pith. machine review for the scientific record. sign in

arxiv: 2605.11610 · v1 · submitted 2026-05-12 · ❄️ cond-mat.mtrl-sci · physics.comp-ph

Recognition: no theorem link

Fast and Accurate Prediction of Lattice Thermal Conductivity via Machine Learning Surrogates

Jianghai Wang, Junichiro Shiomi, Kedar Hippalgaonkar, Martin Hoffmann Petersen, Masato Ohnishi, Masatoshi Hanai, Michimasa Morita, Ruiming Zhu, Shuya Yamazaki, Tomiya Yamamoto, Toyotaro Suzumura, Wei Nong, Zekun Ren, Zeyu Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-13 01:38 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci physics.comp-ph
keywords lattice thermal conductivitymachine learning surrogatesMLIP embeddingsout-of-distribution generalizationthermoelectric materialsphonon propertieshigh-throughput screeningcrystal symmetries
0
0 comments X

The pith

Machine learning surrogate models predict lattice thermal conductivity accurately enough for high-throughput material screening, with MLIP embeddings strong inside known ranges and deep networks better for novel low-conductivity cases.

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

This paper benchmarks fifteen surrogate models for predicting lattice thermal conductivity on the Phonix database of 6,966 first-principles anharmonic phonon entries. Models are grouped as physical-informed feature methods, end-to-end deep neural networks, and pre-trained MLIP embeddings, then evaluated on random splits, space-group disjoint splits, and out-of-distribution splits keyed to extreme κ_lat values. Results indicate that MLIP-embedded models perform best for interpolation within familiar regions while deep networks such as ALiEGNN maintain higher robustness when extrapolating to unseen low-κ_lat materials. The surrogates deliver orders-of-magnitude speed gains over direct simulations, supporting efficient screening of thermoelectric materials despite reduced accuracy.

Core claim

Among the tested surrogates, MLIP-embedded models excel at interpolation within well-sampled regions of the Phonix database while deep neural network models, especially ALiEGNN, demonstrate superior robustness in out-of-distribution regimes that matter for discovering novel low-κ_lat materials. Systematic tests across random, symmetry-unseen, and property-based splits show consistent performance degradation when structural representations are reduced, yet all approaches cut computational costs dramatically compared with full first-principles calculations.

What carries the argument

Categorization of fifteen surrogate models into physical-informed feature ML, end-to-end deep neural networks, and MLIP-embeddings, evaluated across random, space-group disjoint, and κ_lat-based out-of-distribution data splits.

If this is right

  • MLIP-embedded models are preferred for interpolating lattice thermal conductivity inside well-sampled chemical and symmetry regions.
  • Deep neural networks like ALiEGNN are more reliable for extrapolating to out-of-distribution low-κ_lat materials needed for new thermoelectric discovery.
  • All surrogate models reduce computational costs by orders of magnitude relative to direct anharmonic phonon calculations.
  • Performance drops when structural input representations are simplified, regardless of model type.
  • The speed gains support integration into generative design workflows for thermoelectric materials with only modest accuracy loss.

Where Pith is reading between the lines

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

  • The benchmark results could guide selection of models for predicting other anharmonic phonon properties such as thermal expansion or phonon lifetimes.
  • Embedding the fastest surrogates into generative models might further speed up identification of high-performance thermoelectrics.
  • Testing the same splits on databases with greater chemical diversity would better reveal the true limits of current generalization claims.
  • Hybrid architectures that combine MLIP embeddings with deep network layers may balance interpolation strength and OOD robustness.

Load-bearing premise

The Phonix database and the chosen out-of-distribution splits based on κ_lat values sufficiently represent the challenges of generalizing to novel chemical spaces and unseen crystal symmetries.

What would settle it

A newly synthesized crystal with low lattice thermal conductivity lying outside the Phonix distribution where ALiEGNN predictions deviate substantially from direct first-principles results would falsify the claimed OOD robustness.

Figures

Figures reproduced from arXiv: 2605.11610 by Jianghai Wang, Junichiro Shiomi, Kedar Hippalgaonkar, Martin Hoffmann Petersen, Masato Ohnishi, Masatoshi Hanai, Michimasa Morita, Ruiming Zhu, Shuya Yamazaki, Tomiya Yamamoto, Toyotaro Suzumura, Wei Nong, Zekun Ren, Zeyu Wang.

Figure 1
Figure 1. Figure 1: Material distribution in Phonix. (a) Distribution and percentage of crystal [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dataset split for the benchmark. (a) Random 80%:20% split. (b) Space-group [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the surrogate models categorized into three groups, with the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The performance across the 15 surrogate models for each individual data split, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Schematic of the information passing layer in ALiEGNN, which jointly [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Schematic of the MLP architecture used to predict the lattice thermal [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
read the original abstract

The appearance of generative models has opened vast chemical spaces in the design of functional materials. Although machine learning interatomic potentials (MLIPs) have substantially accelerated phonon calculations, high-fidelity prediction of lattice thermal conductivity \k{appa}lat still requires accurate treatment of anharmonic interactions, which remains a key challenge for existing potentials across novel chemical spaces. To address this challenge, we present a comprehensive benchmark of 15 surrogate models for predicting \k{appa}lat using the Phonix database, which contains 6,966 entries with anharmonic phonon properties derived from first-principles calculations. Firstly, We categorize these surrogate models into three distinct groups: Physical-informed feature descriptors combined with ML models, end-to-end deep neural networks, and pre-trained MLIP-embeddings combined with ML models. By evaluating model performance across random, space-group disjoint (testing generalization to unseen crystal symmetries), and Out-Of-Distribution splits (OOD dataset that testing extrapolation to property regimes beyond the training range) based on \k{appa}lat, we probe both interpolation and exploration capabilities. Our results reveal that MLIP-embedded models excel in interpolation within well-sampled regions, deep neural network models especially ALiEGNN demonstrate superior robustness in OOD regimes critical for discovering novel low-\k{appa}lat. Additionally, we find a systematic degradation in performance when the structural representation is reduced. Although surrogate models exhibit lower accuracy than direct simulations using first-principles calculation, they reduce computational costs by orders of magnitude, enabling efficient high-throughput screening of thermoelectric materials with minimal loss in generative design workflows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper benchmarks 15 surrogate models for predicting lattice thermal conductivity (κ_lat) on the Phonix database of 6,966 first-principles anharmonic phonon entries. Models are grouped into physical-informed feature descriptors with ML, end-to-end deep neural networks (highlighting ALiEGNN), and pre-trained MLIP embeddings with ML. Performance is assessed on random splits, space-group disjoint splits, and OOD splits defined by κ_lat thresholds, with claims that MLIP-embedded models excel at interpolation while ALiEGNN shows superior robustness in OOD regimes for novel low-κ_lat discovery, at orders-of-magnitude lower cost than direct first-principles calculations.

Significance. If the OOD generalization claims hold under chemical novelty, the work would provide a valuable, practical toolkit for high-throughput screening in generative thermoelectric materials design. The multi-split evaluation protocol (random, symmetry-disjoint, property-OOD) is a clear methodological strength that allows systematic probing of interpolation versus extrapolation. The scale of the Phonix database and focus on anharmonic properties also strengthen the benchmark's relevance.

major comments (2)
  1. [Abstract and Methods (dataset splits)] Abstract and Methods (OOD splits definition): The OOD regime is constructed solely by thresholding on κ_lat values (property-regime extrapolation) and space-group disjoint splits (symmetry extrapolation). No quantitative metrics are reported for chemical novelty, such as element-set overlap, compositional Tanimoto distance, or stoichiometry uniqueness between train and test partitions. This is load-bearing for the central claim that ALiEGNN demonstrates 'superior robustness in OOD regimes critical for discovering novel low-κ_lat' materials, because performance gains could reflect within-chemistry interpolation rather than the claimed extrapolation to unseen chemical spaces required for generative design workflows.
  2. [Results] Results section (model comparisons): The superiority of ALiEGNN in OOD is presented qualitatively, but without reported statistical significance tests, error bars on metrics, or ablation on hyperparameter sensitivity, it is difficult to establish that the observed robustness is robust rather than sensitive to the specific train/test partitioning or model configuration. This weakens the quantitative support for preferring ALiEGNN over MLIP-embedded alternatives in OOD settings.
minor comments (3)
  1. [Abstract] Abstract: 'Firstly, We categorize' contains an unnecessary capital 'W' and awkward phrasing; 'testing extrapolation' is repeated in the OOD description.
  2. [Methods] The manuscript would benefit from an explicit table or supplementary section listing all 15 models with their exact architectures, hyperparameters, and training protocols to support reproducibility.
  3. [Figures] Figure captions and axis labels should explicitly state the metric (e.g., MAE, R²) and split type for each panel to improve clarity when comparing interpolation versus OOD performance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback, which has helped us identify areas to strengthen the manuscript. We address each major comment below and commit to revisions that enhance the rigor of our claims without altering the core findings.

read point-by-point responses

  1. Referee: Abstract and Methods (OOD splits definition): The OOD regime is constructed solely by thresholding on κ_lat values (property-regime extrapolation) and space-group disjoint splits (symmetry extrapolation). No quantitative metrics are reported for chemical novelty, such as element-set overlap, compositional Tanimoto distance, or stoichiometry uniqueness between train and test partitions. This is load-bearing for the central claim that ALiEGNN demonstrates 'superior robustness in OOD regimes critical for discovering novel low-κ_lat' materials, because performance gains could reflect within-chemistry interpolation rather than the claimed extrapolation to unseen chemical spaces required for generative design workflows.

    Authors: We appreciate the referee's point that explicit chemical novelty metrics would better substantiate the extrapolation claims. Our OOD definition emphasizes property-regime shifts (via κ_lat thresholding) and symmetry generalization (space-group disjoint), which are relevant for screening workflows, but we agree these do not directly quantify compositional novelty. In the revised manuscript, we will compute and report additional metrics for all splits, including element-set overlap fractions, mean compositional Tanimoto distances, and the percentage of unique stoichiometries between train and test partitions. These will be added to the Methods and Results sections to allow readers to evaluate the degree of chemical novelty and clarify the scope of ALiEGNN's robustness for novel low-κ_lat discovery. revision: yes

  2. Referee: Results section (model comparisons): The superiority of ALiEGNN in OOD is presented qualitatively, but without reported statistical significance tests, error bars on metrics, or ablation on hyperparameter sensitivity, it is difficult to establish that the observed robustness is robust rather than sensitive to the specific train/test partitioning or model configuration. This weakens the quantitative support for preferring ALiEGNN over MLIP-embedded alternatives in OOD settings.

    Authors: We agree that quantitative statistical support and sensitivity analysis are needed to make the model comparisons more robust. In the revised version, we will add error bars (standard deviations from repeated training runs with different random seeds) to all reported metrics in the OOD evaluations. We will also include statistical significance tests (e.g., paired t-tests or Wilcoxon signed-rank tests with p-values) comparing ALiEGNN against the leading MLIP-embedded models. Additionally, we will include a hyperparameter ablation study for ALiEGNN and the top-performing alternatives to confirm that the OOD advantages persist across reasonable configuration ranges and are not artifacts of the specific partitioning or tuning. revision: yes

Circularity Check

0 steps flagged

No circularity: benchmark uses independent held-out test splits by construction

full rationale

The paper trains 15 surrogate models (feature-based ML, end-to-end DNNs including ALiEGNN, and MLIP-embeddings) on subsets of the 6,966-entry Phonix database and reports accuracy on explicitly constructed held-out test sets: random splits, space-group-disjoint splits, and OOD splits defined by κ_lat thresholds. These splits are formed by partitioning the data prior to training, so test predictions are independent of the training inputs by design and do not reduce to any fitted parameter or input feature. No equations, self-citations, ansatzes, or uniqueness claims are invoked to derive the reported performance numbers; the central results are empirical error metrics on external test data. The evaluation chain therefore remains self-contained and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the empirical performance of trained surrogate models evaluated on held-out data; no new physical axioms, free parameters fitted within the paper, or invented entities are introduced beyond standard machine learning practices.

pith-pipeline@v0.9.0 · 5653 in / 1141 out tokens · 36394 ms · 2026-05-13T01:38:12.226873+00:00 · methodology

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

50 extracted references · 50 canonical work pages · 1 internal anchor

  1. [2]

    Schleder G R, Padilha A C M, Acosta C M, Costa M and Fazzio A 2019 Journal of Physics: Materials2032001 ISSN 2515-7639 URLhttp://dx.doi.org/10.1088/ 2515-7639/ab084b

    work page 2019

  2. [3]

    Butler K T, Davies D W, Cartwright H, Isayev O and Walsh A 2018 Nature559 547–555 ISSN 1476-4687 URLhttp://dx.doi.org/10.1038/s41586-018-0337-2

    work page doi:10.1038/s41586-018-0337-2 2018

  3. [4]

    Oganov A R, Pickard C J, Zhu Q and Needs R J 2019 Nature Reviews Materials4 331–348 ISSN 2058-8437 URLhttp://dx.doi.org/10.1038/s41578-019-0101-8

    work page doi:10.1038/s41578-019-0101-8 2019

  4. [5]

    Charmousis, E.J

    Seko A, Togo A, Hayashi H, Tsuda K, Chaput L and Tanaka I 2015Physical Review Letters115ISSN 1079-7114 URLhttp://dx.doi.org/10.1103/PhysRevLett. 115.205901

    work page doi:10.1103/physrevlett

  5. [6]

    Chen Z, Andrejevic N, Smidt T, Ding Z, Xu Q, Chi Y, Nguyen Q T, Alatas A, Kong J and Li M 2021 Advanced Science8ISSN 2198-3844 URLhttp: //dx.doi.org/10.1002/advs.202004214

    work page doi:10.1002/advs.202004214 2021

  6. [7]

    Fang S, Geiger M, Checkelsky J G and Smidt T 2024 Phonon predictions with e(3)- equivariant graph neural networks URLhttps://arxiv.org/abs/2403.11347 REFERENCES12

    work page arXiv 2024

  7. [8]

    org/10.1038/s43588-024-00661-0

    Okabe R, Chotrattanapituk A, Boonkird A, Andrejevic N, Fu X, Jaakkola T S, Song Q, Nguyen T, Drucker N, Mu S, Wang Y, Liao B, Cheng Y and Li M 2024 Nature Computational Science4522–531 ISSN 2662-8457 URLhttp://dx.doi. org/10.1038/s43588-024-00661-0

    work page doi:10.1038/s43588-024-00661-0 2024

  8. [9]

    Yang H, Hu C, Zhou Y, Liu X, Shi Y, Li J, Li G, Chen Z, Chen S, Zeni C, Horton M, Pinsler R, Fowler A, Z¨ ugner D, Xie T, Smith J, Sun L, Wang Q, Kong L, Liu C, Hao H and Lu Z 2024 arXiv preprint arXiv:2405.04967 URL https://arxiv.org/abs/2405.04967

    work page arXiv 2024

  9. [10]

    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, Berger F, Bernstein N, Bhowmik A, Blau S M, C˘ arare V, Darby J P, De S, Della Pia F, Deringer V L, Elijoˇ sius R, El-Machachi Z, Falcioni F, Fako E, Ferrari A C, Genreith-Schriever A, George J, Goodall R E A, Grey C P, Grigorev P, Han S, Ha...

    work page arXiv

  10. [11]

    1038/s42256-023-00716-3

    Deng B, Zhong P, Jun K, Riebesell J, Han K, Bartel C J and Ceder G 2023 Nature Machine Intelligence51031–1041 ISSN 2522-5839 URLhttp://dx.doi.org/10. 1038/s42256-023-00716-3

    work page 2023

  11. [12]

    Unke O T, Chmiela S, Sauceda H E, Gastegger M, Poltavsky I, Sch¨ utt K T, Tkatchenko A and M¨ uller K R 2021 Chemical Reviews12110142–10186 ISSN 1520-6890 URLhttp://dx.doi.org/10.1021/acs.chemrev.0c01111

    work page doi:10.1021/acs.chemrev.0c01111 2021

  12. [13]

    Zeni C, Pinsler R, Z¨ ugner D, Fowler A, Horton M, Fu X, Wang Z, Shysheya A, Crabb´ e J, Ueda S, Sordillo R, Sun L, Smith J, Nguyen B, Schulz H, Lewis S, Huang C W, Lu Z, Zhou Y, Yang H, Hao H, Li J, Yang C, Li W, Tomioka R and Xie T 2025 Nature639624–632 ISSN 1476-4687 URLhttp://dx.doi.org/10.1038/ s41586-025-08628-5

    work page 2025

  13. [14]

    org/abs/2402.03992

    Jiao R, Huang W, Liu Y, Zhao D and Liu Y 2024 ICLR URLhttps://arxiv. org/abs/2402.03992

    work page arXiv 2024

  14. [15]

    Kazeev N, Nong W, Romanov I, Zhu R, Ustyuzhanin A, Yamazaki S and Hippalgaonkar K 2025 Wyckoff transformer: Generation of symmetric crystals Proceedings of the 46nd International Conference on Machine Learning (ICML) (arXiv) URLhttps://arxiv.org/abs/2503.02407

    work page arXiv 2025

  15. [16]

    Wang Z, Ma J, Hu R and Luo X 2023 Applied Physics Letters122ISSN 1077-3118 URLhttp://dx.doi.org/10.1063/5.0142150

    work page doi:10.1063/5.0142150 2023

  16. [17]

    Merchant A, Batzner S, Schoenholz S S, Aykol M, Cheon G and Cubuk E D REFERENCES13 2023 Nature62480–85 ISSN 1476-4687 URLhttp://dx.doi.org/10.1038/ s41586-023-06735-9

    work page 2023

  17. [18]

    Shi X L, Li N H, Li M and Chen Z G 2025 Chemical Reviews1257525–7724 ISSN 1520-6890 URLhttp://dx.doi.org/10.1021/acs.chemrev.5c00060

    work page doi:10.1021/acs.chemrev.5c00060 2025

  18. [19]

    Wang J, Yin Y, Che C and Cui M 2025 Energies182122 ISSN 1996-1073 URL http://dx.doi.org/10.3390/en18082122

    work page doi:10.3390/en18082122 2025

  19. [20]

    Wang Q, Zhou Z, Liu C, Zheng Y, Shi Z, Wei B, Zhang W, Nan C W and Lin Y H 2025 Chemical Society Reviews ISSN 1460-4744 URLhttp://dx.doi.org/ 10.1039/d5cs01078k

    work page doi:10.1039/d5cs01078k 2025

  20. [21]

    Snyder G J and Toberer E S 2008 Nature Materials7105–114 ISSN 1476-4660 URLhttp://dx.doi.org/10.1038/nmat2090

    work page doi:10.1038/nmat2090 2008

  21. [22]

    Ming H, Luo Z Z, Zou Z and Kanatzidis M G 2025Chemical Reviews1253932–3975 ISSN 1520-6890 URLhttp://dx.doi.org/10.1021/acs.chemrev.4c00786

    work page doi:10.1021/acs.chemrev.4c00786

  22. [23]

    He J and Tritt T M 2017 Science357ISSN 1095-9203 URLhttp://dx.doi.org/ 10.1126/science.aak9997

    work page doi:10.1126/science.aak9997 2017

  23. [24]

    1039/d4cc04660a

    Jong U G, Ryu C, Hwang J M, Kim S H, Ju I G and Yu C J 2024 Chemical Communications6013400–13403 ISSN 1364-548X URLhttp://dx.doi.org/10. 1039/d4cc04660a

    work page 2024

  24. [25]

    Gu X, Fan Z and Bao H 2021 Journal of Applied Physics130ISSN 1089-7550 URL http://dx.doi.org/10.1063/5.0069175

    work page doi:10.1063/5.0069175 2021

  25. [26]

    Wang A Y T, Kauwe S K, Murdock R J and Sparks T D 2021 npj Computational Materials7ISSN 2057-3960 URLhttp://dx.doi.org/10.1038/ s41524-021-00545-1

    work page 2021

  26. [27]

    Xie T and Grossman J C 2018 Physical Review Letters120ISSN 1079-7114 URL http://dx.doi.org/10.1103/PhysRevLett.120.145301

    work page doi:10.1103/physrevlett.120.145301 2018

  27. [28]

    org/abs/2412.05948

    Zeng Y, Cao W, Zuo Y, Peng T, Hou Y, Miao L, Wang Z and Shi J 2024 Accelerating the discovery of materials with expected thermal conductivity via a synergistic strategy of dft and interpretable deep learning URLhttps://arxiv. org/abs/2412.05948

    work page arXiv 2024

  28. [29]

    Ohnishi M, Deng T, Torres P, Xu Z, Tadano T, Zhang H, Nong W, Hanai M, Wang Z, Morita M, Tian Z, Hu M, Ruan X, Yoshida R, Suzumura T, Lindsay L, McGaughey A J H, Luo T, Hippalgaonkar K and Shiomi J 2026 npj Computational Materials12ISSN 2057-3960 URLhttp://dx.doi.org/10.1038/ s41524-026-02033-w

    work page 2026

  29. [30]

    Jain A, Ong S P, Hautier G, Chen W, Richards W D, Dacek S, Cholia S, Gunter D, Skinner D, Ceder G and Persson K A 2013 APL Materials1ISSN 2166-532X URLhttp://dx.doi.org/10.1063/1.4812323

    work page doi:10.1063/1.4812323 2013

  30. [31]

    Kresse G and Furthm¨ uller J 1996Physical Review B5411169–11186 ISSN 1095- 3795 URLhttp://dx.doi.org/10.1103/PhysRevB.54.11169 REFERENCES14

    work page doi:10.1103/physrevb.54.11169

  31. [32]

    Tadano T, Gohda Y and Tsuneyuki S 2014 Journal of Physics: Condensed Matter 26225402 ISSN 1361-648X URLhttp://dx.doi.org/10.1088/0953-8984/26/ 22/225402

    work page doi:10.1088/0953-8984/26/ 2014

  32. [33]

    Shiomi J, Esfarjani K and Chen G 2011 Physical Review B84ISSN 1550-235X URLhttp://dx.doi.org/10.1103/PhysRevB.84.104302

    work page doi:10.1103/physrevb.84.104302 2011

  33. [34]

    Esfarjani K, Chen G and Stokes H T 2011 Physical Review B84ISSN 1550-235X URLhttp://dx.doi.org/10.1103/PhysRevB.84.085204

    work page doi:10.1103/physrevb.84.085204 2011

  34. [35]

    Togo A 2015 Phonon database at kyoto university https://github.com/atztogo/phonondb/tree/main

    work page 2015

  35. [36]

    McInnes L, Healy J and Melville J 2018 Umap: Uniform manifold approximation and projection for dimension reduction URLhttps://arxiv.org/abs/1802. 03426

    work page 2018

  36. [37]

    Ward L, Dunn A, Faghaninia A, Zimmermann N E, Bajaj S, Wang Q, Montoya J, Chen J, Bystrom K, Dylla M, Chard K, Asta M, Persson K A, Snyder G J, Foster I and Jain A 2018 Computational Materials Science15260–69 ISSN 0927-0256 URL http://dx.doi.org/10.1016/j.commatsci.2018.05.018

    work page doi:10.1016/j.commatsci.2018.05.018 2018

  37. [38]

    Neumann M, Gin J, Rhodes B, Bennett S, Li Z, Choubisa H, Hussey A and Godwin J 2024 Orb: A fast, scalable neural network potential (Preprint2410.22570) URL https://arxiv.org/abs/2410.22570

    work page arXiv 2024

  38. [39]

    Rhodes B, Vandenhaute S, ˇSimkus V, Gin J, Godwin J, Duignan T and Neumann M 2025 Orb-v3: atomistic simulation at scale (Preprint2504.06231) URLhttps: //arxiv.org/abs/2504.06231

    work page arXiv 2025

  39. [40]

    Kim S Y, Park Y J and Li J 2025 Leveraging neural network interatomic potentials for a foundation model of chemistry URLhttps://arxiv.org/abs/2506.18497

    work page arXiv 2025

  40. [41]

    Zhang C, Zhang D, Peng A, Guo M, Zhang Y, Wang L, Ke G, Zhang L, Li T and Wang H 2026 Multi-task fine-tuning enables robust out-of-distribution generalization in atomistic models URLhttps://arxiv.org/abs/2601.08486

    work page arXiv 2026

  41. [42]

    Omee S S, Fu N, Dong R, Hu M and Hu J 2024 npj Computational Materials10 ISSN 2057-3960 URLhttp://dx.doi.org/10.1038/s41524-024-01316-4

    work page doi:10.1038/s41524-024-01316-4 2024

  42. [43]

    Liu Z, Wang Y, Vaidya S, Ruehle F, Halverson J, Soljaˇ ci´ c M, Hou T Y and Tegmark M 2024 Kan: Kolmogorov-arnold networks URLhttps://arxiv.org/abs/2404. 19756

    work page 2024

  43. [44]

    Gilmer J, Schoenholz S S, Riley P F, Vinyals O and Dahl G E 2017 Neural message passing for quantum chemistry URLhttps://arxiv.org/abs/1704.01212

    work page arXiv 2017

  44. [45]

    Lim Y F, Ng C K, Vaitesswar U and Hippalgaonkar K 2021 Advanced Intelligent Systems3ISSN 2640-4567 URLhttp://dx.doi.org/10.1002/aisy.202100101

    work page doi:10.1002/aisy.202100101 2021

  45. [46]

    Petersen M H, Zhu R, Dai H, Aggarwal S, Wei N, Chen A P, Bhowmik A, Lastra J M G and Hippalgaonkar K 2025 Dis-gen: Disordered crystal structure generation URLhttps://arxiv.org/abs/2507.18275 REFERENCES15

    work page arXiv 2025

  46. [47]

    Zhu R, Nong W, Yamazaki S and Hippalgaonkar K 2024 Matter73469–3488 ISSN 2590-2385 URLhttp://dx.doi.org/10.1016/j.matt.2024.05.042

    work page doi:10.1016/j.matt.2024.05.042 2024

  47. [48]

    Hollmann N, M¨ uller S, Purucker L, Krishnakumar A, K¨ orfer M, Hoo S B, Schirrmeister R T and Hutter F 2025 Nature637319–326

    work page 2025

  48. [49]

    Batatia I, Kovacs D P, Simm G, Ortner C and Cs´ anyi G 2022 Advances in Neural Information Processing Systems3511423–11436 URLhttps://arxiv.org/abs/ 2206.07697

    work page arXiv 2022

  49. [50]

    Adam: A Method for Stochastic Optimization

    Kingma D P and Ba J 2014 arXiv preprint arXiv:1412.6980 URLhttps://arxiv. org/abs/1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  50. [51]

    Surrogate Models A.1

    Mianzhi P, Li J, Ouyang Y, Ma W Y, Zhang J and Zhou H 2025 ON EXTRAPOLATION IN MATERIAL PROPERTY REGRESSION URLhttps: //openreview.net/forum?id=czVzzXPCkw REFERENCES16 A. Surrogate Models A.1. Physical-informed features + MLP, XGBoost, KAN A.1.1. Feature EngineeringAll three models utilize identical feature representations to enable direct performance com...

    work page 2025