arxiv: 2605.09495 · v1 · submitted 2026-05-10 · ⚛️ physics.chem-ph · cs.LG· physics.comp-ph

Recognition: 2 theorem links

· Lean Theorem

Enabling Structure-Only Initialization and Out-of-Distribution Generalization in GNN-based Molecular Dynamics Simulators

S. A. Shteingolts , Salman N. Salman , Dan Mendels

Authors on Pith no claims yet

Pith reviewed 2026-05-12 02:57 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cs.LGphysics.comp-ph

keywords graph neural networksmolecular dynamicsout-of-distribution generalizationstructure-only initializationbarostatphysics-informed optimizationelastic networksuniaxial compression

0 comments

The pith

GNN molecular dynamics simulators achieve stable single-structure initialization and out-of-distribution generalization through physics-based optimization and a differentiable barostat.

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

The paper develops two complementary techniques so that graph neural network simulators for molecular dynamics can begin accurate predictions from a lone static configuration instead of a sequence of prior states. An inference-time optimization step enforces physical consistency on each predicted step while a new GNN-based barostat, kept differentiable, tracks evolving system dimensions and pressure. When tested on uniaxial compression of disordered elastic networks that span many geometries and mechanical responses, the pair of additions produces stable long rollouts and correct behavior even in regimes whose dynamics differ markedly from the training set. A reader would care because these changes open the door to using fast, differentiable learned simulators directly in inverse design loops where candidate structures must be evaluated without first generating classical trajectories.

Core claim

By pairing an inference-time physics-based optimization framework that constrains each rollout step to remain physically consistent with a differentiable GNN-based barostat that accurately follows system dimensions and pressure, GNN-based molecular dynamics simulators attain stable structure-only initialization and reliable out-of-distribution generalization, including on elastic networks whose microscopic dynamics are more complex than those encountered during training.

What carries the argument

Inference-time physics-based optimization that enforces physical consistency at every rollout step together with the differentiable GNN-based barostat for tracking dimensions and pressure; these two mechanisms jointly remove the need for temporal context and preserve accuracy outside the training distribution.

If this is right

Simulations can be initialized directly from candidate structures without prior trajectory data, enabling inverse design workflows.
Rollouts remain stable over extended times without divergence or unphysical artifacts.
The model handles unseen geometries, Poisson ratios, and microscopic behaviors in elastic networks.
GNN simulators become practical tools for materials discovery and structural optimization.

Where Pith is reading between the lines

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

The same physics-constraint approach could be ported to other dynamical systems such as fluids or crystals by swapping the relevant physical priors.
Low-overhead optimization might allow GNN simulators to supplement or replace segments of traditional molecular dynamics pipelines in screening tasks.
Coupling the barostat to temperature or other ensemble controls would extend the method to broader thermodynamic conditions.
Full differentiability opens direct gradient-based optimization of structures for target macroscopic properties.

Load-bearing premise

Repeated application of the inference-time physics optimization does not accumulate errors or bias the learned dynamics, and the GNN barostat correctly predicts macroscopic pressure responses even in out-of-distribution regimes.

What would settle it

A long rollout begun from a single structure in an out-of-distribution regime with distinct complex dynamics that produces unphysical states, instability, or macroscopic pressure mismatches would falsify the claim of reliable generalization and stability.

Figures

Figures reproduced from arXiv: 2605.09495 by Dan Mendels, Salman N. Salman, S. A. Shteingolts.

**Figure 1.** Figure 1: FIG. 1. Examples of disordered elastic network (DEN) topologies. (a) - a non-auxetic DEN with high Poisson ratio, (b) - auxetic DEN [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: FIG. 2. Schematic overview of the two proposed approaches for addressing the SOI problem of GNN-based simulators. The first approach (top) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Performance of a position-only GNN simulator after a 10- [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Performance of structure-only initialization methodologies [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6. Performance of bootstrapped GNN simulator (a, b) and simu [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5. Performance of structure-only initialization methodologies [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 7.** Figure 7: FIG. 7. Comparison between dynamically calculated (orange) and [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

**Figure 1.** Figure 1: FIG. 1. (a) - Parity plot showing the performance of padded GNN simulator with (blue) and without (orange) [PITH_FULL_IMAGE:figures/full_fig_p019_1.png] view at source ↗

**Figure 2.** Figure 2: FIG. 2. (a) - Parity plot showing the performance of padded GNN simulator with (blue) and without (orange) [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Comparison between two GNN simulator models trained using a one-step training (OST) strategy [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Comparison between two simulator cascades trained using a one-step training (OST) strategy (blue) [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗

**Figure 7.** Figure 7: FIG. 7. Performance of MD-bootstrapped GNN simulator on stiffness-tuned data with (blue) and without [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗

**Figure 8.** Figure 8: FIG. 8. Performance of simulator cascade on stiffness-tuned data with (blue) and without (orange) ITPO [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗

**Figure 9.** Figure 9: FIG. 9. Parity plot (a) and mean-squared position error ( [PITH_FULL_IMAGE:figures/full_fig_p030_9.png] view at source ↗

**Figure 10.** Figure 10: FIG. 10. Parity plot (a) and mean-squared position error ( [PITH_FULL_IMAGE:figures/full_fig_p031_10.png] view at source ↗

**Figure 11.** Figure 11: FIG. 11. Parity plot (a) and mean-squared position error ( [PITH_FULL_IMAGE:figures/full_fig_p031_11.png] view at source ↗

read the original abstract

Machine learning-based simulators offer the potential to model the dynamics of complex systems more efficiently than classical approaches, while retaining differentiability, a key property for materials design. Graph neural network (GNN)-based simulators have shown strong performance across a range of physical domains, including molecular dynamics. However, their reliance on temporal context for accurate prediction limits their use in inverse design settings, where simulations must be initialized from a single static configuration. Moreover, inverse design requires robust out-of-distribution (OOD) generalization, as candidate structures typically lie outside the training domain. Here, we address both challenges by introducing two complementary strategies that enable stable and accurate structure-only initialization of GNN-based simulations. To directly target OOD generalization, we propose an inference-time physics-based optimization framework that constrains model predictions to remain physically consistent during rollout. In addition, we introduce a differentiable, GNN-based barostat that enables accurate tracking of system dimensions and pressure, critical for capturing macroscopic responses and supporting OOD generalization. We evaluate these approaches in the context of uniaxial compression of disordered elastic networks spanning a broad range of geometries, Poisson ratios, and microscopic behaviors. We find that, together, these methods substantially improve rollout stability and enable reliable OOD generalization, including regimes with distinct, more complex dynamics than those in the training data. These results show that, when properly initialized and constrained, GNN-based simulators can serve as efficient and generalizable tools for materials discovery and structural optimization, advancing their use in materials, molecular, and dynamical system design.

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.

Desk Editor's Note private letter to a colleague

The paper adds inference-time physics optimization for structure-only starts and a GNN barostat to improve stability and OOD generalization in GNN molecular dynamics, but the claims rest on high-level assertions without visible numbers or controls.

read the letter

The main things to know are that the authors target two practical limits in GNN-based MD simulators: the need for temporal history to initialize and weak performance on structures outside the training distribution. They propose an inference-time physics optimization to enforce consistency during rollout from a single static config, plus a differentiable GNN barostat to track pressure and box dimensions more accurately in new regimes. The testbed is uniaxial compression of disordered elastic networks with varied geometries, Poisson ratios, and local behaviors, which lets them check generalization to more complex dynamics than seen in training. This combination is presented as enabling reliable long rollouts for inverse design tasks. The paper does a reasonable job naming the bottlenecks and showing why a pure learned simulator struggles with them. The barostat in particular addresses a macroscopic property that many GNN models ignore, and the idea of adding physics constraints at inference time is a direct way to stabilize predictions without retraining. The evaluation domain is also sensible for materials applications. The softer parts are the missing details on results and implementation. The description gives no quantitative gains, error bars, baseline comparisons, or ablations that isolate the optimizer versus the barostat. Without those, it is hard to judge whether the reported stability comes from the GNN learning better or from the added constraints overriding model errors. The stress-test point about possible error accumulation or bias in OOD rollouts still applies on the available information, since there are no specifics on optimization horizon, loss terms, or how the barostat was checked in the more complex cases. If the full manuscript supplies those controls and shows the components actually complement each other, the contribution strengthens. This is for people already working on ML simulators for molecules or materials who need practical fixes for design workflows. A reader looking for engineering ideas to extend GNN-MD to static-start and OOD settings would pick up usable techniques, though they would still need to implement and test the pieces themselves. It deserves a serious referee because the problems are common and the proposed fixes are concrete enough to evaluate. I would send it to review rather than desk reject, with the expectation that reviewers will request the quantitative validation and implementation details that are not yet visible.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces an inference-time physics-based optimization framework and a differentiable GNN-based barostat to enable structure-only initialization and out-of-distribution generalization in GNN-based molecular dynamics simulators. Evaluated on uniaxial compression of disordered elastic networks spanning a range of geometries, Poisson ratios, and microscopic behaviors, the combined methods are claimed to substantially improve rollout stability and enable reliable OOD generalization, including to regimes with distinctly more complex dynamics than the training data.

Significance. If the central claims hold under detailed scrutiny, this work would be significant for hybrid ML-physics modeling in materials science. It directly tackles two practical barriers to using GNN simulators for inverse design—dependence on temporal context for initialization and poor OOD performance—by adding physics constraints at inference time and a learnable barostat. Successful validation could make such simulators more viable for structural optimization tasks where candidate configurations lie outside the training distribution.

major comments (3)

[Abstract] Abstract: the claim of 'substantially improve rollout stability and enable reliable OOD generalization' is presented without any quantitative metrics, error bars, baseline comparisons, or ablation results. The evaluation is described only at a high level, leaving it unclear whether the reported gains arise from the GNN itself or from the added optimizer overriding deficiencies.
[Methods] Methods (inference-time physics-based optimization): the central claim requires that per-step optimization maintains consistency without accumulating errors or systematically biasing GNN predictions over long rollouts. No formulation details (constraint type, optimization horizon, loss terms) or supporting analysis/ablations are referenced, so it is impossible to assess whether the approach satisfies the skeptic's concern about error accumulation or bias in OOD regimes.
[Methods] Methods (differentiable GNN barostat): the assertion that the barostat 'accurately tracks system dimensions and pressure' and supports OOD generalization to more complex dynamics needs explicit validation (e.g., pressure-response curves or comparisons against reference barostats in OOD test cases). Without this, the macroscopic-response component of the OOD claim remains unverified.

minor comments (1)

[Abstract] Abstract: consider adding one sentence specifying the range of network sizes, Poisson ratios, or microscopic interaction types used in the uniaxial-compression experiments to better convey the breadth of the test regime.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. The comments highlight important areas for clarification and strengthening of the presentation. We have revised the manuscript to address each point, as described below.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of 'substantially improve rollout stability and enable reliable OOD generalization' is presented without any quantitative metrics, error bars, baseline comparisons, or ablation results. The evaluation is described only at a high level, leaving it unclear whether the reported gains arise from the GNN itself or from the added optimizer overriding deficiencies.

Authors: We agree that the abstract would benefit from greater specificity to support the claims. In the revised manuscript we have updated the abstract to include concise quantitative indicators drawn from the results (e.g., stability improvement factors and OOD error reductions relative to baselines) together with explicit references to the relevant figures, tables, and ablations. The ablations demonstrate that the reported gains require both the GNN and the physics-based optimizer; removing the optimizer produces clear degradation in OOD regimes. revision: yes
Referee: [Methods] Methods (inference-time physics-based optimization): the central claim requires that per-step optimization maintains consistency without accumulating errors or systematically biasing GNN predictions over long rollouts. No formulation details (constraint type, optimization horizon, loss terms) or supporting analysis/ablations are referenced, so it is impossible to assess whether the approach satisfies the skeptic's concern about error accumulation or bias in OOD regimes.

Authors: We have expanded the Methods section with a dedicated subsection that fully specifies the inference-time optimization: constraint types (energy and linear-momentum conservation), optimization horizon (single-step per rollout step with optional short-horizon lookahead), and loss terms (GNN prediction loss plus weighted soft-constraint penalties). We have also added analysis and ablations in the Results and Supplementary Information that quantify error accumulation over long rollouts and confirm the absence of systematic bias in OOD test regimes, as measured by conserved quantities and macroscopic observables. revision: yes
Referee: [Methods] Methods (differentiable GNN barostat): the assertion that the barostat 'accurately tracks system dimensions and pressure' and supports OOD generalization to more complex dynamics needs explicit validation (e.g., pressure-response curves or comparisons against reference barostats in OOD test cases). Without this, the macroscopic-response component of the OOD claim remains unverified.

Authors: We accept that explicit validation of the barostat is required. The revised manuscript now includes pressure-response curves and direct comparisons against reference barostats (Berendsen and Parrinello-Rahman) on both in-distribution and OOD test sets. These results confirm accurate tracking of system dimensions and pressure and show that the GNN barostat enables the reported OOD generalization to more complex dynamics without introducing macroscopic artifacts. revision: yes

Circularity Check

0 steps flagged

No significant circularity; methods presented as independent additive techniques

full rationale

The abstract and available description introduce two new strategies—an inference-time physics-based optimization framework and a differentiable GNN-based barostat—as complementary additions to enable structure-only initialization and OOD generalization in GNN simulators. These are framed as novel contributions evaluated empirically on uniaxial compression of disordered elastic networks, with no equations, derivations, or self-referential reductions shown that would equate outputs to fitted inputs or prior self-citations by construction. The central claims rest on the proposed methods' ability to improve stability and generalization rather than any tautological renaming or load-bearing self-citation chain. The derivation chain is therefore self-contained with independent empirical content.

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 physics-based optimization and barostat are described at a conceptual level without implementation details.

pith-pipeline@v0.9.0 · 5593 in / 1117 out tokens · 43862 ms · 2026-05-12T02:57:11.384412+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

inference-time physics-based optimization framework that constrains model predictions to remain physically consistent during rollout... composite loss function consisting of an anchor term L_anchor and a weighted physics term L_physics
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

differentiable, GNN-based barostat... Py = P_K_y + P_V_y ... box acceleration a_t_y

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

299 extracted references · 299 canonical work pages

[1]

Frenkel, Daan and Smit, Berend , doi =

work page
[2]

Andersen, Hans C. , doi =. The Journal of Chemical Physics , month =

work page
[3]

and Zhang, Yuhong and Pastor, Richard W

Feller, Scott E. and Zhang, Yuhong and Pastor, Richard W. and Brooks, Bernard R. , doi =. The Journal of Chemical Physics , month =

work page
[4]

, file =

Tuckerman, Mark E. , file =

work page
[5]

and Tobias, Douglas J

Martyna, Glenn J. and Tobias, Douglas J. and Klein, Michael L. , doi =. The Journal of Chemical Physics , month =

work page
[6]

Advanced Materials , keywords =

Zheng, Xiaoyang and Zhang, Xubo and Chen, Ta‐Te and Watanabe, Ikumu , doi =. Advanced Materials , keywords =

work page
[7]

and Howard, Michael P

Sherman, Zachary M. and Howard, Michael P. and Lindquist, Beth A. and Jadrich, Ryan B. and Truskett, Thomas M. , doi =. The Journal of Chemical Physics , month =. arXiv , arxivId =:2004.00181 , file =

work page arXiv 2004
[8]

Materials , keywords =

Wang, Jia and Wang, Yingxue and Chen, Yanan , doi =. Materials , keywords =

work page
[9]

Manoharan, Naveen Raj and Iqbal, Hassan and Kumar, Krishna , eprint =

work page
[10]

Yang, Kai and Huang, Yuqi and Tao, Junheng and Wang, Wanyu and Wu, Qitian , eprint =

work page
[11]

Nano Letters , keywords =

Yang, Shuo and Yang, Yunhao and Huang, Chen and Bai, Runnan and Cao, Leitao and Ren, Jing and Ling, Shengjie , doi =. Nano Letters , keywords =

work page
[12]

npj Computational Materials , month =

Maurizi, Marco and Gao, Chao and Berto, Filippo , doi =. npj Computational Materials , month =

work page
[13]

and Huang, Guoliang , doi =

Chen, Jiaji and Miao, Xuanbo and Ma, Hongbin and Hopkins, Jonathan B. and Huang, Guoliang , doi =. Materials & Design , keywords =. arXiv , arxivId =:2404.11785 , file =

work page arXiv
[14]

Gowen, Savannah D. , doi =. Soft Matter , number =

work page
[15]

and Feng, Jingchen and MacKintosh, Fred C

Shivers, Jordan L. and Feng, Jingchen and MacKintosh, Fred C. , doi =. Physical Review X , keywords =. arXiv , arxivId =:2407.19563 , file =

work page arXiv
[16]

Allen, Kelsey R. and. Advances in Neural Information Processing Systems , number =

work page
[17]

Hoang, Tai and Trenta, Alessandro and Gravina, Alessio and Freymuth, Niklas and Becker, Philipp and Bacciu, Davide and Neumann, Gerhard , eprint =

work page
[18]

Sharma, Vinay and Fink, Olga , eprint =

work page
[20]

Immanuel, Judah and Mahata, Avik and Maiti, Aniruddha , eprint =

work page
[21]

Nanoscale , number =

Tian, Jie and Tang, Keke and Chen, Xianyan and Wang, Xianqiao , doi =. Nanoscale , number =

work page
[22]

arXiv , arxivId =:2307.08423 , file =

Zhang, Xuan and Wang, Limei and Helwig, Jacob and Luo, Youzhi and Fu, Cong and Xie, YuQing Yaochen and Liu, Meng and Lin, Yuchao and Xu, Zhao and Yan, Keqiang and Adams, Keir and Weiler, Maurice and Li, Xiner and Fu, Tianfan and Wang, Yucheng and Strasser, Alex and Yu, Haiyang and Xie, YuQing Yaochen and Fu, Xiang and Xu, Shenglong and Liu, Yi and Du, Yua...

work page arXiv
[23]

and Stash, Adam I

Shteingolts, Sergey A. and Stash, Adam I. and Tsirelson, Vladimir G. and Fayzullin, Robert R. , doi =. Chemistry - A European Journal , keywords =

work page
[24]

Jarzembska, Katarzyna N. and. Acta Crystallographica Section B: Structural Science, Crystal Engineering and Materials , keywords =. doi:10.1107/S2052520617005534 , file =

work page doi:10.1107/s2052520617005534
[25]

Angewandte Chemie - International Edition , keywords =

Himmel, Daniel and Trapp, Nils and Krossing, Ingo and Altmannshofer, Sandra and Herz, Verena and Eickerling, Georg and Scherer, Wolfgang , doi =. Angewandte Chemie - International Edition , keywords =

work page
[26]

npj Computational Materials , month =

Omee, Sadman Sadeed and Fu, Nihang and Dong, Rongzhi and Hu, Ming and Hu, Jianjun , doi =. npj Computational Materials , month =. arXiv , arxivId =:2401.08032 , file =

work page arXiv
[27]

Journal of Applied Crystallography , month =

Stash, Adam and Tsirelson, Vladimir , doi =. Journal of Applied Crystallography , month =

work page
[28]

and Tournat, Vincent and Bertoldi, Katia , doi =

Bordiga, Giovanni and Medina, Eder and Jafarzadeh, Sina and Boesch, Cyrill and Adams, Ryan P. and Tournat, Vincent and Bertoldi, Katia , doi =. Nature Materials , number =. arXiv , arxivId =:2403.08078 , issn =

work page arXiv
[29]

Acta Crystallographica Section A Foundations of Crystallography , keywords =

Tsirelson, Vladimir and Stash, Adam , doi =. Acta Crystallographica Section A Foundations of Crystallography , keywords =

work page
[30]

Bone, Richard G.A. A. and Bader, Richard F.W. W. , doi =. The Journal of Physical Chemistry , month =

work page
[31]

Bader, Richard F.W. W. , doi =. Journal of Physical Chemistry A , keywords =

work page
[32]

and Bader, Richard F.W

Fradera, Xavier and Austen, Maggie A. and Bader, Richard F.W. W. , doi =. Journal of Physical Chemistry A , month =

work page
[33]

Zeitschrift f

H. Zeitschrift f. doi:10.1515/zkri-2018-2060 , file =

work page doi:10.1515/zkri-2018-2060 2018
[34]

and Zhurova, Elizabeth A

Zhurov, Vladimir V. and Zhurova, Elizabeth A. and Stash, Adam I. and Pinkerton, A. Alan , doi =. Acta Crystallographica Section A: Foundations of Crystallography , keywords =

work page
[35]

Acta Crystallographica Section A Foundations of Crystallography , month =

Restori, Renzo and Schwarzenbach, Dieter , doi =. Acta Crystallographica Section A Foundations of Crystallography , month =

work page
[36]

and Shteingolts, Sergey A

Fayzullin, Robert R. and Shteingolts, Sergey A. and Lodochnikova, Olga A. and Mamedova, Vera L. and Korshin, Dmitry E. and Mamedov, Vakhid A. , doi =. CrystEngComm , number =

work page
[37]

and Sheldrick, George M

Meindl, Kathrin and Henn, Julian and Kocher, Nikolaus and Leusser, Dirk and Zachariasse, Klaas A. and Sheldrick, George M. and Koritsanszky, Tibor and Stalke, Dietmar , doi =. Journal of Physical Chemistry A , number =

work page
[38]

Tsirelson, Vladimir G. , doi =. Acta Crystallographica Section B Structural Science , keywords =

work page
[39]

and Fayzullin, Robert R

Shteingolts, Sergey A. and Fayzullin, Robert R. , doi =. Crystal Growth & Design , month =

work page
[40]

Zhang, Jintu and Bonati, Luigi and Trizio, Enrico and Zhang, Odin and Kang, Yu and Hou, TingJun and Parrinello, Michele , eprint =

work page
[41]

Proceedings of the National Academy of Sciences , keywords =

Yu, Rose and Wang, Rui , doi =. Proceedings of the National Academy of Sciences , keywords =

work page
[42]

and Ceriotti, Michele , doi =

Cheng, Bingqing and Mazzola, Guglielmo and Pickard, Chris J. and Ceriotti, Michele , doi =. Nature , month =. arXiv , arxivId =:1906.03341 , issn =

work page arXiv 1906
[43]

Proceedings of the AAAI Conference on Artificial Intelligence , keywords =

Wang, Haoxiang and Yu, Tao and Yang, Tianwei and Qiao, Hui and Dai, Qionghai , doi =. Proceedings of the AAAI Conference on Artificial Intelligence , keywords =

work page
[44]

Nature Communications , month =

Nikolados, Evangelos-Marios and Wongprommoon, Arin and Aodha, Oisin Mac and Cambray, Guillaume and Oyarz. Nature Communications , month =. doi:10.1038/s41467-022-34902-5 , isbn =

work page doi:10.1038/s41467-022-34902-5
[45]

Computing and Software for Big Science , month =

Brehmer, Johann and Kling, Felix and Espejo, Irina and Cranmer, Kyle , doi =. Computing and Software for Big Science , month =. arXiv , arxivId =:1907.10621 , isbn =

work page arXiv 1907
[46]

The European Physical Journal E , keywords =

Rens, Robbie and Lerner, Edan , doi =. The European Physical Journal E , keywords =. arXiv , arxivId =:1904.07054 , issn =

work page arXiv 1904
[47]

Nature Reviews Physics , month =

Gilpin, William , doi =. Nature Reviews Physics , month =. arXiv , arxivId =:2311.04128 , issn =

work page arXiv
[48]

and Keck, T

Bapst, V. and Keck, T. and Grabska-Barwi. Nature Physics , month =. doi:10.1038/s41567-020-0842-8 , issn =

work page doi:10.1038/s41567-020-0842-8
[49]

and Knudsen, Kenneth and Weber, Hans-Peter Peter and Pattison, Philip and Schwarzenbach, Dieter , doi =

Birkedal, Henrik and Madsen, Dennis and Mathiesen, Ragnvald H. and Knudsen, Kenneth and Weber, Hans-Peter Peter and Pattison, Philip and Schwarzenbach, Dieter , doi =. Acta Crystallographica Section A: Foundations of Crystallography , keywords =

work page
[50]

Nature Communications , month =

Gao, Han and Kaltenbach, Sebastian and Koumoutsakos, Petros , doi =. Nature Communications , month =. arXiv , arxivId =:2402.17157 , isbn =

work page arXiv
[51]

Zu, Mengjie and Desai, Aayush and Goodrich, Carl P. , doi =. Physical Review Letters , keywords =. arXiv , arxivId =:2412.05031 , file =

work page arXiv
[52]

Zu, Mengjie and Goodrich, Carl P. , doi =. Communications Materials , month =. arXiv , arxivId =:2404.15101 , file =

work page arXiv
[53]

Bastek, Jan Hendrik and Kochmann, Dennis M. , doi =. Nature Machine Intelligence , month =. arXiv , arxivId =:2305.19836 , file =

work page arXiv
[54]

Fu, Xiang and Rebello, Nathan J and Olsen, Bradley D , file =

work page
[55]

Proceedings of Machine Learning Research , month =

Shi, Chence and Luo, Shitong and Xu, Minkai and Tang, Jian , eprint =. Proceedings of Machine Learning Research , month =

work page
[56]

and Olsen, Bradley D

Fu, Xiang and Xie, Tian and Rebello, Nathan J. and Olsen, Bradley D. and Jaakkola, Tommi , eprint =

work page
[57]

Advances in Neural Information Processing Systems , month =

Song, Yang and Ermon, Stefano , eprint =. Advances in Neural Information Processing Systems , month =

work page
[58]

Bao, Shouyang and Feng, Jing and Xu, Xiaobin and al - and Fu, Qiang and Wei, Yongchao and Shlomi, Jonathan and Battaglia, Peter and Vlimant, Jean-Roch , doi =. Mach. Learn.: Sci. Technol , keywords =

work page
[59]

Cite this: Mol

Meredig, Bryce and Antono, Erin and Church, Carena and Hutchinson, Maxwell and Ling, Julia and Paradiso, Sean and Blaiszik, Ben and Foster, Ian and Gibbons, Brenna and Hattrick-Simpers, Jason and Mehta, Apurva and Ward, Logan , doi =. Cite this: Mol. Syst. Des. Eng , pages =

work page
[60]

Li, Kangming and Decost, Brian and Choudhary, Kamal and Greenwood, Michael and Hattrick-Simpers, Jason , doi =

work page
[61]

Kauwe, Steven K and Graser, Jake and Murdock, Ryan and Sparks, Taylor D , doi =

work page
[62]

International Journal of Precision Engineering and Manufacturing , keywords =

Song, Jun Ho and Lee, Jae Hoon and Kim, Namjung and Min, Kyoungmin , doi =. International Journal of Precision Engineering and Manufacturing , keywords =

work page
[63]

Biochemical Engineering Journal , pages =

Mowbray, Max and Savage, Thomas and Wu, Chufan and Song, Ziqi and Cho, Anye and Rio-Chanona, Ehecatl A Del and Zhang, Dongda , doi =. Biochemical Engineering Journal , pages =

work page
[64]

Operations Research/ Computer Science Interfaces Series , pages =

Iooss, Bertrand and Lema. Operations Research/ Computer Science Interfaces Series , pages =. doi:10.1007/978-1-4899-7547-8_5/FIGURES/7 , eprint =

work page doi:10.1007/978-1-4899-7547-8_5/figures/7
[65]

Technometrics , number =

Morris, Max D , doi =. Technometrics , number =

work page
[66]

and Neeves, Keith B

Fogelson, Aaron L. and Neeves, Keith B. , doi =. Annual Review of Fluid Mechanics , keywords =

work page
[67]

and Stewart, Garrick C

Mehra, Mandeep R. and Stewart, Garrick C. and Uber, Patricia A. , doi =. The Journal of Heart and Lung Transplantation , keywords =

work page
[68]

and Denney, William S

Chatterjee, Manash S. and Denney, William S. and Jing, Huiyan and Diamond, Scott L. , doi =. PLoS Computational Biology , month =

work page
[69]

Artificial Intelligence Review , keywords =

Zhao, Yingxue and Li, Haoran and Zhou, Haosu and Attar, Hamid Reza and Pfaff, Tobias and Li, Nan , doi =. Artificial Intelligence Review , keywords =. arXiv , arxivId =:2407.11060 , file =

work page arXiv
[70]

Proceedings of Machine Learning Research , pages =

Xu, Minkai and Han, Jiaqi and Lou, Aaron and Kossaifi, Jean and Ramanathan, Arvind and Azizzadenesheli, Kamyar and Leskovec, Jure and Ermon, Stefano and Anandkumar, Anima , eprint =. Proceedings of Machine Learning Research , pages =

work page
[71]

Computational Mechanics , keywords =

Herrmann, Leon and Kollmannsberger, Stefan , doi =. Computational Mechanics , keywords =

work page
[72]

Nature Machine Intelligence , month =

Atz, Kenneth and Grisoni, Francesca and Schneider, Gisbert , doi =. Nature Machine Intelligence , month =. arXiv , arxivId =:2107.12375 , file =

work page arXiv
[73]

Journal of Computational Physics , keywords =

Shu, Dule and Li, Zijie and. Journal of Computational Physics , keywords =. doi:10.1016/j.jcp.2023.111972 , eprint =

work page doi:10.1016/j.jcp.2023.111972 2023
[74]

Luo, Xiao and Luo, Junyu and Jiang, Huiyu and Ju, Wei and Yang, Carl and Zhang, Ming and Sun, Yizhou , file =

work page
[75]

Journal of Medicine (Cincinnati) , number =

To, Trategies and After, Utcomes and Loss, T H E Problem O F Late Allograft , file =. Journal of Medicine (Cincinnati) , number =

work page
[76]

doi:10.1007/s10237-019-01134-4 , file =

Biomechanics and Modeling in Mechanobiology , keywords =. doi:10.1007/s10237-019-01134-4 , file =

work page doi:10.1007/s10237-019-01134-4
[77]

Schneider, Ludwig and de Pablo, Juan J. , doi =. Macromolecules , month =

work page
[78]

Journal of Computational Physics , month =

Barwey, Shivam and Shankar, Varun and Viswanathan, Venkatasubramanian and Maulik, Romit , doi =. Journal of Computational Physics , month =

work page
[79]

and Weare, Jonathan and Dinner, Aaron R

Pengmei, Zihan and Lorpaiboon, Chatipat and Guo, Spencer C. and Weare, Jonathan and Dinner, Aaron R. , eprint =

work page
[80]

Science Advances , month =

Luan, Shengzhi and Chen, Enze and John, Joel and Gaitanaros, Stavros , doi =. Science Advances , month =. arXiv , arxivId =:2304.04809 , file =

work page arXiv
[81]

and Stash, Adam I

Tsirelson, Vladimir G. and Stash, Adam I. and Karasiev, Valentin V. and Liu, Shubin , doi =. Computational and Theoretical Chemistry , keywords =

work page

Showing first 80 references.