pith. sign in

arxiv: 2604.07979 · v1 · submitted 2026-04-09 · ❄️ cond-mat.mtrl-sci · physics.comp-ph

Differentiable hybrid force fields support scalable autonomous electrolyte discovery

Xintian Wang , Junmin Chen , Zhuoying Zhu , Peichen Zhong This is my paper

Pith reviewed 2026-05-10 17:20 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci physics.comp-ph
keywords hybrid force fieldsdifferentiable molecular dynamicselectrolyte discoveryautonomous discoverymachine learning potentialsenergy decomposition analysisclosed-loop simulationbattery materials
0
0 comments X

The pith

Differentiable hybrid force fields combine physical models with neural corrections to enable fast, accurate, and calibratable simulations for autonomous electrolyte discovery.

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

The paper establishes that differentiable hybrid force fields, which merge physically motivated functional forms with neural-network short-range corrections, resolve the trilemma of speed for high-throughput screening, accuracy for quantitative predictions, and calibratability for refinement. This architecture supports closed-loop autonomous electrolyte discovery by allowing bottom-up ab initio parameterization and top-down experimental fine-tuning. A sympathetic reader would care because it addresses the limitations of classical force fields relying on error cancellation and machine learning potentials being too slow or lacking long-range physics.

Core claim

Differentiable hybrid force fields resolve the trilemma of being fast enough for high-throughput screening, accurate enough for quantitative property prediction, and calibratable enough for online refinement by fusing physically motivated functional forms with neural-network short-range corrections. Grounded in Energy Decomposition Analysis, models such as PhyNEO-Electrolyte and ByteFF-Pol achieve zero-shot generalization to bulk phases with throughputs of tens of ns/day for 10,000-atom systems. Their physical skeletons provide a well-conditioned parameter space for differentiable molecular dynamics, enabling a dual-calibration paradigm that integrates physics-grounded simulation with calibr

What carries the argument

Differentiable hybrid force fields that integrate Energy Decomposition Analysis-grounded physical skeletons with neural short-range corrections to support differentiable molecular dynamics and dual calibration.

If this is right

  • High-throughput screening of electrolytes becomes possible at throughputs of tens of ns/day for 10,000-atom systems.
  • Quantitative property predictions are achieved without heavy reliance on error cancellation.
  • Dual calibration from ab initio data and macroscopic experiments supports online refinement.
  • The architecture enables closed-loop autonomous discovery by integrating simulation with experimental feedback.

Where Pith is reading between the lines

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

  • The hybrid approach could extend to simulation challenges in other materials domains facing similar speed-accuracy-calibration conflicts.
  • Limits of zero-shot generalization might be probed by testing on electrolyte compositions far from the training distribution.
  • Combining these models with robotic experimental loops could shorten discovery cycles beyond the paper's outlined digital-twin concept.

Load-bearing premise

The physical skeletons of these hybrid models provide a well-conditioned parameter space for differentiable molecular dynamics and the models achieve reliable zero-shot generalization to bulk phases without post-hoc adjustments.

What would settle it

A simulation run where gradient-based calibration on the hybrid model produces unstable trajectories or where fine-tuned predictions deviate from measured bulk electrolyte properties such as conductivity would show the central claim does not hold.

Figures

Figures reproduced from arXiv: 2604.07979 by Junmin Chen, Peichen Zhong, Xintian Wang, Zhuoying Zhu.

Figure 1
Figure 1. Figure 1: (a) Representative electrolyte design space: salts, solvents, and additives spanning the combinatorial formu [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Computational-experimental workflow for autonomous electrolyte discovery. The workflow begins with the hybrid [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
read the original abstract

Autonomous electrolyte discovery demands a computational engine that satisfies a critical trilemma: it must be fast enough for high-throughput screening, accurate enough for quantitative property prediction, and calibratable enough for online refinement. Classical empirical force fields (FFs) are fast but rely heavily on error cancellation, while standard machine learning interatomic potentials (MLIPs) are computationally expensive, lack rigorous long-range physics, and resist gradient-based calibration. In this Perspective, we highlight that differentiable hybrid FFs resolve this trilemma by fusing physically motivated functional forms with neural-network short-range corrections. Grounded in Energy Decomposition Analysis (EDA), state-of-the-art models such as PhyNEO-Electrolyte and ByteFF-Pol achieve zero-shot generalization to bulk phases, delivering throughputs on the order of tens of ns/day (up to $\sim$50 ns/day, depending on model complexity) for 10,000-atom systems. Crucially, their physical skeletons provide a well-conditioned parameter space for differentiable molecular dynamics (dMD). This enables a dual-calibration paradigm: bottom-up \textit{ab initio} parameterization combined with top-down fine-tuning from macroscopic experimental observables. We propose that this architecture meets the requirements of a ``ChemRobot-ready'' digital twin by integrating physics-grounded simulation with experimentally calibratable refinement, thereby enabling closed-loop autonomous electrolyte discovery.

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 / 1 minor

Summary. This Perspective argues that differentiable hybrid force fields resolve the trilemma of speed, accuracy, and calibratability for autonomous electrolyte discovery. By fusing EDA-grounded physical functional forms with neural short-range corrections, models such as PhyNEO-Electrolyte and ByteFF-Pol are claimed to deliver zero-shot bulk-phase generalization at throughputs of tens of ns/day (up to ~50 ns/day) on 10k-atom systems while supporting differentiable MD for dual bottom-up/top-down calibration, thereby enabling closed-loop 'ChemRobot-ready' workflows.

Significance. If the performance and conditioning claims hold, the work could meaningfully advance scalable computational electrolyte design by bridging classical FFs and MLIPs. The emphasis on physics-grounded skeletons that remain perturbative and well-conditioned for dMD calibration is a potentially useful framing for the community, though the Perspective introduces no new benchmarks or derivations.

major comments (2)
  1. [Abstract] Abstract: the central trilemma-resolution claim rests on zero-shot bulk-phase accuracy and ~50 ns/day throughput for 10k-atom electrolyte systems, yet the manuscript supplies no new error metrics (e.g., on densities, ionic conductivities, or solvation free energies), error bars, or direct comparisons to baselines. As a Perspective, this leaves the quantitative support entirely dependent on external citations whose applicability to autonomous workflows is not re-examined here.
  2. [Abstract] Abstract: the assertion that 'physical skeletons provide a well-conditioned parameter space for differentiable molecular dynamics' is load-bearing for the proposed dual-calibration paradigm, but the text contains no quantitative evidence such as Hessian eigenvalues, conditioning numbers, or calibration stability examples for the hybrid parameters.
minor comments (1)
  1. The phrase 'ChemRobot-ready digital twin' is introduced without a precise definition or operational criteria that would allow readers to evaluate the claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our Perspective. As this is a forward-looking discussion rather than a research article presenting new data, we synthesize concepts and results from the cited literature. We address each major comment below and indicate where revisions will clarify the manuscript's reliance on external citations.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central trilemma-resolution claim rests on zero-shot bulk-phase accuracy and ~50 ns/day throughput for 10k-atom electrolyte systems, yet the manuscript supplies no new error metrics (e.g., on densities, ionic conductivities, or solvation free energies), error bars, or direct comparisons to baselines. As a Perspective, this leaves the quantitative support entirely dependent on external citations whose applicability to autonomous workflows is not re-examined here.

    Authors: We agree that the Perspective introduces no new benchmarks or error metrics, as its purpose is to highlight the conceptual resolution of the speed-accuracy-calibratability trilemma through differentiable hybrid force fields. The cited throughput (up to ~50 ns/day on 10k-atom systems) and zero-shot bulk-phase generalization claims are drawn directly from the original PhyNEO-Electrolyte and ByteFF-Pol publications. In revision, we will expand the abstract and relevant sections to explicitly reference the specific metrics, error bars, and baseline comparisons reported in those works, while adding a brief discussion of their applicability to closed-loop autonomous workflows. This strengthens the manuscript without requiring new computations outside the Perspective scope. revision: partial

  2. Referee: [Abstract] Abstract: the assertion that 'physical skeletons provide a well-conditioned parameter space for differentiable molecular dynamics' is load-bearing for the proposed dual-calibration paradigm, but the text contains no quantitative evidence such as Hessian eigenvalues, conditioning numbers, or calibration stability examples for the hybrid parameters.

    Authors: The claim follows from the EDA-grounded design, in which physical functional forms capture long-range interactions and neural corrections are restricted to short-range perturbations, preserving parameter conditioning by construction. While the Perspective text does not contain new quantitative diagnostics such as Hessian eigenvalues or conditioning numbers, these properties are demonstrated in the supporting literature for the models discussed. We will revise the manuscript to include targeted citations to the relevant conditioning and stability results from the original papers, along with a concise explanatory clause in the abstract or main text to better support the dual bottom-up/top-down calibration paradigm. revision: partial

Circularity Check

0 steps flagged

No significant circularity; perspective argument is self-contained via external citations

full rationale

The manuscript is a perspective that argues differentiable hybrid FFs (via cited models PhyNEO-Electrolyte and ByteFF-Pol) resolve the speed-accuracy-calibratability trilemma for electrolyte discovery. The provided text contains no equations, no fitted parameters, no predictions derived from inputs, and no self-definitional loops or ansatzes. Claims about zero-shot generalization and well-conditioned dMD parameter spaces are attributed to prior published models rather than derived or renamed within this paper. Per the hard rules, self-citation is normal and does not constitute circularity unless a load-bearing step explicitly reduces to an unverified self-citation by construction; no such reduction is exhibited here. The derivation chain is therefore independent of the present manuscript's inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The perspective relies on domain assumptions about energy decomposition analysis and the suitability of hybrid architectures for differentiable MD, without introducing new free parameters or entities in this document.

axioms (1)
  • domain assumption Energy Decomposition Analysis (EDA) provides a valid grounding for separating physical long-range terms from short-range corrections in hybrid force fields.
    Invoked in the abstract as the basis for the hybrid models.

pith-pipeline@v0.9.0 · 5543 in / 1323 out tokens · 49827 ms · 2026-05-10T17:20:08.438812+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

73 extracted references · 73 canonical work pages · 1 internal anchor

  1. [1]

    Differentiable hybrid force fields support scalable autonomous electrolyte discovery

    and ByteFF-Pol [16], offer a principled resolution: by grounding the PES in EDA-decomposed physical com- ponents [17], one can (i) achieve zero-shot transferable accuracy (i.e., predictability of unseen molecules), (ii) retain the throughput of a semi-analytical model, and (iii) expose a well-conditioned parameter space for both bottom-up and top-down cal...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    density alignment

    or modified Buckingham functions [16], and leaving only a smaller residual for the neural correction. Dimer interactions as transferable training tar- gets.A ChemRobot-ready potential must maintain predictive accuracy across diverse condensed-phase en- vironments. On bulk benchmarks within the training distribution, current MLIPs can match or slightly exc...

    work page

  3. [3]

    Hannah, Y

    D. Hannah, Y. Zhang, X. Li, D. Dong, J. Han, G. Park, H. Gan, B. Liu, K. Liu, Q. Hu,et al., Searching for ideal electrolytes in the molecular universe, The Electrochem- ical Society Interface34, 35 (2025)

    work page 2025

  4. [4]

    S. C. Kim, J. Wang, R. Xu, P. Zhang, Y. Chen, Z. Huang, Y. Yang, Z. Yu, S. T. Oyakhire, W. Zhang, L. C. Green- burg, M. S. Kim, D. T. Boyle, P. Sayavong, Y. Ye, J. Qin, Z. Bao, and Y. Cui, High-entropy electrolytes for practi- cal lithium metal batteries, Nature Energy8, 814 (2023)

    work page 2023

  5. [5]

    N. Yao, X. Chen, Z.-H. Fu, and Q. Zhang, Applying clas- sical, ab initio, and machine-learning molecular dynamics simulations to the liquid electrolyte for rechargeable bat- teries, Chem. Rev.122, 10970 (2022)

    work page 2022

  6. [6]

    Xu, Nonaqueous liquid electrolytes for lithium-based rechargeable batteries, Chem

    K. Xu, Nonaqueous liquid electrolytes for lithium-based rechargeable batteries, Chem. Rev.104, 4303 (2004)

    work page 2004

  7. [7]

    Y. S. Meng, V. Srinivasan, and K. Xu, Designing better electrolytes, Science378, eabq3750 (2022)

    work page 2022

  8. [8]

    Bedrov, J.-P

    D. Bedrov, J.-P. Piquemal, O. Borodin, A. D. MacKerell, B. Roux, and C. Schr¨ oder, Molecular dynamics simula- tions of ionic liquids and electrolytes using polarizable force fields, Chem. Rev.119, 7940 (2019)

    work page 2019

  9. [9]

    W. L. Jorgensen, D. S. Maxwell, and J. Tirado-Rives, Development and testing of the OPLS all-atom force field on conformational energetics and properties of organic liquids, J. Am. Chem. Soc.118, 11225 (1996)

    work page 1996

  10. [10]

    J. Wang, R. M. Wolf, J. W. Caldwell, P. A. Kollman, and D. A. Case, Development and testing of a general amber force field, J. Comput. Chem.25, 1157 (2004)

    work page 2004

  11. [11]

    J. Chen, Q. Gao, M. Huang, and K. Yu, Application of modern artificial intelligence techniques in the devel- opment of organic molecular force fields, Phys. Chem. Chem. Phys.27, 2294 (2025)

    work page 2025

  12. [12]

    D. M. Anstine and O. Isayev, Machine learning inter- atomic potentials and long-range physics, J. Phys. Chem. A127, 2417 (2023)

    work page 2023

  13. [13]

    S. Yue, M. C. Muniz, M. F. Calegari Andrade, L. Zhang, R. Car, and A. Z. Panagiotopoulos, When do short-range atomistic machine-learning models fall short?, J. Chem. Phys.154, 034111 (2021)

    work page 2021

  14. [14]

    D. Kim, X. Wang, S. Vargas, P. Zhong, D. S. King, T. J. Inizan, and B. Cheng, A Universal Augmentation Frame- work for Long-Range Electrostatics in Machine Learning Interatomic Potentials, Journal of Chemical Theory and Computation21, 12709 (2025)

    work page 2025

  15. [15]

    X. Fu, Z. Wu, W. Wang, T. Xie, S. Keten, R. Gomez- Bombarelli, and T. Jaakkola, Forces are not enough: Benchmark and critical evaluation for machine learning force fields with molecular simulations, arXiv preprint (2023), arXiv:2210.07237

    work page arXiv 2023

  16. [16]

    Chen and K

    J. Chen and K. Yu, PhyNEO: A neural-network- enhanced physics-driven force field development work- flow for bulk organic molecule and polymer simulations, J. Chem. Theory Comput.20, 253 (2024)

    work page 2024

  17. [17]

    J. Chen, Q. Gao, Y. Lin, M. Huang, Z. Cheng, W. Feng, J. Huang, B. Wang, and K. Yu, A Hybrid Physics-Driven Neural Network Force Field for Liquid Electrolytes, Jour- nal of Chemical Theory and Computation22, 3011 (2026)

    work page 2026

  18. [18]

    Zheng, X

    T. Zheng, X. Xu, Z. Wang, Z. Yang, Y. Wang, X. Han, Z. Mu, Z. Zhang, S. Liu, S. Gong, K. Yu, and W. Yan, Bridging quantum mechanics to organic liquid proper- ties via a universal force field, arXiv preprint (2025), arXiv:2508.08575

    work page arXiv 2025

  19. [19]

    J. R. Schmidt, K. Yu, and J. G. McDaniel, Transferable next-generation force fields from simple liquids to com- plex materials, Acc. Chem. Res.48, 548 (2015)

    work page 2015

  20. [20]

    X. Wang, Y. Xu, H. Zheng, and K. Yu, A scalable graph neural network method for developing an accurate force field of large flexible organic molecules, J. Phys. Chem. Lett.12, 7982 (2021)

    work page 2021

  21. [21]

    R. Z. Khaliullin, E. A. Cobar, R. C. Lochan, A. T. Bell, and M. Head-Gordon, Unravelling the origin of inter- molecular interactions using absolutely localized molecu- lar orbitals, J. Phys. Chem. A111, 8753 (2007)

    work page 2007

  22. [22]

    Bradbury, R

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. Van- derPlas, S. Wanderman-Milne, and Q. Zhang, JAX: com- posable transformations of Python+NumPy programs (2018)

    work page 2018

  23. [23]

    Han and K

    B. Han and K. Yu, Refining potential energy surface through dynamical properties via differentiable molecu- lar simulation, Nat. Commun.16, 816 (2025)

    work page 2025

  24. [24]

    D. P. Kov´ acs, J. H. Moore, N. J. Browning, I. Bata- tia, J. T. Horton, Y. Pu, V. Kapil, W. C. Witt, I.-B. 8 Magd˘ au, D. J. Cole, and G. Cs´ anyi, MACE-OFF: Short- range transferable machine learning force fields for or- ganic molecules, J. Am. Chem. Soc.147, 17598 (2025)

    work page 2025

  25. [25]

    S. Gong, Y. Zhang, Z. Mu, Z. Pu, H. Wang, X. Han, Z. Yu, M. Chen, T. Zheng, Z. Wang,et al., A predictive machine learning force-field framework for liquid elec- trolyte development, Nat. Mach. Intell.7, 543 (2025)

    work page 2025

  26. [26]

    Essmann, L

    U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee, and L. G. Pedersen, A smooth particle mesh Ewald method, J. Chem. Phys.103, 8577 (1995)

    work page 1995

  27. [27]

    Huang, A

    J. Huang, A. C. Simmonett, F. C. Pickard, A. D. MacK- erell, and B. R. Brooks, Mapping the Drude polarizable force field onto a multipole and induced dipole model, J. Chem. Phys.147, 161702 (2017)

    work page 2017

  28. [28]

    Musaelian, S

    A. Musaelian, S. Batzner, A. Johansson, L. Sun, C. J. Owen, M. Kornbluth, and B. Kozinsky, Learning local equivariant representations for large-scale atomistic dy- namics, Nature Communications14, 579 (2023)

    work page 2023

  29. [29]

    A. S. Goodfellow and B. N. Nguyen, Graph-Based Inter- nal Coordinate Analysis for Transition State Character- ization, Journal of Chemical Theory and Computation 22, 2348 (2026)

    work page 2026

  30. [30]

    Baldwin, Domantas Kuryla, Joseph Hart, Elliott Kasoar, Alin M

    I. Batatia, W. J. Baldwin, D. Kuryla, J. Hart, E. Ka- soar, A. M. Elena, H. Moore, M. J. Gawkowski, B. X. Shi, V. Kapil, P. Kourtis, I.-B. Magd˘ au, and G. Cs´ anyi, MACE-POLAR-1: A polarisable electrostatic foundation model for molecular chemistry, arXiv preprint (2026), arXiv:2602.19411

    work page arXiv 2026

  31. [31]

    X. Zhu, M. Riera, E. F. Bull-Vulpe, and F. Paesani, MB- pol(2023): Sub-chemical accuracy for water simulations from the gas to the liquid phase, J. Chem. Theory Com- put.19, 3551 (2023)

    work page 2023

  32. [32]

    A. K. Picha, M. Wieder, and S. Boresch, Transferable neural network potentials and condensed phase proper- ties, Journal of Chemical Information and Modeling65, 9483 (2025)

    work page 2025

  33. [33]

    S. P. Niblett, P. Kourtis, I.-B. Magd˘ au, C. P. Grey, and G. Cs´ anyi, Transferability of Data Sets between Machine- Learned Interatomic Potential Algorithms, Journal of Chemical Theory and Computation21, 6096 (2025)

    work page 2025

  34. [34]

    M. J. Van Vleet, A. J. Misquitta, and J. R. Schmidt, New angles on standard force fields: Toward a general approach for treating atomic-level anisotropy, J. Chem. Theory Comput.14, 739 (2018)

    work page 2018

  35. [35]

    A. J. Misquitta and A. J. Stone, ISA-Pol: Distributed polarizabilities and dispersion models from a basis-space implementation of the iterated stockholder atoms proce- dure, Theor. Chem. Acc.137, 153 (2018)

    work page 2018

  36. [36]

    A. J. Misquitta, A. J. Stone, and F. Fazeli, Distributed multipoles from a robust basis-space implementation of the iterated stockholder atoms procedure, J. Chem. The- ory Comput.10, 5405 (2014)

    work page 2014

  37. [37]

    J. G. McDaniel, K. Yu, and J. R. Schmidt, Ab initio, physically motivated force fields for CO 2 adsorption in zeolitic imidazolate frameworks, J. Phys. Chem. C116, 1892 (2012)

    work page 2012

  38. [38]

    Zheng, A

    T. Zheng, A. Wang, X. Han, Y. Xia, X. Xu, J. Zhan, Y. Liu, Y. Chen, Z. Wang, X. Wu,et al., Data-driven parametrization of molecular mechanics force fields for expansive chemical space coverage, Chem. Sci.16, 2730 (2025)

    work page 2025

  39. [39]

    Illarionov, S

    A. Illarionov, S. Sakipov, L. Pereyaslavets, I. V. Kurnikov, G. Kamath, O. Butin, E. Voronina, I. Ivah- nenko, I. Leontyev, G. Nawrocki, M. Darkhovskiy, M. Ol- evanov, Y. K. Cherniavskyi, C. Lock, S. Greenslade, S. K. R. S. Sankaranarayanan, M. G. Kurnikova, J. Potoff, R. D. Kornberg, M. Levitt, and B. Fain, Accurate repre- sentation of intermolecular int...

    work page 2023

  40. [40]

    Kamath, A

    G. Kamath, A. Illarionov, S. Sakipov, L. Pereyaslavets, I. V. Kurnikov, O. Butin, E. Voronina, I. Ivahnenko, I. Leontyev, G. Nawrocki, M. Darkhovskiy, M. Olevanov, Y. K. Cherniavskyi, C. Lock, S. Greenslade, Y. Chen, R. D. Kornberg, M. Levitt, and B. Fain, Combining force fields and neural networks for an accurate representa- tion of bonded interactions, ...

    work page 2024

  41. [41]

    Pl´ e, L

    T. Pl´ e, L. Lagard` ere, and J.-P. Piquemal, FeNNol: An efficient and flexible library for building force-field- enhanced neural network potentials, J. Chem. Phys.161, 042502 (2024), arXiv:2301.08734

    work page arXiv 2024

  42. [42]

    L. Yang, J. Li, F. Chen, and K. Yu, A transferrable range- separated force field for water: Combining the power of both physically-motivated models and machine learning techniques, J. Chem. Phys.157, 214108 (2022)

    work page 2022

  43. [43]

    Q. Gao, J. Chen, and K. Yu, Refinement and performance benchmark for range-separated water force field, arXiv preprint (2026), arXiv:2601.18416

    work page arXiv 2026

  44. [44]

    X. Wang, J. Li, L. Yang, F. Chen, Y. Wang, J. Chang, J. Chen, W. Feng, L. Zhang, and K. Yu, DMFF: An open-source automatic differentiable platform for molec- ular force field development and molecular dynamics sim- ulation, J. Chem. Theory Comput.19, 5897 (2023)

    work page 2023

  45. [45]

    S. S. Schoenholz and E. D. Cubuk, JAX-MD: A frame- work for differentiable physics, inAdvances in Neural In- formation Processing Systems, Vol. 33 (2020) pp. 11428– 11441

    work page 2020

  46. [46]

    Doerr, M

    S. Doerr, M. Majewski, A. P´ erez, A. Kr¨ amer, C. Clementi, F. Noe, T. Giorgino, and G. De Fabritiis, TorchMD: A deep learning framework for molecular sim- ulations, J. Chem. Theory Comput.17, 2355 (2021)

    work page 2021

  47. [47]

    Christiansen, T

    H. Christiansen, T. Maruyama, F. Errica, V. Zaverkin, M. Takamoto, and F. Alesiani, Fast, modular, and differentiable framework for machine learning-enhanced molecular simulations, The Journal of Chemical Physics 163(2025)

    work page 2025

  48. [48]

    Fuchs, S

    P. Fuchs, S. Thaler, S. R¨ ocken, and J. Zavadlav, chem- train: Learning deep potential models via automatic differentiation and statistical physics, arXiv preprint (2024), arXiv:2408.15852

    work page arXiv 2024

  49. [49]

    Thaler and J

    S. Thaler and J. Zavadlav, Learning neural network po- tentials from experimental data via differentiable trajec- tory reweighting, Nat. Commun.12, 6884 (2021)

    work page 2021

  50. [50]

    B. Jin, B. Han, W. Feng, K. Yu, and S. Xu, Automatic refinement of force fields based on phase diagrams, arXiv preprint (2025), arXiv:2510.16778

    work page arXiv 2025

  51. [51]

    R¨ ocken, J

    S. R¨ ocken, J. Zavadlav,et al., Refining machine learn- ing potentials through thermodynamic theory of phase transitions, arXiv preprint (2025), arXiv:2512.03974

    work page arXiv 2025

  52. [52]

    L. Metz, C. D. Freeman, S. S. Schoenholz, and T. Kach- man, Gradients are not all you need, arXiv preprint (2021), arXiv:2111.05803

    work page arXiv 2021

  53. [53]

    A. Dave, J. Mitchell, S. Burke, H. Lin, J. Whitacre, and V. Viswanathan, Autonomous optimization of non- aqueous Li-ion battery electrolytes via robotic experi- mentation and machine learning coupling, Nat. Commun. 13, 5454 (2022). 9

    work page 2022

  54. [54]

    Zhuet al., Differentiable modeling and optimization of non-aqueous Li-based battery electrolyte solutions using geometric deep learning, Nat

    S. Zhuet al., Differentiable modeling and optimization of non-aqueous Li-based battery electrolyte solutions using geometric deep learning, Nat. Commun.15, 8649 (2024)

    work page 2024

  55. [55]

    D. A. McQuarrie,Statistical Mechanics(Harper & Row, New York, 1976)

    work page 1976

  56. [56]

    Zhang, S

    Y. Zhang, S. Ye, J. Zhang, C. Hu, J. Jiang, and B. Jiang, Efficient and accurate simulations of vibrational and elec- tronic spectra with symmetry-preserving neural network models for tensorial properties, J. Phys. Chem. B124, 7284 (2020)

    work page 2020

  57. [57]

    Cheng, H

    Z. Cheng, H. Bi, S. Liu, J. Chen, A. J. Misquitta, and K. Yu, Developing a differentiable long-range force field for proteins with E(3) neural network-predicted asymp- totic parameters, J. Chem. Theory Comput.20, 5598 (2024)

    work page 2024

  58. [58]

    Zhong, D

    P. Zhong, D. Kim, D. S. King, and B. Cheng, Ma- chine learning interatomic potential can infer electrical response, npj Computational Materials11, 384 (2025)

    work page 2025

  59. [59]

    Zhang, Z

    B. Zhang, Z. Zhu, H. Li, J. Cao, and J. Jiang, Revolu- tionizing Chemistry and Material Innovation: An Iter- ative Theoretical-Experimental Paradigm Leveraged by Robotic AI Chemists, CCS Chemistry7, 345 (2025)

    work page 2025

  60. [60]

    Y. Shen, L. Wang, Y. Huang, X. Zhang, M. Huang, H. Li, J. He, A. Cai, Y. Wang, P. E. S. Smith, J. Jiang, Z. Zhu, and L. Chen, Unlocking azobenzene isomerization mech- anismsviaan LLM agent-driven workflow integrating simulation, experiment, and machine learning, Chemical Science , 10.1039.D5SC08794E (2026)

    work page 2026

  61. [61]

    Y. Sun, F. Xu, H. Liang, X. Fan, G. Wan, W. Zhong, J. Jiang, X. Li, and L. Chen, MOSES: combining auto- mated ontology construction with a multi-agent system for explainable chemical knowledge reasoning, AI for Sci- ence2, 015001 (2026)

    work page 2026

  62. [62]

    T. Song, M. Luo, X. Zhang, L. Chen, Y. Huang, J. Cao, Q. Zhu, D. Liu, B. Zhang, G. Zou, G. Zhang, F. Zhang, W. Shang, Y. Fu, J. Jiang, and Y. Luo, A Multiagent-Driven Robotic AI Chemist Enabling Au- tonomous Chemical Research On Demand, Journal of the American Chemical Society147, 12534 (2025)

    work page 2025

  63. [63]

    C. Ye, S. Tu, S.-J. Zhang, C. Wang, and S.-Z. Qiao, Har- nessing interfacial solvation structure for next-generation secondary batteries, Nature Energy11, 167 (2026)

    work page 2026

  64. [64]

    Campbell, S

    Q. Campbell, S. Cox, J. Medina, B. Watterson, and A. D. White, MDCrow: automating molecular dynamics work- flows with large language models, Machine Learning: Sci- ence and Technology7, 025037 (2026)

    work page 2026

  65. [65]

    Ding, J.-M

    L. Ding, J.-M. Carrillo, and C. Do, ToPolyAgent: AI agents for coarse-grained bead-spring topological poly- mer simulations, Digital Discovery5, 901 (2026)

    work page 2026

  66. [66]

    Guilbert, C

    S. Guilbert, C. Masschelein, J. Goumaz, B. Naida, and P. Schwaller, DynaMate: An Autonomous Agent for Protein-Ligand Molecular Dynamics Simulations (2025), version Number: 1

    work page 2025

  67. [67]

    Z. Shi, H. A, Y. Shao, D. Huang, H. An, C. Xin, H. Shen, Z. Wang, Y. Na, G. Huang, and X. Jing, MDAgent2: Large Language Model for Code Generation and Knowl- edge Q&A in Molecular Dynamics (2026), version Number: 4

    work page 2026

  68. [68]

    T. Liu, N. Astorga, N. Seedat, and M. van der Schaar, Large Language Models to Enhance Bayesian Optimiza- tion, inThe Twelfth International Conference on Learn- ing Representations(2024)

    work page 2024

  69. [69]

    Kristiadi, F

    A. Kristiadi, F. Strieth-Kalthoff, M. Skreta, P. Poupart, A. Aspuru-Guzik, and G. Pleiss, A Sober Look at LLMs for Material Discovery: Are They Actually Good for Bayesian Optimization Over Molecules?, inProceedings of the 41st International Conference on Machine Learn- ing(2024) pp. 25603–25622

    work page 2024

  70. [70]

    H. Wang, M. Skreta, C. T. Ser, W. Gao, L. Kong, F. Strieth-Kalthoff, C. Duan, Y. Zhuang, Y. Yu, Y. Zhu, et al., Efficient Evolutionary Search Over Chemical Space with Large Language Models, inThe Thirteenth Interna- tional Conference on Learning Representations(2025)

    work page 2025

  71. [71]

    J. Gan, P. Zhong, Y. Du, Y. Zhu, C. Duan, H. Wang, D. Schwalbe-Koda, C. P. Gomes, K. A. Persson, and W. Wang, Matllmsearch: Crystal structure discov- ery with evolution-guided large language models, arXiv preprint arXiv:2502.20933 (2025)

    work page arXiv 2025

  72. [72]

    Y. Du, B. Yu, T. Liu, T. Shen, J. Chen, J. G. Rittig, K. Sun, Y. Zhang, Z. Song, B. Zhou,et al., Accelerat- ing Scientific Discovery with Autonomous Goal-evolving Agents, arXiv preprint arXiv:2512.21782 (2025)

    work page arXiv 2025

  73. [73]

    W. Feng, L. Zhang, Y. Cheng, J. Wu, C. Wei, J. Zhang, and K. Yu, Screening and design of aqueous zinc bat- tery electrolytes based on the multimodal optimization of molecular simulation, The Journal of Physical Chem- istry Letters16, 3326 (2025)

    work page 2025