Pith. sign in

REVIEW 3 major objections 7 minor 76 references

The external-potential matrix, together with its powers, is a sufficient, symmetry-adapted input for learning molecular energies, dipoles, and Fock/density matrices.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 22:54 UTC pith:O4NBFM5Y

load-bearing objection A conceptually clean framework with a solid equivariance core, but the long-range claim needs a V^1-only control and the benchmarks need more care. the 3 major comments →

arxiv 2602.15345 v2 pith:O4NBFM5Y submitted 2026-02-17 physics.chem-ph physics.comp-ph

Machine learning of electronic structure and atomistic properties from the external potential

Jigyasa Nigam , Tess Smidt , Genevi\`eve Dusson This is my paper
classification physics.chem-ph physics.comp-ph
keywords external potentialatomic orbital basisequivariant message passingoperator learningFock matrixdensity matrixlong-range interactionsHohenberg-Kohn theorem
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes that the nuclear (external) potential, represented as a matrix V in an atomic-orbital basis, is a sufficient and symmetry-adapted input for machine learning of molecular properties and electronic operators. Its central move is to feed the model not just V but also its powers V², V³, ...; matrix multiplication contracts over orbital and atomic indices, which implements equivariant message passing and carries information across the whole molecule, so long-range interactions are captured without any cutoff. The paper shows that linear models built on V match or beat the common SOAP descriptor for energies and dipole moments, and that nonlinear versions can learn Fock and density matrices from which many observables follow. If this holds, a single object unifies local geometric descriptors, equivariant graph learning, and operator prediction, with a parameter-free route to nonlocal effects.

Core claim

On its own terms, the paper's claim is that the finite-basis representation V of the nuclear attraction operator is a hierarchically body-ordered, O(3)-equivariant representation of a molecular geometry, and that polynomial expansions in V are sufficient to learn both ground-state properties and one-particle electronic operators. The key observation is that matrix products of equivariant matrices remain equivariant: the Wigner-D rotations at the intermediate atomic center cancel, so V^κ transforms like V with no additional symmetrization. Each multiplication is a message-passing step over the atomic graph, and repeated products extend the reachable neighborhood one hop at a time, giving a sc

What carries the argument

The central object is the external-potential matrix V, whose elements are nuclear-attraction integrals over pairs of atom-centered orbitals. Its powers V^κ are the load-bearing mechanism: each matrix multiplication sums over intermediate orbital and atomic indices, which the paper shows is exactly equivariant message passing in which the Wigner-D rotation matrices at the intermediate center cancel via orthogonality. The same matrix, decomposed into irreducible O(3) blocks by Clebsch-Gordan coupling, yields hierarchical body-ordered features analogous to atom-centered density correlations, with matrix powers systematically increasing the body order and spatial reach.

Load-bearing premise

The load-bearing premise is that the external-potential matrix V uniquely determines the target quantities within the data distribution — but Hohenberg-Kohn uniqueness holds only for a fixed electron count, and the models never condition on N, so identical V for different charge states (or distinct geometries under finite angular truncation) makes the learned map many-to-one.

What would settle it

Compute V in a fixed AO basis for the same geometry in two different charge states; if the two states converge to materially different Fock/density matrices, then a model trained on V alone without charge conditioning must fail on at least one of them. More directly, train an Op2Op model on a mixed-charge dataset and check whether predictions for a shared geometry collapse to a single matrix rather than tracking the charge state.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If V is sufficient, a single representation can serve both property prediction and operator learning, removing the need for separate descriptor designs for each target.
  • Matrix-power message passing naturally covers long-range interactions without cutoffs, since each power extends the receptive field by one hop and ultimately couples all atoms in the structure.
  • Because products of equivariant matrices are equivariant by construction, symmetry is preserved without explicit symmetrization at every layer, simplifying equivariant message passing.
  • Learning an effective operator in a compact basis and optimizing observables derived from it mitigates the strong error amplification that direct prediction of large Fock/density matrices suffers upon diagonalization.
  • V-based features are obtained from one-electron integral evaluations that are already part of any self-consistent-field calculation, so the input is available at essentially no extra cost.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same construction could be applied to other one-body operators (kinetic energy, overlap) or analytic functions of V; polynomial truncation is a crude stand-in for an operator exponential, and the learned coefficients might be replaced by flexible spectral filters to improve expressivity.
  • The charge-state ambiguity is a natural stress test: because Hohenberg–Kohn uniqueness holds per fixed particle number, a dataset containing multiple charge states of the same geometry would break the V-to-operator map unless electron count is supplied as an explicit input.
  • The framework turns basis-set choice into a tunable angular/radial resolution hyperparameter; a systematic scan of input basis sets could clarify how angular truncation limits distinguishability, complementing the paper's octahedron degeneracy examples.
  • Because V powers couple all atoms, the model could be extended to periodic systems or large fragments where the matrix is sparse and block-structured, with the polynomial order acting as an effective interaction range.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes using the external (nuclear) potential in an atomic-orbital basis, V, as the central input for machine-learning models of molecular properties and electronic-structure operators. It develops a symmetry-adapted decomposition of V (Section II.A), shows that powers of V are equivariant and implement message passing through matrix multiplication (Section II.B and SM §III), and connects V elements to body-ordered atom-centered descriptors (SM §II). The authors demonstrate linear Op2Prop models from V against SOAP on water and QM7b subsets, a water-dimer long-range experiment, and Op2Op models mapping V to Fock/density matrices, including effective downfolded operators. The central claim is that V unifies local descriptors, equivariant message passing, and operator learning, with efficient long-range handling.

Significance. If the claims hold, the paper provides a genuinely unifying perspective: a single matrix-valued object that simultaneously behaves as a local body-ordered descriptor, an equivariant message-passing operator, and a physically motivated input for operator learning. The theoretical core is a strength: the equivariance of matrix products follows cleanly from Wigner-matrix orthogonality (SM §III), the three-body reading of V elements follows from linearity in the nuclear charge density (SM §II, Eq. S13), and the message-passing interpretation via 6j recoupling is concrete. The empirical demonstrations are sensible, though many are on small datasets and simple linear models. The most distinctive empirical claim — that successive products of V enable long-range effects — currently lacks the control needed to attribute the observed behavior to the matrix powers rather than to V itself. The manuscript would be strengthened by tighter phrasing of the HK-based sufficiency claim in finite bases and by explicit treatment of electron number as a conditioning variable.

major comments (3)
  1. [III.D, Fig. 5; §II.B, Eq. (11)] The long-range claim is not supported without a V^1-only control. Each V element (Eq. 2) already contains a sum over all nuclei with a Coulomb 1/|r−R_a| kernel; for a distant nucleus a, the leading contribution to a localized AO pair is approximately −Z_a S_{αβ}/|R_a − R_ij|, so V alone is cutoff-free and contains long-range information. Fig. 5 compares SOAP (4 Å cutoff) with features from V, V^2, and V^3, so it cannot isolate the effect of 'successive products.' Please add κmax=1 (and ideally κmax=2) curves and discuss whether the tail is due to V^1 rather than to message passing.
  2. [§II (HK1 paragraph); §III.B, III.F (QM7b)] HK1 is invoked only 'for a fixed particle number and a nondegenerate ground state,' but the manuscript never states how the electron number is supplied to the model. If the QM7b subset includes more than one charge state for the same geometry, the same V matrix would map to different converged Fock/density matrices, making the learned map many-to-one. Even if all molecules in the subset are neutral, the conditioning should be stated explicitly and, if needed, encoded (e.g., by adding total charge as an input). Please clarify the charge-state composition of the dataset or filter it accordingly.
  3. [III.A, Fig. 3 and following text] The paper's own analysis shows that for any finite AO angular truncation, distinct geometries can produce identical V elements ('for any finite angular truncation, it is possible to construct an example that would lead to degeneracy'). This means the representation is not injective in practice, and the abstract's 'sufficient' input language should be qualified to the complete-basis limit or to the empirical approximation learned from data. The authors acknowledge the limitation in the body, but the framing in the abstract and conclusions should be adjusted so that readers do not infer exact sufficiency for finite bases.
minor comments (7)
  1. [Throughout] Several cross-references appear as 'SM Section??' (e.g., in Sections II, III.A, and III.B). These placeholders must be resolved.
  2. [Eq. (8), Eq. (13)] The summation over the q' index is implicit in Eq. (8) and Eq. (13). Please write the sum explicitly to avoid ambiguity.
  3. [Eq. (16)] The gating function g^{(γ)} notation is unclear: does it depend on the output feature index q? It would help to specify the MLP input/output dimensions explicitly.
  4. [III.D, dimer experiment] The y-axis label 'interaction energy (total energy baselined by the mean E)' is not an interaction energy in the usual sense. Please specify the baseline (e.g., separated monomers) or relabel the quantity as a relative total energy.
  5. [III.B–III.F] Reported RMSE values lack error bars or repeated-split statistics. Given the small validation sets (e.g., 200 structures), some numbers may not be stable. Please provide uncertainties or at least a statement about variance across splits.
  6. [III.D, QM7b SOAP comparison] The SOAP hyperparameters for QM7b are not stated in the main text; in particular, the cutoff radius is not given. Please report the exact SOAP settings so the comparison is reproducible.
  7. [Data availability] The statement says code will be made available 'on publication.' If a repository exists, a link or DOI would strengthen reproducibility.

Circularity Check

0 steps flagged

No circularity: V^k equivariance and body-order derivations are self-contained; the long-range attribution in the dimer experiment lacks a V^1-only control but is not a circular step.

full rationale

The paper's load-bearing derivations are self-contained and do not reduce to their inputs. The equivariance of matrix powers is derived directly from Wigner-D orthogonality in Eqs. (S19)-(S29), and the irreps/permutation construction in Eq. (5) is re-derived rather than imported. The polynomial expansion in Eq. (11) is an explicit ansatz with tunable weights; all Op2Prop and Op2Op results are openly supervised regressions evaluated on held-out validation sets, so no fitted parameter is renamed as a prediction. The Hohenberg-Kohn citation is an external theorem used as motivation, and the paper itself states the fixed-particle-number/nondegenerate condition, so the omission of N as an input in the QM7b example is a well-posedness concern, not a circular reduction. Self-citations such as Ref. [22] are used for standard angular-momentum algebra that the SM re-derives, so they are not load-bearing. The main empirical concern is Fig. 5: the dimer tail is produced with features V, V^2, V^3, and V^1 already contains cutoff-free Coulomb sums, so the claim that successive matrix powers enable the long-range behavior lacks a V^1-only ablation; this is a missing control and attribution gap, not an equation-level equivalence. The paper also explicitly acknowledges finite-basis degeneracies in Fig. 3, which is an expressivity limitation rather than circularity. No circular step can be exhibited from the text.

Axiom & Free-Parameter Ledger

7 free parameters · 8 axioms · 0 invented entities

The framework introduces no new physical entities — no new particles, forces, or conserved quantities. Inputs are standard one-electron integrals (V, S, T) and targets are standard DFT/LDA quantities. The derivations rest on standard angular-momentum algebra (Wigner-D orthogonality, CG/6j recoupling) and on domain conventions (closed-shell idempotency, DFT/LDA as ground truth, AO locality/decay). The genuinely paper-specific assumptions are architectural (blockwise block-diagonal maps Eq. 8/13; polynomial truncation κmax; gating MLP Eq. 16) and the fixed-electron-number condition for the V→M map, which is stated but never enforced. The mathematical statements are self-contained; the empirical performance depends on roughly a dozen free choices, several of them unreported (SOAP hyperparameters, loss weights, projection threshold ε, QM7b subset composition).

free parameters (7)
  • Blockwise linear weights w^(γ) and w^(γ,κ) = fitted per task via Eq. 8/13 loss
    Standard supervised regression weights mapping V-blocks to property/operator blocks; openly fitted, not derived.
  • Matrix-power order κmax = 2 or 3 (water: 3 for properties, 2 for effective Op2Op; QM7b: 3)
    Chosen by hand; §III.E shows κ=3 improves Fock MSE almost fivefold over κ=1 — a tunable expressivity knob, not derived.
  • Input basis B_V = cc-pVTZ, cc-pVQZ, STO-3G, def2-SVP, def2-TZVP depending on experiment
    Basis resolution of V is a declared tunable hyperparameter (§II; Fig. 6 demonstrates its effect on accuracy).
  • Effective-output basis B_M and projection threshold ε = B_M = STO-3G; ε not reported
    Downfolded Fock basis and adaptive projection threshold (SM §VI) determine the effective Op2Op results.
  • SOAP baseline hyperparameters = not reported (cutoff 4 Å in dimer experiment)
    SOAP r_cut, n_max, l_max promised 'in the SM' but absent from the provided SM; needed to judge the 'matches or exceeds' claim.
  • Loss weights ω_y (Eq. 17) = not reported
    Weights for eigenvalue/charge/dipole terms in the effective Op2Op loss; values unspecified.
  • Gating MLP g^(γ) parameters = not reported
    The nonlinear gating network in Eq. 16; architecture and hyperparameters not detailed.
axioms (8)
  • standard math Wigner-D orthogonality: Σ_m D†_{m'a}(R) D_{a m''}(R) = δ_{m'm''}
    Used in SM §III (Eq. S24-S29) to prove rotational equivariance of matrix products V^κ.
  • standard math Clebsch-Gordan coupling and Wigner-6j recoupling are unitary and closed
    Used to decompose V into O(3) irreps (Eq. 5) and to interpret matrix products as invariant propagation at intermediate centers (SM Eq. S30-S32).
  • domain assumption HK1: for fixed N and nondegenerate ground state, V uniquely determines ground-state observables
    Motivates the V→property/operator maps (§II) but is not constructive; its fixed-N caveat is not enforced in the QM7b application (§III.F).
  • domain assumption DFT/LDA (PySCF) reference values are ground truth
    All labels (E, μ, H, P, ε) are LDA in def2-TZVP/def2-SVP/STO-3G; model accuracy is relative to this level of theory.
  • domain assumption AO matrix elements of V decay with interatomic separation
    Justifies the locality analogies to SOAP/ACE; verified numerically in SM §IV (6 Å truncation error ~7e-6 a.u. for H_STO-3G).
  • domain assumption Closed-shell idempotency PSP = 2P (occupation f ∈ {0,2})
    Used to define Grassmannian diagnostics (SM §V); valid only for closed-shell single determinants.
  • ad hoc to paper Blockwise ansatz: target block (γ, λσ) depends only on same-γ blocks of V and its powers
    Eq. 8/13 restrict the linear map to be block-diagonal in species-pair and angular momentum; cross-block coupling enters only through V^κ. This is an architectural ansatz, not derived.
  • ad hoc to paper Fixed electron number per molecule
    The V→M map is well-defined only at fixed N (HK1 caveat stated in §II); the QM7b set contains charge variants and the paper does not state how N is conditioned.

pith-pipeline@v1.3.0-alltime-deepseek · 29794 in / 29138 out tokens · 295275 ms · 2026-08-02T22:54:33.008938+00:00 · methodology

0 comments
read the original abstract

Electronic structure calculations remain a major bottleneck in atomistic simulations and, not surprisingly, have attracted significant attention in machine learning (ML). Most existing approaches learn a direct map from molecular geometries, typically represented as graphs or encoded local environments, to molecular properties or use ML as a surrogate for electronic structure theory by targeting quantities such as Fock or density matrices expressed in an atomic orbital (AO) basis. Inspired by the Hohenberg-Kohn theorem, in this work, we propose an operator-centered framework in which the external (nuclear) potential, expressed in an AO basis, serves as the model input. From this operator, we construct hierarchical, body-ordered representations of atomic configurations that closely mirror the principles underlying several popular atom-centered descriptors. At the same time, the matrix-valued nature of the external potential provides a natural connection to equivariant message-passing neural networks. In particular, we show that successive products of the external potential provide a scalable route to equivariant message passing and enable an efficient description of long-range effects. We demonstrate that this approach can be used to model molecular properties, such as energies and dipole moments, from the external potential, or learn effective operator-to-operator maps, including mappings to the Fock matrix and the reduced density matrix from which multiple molecular observables can be simultaneously derived.

Figures

Figures reproduced from arXiv: 2602.15345 by Genevi\`eve Dusson, Jigyasa Nigam, Tess Smidt.

Figure 1
Figure 1. Figure 1: FIG. 1: Schematic overview of different models considered in this work. For each molecular configuration, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Decomposition of AO matrix representations of [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Examples demonstrating the role of basis [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Considering only the invariant features [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Interaction energy for water dimers (total [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Relative RMSE per matrix element (Eq. (20), [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7: Relative RMSE for the validation set on [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Relative RMSE on L¨owdin charges ( [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9: Relative RMSE on the total energy ( [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10: Relative RMSE on L¨owdin charges ( [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

76 extracted references · 10 linked inside Pith

  1. [1]

    Op2Op models We now consider nonlinear models that incorporate ma- trix products and nonlinear gating as described in Sec- tion II D. For these models, we use cc-pVTZ as the in- put basisB V to express the external potential, and em- ploy the same architectures for models targeting both HandP, without attempting to optimize hyperparam- eters. We supervise...

  2. [2]

    Effective Op2Op models As discussed in Section II E, we can model the Fock op- erator on an effective basisB M even though the target properties are computed with a larger basisB y, allowing the choice of basis set representation ofMto be treated as a model hyperparameter. To do so requires a differ- entiable evaluation of the properties from the predicte...

  3. [3]

    Op2Op models Rather than directly predict the full Fock matrixH, we target only the nonlocal exchange-correlation correction, which is obtained as the difference ofHand the non- interacting HamiltonianH core =T+V, whereTis the kinetic energy operator. This decomposition allows the model to focus on learning the exchange-correlation ef- fects, which are th...

  4. [4]

    Effective Op2Op models The benefit of the effectiveOp2Opmodels becomes most apparent for this example. As previously observed, for identicalVinputs and network architectures, the accu- racy ofOp2Opmodels degrades systematically as the target basis expands from STO-3G to def2-TZVP, and derived observables are especially sensitive to these er- rors. Instead...

  5. [5]

    Generalized Neural- Network Representation of High-Dimensional Potential- Energy Surfaces,

    J¨ org Behler and Michele Parrinello, “Generalized Neural- Network Representation of High-Dimensional Potential- Energy Surfaces,” Phys. Rev. Lett.98, 146401 (2007)

    2007

  6. [6]

    Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Elec- trons,

    Albert P. Bart´ ok, Mike C. Payne, Risi Kondor, and G´ abor Cs´ anyi, “Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Elec- trons,” Phys. Rev. Lett.104, 136–403 (2010)

    2010

  7. [7]

    E(3)-Equivariant Graph Neural Networks for Data-efficient and Accurate Interatomic Potentials,

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. Schmidt, and B. Kozinsky, “E(3)-Equivariant Graph Neural Networks for Data-efficient and Accurate Interatomic Potentials,” Nature Communications13(2022)

    2022

  8. [8]

    MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields,

    Ilyes Batatia, David P Kovacs, Gregor Simm, Christoph Ortner, and Gabor Csanyi, “MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields,” inNeurIPS, Vol. 35, edited by S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Curran Associates, Inc., 2022) pp. 11423–11436

    2022

  9. [9]

    Fast and Ac- curate Modeling of Molecular Atomization Energies with Machine Learning,

    Matthias Rupp, Alexandre Tkatchenko, Klaus-Robert M¨ uller, and O. Anatole von Lilienfeld, “Fast and Ac- curate Modeling of Molecular Atomization Energies with Machine Learning,” Phys. Rev. Lett.108, 058301 (2012)

    2012

  10. [10]

    Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Poten- tials,

    Alexander V. Shapeev, “Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Poten- tials,” Multiscale Modeling & Simulation14, 1153–1173 (2016)

    2016

  11. [11]

    Atomic Cluster Expansion for Accurate and Transferable Interatomic Potentials,

    Ralf Drautz, “Atomic Cluster Expansion for Accurate and Transferable Interatomic Potentials,” Phys. Rev. B 99, 014104 (2019)

    2019

  12. [12]

    Four generations of high-dimensional neu- ral network potentials,

    Jorg Behler, “Four generations of high-dimensional neu- ral network potentials,” Chemical Reviews121, 10037– 10072 (2021)

    2021

  13. [13]

    Learning Local Equivariant Representa- tions for Large-Scale Atomistic Dynamics,

    Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J Owen, Mordechai Kornbluth, and Boris Kozinsky, “Learning Local Equivariant Representa- tions for Large-Scale Atomistic Dynamics,” Nature Com- munications14, 579 (2023)

    2023

  14. [14]

    Predicting Molecular Dipole Moments by Combining Atomic Partial Charges and Atomic Dipoles,

    Max Veit, David M. Wilkins, Yang Yang, Robert A. DiStasio, and Michele Ceriotti, “Predicting Molecular Dipole Moments by Combining Atomic Partial Charges and Atomic Dipoles,” J. Chem. Phys.153, 024–113 (2020)

    2020

  15. [15]

    Molecular Dipole Moment Learning via Rotationally Equivariant Derivative Kernels in Molecular-orbital- based Machine Learning,

    Jiace Sun, Lixue Cheng, and III Miller, Thomas F., “Molecular Dipole Moment Learning via Rotationally Equivariant Derivative Kernels in Molecular-orbital- based Machine Learning,” The Journal of Chemical Physics157, 104–109 (2022)

    2022

  16. [16]

    Physically Inspired Deep Learning of Molecular Excitations and Photoemission Spectra,

    Julia Westermayr and Reinhard J. Maurer, “Physically Inspired Deep Learning of Molecular Excitations and Photoemission Spectra,” Chem. Sci.12, 10755–10764 (2021)

    2021

  17. [17]

    Electronic excited states from physically constrained machine learning,

    Edoardo Cignoni, Divya Suman, Jigyasa Nigam, Lorenzo Cupellini, Benedetta Mennucci, and Michele Ceriotti, “Electronic excited states from physically constrained machine learning,” ACS Central Science10, 637–648 (2024)

    2024

  18. [18]

    SPA ˆHM: The Spectrum of Approximated Hamiltonian Matrices representations,

    Alberto Fabrizio, Ksenia R Briling, and Clemence Corminboeuf, “SPA ˆHM: The Spectrum of Approximated Hamiltonian Matrices representations,” Digital Discov- ery , 286–294 (2022)

    2022

  19. [19]

    Transferability in Machine Learning for Elec- tronic Structure via the Molecular Orbital Basis,

    Matthew Welborn, Lixue Cheng, and Thomas F Miller III, “Transferability in Machine Learning for Elec- tronic Structure via the Molecular Orbital Basis,” Jour- nal of Chemical Theory and Computation14, 4772–4779 (2018)

    2018

  20. [20]

    OrbNet: Deep Learning for Quantum Chemistry Using Symmetry-adapted Atomic-orbital Features,

    Zhuoran Qiao, Matthew Welborn, Animashree Anand- kumar, Frederick R Manby, and Thomas F Miller III, “OrbNet: Deep Learning for Quantum Chemistry Using Symmetry-adapted Atomic-orbital Features,” The Jour- nal of Chemical Physics153, 124–111 (2020)

    2020

  21. [21]

    Transferable machine-learning model of the electron density,

    Andrea Grisafi, Alberto Fabrizio, Benjamin Meyer, David M Wilkins, Clemence Corminboeuf, and Michele Ceriotti, “Transferable machine-learning model of the electron density,” ACS Central Science5, 57–64 (2018)

    2018

  22. [22]

    A Recipe for Cracking the Quantum Scaling Limit with Machine Learned Electron Densities,

    Joshua A Rackers, Lucas Tecot, Mario Geiger, and Tess E Smidt, “A Recipe for Cracking the Quantum Scaling Limit with Machine Learned Electron Densities,” Machine Learning: Science and Technology4, 015–027 (2023)

    2023

  23. [23]

    Higher-order equivariant neural networks for charge density prediction in materials,

    Teddy Koker, Keegan Quigley, Eric Taw, Kevin Tibbetts, and Lin Li, “Higher-order equivariant neural networks for charge density prediction in materials,” npj Compu- tational Materials10, 161 (2024)

    2024

  24. [24]

    Machine-learned Approximations to Density Functional Theory Hamilto- nians,

    Ganesh Hegde and R Chris Bowen, “Machine-learned Approximations to Density Functional Theory Hamilto- nians,” Scientific reports7, 1–11 (2017)

    2017

  25. [25]

    Unifying Machine Learning and Quantum Chemistry with a Deep Neural Network for Molecular Wavefunctions,

    Kristoff T. Sch¨ utt, Michael Gastegger, Alexandre Tkatchenko, Klaus-Robert M¨ uller, and Reinhard J. Maurer, “Unifying Machine Learning and Quantum Chemistry with a Deep Neural Network for Molecular Wavefunctions,” Nat Commun10, 50–24 (2019)

    2019

  26. [26]

    Equivariant Representations for Molecular Hamiltoni- ans andN-Center Atomic-Scale Properties,

    Jigyasa Nigam, Michael J. Willatt, and Michele Ceriotti, “Equivariant Representations for Molecular Hamiltoni- ans andN-Center Atomic-Scale Properties,” J. Chem. Phys.156, 014–115 (2022)

    2022

  27. [27]

    SE(3)-equivariant Prediction of Molecular Wavefunc- tions and Electronic Densities,

    Oliver T. Unke, Mihail Bogojeski, Michael Gastegger, Mario Geiger, Tess Smidt, and Klaus-Robert M¨ uller, “SE(3)-equivariant Prediction of Molecular Wavefunc- tions and Electronic Densities,” NeurIPS , 14434–14447 (2021)

    2021

  28. [28]

    Equivariant Analytical Mapping of First Principles Hamiltonians to Accurate and Transferable Materials Models,

    Liwei Zhang, Berk Onat, Genevi` eve Dusson, Adam Mc- Sloy, Gautam Anand, Reinhard J Maurer, Christoph Or- tner, and James R Kermode, “Equivariant Analytical Mapping of First Principles Hamiltonians to Accurate and Transferable Materials Models,” npj Computational Materials8, 158 (2022)

    2022

  29. [29]

    Harmonizing Covariance and Expressiveness for Deep Hamiltonian Regression in Crys- talline Material Research: a Hybrid Cascaded Regression Framework,

    Shi Yin, Xudong Zhu, Tianyu Gao, Haochong Zhang, Feng Wu, and Lixin He, “Harmonizing Covariance and Expressiveness for Deep Hamiltonian Regression in Crys- talline Material Research: a Hybrid Cascaded Regression Framework,” arXiv:2401.00744 (2024)

    Pith/arXiv arXiv 2024

  30. [30]

    Universal Machine Learning Kohn-Sham Hamil- tonian for Materials,

    Yang Zhong, Jihui Yang, Hongjun Xiang, and Xingao Gong, “Universal Machine Learning Kohn-Sham Hamil- tonian for Materials,” arXiv:2402.09251 (2024). 17

    Pith/arXiv arXiv 2024

  31. [31]

    GEARS H: Accurate Machine-learned Hamiltonians for Next-Generation Device-scale Model- ing,

    Anubhab Haldar, Ali K Hamze, Nikhil Sivadas, and Yongwoo Shin, “GEARS H: Accurate Machine-learned Hamiltonians for Next-Generation Device-scale Model- ing,” arXiv:2506.10298 (2025)

    arXiv 2025

  32. [32]

    Slakonet: A unified slater-koster tight-binding framework using neural network infrastruc- ture for the periodic table,

    Kamal Choudhary, “Slakonet: A unified slater-koster tight-binding framework using neural network infrastruc- ture for the periodic table,” The Journal of Physical Chemistry Letters16, 11109–11119 (2025)

    2025

  33. [33]

    Graph2Mat: Universal Graph to Matrix Conversion for Electron Density Prediction,

    Pol Febrer, Peter Jørgensen, Miguel Pruneda, Al- berto Garcia, Pablo Ordejon, and Arghya Bhowmik, “Graph2Mat: Universal Graph to Matrix Conversion for Electron Density Prediction,” (2024)

    2024

  34. [34]

    A Symmetry-preserving and Transferable Represen- tation for Learning the Kohn-Sham Density Matrix,

    Liwei Zhang, Patrizia Mazzeo, Michele Nottoli, Edoardo Cignoni, Lorenzo Cupellini, and Benjamin Stamm, “A Symmetry-preserving and Transferable Represen- tation for Learning the Kohn-Sham Density Matrix,” arXiv:2503.08400 (2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Image Super-resolution Inspired Electron Density Prediction,

    Chenghan Li, Or Sharir, Shunyue Yuan, and Garnet K Chan, “Image Super-resolution Inspired Electron Density Prediction,” arXiv:2402.12335 (2024)

    Pith/arXiv arXiv 2024

  36. [36]

    Pushing the frontiers of density functionals by solving the fractional electron problem,

    James Kirkpatrick, Brendan McMorrow, David HP Tur- ban, Alexander L Gaunt, James S Spencer, Alexan- der GDG Matthews, Annette Obika, Louis Thiry, Meire Fortunato, David Pfau,et al., “Pushing the frontiers of density functionals by solving the fractional electron problem,” Science374, 1385–1389 (2021)

    2021

  37. [37]

    Accurate and scalable exchange-correlation with deep learning,

    Giulia Luise, Chin-Wei Huang, Thijs Vogels, Derk P Kooi, Sebastian Ehlert, Stephanie Lanius, Klaas J H Giesbertz, Amir Karton, Deniz Gunceler, Megan Stanley, Wessel P Bruinsma, Lin Huang, Xinran Wei, Jos´ e Gar- rido Torres, Abylay Katbashev, B´ alint M´ at´ e, S´ ekou- Oumar Kaba, Roberto Sordillo, Yingrong Chen, David B Williams-Young, Christopher M Bis...

    Pith/arXiv arXiv 2025

  38. [38]

    Exploring the Design Space of Machine Learning Models for Quantum Chemistry with a Fully Differentiable Framework,

    Divya Suman, Jigyasa Nigam, Sandra Saade, Paolo Pe- golo, Hanna T¨ urk, Xing Zhang, Garnet Kin-Lic Chan, and Michele Ceriotti, “Exploring the Design Space of Machine Learning Models for Quantum Chemistry with a Fully Differentiable Framework,” Journal of Chemical Theory and Computation (2025)

    2025

  39. [39]

    Approaching coupled-cluster accuracy for molecular electronic structures with multi-task learn- ing,

    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 learn- ing,” Nature Computational Science , 1–11 (2024)

    2024

  40. [40]

    dxtb—an efficient and fully differen- tiable framework for extended tight-binding,

    Marvin Friede, Christian H¨ olzer, Sebastian Ehlert, and Stefan Grimme, “dxtb—an efficient and fully differen- tiable framework for extended tight-binding,” The Jour- nal of Chemical Physics161(2024)

    2024

  41. [41]

    Inhomogeneous electron gas,

    Pierre Hohenberg and Walter Kohn, “Inhomogeneous electron gas,” Physical review136, B864 (1964)

    1964

  42. [42]

    Physics- Inspired Structural Representations for Molecules and Materials,

    Felix Musil, Andrea Grisafi, Albert P. Bart´ ok, Christoph Ortner, G´ abor Cs´ anyi, and Michele Ceriotti, “Physics- Inspired Structural Representations for Molecules and Materials,” Chem. Rev.121, 9759–9815 (2021)

    2021

  43. [43]

    Ten- sor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds,

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley, “Ten- sor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds,” arxiv:1802.08219 (2018)

    Pith/arXiv arXiv 2018

  44. [44]

    Neural message pass- ing for quantum chemistry,

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl, “Neural message pass- ing for quantum chemistry,” inInt. Conf. Mach. Learn. (2017) pp. 1263–1272

    2017

  45. [45]

    Bypass- ing the kohn-sham equations with machine learning,

    Felix Brockherde, Leslie Vogt, Li Li, Mark E Tucker- man, Kieron Burke, and Klaus-Robert M¨ uller, “Bypass- ing the kohn-sham equations with machine learning,” Na- ture communications8, 872 (2017)

    2017

  46. [46]

    Machine learning electronic struc- ture methods based on the one-electron reduced density matrix,

    Xuecheng Shao, Lukas Paetow, Mark E Tuckerman, and Michele Pavanello, “Machine learning electronic struc- ture methods based on the one-electron reduced density matrix,” Nature communications14, 6281 (2023)

    2023

  47. [47]

    Learning the one-electron reduced density matrix at scf convergence thresholds,

    Bhaskar Rana, Nicolas Viot, Jessica A Martinez B, Xuecheng Shao, Pablo Ramos, and Michele Pavanello, “Learning the one-electron reduced density matrix at scf convergence thresholds,” Journal of Chemical Theory and Computation (2025)

    2025

  48. [48]

    Atom-density representations for machine learning,

    Michael J. Willatt, F´ elix Musil, and Michele Ceriotti, “Atom-density representations for machine learning,” J. Chem. Phys.150, 154110 (2019)

    2019

  49. [49]

    Unified theory of atom-centered representations and message-passing machine-learning schemes,

    Jigyasa Nigam, Sergey Pozdnyakov, Guillaume Fraux, and Michele Ceriotti, “Unified theory of atom-centered representations and message-passing machine-learning schemes,” J. Chem. Phys.156, 204115 (2022)

    2022

  50. [50]

    Atomic cluster expansion: Complete- ness, efficiency and stability,

    Genevieve Dusson, Markus Bachmayr, G´ abor Cs´ anyi, Ralf Drautz, Simon Etter, Cas van Der Oord, and Christoph Ortner, “Atomic cluster expansion: Complete- ness, efficiency and stability,” Journal of Computational Physics454, 110946 (2022)

    2022

  51. [51]

    On representing chemical environments,

    Albert P. Bart´ ok, Risi Kondor, and G´ abor Cs´ anyi, “On representing chemical environments,” Phys. Rev. B87, 184115 (2013)

    2013

  52. [53]

    Semi-supervised classifica- tion with graph convolutional networks,

    TN Kipf and Max Welling, “Semi-supervised classifica- tion with graph convolutional networks,” arXiv preprint arXiv:1609.02907 (2016)

    Pith/arXiv arXiv 2016

  53. [54]

    Recursive evaluation and iterative contraction ofN -body equivariant features,

    Jigyasa Nigam, Sergey Pozdnyakov, and Michele Ceri- otti, “Recursive evaluation and iterative contraction ofN -body equivariant features,” J. Chem. Phys.153, 121101 (2020)

    2020

  54. [55]

    Gaussian process regression for materials and molecules,

    Volker L Deringer, Albert P Bart´ ok, Noam Bern- stein, David M Wilkins, Michele Ceriotti, and G´ abor Cs´ anyi, “Gaussian process regression for materials and molecules,” Chemical reviews121, 10073–10141 (2021)

    2021

  55. [56]

    On the completeness of atomic structure rep- resentations,

    Sergey N Pozdnyakov, Michael J Willatt, Albert P Bart´ ok, Christoph Ortner, G´ abor Cs´ anyi, and Michele Ceriotti, “On the completeness of atomic structure rep- resentations,” arXiv preprint arXiv:2001.11696 (2020)

    Pith/arXiv arXiv 2001

  56. [57]

    Symmetry-Adapted Machine Learn- ing for Tensorial Properties of Atomistic Systems,

    Andrea Grisafi, David M. Wilkins, G´ abor Cs´ anyi, and Michele Ceriotti, “Symmetry-Adapted Machine Learn- ing for Tensorial Properties of Atomistic Systems,” Phys. Rev. Lett.120, 036002 (2018)

    2018

  57. [58]

    Dataset: Symmetry-Adapted Ma- chine Learning for Tensorial Properties of Atomistic Systems,

    Andrea Grisafi, David M. Wilkins, Gabor Cs´ anyi, and Michele Ceriotti, “Dataset: Symmetry-Adapted Ma- chine Learning for Tensorial Properties of Atomistic Systems,”http://doi.org/10.24435/materialscloud: 2018.0009/v1(2018)

    arXiv 2018

  58. [59]

    Machine Learning of Molecular Elec- tronic Properties in Chemical Compound Space,

    Gr´ egoire Montavon, Matthias Rupp, Vivekanand Go- bre, Alvaro Vazquez-Mayagoitia, Katja Hansen, Alexan- dre Tkatchenko, Klaus-Robert M¨ uller, and O Anatole Von Lilienfeld, “Machine Learning of Molecular Elec- tronic Properties in Chemical Compound Space,” New Journal of Physics15, 095–003 (2013). 18

    2013

  59. [60]

    PySCF: the Python-based Simulations of Chemistry Framework,

    Qiming Sun, Timothy C Berkelbach, Nick S Blunt, George H Booth, Sheng Guo, Zhendong Li, Junzi Liu, James D McClain, Elvira R Sayfutyarova, Sandeep Sharma,et al., “PySCF: the Python-based Simulations of Chemistry Framework,” Wiley Interdisciplinary Reviews: Computational Molecular Science8, e1340 (2018)

    2018

  60. [61]

    The role of feature space in atomistic learning,

    Alexander Goscinski, Guillaume Fraux, Giulio Im- balzano, and Michele Ceriotti, “The role of feature space in atomistic learning,” Machine Learning: Science and Technology2, 025028 (2021)

    2021

  61. [62]

    Equivariant electronic hamilto- nian prediction with many-body message passing,

    Chen Qian, Valdas Vitartas, James Kermode, and Reinhard J Maurer, “Equivariant electronic hamilto- nian prediction with many-body message passing,” arXiv preprint arXiv:2508.15108 (2025)

    arXiv 2025

  62. [63]

    Differentiable quantum chemistry with pyscf for molecules and materi- als at the mean-field level and beyond,

    Xing Zhang and Garnet Kin-Lic Chan, “Differentiable quantum chemistry with pyscf for molecules and materi- als at the mean-field level and beyond,” The Journal of Chemical Physics157, 204801 (2022)

    2022

  63. [64]

    A smooth basis for atomistic machine learning,

    Filippo Bigi, Kevin K. Huguenin-Dumittan, Michele Ce- riotti, and David E. Manolopoulos, “A smooth basis for atomistic machine learning,” J. Chem. Phys.157, 234101 (2022)

    2022

  64. [65]

    Compressing local atomic neighbourhood descriptors,

    James P Darby, James R Kermode, and G´ abor Cs´ anyi, “Compressing local atomic neighbourhood descriptors,” npj Computational Materials8, 166 (2022)

    2022

  65. [66]

    Optimal ra- dial basis for density-based atomic representations,

    Alexander Goscinski, F´ elix Musil, Sergey Pozdnyakov, Jigyasa Nigam, and Michele Ceriotti, “Optimal ra- dial basis for density-based atomic representations,” The Journal of Chemical Physics155(2021). Machine learning electronic structure and atomistic properties from the external potential Supplementary Material Jigyasa Nigam,1,∗ Tess Smidt,1 and Geneviè...

    2021

  66. [67]

    Nigam, M

    J. Nigam, M. J. Willatt, and M. Ceriotti, Equivariant Representations for Molecular Hamiltonians and N-Center Atomic-Scale Properties, J. Chem. Phys.156, 014 (2022)

    2022

  67. [68]

    Nigam, S

    J. Nigam, S. Pozdnyakov, and M. Ceriotti, Recursive evaluation and iterative contraction ofN-body equivariant features, J. Chem. Phys.153, 121101 (2020)

    2020

  68. [69]

    M. J. Willatt, F. Musil, and M. Ceriotti, Atom-density representations for machine learning, J. Chem. Phys.150, 154110 (2019)

    2019

  69. [70]

    P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zambaldi, M. Malinowski, A. Tac- chetti, D. Raposo, A. Santoro, R. Faulkner,et al., Relational inductive biases, deep learning, and graph networks, arXiv preprint arXiv:1806.01261 (2018)

    Pith/arXiv arXiv 2018

  70. [71]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, Neural message passing for quantum chemistry, inInt. Conf. Mach. Learn.(2017) pp. 1263–1272

    2017

  71. [72]

    Jiang, C

    Q. Jiang, C. Wang, M. Lones, and W. Pang, Demystifying mpnns: Message passing as merely efficient matrix multiplication, arXiv preprint arXiv:2502.00140 (2025)

    Pith/arXiv arXiv 2025

  72. [73]

    Godsil and G

    C. Godsil and G. Royle,Algebraic Graph Theory (Graduate Texts in Mathematics 207) ISBN 0-387- 95220-9 (Springer, New York, 2001)., Vol. 34 (Cambridge University Press, 2002) pp. 619–631

    2001

  73. [74]

    L. C. Biedenharn and J. D. Louck,The Racah-Wigner algebra in Quantum Theory, Encyclopedia of Mathematics and its Applications(Addison-Wesley, 1981)

    1981

  74. [75]

    S. F. Boys, Electronic wave functions-i. a general method of calculation for the stationary states of any molecular system, Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences200, 542 (1950)

    1950

  75. [76]

    Obara and A

    S. Obara and A. Saika, Efficient recursive computation of molecular integrals over cartesian gaussian functions, The Journal of chemical physics84, 3963 (1986)

    1986

  76. [77]

    Pu and Q

    Z. Pu and Q. Sun, Enhancing pyscf-based quantum chemistry simulations with modern hardware, algorithms, and python tools, APL Computational Physics1, 016101 (2025). 16 35

    2025