Pith. sign in

REVIEW 3 major objections 5 minor 80 references

A graph neural network can learn the effective magnetic energy of metallic magnets from electronic data and drive spin dynamics without repeatedly solving the electronic problem.

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 · grok-4.5

2026-07-31 04:17 UTC pith:2SJO46HM

load-bearing objection Solid incremental methods paper: a tailored GNN force field that actually matches electronic torques and LLG dynamics across three s–d textures, including chiral coarsening. the 3 major comments →

arxiv 2607.28537 v1 pith:2SJO46HM submitted 2026-07-30 cond-mat.str-el cs.LGphysics.comp-ph

Graph Neural Network Force Fields for Spin Dynamics in Metallic Magnets

Ali Rayat , Yunhao Fan , Gia-Wei Chern This is my paper
classification cond-mat.str-el cs.LGphysics.comp-ph
keywords graph neural networksmagnetic force fieldsitinerant magnetsspin dynamicsLandau-Lifshitz-Gilberts-d modelmachine-learned potentialschiral domain coarsening
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.

In metallic magnets, spin motion is set by forces that come from itinerant electrons, so realistic spin simulations usually recompute the electronic state at every step. This paper shows that a graph neural network can learn the effective magnetic energy functional directly from those electronic calculations and then supply the spin torques by automatic differentiation. The network is built from spin-rotation-invariant bond and triangle features and updated by message passing on the magnetic lattice, so it respects the main symmetries of weak-spin-orbit itinerant magnets. On collinear Néel, coplanar 120°, and noncoplanar tetrahedral orders, the learned torques match electronic references closely, and the resulting Landau–Lifshitz dynamics—including structure factors, correlations, and chiral-domain coarsening—tracks direct electronic simulations. If the approach holds more broadly, large-scale nonequilibrium magnetism in metals becomes accessible at a fraction of the usual electronic cost.

Core claim

A symmetry-preserving graph neural network trained on electronic s–d calculations learns the configuration-dependent magnetic energy of itinerant magnets well enough that torques from automatic differentiation reproduce electronically generated torques, and Landau–Lifshitz dynamics driven by those torques agrees with direct electronic simulations for collinear, noncollinear, and noncoplanar orders—including chiral-domain coarsening.

What carries the argument

GNN magnetic force field: lattice graph with SO(3)- and time-reversal-invariant inputs (bond spin products and triangle chirality/alignment invariants), shell-aware message passing on nodes and edges, site-energy readout summed to a total energy, and torques from autodiff of that energy used in LLG dynamics.

Load-bearing premise

Electrons are assumed to stay equilibrated with the instantaneous classical spins, so a single static energy landscape fully sets the torques, and spin-orbit coupling is weak enough that global spin rotations remain a symmetry.

What would settle it

Run the same thermal-quench Landau–Lifshitz trajectories with direct electronic fields (exact diagonalization or kernel polynomial method) and with the trained GNN; a clear mismatch in local torques, structure-factor growth, spin correlations, or chiral-domain size L(t) on Néel, 120°, or tetrahedral benchmarks would falsify the claim.

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

If this is right

  • Repeated electronic solves can be replaced by a trained GNN during long spin-dynamics runs in weak-anisotropy itinerant magnets.
  • Collinear, coplanar, and noncoplanar textures—including Z2 chiral domains with L(t)∼t coarsening—can be simulated at larger length and time scales than direct electronic methods allow.
  • The same force-field philosophy can be retrained when better electronic solvers or multi-orbital models supply the labels, without changing the graph readout structure.
  • Extending the architecture to intertwined spin–lattice symmetries is the stated route toward materials with strong spin-orbit coupling.

Where Pith is reading between the lines

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

  • Training on quench trajectories may under-sample rare textures (skyrmions, multi-Q defects), so transfer to driven or topological dynamics would need targeted data, not just more of the same relaxations.
  • Because the loss is on torques rather than absolute energies, thermodynamically derived quantities (free-energy differences, barrier heights) could be less reliable than dynamical trajectories even when LLG looks correct.
  • Shell-limited message passing implies a finite interaction range; systems with truly macroscopic RKKY tails may need deeper graphs or longer-range edges than the benchmarks used here.

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 / 5 minor

Summary. The manuscript introduces a graph neural network (GNN) magnetic force-field framework that learns an effective SO(3)-invariant magnetic energy functional for itinerant spin dynamics directly from electronic s–d calculations. Node and edge features are initialized from bond correlations and plaquette invariants (χ², Λ), refined by shell-aware message passing, and read out into a site-decomposed energy from which torques are obtained by automatic differentiation and used in stochastic LLG dynamics. The method is benchmarked on three adiabatic s–d settings—square-lattice Néel, triangular 120°, and noncoplanar tetrahedral order—showing high torque accuracy (R² ≈ 0.96–0.9995) and close agreement with ED/KPM–LLG reference trajectories for structure factors, spin correlations, chiral-domain morphology, L(t)∼t coarsening, and dynamical scaling collapse.

Significance. If the reported matches hold, the work provides a practical and physically motivated route to large-scale nonequilibrium spin dynamics in metallic magnets without repeated electronic solves, analogous to ML interatomic potentials. Strengths include symmetry-preserving plaquette features tied to Berry-phase geometry, torque-supervised training with autodiff-consistent fields, and stringent collective benchmarks (especially faceted chiral-domain coarsening and scaling collapse) against ED/KPM references across collinear, coplanar, and noncoplanar orders. The adiabatic, weak-SOC scope is stated clearly. The contribution is incremental relative to prior descriptor-based magnetic force fields from related work, but the GNN formulation and multi-order dynamical validation are a solid step for the field.

major comments (3)
  1. [Sec. III.C, Table I] Sec. III.C and Table I: the text claims a “transferable magnetic force-field framework,” yet three separately optimized models are trained (distinct node/edge widths, depths, learning rates, and 4.4M–8.4M parameters). No cross-phase or out-of-distribution transfer test is reported (e.g., a model trained on 120° data evaluated on tetrahedral configurations, or a single shared architecture). Either demonstrate cross-texture transfer or qualify the claim to “a reusable GNN methodology with phase-specific training.”
  2. [Introduction; Sec. III.B–C] Introduction and Sec. III: the manuscript asserts that the GNN is “more flexible and systematically improvable” than prior descriptor-based magnetic force fields [40–46], including related work on the same s–d benchmarks and chiral coarsening. No head-to-head torque or dynamics comparison on identical datasets/metrics is provided. A quantitative baseline (even against the authors’ earlier descriptor models on Néel/120°/tetrahedral) is needed to substantiate the architectural advantage beyond qualitative framing.
  3. [Sec. II.C, Eq. (14)] Eq. (14) and Sec. II.C: training supervises only torques τ_i = S_i × H_i, leaving the longitudinal field (and thus absolute energy differences) underconstrained. Dynamics tests support trajectory fidelity, but the paper still markets an “energy functional.” Please report held-out energy and/or full-field errors (or show that relative energies along trajectories match ED/KPM) so readers can judge whether E_θ is a faithful surrogate or only a torque engine.
minor comments (5)
  1. [Appendix B–C] No code, trained weights, or dataset release is mentioned. For a methods paper whose central claim is empirical surrogate fidelity, a repository or supplementary data statement would substantially improve reproducibility.
  2. [Table II; Appendix A] Table II: microscopic s–d parameters (t_ij, J, filling, electronic temperature) used to generate labels are not listed in the main text or tables; only LLG α and Δt appear. Please state them explicitly.
  3. [Figs. 2–4; Appendix C] Figs. 2–4: parity plots and error histograms are clear, but axis units/normalization for torques and the precise train/test split protocol (trajectory-wise vs configuration-wise) should be stated to rule out leakage across correlated snapshots.
  4. [Sec. III heading; figure captions] Minor typography: “V ALIDA TION”, “N´ eel”, and occasional spacing artifacts in headings; unify “ED–LLG” vs “ED-LLG” hyphenation.
  5. [Fig. 1; Table I] Fig. 1 message-passing schematic is helpful; defining NV, NE and the number of shells r used in production models in the caption or Table I would aid reimplementation.

Circularity Check

0 steps flagged

No significant circularity: supervised GNN surrogate trained on external electronic torque labels, validated on held-out torques and integrated dynamics.

full rationale

The paper’s load-bearing chain is standard ML force-field methodology, not a closed definitional loop. Reference torques and fields are generated from an independent microscopic solver (s–d Hamiltonian via ED or KPM and Hellmann–Feynman; Eq. (15), App. A), then used as supervised labels. The GNN learns a scalar energy E[{S_i}] whose automatic derivatives yield torques; training minimizes MSE against those external labels (Eq. 14), with explicit train/test splits (App. B–C). Reported R²/MSE/MAE and parity plots are therefore out-of-sample surrogate accuracy, not quantities fixed by construction from the fit. Nonequilibrium benchmarks (structure factors, spin correlations, chiral-domain L(t) and dynamical scaling; Figs. 2–5) compare integrated GNN–LLG trajectories to direct ED/KPM–LLG on the same model—the usual force-field test that the learned landscape remains faithful under differentiation and time integration. Self-citations to the authors’ prior ML-magnetism and coarsening work supply method context and a previously reported L(t)∼t law that is re-checked here against direct electronic runs in this manuscript; they do not replace the electronic labels or force the central accuracy claims. No step reduces a claimed prediction to its own fitted input or to an unverified self-citation uniqueness theorem. Scope limits (adiabatic Hellmann–Feynman functional, weak SOC) are stated physics assumptions, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 1 invented entities

The central claim rests on standard condensed-matter modeling choices (classical spins, adiabatic electrons, s–d Hamiltonian), standard ML training practice, and a large set of architecture hyperparameters chosen per magnetic phase. No new physical particle or force is postulated; the 'entity' is the learned surrogate energy. Load-bearing physics assumptions are the adiabatic Hellmann–Feynman torque picture and approximate global SO(3) spin symmetry without SOC.

free parameters (6)
  • Per-phase GNN width schedules (node/edge channel lists) = e.g. Néel nodes [1024…256]; tetrahedral [1280…256]
    Table I uses different node/edge feature size towers for Néel, 120°, and tetrahedral models; these capacities are chosen/tuned rather than derived.
  • Learning rate, weight decay, batch size, epoch count = lr 5e-4 or 3e-4; wd 1e-3; batch 6–8; 100–120 epochs
    Training hyperparameters selected per benchmark (Table I); affect achieved torque errors.
  • Message-passing depth / shell weighting networks φ_V, φ_E
    Gating MLPs and number of layers control receptive field; not fixed by theory.
  • Readout MLP architecture ψ = [256, 128, 64, 16, 1]
    Local energy head [256,128,64,16,1] is a design choice shared across phases.
  • LLG damping α and integration timestep Δt in dataset generation = α=0.1 or 0.075; Δt=0.005–0.025
    Table II: α and Δt differ by phase and shape the sampled configuration distribution used as training data.
  • s–d microscopic parameters (t_ij, J) and filling/temperature in electronic solves
    Define the target energy landscape; treated as fixed inputs of the benchmark Hamiltonians (not always numerically listed in the main text).
axioms (7)
  • domain assumption Adiabatic approximation: electrons stay in instantaneous equilibrium with classical spins, yielding torques from a static free-energy functional via Hellmann–Feynman.
    Stated in Introduction and Appendix A; required for any energy-based force field of this type.
  • domain assumption Localized moments may be treated as classical unit spins evolving under stochastic LLG.
    Throughout Sec. II.C and benchmarks; standard atomistic spin dynamics premise.
  • domain assumption In the target materials regime, global SO(3) spin rotation symmetry holds and spin–orbit-induced anisotropies are weak or subdominant.
    Sec. II opening; justifies invariant (not fully equivariant spin–lattice) features and limits claimed scope.
  • domain assumption The one-band s–d Hamiltonian is a sufficient generator of electronically mediated interactions for the benchmark orders considered.
    Sec. III.A Eq. (15); all labels come from this model (ED or KPM).
  • domain assumption Automatic differentiation of a site-summed scalar energy produces torques sufficiently accurate for long-time integrated dynamics if torque MSE is minimized.
    Sec. II.C training loss (14); standard energy-conserving ML-FF logic applied to spins.
  • standard math Standard message-passing GNN universal-approximation / expressivity assumptions on graphs with shell-wise aggregation.
    Implicit in Sec. II.A architecture; no new theorem proved.
  • domain assumption KPM Chebyshev reconstruction yields reference density matrices accurate enough to serve as ground-truth torques on 96×96 lattices.
    Appendix A; tetrahedral labels depend on KPM fidelity controlled in cited prior work.
invented entities (1)
  • GNN magnetic force-field surrogate E_θ[{S_i}] with χ²/Λ plaquette node features no independent evidence
    purpose: Replace repeated electronic solves by a learned energy functional for torque evaluation in LLG.
    Methodological construct, not a new physical field; existence is demonstrated only inside the trained s–d benchmarks.

pith-pipeline@v1.2.0-daily-grok45 · 48544 in / 4092 out tokens · 88472 ms · 2026-07-31T04:17:32.881054+00:00 · methodology

0 comments
read the original abstract

Metallic magnets exhibit complex spin dynamics governed by electronically generated interactions. Predictive simulations of such dynamics typically require repeated solutions of an underlying electronic problem throughout the time evolution, creating a major computational bottleneck. Here we introduce a graph neural network (GNN) magnetic force-field framework that learns the effective magnetic energy functional governing itinerant spin dynamics directly from electronic calculations. Conceptually analogous to machine-learned interatomic potentials, the proposed framework enables efficient evaluation of spin torques while capturing the nonlinear and spatially extended interactions generated by itinerant electrons. We benchmark the method on representative metallic magnetic systems exhibiting collinear, noncollinear, and noncoplanar magnetic order. The learned force fields accurately reproduce electronically generated spin torques and yield nonequilibrium spin dynamics in excellent agreement with direct electronic simulations. Our results establish graph neural networks as a powerful framework for machine-learned magnetic force fields, providing a pathway toward predictive large-scale simulations of nonequilibrium magnetism across multiple length and time scales.

Figures

Figures reproduced from arXiv: 2607.28537 by Ali Rayat, Gia-Wei Chern, Yunhao Fan.

Figure 1
Figure 1. Figure 1: FIG. 1. Schematic illustration of the graph-neural-network architecture. (a) The magnetic lattice is represented as a graph [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Benchmark of the GNN force field for the square-lattice N´eel antiferromagnet. (a) Comparison of torques [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Benchmark of the graph-neural-network force field for the triangular-lattice 120 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Benchmark of the graph-neural-network force field for the noncoplanar tetrahedral state on the triangular lattice. (a) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Further benchmark of the GNN force field for the non [PITH_FULL_IMAGE:figures/full_fig_p009_5.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

80 extracted references · 10 linked inside Pith

  1. [1]

    Nagaosa and Y

    N. Nagaosa and Y. Tokura, Topological properties and dynamics of magnetic skyrmions, Nature Nanotechnol- ogy8, 899 (2013)

    2013

  2. [2]

    ˇZuti´ c, J

    I. ˇZuti´ c, J. Fabian, and S. Das Sarma, Spintronics: Fun- damentals and applications, Reviews of Modern Physics 76, 323 (2004)

    2004

  3. [3]

    Chappert, A

    C. Chappert, A. Fert, and F. N. Van Dau, The emergence of spin electronics in data storage, Nature Materials6, 813 (2007)

    2007

  4. [4]

    Jungwirth, X

    T. Jungwirth, X. Marti, P. Wadley, and J. Wunderlich, Antiferromagnetic spintronics, Nature Nanotechnology 11, 231 (2016)

    2016

  5. [5]

    Tokura, M

    Y. Tokura, M. Kawasaki, and N. Nagaosa, Emergent functions of quantum materials, Nature Physics13, 1056 (2017)

    2017

  6. [6]

    vortices

    A. N. Bogdanov and D. Yablonskii, Thermodynamically 13 stable “vortices” in magnetically ordered crystals. the mixed state of magnets, Zh. Eksp. Teor. Fiz95, 178 (1989)

    1989

  7. [7]

    U. K. R¨ oßler, A. N. Bogdanov, and C. Pfleiderer, Sponta- neous skyrmion ground states in magnetic metals, Nature 442, 797 (2006)

    2006

  8. [8]

    M¨ uhlbauer, B

    S. M¨ uhlbauer, B. Binz, F. Jonietz, C. Pfleiderer, A. Rosch, A. Neubauer, R. Georgii, and P. B¨ oni, Skyrmion lattice in a chiral magnet, Science323, 915 (2009)

    2009

  9. [9]

    X. Z. Yu, N. Kanazawa, Y. Onose, K. Kimoto, W. Z. Zhang, S. Ishiwata, Y. Matsui, and Y. Tokura, Near room-temperature formation of a skyrmion crystal in thin-films of the helimagnet fege, Nature Materials10, 106 (2011)

    2011

  10. [10]

    R. F. L. Evans, W. J. Fan, P. Chureemart, T. A. Ostler, M. O. A. Ellis, and R. W. Chantrell, Atomistic spin model simulations of magnetic nanomaterials, Journal of Physics: Condensed Matter26, 103202 (2014)

    2014

  11. [11]

    Eriksson, A

    O. Eriksson, A. Bergman, L. Bergqvist, and J. Hellsvik, Atomistic spin dynamics: Foundations and applications, Oxford University Press (2017)

    2017

  12. [12]

    Car and M

    R. Car and M. Parrinello, Unified approach for molecular dynamics and density-functional theory, Physical Review Letters55, 2471 (1985)

    1985

  13. [13]

    Marx and J

    D. Marx and J. Hutter,Ab Initio Molecular Dynamics: Basic Theory and Advanced Methods(Cambridge Uni- versity Press, 2009)

    2009

  14. [14]

    Behler and M

    J. Behler and M. Parrinello, Generalized neural-network representation of high-dimensional potential-energy sur- faces, Phys. Rev. Lett.98, 146401 (2007)

    2007

  15. [15]

    A. P. Bart´ ok, M. C. Payne, R. Kondor, and G. Cs´ anyi, Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons, Phys. Rev. Lett.104, 136403 (2010)

    2010

  16. [16]

    Z. Li, J. R. Kermode, and A. De Vita, Molecular dy- namics with on-the-fly machine learning of quantum- mechanical forces, Phys. Rev. Lett.114, 096405 (2015)

    2015

  17. [17]

    A. V. Shapeev, Moment tensor potentials: A class of sys- tematically improvable interatomic potentials, Multiscale Modeling & Simulation14, 1153 (2016)

    2016

  18. [18]

    V. Botu, R. Batra, J. Chapman, and R. Ramprasad, Ma- chine learning force fields: Construction, validation, and outlook, The Journal of Physical Chemistry C121, 511 (2017)

    2017

  19. [19]

    J. S. Smith, O. Isayev, and A. E. Roitberg, Ani-1: an extensible neural network potential with dft accuracy at force field computational cost, Chem. Sci.8, 3192 (2017)

    2017

  20. [20]

    Zhang, J

    L. Zhang, J. Han, H. Wang, R. Car, and W. E, Deep potential molecular dynamics: A scalable model with the accuracy of quantum mechanics, Phys. Rev. Lett.120, 143001 (2018)

    2018

  21. [21]

    Lubbers, J

    N. Lubbers, J. S. Smith, and K. Barros, Hierarchical modeling of molecular energies using a deep neural net- work, The Journal of Chemical Physics148, 241715 (2018)

    2018

  22. [22]

    V. L. Deringer, M. A. Caro, and G. Csanyi, Machine learning interatomic potentials as emerging tools for ma- terials science, Advanced Materials31, 1902765 (2019)

    2019

  23. [23]

    Chmiela, A

    S. Chmiela, A. Tkatchenko, H. E. Sauceda, I. Poltavsky, K. T. Sch¨ utt, and K.-R. M¨ uller, Machine learning of ac- curate energy-conserving molecular force fields, Science Advances3, e1603015 (2017)

    2017

  24. [24]

    Chmiela, H

    S. Chmiela, H. E. Sauceda, K.-R. M¨ uller, and A. Tkatchenko, Towards exact molecular dynamics sim- ulations with machine-learned force fields, Nature Com- munications9, 3887 (2018)

    2018

  25. [25]

    H. E. Sauceda, M. Gastegger, S. Chmiela, K.-R. M¨ uller, and A. Tkatchenko, Molecular force fields with gradient- domain machine learning (GDML): Comparison and syn- ergies with classical force fields, The Journal of Chemical Physics153, 124109 (2020)

    2020

  26. [26]

    J. Ma, P. Zhang, Y. Tan, A. W. Ghosh, and G.-W. Chern, Machine learning electron correlation in a disor- dered medium, Phys. Rev. B99, 085118 (2019)

    2019

  27. [27]

    Zhang, S

    P. Zhang, S. Zhang, and G.-W. Chern, Descriptors for machine learning model of generalized force field in con- densed matter systems (2022), arXiv:2201.00798 [cond- mat.str-el]

    Pith/arXiv arXiv 2022

  28. [28]

    Zhang, P

    S. Zhang, P. Zhang, and G.-W. Chern, Anomalous phase separation in a correlated electron system: Machine- learning enabled large-scale kinetic monte carlo simula- tions, Proceedings of the National Academy of Sciences 119, e2119957119 (2022)

    2022

  29. [29]

    Cheng, S

    C. Cheng, S. Zhang, and G.-W. Chern, Machine learn- ing for phase ordering dynamics of charge density waves, Phys. Rev. B108, 014301 (2023)

    2023

  30. [30]

    Y.-H. Liu, S. Zhang, P. Zhang, T.-K. Lee, and G.- W. Chern, Machine learning predictions for local elec- tronic properties of disordered correlated electron sys- tems, Phys. Rev. B106, 035131 (2022)

    2022

  31. [31]

    Z. Tian, S. Zhang, and G.-W. Chern, Machine learning for structure-property mapping of Ising models: Scala- bility and limitations, Phys. Rev. E108, 065304 (2023)

    2023

  32. [32]

    Ghosh, S

    S. Ghosh, S. Zhang, C. Cheng, and G.-W. Chern, Kinet- ics of orbital ordering in cooperative jahn-teller models: Machine-learning enabled large-scale simulations, Phys. Rev. Mater.8, 123602 (2024)

    2024

  33. [33]

    Fan and G.-W

    Y. Fan and G.-W. Chern, Equivariant neural net- works for force-field models of lattice systems (2026), arXiv:2601.04104 [cond-mat.str-el]

    arXiv 2026

  34. [34]

    Eckhoff and J

    M. Eckhoff and J. Behler, High-dimensional neural network potentials for magnetic systems using spin- dependent atom-centered symmetry functions, npj Com- putational Materials7, 170 (2021)

    2021

  35. [35]

    A. S. Kotykhov, K. Gubaev, V. Sotskov, C. Tantardini, M. Hodapp, A. V. Shapeev, and I. S. Novikov, Fitting to magnetic forces improves the reliability of magnetic mo- ment tensor potentials, Computational Materials Science 245, 113331 (2024)

    2024

  36. [36]

    H. Yu, Y. Zhong, L. Hong, C. Xu, W. Ren, X. Gong, and H. Xiang, Spin-dependent graph neural network potential for magnetic materials, Phys. Rev. B109, 144426 (2024)

    2024

  37. [37]

    H. Yu, B. Liu, Y. Zhong, L. Hong, J. Ji, C. Xu, X. Gong, and H. Xiang, Physics-informed time-reversal equivariant neural network potential for magnetic materials, Phys. Rev. B110, 104427 (2024)

    2024

  38. [38]

    Huang, T

    Z. Huang, T. Yang, H. Liu, and B. Xu, Simultaneous optimization of lattice and spin configurations in atomic scale simulation of magnetic materials, Phys. Rev. B111, 134412 (2025)

    2025

  39. [39]

    W. Xu, R. Y. Sanspeur, A. Kolluru, B. Deng, P. Har- rington, S. Farrell, K. Reuter, and J. R. Kitchin, Spin- informed universal graph neural networks for simulating magnetic ordering, Proceedings of the National Academy of Sciences122, e2422973122 (2025)

    2025

  40. [40]

    Zhang and G.-W

    P. Zhang and G.-W. Chern, Arrested phase separation in double-exchange models: Large-scale simulation en- 14 abled by machine learning, Phys. Rev. Lett.127, 146401 (2021)

    2021

  41. [41]

    Zhang and G.-W

    P. Zhang and G.-W. Chern, Machine learning nonequilib- rium electron forces for spin dynamics of itinerant mag- nets, npj Computational Materials9, 32 (2023)

    2023

  42. [42]

    Cheng, S

    X. Cheng, S. Zhang, P. C. H. Nguyen, S. Azarfar, G.-W. Chern, and S. S. Baek, Convolutional neural networks for large-scale dynamical modeling of itinerant magnets, Phys. Rev. Res.5, 033188 (2023)

    2023

  43. [43]

    Y. Fan, S. Zhang, and G.-W. Chern, Coarsening of chi- ral domains in itinerant electron magnets: A machine learning force-field approach, Phys. Rev. B110, 245105 (2024)

    2024

  44. [44]

    Tyberg, Y

    A. Tyberg, Y. Fan, and G.-W. Chern, Machine learning force field model for kinetic monte carlo simulations of itinerant ising magnets, Phys. Rev. B111, 235132 (2025)

    2025

  45. [45]

    Chern, Y

    G.-W. Chern, Y. Fan, S. Zhang, and P. Zhang, Machine- learning force-field models for dynamical simulations of metallic magnets, AIP Advances16, 025031 (2026)

    2026

  46. [46]

    Chern, Y

    G.-W. Chern, Y. Fan, S. Zhang, and P. Zhang, Machine- learning modeling of magnetization dynamics in itinerant magnets, Journal of Magnetism and Magnetic Materials 628, 173679 (2026)

    2026

  47. [47]

    Behler, Atom-centered symmetry functions for con- structing high-dimensional neural network potentials, The Journal of Chemical Physics134, 074106 (2011)

    J. Behler, Atom-centered symmetry functions for con- structing high-dimensional neural network potentials, The Journal of Chemical Physics134, 074106 (2011)

    2011

  48. [48]

    L. M. Ghiringhelli, J. Vybiral, S. V. Levchenko, C. Draxl, and M. Scheffler, Big data of materials science: Criti- cal role of the descriptor, Phys. Rev. Lett.114, 105503 (2015)

    2015

  49. [49]

    A. P. Bart´ ok, R. Kondor, and G. Cs´ anyi, On representing chemical environments, Phys. Rev. B87, 184115 (2013)

    2013

  50. [50]

    Himanen, M

    L. Himanen, M. O. J¨ ager, E. V. Morooka, F. Federici Canova, Y. S. Ranawat, D. Z. Gao, P. Rinke, and A. S. Foster, Dscribe: Library of descriptors for machine learn- ing in materials science, Computer Physics Communica- tions247, 106949 (2020)

    2020

  51. [51]

    Huo and M

    H. Huo and M. Rupp, Unified representation of molecules and crystals for machine learning, Machine Learning: Science and Technology3, 045017 (2022)

    2022

  52. [52]

    Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Phys

    R. Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Phys. Rev. B99, 014104 (2019)

    2019

  53. [53]

    Scarselli, M

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini, The graph neural network model, IEEE Transactions on Neural Networks20, 61 (2009)

    2009

  54. [54]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, Neural message passing for quantum chem- istry, inProceedings of the 34th International Conference on Machine Learning, Vol. 70 (PMLR, 2017) pp. 1263– 1272, arXiv:1704.01212

    Pith/arXiv arXiv 2017

  55. [55]

    W. L. Hamilton, R. Ying, and J. Leskovec, Inductive representation learning on large graphs, inAdvances in Neural Information Processing Systems, Vol. 30 (2017) arXiv:1706.02216

    Pith/arXiv arXiv 2017

  56. [56]

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, How powerful are graph neural networks?, inInternational Conference on Learning Representations (ICLR)(2019) arXiv:1810.00826

    Pith/arXiv arXiv 2019

  57. [57]

    Maron, H

    H. Maron, H. Ben-Hamu, N. Shamir, and Y. Lipman, In- variant and equivariant graph networks, inInternational Conference on Learning Representations (ICLR)(2019) arXiv:1812.09902

    Pith/arXiv arXiv 2019

  58. [58]

    Xie and J

    T. Xie and J. C. Grossman, Crystal graph convolutional neural networks for an accurate and interpretable predic- tion of material properties, Phys. Rev. Lett.120, 145301 (2018)

    2018

  59. [59]

    K. T. Sch¨ utt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, and K.-R. M¨ uller, Schnet – a deep learn- ing architecture for molecules and materials, The Journal of Chemical Physics148, 241722 (2018)

    2018

  60. [60]

    Choudhary and B

    K. Choudhary and B. DeCost, Atomistic line graph neu- ral network for improved materials property predictions, npj Computational Materials7, 185 (2021)

    2021

  61. [61]

    M. Dai, M. F. Demirel, Y. Liang, and J.-M. Hu, Graph neural networks for an accurate and interpretable pre- diction of the properties of polycrystalline materials, npj Computational Materials7, 103 (2021)

    2021

  62. [62]

    Reiser, M

    P. Reiser, M. Neubert, A. Eberhard, L. Torresi, C. Zhou, C. Shao, H. Metni, C. van Hoesel, H. Schopmans, T. Sommer, and P. Friederich, Graph neural networks for materials science and chemistry, Communications Mate- rials3, 93 (2022)

    2022

  63. [63]

    Batatia, D

    I. Batatia, D. P. Kov´ acs, G. N. C. Simm, C. Ortner, and G. Cs´ anyi, Mace: Higher order equivariant mes- sage passing neural networks for fast and accurate force fields, inAdvances in Neural Information Processing Sys- tems, Vol. 35 (2022) pp. 11423–11436, arXiv:2206.07697 [stat.ML]

    Pith/arXiv arXiv 2022

  64. [64]

    T. S. Cohen and M. Welling, Group equivariant con- volutional networks, inProceedings of the 33rd Interna- tional Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 48 (PMLR, 2016) pp. 2990–2999, arXiv:1602.07576 [cs.LG]

    Pith/arXiv arXiv 2016

  65. [65]

    T. S. Cohen, M. Geiger, M. Weiler, and M. Welling, A general theory of equivariant cnns on homogeneous spaces, inAdvances in Neural Information Processing Systems, Vol. 31 (2018) arXiv:1811.02017 [cs.LG]

    Pith/arXiv arXiv 2018

  66. [66]

    Weiler, M

    M. Weiler, M. Geiger, M. Welling, W. Boomsma, and T. S. Cohen, 3d steerable cnns: Learning rotationally equivariant features in volumetric data, inAdvances in Neural Information Processing Systems, Vol. 31 (Curran Associates, Inc., 2018) arXiv:1807.02547 [cs.LG]

    Pith/arXiv arXiv 2018

  67. [67]

    Kondor, The principles behind equivariant neural net- works for physics and chemistry, Proceedings of the Na- tional Academy of Sciences122, e2415656122 (2025)

    R. Kondor, The principles behind equivariant neural net- works for physics and chemistry, Proceedings of the Na- tional Academy of Sciences122, e2415656122 (2025)

    2025

  68. [68]

    Batzner, A

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials, Nature Communications13, 2453 (2022)

    2022

  69. [69]

    Kaba and S

    S.-O. Kaba and S. Ravanbakhsh, Equivariant networks for crystal structures, inAdvances in Neural Informa- tion Processing Systems, Vol. 35 (2022) pp. 300–314, arXiv:2211.15420 [cs.LG]

    Pith/arXiv arXiv 2022

  70. [70]

    Musaelian, S

    A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen, M. Kornbluth, and B. Kozinsky, Learning local equivariant representations for large-scale atomistic dy- namics, Nature Communications14, 579 (2023)

    2023

  71. [71]

    X. Gong, H. Li, N. Zou, R. Xu, W. Duan, and Y. Xu, General framework for e(3)-equivariant neural network representation of density functional theory hamiltonian, Nature Communications14, 2848 (2023)

    2023

  72. [72]

    Batatia, S

    I. Batatia, S. Batzner, D. P. Kov´ acs, A. Musaelian, G. N. C. Simm, R. Drautz, C. Ortner, B. Kozinsky, and G. Cs´ anyi, The design space of e(3)-equivariant atom- centred interatomic potentials, Nature Machine Intelli- gence7, 56 (2025). 15

    2025

  73. [73]

    Z. Yang, X. Wang, Y. Li, Q. Lv, C. Y.-C. Chen, and L. Shen, Efficient equivariant model for machine learning interatomic potentials, npj Computational Materials11, 49 (2025)

    2025

  74. [74]

    Zener, Interaction between the d shells in the transi- tion metals, Physical Review81, 440 (1951)

    C. Zener, Interaction between the d shells in the transi- tion metals, Physical Review81, 440 (1951)

    1951

  75. [75]

    de Gennes, Effects of double exchange in magnetic crystals, Physical Review118, 141 (1960)

    P.-G. de Gennes, Effects of double exchange in magnetic crystals, Physical Review118, 141 (1960)

    1960

  76. [76]

    P. W. Anderson, Localized magnetic states in metals, Phys. Rev.124, 41 (1961)

    1961

  77. [77]

    Weiße, G

    A. Weiße, G. Wellein, A. Alvermann, and H. Fehske, The kernel polynomial method, Rev. Mod. Phys.78, 275 (2006)

    2006

  78. [78]

    Barros and Y

    K. Barros and Y. Kato, Efficient langevin simulation of coupled classical fields and fermions, Phys. Rev. B88, 235101 (2013)

    2013

  79. [79]

    Wang, G.-W

    Z. Wang, G.-W. Chern, C. D. Batista, and K. Bar- ros, Gradient-based stochastic estimation of the density matrix, The Journal of Chemical Physics148, 094107 (2018)

    2018

  80. [80]

    Ozawa, S

    R. Ozawa, S. Hayami, K. Barros, and Y. Motome, Shape of magnetic domain walls formed by coupling to mobile charges, Phys. Rev. B96, 094417 (2017)

    2017