Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Distillation of atomistic foundation models across architectures and chemical domains

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims a general three-step protocol that distils any atomistic foundation model into a fast, domain-specific machine-learning potential using only tens of DFT-labelled structures and about ten thousand teacher-labelled…

desk verdict The water proof-of-concept is real and the MD-free distillation pipeline is a useful engineering contribution, but the abstract oversells reactive and extreme-condition applications that fail exactly where the teacher is wrong. read the letter →

arxiv 2506.10956 v1 pith:ZPIS7UQG submitted 2025-06-12 physics.comp-ph

John L. A. Gardner , Daniel F. Thomas du Toit , Chiheb Ben Mahmoud , Zoé Faure Beaulieu , Veronika Juraskova , Laura-Bianca Paşca , Louise A. M. Rosset , Fernanda Duarte
show 3 more authors
Fausto Martelli Chris J. Pickard Volker L. Deringer
This is my paper
classification physics.comp-ph
keywords atomisticfoundationmodelsknowledgedistillationmachine-learnedinteratomicpotentialssyntheticdatafine-tuningmoleculardynamicsatomicclusterexpansiongraphneuralnetworks
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 claims that the knowledge in a large atomistic foundation model can be transferred into a small, fast machine-learned interatomic potential by a three-step protocol: fine-tune the foundation model on a few dozen domain-specific DFT-labelled structures, use the fine-tuned model to generate and label about ten thousand synthetic structures, and train a compact student potential on those labels. The authors demonstrate distilled water potentials with DFT force errors of $37$--$51$ meV/Å, compared with $32$ meV/Å for the fine-tuned teacher, at over ten times higher inference speed for graph-network students and over one hundred times for the atomic cluster expansion student. They apply the same recipe to liquid water, dense metallic hydrogen, porous silica, a hybrid perovskite, and an organic S$_N$2 reaction, and report stable molecular dynamics with the distilled models. If correct, the protocol turns expensive foundation models into cheap, domain-specific potentials that can be trained in hours on a single GPU.

What carries the argument

The load-bearing mechanism is the rattle-relax-repeat synthetic-data generator, implemented in open-sourced, GPU-accelerated code. Starting from a small seed set, each iteration samples a parent structure with probability mixing a Boltzmann energy weight and a generation bias, rattles positions and cell by Gaussian noise, relaxes with the fine-tuned teacher's forces using a Robbins-Monro-style update, and adds the child to a family tree; each new structure is labelled by the teacher, not DFT. This cheaply produces roughly one uncorrelated structure per second and about five model calls per structure, at least an order of magnitude cheaper than MD-based sampling. The student models are then fitted with PaiNN, TensorNet, ACE, or EDDP architectures with around 40k parameters or fewer, and the paper shows that smaller radial cut-offs than the teacher's can be used without much accuracy loss.

What would settle it

Run a well-tempered metadynamics simulation of the distilled S$_N$2 PaiNN model past the reported 600 ps and compare the free-energy surface with direct DFT calculations on the same collective variables; persistent unphysical Cl-I or C-I contact structures that DFT would rule out would show the teacher's blind spot propagating through distillation. Alternatively, compute the student's force MAE relative to DFT on hydrogen configurations above 1.6 g/cm$^3$, where the paper already reports divergence from the ground truth.

Watch

Extended reading notes

Core claim

The central claim is that an atomistic foundation model, after a short fine-tuning on as few as 25 domain-specific structures with DFT labels, can serve as a teacher that labels a large synthetic dataset, and that any smaller architecture trained on those labels inherits the teacher's accuracy while being much faster. The proof-of-concept distils MACE-MP-0b3 into TensorNet, PaiNN, and ACE water potentials; the students show component-wise force MAEs of $37$, $39$, and $51$ meV/Å against DFT, close to the teacher's $32$ meV/Å, and remain stable in molecular dynamics. The claim is that this transfer works across architectures and across chemical domains because the synthetic structures are generated by a rattle-relax-repeat protocol that explores the local potential-energy landscape without running MD. The authors also report that student errors relative to the teacher are negatively correlated with teacher errors relative to DFT, so distilled students tend to land closer to the true surface than a naive error analysis would predict.

Load-bearing premise

The protocol assumes the fine-tuned teacher's synthetic labels are a faithful surrogate for the true potential-energy surface everywhere the student will be used; the paper's S$_N$2 application shows this premise can fail, since both teacher and student produce unphysical products after 600 ps in regions outside the teacher's training set.

Editorial extensions

If this is right

  • Users can obtain near-foundation-model accuracy on a desired chemical system after fine-tuning with about 25 DFT-labelled structures, without needing server-grade hardware.
  • Distilled student potentials run 10x to 100x faster than the teacher, and can scale to system sizes where the foundation model runs out of memory.
  • The recipe is architecture-agnostic: any ASE-compatible teacher model can be distilled, and any fitting software can produce the student.
  • The full workflow from foundation model to distilled potential took about 7.5 hours on a single mid-level GPU for the water case.
  • Distillation transfers accuracy but also inherits the teacher's systematic blind spots, so out-of-domain behaviour must be checked for each application.

Reading between the lines

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

  • A testable extension is whether the reported negative correlation between student-vs-teacher and teacher-vs-DFT errors holds for other foundation models and domains; if it is general, distillation would systematically improve data efficiency beyond what the teacher's own error would suggest.
  • The protocol could be pointed at beyond-DFT ground truth, such as coupled-cluster labels, by fine-tuning a teacher on a very small set, which the paper notes is attractive for expensive reference methods.
  • Because the synthetic data are generated without MD, the method avoids requiring a stable teacher for sampling, but the family-tree protocol must cover the intended application region or the student inherits the teacher's blind spots.
  • An open comparison is distillation versus direct training on actively selected DFT labels at equal labelling cost; the paper does not perform this comparison, so it remains unclear how much of the speed-up comes from the teacher's inductive bias rather than simply from having more effective training points.
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 / 6 minor

Summary. The manuscript proposes a three-step protocol for distilling atomistic foundation models (FMs) into smaller, faster student potentials: (1) fine-tune an FM on a small number of DFT-labelled structures; (2) generate a large synthetic dataset by rattling and relaxing structures with the fine-tuned teacher; (3) train student models of arbitrary architecture on the synthetic labels. The proof-of-concept on liquid water shows distilled PaiNN, TensorNet, and ACE models with DFT force MAEs of 37–51 meV/Å versus 32 meV/Å for the teacher, at 10–100× speed-ups and stable MD. Ablation studies examine synthetic-data scaling, cutoff radii, and fine-tuning set size. Applications to hydrogen, porous silica, MAPI, and an SN2 reaction in explicit solvent are reported. The SN2 and high-density hydrogen cases fail outside the teacher's domain, and the paper acknowledges these limitations in the applications section.

Significance. If validated, the protocol would be practically valuable: it lets users obtain near-FM accuracy on modest hardware with only tens of DFT-labelled structures, and the authors provide open-source implementations (augment-atoms, graph-pes, XPOT). The paper's strengths include the hold-out DFT test sets, the comparison against directly trained models, the architecture diversity of students (PaiNN, TensorNet, ACE, EDDP), the multi-FM evaluation (MACE-MP-0b3, MatterSim, orb-v3, MACE-OFF24), and the physics-based validation of liquid water (RDF, ring statistics, hydrogen-bonding). The reported anti-correlation in Fig. 4b is a thought-provoking observation, though its interpretation remains unresolved. The main reservation is that the demonstrated successes are confined to domains where the teacher is accurate; the reactive and extreme-condition examples show the student inherits teacher errors, so the 'any domain' claim overstates the evidence.

major comments (3)
  1. [Applications (An organic reaction in explicit solvent; Metallic hydrogen)] The abstract and Introduction claim the protocol 'quickly produces fast, accurate, and stable student MLIPs' for 'any user-specified chemical domain.' The paper's own applications contradict this for reactive and out-of-domain cases. In the SN2 example, the distilled PaiNN model yields unphysical reaction products after about 600 ps, and the authors state that 'the teacher and student models used here still suffer from the inaccurate description of strong attractive and repulsive interactions... which are outside of the teacher's training set.' The metallic-hydrogen EOS diverges from the DFT reference for densities above the 1.6 g/cm3 fine-tuning threshold (Fig. 5a), with several MD trajectories failing despite reduced timesteps. These results demonstrate that the student's accuracy is bounded by the teacher's and that the protocol does not, as written, 'target any user-specified chemical domain.' The central claim should be qualified to specify the condition that the teacher is reliable on the target domain, and the manuscript should provide guidance for detecting out-of-domain operation.
  2. [Methods (Synthetic data generation)] The synthetic-data generator relaxes rattled structures using the teacher's own forces. This concentrates the generated dataset in minima of the teacher's potential-energy surface. If the teacher is inaccurate in a region—e.g., the over-stabilized carbocation seen for MACE-OFF23 in the SN2 attempt, or the repulsive Cl–I interactions attributed to MACE-OFF24—those incorrect regions are labeled and fed to the student, which learns them faithfully. The protocol therefore has no mechanism to detect or correct teacher error, and the 'general, architecture-agnostic' claim requires the additional assumption that the teacher is accurate across the intended operating domain. The authors should state this assumption explicitly and, ideally, propose a practical check (e.g., comparison with a small DFT validation set or an uncertainty metric) to flag regions where the teacher is unreliable.
  3. [Ablation studies (Fig. 4b)] In §Ablation (Fig. 4b), the authors report a negative correlation between the teacher's error against DFT and the student's error against the teacher, and interpret it as 'inherent bias towards the true PES' that 'improves the data efficiency of synthetic distillation.' No correlation coefficient, significance test, or error analysis is provided, and the origin is acknowledged to require 'a more systematic study.' Given that this anti-correlation is used to rationalize the method's data efficiency, the authors should either provide a quantitative statistical characterization (e.g., Spearman correlation with confidence intervals across architectures and dataset sizes) or explicitly label this as a preliminary observation that is not required for the protocol's practical validity.
minor comments (6)
  1. [Proof-of-concept (Fig. 1b) and Fig. 2] The speed-up factors (>10x, >100x) in Fig. 1b should specify the system size at which they were measured, since Fig. 2 shows that the teacher's memory requirements make the relative speed strongly system-size dependent.
  2. [Applications (An organic reaction in explicit solvent)] The sentence 'The fine-tuned MACE-OFF24 model accurately predicted the concerted SN2 mechanism. However, after 800 ps, dynamics yielded unphysical structures' is contradictory; please qualify 'accurately' (e.g., 'initially') or restructure the sentence.
  3. [Ablation studies (Fig. 4b)] The caption note that 'the remaining EDDP models behave similarly up to 400 meV/Å' is difficult to evaluate because the plot is truncated at 130 meV/Å; consider including the full range or a supplementary panel.
  4. [Methods (Synthetic data generation)] Please clarify how the generation number G_i is defined and whether the minimum atomic displacement of 0.2 Å between structures (cited in the Fig. 6 caption) is enforced explicitly during sampling.
  5. [Proof-of-concept and Ablation studies] The paper reports only 'best-of-three' results without variance; reporting mean and standard deviation across random seeds would provide a more robust comparison.
  6. [Discussion] The statement that the approach 'can provide improved access to accurate, performant, and robust MLIPs' should be tempered given the acknowledged failures in reactive and extreme-condition domains, as those examples show the student models are not robust beyond the teacher's reliable domain.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: students are trained on teacher labels by construction, but accuracy, speed, and stability claims are validated against independent DFT data and physical benchmarks.

full rationale

The paper's derivation chain is: fine-tune a foundation model on a small DFT-labelled set; generate and label a larger synthetic set with the fine-tuned teacher; train smaller student models on that synthetic set; and evaluate students on a held-out DFT test set plus physical MD validation. Training students to reproduce teacher labels is the definition of distillation, not a disguised restatement of the paper's conclusions. The paper explicitly states that the fine-tuned FM, 'not DFT!', labels the synthetic structures (Methods, Synthetic data generation), and separately reports student force MAEs against held-out DFT data (37/39/51 meV/Å for TensorNet/PaiNN/ACE vs 32 meV/Å for the fine-tuned teacher). Speed-ups are measured directly on hardware (Fig. 2), and stability is assessed via RDFs, ring statistics, tetrahedral order, and hydrogen-bonding analyses (Fig. 3), none of which are encoded in the training objective. The ablation 'scaling laws' are empirical: the reported anti-correlation between F_FM−F_DFT and F_distilled−F_FM is a measured property, not an input assumption. The SN2 and high-density hydrogen failures are explicitly acknowledged limitations ('the teacher and student models used here still suffer from the inaccurate description of strong attractive and repulsive interactions...'), which weaken the generality claim but do not make the derivation circular. Self-citations to prior synthetic-data and validation work (Refs 38, 39, 44, 59) support methodology, but the load-bearing validation here is against independent DFT and experimental data, so no circularity is present.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The protocol rests on the quality of teacher labels and the coverage of the synthetic sampling; no new physical entities are introduced. Several user-defined hyperparameters control the augmentation and are not fitted to the test set. The only hard assumption is that teacher synthetic labels approximate DFT well enough in the target region.

free parameters (7)
  • beta (explore/exploit weight)
    User-defined hyperparameter in parent sampling probability (Methods, P_i formula); value not stated in text.
  • sigma_A
    Standard deviation for cell rattling in augmentation protocol; user-defined, not fitted.
  • sigma_B
    Standard deviation for position rattling; also used in relaxation step size; user-defined.
  • M (max relaxation steps)
    Maximum number of relaxation steps in rattle-relax-repeat; user-defined.
  • synthetic dataset size = 10k (water), 5k (applications)
    Number of synthetic structures labeled by the teacher; chosen by hand; affects accuracy and cost.
  • radial cutoffs = 4.5 Å PaiNN/TensorNet, 5.5 Å ACE/EDDP, 2 Å hydrogen PaiNN
    Chosen by hand; ablation shows modest sensitivity for water.
  • dimer sampling interval = [0.6d, 3d]
    Interval for sampling dimer distances in synthetic set; chosen ad hoc.
assumptions (5)
  • domain assumption Fine-tuned FM labels are a valid proxy for DFT labels when training the student.
    The entire pipeline labels synthetic structures with the teacher, not DFT; if the teacher's PES is inaccurate in a region, the student inherits the error. Explicit failure in the SN2 application.
  • domain assumption Rattle-relax-repeat sampling covers the relevant configuration space of the target domain.
    The protocol explores by rattling and relaxing from a small seed set; if the seeds or perturbations miss important basins, the student will be incomplete. The hydrogen OOD divergence illustrates this limitation.
  • domain assumption Student architectures are expressive enough to represent the teacher's PES in the target domain.
    PaiNN, TensorNet, ACE, EDDP are used; performance varies (ACE over-structures water), so expressivity and inductive bias matter.
  • domain assumption The hold-out DFT test set is representative and independent.
    Error estimates rely on the test set, e.g., 1,563 water structures; if it is biased, the reported MAEs are misleading.
  • standard math Stochastic approximation relaxation (Robbins-Monro inspired) converges to a useful local minimum for each rattled structure.
    The update rule is heuristic with early stopping; no proof of coverage or convergence is given, but standard optimization assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Distillation of atomistic foundation models across architectures and chemical domains." pith.science (2026). https://pith.science/paper/ZPIS7UQG

@misc{pith2026250610956,
  author       = {Pith},
  title        = {Pith review of: Distillation of atomistic foundation models across architectures and chemical domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZPIS7UQG}},
  note         = {Machine review of arXiv:2506.10956}
}
abstract

Machine-learned interatomic potentials have transformed computational research in the physical sciences. Recent atomistic `foundation' models have changed the field yet again: trained on many different chemical elements and domains, these potentials are widely applicable, but comparably slow and resource-intensive to run. Here we show how distillation via synthetic data can be used to cheaply transfer knowledge from atomistic foundation models to a range of different architectures, unlocking much smaller, more efficient potentials. We demonstrate speed-ups of $> 10\times$ by distilling from one graph-network architecture into another, and $> 100\times$ by leveraging the atomic cluster expansion framework. We showcase applicability across chemical and materials domains: from liquid water to hydrogen under extreme conditions; from porous silica and a hybrid halide perovskite solar-cell material to modelling organic reactions. Our work shows how distillation can support the routine and computationally efficient use of current and future atomistic foundation models in real-world scientific research.

Figures

Figures reproduced from arXiv: 2506.10956 by the authors.

Figure 1
Figure 1. Distilling atomistic foundation models via synthetic data. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Computational efficiency. We show the inference costs for different water models. Speeds are measured during 300 K NVT simulations run on a single Nvidia RTX A6000 GPU. The (fine￾tuned) MACE-MP-0b3 model is memory-intensive, and quickly runs out of memory (OOM) as the system size increases beyond 1,000 atoms. In contrast, the smaller, distilled GNN models allow us to scale up system size by up to another 2 orders of… view at source ↗
Figure 3
Figure 3. Physics-guided validation of distilled MLIPs. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Ablation studies. All panels relate to the MACE-MP-0b3 foundation model (FM) fine-tuned on water and distilled into various MLIP architectures (see Proof-of-concept section). (a) Learning curves. We test the number of synthetic data labels required for distillation: en…
Figure 5
Figure 5. Figure 5: Applications of distilled MLIPs across different domains. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Synthetic data generation. (a) Our protocol creates a ‘family tree’ of structures for each seed structure: in all panels, we colour-code each structure based on its generation. By incorporating an element of exploration into our structure selection step, we preferentia…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Heterogeneous Ensemble Enables a Universal Uncertainty Metric for Atomistic Foundation Models

    cond-mat.mtrl-sci 2025-07 conditional novelty 5.0 of 10

    Weights from OMat24 force errors turn an eleven-model heterogeneous uMLIP ensemble into an uncertainty metric U that correlates with true force errors across material families and drives low-DFT distillation.

Reference graph

Works this paper leans on

115 extracted references · 54 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bommasani, R. et al. On the opportunities and risks of foundation models. ArXiv (2021). URL https://crfm.stanford.edu/assets/report.pdf

    2021

  2. [2]

    Brown, T. B. et al. Language Models are Few-Shot Learners (2020). Preprint athttps: //arxiv.org/abs/2005.14165

    arXiv 2020

  3. [3]

    Kirillov, A. et al. Segment Anything (2023). Preprint at https://arxiv.org/abs/ 2304.02643

    arXiv 2023

  4. [4]

    Zhou, Y . et al. A foundation model for generalizable disease detection from retinal images. Nature 622, 156–163 (2023)

    2023

  5. [5]

    Parker, L. et al. AstroCLIP: A cross-modal foundation model for galaxies. Monthly Notices of the Royal Astronomical Society 531, 4990–5011 (2024)

    2024

  6. [6]

    Hollmann, N. et al. Accurate predictions on small data with a tabular foundation model. Nature 637, 319–326 (2025)

    2025

  7. [7]

    Fu, X. et al. A foundation model of transcription across human cell types. Nature 637, 965–973 (2025)

    2025

  8. [8]

    Pyzer-Knapp, E. O. et al. Foundation models for materials discovery – current state and future directions. npj Comput. Mater. 11, 61 (2025)

    2025

Show all 115 references
  1. [9]

    Bodnar, C. et al. A foundation model for the Earth system. Nature 641, 1180–1187 (2025)

    2025

  2. [10]

    Batatia, I. et al. A foundation model for atomistic materials chemistry (2024). Preprint at https://arxiv.org/abs/2401.00096

    2024 arXiv

  3. [11]

    Yang, H. et al. MatterSim: A Deep Learning Atomistic Model Across Elements, Tem- peratures and Pressures (2024). Preprint at https://arxiv.org/abs/2405.04967

    2024 arXiv

  4. [12]

    Neumann, M. et al. Orb: A Fast, Scalable Neural Network Potential (2024). Preprint at https://arxiv.org/abs/2410.22570

    2024 arXiv

  5. [13]

    Rhodes, B. et al. Orb-v3: Atomistic simulation at scale (2025). Preprint at https: //arxiv.org/abs/2504.06231

    2025 arXiv

  6. [14]

    Barroso-Luque, L. et al. Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models (2024). Preprint at https://arxiv.org/abs/2410.12771

    2024 arXiv

  7. [15]

    S., Isayev, O

    Smith, J. S., Isayev, O. & Roitberg, A. E. ANI-1: An extensible neural network potential with DFT accuracy at force field computational cost. Chem. Sci. 8, 3192–3203 (2017). 23

    2017

  8. [16]

    S., Leszczynski, J

    Zubatyuk, R., Smith, J. S., Leszczynski, J. & Isayev, O. Accurate and transferable multitask prediction of chemical properties with an atoms-in-molecules neural network. Sci. Adv. 5, eaav6490 (2019)

    2019

  9. [17]

    Kov ´acs, D. P. et al. MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules. J. Am. Chem. Soc. 147, 17598–17611 (2025)

    2025

  10. [18]

    Shoghi, N. et al. From Molecules to Materials: Pre-training Large Generalizable Mod- els for Atomic Property Prediction (2024). Preprint at https://arxiv.org/abs/ 2310.16802

    2024 arXiv

  11. [19]

    Zhang, S. et al. Exploring the frontiers of condensed-phase chemistry with a general reactive machine learning potential. Nat. Chem. 16, 727–734 (2024)

    2024

  12. [20]

    Zhang, D. et al. DPA-2: A large atomic model as a multi-task learner. npj Comput. Mater. 10, 293 (2024)

    2024

  13. [21]

    & Smidt, T

    Liao, Y .-L., Wood, B., Das, A. & Smidt, T. EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations (2024). Preprint at https: //arxiv.org/abs/2306.12059

    2024 arXiv

  14. [22]

    Fu, X. et al. Learning Smooth and Expressive Interatomic Potentials for Physical Prop- erty Prediction (2025). Preprint at https://arxiv.org/abs/2502.12147

    2025 arXiv

  15. [23]

    Batatia, I. et al. The design space of E(3)-equivariant atom-centred interatomic poten- tials. Nat. Mach. Intell. 7, 56–67 (2025)

    2025

  16. [24]

    & Simoncelli, M

    P ´ota, B., Ahlawat, P., Cs ´anyi, G. & Simoncelli, M. Thermal Conductivity Predic- tions with Foundation Atomistic Models (2024). Preprint at https://arxiv.org/ abs/2408.00755

    2024 arXiv

  17. [25]

    Riebesell, J. et al. Matbench Discovery – A framework to evaluate machine learn- ing crystal stability predictions (2024). Preprint at https://arxiv.org/abs/2308. 14920

    2024

  18. [26]

    M., Mori, T

    Shiota, T., Ishihara, K., Do, T. M., Mori, T. & Mizukami, W. Taming Multi-Domain, -Fidelity Data: Towards Foundation Models for Atomistic Scale Simulations (2024). Preprint at https://arxiv.org/abs/2412.13088

    2024

  19. [27]

    Schmidt, J., Wang, H.-C., Cerqueira, T. F. T., Botti, S. & Marques, M. A. L. A dataset of 175k stable and metastable materials calculated with the PBEsol and SCAN functionals. Sci. Data 9, 64 (2022)

    2022

  20. [28]

    Kaplan, A. D. et al. A Foundational Potential Energy Surface Dataset for Materials (2025). Preprint at https://arxiv.org/abs/2503.04070. 24

    2025

  21. [29]

    Levine, D. S. et al. The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models (2025). Preprint at https://arxiv.org/abs/2505.08762

    2025

  22. [30]

    & Dean, J

    Hinton, G., Vinyals, O. & Dean, J. Distilling the Knowledge in a Neural Network. arXiv:1503.02531 [cs, stat] (2015). Preprint at https://arxiv.org/abs/1503. 02531

    2015 arXiv

  23. [31]

    Allen-Zhu, Z. & Li, Y . Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learning (2023). Preprint at https://arxiv.org/abs/ 2012.09816

    2023 arXiv

  24. [32]

    & Zhang, Q

    Cheng, X., Rao, Z., Chen, Y . & Zhang, Q. Explaining Knowledge Distillation by Quan- tifying the Knowledge (2020). Preprint at https://arxiv.org/abs/2003.03622

    2020 arXiv

  25. [33]

    Touvron, H. et al. LLaMA: Open and Efficient Foundation Language Models (2023). Preprint at https://arxiv.org/abs/2302.13971

    2023 arXiv

  26. [34]

    & Gao, J

    Peng, B., Li, C., He, P., Galley, M. & Gao, J. Instruction Tuning with GPT-4 (2023). Preprint at https://arxiv.org/abs/2304.03277

    2023 arXiv

  27. [35]

    Liu, Z. et al. Wisdom of Committee: Distilling from Foundation Model to Special- izedApplication Model (2024). Preprint at https://arxiv.org/abs/2402.14035

    2024 arXiv

  28. [36]

    Vemulapalli, R. et al. Knowledge Transfer from Vision Foundation Models for Effi- cient Training of Small Task-specific Models. In Proceedings of the 41st International Conference on Machine Learning, 49345–49367 (2024)

    2024

  29. [37]

    A data-efficient strategy for building high-performing medical foundation models

    Sun, Y .et al. A data-efficient strategy for building high-performing medical foundation models. Nat. Biomed. Eng. 9, 539–551 (2025)

    2025

  30. [38]

    Gardner, J. L. A., Beaulieu, Z. F. & Deringer, V . L. Synthetic data enable experiments in atomistic machine learning. Digital Discovery 2, 651–662 (2023)

    2023

  31. [39]

    Morrow, J. D. & Deringer, V . L. Indirect learning and physically guided validation of interatomic potential models. J. Chem. Phys. 157 (2022)

    2022

  32. [40]

    Matin, S. et al. Teacher-student training improves accuracy and efficiency of machine learning inter-atomic potentials (2025). Preprint at https://arxiv.org/abs/2502. 05379

    2025

  33. [41]

    & Krishnapriyan, A

    Amin, I., Raja, S. & Krishnapriyan, A. Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians (2025). Preprint at https://arxiv.org/abs/2501.09009

    2025 arXiv

  34. [42]

    & Gasteiger, J

    Ekstr ¨om Kelvinius, F., Georgiev, D., Toshev, A. & Gasteiger, J. Accelerating Molecular 25 Graph Neural Networks via Knowledge Distillation. Advances in Neural Information Processing Systems 36, 25761–25792 (2023)

    2023

  35. [43]

    Matin, S. et al. Ensemble Knowledge Distillation for Machine Learning Interatomic Potentials (2025). Preprint at https://arxiv.org/abs/2503.14293

    2025 arXiv

  36. [44]

    Gardner, J. L. A., Baker, K. T. & Deringer, V . L. Synthetic pre-training for neural- network interatomic potentials. Mach. Learn.: Sci. Technol. 5, 015003 (2024)

    2024

  37. [45]

    & Zhong, Z

    Wang, R., Gao, Y ., Wu, H. & Zhong, Z. PFD: Automatically Generating Machine Learning Force Fields from Universal Models (2025). Preprint at https://arxiv. org/abs/2502.20809

    2025

  38. [46]

    & Parrinello, M

    Behler, J. & Parrinello, M. Generalized Neural-Network Representation of High- Dimensional Potential-Energy Surfaces. Phys. Rev. Lett. 98, 146401 (2007)

    2007

  39. [47]

    P., Payne, M

    Bart ´ok, A. P., Payne, M. C., Kondor, R. & Cs ´anyi, G. Gaussian Approximation Po- tentials: The Accuracy of Quantum Mechanics, without the Electrons. Phys. Rev. Lett. 104, 136403 (2010)

    2010

  40. [48]

    Batzner, S. et al. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nat. Commun. 13, 2453 (2022)

    2022

  41. [49]

    & Ong, S

    Chen, C. & Ong, S. P. A universal graph deep learning interatomic potential for the periodic table. Nat. Comput. Sci. 2, 718–728 (2022)

    2022

  42. [50]

    P., Simm, G

    Batatia, I., Kov ´acs, D. P., Simm, G. N. C., Ortner, C. & Cs´anyi, G. MACE: Higher Or- der Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields (2023). Preprint at https://arxiv.org/abs/2206.07697

    2023 arXiv

  43. [51]

    Merchant, A. et al. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023)

    2023

  44. [52]

    Deng, B. et al. CHGNet as a pretrained universal neural network potential for charge- informed atomistic modelling. Nat. Mach. Intell. 5, 1031–1041 (2023)

    2023

  45. [53]

    & Scuseria, G

    Ernzerhof, M. & Scuseria, G. E. Assessment of the Perdew–Burke–Ernzerhof exchange-correlation functional. J. Chem. Phys. 110, 5029–5036 (1999)

    1999

  46. [54]

    & Barone, V

    Adamo, C. & Barone, V . Toward reliable density functional methods without adjustable parameters: The PBE0 model. J. Chem. Phys. 110, 6158–6170 (1999)

    1999

  47. [55]

    & Krieg, H

    Grimme, S., Antony, J., Ehrlich, S. & Krieg, H. A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu. J. Chem. Phys. 132 (2010). 26

    2010

  48. [56]

    & Grimme, S

    Goerigk, L. & Grimme, S. A thorough benchmark of density functional methods for general main group thermochemistry, kinetics, and noncovalent interactions. Phys. Chem. Chem. Phys. 13, 6670 (2011)

    2011

  49. [57]

    A., Behler, J., Dellago, C

    Cheng, B., Engel, E. A., Behler, J., Dellago, C. & Ceriotti, M. Ab initio thermodynam- ics of liquid and solid water. Proc. Natl. Acad. Sci. U. S. A. 116, 1110–1115 (2019)

    2019

  50. [58]

    Kaur, H. et al. Data-efficient fine-tuning of foundational models for first-principles quality sublimation enthalpies. Faraday Discuss. 256, 120–138 (2025)

    2025

  51. [59]

    D., Gardner, J

    Morrow, J. D., Gardner, J. L. A. & Deringer, V . L. How to validate machine-learned interatomic potentials. J. Chem. Phys. 158, 121501 (2023)

    2023

  52. [60]

    Errington, J. R. & Debenedetti, P. G. Relationship between structural order and the anomalies of liquid water. Nature 409, 318–321 (2001)

    2001

  53. [61]

    A., Jr., Santra, B., Li, Z., Wu, X

    DiStasio, R. A., Jr., Santra, B., Li, Z., Wu, X. & Car, R. The individual and collective effects of exact exchange and dispersion interactions on the ab initio structure of liquid water. J. Chem. Phys. 141, 084502 (2014)

    2014

  54. [62]

    B., Benmore, C

    Skinner, L. B., Benmore, C. J., Neuefeind, J. C. & Parise, J. B. The structure of water around the compressibility minimum. J. Chem. Phys. 141, 214507 (2014)

    2014

  55. [63]

    Pickard, C. J. Ephemeral data derived potentials for random structure search. Phys. Rev. B 106, 014102 (2022)

    2022

  56. [64]

    The Interiors of Giant Planets: Models and Outstanding Questions

    Guillot, T. The Interiors of Giant Planets: Models and Outstanding Questions. Annu. Rev. Earth Planet. Sci. 33, 493–530 (2005)

    2005

  57. [65]

    & Redmer, R

    Helled, R., Mazzola, G. & Redmer, R. Understanding dense hydrogen at planetary conditions. Nat. Rev. Phys. 2, 562–574 (2020)

    2020

  58. [66]

    Celliers, P. M. et al. Insulator-metal transition in dense fluid deuterium. Science 361, 677–682 (2018)

    2018

  59. [67]

    & Ceriotti, M

    Ben Mahmoud, C., Grasselli, F. & Ceriotti, M. Predicting hot-electron free energies from ground-state data. Phys. Rev. B 106, L121116 (2022)

    2022

  60. [68]

    V ., Dufty, J

    Karasiev, V . V ., Dufty, J. W. & Trickey, S. Nonempirical Semilocal Free-Energy Density Functional for Matter under Extreme Conditions. Phys. Rev. Lett. 120, 076401 (2018)

    2018

  61. [69]

    Bonitz, M. et al. Ab initio simulation of warm dense matter. Phys. Plasmas 27, 042710 (2020)

    2020

  62. [70]

    & Fricke, J

    Gross, J., Reichenauer, G. & Fricke, J. Mechanical properties of SiO2 aerogels. J. Phys. D: Appl. Phys. 21, 1447 (1988). 27

    1988

  63. [71]

    Abdusalamov, R. et al. Modeling and Simulation of the Aggregation and the Structural and Mechanical Properties of Silica Aerogels.J. Phys. Chem. B125, 1944–1950 (2021)

    2021

  64. [72]

    & Park, C

    Liu, J., Buahom, P., Lu, C., Yu, H. & Park, C. B. Microscopic revelation of the solid–gas coupling and Knudsen effect on the thermal conductivity of silica aerogel with inter- connected pores. Sci. Rep. 12, 21034 (2022)

    2022

  65. [73]

    Iswar, S. et al. Dense and strong, but superinsulating silica aerogel. Acta Mater. 213, 116959 (2021)

    2021

  66. [74]

    B., Rosset, L

    Mahmoud, C. B., Rosset, L. A. M., Yates, J. R. & Deringer, V . L. Graph-neural-network predictions of solid-state NMR parameters from spherical tensor decomposition (2024). Preprint at https://arxiv.org/abs/2412.15063

    2024 arXiv

  67. [75]

    C., Rohrer, J., Albe, K

    Erhard, L. C., Rohrer, J., Albe, K. & Deringer, V . L. Modelling atomic and nanoscale structure in the silicon–oxygen system through active machine learning. Nat. Commun. 15, 1927 (2024)

    2024

  68. [76]

    Visualization and analysis of atomistic simulation data with OVITO-the Open Visualization Tool

    Stukowski, A. Visualization and analysis of atomistic simulation data with OVITO-the Open Visualization Tool. Model. Simul. Mater. Sci. Eng.18, 015012 (2010)

    2010

  69. [77]

    M., Teuscher, J., Miyasaka, T., Murakami, T

    Lee, M. M., Teuscher, J., Miyasaka, T., Murakami, T. N. & Snaith, H. J. Efficient Hybrid Solar Cells Based on Meso-Superstructured Organometal Halide Perovskites. Science 338, 643–647 (2012)

    2012

  70. [78]

    Whitfield, P. S. et al. Structures, Phase Transitions and Tricritical Behavior of the Hybrid Perovskite Methyl Ammonium Lead Iodide. Sci. Rep. 6, 35685 (2016)

    2016

  71. [79]

    L ´opez, C. A. et al. Enhanced stability in CH 3NH3PbI3 hybrid perovskite from mechano-chemical synthesis: structural, microstructural and optoelectronic characteri- zation. Sci. Rep. 10, 11228 (2020)

    2020

  72. [80]

    Eames, C. et al. Ionic transport in hybrid lead iodide perovskite solar cells. Nat. Commun. 6, 7497 (2015)

    2015

  73. [81]

    & Nan, C

    Wang, S., Qiao, J., Xu, Y ., Guo, X. & Nan, C. First-principles study on CH 3NH3PbI3 materials for perovskite solar cells. Physica B: Condens. Matter 550, 347–353 (2018)

    2018

  74. [82]

    P., Goh, T

    Ong, K. P., Goh, T. W., Xu, Q. & Huan, A. Structural Evolution in Methylammonium Lead Iodide CH3NH3PbI3. J. Phys. Chem. A 119, 11033–11038 (2015)

    2015

  75. [83]

    Mixed MAPbI 3−xBrx database for force field training

    Bokdam, M. Mixed MAPbI 3−xBrx database for force field training. https://doi. org/10.4121/21878661.v2 (2023). Dataset

    2023 doi

  76. [84]

    Fykouras, K. et al. Disorder to order: how halide mixing in MAPbI 3−xBrx perovskites restricts MA dynamics. J. Mater. Chem. A11, 4587–4597 (2023). 28

    2023

  77. [85]

    & Miyasaka, T

    Nakamura, Y ., Shibayama, N., Hayashida, H., Sugimoto, K. & Miyasaka, T. Thermal Expansion Behavior of Halide Perovskite Single Crystals Across a Broad Temperature Range. Eur. J. Inorg. Chem.28, e202400682 (2025)

    2025

  78. [86]

    & Bokdam, M

    Jinnouchi, R., Lahnsteiner, J., Karsai, F., Kresse, G. & Bokdam, M. Phase Transitions of Hybrid Perovskites Simulated by Machine-Learning Force Fields Trained on the Fly with Bayesian Inference. Phys. Rev. Lett. 122, 225701 (2019)

    2019

  79. [87]

    Mattoni, A., Filippetti, A., Saba, M. I. & Delugas, P. Methylammonium Rotational Dynamics in Lead Halide Perovskite by Classical Molecular Dynamics: The Role of Temperature. J. Phys. Chem. C 119, 17421–17428 (2015)

    2015

  80. [88]

    Liang, X. et al. Structural dynamics descriptors for metal halide perovskites. J. Phys. Chem. C 127, 19141–19151 (2023)

    2023

  81. [89]

    Conant, J. B. & Hussey, R. E. The relation between the structure of organic halides and the speeds of their reaction with inorganic iodides. II. A study of the alkyl chlorides. J. Am. Chem. Soc. 47, 476–488 (1925)

    1925

  82. [90]

    & Wothers, P.Organic Chemistry (Oxford Univer- sity Press, Oxford, 2001)

    Clayden, J., Greeves, N., Warren, S. & Wothers, P.Organic Chemistry (Oxford Univer- sity Press, Oxford, 2001)

    2001

  83. [91]

    & Duarte, F

    Zhang, H., Juraskova, V . & Duarte, F. Modelling chemical processes in explicit solvents with machine learning potentials. Nat. Commun. 15, 6114 (2024)

    2024

  84. [92]

    & Duarte, F

    Vitartas, V ., Zhang, H., Juraskova, V ., Johnston-Wood, T. & Duarte, F. Active learning meets metadynamics: Automated workflow for reactive machine learning potentials (2025) Preprint at https://doi.org/10.26434/chemrxiv-2024-twmlz-v2

    2025 doi

  85. [93]

    & Fung, V

    Kong, L., Shoghi, N., Hu, G., Li, P. & Fung, V . MatterTune: An Integrated, User- Friendly Platform for Fine-Tuning Atomistic Foundation Models to Accelerate Materi- als Simulation and Discovery (2025)

    2025

  86. [94]

    Novelli, P. et al. Fine-tuning Foundation Models for Molecular Dynamics: A Data- Efficient Approach with Random Features. Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) (2024)

    2024

  87. [95]

    G., Allen, C

    Radova, M., Stark, W. G., Allen, C. S., Maurer, R. J. & Bart´ok, A. P. Fine-tuning foun- dation models of materials interatomic potentials with frozen transfer learning (2025). Preprint at https://arxiv.org/abs/2502.15582

    2025 arXiv

  88. [96]

    Mazitov, A. et al. PET-MAD, a universal interatomic potential for advanced materials modeling (2025). Preprint at https://arxiv.org/abs/2503.14118

    2025 arXiv

  89. [97]

    Gardner, J. L. A. graph-pes: train and use graph-based Machine-Learned models of 29 Potential Energy Surfaces. https://github.com/jla-gardner/graph-pes (2024). Version 0.0.10

    2024

  90. [98]

    Hjorth Larsen, A. et al. The atomic simulation environment—a Python library for working with atoms. J. Phys.: Condens. Matter 29, 273002 (2017)

    2017

  91. [99]

    Thompson, A. P. et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Comput. Phys. Commun. 271, 108171 (2022)

    2022

  92. [100]

    Gangan, A. S. et al. Torch-Sim (2025). URL https://github.com/Radical-AI/ torch-sim

    2025

  93. [101]

    & Monro, S

    Robbins, H. & Monro, S. A Stochastic Approximation Method. Ann. Math. Stat. 22, 400–407 (1951)

    1951

  94. [102]

    T., Unke, O

    Sch ¨utt, K. T., Unke, O. T. & Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra (2021). Preprint at https: //arxiv.org/abs/2102.03150

    2021 arXiv

  95. [103]

    & de Fabritiis, G

    Simeon, G. & de Fabritiis, G. TensorNet: Cartesian Tensor Representations for Efficient Learning of Molecular Potentials (2023). Preprint athttps://arxiv.org/abs/2306. 06482

    2023

  96. [104]

    Ziegler, J. F. & Biersack, J. P. The Stopping and Range of Ions in Matter. In Bromley, D. A. (ed.) Treatise on Heavy-Ion Science: Volume 6: Astrophysics, Chemistry, and Condensed Matter, 93–129 (Boston, MA, 1985)

    1985

  97. [105]

    Thomas du Toit, D. F. & Deringer, V . L. Cross-platform hyperparameter optimization for machine learning interatomic potentials. J. Chem. Phys. 159, 024803 (2023)

    2023

  98. [106]

    Lysogorskiy, Y . et al. Performant implementation of the atomic cluster expansion (PACE) and application to copper and silicon. npj Comput. Mater. 7, 1–12 (2021)

    2021

  99. [107]

    Salzbrenner, P. T. et al. Developments and further applications of ephemeral data de- rived potentials. J. Chem. Phys. 159, 144801 (2023)

    2023

  100. [108]

    Litman, Y . et al. i-pi 3.0: A flexible and efficient framework for advanced atomistic simulations. J. Chem. Phys. 161, 062504 (2024)

    2024

  101. [109]

    & Holian, B

    Melchionna, S., Ciccotti, G. & Holian, B. L. Hoover NPT dynamics for systems varying in shape and size. Mol. Phys. 78, 533–544 (1993)

    1993

  102. [110]

    Constrained systems and statistical distribution

    Melchionna, S. Constrained systems and statistical distribution. Phys. Rev. E61, 6165– 6170 (2000). 30

    2000

  103. [111]

    Software update: The ORCA program system-Version 5.0

    Neese, F. Software update: The ORCA program system-Version 5.0. Wiley Interdisci- plinary Reviews: Computational Molecular Science 12, 1–15 (2022)

    2022

  104. [112]

    & Goerigk, L

    Najibi, A. & Goerigk, L. The nonlocal kernel in van der Waals density functionals as an additive correction: An extensive analysis with special emphasis on the B97M-V and ωB97M-V approaches. J. Chem. Theory Comput. 14, 5725–5738 (2018)

    2018

  105. [113]

    A., Johnston-Wood, T., Zhang, H

    Young, T. A., Johnston-Wood, T., Zhang, H. & Duarte, F. Reaction dynamics of Diels–Alder reactions from machine learned potentials. Phys. Chem. Chem. Phys. 24, 20820–20827 (2022)

    2022

  106. [114]

    A., Bonomi, M., Branduardi, D., Camilloni, C

    Tribello, G. A., Bonomi, M., Branduardi, D., Camilloni, C. & Bussi, G. PLUMED 2: New feathers for an old bird. Comput. Phys. Commun. 185, 604–613 (2014)

    2014

  107. [115]

    & Parrinello, M

    Bussi, G., Donadio, D. & Parrinello, M. Canonical sampling through velocity rescaling. J. Chem. Phys. 126, 014101 (2007). 31

    2007

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.