detailed in
Figure S1. The full network and the assembly of the Hamiltonian. Gray: fixed inputs and operations; white: learned networks; blue, aqua and orange: the local potential and charges, the projectors and the channels, as in main-text Figure 1. The encoder features , and feed the environment term, the charges and the per-atom corrections of both projector sets; the router weights mix the four environment readouts; a hypernetwork maps the element features to the specie potentials and to the specie projector strengths .
Figure 1f. The highest training errors are predominantly concentrated in Mo, W, S, Se, and Te, originating primarily from layered mixed Mo/W dichalcogenide structures containing large vacuum gaps in the periodic supercell[56]. Specifically, 38 of the 64 evaluated training structures containing Mo or W exhibit these elevated median errors of 0.18 eV. After excluding these layered systems, the median training errors are reduced to 0.04 eV for Mo and W, and 0.05–0.06 eV for S, Se, and Te. Hence, these errors are specific to layered stacks with large vacuum gaps rather than to Mo, W or the chalcogens in general (Supplementary Section 6).
Results
Model evaluation across 63 elements. Figure 2 summarizes the performance of the final model on the 2,050 held-out structures (containing up to 24 sites) whose chemical systems were completely absent from the training set. The median band-energy error within the evaluation window is 0.061 eV, with 74% of all structures falling within an error of 0.1 eV. Representative band structures of held-out compounds, selected by error percentile ranging from the median to the 90th percentile, are presented in Figure S2. When resolved by element (Figure 2a), the largest errors occur on the first-row elements H, F, N, O, and C (0.09–0.11 eV). Across all held-out structures, the predicted band gaps yield a median error of 0.102 eV and a mean error of 0.16 eV (Figure 2b), exhibiting consistent performance across diverse chemical domains. Evaluation on structure types absent from the training set and larger unit cells containing 25–60 sites shows slightly higher errors (Table S1). Overall, held-out and training errors remain closely aligned, with test-to-train error ratios ranging between 0.76 and 1.38 across the various -block groups, confirming that the model avoids overfitting. Despite operating with a highly compact parameter space of 78,964 weights, EpNet achieves a good balance between low model complexity and high predictive accuracy across the periodic table.

Figure 2. Accuracy on chemical systems absent from training. a, Median band error (meV) of the held-out structures that contain each element. b, Band gaps of the 964 gapped held-out structures.
Figure S2. Band structures of held-out compounds. PBE (gray) and EpNet (blue, dashed) for structures whose chemical systems are absent from training, chosen by rank among the held-out structures of up to 3,000 plane waves: a,b, the median and 90th-percentile gapped compounds; c,d, the median gapped and early- compounds; e,f, the median and open-shell structures; g,h, the 75th- and 90th-percentile gapped structures with elements; i, the 90th-percentile metal with elements.
| Set | Structures | Band error (eV) | ≤ 0.1 eV | Gapped | Gap error (eV) |
| Held-out chemical systems | 2,050 | 0.061 | 74% | 964 | 0.102 |
| only elements | 639 | 0.065 | 71% | 466 | 0.084 |
| with elements | 555 | 0.057 | 76% | 226 | 0.109 |
| with early elements | 589 | 0.058 | 76% | 196 | 0.131 |
| with elements | 423 | 0.050 | 81% | 82 | 0.102 |
| with open-shell elements | 177 | 0.056 | 82% | 14 | 0.103 |
| Held-out structure types | 953 | 0.076 | 62% | 377 | 0.110 |
| Sealed chemical systems | 1,210 | 0.061 | 74% | 605 | 0.094 |
| Held-out, 25–60 sites, only | 575 | 0.066 | 69% | 541 | 0.112 |
| Held-out, 25–60 sites, with | 471 | 0.073 | 64% | 350 | 0.136 |
Table S1. Table S1. Accuracy on held-out structures. Band error is the mean absolute error of the band energies within eV of midgap or the Fermi level, median over structures. Gap error is the median absolute error of the band gap over gapped structures. A structure counts in every group whose elements it contains: (Cu, Ag, Au, Zn, Cd, Hg), early (Sc, Y, Ti, Zr, Hf, V, Nb, Ta), (Mo, Tc, Ru, Rh, Pd, W, Re, Os, Ir, Pt), open-shell (Cr, Mn, Fe, Co, Ni).
Learned local potentials of s/p elements. We attribute much of this accuracy to the constraint that the learned potentials remain physical pseudopotentials in our design. Figure 3a shows that the base potentials of the s/p elements stay smooth, finite at as for screened ions and decaying by 8 Å, and differ from the empty cores they started from mainly below 3 Å. The first node of a base potential is the wavevector at which first changes sign, from attractive to repulsive, measuring the size of the ion core. For an empty core of radius , it lies at . Across all 34 s/p elements, the learned first node follows that of the starting empty core, with a Pearson correlation coefficient (Figure 3b). Beyond these base potentials, the chemistry-specific part of the model lies in the environment potential and the charges. The learned environment potential also follows bonding (see Figure 3c). For carbon its spherical part shifts the potential near Å by about eV·Å, a tenth of the base potential there ( eV·Å), in opposite directions for twofold and fourfold coordination. Its anisotropic part is 1.6–1.8 eV·Å for twofold and threefold carbon but almost vanishes (0.13 eV·Å) for fourfold carbon, as it should for tetrahedral sites, whose symmetry forbids dipole and quadrupole terms. Figure 3d shows the learned atomic charges of the held-out s/p compounds against their electronegativity relative to the cell mean. Their median magnitude is 0.10 , far below formal oxidation states, yet they correlate strongly with the atom’s electronegativity (). Switching them off in the trained model shows that on 240 ionic held-out structures the median band error rises from 0.063 to 0.152 eV, 77% of the structures become worse by more than 0.01 eV, and the median gap error of the 167 gapped ones rises from 0.077 to 0.393 eV. In the meantime, the metals change less (0.044 to 0.091 eV). This confirms that the charges are essential for ionic compounds, and that the model has learned to use them correctly.
Figure 3. Learned local potentials of elements. a, Base potentials of first-row, group IV and V and ionic elements (solid) and the screened empty cores they started from (dashed). b, First node of against that of the empty core, with , for the 34 elements. c, Environment potential of the carbon atoms of scored compounds, by coordination number: spherical part (median, solid; interquartile range, shading) and anisotropic part (root mean square over directions, dashed). d, Charges of the atoms of the held-out compounds against their electronegativity relative to the cell mean.
Learned potentials of elements.
For the elements, the essential physics is built in before any learning. The width of a band is set by the overlap of orbitals on neighboring atoms, which lies in the orbital tail (Figure 4a). EpNet therefore builds each channel from the free atom’s orbital and keeps its all-electron tail unchanged; the norm missing inside the core radius is restored by the overlap operator, and the channel has no spurious bound states[57]. Before any training, this construction places the -band widths of seven compounds within 1.13–1.26 times PBE (Supplementary S3, Figure S3 and Table S2).
Figure 4. Learned potentials and channels of elements. a, Cu orbital: all-electron (black), the ultrasoft pseudo-orbital that keeps the all-electron tail (blue) and a norm-conserving pseudo-orbital obtained by renormalizing it (orange); shading: the norm restored by the overlap operator. b, Learned -level shifts: per element (dots) and, with the per-atom shift, the 10th–90th percentiles over the atoms of the scored compounds (bars). c, Environment potential of atoms by coordination: spherical part (median, solid; interquartile range, shading) and anisotropic part (root mean square over directions, dashed). d, Charges of the atoms of held-out compounds against their electronegativity relative to the cell mean.
Figure S3. Building the channel from the atom (full construction tests). a, Cu orbital: all-electron (black), the ultrasoft pseudo-orbital that keeps the all-electron tail (blue) and a norm-conserving pseudo-orbital obtained by renormalizing it (orange). The shading is the norm restored by the overlap operator. b, Cu local potential: the neutral all-electron atom beyond (blue) and the shallow -valence empty core (orange). c, Error of the Cu channel in scattering energy against the all-electron atom, at 2.4 Å, for one and two reference energies. d,e, Zero-shot test of the element at its atomic construction and anion potentials trained on compounds only, for seven compounds and five constructions. d, Band error in the ±2 eV window. e, Width of the band relative to PBE after shifting the level onto the PBE -band center.
| Sc | −4.30 | −1.52 | −3.43 | 1.19 | 1.23 | 1.22 | 0.48 | 1.4 | 6 | 0.088 | 28 | 0.97 |
| Ti | −4.60 | −1.54 | −4.46 | 1.11 | 1.14 | 1.24 | 0.59 | 1.2 | 3 | – | 18 | 0.95 |
| V | −4.85 | −1.53 | −5.38 | 1.06 | 1.07 | 1.27 | 0.68 | 1.2 | 3 | – | 15 | 0.96 |
| Cr | −4.20 | −1.08 | −3.13 | 1.01 | 1.03 | 1.24 | 0.65 | 1.2 | 3 | – | 16 | 0.96 |
| Mn | −5.29 | −1.47 | −7.01 | 0.95 | 0.95 | 1.38 | 0.82 | 1.2 | 3 | – | 9 | 0.98 |
| Fe | −5.49 | −1.44 | −7.75 | 0.90 | 0.90 | 1.38 | 0.86 | 1.2 | 3 | – | 8 | 0.98 |
| Co | −5.68 | −1.39 | −8.46 | 0.87 | 0.87 | 1.38 | 0.88 | 1.2 | 3 | – | 7 | 0.99 |
| Ni | −5.86 | −1.34 | −9.14 | 0.82 | 0.82 | 1.38 | 0.90 | 1.2 | 3 | – | 6 | 0.99 |
| Cu | −4.86 | −0.79 | −5.33 | 0.79 | 0.82 | 1.38 | 0.88 | 1.2 | 3 | – | 6 | 0.99 |
| Zn | −6.22 | −1.23 | −10.43 | 0.77 | 0.77 | 1.38 | 0.93 | 1.2 | 3 | – | 5 | 1.00 |
| Y | −4.23 | −1.50 | −2.65 | 1.36 | 1.49 | 0.77 | 0.02 | 1.8 | 3 | 0.025 | 39 | 1.00 |
| Zr | −4.59 | −1.55 | −3.74 | 1.30 | 1.39 | 1.16 | 0.21 | 1.8 | 3 | – | 33 | 0.99 |
| Nb | −4.20 | −1.22 | −3.21 | 1.23 | 1.33 | 1.19 | 0.27 | 1.4 | 3 | 0.15 | 30 | 0.92 |
| Mo | −4.33 | −1.16 | −3.94 | 1.17 | 1.26 | 1.22 | 0.37 | 1.4 | 3 | 0.15 | 28 | 0.92 |
| Tc | −5.26 | −1.52 | −6.81 | 1.11 | 1.18 | 1.27 | 0.52 | 1.2 | 3 | – | 22 | 0.93 |
| Ru | −4.52 | −1.04 | −5.40 | 1.08 | 1.15 | 1.27 | 0.55 | 1.2 | 3 | – | 21 | 0.93 |
| Rh | −4.59 | −0.97 | −6.15 | 1.06 | 1.10 | 1.32 | 0.64 | 1.2 | 3 | – | 16 | 0.94 |
| Pd | −3.62 | −0.36 | −4.31 | 1.03 | 1.08 | 1.30 | 0.64 | 1.2 | 3 | – | 17 | 0.94 |
| Ag | −4.72 | −0.85 | −7.66 | 0.98 | 1.03 | 1.38 | 0.76 | 1.2 | 3 | – | 13 | 0.95 |
| Cd | −5.96 | −1.33 | −11.90 | 0.93 | 0.98 | 1.38 | 0.83 | 1.2 | 3 | – | 12 | 0.95 |
| Hf | −5.26 | −1.47 | −2.88 | 1.18 | 1.35 | 1.14 | 0.15 | 1.6 | 3 | 0.15 | 36 | 0.98 |
| Ta | −5.62 | −1.48 | −3.79 | 1.16 | 1.29 | 1.19 | 0.23 | 1.4 | 3 | 0.15 | 30 | 0.92 |
| W | −5.91 | −1.47 | −4.69 | 1.11 | 1.26 | 1.22 | 0.31 | 1.2 | 3 | 0.15 | 28 | 0.93 |
| Re | −6.16 | −1.44 | −5.58 | 1.08 | 1.20 | 1.24 | 0.39 | 1.2 | 3 | – | 26 | 0.92 |
| Os | −6.38 | −1.40 | −6.47 | 1.06 | 1.17 | 1.27 | 0.47 | 1.2 | 3 | – | 23 | 0.92 |
| Ir | −6.59 | −1.35 | −7.36 | 1.01 | 1.12 | 1.30 | 0.54 | 1.2 | 3 | – | 21 | 0.92 |
| Pt | −5.96 | −0.96 | −6.38 | 1.01 | 1.10 | 1.30 | 0.56 | 1.2 | 3 | – | 20 | 0.93 |
| Au | −6.08 | −0.89 | −7.12 | 0.98 | 1.08 | 1.38 | 0.67 | 1.2 | 3 | – | 16 | 0.93 |
| Hg | −7.13 | −1.19 | −10.08 | 0.93 | 1.03 | 1.38 | 0.73 | 1.2 | 3 | – | 16 | 0.94 |
Table S2. Atomic channels of the 29 elements (the tables the model uses; built from the neutral, scalar-relativistic all-electron atom in the local-density approximation, as in equations (S4)–(S9)). , , : reference energies of the , and channels, the all-electron levels of the atom; : pseudization radii; : norm deficit of the pseudo-orbital, restored by the overlap operator; : radius beyond which the local potential is the atom’s own; : second reference energy above ; : bound of the learned orbital stretch , ; –: not learned (); : error of the pseudo-atom’s level at 20 Ry against the all-electron one; : smallest eigenvalue of the pseudo-atom’s overlap.
After model training, the corrections learned on top of this construction are small, and most of them follow physical trends. The base potentials stay within 0.2% (root mean square) of their atomic tables, and the orbital stretch within 1% of the atomic orbital. The learned -level shifts (Figure 4b) are a few tenths of an eV, the size of the correction between a free neutral atom and a ion screened in a crystal. They sort by shell filling. Per element, they are slightly negative for the filled shells (median −0.13 eV) and positive for the open-shell and the elements (medians +0.17 and +0.16 eV). Per atom, they are largest on the most cationic sites. The environment potential (Figure 4c) is more repulsive at more crowded sites, as the overlap of neighboring charge densities makes it: its spherical part near Å rises from 11 eV·Å at fourfold sites to 20–23, 29 and 37 eV·Å at six-, eight- and twelvefold sites. Its anisotropic part obeys the crystal-field selection rules: it is small at tetrahedral and octahedral sites (0.9–1.1 eV·Å), whose cubic symmetry forbids dipole and quadrupole terms, and larger at trigonal prisms (1.6 eV·Å), where a quadrupole term splits from the in-plane and orbitals. However, the charges of the atoms are not physical (Figure 4d): 85% of those in held-out compounds are negative (median −0.12 ), and they correlate only weakly with electronegativity (, against −0.91 for atoms). A negative charge raises the potential of the atom, and the model appears to use it, with the mostly positive -level shifts, to lift a level that sits too deep; constraining the charges to be non-negative lost the fit of the gaps, so we read them as effective parameters.
Application to large-scale materials screening. We next apply EpNet to the large-scale Alexandria PBE database[9], from which we extracted 600,617 non-magnetic structures that are completely absent from the Materials Project, span the 63 elements, contain up to 24 sites and lie within 0.1 eV/atom of the convex hull. Since Alexandria gives band gaps but not band structures, EpNet scored all of them based on the gap values and the metal/insulator classification. For a fair comparison with other methods, we also drew a sample of 5,000, 1,000 per group to score tight-binding models below.
Figure 5a shows that over all 600,617 structures EpNet’s median gap error is 0.108 eV over the 124,416 gapped ones, 98.9% are correctly classified as metal or insulator, and the distribution of gaps matches Alexandria’s. On structures present in both databases, the Alexandria and Materials Project gaps differ by a median of only 0.03 eV, so Alexandria is a fair external test for a model trained on Materials Project labels. Figure 5b,Figure 5c compares EpNet with existing models on the same structures. Two graph networks trained on Materials Project PBE gaps, ALIGNN[14] and MEGNet[13], are fast to run on the whole pool, but they only yield the scalar gap values and their median gap errors are 0.279 and 0.555 eV, 1.8–3.0 and 3.1–5.7 times EpNet’s in the five groups, and they classify many metals as insulators (89% and 88%, against 99% for EpNet). Two tight-binding models also return full band structures. On the 5,000-structure sample their median gap errors are 0.282 eV (DFTB+[58] with the PTBP parameters[29]), against 0.105 eV for EpNet on the same structures, and 0.424 eV (ThreeBodyTB[30], on the 1,597 structures within its fit), against 0.091 eV. Hence, EpNet is the most accurate model in every group.

Figure 5. Benchmarking unseen Alexandria materials and comparison with other models. Alexandria structures absent from the Materials Project: all 600,617 in a; in b and c, all of them for EpNet, ALIGNN and MEGNet and the 5,000 test structures for the two tight-binding models marked with † (ThreeBodyTB on the 1,597 within its fit). a, EpNet (screening mode) against Alexandria band gaps, structures per bin. b, Median gap error of the gapped structures, by group. c, Mean gap error over all structures against time per structure on 8 CPU cores; labels, metal/insulator agreement; open marker, EpNet on one GPU.
In addition to its predictive accuracy, EpNet is compelling due to its low computational cost (Figure 5c). For gapped systems, EpNet requires approximately 12 s per structure on 8 CPU cores, compared to 4 s and 12 s for tight-binding models on identical CPU hardware, and roughly 0.1 s for graph neural networks, which yield only the band gap. However, GPU acceleration significantly reduces this runtime. Using EpNet, screening 614,729 structures of the Alexandria database required 18.5 h across 16 GPUs, a median of 0.9 s per structure on a single GPU. We further benchmarked EpNet’s computational cost against standard DFT solvers using two distinct test sets (Figure S4 and Figure S5). On 24 held-out structures containing 2–24 atoms, computing a converged PBE band structure using Quantum ESPRESSO[59, 60] with SSSP pseudopotentials[61] required 3–984 s. This corresponds to a median wall-clock time 99 times longer than EpNet on a single GPU, and 12 times longer than EpNet on the same 16 CPU cores (Figure S4). The speedup grows with cell size up to 24 atoms (Figure S4); for larger cells it becomes a constant factor, as both codes scale as – (Figure S5). On FCC supercells of Al and Cu containing more than 64 atoms, EpNet on a single GPU is 110–200 times faster for Al and 420–470 times faster for Cu per k-point than Quantum ESPRESSO on 16 CPU cores, requiring only 10 s and 82 s per k-point for 512-atom supercells of Al and Cu, respectively. EpNet’s efficiency stems from bypassing the self-consistent field cycle, employing a significantly smaller plane-wave basis set (a 12–20 Ry kinetic energy cutoff compared to 35–90 Ry in standard DFT), and leveraging full GPU parallelization. Table S3 also provides a detailed comparison of EpNet against several other machine-learning and semi-empirical electronic structure methods.
Figure S4. Band-structure cost against cell size, Quantum ESPRESSO and EpNet. 24 held-out structures of 2–24 atoms (filled, only elements; open, with elements), with the same cells and points: QE 7.3 and EpNet on one NVIDIA H100 and on the same 16 cores. a, Wall time per band structure. b, Wall time per point. Lines, power-law fits , exponents in the legend.
Figure S5. Cost per k point against cell size on FCC supercells. supercells (–512 atoms) of the primitive cells of a, Al and b, Cu, each with the same eight points, so that only the cell size changes: QE (bands run after its self-consistent field) on 16 Intel Xeon Gold 6248R cores and EpNet on the same 16 cores and on one NVIDIA H100. QE runs of Al with 512 atoms and of Cu with 216 or more atoms are omitted, so the QE fit for Cu uses two cells (64 and 125 atoms). Lines, power-law fits over ; dotted, through the largest QE cell.
| Approach and models | Input | Output | Training labels | Reported accuracy |
| Property models: ALIGNN[14] MEGNet[12, 13] coGN[15] | structure | band gap or density of states | gaps or densities of states | gap, mean error 0.16 eV (coGN on Matbench[16], PBE) |
| Universal learned DFT Hamiltonians: DeepH-UMM[23], UniHamGNN[24], NextHAM[25] | structure, plus the DFT code’s basis and overlap matrices or a starting Hamiltonian | Hamiltonian in an atomic-orbital basis | self-consistent Hamiltonian matrices | matrix elements, 1.3–2.2 meV |
| Tight binding: DFTB+[58] with PTBP[29], ThreeBodyTB[30] g-xTB[31] | structure | Hamiltonian in a minimal basis | DFT energies and band structures | gap, mean error 0.67–0.82 eV against DFT[32] |
| Learned EPMs: Kim and Son[36] Kang et al.[37] Lin et al.[38] | structure | plane-wave Hamiltonian | DFT potentials, or band energies (DFT or GW) | single elements or material families |
| EpNet (this work) | structure | plane-wave Hamiltonian | band energies | held-out chemical systems: band error 0.061 eV, gap 0.102 eV |
Table S3. Comparison of selected existing machine learning models for electronic structure calculations.
Conclusions
We have demonstrated that a physics-based empirical pseudopotential network with 78,964 parameters captures the electronic band structures of crystals across the , , and blocks of the periodic table. The central innovation is a Hamiltonian decomposed into distinct local and nonlocal operators. In this design, a universal local potential for all elements is learned from atomic features and local structural environments, while the nonlocal channels use learned projectors for the extended and states and projectors built from the free atom for the localized states. On chemical systems absent from the training set, EpNet, without self-consistent iterations or reliance on a DFT solver, predicts complete band structures directly from crystal coordinates with median errors of 0.061 eV for band energies and 0.102 eV for band gaps. Crucially, benchmarks on an independent database show band-gap errors 2–5 times smaller than those of direct property-prediction networks and tight-binding models.
The remaining errors are concentrated within a few specific chemical regimes (Supplementary § 16 and Figure S6). Furthermore, the model currently inherits the domain limits of its training labels. The present model is limited to PBE band energies of non-magnetic crystals without spin–orbit coupling. However, because EpNet relies solely on band energies for training, it can be readily extended or fine-tuned on hybrid-functional[26, 62], [38, 63], or experimental[35] band data without requiring full Hamiltonian matrices. Looking forward, -block elements can be incorporated using the same atomic construction established for channels, and the underlying Hamiltonian formulation can be extended to enable application to absolute band alignment and electron–phonon coupling calculations[38, 64, 65, 66, 67].
Figure S6. Two limits of the model, on held-out chemical systems. a, in the fluorite structure (mp-1008677, , Ti eightfold), conventional cell. b, Its band structure, EpNet (blue, dashed) against PBE (gray), aligned at the valence-band maximum. c, (mp-1023928), a bilayer, viewed edge-on to the layers over two cells along the layer normal, with the vacuum between periodic images and the van der Waals gap marked. d, Its band structure, as in b; arrows, the direct gap at .
Methods
Hamiltonian and basis
The basis is centered on each , , and is folded into the first Brillouin zone before is built[39]. The kinetic-energy cutoff depends on the element, 12 Ry by default. Compact states need more plane waves than heavier states, so we used 16 Ry for B and C and 20 Ry for N, O, F and all elements. Like other plane-wave codes, each cell uses the largest cutoff of its elements. In the first stage of training, on the compounds, a basis-consistency term, the mean absolute difference of the bands at four points between the cell’s cutoff and 20 Ry after removing a rigid shift, was added to the loss with weight 1 for one cell per GPU and step, so that the learned potential does not exploit the truncation of the basis. The component of the local potential is set to zero, which fixes the energy reference. Band energies are therefore compared after alignment. For cells with atoms the generalized problem is solved through the Cholesky factor of on the dense path, and through the exact low-rank (Löwdin form[68]) on the iterative path.
Specie potentials
Base potentials. Each element is described by 15 features: a constant, , period/7, group/18, the , and valence occupancies, one-hot , and block labels, the Pauling electronegativity , the covalent radius[69] Å, , the model valence and the number of model electrons, which separates valence shells (Zn–Hg) from the filled semicore shells of Ga–Bi. A hypernetwork, a multilayer perceptron with one hidden layer of 32 units with sigmoid linear unit (SiLU) activations, maps these features to the specie projector strengths and to the coefficients of the specie potential,
with in Å. The second sum, a high- tail, is needed because the 20 Ry basis couples plane waves up to Å. The coefficients are the sum of the hypernetwork output, a small per-element residual with an penalty () and a fixed prior. For elements, the prior is a Thomas–Fermi-screened Ashcroft empty-core pseudopotential[48]. It is set by the element’s valence and its core radius , a fixed multiple of the covalent radius , since pseudopotential core radii scale linearly with atomic radii[70]. Because the prior depends only on element features, it is defined for every element. For elements it is the tabulated all-electron local potential described below, and the learned deviation from it enters scaled by 0.01. The hypernetwork’s output layer and the residuals start at zero.
Environment terms and charges. We extended the environment encoder from the previous EpNet-Si[39] to many species, using neighbor moments up to within 6 Å under periodic boundary conditions, 8 radial functions, a single gated pass, and 64 channels of invariant (), vector () and traceless rank-2 () features. The species embedding is a learned linear map of the element features plus a penalized per-species residual ( weight ), so an element enters through its features rather than through an independent vector. The norms of and that feed the gates are smoothed as with , which removes the cusp of at high-symmetry sites, where , and keeps second derivatives of the bands with respect to atomic displacements well defined. Four readouts map the encoder features to form factors with components and the phases that keep , with an additional 12-term cosine basis in the monopole channel. A router mixes them with per-atom softmax weights from a multilayer perceptron (one hidden layer of 16 units) on eight standardized inputs: the logarithm of the local valence-electron density at , 3, 4, 5, 6, 8 and 10 Å, normalized by e/Å, where is a smooth neighbor count, and the mean electronegativity difference to the neighbors within 4 Å. A linear head on the invariant features gives raw charges whose cell mean is subtracted, so that holds exactly, while the charges enter only through the Coulomb potential of equation (2) ( Å), and the head starts at zero.
Nonlocal projectors
s/p elements. The nonlocal term of equation (1) is , and the first term is
with , and Gaussian radial projectors carrying the angular factor of channel [49]. The two radial functions of channel have real-space radii of 1.2 and 0.7 times , the peak radius of the element’s valence or orbital in the neutral all-electron atom computed in the local-density approximation (LDA)[71], multiplied by a learned global factor per radial function and a learned factor per element and channel. First-row atoms (B–F) have and projectors only, because a -like channel at their compact radius produced spurious states in development; for Ca, Sr, Ba and Cs the channel sits at the radius of the lowest empty orbital, which forms their conduction-band edge, and for the other elements at the radius. The specie strengths come from the hypernetwork and the per-atom corrections from a zero-initialized multilayer perceptron (one hidden layer of 16 units) on the 192 encoder invariants and two coordination numbers; one strength per multiplet keeps the projector rotationally invariant.
elements. Unlike elements, each element is solved as a neutral, spherical, scalar-relativistic (zeroth-order regular approximation)[50] LDA atom in its experimental ground-state configuration; valence eigenvalues agree with the NIST scalar-relativistic LDA tables [72, 73] to within 0.041 eV. Inside the orbital is , matched to the all-electron orbital in value, slope and curvature; is the smallest radius for which the plane-wave components above 20 Ry carry less than 0.05 eV of kinetic energy (1.14–1.38 Å for , 0.77 Å for Y; Table S2). The orbital is not renormalized; the norm deficits range from 0.02 (Y) to 0.93 (Zn). Beyond (1.2–1.8 Å) the local potential equals the neutral atom’s self-consistent potential and inside it is a smooth polynomial matched in value, slope and curvature. The neutral-atom potential extends to 5–6 Å, farther than the cosine basis of the learned potentials can represent. It therefore enters as a fixed table of (1,601 points up to 24.7 Å), which reproduces the potential in real space to 0.003 eV. The radius is the smallest at which every channel passes the checks described below. The and channels use one reference energy and the channel two, and eV (+6 eV for Sc), the second from the all-electron solution at that energy pseudized at the same (Supplementary equations S7–S9). Several choices keep the construction numerically stable. The second orbital is normalized within the core, and is replaced by its positive-semidefinite part, a change below 0.003. The learned -level shifts enter through rather than through , so that they cannot drive toward singularity. The orbital stretch is learned only for the seven elements whose operator norms stay within three times their atomic values over the allowed range (Sc, Y, Nb, Mo, Hf, Ta, W). The channel lies close to its Kleinman–Bylander singular point and is not learned. The pseudo-atom of every element, solved at 20 Ry, is tested in five ways: (i) it reproduces the all-electron levels within 0.04 eV; (ii) its overlap is positive definite (); (iii) no channel has a bound state below its reference level; (iv) its tail is within 2% of the all-electron orbital (at 30 Ry); and (v) its and scattering errors are 0.01–0.20 eV from 4 eV below to 6 eV above the reference levels, except for the diffuse channels of Y and Zr (0.23–0.53 eV). We accept these two exceptions because the simpler local potential they would otherwise use gives far larger errors in crystals. The channels enter the projector slots of atoms; the Gaussian projector used for atoms is switched off for them, since in single-crystal fits it allowed the fit to expel the state. The augmentation term adds to , with the cell’s local potential, minus the atom’s own, averaged over the augmentation density . A per-atom -level shift comes from a small network (one hidden layer of 16 units, zero-initialized output) on the atom’s invariant features.
Valence electron condition, data and splits
The model describes valence electrons only: counts the valence and electrons of elements and the electrons of elements (for example Ti: 4, Fe: 8, Pt: 10). For each reference structure, the bands of the semicore shells that the Materials Project datasets treat as valence are counted from their valences and removed from the bottom of the spectrum. Structures with a non-integer count, or with a removed band within 4 eV of the band edge or Fermi level, are rejected. In the retained structures, every removed band lies at least 4.5 eV below the band edge or Fermi level, so the bands near the gap are untouched. Every gapped structure confirmed the valence table. The data are the PBE line-mode band structures of the Materials Project computed with VASP[74] along standard paths [75]. We exclude spin-polarized or calculations and the block systems, which leaves 22,553 structures of up to 24 sites (consisting of 8,464 with only elements and 14,089 with elements). Structures are assigned to splits by hashing first their chemical system and then their prototype. The structures were assigned first, and their splits were kept when the structures were added. Within the training split, 20% of the chemical systems are held out from the loss as an inner validation set and 10% as a checkpoint-selection set.
Loss, eigensolver and training
The loss of a structure is the mean absolute error of its band energies within of midgap (gapped structures, aligned at the valence-band maximum) or of the Fermi level (metals), with the band-edge bands weighted by 5, plus the absolute error of the predicted gap and an all-occupied-band term with weight 0.3. For gapped structures with open-shell atoms the labels include the lowest conduction bands, the empty manifold, and the window extends to their top. Only the lowest bands are computed: each structure keeps its eigenvectors from the previous visit, which seed Chebyshev-filtered subspace iteration[76], and above 3,000 plane waves the Hamiltonian is never formed, the local potential being applied on a fast Fourier transform (FFT) grid and the nonlocal and overlap terms in low-rank form. Gradients with respect to a model parameter follow from the Hellmann–Feynman theorem[77] in its generalized form,
Training uses AdamW[78] without weight decay, batches of 16 structures and a cosine learning-rate decay, with separate rates for the species parameters. Training began with the compounds: 17,800 steps from scratch on structures of up to 24 sites. Three warm-start stages followed (1,500, 2,500 and 2,500 steps), which added a high- tail to the specie potentials and 596 structures of 25–60 sites rich in N and C–H. Throughout these stages, a penalty kept each cell’s bands nearly unchanged when its basis was raised to 20 Ry. The elements were then added to the species table and trained in stages, each starting from the previous one and each with criteria fixed before it ran (Supplementary Information). First, cells of up to 3 sites introduced the groups one at a time: , early , and open-shell elements, each stage with cells of up to 8 sites as anchors (3,000 steps per stage, learning rate ). Next, all cells of up to 8 sites were trained for 22,000 steps (learning rates , then ). Then all 12,136 training cells of up to 24 sites and 8,700 plane waves were trained for 11,500 steps (learning rate ), with the new cells above 8 sites weighted by 4. A final 22,000 steps on the same cells used half the learning rate, with the cells weighted by 1.5 and the new cells by 1. The basis-consistency penalty of the first stage was not used in these stages. Each structure contributes a subset of its path that always includes the path vertices and the band extrema (median 48 k points for cells of up to 8 sites and 33 for larger cells; 32 and 16 in the final stage). The final stage took 17 h on four NVIDIA H100 GPUs for one model and 35 h on four L40S GPUs for the other.
Evaluation
Structures are scored on low-resolution band paths with 64 k points. The band error is the mean absolute error within of midgap (gapped) or the Fermi level (metals, after the median shift within that window); the gap error is the difference of the path gaps. Alexandria PBE (release 2025.07.02, 5.78 million entries) was filtered to the training scope, which leaves 615,564 structures, of which 600,617 match no in-scope Materials Project structure according to pymatgen’s structure matcher [79]. From these, 1,000 per group were drawn, and the sample and metrics were fixed before scoring. EpNet’s gap is computed over a standard band path [75] and a -centered mesh with 0.12 spacing, while a structure is metallic when the gap closes or the electron count is odd.
Setting of other models: For the benchmark of Alexandria data, we also employed (i) MEGNet; (ii) ALIGNN models trained on Materials Project PBE gaps; (iii) DFTB+ 25.1 with the complete PTBP parameter set; and (iv) ThreeBodyTB 1.1.1 (fitted to PBEsol). These models ran on 8 CPU cores (ALIGNN and MEGNet in batches, their total time divided by the number of structures), while EpNet was run on both 8 CPU cores and one GPU. The comparison with the plane-wave DFT code is described in Supplementary Section 4 (Figure S4 and Figure S5).
EpNet band structure screen mode. For screening, EpNet takes the gap from the band path alone, thinned to every fourth point (every third for the first 28,000 structures of the screen) with all vertices kept, and computes the bands at successive points by Chebyshev-filtered subspace iteration started from the previous point’s eigenvectors with a tolerance of 5 meV. The screen covered all in-scope structures except the 835 that match sealed Materials Project structures, 614,729 in total, in 285 GPU-hours (18.5 h on up to 16 GPUs: A40, A100 and H100).
Data and code availability
The code is available at https://github.com/MaterSim/EpNet-spd under the MIT license, together with the trained models (both seeds and every checkpoint of the training curriculum), the recorded arguments of every training stage, the training manifest with its splits, the atomic channels and the per-structure scores of the final model. The training labels are Materials Project [3] band structures, which the repository’s scripts download, and the external test uses the Alexandria database [9]. The web application at https://mmi.charlotte.edu/epnet gives access to the final model.
Acknowledgments
This research was sponsored by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, and the Established Program to Stimulate Competitive Research (EPSCoR) under the DOE Early Career Award No. DE-SC0024866. Computations were performed on the resources of University Research Computing at the University of North Carolina at Charlotte.
Competing interests
The author declares no competing interests.
Supplementary Information
The network, the Hamiltonian and the training loss
Figure S1 shows the complete model from the crystal structure to the band energies, including the element features, the router and the equivariant encoder, the operators they produce, and the assembly of the plane-wave Hamiltonian.
Training loss. The model is trained on the PBE band energies of each structure at the points of its band path; are the predicted energies, is the highest occupied band of the valence electron count and . For a gapped structure the predicted bands are aligned at the valence-band maximum, with , and the loss is
The window holds the states with , being the midgap, together with the two band-edge bands at every , so that the edges are fitted whatever the gap; for and 1 otherwise. is the signed predicted gap, negative when the bands overlap, so that a closed gap still has a gradient; and . For gapped structures with open-shell atoms, whose labels include the lowest conduction bands, is instead with , the top of the empty manifold. For a metal the window holds the states within 2 eV of the PBE Fermi level, the prediction is aligned by the rigid shift , which is not differentiated, and there is no gap term:
Each step draws a batch of training structures without replacement, with probability proportional to , where is the number of training structures in the chemical system of , so that large families do not dominate, and is the stage weight of Methods (4 for the new cells above 8 sites in the 24-site stage, 1.5 for the s/p cells in the final stage, 1 otherwise), and minimizes
where are the per-element residuals of the specie-potential coefficients () and the per-species residuals of the encoder’s species embedding (). In the first stage, on the s/p compounds, a basis-consistency term , with at four points , was added with weight 1 for one structure per GPU and step. Gradients of follow from the generalized Hellmann–Feynman theorem.
Accuracy and band structures of held-out structures
Table S1 gives the accuracy of the final model in three major groups, including (i) held-out structure types, (ii) the sealed chemical systems and (iii) cells of 25–60 sites. By group the median gap error is 0.109 eV (), 0.131 eV (early ), 0.102 eV (4/5), 0.103 eV (open-shell 3, 14 gapped structures) and 0.084 eV (s/p). Restricted to compounds within 0.1 eV/atom of the convex hull, the median gap errors are 0.105, 0.122 and 0.100 eV for the three larger groups. Structure types absent from training are harder (0.076 eV), and the difference is concentrated in one family. Held-out structure types containing 4/5 elements are dominated by metastable mixed Mo/W chalcogenide stacks (141 of 314 structures, mostly 0.07–0.23 eV/atom above the hull), with a median band error of 0.22 eV; for the structures within 0.05 eV/atom of the hull the error is 0.067 eV. The sealed chemical systems, set aside before training and scored once with the final model, confirm these numbers: over 1,210 structures the median band error is 0.061 eV (74% within 0.1 eV), between 0.049 and 0.068 eV in every group, and the gap error 0.094 eV over 605 gapped structures. Larger cells of 25–60 sites and up to about 25,000 plane waves, beyond the 24 sites and 8,700 plane waves of training, are somewhat harder. For these cases, the band and gap errors are 0.066 and 0.112 eV for 575 held-out s/p structures and 0.073 and 0.136 eV for 471 with elements, the gap error rising most for 4/5 compounds (0.161 eV). This suggests that the model may not have fully learned either the atomic environment or long-range interactions of the larger cells. To further boost the model’s performance, one could consider including more large-cell structures in the training set or employing data augmentation techniques to better capture the physics of these systems.
Figure S2 shows band structures of held-out compounds, including the median and 90th-percentile s/p structures, the median structure of each group, and the 75th- and 90th-percentile structures.
Atomic construction of channels
We first tested whether the operator used for the elements, with its Kleinman–Bylander projector placed at the -orbital radius, can represent -electron band structures at all. Fitted to each of 45 small crystals spanning the block separately, it reached 0.1 eV on at most 12 of them. This motivated us to consider the atomic construction of orbitals.
Figure S3 shows the tests of that construction. For seven compounds we placed the element at its atomic construction, with no environment terms, next to anion potentials trained on compounds only (Figure S3,Figure S3). A norm-conserving orbital, renormalized inside the core radius, scales up the all-electron tail (2.9 times for Cu, see Figure S3), and the bands come out 3.7–10.6 times too wide even after their center is placed correctly (Figure S3). Keeping the tail and restoring the norm with the overlap operator reduces the – overlap to its physical size. Two further ingredients are each needed: the atom’s own local potential (Figure S3) and a second reference energy (Figure S3). With all three, the zero-shot -band widths come within 1.13–1.26 times PBE for all seven compounds, with band errors of 0.05–0.35 eV before any training. For all 29 elements the constructed pseudo-atom reproduces the all-electron level within 0.04 eV, with a positive-definite overlap and no spurious bound states[57]. Table S2 lists the parameters of the channels of every element, and equations (S4)-(S9) below show how they enter. On 12 of these crystals, fitted one at a time, the atomic channel reached 0.1 eV on 9, against 5 for the learned projector; trained jointly on compounds, it transferred to held-out chemical systems with a median band error of 0.14 eV, against 0.87 eV for the learned projector.
Construction of the atomic channels. The parameters of Table S2 enter the channels as follows. The all-electron atom gives the valence levels (, and ), which are the reference energies, and the radial orbitals , normalized to one. Inside a core radius each orbital is replaced by a smooth, nodeless polynomial,
with , and fixed by the value, slope and curvature of at ; is the smallest radius at which is nodeless and its plane-wave components above 20 Ry carry less than 0.05 eV of kinetic energy. The pseudo-orbital is not renormalized: beyond it is the all-electron orbital, and inside it lacks the norm
which is for the channel (0.01–0.02 for and ). The local potential is the atom’s self-consistent potential beyond and an even polynomial inside,
matched in value, slope and curvature at ; it is the fixed table in the element’s (Methods), to which the learned part is added. The channel has a second reference: the all-electron solution regular at the origin at , pseudized by equation (S4) at the same . With labeling the two references and and all integrals taken within Å,
where and (in the model is replaced by its positive-semidefinite part, a change below 0.003). The projectors are dual to the pseudo-orbitals, , so the channel and its overlap,
give and . At both reference energies, the pseudo-atom therefore has the all-electron level, the all-electron orbital beyond and the all-electron norm. The and channels are the one-reference case, with , and .
In the crystal, with the spherical harmonics is transformed to and placed on every atom, and equations (S7)–(S9) are re-evaluated with the current at every training step, so that they hold throughout. Two corrections are learned on top of them: (i) a -level shift per element and per atom, applied as so that it does not enter ; and (ii), for the seven elements with a value of , a stretch of the orbitals, with and learned, so that . In addition, the augmentation term is added; it has no parameters of its own. and were set for conditioning before any training: is the offset among 3, 4, 5 and 6 eV that minimizes the largest eigenvalue of , with , and the largest stretch (in steps of 0.0125, up to 0.15) over which the largest eigenvalues of and stay within three times their atomic values; below 0.025 the stretch is fixed. The last two columns of Table S2 are checks, not inputs: the pseudo-atom built from equations (S4)–(S9) is solved in a 20 Ry plane-wave basis, is the error of its level against the all-electron one, and is the smallest eigenvalue of its overlap matrix, which must be positive for to be well posed.
Cost against cell size
We selected 24 held-out structures, 4 per size range (1–2, 3–4, 5–8, 9–12, 13–16 and 17–24 sites), half of them with only elements. They were computed with QE 7.3 on their standardized primitive cells, using PBE, SSSP 1.3 efficiency pseudopotentials with their recommended cutoffs and Marzari–Vanderbilt smearing of 0.01 Ry. Each calculation is a self-consistent run on a -centered grid with 0.25 Å spacing followed by a bands run on EpNet’s path points, with 16 MPI processes in 4 pools on one Intel Xeon Gold 6248R node. The resulting gaps agree with the Materials Project’s to a median of 0.018 eV. EpNet solves the same cells and points, on one NVIDIA H100 and on 16 cores of the same node type. On the CPU, the path is split into four contiguous parts run by 4 processes of 4 threads, as QE’s pools split the points; this is a median 1.5 times faster than one process of 16 threads. EpNet times include setup and are taken after a first warm-up evaluation. Figure S4 compares the wall time of a band structure in QE and EpNet for 24 held-out structures of 2–24 atoms with the same cells and points. Overall, EpNet on one GPU is 1–2 orders of magnitude faster than QE, and the speedup grows with cell size up to 24 atoms; for the 16–24-atom cells EpNet takes 4–7 s per band structure and QE 345–984 s.
Figure S5 isolates the dependence on cell size with FCC supercells of Al and Cu of 1–512 atoms that share the same eight points. For the scaling test (Figure S5), we built supercells (–8) of the Materials Project primitive cells of Al and Cu (mp-134, mp-30). Every supercell was given the same eight Cartesian points, and seven steps of 0.05 Å along a reciprocal-lattice vector of the primitive cell, and was computed with the same settings. QE runs of Al with 512 atoms and of Cu with 216 or more atoms are omitted due to high costs. Time per point is the bands-run time (QE) or the total time (EpNet) divided by eight. Above 64 atoms, the time per point of both codes grows as –, so EpNet’s advantage is a constant factor rather than a lower exponent. Overall, EpNet on the H100 is 2–3 orders of magnitude faster than QE on 16 cores for large cells.
Comparison with alternative models
Limitations of the current EpNet model
Among the elements the first-row elements H, F, N, O and C have the largest held-out errors (main-text Figure 2a), and on Alexandria the gaps of structures with hydrogen are the least accurate (0.19 eV), the worst being complex hydrides of metals.
Among the elements, the largest gap errors are from oxides and fluorides of high coordination, such as (Figure S6,Figure S6). In these structures, the level does not drop as far as in PBE, and the gap is too large: in fluorite (0.31 eV/atom above the hull) the O valence bands are reproduced to 0.03 eV, but the empty Ti bands lie 0.76 eV too high, and the gap is 1.78 eV against 1.06 eV. These polymorphs lie 0.3–0.9 eV/atom above the hull and have no counterpart in the training data.
Another limit is the layered Mo/W chalcogenide stacks. Almost all of these cells (663 of the 668 Mo/W chalcogenides of space group ) are bilayer or trilayer slabs, separated from their periodic images by about 13 Å of vacuum. In the held-out bilayer of Figure S6,Figure S6, the direct gap at is 0.64 eV too small, because the lowest conduction band there is 0.61 eV too low, while the direct gap at K is 0.15 eV too small and the fundamental gap is 1.05 eV against 1.28 eV.
The first two groups’ errors are likely to be fixable by including representative structures in the training set, while the last group may require a separate treatment of van der Waals interactions for the slabs. Finally, the charges of the atoms are effective parameters rather than physical charges (main text, Figure 4d). If the charges need to be physical for further applications (e.g., the electron density extraction from learned Hamiltonian), the models need to be retrained.
References
- [1]P. Hohenberg and W. Kohn. Inhomogeneous electron gas. Phys. Rev., 136:B864, 1964. doi: 10.1103/PhysRev.136.B864.DOI
- [2]W. Kohn and L. J. Sham. Self-consistent equations including exchange and correlation effects. Phys. Rev., 140:A1133, 1965. doi: 10.1103/PhysRev.140.A1133.DOI
- [3]Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, and Kristin A. Persson. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Mater., 1:011002, 2013. doi: 10.1063/1.4812323.DOI
- [4]Stefano Curtarolo et al. AFLOW: An automatic framework for high-throughput materials discovery. Comput. Mater. Sci., 58:218, 2012. doi: 10.1016/j.commatsci.2012.02.005.DOI
- [5]Scott Kirklin, James E. Saal, Bryce Meredig, Alex Thompson, Jeff W. Doak, Muratahan Aykol, Stephan Rühl, and Chris Wolverton. The open quantum materials database (OQMD): assessing the accuracy of DFT formation energies. npj Comput. Mater., 1:15010, 2015. doi: 10.1038/npjcompumats.2015.10.DOI
- [6]Kamal Choudhary et al. The joint automated repository for various integrated simulations (JARVIS) for data-driven materials design. npj Comput. Mater., 6:173, 2020. doi: 10.1038/s41524-020-00440-1.DOI
- [7]Sten Haastrup et al. The Computational 2D Materials Database: high-throughput modeling and discovery of atomically thin crystals. 2D Mater., 5:042002, 2018. doi: 10.1088/2053-1583/aacfc1.DOI
- [8]Claudia Draxl and Matthias Scheffler. The NOMAD laboratory: from data sharing to artificial intelligence. J. Phys. Mater., 2:036001, 2019. doi: 10.1088/2515-7639/ab13bb.DOI
- [9]Jonathan Schmidt, Noah Hoffmann, Hai-Chen Wang, Pedro Berlido, Pedro J. M. A. Carriço, Tiago F. T. Cerqueira, Silvana Botti, and Miguel A. L. Marques. Machine-learning-assisted determination of the global zero-temperature phase diagram of materials. Adv. Mater., 35:2210788, 2023. doi: 10.1002/adma.202210788.DOI
- [10]H. J. Kulik et al. Roadmap on machine learning in electronic structure. Electron. Struct., 4:023004, 2022. doi: 10.1088/2516-1075/ac572f.DOI
- [11]Tian Xie and Jeffrey C. Grossman. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Phys. Rev. Lett., 120:145301, 2018. doi: 10.1103/physrevlett.120.145301.DOI
- [12]Chi Chen, Weike Ye, Yunxing Zuo, Chen Zheng, and Shyue Ping Ong. Graph networks as a universal machine learning framework for molecules and crystals. Chem. Mater., 31:3564, 2019. doi: 10.1021/acs.chemmater.9b01294.DOI
- [13]C. Chen, Y. Zuo, W. Ye, X. Li, and S. P. Ong. Learning properties of ordered and disordered materials from multi-fidelity data. Nat. Comput. Sci., 1:46–53, 2021. doi: 10.1038/s43588-020-00002-x.DOI
- [14]Kamal Choudhary and Brian DeCost. Atomistic line graph neural network for improved materials property predictions. npj Comput. Mater., 7:185, 2021. doi: 10.1038/s41524-021-00650-1.DOI
- [15]Robin Ruff, Patrick Reiser, Jan Stühmer, and Pascal Friederich. Connectivity optimized nested line graph networks for crystal structures. Digital Discovery, 3, 2024. doi: 10.1039/D4DD00018H.DOI
- [16]Alexander Dunn, Qi Wang, Alex Ganose, Daniel Dopp, and Anubhav Jain. Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm. npj Comput. Mater., 6:138, 2020. doi: 10.1038/s41524-020-00406-3.DOI
- [17]K. T. Schütt, M. Gastegger, A. Tkatche nko, K.-R. Müller, and R. J. Maurer. Unifying machine learning and quantum chemistry with a deep neural network for molecular wavefunctions. Nat. Commun., 10:5024, 2019. doi: 10.1038/s41467-019-12875-2.
- [18]He Li, Zun Wang, Nianlong Zou, Meng Ye, Runzhang Xu, Xiaoxun Gong, Wenhui Duan, and Yong Xu. Deep-learning density functional theory Hamiltonian for efficient ab initio electronic-structure calculation. Nat. Comput. Sci., 2:367, 2022. doi: 10.1038/s43588-022-00265-6.DOI
- [19]Xiaoxun Gong, He Li, Nianlong Zou, Runzhang Xu, Wenhui Duan, and Yong Xu. General framework for E(3)-equivariant neural network representation of density functional theory Hamiltonian. Nat. Commun., 14:2848, 2023. doi: 10.1038/s41467-023-38468-8.DOI
- [20]Yang Zhong, Hongyu Yu, Mao Su, Xingao Gong, and Hongjun Xiang. Transferable equivariant graph neural networks for the Hamiltonians of molecules and solids. npj Comput. Mater., 9:182, 2023. doi: 10.1038/s41524-023-01130-4.DOI
- [21]Qiangqiang Gu et al. Deep learning tight-binding approach for large-scale electronic simulations at finite temperatures with ab initio accuracy. Nat. Commun., 15:6772, 2024. doi: 10.1038/s41467-024-51006-4.DOI
- [22]Chen Qian, Valdas Vitartas, James Kermode, and Reinhard J. Maurer. Equivariant electronic Hamiltonian prediction with many-body message passing. npj Comput. Mater., 12:169, 2026. doi: 10.1038/s41524-026-02020-1.DOI
- [23]Yuxiang Wang, Yang Li, Zechen Tang, He Li, Zilong Yuan, Honggeng Tao, Nianlong Zou, Ting Bao, Xinghao Liang, Zezhou Chen, Shanghua Xu, Ce Bian, Zhiming Xu, Chong Wang, Chen Si, Wenhui Duan, and Yong Xu. Universal materials model of deep-learning density functional theory Hamiltonian. Sci. Bull., 2024. doi: 10.1016/j.scib.2024.06.011.DOI
- [24]Yang Zhong, Hongyu Yu, Jihui Yang, Xingyu Guo, Hongjun Xiang, and Xingao Gong. Universal machine learning Kohn–Sham Hamiltonian for materials. Chin. Phys. Lett., 41:077103, 2024. doi: 10.1088/0256-307X/41/7/077103.DOI
- [25]Shi Yin, Zujian Dai, Xinyang Pan, and Lixin He. Advancing universal deep learning for electronic-structure Hamiltonian prediction of materials. In International Conference on Learning Representations (ICLR), 2026. doi: 10.48550/arXiv.2509.19877. arXiv:2509.19877.DOI
- [26]Zechen Tang, He Li, Peize Lin, Xiaoxun Gong, Gan Jin, Lixin He, Hong Jiang, Xingguo Ren, Wenhui Duan, and Yong Xu. A deep equivariant neural network approach for efficient hybrid density functional calculations. Nat. Commun., 15:8815, 2024. doi: 10.1038/s41467-024-53028-4.DOI
- [27]Xiaoxun Gong, Steven G. Louie, Wenhui Duan, and Yong Xu. Generalizing deep learning electronic structure calculation to the plane-wave basis. Nat. Comput. Sci., 2024. doi: 10.1038/s43588-024-00701-9.DOI
- [28]Hao Tang, Brian Xiao, Wenhao He, Pero Subasic, Avetik R. Harutyunyan, Yao Wang, Fang Liu, Haowei Xu, and Ju Li. Approaching coupled-cluster accuracy for molecular electronic structures with multi-task learning. Nat. Comput. Sci., 5:144, 2025. doi: 10.1038/s43588-024-00747-9.DOI
- [29]Maximilian Cui, Karsten Reuter, and Johannes T. Margraf. Obtaining robust density functional tight-binding parameters for solids across the periodic table. J. Chem. Theory Comput., 20:5276–5290, 2024. doi: 10.1021/acs.jctc.4c00228.DOI
- [30]Kevin F. Garrity and Kamal Choudhary. Fast and accurate prediction of material properties with three-body tight-binding model for the periodic table. Phys. Rev. Mater., 7:044603, 2023. doi: 10.1103/physrevmaterials.7.044603.DOI
- [31]Thomas Froitzheim, Marcel Müller, Andreas Hansen, and Stefan Grimme. g-xTB: A general-purpose extended tight-binding electronic structure method for the elements H to Lr (Z=1–103), 2025. Chem-Rxiv preprint, doi: 10.26434/chemrxiv-2025-bjxvt.DOI
- [32]In Jun Park and Kamal Choudhary. CHIPS-TB: Evaluating tight-binding models for metals, semiconductors, and insulators, 2025. arXiv:2511.20740, doi: 10.48550/arXiv.2511.20740.arxiv.org/abs/2511.20740
- [33]Marvin L. Cohen and T. K. Bergstresser. Band structures and pseudopotential form factors for fourteen semiconductors of the diamond and zinc-blende structures. Phys. Rev., 141:789, 1966. doi: 10.1103/PhysRev.141.789.DOI
- [34]James R. Chelikowsky and Marvin L. Cohen. Nonlocal pseudopotential calculations for the electronic structure of eleven diamond and zinc-blende semiconductors. Phys. Rev. B, 14:556, 1976. doi: 10.1103/PhysRevB.14.556.DOI
- [35]Marvin L. Cohen and James R. Chelikowsky. Electronic Structure and Optical Properties of Semiconductors. Springer, Berlin, 1988. doi: 10.1007/978-3-642-97080-1.DOI
- [36]Rokyeon Kim and Young-Woo Son. Transferable empirical pseudopotentials from machine learning. Phys. Rev. B, 109:045153, 2024. doi: 10.1103/PhysRevB.109.045153.DOI
- [37]Sungmo Kang, Rokyeon Kim, Seungwu Han, and Young-Woo Son. Electronic structures of crystalline and amorphous GeSe and GeSbTe compounds using machine learning empirical pseudopotentials. APL Mach. Learn., 3:036108, 2025. doi: 10.1063/5.0270774.DOI
- [38]Kailai Lin, Matthew J. Coley-O’Rourke, and Eran Rabani. Deep-learning atomistic semi-empirical pseudopotential model for nanomaterials. npj Comput. Mater., 11:381, 2025. doi: 10.1038/s41524-025-01862-5.DOI
- [39]Qiang Zhu. EpNet-Si: Atomic form factors learned from band structures, 2026. arXiv:2610.14992, doi: 10.48550/arXiv.2610.14992.DOI
- [40]Volker Heine. s-d interaction in transition metals. Phys. Rev., 153:673, 1967. doi: 10.1103/PhysRev.153.673.DOI
- [41]D. G. Pettifor. Accurate resonance-parameter approach to transition-metal band structure. Phys. Rev. B, 2:3031, 1970. doi: 10.1103/physrevb.2.3031.DOI
- [42]Walter A. Harrison and Sverre Froyen. Universal linear-combination-of-atomic-orbitals parameters for d-state solids. Phys. Rev. B, 21:3214, 1980. doi: 10.1103/PhysRevB.21.3214.DOI
- [43]C. Y. Fong and Marvin L. Cohen. Energy band structure of copper by the empirical pseudopotential method. Phys. Rev. Lett., 24:306, 1970. doi: 10.1103/physrevlett.24.306.DOI
- [44]Lin-Wang Wang and Alex Zunger. Local-density-derived semiempirical pseudopotentials. Phys. Rev. B, 51:17398, 1995. doi: 10.1103/physrevb.51.17398.DOI
- [45]Huaxiang Fu and Alex Zunger. Local-density-derived semiempirical nonlocal pseudopotentials for InP with applications to large quantum dots. Phys. Rev. B, 55:1642, 1997. doi: 10.1103/physrevb.55.1642.DOI
- [46]J. R. Cárdenas and G. Bester. Atomic effective pseudopotentials for semiconductors. Phys. Rev. B, 86:115332, 2012. doi: 10.1103/physrevb.86.115332.DOI
- [47]Raj Kumar Paudel, Chung-Yuan Ren, and Yia-Chung Chang. Semi-empirical pseudopotential method for monolayer transition metal dichalcogenides, 2025. arXiv:2506.11360, doi: 10.48550/arXiv.2506.11360.arxiv.org/abs/2506.11360
- [48]N. W. Ashcroft. Electron-ion pseudopotentials in metals. Phys. Lett., 23:48, 1966. doi: 10.1016/0031-9163(66)90251-4.DOI
- [49]Leonard Kleinman and D. M. Bylander. Efficacious form for model pseudopotentials. Phys. Rev. Lett., 48:1425, 1982. doi: 10.1103/PhysRevLett.48.1425.DOI
- [50]E. van Lenthe, E. J. Baerends, and J. G. Snijders. Relativistic regular two-component Hamiltonians. J. Chem. Phys., 99:4597, 1993. doi: 10.1063/1.466059.DOI
- [51]David Vanderbilt. Soft self-consistent pseudopotentials in a generalized eigenvalue formalism. Phys. Rev. B, 41:7892, 1990. doi: 10.1103/physrevb.41.7892.DOI
- [52]D. R. Hamann. Optimized norm-conserving Vanderbilt pseudopotentials. Phys. Rev. B, 88:085117, 2013. doi: 10.1103/physrevb.88.085117.DOI
- [53]John P. Perdew, Kieron Burke, and Matthias Ernzerhof. Generalized gradient approximation made simple. Phys. Rev. Lett., 77:3865, 1996. doi: 10.1103/PhysRevLett.77.3865.DOI
- [54]Bryce Meredig, Erin Antono, Carena Church, Maxwell Hutchinson, Julia Ling, Sean Paradiso, Ben Blaizsik, Ian Foster, Brenna Gibbons, Jason Hattrick-Simpers, Apurva Mehta, and Logan Ward. Can machine learning identify the next high-temperature superconductor? Examining extrapolation performance for materials discovery. Mol. Syst. Des. Eng., 3:819–825, 2018. doi: 10.1039/c8me00012c.DOI
- [55]Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML), pages 41–48, 2009. doi: 10.1145/1553374.1553380.DOI
- [56]Lindsay Bassman Oftelie, Pankaj Rajak, Rajiv K. Kalia, Aiichiro Nakano, Fei Sha, Jifeng Sun, David J. Singh, Muratahan Aykol, Patrick Huck, Kristin Persson, and Priya Vashishta. Active learning for accelerated design of layered materials. npj Comput. Mater., 4:74, 2018. doi: 10.1038/s41524-018-0129-0.DOI
- [57]Xavier Gonze, Roland Stumpf, and Matthias Scheffler. Analysis of separable potentials. Phys. Rev. B, 44:8503, 1991. doi: 10.1103/physrevb.44.8503.DOI
- [58]B. Hourahine, B. Aradi, V. Blum, F. Bonafé, A. Buccheri, C. Camacho, et al. DFTB+, a software package for efficient approximate density functional theory based atomistic simulations. J. Chem. Phys., 152:124101, 2020. doi: 10.1063/1.5143190.DOI
- [59]Paolo Giannozzi et al. QUANTUM ESPRESSO: a modular and open-source software project for quantum simulations of materials. J. Phys.: Condens. Matter, 21:395502, 2009. doi: 10.1088/0953-8984/21/39/395502.DOI
- [60]P. Giannozzi et al. Advanced capabilities for materials modelling with Quantum ESPRESSO. J. Phys.: Condens. Matter, 29:465901, 2017. doi: 10.1088/1361-648X/aa8f79.DOI
- [61]G. Prandini, A. Marrazzo, I. E. Castelli, N. Mounet, and N. Marzari. Precision and efficiency in solid-state pseudopotential calculations. npj Comput. Mater., 4:72, 2018. doi: 10.1038/s41524-018-0127-2.DOI
- [62]Santosh Adhikari, Jacob Clary, Ravishankar Sundararaman, Charles Musgrave, Derek Vigil-Fowler, and Christopher Sutton. Accurate prediction of HSE06 band structures for a diverse set of materials using δ-learning. Chem. Mater., 35:8397, 2023. doi: 10.1021/acs.chemmater.3c01131.DOI
- [63]Nikolaj Rørbæk Knøsgaard and Kristian Sommer Thygesen. Representing individual electronic states for machine learning GW band structures of 2D materials. Nat. Commun., 13:468, 2022. doi: 10.1038/s41467-022-28122-0.DOI
- [64]Feliciano Giustino. Electron-phonon interactions from first principles. Rev. Mod. Phys., 89:015003, 2017. doi: 10.1103/revmodphys.89.015003.DOI
- [65]He Li, Zechen Tang, Jingheng Fu, Wen-Han Dong, Nianlong Zou, Xiaoxun Gong, Wenhui Duan, and Yong Xu. Deep-learning density functional perturbation theory. Phys. Rev. Lett., 132:096401, 2024. doi: 10.1103/physrevlett.132.096401.DOI
- [66]Yang Zhong, Shixu Liu, Binhua Zhang, Zhiguo Tao, Yuting Sun, Weიბin Chu, Xin-Gao Gong, Ji-Hui Yang, and Hongjun Xiang. Accelerating the calculation of electron–phonon coupling strength with machine learning. Nat. Comput. Sci., 4:615, 2024. doi: 10.1038/s43588-024-00668-7.DOI
- [67]Qiang Zhu. Electron–phonon coupling without density-functional perturbation theory: a learned pseudopotential Hamiltonian and a fine-tuned interatomic potential. companion manuscript, in preparation, 2026.
- [68]Per-Olov Löwdin. On the non-orthogonality problem connected with the use of atomic wave functions in the theory of molecules and crystals. J. Chem. Phys., 18:365, 1950. doi: 10.1063/1.1747632.DOI
- [69]Beatriz Cordero, Verónica Gómez, Ana E. Platero-Prats, Marc Revés, Jorge Echeverría, Eduard Cremades, Flavia Barragán, and Santiago Alvarez. Covalent radii revisited. Dalton Trans., page 2832, 2008. doi: 10.1039/b801115j.DOI
- [70]Alex Zunger. Pseudopotential and all-electron atomic core size scales. J. Chem. Phys., 74:4209, 1981. doi: 10.1063/1.441556.DOI
- [71]J. P. Perdew and Alex Zunger. Self-interaction correction to density-functional approximations for many-electron systems. Phys. Rev. B, 23:5048–5079, 1981. doi: 10.1103/PhysRevB.23.5048.DOI
- [72]Svetlana Kotochigova, Zachary H. Levine, Eric L. Shirley, M. D. Stiles, and Charles W. Clark. Local-density-functional calculations of the energy of atoms. Phys. Rev. A, 55:191–199, 1997. doi: 10.1103/PhysRevA.55.191. Erratum: Phys. Rev. A 56, 5191 (1997).DOI
- [73]S. Kotochigova, Z. H. Levine, E. L. Shirley, M. D. Stiles, and C. W. Clark. Atomic reference data for electronic structure calculations. NIST Standard Reference Database 141, 2009. doi: 10.18434/T4ZP4F.
- [74]G. Kresse and J. Furthmüller. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B, 54:11169, 1996. doi: 10.1103/physrevb.54.11169.DOI
- [75]Wahyu Setyawan and Stefano Curtarolo. High-throughput electronic band structure calculations: Challenges and tools. Comput. Mater. Sci., 49:299, 2010. doi: 10.1016/j.commatsci.2010.05.010.DOI
- [76]Yunkai Zhou, Yousef Saad, Murilo L. Tiago, and James R. Chelikowsky. Self-consistent-field calculations using Chebyshev-filtered subspace iteration. J. Comput. Phys., 219:172, 2006. doi: 10.1016/j.jcp.2006.03.017.DOI
- [77]R. P. Feynman. Forces in molecules. Phys. Rev., 56:340, 1939. doi: 10.1103/physrev.56.340.DOI
- [78]Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. doi: 10.48550/arXiv.1711.05101.DOI
- [79]Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffrey Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L. Chevrier, Kristin A. Persson, and Gerbrand Ceder. Python materials genomics (pymatgen): A robust, open-source python library for materials analysis. Comput. Mater. Sci., 68:314, 2013. doi: 10.1016/j.commatsci.2012.10.028.DOI








