pith. the verified trust layer for science. sign in

arxiv: 2509.14205 · v2 · submitted 2025-09-17 · ⚛️ physics.chem-ph

Teachers that teach the irrelevant: Pre-training machine learned interaction potentials with classical force fields for robust molecular dynamics simulations

Eric C.-Y. Yuan , Teresa Head-Gordon This is my paper

Pith reviewed 2026-05-18 15:56 UTC · model grok-4.3

classification ⚛️ physics.chem-ph
keywords machine learned interaction potentialspre-trainingmolecular dynamicsclassical force fieldsab initio fine-tuningmetadynamicsliquid waterhydrogen combustion
0
0 comments X p. Extension

The pith

Pre-training machine learned interaction potentials on classical force field data for single molecules then fine-tuning on limited ab initio labels yields stable molecular dynamics and metadynamics simulations.

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

The paper proposes using inexpensive classical force field data for single molecules to pre-train machine learned interaction potentials before fine-tuning them with a small amount of ab initio data that captures intermolecular forces and reactivity. This two-stage process is intended to address the numerical instabilities that arise in molecular dynamics when models encounter untrained regions of the potential energy surface due to insufficient high-quality data. If the approach works as claimed, it would make high-fidelity simulations more accessible by minimizing the need for vast quantities of computationally demanding quantum mechanical calculations while maintaining or improving accuracy and stability compared to direct training methods. A reader would care because current limitations in data availability often restrict the use of machine learned potentials to well-sampled systems.

Core claim

The authors claim that pre-training on low-quality single-molecule non-reactive force field data followed by data-efficient ab initio fine-tuning allows for stable and accurate molecular dynamics and metadynamics simulations of gas phase molecules, liquid water, and hydrogen combustion reactions, in contrast to models trained from scratch.

What carries the argument

The pre-training learning scheme that uses classical force field data to teach basic intramolecular features before introducing intermolecular and reactive properties in the fine-tuning stage.

If this is right

  • Stable molecular dynamics simulations for gas phase molecules even in new potential energy surface regions.
  • Accurate reproduction of liquid water properties in simulations.
  • Reliable modeling of reactive events in hydrogen combustion.
  • More efficient use of limited ab initio training data for potential learning.
  • Improved stability in metadynamics simulations for free energy calculations.

Where Pith is reading between the lines

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

  • This pre-training strategy might be adapted for studying larger biomolecules by leveraging existing force field libraries.
  • It could facilitate the discovery of new reaction pathways by enabling longer and more stable reactive simulations.
  • The separation of training stages may allow for better understanding of how different physical interactions are encoded in the model.
  • Testing the approach on systems with different types of intermolecular forces could reveal its broader applicability.

Load-bearing premise

That pre-training exclusively on single-molecule non-reactive classical force field data will not introduce biases or instabilities preventing effective learning of intermolecular interactions and reactive properties in the fine-tuning stage.

What would settle it

Observing whether a model pre-trained on force fields and then fine-tuned exhibits fewer numerical instabilities or unphysical trajectories than a from-scratch model when running extended molecular dynamics on a hydrogen combustion reaction.

Figures

Figures reproduced from arXiv: 2509.14205 by Eric C.-Y. Yuan, Teresa Head-Gordon.

Figure 1
Figure 1. Figure 1: Force field strategy of sampling high energy and unphysical data for pre-training a MLIP with subsequent fine-tuning. (a) The general workflow for chemical dataset construction can be divided into sampling and labeling. We use rattling to systematically sample high energy conformations, as well as using physics-based FFs to label the data to ensure data coverage in unphysical regions. (b) Compared to accum… view at source ↗
Figure 2
Figure 2. Figure 2: MD simulation stability improved by FFPT for aspirin. (a,b) MD failures can occur with or without hitting a hole on the PES. (c) FFPT greatly improves the MD stability compared to an MLIP trained from scratch. (d) The stability improvement does not come from the ID accuracy. Even with more training data and lower error, the MD stability does not improve correspondingly. This is a direct result from the wro… view at source ↗
Figure 3
Figure 3. Figure 3: Bulk water simulation stability improved by monomer FF pre-training. (a) The MLIP trained from scratch has holes in the PES unlike the FFPT-FT for the water monomer. (b) In the condensed phase simulation using the MLIP trained from scratch, water molecules can adopt a near-linear conformation which leads to collisions with neighboring waters. (c) By pre-training on a one-body FF and fine-tuning with bulk w… view at source ↗
Figure 4
Figure 4. Figure 4: Hydrogen combustion reactions improved by non-reactive FFPT illustrated using reaction 9 HO2 −−→ H + O2. (a) When pre-trained on non-reactive FFs for reactant and products, the FFPT model can learn an effective interpolation over the course of reaction described by the O1-H3 order parameter (blue). While not quantitatively accurate, it can be accurately fine-tuned using high-quality positive examples from … view at source ↗
read the original abstract

Machine learned interaction potentials (MLIPs) have become a critical component of large-scale, high-quality simulations for a range of chemical and biochemical systems. Yet, despite their in-distribution accuracy, molecular dynamics simulations using MLIPs exhibit numerical instabilities due to underlying data insufficiencies when encountering new regions of the potential energy surface. Here we propose a pre-training learning scheme that uses low-quality, practically free, single-molecule non-reactive force field data while all intermolecular interactions and reactive properties are learned at a fine-tuning stage with a small amount of computationally more expensive labels. We show that the force field pre-training approach followed by data efficient ab initio fine tuning allows for stable and accurate molecular dynamics and metadynamics simulations of gas phase molecules, liquid water, and hydrogen combustion reactions compared to models trained from scratch.

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 / 2 minor

Summary. The manuscript proposes pre-training machine-learned interaction potentials (MLIPs) exclusively on low-cost, single-molecule non-reactive classical force field data, followed by data-efficient fine-tuning on a small set of ab initio labels to capture intermolecular interactions and reactive properties. It claims that this two-stage procedure produces MLIPs that enable stable and accurate molecular dynamics and metadynamics simulations for gas-phase molecules, liquid water, and hydrogen combustion reactions, outperforming models trained from scratch on ab initio data alone.

Significance. If the empirical results hold under scrutiny, the approach offers a practical route to more data-efficient and robust MLIPs by delegating intramolecular non-reactive physics to essentially free classical force fields while reserving expensive ab initio data for the physically critical intermolecular and reactive regimes. This could lower barriers to high-quality simulations of reactive and condensed-phase systems where collecting sufficient ab initio training data remains prohibitive.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (results): the central stability claim for liquid water and H2 combustion is presented without quantitative metrics (e.g., energy/force RMSE, fraction of unstable trajectories, or survival time in metadynamics) or error bars; the comparison to scratch-trained models therefore cannot be evaluated for statistical significance or effect size.
  2. [§2.2] §2.2 (fine-tuning procedure): the manuscript does not report an ablation that isolates whether the classical pre-training priors persist in regions outside the fine-tuning distribution (e.g., bond-dissociation coordinates or high-density liquid configurations). Without such diagnostics, it remains possible that observed stability gains arise from the fine-tuning data distribution rather than from the pre-training strategy itself.
minor comments (2)
  1. [Eq. (3)] Notation for the loss function in Eq. (3) mixes force-field and ab initio labels without an explicit subscript; this should be clarified to avoid reader confusion when comparing pre-training and fine-tuning stages.
  2. [Figure 4] Figure 4 caption should state the number of independent MD runs and the exact criterion used to declare a trajectory 'unstable'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have helped us improve the clarity and rigor of our manuscript. We agree that quantitative metrics and an explicit ablation are valuable additions and have incorporated both in the revised version to better support our claims.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (results): the central stability claim for liquid water and H2 combustion is presented without quantitative metrics (e.g., energy/force RMSE, fraction of unstable trajectories, or survival time in metadynamics) or error bars; the comparison to scratch-trained models therefore cannot be evaluated for statistical significance or effect size.

    Authors: We agree that the presentation would benefit from explicit quantitative metrics to allow direct statistical comparison. In the revised manuscript we have added energy and force RMSE values (with standard deviations over three independent training runs) for both pre-trained and scratch-trained models on held-out test sets for liquid water and the hydrogen combustion system. We also report the fraction of trajectories that remained stable for at least 100 ps (averaged over 20 independent MD runs with error bars) and the mean survival time before instability in metadynamics simulations. These new results are summarized in a table in §3 and referenced in the abstract; they show a statistically significant reduction in instability for the pre-trained models. revision: yes

  2. Referee: [§2.2] §2.2 (fine-tuning procedure): the manuscript does not report an ablation that isolates whether the classical pre-training priors persist in regions outside the fine-tuning distribution (e.g., bond-dissociation coordinates or high-density liquid configurations). Without such diagnostics, it remains possible that observed stability gains arise from the fine-tuning data distribution rather than from the pre-training strategy itself.

    Authors: We thank the referee for highlighting the need to isolate the contribution of pre-training. We have added an ablation study to §2.2 in which we train an otherwise identical model from scratch on the same small ab initio fine-tuning set and compare its performance to the pre-trained-then-fine-tuned model. The new results show that the pre-trained model maintains lower force errors and higher stability when evaluated on out-of-distribution configurations (extended bond lengths up to 3 Å and liquid densities 20 % above the training range), whereas the scratch-trained model exhibits rapid error growth and frequent instabilities in these regimes. This supports that the classical priors persist and contribute to robustness beyond the fine-tuning distribution. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical pre-training plus fine-tuning procedure

full rationale

The paper describes a two-stage training procedure: pre-train MLIPs on single-molecule classical force field data, then fine-tune on limited ab initio labels for intermolecular and reactive properties. The central claim is that this yields more stable MD and metadynamics simulations than scratch training, presented as an empirical outcome. No equations or steps reduce a 'prediction' to a fitted parameter by construction, nor does any load-bearing premise rest on self-citation chains or imported uniqueness theorems. The derivation chain is self-contained as a practical training recipe whose success is measured against external simulation benchmarks rather than internal redefinitions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated assumption that classical force field pre-training transfers usefully to the target regime without detailed justification of transferability.

pith-pipeline@v0.9.0 · 5672 in / 1195 out tokens · 38094 ms · 2026-05-18T15:56:02.974852+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

86 extracted references · 86 canonical work pages · 1 internal anchor

  1. [1]

    Constructing high-dimensional neural net work potentials: A tutorial review

    J¨ org Behler. Constructing high-dimensional neural net work potentials: A tutorial review. International Journal of Quantum Chemistry, 115:1032–1050, 8 2015

    work page 2015

  2. [2]

    Atomic cluster expansion for accurate and t ransferable interatomic potentials

    Ralf Drautz. Atomic cluster expansion for accurate and t ransferable interatomic potentials. Phys. Rev. B, 99:014104, 1 2019

    work page 2019

  3. [3]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Ori ol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. 34th International Conference on Machine Learning, ICML 2017, 3:2053–2070, 4 2017

    work page 2017

  4. [4]

    Smith, and Kipton Barros

    Nicholas Lubbers, Justin S. Smith, and Kipton Barros. Hier archical modeling of molecular energies using a deep neural network. The Journal of Chemical Physics, 148(24):241715, 2018

    work page 2018

  5. [5]

    Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger , Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Bor is Kozinsky. E(3)- equivariant graph neural networks for data-efficient and acc urate interatomic potentials. Nat. Comm. 2022 13:1, 13:1–11, 1 2021. 12

    work page 2022

  6. [6]

    Stein, Farnaz Heidar-Zadeh, Meili Liu, Martin Head-Gordon, L uke Bertels, Hongxia Hao, Itai Leven, and Teresa Head Gordon

    Mojtaba Haghighatlari, Jie Li, Xingyi Guan, Oufan Zhang, Aks haya Das, Christopher J. Stein, Farnaz Heidar-Zadeh, Meili Liu, Martin Head-Gordon, L uke Bertels, Hongxia Hao, Itai Leven, and Teresa Head Gordon. Newtonnet: a newtonian mes sage passing network for deep learning of interatomic potentials and forces. Digital Discovery, 1:333–343, 6 2022

    work page 2022

  7. [7]

    Elena , D´ avid P

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena , D´ avid P. Kov´ acs, Janosh Riebesell, Xavier R. Advincula, Mark Asta, Matthew Avaylon, Will iam J. Baldwin, Fabian Berger, Noam Bernstein, Arghya Bhowmik, Filippo Bigi, S amuel M. Blau, Vlad C˘ arare, Michele Ceriotti, Sanggyu Chong, James P. Darby, Sa ndip De, Flaviano Della Pia, Volker L. Deri...

    work page 2025

  8. [8]

    Yuan, Eric C.-Y

    Eric C.Y. Yuan, Eric C.-Y. Yuan, Yunsheng Liu, Junmin Chen, P eichen Zhong, Sanjeev Raja, Tobias Kreiman, Santiago Vargas, Wenbin Xu, Martin Head -Gordon, Chao Yang, Samuel Blau, Bingqing Cheng, Aditi Krishnapriyan, and Teres a Head-Gordon. Foundation models for atomistic simulation of chemistry and materials . Nature Review Chemistry, 2025

    work page 2025

  9. [9]

    Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, Joh n R

    Brandon M. Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, Joh n R. Kitchin, Daniel S. Levine, Kyle Michel, Anuroop Sriram, Taco Cohen, Abhishek Das , Ammar Rizvi, Sushree Jagriti Sahoo, Zachary W. Ulissi, and C. Lawrence Zitni ck. Uma: A family of universal models for atoms, 2025

    work page 2025

  10. [10]

    A foundation model for atomistic materials chemistry

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M Elena , D´ avid P Kov´ acs, Janosh Riebesell, Xavier R Advincula, Mark Asta, Matthew Avaylon, Willi am J Baldwin, Fabian Berger, Noam Bernstein, Arghya Bhowmik, Samuel M Blau, Vlad C˘ arare, James P Darby, Sandip De, Della Pia, Volker L Deringer, Rokas Elijoˇ sius, Zakariya El-Machachi, Fabio Fal- cioni, ...

    work page 2023

  11. [11]

    Harry Moore, Nicholas J

    D´ avid P´ eter Kov´ acs, J. Harry Moore, Nicholas J. Browning , Ilyes Batatia, Joshua T. Horton, Yixuan Pu, Venkat Kapil, William C. Witt, Ioan-Bogdan Ma gd˘ au, Daniel J. Cole, and G´ abor Cs´ anyi. Mace-off: Short-range transferable machine learning force fields for organic molecules. Journal of the American Chemical Society, 147(21):17598–17611, 2025

    work page 2025

  12. [12]

    Anstine, Roman Zubatyuk, and Olexandr Isayev

    Dylan M. Anstine, Roman Zubatyuk, and Olexandr Isayev. Ai mnet2: a neural network potential to meet your neutral, charged, organic, and eleme ntal-organic needs. Chemical Science, 16(23):10228–10244, 2025

    work page 2025

  13. [13]

    Levine, Muhammed Shuaibi, Evan Walter Clark S potte-Smith, Michael G

    Daniel S. Levine, Muhammed Shuaibi, Evan Walter Clark S potte-Smith, Michael G. Tay- lor, Muhammad R. Hasyim, Kyle Michel, Ilyes Batatia, G´ abor C s´ anyi, Misko Dzamba, Peter Eastman, Nathan C. Frey, Xiang Fu, Vahe Gharakhanyan, Adi ti S. Krishnapriyan, Joshua A. Rackers, Sanjeev Raja, Ammar Rizvi, Andrew S. Rosen, Za chary Ulissi, San- tiago Vargas, C....

    work page 2025

  14. [14]

    Smith, Benjamin T

    Justin S. Smith, Benjamin T. Nebgen, Roman Zubatyuk, Nicho las Lubbers, Christian Dev- ereux, Kipton Barros, Sergei Tretiak, Olexandr Isayev, and Adrian E. Roitberg. Approach- ing coupled cluster accuracy with a general-purpose neural network potential through transfer learning. Nature Communications 2019 10:1, 10:1–8, 7 2019

    work page 2019

  15. [15]

    Using metadynamics to build neural network potentials for reactive events: the case of urea decomposition in water

    Manyi Yang, Luigi Bonati, Daniela Polino, and Michele P arrinello. Using metadynamics to build neural network potentials for reactive events: the case of urea decomposition in water. Catalysis Today, 387:143–149, 2022

    work page 2022

  16. [16]

    Heindel, Taehee Ko, Chao Yang, and Te resa Head-Gordon

    Xingyi Guan, Joseph P. Heindel, Taehee Ko, Chao Yang, and Te resa Head-Gordon. Us- ing machine learning to go beyond potential energy surface b enchmarking for chemical reactivity. Nature Computational Science 2023 3:11, 3:965–974, 11 2023

    work page 2023

  17. [17]

    Forces are not enough: Bench mark and critical evalua- tion for machine learning force fields with molecular simula tions

    Xiang Fu, Zhenghao Wu, Wujie Wang, Tian Xie, Microsoft Res earch, Rafael Gomez- Bombarelli, and Tommi Jaakkola. Forces are not enough: Bench mark and critical evalua- tion for machine learning force fields with molecular simula tions. 10 2022

    work page 2022

  18. [18]

    Sina Stocker, Johannes Gasteiger, Florian Becker, Steph an G¨ unnemann, and Johannes T. Margraf. How robust are modern graph neural network potentia ls in long and hot molecular dynamics simulations? Machine Learning: Science and Technology, 3:045010, 11 2022

    work page 2022

  19. [19]

    Data efficiency and extrapolation trends in neu- ral network interatomic potentials

    Joshua A Vita and Daniel Schwalbe-Koda. Data efficiency and extrapolation trends in neu- ral network interatomic potentials. Machine Learning: Science and Technology, 4:035031, 8 2023. 14

    work page 2023

  20. [20]

    Morrow, John L.A

    Joe D. Morrow, John L.A. Gardner, and Volker L. Deringer. How to validate machine- learned interatomic potentials. Journal of Chemical Physics, 158:121501, 3 2023

    work page 2023

  21. [21]

    Smedskjaer, Sayan Ranu, and N

    Vaibhav Bihani, Sajid Mannan, Utkarsh Pratiush, Tao Du, Zhimin Chen, Santiago Miret, Matthieu Micoulaut, Morten M. Smedskjaer, Sayan Ranu, and N. M.Anoop Krishnan. Egraffbench: evaluation of equivariant graph neural networ k force fields for atomistic sim- ulations. Digital Discovery, 3:759–768, 4 2024

    work page 2024

  22. [22]

    Smith, Ben Nebgen, Nicholas Lubbers, Olexandr Is ayev, and Adrian E

    Justin S. Smith, Ben Nebgen, Nicholas Lubbers, Olexandr Is ayev, and Adrian E. Roitberg. Less is more: Sampling chemical space with active learning. Journal of Chemical Physics, 148:241733, 6 2018

    work page 2018

  23. [23]

    Committee neural network potentials control generalization errors and enable activ e learning

    Christoph Schran, Krystof Brezina, and Ondrej Marsale k. Committee neural network potentials control generalization errors and enable activ e learning. Journal of Chemical Physics, 153:104105, 9 2020

    work page 2020

  24. [24]

    Torrisi, Simon Batzner , Yu Xie, Lixin Sun, Alexie M

    Jonathan Vandermause, Steven B. Torrisi, Simon Batzner , Yu Xie, Lixin Sun, Alexie M. Kolpak, and Boris Kozinsky. On-the-fly active learning of in terpretable bayesian force fields for atomistic rare events. npj Computational Materials 2020 6:1, 6:1–11, 3 2020

    work page 2020

  25. [25]

    Thiemann, Patrick Rowe, Er ich A

    Christoph Schran, Fabian L. Thiemann, Patrick Rowe, Er ich A. M¨ uller, Ondrej Marsalek, and Angelos Michaelides. Machine learning potentials for co mplex aqueous systems made simple. Proceedings of the National Academy of Sciences of the United States of America, 118:e2110077118, 9 2021

    work page 2021

  26. [26]

    Se arching configurations in uncertainty space: Active learning of high-dimensional neu ral network reactive potentials

    Qidong Lin, Liang Zhang, Yaolong Zhang, and Bin Jiang. Se arching configurations in uncertainty space: Active learning of high-dimensional neu ral network reactive potentials. Journal of Chemical Theory and Computation, 17:2691–2701, 5 2021

    work page 2021

  27. [27]

    Smith, and Benjamin Nebgen

    Maksim Kulichenko, Kipton Barros, Nicholas Lubbers, Yin g Wai Li, Richard Messerly, Sergei Tretiak, Justin S. Smith, and Benjamin Nebgen. Uncertai nty-driven dynamics for active learning of interatomic potentials. Nature Computational Science 2023 3:3, 3:230– 239, 3 2023

    work page 2023

  28. [28]

    Sauceda , Igor Poltavsky, Kristof T

    Stefan Chmiela, Alexandre Tkatchenko, Huziel E. Sauceda , Igor Poltavsky, Kristof T. Sch¨ utt, and Klaus Robert M¨ uller. Machine learning of accurate energy-conserving molec- ular force fields. Science Advances, 3, 5 2017

    work page 2017

  29. [29]

    Sauceda, Klaus Robert M¨ uller , and Alexandre Tkatchenko

    Stefan Chmiela, Huziel E. Sauceda, Klaus Robert M¨ uller , and Alexandre Tkatchenko. Towards exact molecular dynamics simulations with machine -learned force fields. Nature Communications 2018 9:1, 9:1–10, 9 2018

    work page 2018

  30. [30]

    Unke, Adil Kabylda, Huziel E

    Stefan Chmiela, Valentin Vassilev-Galindo, Oliver T. Unke, Adil Kabylda, Huziel E. Sauceda, Alexandre Tkatchenko, and Klaus Robert M¨ uller. Accurate global machine learn- ing force fields for molecules with hundreds of atoms. Science Advances, 9, 1 2023

    work page 2023

  31. [31]

    Engel, J¨ org Behler, Christoph D ellago, and Michele Ceriotti

    Bingqing Cheng, Edgar A. Engel, J¨ org Behler, Christoph D ellago, and Michele Ceriotti. Ab initio thermodynamics of liquid and solid water. Proceedings of the National Academy of Sciences of the United States of America, 116:1110–1115, 1 2019. 15

    work page 2019

  32. [32]

    Fine-tuning foundati on models for molecular dynam- ics: A data-efficient approach with random features, 2024

    Pietro Novelli, Luigi Bonati, Pedro J Buigues, Giacomo M eanti, Lorenzo Rosasco, Michele Parrinello, and Massimiliano Pontil. Fine-tuning foundati on models for molecular dynam- ics: A data-efficient approach with random features, 2024

    work page 2024

  33. [33]

    Stability-aware training of machine learning force fields w ith differentiable boltzmann es- timators

    Sanjeev Raja, Ishan Amin, Fabian Pedregosa, Google Deep mind, and Aditi Krishnapriyan. Stability-aware training of machine learning force fields w ith differentiable boltzmann es- timators. 2 2024

    work page 2024

  34. [34]

    Online test-time adaptation for be tter generalization of inter- atomic potentials to out-of-distribution data

    Taoyong Cui, Chenyu Tang, Dongzhan Zhou, Yuqiang Li, Xin gao Gong, Wanli Ouyang, Mao Su, and Shufei Zhang. Online test-time adaptation for be tter generalization of inter- atomic potentials to out-of-distribution data. Nature Communications 2025 16:1, 16:1–11, 2 2025

    work page 2025

  35. [35]

    John L. A. Gardner, Daniel F. Thomas du Toit, Chiheb Ben Mahm oud, Zo´ e Faure Beaulieu, Veronika Juraskova, Laura-Bianca Pa¸ sca, Louise A. M. Rosset , Fernanda Duarte, Fausto Martelli, Chris J. Pickard, and Volker L. Deringer. Distilla tion of atomistic foundation models across architectures and chemical domains, 2025

    work page 2025

  36. [36]

    Peikun Zheng, Roman Zubatyuk, Wei Wu, Olexandr Isayev, and Pavlo O. Dral. Arti- ficial intelligence-enhanced quantum chemical method with broad applicability. Nature Communications 2021 12:1, 12:1–13, 12 2021

    work page 2021

  37. [37]

    Richardson, and Markus Meuwly

    Silvan K¨ aser, Jeremy O. Richardson, and Markus Meuwly. Transfer learning for affordable and high-quality tunneling splittings from instanton calc ulations. Journal of Chemical Theory and Computation, 18:6840–6850, 11 2022

    work page 2022

  38. [38]

    Chen, Joonho Lee, Hong Zhou Ye, Timothy C

    Michael S. Chen, Joonho Lee, Hong Zhou Ye, Timothy C. Berke lbach, David R. Reichman, and Thomas E. Markland. Data-efficient machine learning pote ntials from transfer learning of periodic correlated electronic structure methods: Liqu id water at afqmc, ccsd, and ccsd(t) accuracy. Journal of Chemical Theory and Computation, 19:4510–4519, 7 2023

    work page 2023

  39. [39]

    Transfer learning for chemically accurate interatomic neural netwo rk potentials

    Viktor Zaverkin, David Holzm¨ uller, Luca Bonfirraro, and Johannes K¨ astner. Transfer learning for chemically accurate interatomic neural netwo rk potentials. Physical Chemistry Chemical Physics, 25:5383–5396, 2 2023

    work page 2023

  40. [40]

    Transfer learning for molecular property predic- tions from small datasets

    Thorren Kirschbaum and Annika Bande. Transfer learning for molecular property predic- tions from small datasets. AIP Advances, 14:105119, 10 2024

    work page 2024

  41. [41]

    E. O. Khazieva, N. M. Chtchelkatchev, and R. E. Ryltsev. T ransfer learning for accurate description of atomic transport in al-cu melts. The Journal of chemical physics, 161:174101, 11 2024

    work page 2024

  42. [42]

    Luan, Benjamin T

    Luan G. Luan, Benjamin T. Nebgen, Alice E.A. Allen, Brenden W. Hamilton, Sakib Matin, Justin S. Smith, and Richard A. Messerly. Improving bond disso ciations of reactive machine learning potentials through physics-constrained data aug mentation. Journal of Chemical Information and Modeling, 65, 2 2025

    work page 2025

  43. [43]

    Karls, Mingjian Wen, Ilia A

    Zeren Shui, Daniel S. Karls, Mingjian Wen, Ilia A. Nikifor ov, Ellad B. Tadmor, and George Karypis. Injecting domain knowledge from empirical intera tomic potentials to neural networks for predicting material properties. Advances in Neural Information Processing Systems, 35, 10 2022. 16

    work page 2022

  44. [44]

    A multiple- fidelity method for accurate simulation of mos2 properties u sing jax-reaxff and neural network potentials

    Kehan Wang, Longkun Xu, Wei Shao, Haishun Jin, Qiang Wang, a nd Ming Ma. A multiple- fidelity method for accurate simulation of mos2 properties u sing jax-reaxff and neural network potentials. Journal of Physical Chemistry Letters, 15:371–379, 1 2024

    work page 2024

  45. [45]

    S ynthetic pre-training for neural-network interatomic potentials

    John L A Gardner, Kathryn T Baker, and Volker L Deringer. S ynthetic pre-training for neural-network interatomic potentials. Machine Learning: Science and Technology, 5:015003, 1 2024

    work page 2024

  46. [46]

    Gardner, Zo´ e Faure Beaulieu, and Volker L

    John L.A. Gardner, Zo´ e Faure Beaulieu, and Volker L. Deri nger. Synthetic data enable experiments in atomistic machine learning. Digital Discovery, 2:651–662, 6 2023

    work page 2023

  47. [47]

    Alkhulaifi, F

    A. Alkhulaifi, F. Alsahli, and I. Ahmad. Knowledge distillati on in deep learning and its applications. PeerJ Comput Sci, 7:e474, 2021

    work page 2021

  48. [48]

    Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models

    Luis Barroso-Luque, Muhammed Shuaibi, Xiang Fu, Brando n M. Wood, Misko Dzamba, Meng Gao, Ammar Rizvi, C. Lawrence Zitnick, and Zachary W. Uliss i. Open materials 2024 (omat24) inorganic materials dataset and models. arXiv preprint arXiv:2410.12771, 10 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [49]

    Pre- training via denoising for molecular property prediction

    Sheheryar Zaidi, Michael Schaarschmidt, James Martens , Hyunjik Kim, Yee Whye Teh, Alvaro Sanchez-Gonzalez, Peter Battaglia, Razvan Pascanu, and Jonathan Godwin. Pre- training via denoising for molecular property prediction. 11th International Conference on Learning Representations, ICLR 2023, 5 2022

    work page 2023

  50. [50]

    Krishnapriyan

    Tobias Kreiman and Aditi S. Krishnapriyan. Understandin g and Mitigating Distribution Shifts For Machine Learning Force Fields. arXiv preprint arXiv:2503.08674, March 2025. arXiv:2503.08674 [cs]

    work page arXiv 2025

  51. [51]

    Bull-Vulpe, and Francesc o Paesani

    Xuanyu Zhu, Marc Riera, Ethan F. Bull-Vulpe, and Francesc o Paesani. Mb-pol(2023): Sub-chemical accuracy for water simulations from the gas to the liquid phase. Journal of Chemical Theory and Computation, 19(12):3551–3566, 2023

    work page 2023

  52. [52]

    q-aqua: A many-body ccsd (t) water potential, including fou r-body interactions, demon- strates the quantum nature of water from clusters to the liqu id phase

    Qi Yu, Chen Qu, Paul L Houston, Riccardo Conte, Apurba Nandi , and Joel M Bowman. q-aqua: A many-body ccsd (t) water potential, including fou r-body interactions, demon- strates the quantum nature of water from clusters to the liqu id phase. J. Phys. Chem. Lett., 13(22):5068–5074, 2022

    work page 2022

  53. [53]

    J. P. Heindel, S. Sami, and T. Head-Gordon. Completely mult ipolar model as a general framework for many-body interactions as illustrated for wa ter. J Chem Theory Comput, 20(19):8594–8608, 2024

    work page 2024

  54. [54]

    Stein, Farnaz Heid ar-Zadeh, Luke Bertels, Meili Liu, Mojtaba Haghighatlari, Jie Li, Oufan Zhang, Hongxia Hao, Itai Leven, Martin Head-Gordon, and Teresa Head-Gordon

    Xingyi Guan, Akshaya Das, Christopher J. Stein, Farnaz Heid ar-Zadeh, Luke Bertels, Meili Liu, Mojtaba Haghighatlari, Jie Li, Oufan Zhang, Hongxia Hao, Itai Leven, Martin Head-Gordon, and Teresa Head-Gordon. A benchmark dataset for hydrogen combustion. Scientific Data 2022 9:1, 9:1–7, 5 2022

    work page 2022

  55. [55]

    Menger, Shirin Faraji, Ria Broer, and Remco W.A

    Selim Sami, Maximilian F.S.J. Menger, Shirin Faraji, Ria Broer, and Remco W.A. Havenith. Q-force: Quantum mechanically augmented molecul ar force fields. Journal of Chemical Theory and Computation, 17:4946–4960, 8 2021. 17

    work page 2021

  56. [56]

    Using metadynamics t o explore complex free-energy landscapes

    Giovanni Bussi and Alessandro Laio. Using metadynamics t o explore complex free-energy landscapes. Nature Reviews Physics, 2(44):200–212, Apr 2020

    work page 2020

  57. [57]

    Mitchell Messerly, Sakib Matin, Alice E. A. Allen, Benjami n Nebgen, Kipton Barros, Justin S. Smith, Nicholas Lubbers, and Richard Messerly. Mult i-fidelity learning for inter- atomic potentials: Low-level forces and high-level energi es are all you need, 2025

    work page 2025

  58. [58]

    Noah Hoffmann, Jonathan Schmidt, Silvana Botti, and Miguel A. L. Marques. Trans- fer learning on large datasets for the accurate prediction o f material properties. Digital Discovery, 2(5):1368–1379, 2023

    work page 2023

  59. [59]

    Learner: A transfer learning method for low-rank matrix estimation, 2025

    Sean McGrath, Cenhao Zhu, Ryan O’Dea, Min Guo, and Rui Du an. Learner: A transfer learning method for low-rank matrix estimation, 2025

    work page 2025

  60. [60]

    Dotson, Rai mondas Galvelis, John E

    Peter Eastman, Pavan Kumar Behara, David L. Dotson, Rai mondas Galvelis, John E. Herr, Josh T. Horton, Yuezhi Mao, John D. Chodera, Benjamin P. Pritcha rd, Yuanqing Wang, Gianni De Fabritiis, and Thomas E. Markland. Spice, a datase t of drug-like molecules and peptides for training machine learning potentials. Scientific Data 2022 10:1, 10:1–11, 1 2023

    work page 2022

  61. [61]

    Irwin, Khanh G

    John J. Irwin, Khanh G. Tang, Jennifer Young, Chinzorig Dan darchuluun, Benjamin R. Wong, Munkhzul Khurelbaatar, Yurii S. Moroz, John Mayfield, a nd Roger A. Sayle. Zinc20 - a free ultralarge-scale chemical database for liga nd discovery. Journal of Chemical Information and Modeling, 60:6065–6073, 12 2020

    work page 2020

  62. [62]

    Tingle, Khanh G

    Benjamin I. Tingle, Khanh G. Tang, Mar Castanon, John J. Gu tierrez, Munkhzul Khurel- baatar, Chinzorig Dandarchuluun, Yurii S. Moroz, and John J. I rwin. Zinc-22 – a free multi- billion-scale database of tangible compounds for ligand di scovery. Journal of Chemical Information and Modeling, 63:1166–1176, 2 2023

    work page 2023

  63. [63]

    Uni-mol2: Exploring molecular pretraining model a t scale

    Xiaohong Ji, Zhen Wang, Zhifeng Gao, Hang Zheng, Linfeng Zh ang, Guolin Ke, and Weinan E. Uni-mol2: Exploring molecular pretraining model a t scale. 6 2024

    work page 2024

  64. [64]

    Blum, and Je an Louis Reymond

    Lars Ruddigkeit, Ruud Van Deursen, Lorenz C. Blum, and Je an Louis Reymond. Enu- meration of 166 billion organic small molecules in the chemi cal universe database gdb-17. Journal of Chemical Information and Modeling, 52:2864–2875, 11 2012

    work page 2012

  65. [65]

    Knowledge graph-enhanced molecular co ntrastive learning with functional prompt

    Yin Fang, Qiang Zhang, Ningyu Zhang, Zhuo Chen, Xiang Zhuan g, Xin Shao, Xiaohui Fan, and Huajun Chen. Knowledge graph-enhanced molecular co ntrastive learning with functional prompt. Nature Machine Intelligence, 5(5):542–553, 2023

    work page 2023

  66. [66]

    Automated 3 d pre-training for molecular property prediction

    Xu Wang, Huan Zhao, Weiwei Tu, and Quanming Yao. Automated 3 d pre-training for molecular property prediction. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 23:2419–2430, 6 2023

    work page 2023

  67. [67]

    Ene rgy-motivated equiv- ariant pretraining for 3d molecular graphs

    Rui Jiao, Jiaqi Han, Wenbing Huang, Yu Rong, and Yang Liu. Ene rgy-motivated equiv- ariant pretraining for 3d molecular graphs. Proceedings of the 37th AAAI Conference on Artificial Intelligence, AAAI 2023, 37:8096–8104, 7 2022. 18

    work page 2023

  68. [68]

    Molecular geome try pretraining with se(3)-invariant denoising distance matching

    Shengchao Liu, Hongyu Guo, and Jian Tang. Molecular geome try pretraining with se(3)-invariant denoising distance matching. 11th International Conference on Learning Representations, ICLR 2023, 6 2022

    work page 2023

  69. [69]

    Uni-mol: A universal 3d molecul ar representation learning framework

    Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongt eng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni-mol: A universal 3d molecul ar representation learning framework. 2 2018

    work page 2018

  70. [70]

    Jorgensen, Jayaraman Chandrasekhar, Jeffry D

    William L. Jorgensen, Jayaraman Chandrasekhar, Jeffry D. Ma dura, Roger W. Impey, and Michael L. Klein. Comparison of simple potential functions for simulating liquid water. The Journal of Chemical Physics, 79:926–935, 7 1983

    work page 1983

  71. [71]

    Denoising diffusio n probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusio n probabilistic models. Advances in Neural Information Processing Systems, 2020-December, 6 2020

    work page 2020

  72. [72]

    Kingma and Jimmy Lei Ba

    Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stoc hastic optimization. 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, 12 2014

    work page 2015

  73. [73]

    Castelli, Rune Christensen, Marcin Du/suppress lak, Jesper Friis, Michael N

    Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, I vano E. Castelli, Rune Christensen, Marcin Du/suppress lak, Jesper Friis, Michael N. Groves,Bjørk Hammer, Cory Har- gus, Eric D. Hermes, Paul C. Jennings, Peter Bjerre Jensen, James Kermode, John R. Kitchin, Esben Leonhard Kolsbjerg, Joseph Kubal, Kristen Ka asbjerg, Steen Lysgaard, J´ on Bergma...

    work page 2017

  74. [74]

    Tribello, Massimiliano Bonomi, Davide Brandu ardi, Carlo Camilloni, and Gio- vanni Bussi

    Gareth A. Tribello, Massimiliano Bonomi, Davide Brandu ardi, Carlo Camilloni, and Gio- vanni Bussi. Plumed 2: New feathers for an old bird. Computer Physics Communications, 185(2):604–613, 2014. 19 Teachers that teach the irrelevant: Pre-training machine learned interaction potentials with classical force fields for robust molecular dynamics simulations Er...

    work page 2014

  75. [75]

    H 2O2 − → 2OH 4 12 6 Substitution

    work page

  76. [76]

    H 2O2+H − → H2O+OH 5 15 9 O-transfer

    work page

  77. [77]

    HO 2+H − → 2OH 4 12 6

    work page

  78. [78]

    HO 2+O − → OH+O2 4 12 6 H-transfer

    work page

  79. [79]

    O+H 2 − → OH+H 3 9 3

    work page

  80. [80]

    H 2+OH − → H2O+H 4 12 6

    work page

Showing first 80 references.