Pith. sign in

REVIEW 3 major objections 4 minor 52 references

Leveraging neural network interatomic potentials for a foundation model of chemistry

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read HackNIP claims that frozen embeddings from a pretrained neural network interatomic potential, paired with a shallow regression model, match or beat state-of-the-art property predictors on Matbench and outperform direct fine-tuning below…

desk verdict Promising hybrid NIP-descriptor recipe, but the headline benchmark numbers are selected on the test folds, so the quantitative claims need a nested-evaluation redo before they can be trusted. read the letter →

arxiv 2506.18497 v1 pith:MC2K6WFD submitted 2025-06-23 cond-mat.mtrl-sci cs.LG

So Yeon Kim , Yang Jeong Park , Ju Li This is my paper
classification cond-mat.mtrl-scics.LG
keywords neuralnetworkinteratomicpotentialstransferlearningstructure-to-propertypredictionMatbenchfeature-basedmachinematerialsinformaticsembeddingdepthfoundationmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that a pretrained neural network interatomic potential can be turned into a general chemistry representation by freezing the network, reading out fixed-length embeddings at a chosen layer, and training a shallow model on top. On eight Matbench regression tasks this hybrid pipeline, called HackNIP with ORB features and a MODNet regressor, matches or beats every leaderboard model on average relative error and achieves the best normalized error on four tasks. The paper further claims that this frozen-feature approach outperforms direct fine-tuning of the same NIP for datasets up to about $10^4$ samples, with the crossover between $10^3$ and $10^4$, and that it transfers to experimental, amorphous, dynamic, and molecular properties. A reader should care because the claim, if right, means high-performance property prediction no longer requires training a deep network for each new property.

What carries the argument

The central object is the depth-truncated NIP embedding: raw coordinates and species labels are pushed through ORB's shared encoder and the first $L$ backbone layers, and the resulting atom-level feature vectors are averaged over all atoms to form one fixed-length descriptor, 256-dimensional for every ORB layer. That descriptor is the sole input to a four-layer MODNet, and the truncation depth $L$ is the control knob that determines how much global, energy-oriented compression the representation has undergone. The paper uses UMAP projections and per-atom Jensen-Shannon divergence across layers to argue that early-to-middle layers keep property-relevant variation that the final layer has squeezed out, which is what makes shallow-layer embeddings more useful for downstream prediction.

What would settle it

Recompute each Matbench task with the embedding depth selected by nested cross-validation on training folds only; if the averaged relative error worsens or the model no longer ranks first, the reported benchmark advantage is in part an artifact of choosing the architecture on test labels.

Watch

Extended reading notes

Core claim

On its own terms, HackNIP establishes that mean-pooled, depth-truncated embeddings from the ORB foundation potential, fed into a small MODNet, give state-of-the-art structure-to-property predictions: the lowest normalized mean absolute error on four of the eight Matbench regression tasks, the best average rank among leaderboard models, $R^2=0.968$ on the phonon-peak frequency task, and 40 meV per unit cell on perovskite formation energy. It also claims that at dataset sizes up to roughly $10^4$ samples the frozen-embedding pipeline beats property-specific fine-tuning of ORB, while at the full band-gap dataset fine-tuned ORB reaches about 0.10 eV and edges out HackNIP. The generality results cover AIMD-computed lithium diffusivity, experimentally measured superconducting critical temperatures, electrolyte Coulombic efficiency, and three molecular classification benchmarks. The layer analysis shows that the most informative ORB embeddings for solid-state tasks usually lie in layers 1 through 7, not the final layer, whose representations collapse into a low-dimensional manifold.

Load-bearing premise

The whole comparison rests on the assumption that choosing the best embedding depth from the same five-fold test experiments and then reporting that performance is a fair evaluation protocol; if that choice leaks test information, the leaderboard ranking, the fine-tuning comparison, and the data-efficiency crossover are all over-optimistic.

Editorial extensions

If this is right

  • Any structure-to-property regression or classification task can be attempted with one frozen NIP plus a shallow model, removing the need for per-property deep-network training.
  • For datasets of up to about $10^4$ samples, the paper claims HackNIP should be preferred over fine-tuning the same NIP, and its advantage grows as the dataset shrinks.
  • Because ORB embeddings are fixed and cheap to evaluate, leaderboard-competitive accuracy is reachable for datasets up to about $10^5$ samples on free-tier cloud hardware.
  • Embedding depth matters: users should sweep early and middle layers instead of defaulting to the last layer, which often gives the worst features.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Since the best embedding depth is selected on the same test folds used for reporting, part of the leaderboard margin may be optimistic; a validation-only layer sweep is the direct check.
  • The per-atom Jensen-Shannon divergence curves could be developed into a feature-importance diagnostic that identifies which atoms carry a property, beyond the aggregate parity plots shown here.
  • Concatenating embeddings from several backbones such as ORB, MACE, and EquiformerV2 is an untested extension that might beat any single backbone, since the paper compares them only as alternatives.
  • The same frozen-embedding recipe could be aimed at generative or inverse design, though the paper only demonstrates prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes HackNIP, a two-stage pipeline that extracts fixed-length, mean-pooled embeddings from pretrained neural network interatomic potentials (ORB, EquiformerV2, MACE) at varying network depths and feeds them into shallow regressors (MODNet, XGBoost, MLP) for structure-to-property prediction. The authors benchmark the method on the eight Matbench regression tasks with structure inputs, compare it against the Matbench leaderboard, study data efficiency against fine-tuned ORB on tasks with more than 10^4 samples, and demonstrate generalizability on Li diffusivity, superconducting critical temperature, electrolyte Coulombic efficiency, and three MoleculeNet classification tasks. They report that the best configuration (ORB + MODNet) achieves the lowest normalized MAE on four of eight Matbench tasks, outperforms direct fine-tuning for datasets up to roughly 10^4 samples, and that shallow-layer ORB embeddings are often more informative than final-layer embeddings. The paper also presents UMAP and Jensen-Shannon divergence analyses to interpret the layer-depth dependence.

Significance. If the reported performance survived an unbiased evaluation, the paper would offer a practically valuable, computationally lightweight alternative to end-to-end graph neural networks and fine-tuned NIPs for small-to-medium materials datasets. The manuscript is clearly written, uses standard public benchmarks, makes its code available on GitHub, and includes a useful analysis of which NIP layers carry transferable features. These are genuine strengths. However, the central quantitative claims rest on a model-selection protocol that uses the Matbench test folds to choose the embedding depth and hyperparameters, so the leaderboard ranking, the state-of-the-art statement, and the data-efficiency crossover are not yet supported as unbiased estimates. The qualitative idea that shallow models on frozen NIP embeddings can be competitive is plausible, but the reported margins and rankings require a valid nested evaluation before they can be accepted.

major comments (3)
  1. [5.5, Fig. 2, Table S1] The protocol described as "We identified the best-performing embedding depth L* and considered its corresponding predictive performance to represent the HackNIP architecture's performance" is a post-hoc model selection on the Matbench test folds. Because the five-fold experiments use Matbench's predefined train/test splits (Section 2.1 and Section 5.5), L* is chosen for each task by the lowest average test MAE across depths 1 through 15. The reported Figure 2 numbers, the nMAE ranking, and the claim of state-of-the-art performance are therefore minima over candidate depths evaluated on the same labels used for reporting, not unbiased estimates of a fixed architecture's accuracy. This overstates the margin over baselines. Please re-run the evaluation with an honest nested split: hold out a test set, or use outer folds that are never touched when selecting L*, hyperparameters, or the regressor pair, and report both the selection-unbiased MAEs and the selected values, including the spread across selected depths.
  2. [5.6, Fig. S1] The same test-label selection is compounded by the Optuna search described in Section 5.6, which runs "50 trials at each of five ORB depths around the best-performing embedding depths (from L*−2 to L*+2)" and scores "each trial by the MAE averaged across outer folds." As written, these outer folds are the same five Matbench folds used for the reported MAEs, so hyperparameters are also tuned on test labels. In addition, the choice of ORB–MODNet as the HackNIP architecture is made from Figure S1, which compares five combinations on the same five-fold test evaluation. The manuscript should state explicitly whether any validation split was ever separated from the test folds used in Figures 2 and 3; if not, the benchmark procedure needs to be corrected before the leaderboard and data-efficiency claims can be made.
  3. [2.2, Fig. 3] The data-efficiency comparison is confounded by the same selection protocol. HackNIP inherits task-specific L* values selected on the full-data Matbench test folds (Table S1), while the fine-tuned ORB baseline is a fixed, unselected protocol; the statement that "MODNet hyperparameters were re-optimized for the smaller subsets" suggests that the subset models are also tuned using the evaluation folds. Thus the crossover between 10^3 and 10^4 samples in Figure 3b, and the 18% bulk-modulus margin, may partly reflect an advantage from having selected the model on the evaluation labels rather than an intrinsic data-efficiency property of the frozen-embedding pipeline. Please evaluate both methods under the same nested or hold-out protocol and report whether the crossover and the per-task margins persist.
minor comments (4)
  1. [Fig. 2b, Eq. (1), Sec. 2.4] The text has repeated typos: "mix-man normalized" in the Fig. 2b caption and near Eq. (1), and "mix-max normalized" near Eq. (4); both should read "min-max normalized."
  2. [Fig. 3a, Discussion] The Fig. 3a caption contains "shesar and bulk moduli," which should be "shear and bulk moduli"; the Discussion contains "more diverse physcis," which should be "more diverse physics."
  3. [5.5, Table S1] The sentence "The optimized hyperparameters for each task are summarized in Supplementary Table S1" is ambiguous about whether those hyperparameters were selected on a validation split or on the test folds; please state clearly which splits were used for model selection throughout.
  4. [2.3, Tc paragraph] The comparison of the reported R2 = 0.768 with Ref. [24]'s R2 = 0.88 would be clearer if the sentence directly stated that the two values are not comparable because of the different target ranges, split sizes, and structures, before explaining why direct comparison is challenging.

Circularity Check

1 steps flagged · score 6.0 of 10

L* is chosen from the same Matbench test folds used to report final MAEs, so HackNIP's leaderboard and data-efficiency numbers are selected statistics rather than independent predictions.

  1. fitted input called prediction [Section 5.5, 'Feature-based machine learning'; also Sections 2.1 and 5.6, Fig. S1, Table S1.]
    "We identified the best-performing embedding depth L* and considered its corresponding predictive performance to represent the HackNIP architecture's performance."

    L* is selected as the depth with the lowest average MAE over Matbench's predefined five-fold test splits, and the same five-fold test MAEs are then reported as HackNIP's benchmark performance (Fig. 2, Table S1). The reported number is therefore, by construction, the minimum of the 15 per-layer test MAEs evaluated for each task, not an unbiased estimate for a fixed architecture. The same test labels also select the ORB–MODNet pair (Fig. S1) and steer the Optuna search in Section 5.6, which is scored by 'MAE averaged across outer folds.' Consequently, the leaderboard ranking and the data-efficiency crossover in Fig. 3b partly encode selection on the evaluation labels.

full rationale

The central supervised pipeline is not circular: MODNet is trained on fixed, frozen ORB embeddings (which were fit to energies, not to the benchmark targets), and test-fold property values are genuine holdout predictions. The paper does not define ORB embeddings in terms of the predicted properties, and no load-bearing argument reduces to a self-citation; reference [33] (Park et al.) appears only as related work, and the electrolyte dataset citation is a data source. The one circularity-adjacent step is the architecture-selection protocol in Sections 5.5–5.6: 'best-performing embedding depth L*' is chosen using the same five-fold Matbench test evaluations whose average MAE is later reported as HackNIP's performance. That makes the reported MAE the minimum over the evaluated depths, and also over the regressor/NIP pairs screened in Fig. S1 and the Optuna trials scored on outer-fold MAE, so the headline leaderboard and data-efficiency claims are selected statistics. Because this selection directly determines the central quantitative claims, the score is 6 rather than 0, although this is a methodological selection-bias problem, not a logical equivalence between the property predictions and the fitted embeddings.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central method rests on a handful of modeling choices: the pretrained ORB encoder is treated as a fixed feature extractor, atom features are mean-pooled, a per-task embedding depth is selected, and MODNet hyperparameters are tuned. The only parameter that directly shapes the headline numbers is L*, which is selected on test folds, a circularity-adjacent choice. No new physical entities are postulated.

free parameters (3)
  • ORB embedding depth L* per task = 1-11 for Matbench, 2/4/1 for diffusion/Tc/CE, 10/15/13 for molecular tasks
    Selected per task as the layer giving the lowest MAE or highest AUC on the test folds; this is a data-dependent model-selection parameter and inflates reported performance.
  • MODNet hyperparameters (learning rate, batch size, depth, width, input features) = Per-task values in Tables S1 and S2
    Tuned with Optuna (50 trials) for each task around the selected embedding depth; they affect reported errors.
  • Fine-tuning hyperparameters (backbone LR, head LR, weight decay, epochs) = backbone_lr=1e-5, head_lr=1e-3, wd=1e-4, epochs=200
    Chosen by hand for the fine-tuning baseline; not tuned per dataset size, which may disadvantage fine-tuning in small-data comparisons.
assumptions (5)
  • domain assumption Mean-pooled atom-level embeddings from ORB layers are sufficient fixed-length descriptors for structure-to-property prediction.
    Section 5.4 states that atom-level features are averaged over all atoms to produce a fixed-length descriptor vector; this discards spatial information not already retained by the ORB encoder.
  • domain assumption ORB's pretrained representations transfer to tasks far from its energy and force training objective, including experimental Tc, Coulombic efficiency, and molecular bioactivity.
    Used throughout Section 2.3; no fine-tuning or re-training of ORB is performed for these tasks.
  • domain assumption The Matbench test splits and literature datasets are fair, unbiased samples of the property distributions.
    Section 5.1 lists the datasets; the paper relies on their labels and splits being correct.
  • ad hoc to paper Selecting L* on the test folds does not materially inflate the reported performance.
    Section 5.5 explicitly identifies the best-performing L* on the five-fold experiments and reports that performance as HackNIP's; this is a test-set selection step.
  • domain assumption Supercell expansion to lattice vectors over 10 or 12 angstroms ensures all relevant neighbor interactions are captured by the NIP graphs.
    Section 5.3 states this preprocessing; if cutoffs are insufficient, the extracted embeddings change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging neural network interatomic potentials for a foundation model of chemistry." pith.science (2026). https://pith.science/paper/MC2K6WFD

@misc{pith2026250618497,
  author       = {Pith},
  title        = {Pith review of: Leveraging neural network interatomic potentials for a foundation model of chemistry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MC2K6WFD}},
  note         = {Machine review of arXiv:2506.18497}
}
read the original abstract

Large-scale foundation models, including neural network interatomic potentials (NIPs) in computational materials science, have demonstrated significant potential. However, despite their success in accelerating atomistic simulations, NIPs face challenges in directly predicting electronic properties and often require coupling to higher-scale models or extensive simulations for macroscopic properties. Machine learning (ML) offers alternatives for structure-to-property mapping but faces trade-offs: feature-based methods often lack generalizability, while deep neural networks require significant data and computational power. To address these trade-offs, we introduce HackNIP, a two-stage pipeline that leverages pretrained NIPs. This method first extracts fixed-length feature vectors (embeddings) from NIP foundation models and then uses these embeddings to train shallow ML models for downstream structure-to-property predictions. This study investigates whether such a hybridization approach, by ``hacking" the NIP, can outperform end-to-end deep neural networks, determines the dataset size at which this transfer learning approach surpasses direct fine-tuning of the NIP, and identifies which NIP embedding depths yield the most informative features. HackNIP is benchmarked on Matbench, evaluated for data efficiency, and tested on diverse tasks including \textit{ab initio}, experimental, and molecular properties. We also analyze how embedding depth impacts performance. This work demonstrates a hybridization strategy to overcome ML trade-offs in materials science, aiming to democratize high-performance predictive modeling.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 40 canonical work pages

  1. [15]

    & Mizukami, W

    Shiota, T., Ishihara, K. & Mizukami, W. Universal neural network potentials as descriptors: towards scalable chemical property prediction using quantum and classical computers.Digital Discovery3, 1714–1728 (2024). URL https:// pubs.rsc.org/en/content/articlelanding/2024/dd/d4dd00098f. Publisher: Royal Society of Chemistry

    work page 2024

  2. [1]

    & Sutskever, I

    Radford, A., Narasimhan, K., Salimans, T. & Sutskever, I. Improving Language Understanding by Generative Pre-Training (2018). 20

    work page 2018

  3. [2]

    Radford, A.et al.Language models are unsupervised multitask learners.OpenAI blog1, 9 (2019)

    work page 2019

  4. [3]

    & Zhang, T

    Radford, A.et al.Meil˘ a, M. & Zhang, T. (eds)Learning transferable visual models from natural language supervision. (eds Meil˘ a, M. & Zhang, T.)International conference on machine learning, 8748–8763. PMLR (PMLR, 2021)

    work page 2021

  5. [4]

    Touvron, H.et al.Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

    arXiv 2023

  6. [5]

    Touvron, H.et al.Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288(2023)

    arXiv 2023

  7. [6]

    M., Schwaller, P., Ortega-Guerrero, A

    Jablonka, K. M., Schwaller, P., Ortega-Guerrero, A. & Smit, B. Leveraging large language models for predictive chemistry.Nature Machine Intelligence1–9 (2024). Publisher: Nature Publishing Group UK London

    work page 2024

  8. [7]

    URL https://www.nature.com/articles/s41586-023-06735-9

    Merchant, A.et al.Scaling deep learning for materials discovery.Nature 624, 80–85 (2023). URL https://www.nature.com/articles/s41586-023-06735-9. Publisher: Nature Publishing Group

    work page 2023

Show all 52 references
  1. [8]

    URL https://www.microsoft.com/en-us/research/publication/ mattersim-a-deep-learning-atomistic-model-across-elements-temperatures-and-pressures/

    Yang, H.et al.MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures (2024). URL https://www.microsoft.com/en-us/research/publication/ mattersim-a-deep-learning-atomistic-model-across-elements-temperatures-and-pressures/

    2024

  2. [9]

    Publisher: Nature Publishing Group UK London

    Deng, B.et al.CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence5, 1031–1041 (2023). Publisher: Nature Publishing Group UK London

    2023

  3. [10]

    URL http://arxiv.org/abs/2410.12771

    Barroso-Luque, L.et al.Open Materials 2024 (OMat24) Inorganic Mate- rials Dataset and Models (2024). URL http://arxiv.org/abs/2410.12771. ArXiv:2410.12771 [cond-mat]

    2024 arXiv

  4. [11]

    P., Simm, G., Ortner, C

    Batatia, I., Kovacs, D. P., Simm, G., Ortner, C. & Csanyi, G. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields.Advances in Neural Information Processing Systems35, 11423– 11436 (2022). URL https://proceedings.neurips.cc/paper fi...

    2022

  5. [12]

    M., Das, A

    Liao, Y.-L., Wood, B. M., Das, A. & Smidt, T. LeCun, Y. (ed.)EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations. (ed.LeCun, Y.) (2023). URL https://openreview.net/forum?id=mCOBKZmrzD

    2023

  6. [13]

    URL http://arxiv.org/abs/2410.22570

    Neumann, M.et al.Orb: A Fast, Scalable Neural Network Potential (2024). URL http://arxiv.org/abs/2410.22570. ArXiv:2410.22570 [cond-mat]. 21

    2024 arXiv

  7. [14]

    URL http://arxiv

    Rhodes, B.et al.Orb-v3: atomistic simulation at scale (2025). URL http://arxiv. org/abs/2504.06231. ArXiv:2504.06231 [cond-mat]

    2025 arXiv

  8. [16]

    & Jain, A

    Dunn, A., Wang, Q., Ganose, A., Dopp, D. & Jain, A. Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm.npj Computational Materials6, 1–10 (2020). URL https://www. nature.com/articles/s41524-020-00406-3. Publisher: Nat...

    2020

  9. [17]

    URL https://www.nature.com/articles/s41524-024-01469-2

    Zheng, H.et al.The ab initio non-crystalline structure database: empowering machine learning to decode diffusivity.npj Computational Materials10, 1–11 (2024). URL https://www.nature.com/articles/s41524-024-01469-2. Publisher: Nature Publishing Group

    2024

  10. [18]

    & Friederich, P

    Sommer, T., Willa, R., Schmalian, J. & Friederich, P. 3DSC - a dataset of superconductors including crystal structures.Scientific Data10, 816 (2023). URL https://www.nature.com/articles/s41597-023-02721-y. Publisher: Nature Publishing Group

    2023

  11. [19]

    C.et al.Data-driven electrolyte design for lithium metal anodes

    Kim, S. C.et al.Data-driven electrolyte design for lithium metal anodes. Proceedings of the National Academy of Sciences120, e2214357120 (2023)

    2023

  12. [20]

    & Denny, R

    Subramanian, G., Ramsundar, B., Pande, V. & Denny, R. A. Computational Modeling ofβ-Secretase 1 (BACE-1) Inhibitors Using Ligand Based Approaches. Journal of Chemical Information and Modeling56, 1936–1949 (2016). URL https: //doi.org/10.1021/acs.jcim.6b00290. Publisher: Americ...

    2016 doi

  13. [21]

    F., Teixeira, A

    Martins, I. F., Teixeira, A. L., Pinheiro, L. & Falcao, A. O. A Bayesian Approach to in Silico Blood-Brain Barrier Penetration Modeling.Journal of Chemical Information and Modeling52, 1686–1697 (2012). URL https://doi.org/10.1021/ ci300124c. Publisher: American Chemical Society

    2012

  14. [22]

    M., Madhukar, N

    Gayvert, K. M., Madhukar, N. S. & Elemento, O. A Data-Driven Approach to Predicting Successes and Failures of Clinical Trials.Cell Chemical Biology 23, 1294–1301 (2016). URL https://www.sciencedirect.com/science/article/pii/ S2451945616302914

    2016

  15. [23]

    URL https://pubs.rsc.org/en/content/articlelanding/ 2018/sc/c7sc02664a

    Wu, Z.et al.MoleculeNet: a benchmark for molecular machine learning.Chemical Science9, 513–530 (2018). URL https://pubs.rsc.org/en/content/articlelanding/ 2018/sc/c7sc02664a. Publisher: Royal Society of Chemistry. 22

    2018

  16. [24]

    Stanev, V.et al.Machine learning modeling of superconducting critical temperature.npj Computational Materials4, 29 (2018)

    2018

  17. [25]

    Computational Materials Science152, 60–69 (2018)

    Ward, L.et al.Matminer: An open source toolkit for materials data mining. Computational Materials Science152, 60–69 (2018)

    2018

  18. [26]

    & Barati Farimani, A

    Wang, Y., Wang, J., Cao, Z. & Barati Farimani, A. Molecular contrastive learning of representations via graph neural networks.Nature Machine Intelligence4, 279–287 (2022)

    2022

  19. [27]

    Publisher: Nature Publishing Group

    Fang, X.et al.Geometry-enhanced molecular representation learning for property prediction.Nature Machine Intelligence4, 127–134 (2022). Publisher: Nature Publishing Group

    2022

  20. [28]

    Olson, G. B. Computational design of hierarchically structured materials.Science 277, 1237–1242 (1997)

    1997

  21. [29]

    Fine, M. E. & Marcus, H. L. Materials science and engineering, an educational discipline.Annual Review of Materials Science24, 1–19 (1994)

    1994

  22. [30]

    S.et al.Engineering atomic-level complexity in high-entropy and complex concentrated alloys.Nature communications10, 2090 (2019)

    Oh, H. S.et al.Engineering atomic-level complexity in high-entropy and complex concentrated alloys.Nature communications10, 2090 (2019)

    2019

  23. [31]

    Han, L.et al.Multifunctional high-entropy materials.Nature Reviews Materials 9, 846–865 (2024)

    2024

  24. [32]

    You, Y.et al.Graph contrastive learning with augmentations.Advances in neural information processing systems33, 5812–5823 (2020)

    2020

  25. [33]

    J., Kim, H., Jo, J

    Park, Y. J., Kim, H., Jo, J. & Yoon, S. Deep contrastive learning of molecular conformation for efficient property prediction.Nature Computational Science3, 1015–1022 (2023). URL https://www.nature.com/articles/s43588-023-00560-w. Publisher: Nature Publishing Group

    2023

  26. [34]

    SMILES, a chemical language and information system

    Weininger, D. SMILES, a chemical language and information system. 1. Intro- duction to methodology and encoding rules.Journal of chemical information and computer sciences28, 31–36 (1988). Publisher: ACS Publications

    1988

  27. [35]

    & Weininger, J

    Weininger, D., Weininger, A. & Weininger, J. L. SMILES. 2. Algorithm for generation of unique SMILES notation.Journal of chemical information and computer sciences29, 97–101 (1989). Publisher: ACS Publications

    1989

  28. [36]

    & Ramsundar, B

    Ahmad, W., Simon, E., Chithrananda, S., Grand, G. & Ramsundar, B. Chemberta-2: Towards chemical foundation models.arXiv preprint arXiv:2209.01712(2022)

    2022 arXiv

  29. [37]

    Liu, Y.et al.Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692(2019). 23

    2019 arXiv

  30. [38]

    & Ngom, A

    Sadeghi, S., Bui, A., Forooghi, A., Lu, J. & Ngom, A. Comparative Analysis of LLaMA and ChatGPT Embeddings for Molecule Embedding.arXiv preprint arXiv:2402.00024(2024)

    2024 arXiv

  31. [39]

    & Barati Farimani, A

    Magar, R., Wang, Y. & Barati Farimani, A. Crystal twins: self-supervised learning for crystalline material property prediction.npj Computational Materials8, 231 (2022). Publisher: Nature Publishing Group UK London

    2022

  32. [40]

    & Deny, S

    Zbontar, J., Jing, L., Misra, I., LeCun, Y. & Deny, S. Meil˘ a, M. & Zhang, T. (eds)Barlow twins: Self-supervised learning via redundancy reduction. (eds Meil˘ a, M. & Zhang, T.)International Conference on Machine Learning, 12310–12320 (PMLR, 2021)

    2021

  33. [41]

    Chen, X. & He, K. Exploring Simple Siamese Representation Learning (2020). URL http://arxiv.org/abs/2011.10566. ArXiv:2011.10566 [cs]

    2020 arXiv

  34. [42]

    Koker, T., Quigley, K., Spaeth, W., Frey, N. C. & Li, L. Graph Contrastive Learning for Materials.arXiv preprint arXiv:2211.13408(2022)

    2022 arXiv

  35. [43]

    & Parrinello, M

    Behler, J. & Parrinello, M. Generalized Neural-Network Representation of High- Dimensional Potential-Energy Surfaces.Physical Review Letters98, 146401 (2007). URL https://link.aps.org/doi/10.1103/PhysRevLett.98.146401. Pub- lisher: American Physical Society

    2007 doi

  36. [44]

    P., Payne, M

    Bart´ ok, A. P., Payne, M. C., Kondor, R. & Cs´ anyi, G. Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons.Phys- ical Review Letters104, 136403 (2010). URL https://link.aps.org/doi/10.1103/ PhysRevLett.104.136403. Publisher: American P...

    2010

  37. [45]

    G., Hoogeboom, E

    Satorras, V. G., Hoogeboom, E. & Welling, M. Meil˘ a, M. & Zhang, T. (eds)E (n) equivariant graph neural networks. (eds Meil˘ a, M. & Zhang, T.)International conference on machine learning, 9323–9332 (PMLR, 2021)

    2021

  38. [46]

    Batzner, S.et al.E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nature communications13, 2453 (2022)

    2022

  39. [47]

    D.et al.Benchmarking machine learning models for predicting lithium ion migration.npj Computational Materials11, 1–12 (2025)

    Dembitskiy, A. D.et al.Benchmarking machine learning models for predicting lithium ion migration.npj Computational Materials11, 1–12 (2025). URL https: //www.nature.com/articles/s41524-025-01571-z. Publisher: Nature Publishing Group

    2025

  40. [48]

    URL https://www.nature.com/articles/s41467-021-26921-5

    Gupta, V.et al.Cross-property deep transfer learning framework for enhanced predictive analytics on small materials data.Nature Communications12, 6595 (2021). URL https://www.nature.com/articles/s41467-021-26921-5. Publisher: Nature Publishing Group. 24

    2021

  41. [49]

    & Ong, S

    Chen, C. & Ong, S. P. A universal graph deep learning interatomic potential for the periodic table.Nature Computational Science2, 718–728 (2022). URL https: //www.nature.com/articles/s43588-022-00349-3. Publisher: Nature Publishing Group

    2022

  42. [50]

    URL https://www.nature.com/articles/ s41524-025-01606-5

    Minami, S.et al.Scaling Law of Sim2Real transfer learning in expanding computational materials databases for real-world predictions.npj Compu- tational Materials11, 1–10 (2025). URL https://www.nature.com/articles/ s41524-025-01606-5. Publisher: Nature Publishing Group

    2025

  43. [51]

    & Rignanese, G.-M

    De Breuck, P.-P., Hautier, G. & Rignanese, G.-M. Materials property prediction for limited datasets enabled by feature selection and joint learning with MOD- Net.npj Computational Materials7, 1–8 (2021). URL https://www.nature.com/ articles/s41524-021-00552-2. Publisher: Natur...

    2021

  44. [52]

    & Koyama, M

    Akiba, T., Sano, S., Yanase, T., Ohta, T. & Koyama, M. Krishnapuram, B.et al. (eds)Optuna: A next-generation hyperparameter optimization framework. (eds Krishnapuram, B.et al.)Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 262...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.