Pith. sign in

REVIEW 3 major objections 5 minor 40 references

A single equivariant neural network can predict Born effective charges together with energy, forces, and stress at near-DFT accuracy, making charge-aware molecular dynamics under electric fields practical.

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 00:55 UTC pith:DZJHGWDF

load-bearing objection SevenNet-Polar is a genuine, well-engineered advance for charge-aware MLIPs; the headline 'no degradation from multitask training' is real but under-supported by a confounded comparison. the 3 major comments →

arxiv 2607.14827 v3 pith:DZJHGWDF submitted 2026-07-16 cond-mat.mtrl-sci cond-mat.dis-nnphysics.comp-ph

SevenNet-Polar for MultiTask Prediction of Energy, Forces, Stress, and Born Effective Charges: Development and Application to ZrO₂, Li₃PO₄, and Perovskites

Anh Khoa Augustin Lu , Shungo Arai , Yutack Park , Seungwu Han , Tsuyoshi Miyazaki , Satoshi Watanabe This is my paper
classification cond-mat.mtrl-sci cond-mat.dis-nnphysics.comp-ph
keywords Born effective chargesequivariant graph neural networksmultitask learningmachine learning interatomic potentialselectric-field molecular dynamicsZrO2Li3PO4perovskites
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 claims that the Born effective charge tensor—a per-atom 3×3 quantity that determines how atoms respond to electric fields—can be predicted accurately by an equivariant graph neural network, and that this prediction can be folded into a single multitask model that also predicts energy, forces, and stress. On ZrO2 and Li3PO4, the best multitask model reaches errors of 1.0 meV/atom (energy), 11.9 meV/Å (forces), 0.05 GPa (stress), and 0.0029 e (Born charges). BEC-only models trained on three materials reach 0.0043 e. The paper also shows that BEC errors follow their own power-law scaling, more slowly than energy/force/stress errors, and that the models transfer to structures absent from training, such as NEB trajectories and a grain boundary. If these numbers hold, charge-aware molecular dynamics under applied electric fields becomes practical on consumer GPUs.

Core claim

On the paper's own terms, the central discovery is that a general Cartesian rank-2 tensor such as the Born effective charge can be learned by decomposing it into its O(3)-irreducible parts—a scalar trace, an antisymmetric pseudo-vector, and a symmetric traceless rank-2 tensor—and predicting these parts with an equivariant message-passing network whose angular resolution goes at least to lmax=2. This encoding lets the same architecture output per-atom BEC tensors together with energy, forces, and stress in one forward pass. The paper's best multitask model reports BEC RMSE of 0.0029 e on ZrO2+Li3PO4, with energy/force/stress errors comparable to state-of-the-art potentials, and the authors co

What carries the argument

The central object is the irreducible representation of the BEC tensor, Γ_BEC = 1×0e + 1×1e + 1×2e, which encodes the nine Cartesian components as equivariant features that rotate correctly. The network uses message passing with a cutoff of 6 Å and 4–5 layers (receptive field up to roughly 30 Å), lmax between 2 and 4, and a readout head that projects back to Cartesian tensors. FlashTP accelerates the expensive Clebsch-Gordan tensor products, reducing overhead so that large-system inference remains feasible.

Load-bearing premise

The headline claim that multitask training does not hurt BEC accuracy rests on comparing two models that differ in both training data and number of tasks, so the same conclusion might not survive a controlled comparison on identical data.

What would settle it

Train SevenNet-PS and SevenNet-PM with the same ZrO2+Li3PO4 data (no perovskites) and identical hyperparameters; if the multitask model's BEC RMSE is meaningfully higher than the specialized model's on the same held-out test set, the 'no degradation' claim is false. A second check: a BEC-only model with lmax=0 should show near-zero scaling exponent, as the paper reports, confirming that angular resolution is the mechanism carrying the BEC accuracy.

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

If this is right

  • A single multitask model can drive NEB and molecular dynamics simulations under an electric field without combining separate interatomic potentials and BEC predictors.
  • BEC values along an oxygen-migration NEB path in defective ZrO2 stay close to DFPT, and a Σ5(310) grain boundary is predicted with RMSE 0.021 e, indicating transferability to unseen defect environments.
  • Training-set scaling exponents (roughly 0.73 for energy, 0.53 for forces, 0.39 and 0.32 for BEC diagonal and off-diagonal terms) imply that BEC accuracy demands more data than energy or forces for the same relative error reduction.
  • With FlashTP, simulations that include BEC prediction can run at about 1 ns/day for thousands of atoms on a consumer GPU and scale up to 1.5 million atoms on 64 server GPUs.
  • The architecture covers ten chemical elements and can be retrained on external datasets from water, NaCl, MAPbI3, SiO2, and BaTiO3 with errors roughly 50–70% below existing scalar baseline models.

Where Pith is reading between the lines

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

  • The comparison supporting 'multitask training does not degrade BEC' changes two variables at once: SevenNet-PM-L omits perovskite structures that SevenNet-PS-L includes. A cleaner test would train both models on identical data and vary only the output heads.
  • Because off-diagonal BEC components scale slower (α≈0.32) than diagonal ones, a universal BEC model will likely require not just more data but targeted sampling of sheared or disordered environments, or higher-order equivariant features.
  • Since the model returns per-atom BEC tensors, coupling them to an applied field in LAMMPS directly yields electric-field forces; a natural next step is simulating ferroelectric switching or domain-wall dynamics in larger cells than previously feasible.
  • Applying the same irreducible-decomposition readout to the macroscopic dielectric tensor ε∞ would let the framework cover non-analytical phonon corrections, a logical extension the paper mentions as future work.

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. This paper introduces SevenNet-Polar, an extension of the SevenNet equivariant graph neural network architecture that predicts Born effective charge tensors, in addition to energy, forces, and stress. The BEC tensor is decomposed into 0e+1e+2e irreducible representations to preserve O(3) equivariance. The paper reports three sizes (S/M/L) of BEC-only specialized models (SevenNet-PS) trained on ZrO2, Li3PO4, and perovskites, with the largest reaching 0.0043 e RMSE on BECs, and three multitask models (SevenNet-PM) trained on ZrO2 and Li3PO4, with the largest reaching 1.0 meV/atom (energy), 11.9 meV/Å (forces), 0.05 GPa (stress), and 0.0029 e (BEC). It also presents scaling-law exponents showing that BEC components converge slower than energy/forces/stress, a hyperparameter ablation, generalization tests on NEB pathways and a ZrO2 grain boundary, external benchmarks on data sets from Schmiedmayer et al. and Falletta et al., and performance measurements enabled by FlashTP, including 1.5-million-atom simulations on 64 GPUs. The paper claims that multitask training does not degrade BEC accuracy.

Significance. If the claims hold, this is a useful contribution: it combines a principled equivariant readout for rank-2 tensorial properties with a practical multitask framework, and it provides open code, open data, an ASE calculator, and a LAMMPS interface. The external benchmarks and the scaling analysis are also valuable as empirical descriptors of BEC learning difficulty. However, the headline claim that 'BEC accuracy is not degraded by multitask training' is not established by the experiments as presented, because the comparison used to support it changes both the task set and the training data simultaneously. The absence of repeated random splits also makes several numerical comparisons difficult to assess. These issues are fixable with additional controlled experiments and uncertainty reporting, so the central methodology remains defensible.

major comments (3)
  1. [Abstract; §2.3.2, Tables 1–2] The claim that 'BEC accuracy is not degraded by multitask training' is not supported by the evidence shown. The only comparison offered is SevenNet-PM-L (trained on ZrO2+Li3PO4 with energy/force/stress/BEC heads, BEC RMSE 0.0029 e) versus SevenNet-PS-L (trained on ZrO2+Li3PO4+perovskites with only a BEC head, BEC RMSE 0.0043 e). These models differ simultaneously in task count and training-set composition, so the statement in §2.3.2 that the lower RMSE 'originates from the absence of the perovskite structures' is a confounded inference. A controlled experiment is needed: the same architecture and data split trained on identical ZrO2+Li3PO4 data with and without the auxiliary E/F/S heads, ideally over multiple seeds. Without this, the abstract's headline claim should be withdrawn or explicitly qualified.
  2. [§4.6; Tables 1–4] All reported RMSE values and scaling exponents come from a single 80/10/10 random split with no repeated seeds or error bars. This matters for several load-bearing comparisons: the 0.0029 e versus 0.0043 e BEC difference, the distinction between diagonal (α≈0.39) and off-diagonal (α≈0.32) scaling exponents, and the claimed hierarchy of exponents. With a single split, these differences could be within run-to-run noise. Please report mean ± standard deviation over multiple seeds/splits, or otherwise quantify the uncertainty of the main metrics and exponents.
  3. [§2.4.3, Table 4] The external benchmark on the Schmiedmayer et al. data set shows ZrO2 solid BEC RMSE of 41.2 m|e| with SevenNet-PS-M, compared with 8.4 m|e| on the paper's own ZrO2 test set (Table 1). This is a five-fold discrepancy on the same compound and is not discussed in the main text. It directly qualifies the abstract's 'remarkable accuracy' and the paper's 'generalizes robustly' claims. The authors should analyze and discuss this discrepancy — whether it arises from distribution shift, different exchange-correlation functionals, data labeling conventions, or defect content — and temper the generalization claims accordingly.
minor comments (5)
  1. [§2.3.2] The sentence 'It should be noted that the BEC RMSE is lower than that of the BEC-only model' is ambiguous: it refers to SevenNet-PS-L on the three-data-set combined model, not to a same-data single-task baseline. Please state the comparison explicitly.
  2. [§3] There is a typo in the Discussion: 'max = 4' should likely be 'lmax = 4'.
  3. [§2.3.3, Table 3] The table header states that Force and Stress RMSE are reported component-wise, but the table appears to report aggregate values. Please clarify the definition, or change the header if the values are aggregate.
  4. [§4.5, Eq. (2)] The empirical loss weights (w_E=1, w_F=0.1, w_S=1e-6, w_BEC=10) and the factor-9 correction in Eq. (S1) are model choices. It would be useful to state explicitly that the scaling exponents and the multitask comparison may depend on these choices, and ideally to test at least one alternative weighting.
  5. [Fig. 7 caption] The caption notes that for a given image number the DFT and model structures differ. This should be explained in the text: the comparison of barrier heights is between two relaxed paths, not pointwise image energies.

Circularity Check

0 steps flagged

No significant circularity: held-out empirical evaluations and descriptive scaling fits; self-citations are benchmarks, not load-bearing proofs.

full rationale

The paper's core claims (BEC RMSEs, multitask accuracy, NEB/grain-boundary transferability, scaling exponents, speed benchmarks) are all empirical evaluations of trained models against held-out DFT/DFPT references, not derivations from the fitted outputs. The scaling laws (RMSE ∝ N_train^-α, Secs. 2.1, 2.3.1) are descriptive power-law fits to the model's own test RMSE; the paper does not present them as first-principles predictions, so fitting the RMSE that generated them is not circular. The BEC readout (Γ_BEC = 1×0e + 1×1e + 1×2e, Sec. 4.4) is an exact, norm-preserving change of basis for the DFPT labels and is not an assumption smuggled in via citation. Self-citations to SevenNet [9], the ZrO2 dataset [20], and the grain-boundary model [17] are used as architecture/dataset sources and benchmarks; none is invoked as a uniqueness theorem or as proof of the present results, and the models are additionally tested on external datasets (Schmiedmayer et al., Falletta et al.). The abstract's claim that 'BEC accuracy is not degraded by multitask training' rests on a comparison of SevenNet-PM-L (ZrO2+Li3PO4, 0.0029 e) with SevenNet-PS-L (ZrO2+Li3PO4+perovskites, 0.0043 e), which is a confounded comparison (training data and task set vary simultaneously). That is a methodological limitation affecting significance, not a circular step: no quantity in this comparison is defined in terms of the conclusion, and the conclusion is not mathematically forced by the loss function. Under the specified circularity criteria, no step reduces to its own input, so the appropriate score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central claims rest on hand-tuned loss weights and architecture choices, plus standard symmetry/DFT assumptions. No new physical entities are postulated.

free parameters (3)
  • Loss weights (w_E, w_F, w_S, w_BEC) = 1.0, 0.1, 1e-6, 10.0
    Set by hand to balance gradient magnitudes (Section 4.5); directly affect multitask performance trade-offs.
  • Architecture hyperparameters (lmax, L, C, cutoff) = lmax=2/3/4, L=4/5, C=32/64, rc=6.0 Å
    Chosen via ablation (Table 3); define model capacity and receptive field.
  • Scaling exponent fit range = Ntrain in [64,512]
    Power-law exponents are fitted only over this range (Figures 1,3); results would change with range choice.
axioms (6)
  • domain assumption Random 80/10/10 split gives unbiased test error
    Used to report all RMSEs; no leak-prevention or repeated-seed analysis supplied.
  • domain assumption DFPT labels are ground truth
    The model is trained and evaluated against DFPT BEC values; any systematic errors in the reference data propagate to the metrics.
  • standard math BEC tensor decomposes as 1x0e + 1x1e + 1x2e irreducible representations
    Standard group-theoretic decomposition of a Cartesian rank-2 tensor under O(3), used in Section 4.4.
  • standard math Equivariant message passing exactly preserves O(3) equivariance
    Relies on the proven equivariance of the SevenNet/e3nn framework; the BEC readout is also constructed to be equivariant.
  • domain assumption Extended receptive field via L layers x cutoff approximates long-range response
    The paper assumes that 4-5 layers with 6 Å cutoff capture the non-local dielectric response; no systematic convergence test is provided.
  • domain assumption Combining PBE and PBESol labels is harmless for BEC-only training
    The SevenNet-PS models are trained on PBE (ZrO2, Li3PO4) and PBESol (perovskite) data without adjustment; functional mismatch is not discussed.

pith-pipeline@v1.3.0-alltime-deepseek · 25106 in / 11444 out tokens · 88934 ms · 2026-08-02T00:55:42.608591+00:00 · methodology

0 comments
read the original abstract

Accurate prediction of the Born effective charge (BEC) tensor is crucial for modeling materials under electric fields but remains computationally expensive. To bridge this gap, we present SevenNet-Polar, an equivariant graph neural network framework based on the SevenNet architecture for fast and accurate BEC predictions. Our BEC-only predictors can achieve an RMSE as low as 0.0043 e on ZrO$_2$, Li$_3$PO$_4$, and perovskites, despite the presence of high-temperature (up to 2,000 K) and defect-laden training data. Our all-in-one multitask models for predicting energy, forces, stress, and BEC in ZrO$_2$ and Li$_3$PO$_4$ achieve high accuracy with an RMSE of 1.0 meV/atom for energy, 12 meV/angstrom for forces, 0.05 GPa for stress, and 0.0029 e for BEC. BEC accuracy is not degraded by multitask training. Scaling analysis reveals distinct exponents for diagonal and off-diagonal BEC components, both of which exhibit less favorable scaling than energy, force and stress errors. SevenNet-Polar generalizes robustly when tested on scenarios containing structural environments absent from the training set, such as along nudged elastic band (NEB) trajectories or grain boundaries in ZrO$_2$. Accelerated by FlashTP, SevenNet-Polar enables simulations containing up to 1.5 million atoms on multi-GPU supercomputers and up to approximately 15,000 atoms on a single consumer-grade GPU. This makes charge-aware molecular dynamics simulations under electric fields more accessible.

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

40 extracted references · 12 canonical work pages

  1. [1]

    The Journal of Chemical Physics103, 4129–4137 (1995) https://doi.org/10.1063/1.469597

    Blank, T.B., Brown, S.D., Calhoun, A.W., Doren, D.J.: Neural network models of potential energy surfaces. The Journal of Chemical Physics103, 4129–4137 (1995) https://doi.org/10.1063/1.469597

    doi:10.1063/1.469597 1995

  2. [2]

    Chemical Physics Letters 395, 210–215 (2004) https://doi.org/10.1016/j.cplett.2004.07.076

    Lorenz, S., Groß, A., Scheffler, M.: Representing high-dimensional potential-energy surfaces for reactions at surfaces by neural networks. Chemical Physics Letters 395, 210–215 (2004) https://doi.org/10.1016/j.cplett.2004.07.076

    work page doi:10.1016/j.cplett.2004.07.076 2004

  3. [3]

    Physical Review Letters98(14), 146401 (2007) https://doi.org/10.1103/PhysRevLett.98.146401

    Behler, J., Parrinello, M.: Generalized neural-network representation of high- dimensional potential-energy surfaces. Physical Review Letters98(14), 146401 (2007) https://doi.org/10.1103/PhysRevLett.98.146401

    doi:10.1103/physrevlett.98.146401 2007

  4. [4]

    In: Proceedings of the 34th International Conference on Machine Learning, pp

    Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural mes- sage passing for quantum chemistry. In: Proceedings of the 34th International Conference on Machine Learning, pp. 1263–1272. PMLR, ??? (2017)

    2017

  5. [5]

    Advances in Neural Information Processing Systems30 (2017)

    Sch¨ utt, K.T., Kindermans, P.-J., Sauceda, H.E., Chmiela, S., Tkatchenko, A., M¨ uller, K.-R.: Schnet: A continuous-filter convolutional neural network for model- ing quantum interactions. Advances in Neural Information Processing Systems30 (2017)

    2017

  6. [6]

    Nature Communications 13(1), 2453 (2022) https://doi.org/10.1038/s41467-022-29939-5

    Batzner, S., Musaelian, A., Sun, L., Geiger, M., Mailoa, J.P., Kornbluth, M., Molinari, N., Smidt, T.E., Kozinsky, B.: E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications 13(1), 2453 (2022) https://doi.org/10.1038/s41467-022-29939-5

    doi:10.1038/s41467-022-29939-5 2022

  7. [7]

    Advances in Neural Information Processing Systems35, 11423–11436 (2022) 30

    Batatia, I., Kovacs, D.P., Simm, G.N.C., Ortner, C., Cs´ anyi, G.: Mace: Higher order equivariant message passing neural networks for fast and accurate force fields. Advances in Neural Information Processing Systems35, 11423–11436 (2022) 30

    2022

  8. [8]

    The Journal of Chemical Physics163(18), 184110 (2025) https://doi.org/10.1063/5.0297006

    Batatia, I., Benner, P., Chiang, Y., Elena, A.M., Kov´ acs, D.P., Riebesell, J., Advincula, X.R., Asta, M., Avaylon, M., Baldwin, W.J.,et al.: A foundation model for atomistic materials chemistry. The Journal of Chemical Physics163(18), 184110 (2025) https://doi.org/10.1063/5.0297006

    doi:10.1063/5.0297006 2025

  9. [9]

    Digital Discovery3(8), 1588–1598 (2024) https: //doi.org/10.1039/D4DD00041G

    Park, Y.-G., Kim, J., Hwang, S., Park, J.-H., Kwak, S.W., Han, B., Cho, E.S., Jun, Y.-S., Han, S.: Sevennet: a universal equivariant graph neural network for materials and molecules. Digital Discovery3(8), 1588–1598 (2024) https: //doi.org/10.1039/D4DD00041G

    work page doi:10.1039/d4dd00041g 2024

  10. [10]

    Kim, J., You, J., Park, Y., Lim, Y., Kang, Y., Kim, J., Jeon, H., Ju, S., Hong, D., Lee, S.Y., Choi, S., Kim, Y., Lee, J.W., Han, S.: Optimizing cross-domain transfer for universal machine learning interatomic potentials. Nat. Commun.17(3432) (2026) https://doi.org/10.1038/s41467-026-70195-8

    doi:10.1038/s41467-026-70195-8 2026

  11. [11]

    Nature Communications16(1), 10653 (2025) https://doi.org/10.1038/s41467-025-53812-y

    Mazitov, A., Bigi, F., Kellner, M., Pegolo, P., Tisi, D., Fraux, G., Pozdnyakov, S., Loche, P., Ceriotti, M.: Pet-mad as a lightweight universal interatomic potential for advanced materials modeling. Nature Communications16(1), 10653 (2025) https://doi.org/10.1038/s41467-025-53812-y

    work page doi:10.1038/s41467-025-53812-y 2025

  12. [12]

    The Journal of Chemical Physics161(24), 244102 (2024) https://doi.org/10.1063/5.0234287

    Kang, S.: How graph neural network interatomic potentials extrapolate: Role of the message-passing algorithm. The Journal of Chemical Physics161(24), 244102 (2024) https://doi.org/10.1063/5.0234287

    work page doi:10.1063/5.0234287 2024

  13. [13]

    In: Proceedings of the 42nd International Conference on Machine Learning, vol

    Lee, S.Y., Kim, H., Park, Y., Jeong, D., Han, S., Park, Y., Lee, J.W.: Flashtp: Fused, sparsity-aware tensor product for machine learning interatomic potentials. In: Proceedings of the 42nd International Conference on Machine Learning, vol. 267, pp. 33143–33156. PMLR, ??? (2025). https://proceedings.mlr.press/v267/ lee25l.html 31

    2025

  14. [14]

    arXiv preprint arXiv:2207.09453 (2022) https://doi.org/10.48550/arXiv

    Geiger, M., Smidt, T., Miller, B.K., Boomsma, W., Friederich, P., Lapchevskyi, K., Weiler, M., Miller, L., Magar, M., Riedel, F.: e3nn: Euclidean neural net- works. arXiv preprint arXiv:2207.09453 (2022) https://doi.org/10.48550/arXiv. 2207.09453

    Pith/arXiv arXiv doi:10.48550/arxiv.2207.09453 2022

  15. [15]

    Journal of Physics: Condensed Matter29(27), 273002 (2017) https://doi.org/10.1088/1361-648X/aa680e

    Larsen, A.H., Mortensen, J.J., Blomqvist, J., Castelli, I.E., Christensen, R., Du lak, M., Friis, J., Groves, M.N., Hammer, B., Hargus, C., Hermes, E.D., Jennings, P.C., Jensen, P.B., Kermode, J., Kitchin, J.R., Kolsbjerg, E.L., Kubal, J., Kaasbjerg, K., Lysgaard, S., Maronsson, J.B., Maxson, T., Olsen, T., Pastewka, L., Peterson, A., Rostgaard, C., Schiø...

    2017

  16. [16]

    Journal of Computational Physics117(1), 1–19 (1995) https://doi.org/10.1006/jcph.1995

    Plimpton, S.: Fast parallel algorithms for short-range molecular dynamics. Journal of Computational Physics117(1), 1–19 (1995) https://doi.org/10.1006/jcph.1995. 1039

    work page doi:10.1006/jcph.1995 1995

  17. [17]

    e-Journal of Surface Science and Nanotechnology23, 323–327 (2025) https://doi.org/10.1380/ejssnt.2025-044

    Arai, S., Shimizu, K., Lu, A.K.A., Masuda, H., Yoshida, H., Watanabe, S.: Charge states of ions around σ5(310)/[001] grain boundary in cubic-zro 2 revealed by first-principles calculations. e-Journal of Surface Science and Nanotechnology23, 323–327 (2025) https://doi.org/10.1380/ejssnt.2025-044

    work page doi:10.1380/ejssnt.2025-044 2025

  18. [18]

    Science and Technology of Advanced Materials: Methods3(1), 2253135 (2023) https://doi.org/10.1080/ 27660400.2023.2253135

    Shimizu, K., Otsuka, R., Hara, M., Minamitani, E., Watanabe, S.: Prediction of born effective charges using neural network to study ion migration under electric fields: applications to crystalline and amorphous Li3PO4. Science and Technology of Advanced Materials: Methods3(1), 2253135 (2023) https://doi.org/10.1080/ 27660400.2023.2253135

    arXiv 2023

  19. [19]

    Scientific Reports 15, 1250 (2025) https://doi.org/10.1038/s41598-025-01250-5

    Kutana, A., Shimizu, K., Watanabe, S., Asahi, R.: Representing born effective 32 charges with equivariant graph convolutional neural networks. Scientific Reports 15, 1250 (2025) https://doi.org/10.1038/s41598-025-01250-5

    work page doi:10.1038/s41598-025-01250-5 2025

  20. [20]

    Lu, A.K.A., Maekawa, N., Ikeda, A., Shimizu, K., Masuda, H., Yoshida, H., Watanabe, S.: Study of the ion mobility in defect-laden ZrO2 under an electric field using neural network with predictions for Born effective charges. Phys. Rev. Materials10(6), 066001 (2026) https://doi.org/10.1103/jcsd-dbl2

    work page doi:10.1103/jcsd-dbl2 2026

  21. [21]

    Nature Communications16(1), 4031 (2025) https://doi.org/10.1038/ s41467-025-59304-1

    Falletta, S., Cepellotti, A., Johansson, A., Tan, C.W., Descoteaux, M.L., Musaelian, A., Owen, C.J., Kozinsky, B.: Unified differentiable learning of electric response. Nature Communications16(1), 4031 (2025) https://doi.org/10.1038/ s41467-025-59304-1

    2025

  22. [22]

    The Journal of Chemical Physics 147(21), 214106 (2017) https://doi.org/10.1063/1.4997242

    Li, W., Ando, Y., Minamitani, E., Watanabe, S.: Study of li atom diffusion in amorphous li3po4 with neural network potential. The Journal of Chemical Physics 147(21), 214106 (2017) https://doi.org/10.1063/1.4997242

    work page doi:10.1063/1.4997242 2017

  23. [23]

    The Journal of Chemical Physics113(22), 9978–9985 (2000) https://doi.org/10.1063/1.1323224

    Henkelman, G., J´ onsson, H.: Improved tangent estimate in the nudged elastic band method for finding minimum energy paths and saddle points. The Journal of Chemical Physics113(22), 9978–9985 (2000) https://doi.org/10.1063/1.1323224

    doi:10.1063/1.1323224 2000

  24. [24]

    The Journal of Chemical Physics113(22), 9901–9904 (2000) https://doi.org/10.1063/1.1329672

    Henkelman, G., Uberuaga, B.P., J´ onsson, H.: A climbing image nudged elastic band method for finding saddle points and minimum energy paths. The Journal of Chemical Physics113(22), 9901–9904 (2000) https://doi.org/10.1063/1.1329672

    doi:10.1063/1.1329672 2000

  25. [25]

    The Journal of Chemical Physics164(24), 244114 (2026) https://doi.org/10.1063/5.0327083

    Schmiedmayer, B., Rittsteuer, A., Hilpert, T., Kresse, G.: Scalar machine learning of tensorial quantities—born effective charges from monopole models. The Journal of Chemical Physics164(24), 244114 (2026) https://doi.org/10.1063/5.0327083

    work page doi:10.1063/5.0327083 2026

  26. [26]

    Zenodo (2026)

    Schmiedmayer, B., Rittsteuer, A., Hilpert, T., Kresse, G.: Dataset for: Scalar machine learning of tensorial quantities – Born effective charges from monopole 33 models. Zenodo (2026). https://doi.org/10.5281/zenodo.19397893 . https://doi. org/10.5281/zenodo.19397893

    work page doi:10.5281/zenodo.19397893 2026

  27. [27]

    Materials & Design265, 115851 (2026) https://doi.org/10

    Chen, P., Mizoguchi, T.: Effect of uniaxial compressive stress on polarization switching and domain wall formation in tetragonal phase BaTiO 3 via machine learning potential. Materials & Design265, 115851 (2026) https://doi.org/10. 1016/j.matdes.2026.115851

    arXiv 2026

  28. [28]

    Scripta Materialia108, 1–5 (2015) https://doi.org/10.1016/j.scriptamat.2015.07

    Togo, A., Tanaka, I.: First principles phonon calculations in materials science. Scripta Materialia108, 1–5 (2015) https://doi.org/10.1016/j.scriptamat.2015.07. 021

    work page doi:10.1016/j.scriptamat.2015.07 2015

  29. [29]

    Journal of Physics: Condensed Matter35, 353001 (2023) https: //doi.org/10.1088/1361-648x/acd831

    Togo, A., Chaput, L., Tadano, T., Tanaka, I.: Implementation strategies in phonopy and phono3py. Journal of Physics: Condensed Matter35, 353001 (2023) https: //doi.org/10.1088/1361-648x/acd831

    doi:10.1088/1361-648x/acd831 2023

  30. [30]

    Nature Machine Intelligence7, 836–847 (2025) https://doi.org/10.1038/s42256-025-01055-1

    Riebesell, J., Goodall, R.E.A., Benner, P., Chiang, Y., Deng, B., Ceder, G., Asta, M., Lee, A.A., Jain, A., Persson, K.A.: A framework to evaluate machine learning crystal stability predictions. Nature Machine Intelligence7, 836–847 (2025) https://doi.org/10.1038/s42256-025-01055-1

    doi:10.1038/s42256-025-01055-1 2025

  31. [31]

    World Scientific, Singapore (1988)

    Varshalovich, D.A., Moskalev, A.N., Khersonskii, V.K.: Quantum Theory of Angular Momentum. World Scientific, Singapore (1988)

    1988

  32. [32]

    Oxford University Press, Oxford, UK (1954)

    Born, M., Huang, K.: Dynamical Theory of Crystal Lattices. Oxford University Press, Oxford, UK (1954)

    1954

  33. [33]

    Gonze, X., Lee, C.: Dynamical matrices, born effective charges, dielectric permittiv- ity tensors, and interatomic force constants from density-functional perturbation theory. Phys. Rev. B55, 10355–10368 (1997) https://doi.org/10.1103/PhysRevB. 55.10355 34

    doi:10.1103/physrevb 1997

  34. [34]

    Nature Communications14(1), 79 (2023) https://doi.org/10.1038/ s41467-022-35364-0

    Musaelian, A., Batzner, S., Johansson, A., Sun, L., Owen, C.J., Kornbluth, M., Kozinsky, B.: Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications14(1), 79 (2023) https://doi.org/10.1038/ s41467-022-35364-0

    2023

  35. [35]

    arXiv preprint arXiv:1802.08219 (2018)

    Thomas, N., Smidt, T., Kearnes, S., Yang, L., Li, L., Kohlhoff, K., Riley, P.: Tensor field networks: Rotation and translation equivariant neural networks for 3D point clouds. arXiv preprint arXiv:1802.08219 (2018)

    Pith/arXiv arXiv 2018

  36. [36]

    Hohenberg, P., Kohn, W.: Inhomogeneous electron gas. Phys. Rev.136, 864–871 (1964) https://doi.org/10.1103/PhysRev.136.B864

    doi:10.1103/physrev.136.b864 1964

  37. [37]

    Kohn, W., Sham, L.J.: Self-consistent equations including exchange and correlation effects. Phys. Rev.140, 1133–1138 (1965) https://doi.org/10.1103/PhysRev.140. A1133

    doi:10.1103/physrev.140 1965

  38. [38]

    Kresse, G., Furthm¨ uller, J.: Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set. Phys. Rev. B54, 11169– 11186 (1996) https://doi.org/10.1103/PhysRevB.54.11169

    doi:10.1103/physrevb.54.11169 1996

  39. [39]

    Physical Review Letters77(18), 3865–3868 (1996) https://doi.org/10

    Perdew, J.P., Burke, K., Ernzerhof, M.: Generalized gradient approximation made simple. Physical Review Letters77(18), 3865–3868 (1996) https://doi.org/10. 1103/PhysRevLett.77.3865

    1996

  40. [40]

    7 . 4 3 3 3 7 4 4 1 0.0 0.0 0.0 7 . 4 3 3 3 7 4 4 1 0.0 0.0 0.0 1 0 . 8 4 1 7 0 2 2 7

    Perdew, J.P., Ruzsinszky, A., Csonka, G.I., Vydrov, O.A., Scuseria, G.E., Con- stantin, L.A., Zhou, X., Burke, K.: Restoring the density-gradient expansion for exchange in solids and surfaces. Physical Review Letters100(13), 136406 (2008) https://doi.org/10.1103/PhysRevLett.100.136406 35 Supplementary Information for: SevenNet-Polar for MultiTask Predicti...

    doi:10.1103/physrevlett.100.136406 2008