pith. sign in

arxiv: 2606.03232 · v1 · pith:ADQTS3TSnew · submitted 2026-06-02 · 💻 cs.LG · cs.AI

GFFMERGE: Efficient Merging of Graph Neural Force Fields and Beyond

Parth Verma , Parv P. Singh , Vipul Garg , Ishita Thakre , N. M. Anoop Krishnan , Sayan Ranu This is my paper

Pith reviewed 2026-06-28 11:13 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords graph neural networksmodel mergingneural force fieldsmolecular dynamicsclosed-form solutionembedding alignmentatomistic simulation
0
0 comments X

The pith

GFFMERGE merges separate GNN force-field models through a closed-form solution that approaches the accuracy of joint training.

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

The paper establishes that GNNs for atomic force fields can be merged without full retraining by exploiting the linear structure inside their message-passing layers. This structure turns merging into a convex embedding-alignment task whose solution has an analytical form. Existing vision and language merging techniques collapse on force-field regression, while the new closed-form method recovers most of the accuracy of training on pooled data. The approach yields 5-27 times faster adaptation across molecular, solid-state, and large graph datasets and supplies a strong starting point for any later fine-tuning. A sympathetic reader cares because retraining foundation models for every new chemistry or material is the dominant cost barrier; removing that barrier would let specialized models be combined on demand.

Core claim

By casting model merging as a convex embedding-alignment problem that admits an analytical solution, GFFMERGE recovers performance on MD17, MD22, LiPS20 and large-scale graph tasks that approaches the gold-standard joint-training baseline, while every prior merging technique designed for images or text fails catastrophically on the same force-field regression task.

What carries the argument

The closed-form analytical solution to the convex embedding-alignment problem obtained by treating message-passing layers as linear maps.

If this is right

  • Specialized force-field models can be composed modularly without repeating full training runs.
  • The closed-form merge alone already beats all tested baselines before any fine-tuning step.
  • The same merge supplies a superior initialization that reaches target accuracy with less additional data and fewer epochs.
  • The method extends from force fields to generic GNN tasks via the companion GNNMERGE formulation.

Where Pith is reading between the lines

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

  • If the linear-layer assumption holds for other message-passing architectures, the same closed-form technique could be applied to graph tasks outside atomistic simulation.
  • The speed-up numbers suggest that foundation-model libraries in chemistry could shift from single monolithic checkpoints to libraries of mergeable expert modules.
  • A direct test would be to merge three or more independently trained models and check whether accuracy continues to track joint training.

Load-bearing premise

The message-passing layers of the GNNs must possess a linear structure that permits the merging task to be written as a convex embedding-alignment problem with an analytical solution.

What would settle it

On the MD17 or LiPS20 benchmarks, measure force and energy errors of a GFFMERGE-merged model versus a model trained from scratch on the union of the two datasets; if the merged-model errors remain substantially larger, the central claim does not hold.

Figures

Figures reproduced from arXiv: 2606.03232 by Ishita Thakre, N. M. Anoop Krishnan, Parth Verma, Parv P. Singh, Sayan Ranu, Vipul Garg.

Figure 1
Figure 1. Figure 1: A visual depiction of the alignment objective in GFFMERGE. The orange and purple ellipses represent the regions where the source GNNs Θ1 and Θ2 produce accurate predictions on Datasets 1 and 2 respectively. GFFMERGE aims to learn a merged model ΘM that embeds nodes closer to their original embeddings, and thereby increasing the likelihood that the new embeddings fall within the ellipses. model outputs (Neu… view at source ↗
Figure 2
Figure 2. Figure 2: Illustrates the idea of independent layer-wise merging. generated by Θi at each interaction layer ℓ. Formally, we consider the objective: Select source model z }| { Xn i=1 I(G ∈ Di) GNN layer z}|{ X L ℓ=1 node zX}|{ ∀v∈V [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: MD17 Ablation Results. Top Row: Test MAE vs. Fine-tuning Epochs for the 5-Task mix (Ethanol, Naphthalene, Salicylic Acid, Uracil, Aspirin). Bottom Row: Test MAE vs. Data limit. Trends are consistent with the LiPS results. (a) M3GNet Energy (Convergence) (b) M3GNet Force (Convergence) (c) Orb Force (Convergence) (d) M3GNet Energy (Efficiency) (e) M3GNet Force (Efficiency) (f) Orb Force (Efficiency) [PITH_F… view at source ↗
Figure 5
Figure 5. Figure 5: MD22 Ablation Results. Top Row: Test MAE vs. Fine-tuning Epochs for the supramolecular DHA + Stachyose task. Bottom Row: Test MAE vs. Data limit. (a) M3GNet Energy (Convergence) (b) M3GNet Force (Convergence) (c) Orb Force (Convergence) (d) M3GNet Energy (Efficiency) (e) M3GNet Force (Efficiency) (f) Orb Force (Efficiency) F.2. Ablation Studies on Unfrozen Layers We investigate the sensitivity of the fine-… view at source ↗
Figure 6
Figure 6. Figure 6: Unfreezing-layer ablation on M3GNet. X-axis indicate the number of last blocks unfrozen during fine-tuning and Y-axis indicates the MAE in log scale. For Ethanol+Malonaldehyde+Aspirin, Force MAE (F) is in kcal/mol/A and Energy (E) in kcal/mol. For ˚ Li2S+Li3P, F is in eV/A and E in eV. ˚ [PITH_FULL_IMAGE:figures/full_fig_p025_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Unfreezing-layer ablation on Orb. X-axis indicate the number of last blocks unfrozen during fine-tuning and Y-axis indicates the MAE in log scale. For Ethanol+Malonaldehyde+Aspirin, Force MAE (F) is in kcal/mol/A and Energy (E) in kcal/mol. For Li ˚ 2S+Li3P, F is in eV/A and E in eV. ˚ G. Additional experiments on Generic GNNs To stress-test the methods, we extend the analysis by merging x models correspon… view at source ↗
Figure 8
Figure 8. Figure 8: Variation of average accuracy of merging methods as the number of models varies. Welling, 2017). Since GNNMERGE operates by aligning the embeddings of the new model with those of the base models, it remains agnostic to the base architectures once the intermediate representations are computed. This allows the new GNN to be optimized solely with respect to its own architecture and the target embeddings. • Di… view at source ↗
read the original abstract

Graph Neural Networks (GNNs) have revolutionized Neural Force Fields for atomistic simulations, achieving near-quantum accuracy at reduced cost, yet adapting these models to new chemical systems requires expensive retraining of foundation models. Inspired by model merging in vision and language processing, we introduce GFFMERGE, the first principled framework for closed-form model merging in GNNs. We exploit the linear structure of message-passing layers and formulate merging as a convex embedding-alignment problem with an analytical solution. Through the first systematic benchmarking of model merging for GNNs, we show that existing methods designed for vision and language catastrophically fail on force field regression, while GFFMERGE recovers performance approaching gold standard joint training. Across molecular (MD17, MD22), solid-state (LiPS20), and large-scale graph benchmarks, GFFMERGE and GNNMERGE (its generic GNN counterpart) achieve 5-27$\times$ speedups while enabling modular composition of specialized models. Remarkably, our closed-form solution alone outperforms all baseline methods before fine-tuning and provides superior initialization for faster, data-efficient convergence.

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 / 0 minor

Summary. The manuscript introduces GFFMERGE, a closed-form model merging framework for graph neural force fields that exploits an assumed linear structure in message-passing layers to recast merging as a convex embedding-alignment problem possessing an analytical solution. It reports the first systematic benchmark of merging methods on GNN force-field regression tasks, claiming that vision/language merging baselines fail catastrophically while GFFMERGE (and its generic GNNMERGE variant) recovers performance approaching joint training on MD17, MD22, LiPS20 and large-scale graph benchmarks, with 5-27× speedups and improved fine-tuning initialization.

Significance. If the linearity assumption holds exactly and the performance recovery is reproducible, the result would enable modular composition of specialized force-field models without full retraining, which is practically valuable for atomistic simulation workflows.

major comments (2)
  1. [Abstract] Abstract: the central claim of an analytical solution rests on message-passing layers possessing an exact linear structure that permits a convex embedding-alignment formulation; standard architectures (SchNet, PaiNN) contain nonlinear MLPs, radial basis functions and activations inside the update, yet no derivation, approximation statement or verification that the closed-form solution remains exact is supplied.
  2. [Abstract] Abstract: performance claims (near-joint-training recovery, 5-27× speedups) are stated without error bars, dataset splits, exclusion criteria or statistical tests, rendering the benchmarking results unverifiable and load-bearing for the assertion that GFFMERGE outperforms all baselines.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the theoretical foundations and experimental reporting. We address each major comment below and commit to revisions that strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of an analytical solution rests on message-passing layers possessing an exact linear structure that permits a convex embedding-alignment formulation; standard architectures (SchNet, PaiNN) contain nonlinear MLPs, radial basis functions and activations inside the update, yet no derivation, approximation statement or verification that the closed-form solution remains exact is supplied.

    Authors: The derivation in Section 3 reformulates the message-passing update by isolating the linear embedding transformation while holding nonlinear components (MLPs, radial bases, activations) fixed during the merge step; this yields the convex alignment problem with a closed-form solution. We agree an explicit approximation statement and verification paragraph would improve clarity and will add both in the revised manuscript, including a short empirical check confirming the solution's effectiveness on the evaluated architectures. revision: yes

  2. Referee: [Abstract] Abstract: performance claims (near-joint-training recovery, 5-27× speedups) are stated without error bars, dataset splits, exclusion criteria or statistical tests, rendering the benchmarking results unverifiable and load-bearing for the assertion that GFFMERGE outperforms all baselines.

    Authors: The abstract is a high-level summary; the full experimental protocol (error bars over five random seeds, literature-standard splits for MD17/MD22/LiPS20, outlier exclusion rules, and statistical comparisons) appears in Sections 4–5. We will revise the abstract to reference this statistical robustness and point readers to the detailed reporting. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained from linearity assumption

full rationale

The paper's central derivation exploits an assumed linear structure in message-passing layers to formulate model merging as a convex embedding-alignment problem possessing an analytical solution. This formulation is presented as novel and independent; the closed-form solution is not obtained by fitting parameters to the target regression data or by renaming prior results. No load-bearing self-citations, self-definitional loops, or fitted-input-as-prediction patterns appear in the provided abstract or description. The benchmarking claims are empirical and separate from the derivation step itself. The linearity assumption may be debatable on validity grounds, but that is outside the scope of circularity analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that message-passing layers are linear enough to admit an exact convex embedding-alignment solution; no free parameters or new entities are introduced in the abstract.

axioms (1)
  • domain assumption Message-passing layers in the GNNs exhibit linear structure permitting formulation as a convex embedding-alignment problem with analytical solution.
    Invoked in the abstract as the basis for the closed-form merging procedure.

pith-pipeline@v0.9.1-grok · 5753 in / 1318 out tokens · 33180 ms · 2026-06-28T11:13:59.155488+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

74 extracted references · 14 canonical work pages

  1. [1]

    Proceedings of the web conference 2020 , pages=

    Graphgen: A scalable approach to domain-agnostic labeled graph generation , author=. Proceedings of the web conference 2020 , pages=

    2020

  2. [2]

    IJCAI , year=

    Graphreach: Position-aware graph neural network using reachability estimations , author=. IJCAI , year=

  3. [3]

    Transactions on Machine Learning Research , issn=

    Training Graph Neural Networks Subject to a Tight Lipschitz Constraint , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Neuromlr: Robust & reliable route recommendation on road networks , author=. Advances in Neural Information Processing Systems , volume=

  5. [5]

    Advances in Neural Information Processing Systems , volume=

    Learning articulated rigid body dynamics with lagrangian graph neural network , author=. Advances in Neural Information Processing Systems , volume=

  6. [6]

    International Conference on Machine Learning , pages=

    Stridernet: A graph reinforcement learning approach to optimize atomic structures on rough energy landscapes , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  7. [7]

    International Conference on Machine Learning , pages=

    Grafenne: learning on graphs with heterogeneous and dynamic feature sets , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  8. [8]

    Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Frigate: Frugal spatio-temporal forecasting on road networks , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  9. [9]

    arXiv preprint arXiv:2402.12937 , year=

    Graphgini: Fostering individual and group fairness in graph neural networks , author=. arXiv preprint arXiv:2402.12937 , year=

    arXiv

  10. [10]

    The Eleventh International Conference on Learning Representations , year=

    Enhancing the inductive biases of graph neural ode for modeling physical systems , author=. The Eleventh International Conference on Learning Representations , year=

  11. [11]

    The Twelfth International Conference on Learning Representations , year=

    BroGNet: Momentum-Conserving Graph Neural Stochastic Differential Equation for Learning Brownian Dynamics , author=. The Twelfth International Conference on Learning Representations , year=

  12. [12]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Persona identification in e-commerce with scarce labels and in-context graph learning , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 , pages=

  13. [13]

    ICLR , year=

    Graph attention networks , author=. ICLR , year=

  14. [14]

    Drug discovery today , volume=

    Graph neural networks for automated de novo drug design , author=. Drug discovery today , volume=. 2021 , publisher=

    2021

  15. [15]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  16. [16]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  17. [17]

    M. J. Kearns , title =

  18. [18]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  19. [19]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  20. [20]

    Suppressed for Anonymity , author=

  21. [21]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  22. [22]

    2023 , eprint=

    A Survey on Oversmoothing in Graph Neural Networks , author=. 2023 , eprint=

    2023

  23. [23]

    Konstantin Rusch and Michael Bronstein and Andreea Deac and Marc Lackenby and Siddhartha Mishra and Petar Veli

    Francesco Di Giovanni and T. Konstantin Rusch and Michael Bronstein and Andreea Deac and Marc Lackenby and Siddhartha Mishra and Petar Veli. How does over-squashing affect the power of. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  24. [24]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  25. [25]

    The Twelfth International Conference on Learning Representations , year=

    ZipIt! Merging Models from Different Tasks without Training , author=. The Twelfth International Conference on Learning Representations , year=

  26. [26]

    The Eleventh International Conference on Learning Representations , year=

    Git Re-Basin: Merging Models modulo Permutation Symmetries , author=. The Eleventh International Conference on Learning Representations , year=

  27. [27]

    The Twelfth International Conference on Learning Representations , year=

    AdaMerging: Adaptive Model Merging for Multi-Task Learning , author=. The Twelfth International Conference on Learning Representations , year=

  28. [28]

    CoRR , volume=

    Chenyu Huang and Peng Ye and Tao Chen and Tong He and Xiangyu Yue and Wanli Ouyang , title=. CoRR , volume=. 2024 , cdate=

    2024

  29. [29]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  30. [30]

    Forty-first International Conference on Machine Learning , year=

    Representation Surgery for Multi-Task Model Merging , author=. Forty-first International Conference on Machine Learning , year=

  31. [31]

    The Eleventh International Conference on Learning Representations , year=

    Editing models with task arithmetic , author=. The Eleventh International Conference on Learning Representations , year=

  32. [32]

    2023 , url=

    Prateek Yadav and Derek Tam and Leshem Choshen and Colin Raffel and Mohit Bansal , booktitle=. 2023 , url=

    2023

  33. [33]

    Proceedings of The 33rd International Conference on Machine Learning , pages =

    Revisiting Semi-Supervised Learning with Graph Embeddings , author =. Proceedings of The 33rd International Conference on Machine Learning , pages =. 2016 , editor =

    2016

  34. [34]

    Open Graph Benchmark: Datasets for Machine Learning on Graphs , url =

    Hu, Weihua and Fey, Matthias and Zitnik, Marinka and Dong, Yuxiao and Ren, Hongyu and Liu, Bowen and Catasta, Michele and Leskovec, Jure , booktitle =. Open Graph Benchmark: Datasets for Machine Learning on Graphs , url =

  35. [35]

    and Liu, Juncheng , title =

    Yang, Renchi and Shi, Jieming and Xiao, Xiaokui and Yang, Yin and Bhowmick, Sourav S. and Liu, Juncheng , title =. 2023 , issue_date =. doi:10.1007/s00778-023-00790-4 , journal =

    work page doi:10.1007/s00778-023-00790-4 2023

  36. [36]

    Relational Representation Learning Workshop, NeurIPS 2018 , year=

    Pitfalls of Graph Neural Network Evaluation , author=. Relational Representation Learning Workshop, NeurIPS 2018 , year=

    2018

  37. [37]

    arXiv preprint arXiv:2007.02901 , year=

    Wiki-CS: A Wikipedia-Based Benchmark for Graph Neural Networks , author=. arXiv preprint arXiv:2007.02901 , year=

    arXiv 2007

  38. [38]

    arXiv preprint arXiv:2310.16802 , year=

    From molecules to materials: Pre-training large generalizable models for atomic property prediction , author=. arXiv preprint arXiv:2310.16802 , year=

    arXiv

  39. [39]

    Inductive Representation Learning on Large Graphs , url =

    Hamilton, Will and Ying, Zhitao and Leskovec, Jure , booktitle =. Inductive Representation Learning on Large Graphs , url =

  40. [40]

    International Conference on Learning Representations (ICLR) , year=

    Semi-Supervised Classification with Graph Convolutional Networks , author=. International Conference on Learning Representations (ICLR) , year=

  41. [41]

    International Conference on Learning Representations , year=

    How Powerful are Graph Neural Networks? , author=. International Conference on Learning Representations , year=

  42. [42]

    Graph Attention Networks

    Veli. Graph Attention Networks. International Conference on Learning Representations , year=

  43. [43]

    The Twelfth International Conference on Learning Representations , year=

    Model Merging by Uncertainty-Based Gradient Matching , author=. The Twelfth International Conference on Learning Representations , year=

  44. [44]

    Advances in Neural Information Processing Systems , editor=

    Merging Models with Fisher-Weighted Averaging , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    2022

  45. [45]

    The Eleventh International Conference on Learning Representations , year=

    Dataless Knowledge Fusion by Merging Weights of Language Models , author=. The Eleventh International Conference on Learning Representations , year=

  46. [46]

    and Ying, Rex and Leskovec, Jure , title =

    Hamilton, William L. and Ying, Rex and Leskovec, Jure , title =. Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =. 2017 , isbn =

    2017

  47. [47]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    NodeFormer: A Scalable Graph Structure Learning Transformer for Node Classification , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  48. [48]

    arXiv preprint arXiv:1908.10084 , year=

    Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks , author=. arXiv preprint arXiv:1908.10084 , year=

    Pith/arXiv arXiv 1908

  49. [49]

    International Conference on Learning Representations , year=

    The Role of Permutation Invariance in Linear Mode Connectivity of Neural Networks , author=. International Conference on Learning Representations , year=

  50. [50]

    arXiv preprint arXiv:2103.09430 , year=

    OGB-LSC: A Large-Scale Challenge for Machine Learning on Graphs , author=. arXiv preprint arXiv:2103.09430 , year=

    arXiv

  51. [51]

    2023 , howpublished =

    Hugging Face , title =. 2023 , howpublished =

    2023

  52. [52]

    2023 , isbn =

    Tan, Qiaoyu and Liu, Ninghao and Huang, Xiao and Choi, Soo-Hyun and Li, Li and Chen, Rui and Hu, Xia , title =. 2023 , isbn =. doi:10.1145/3539597.3570404 , booktitle =

    work page doi:10.1145/3539597.3570404 2023

  53. [53]

    International Conference on Learning Representations , year=

    Editing Models with Task Arithmetic , author=. International Conference on Learning Representations , year=

  54. [54]

    Chemical Reviews , volume=

    Machine learning force fields , author=. Chemical Reviews , volume=. 2021 , publisher=

    2021

  55. [55]

    Nature Communications , volume=

    3-dimensional equivariant graph neural networks for interatomic potentials , author=. Nature Communications , volume=. 2022 , publisher=

    2022

  56. [56]

    2024 , eprint=

    Orb: A Fast, Scalable Neural Network Potential , author=. 2024 , eprint=

    2024

  57. [57]

    Nature Computational Science , year=

    Chen, Chi and Ong, Shyue Ping , title=. Nature Computational Science , year=. doi:10.1038/s43588-022-00349-3 , url=

    work page doi:10.1038/s43588-022-00349-3

  58. [58]

    Scalable Parallel Algorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations , volume =. J. Chem. Theory Comput. , author =. 2024 , pages =. doi:10.1021/acs.jctc.4c00190 , number =

    work page doi:10.1021/acs.jctc.4c00190 2024

  59. [59]

    Ilyes Batatia and David Peter Kovacs and Gregor N. C. Simm and Christoph Ortner and Gabor Csanyi , booktitle=. 2022 , url=

    2022

  60. [60]

    and Torrisi, Steven B

    Owen, Cameron J. and Torrisi, Steven B. and Xie, Yu and Batzner, Simon and Bystrom, Kyle and Coulter, Jennifer and Musaelian, Albert and Sun, Lixin and Kozinsky, Boris , title=. npj Computational Materials , year=. doi:10.1038/s41524-024-01264-z , url=

    work page doi:10.1038/s41524-024-01264-z

  61. [61]

    , title=

    Park, Cheol Woo and Kornbluth, Mordechai and Vandermause, Jonathan and Wolverton, Chris and Kozinsky, Boris and Mailoa, Jonathan P. , title=. npj Computational Materials , year=. doi:10.1038/s41524-021-00543-3 , url=

    work page doi:10.1038/s41524-021-00543-3

  62. [62]

    Chemical Reviews , year=

    Zhang, Odin and Lin, Haitao and Zhang, Xujun and Wang, Xiaorui and Wu, Zhenxing and Ye, Qing and Zhao, Weibo and Wang, Jike and Ying, Kejun and Kang, Yu and Hsieh, Chang-Yu and Hou, Tingjun , title=. Chemical Reviews , year=. doi:10.1021/acs.chemrev.5c00461 , url=

    work page doi:10.1021/acs.chemrev.5c00461

  63. [63]

    Riebesell, Janosh and Goodall, Rhys E. A. and Benner, Philipp and Chiang, Yuan and Deng, Bowen and Ceder, Gerbrand and Asta, Mark and Lee, Alpha A. and Jain, Anubhav and Persson, Kristin A. , title=. Nature Machine Intelligence , year=. doi:10.1038/s42256-025-01055-1 , url=

    work page doi:10.1038/s42256-025-01055-1

  64. [64]

    2024 , url=

    Yi-Lun Liao and Brandon Wood and Abhishek Das* and Tess Smidt* , booktitle=. 2024 , url=

    2024

  65. [65]

    & Smidt, T

    Geiger, Mario and Smidt, Tess , title =. 2022 , copyright =. doi:10.48550/ARXIV.2207.09453 , url =

    work page doi:10.48550/arxiv.2207.09453 2022

  66. [66]

    Sauceda and Igor Poltavsky and Kristof T

    Stefan Chmiela and Alexandre Tkatchenko and Huziel E. Sauceda and Igor Poltavsky and Kristof T. Schütt and Klaus-Robert Müller , title =. Science Advances , volume =. 2017 , doi =

    2017

  67. [67]

    Unke and Adil Kabylda and Huziel E

    Stefan Chmiela and Valentin Vassilev-Galindo and Oliver T. Unke and Adil Kabylda and Huziel E. Sauceda and Alexandre Tkatchenko and Klaus-Robert Müller , title =. Science Advances , volume =. 2023 , doi =

    2023

  68. [68]

    and Ranu, Sayan and Krishnan, N

    Bihani, Vaibhav and Mannan, Sajid and Pratiush, Utkarsh and Du, Tao and Chen, Zhimin and Miret, Santiago and Micoulaut, Matthieu and Smedskjaer, Morten M. and Ranu, Sayan and Krishnan, N. M. Anoop. EGraFFBench: evaluation of equivariant graph neural network force fields for atomistic simulations. Digital Discovery. 2024. doi:10.1039/D4DD00027G

    work page doi:10.1039/d4dd00027g 2024

  69. [69]

    Chen, Zhimin and Du, Tao and Krishnan, N. M. Anoop and Yue, Yuanzheng and Smedskjaer, Morten M. , title=. Nature Communications , year=. doi:10.1038/s41467-025-56322-x , url=

    work page doi:10.1038/s41467-025-56322-x

  70. [70]

    The Journal of Chemical Physics , author =

    A foundation model for atomistic materials chemistry , volume =. The Journal of Chemical Physics , author =. 2025 , pages =. doi:10.1063/5.0297006 , abstract =

    work page doi:10.1063/5.0297006 2025

  71. [71]

    Interatomic potentials , isbn =

    Torrens, Iam , year =. Interatomic potentials , isbn =

  72. [72]

    Current Opinion in Solid State and Materials Science , author =

    A practical guide to machine learning interatomic potentials –. Current Opinion in Solid State and Materials Science , author =. 2025 , pages =. doi:10.1016/j.cossms.2025.101214 , language =

    work page doi:10.1016/j.cossms.2025.101214 2025

  73. [73]

    , title =

    Bauchy, M. , title =. The Journal of Chemical Physics , volume =. 2014 , month =. doi:10.1063/1.4886421 , url =

    work page doi:10.1063/1.4886421 2014

  74. [74]

    2025 , eprint=

    Evaluating Universal Machine Learning Force Fields Against Experimental Measurements , author=. 2025 , eprint=

    2025