pith. sign in

arxiv: 2605.22990 · v1 · pith:D3TI62N2new · submitted 2026-05-21 · ⚛️ physics.chem-ph

Drift-React: One-step Generation of Reaction Pathways via SE(3) Drifting Fields

R\'emi Schlama , Philippe Schwaller This is my paper

Pith reviewed 2026-05-25 05:18 UTC · model grok-4.3

classification ⚛️ physics.chem-ph
keywords reaction pathwaystransition statesgenerative modelsSE(3) equivarianceminimum energy pathwaysone-step generationchemical reactions
0
0 comments X

The pith

Drift-React generates complete reaction pathways in a single forward pass from reactant and product geometries.

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

The paper introduces Drift-React as a way to generate full minimum energy pathways for chemical reactions directly from reactant and product geometries in one step. It does this with an SE(3)-equivariant model that incorporates drifting fields weighted by Sinkhorn during training. This removes the need for repeated force calculations or solving differential equations when making predictions. If it works as described, it would allow much quicker identification of reaction mechanisms and transition states across many reactions.

Core claim

Drift-React is an SE(3)-equivariant generative framework that predicts complete reaction pathways in a single forward pass from only reactant and product geometries. By shifting distribution evolution to training via a Sinkhorn-weighted drifting field, Drift-React eliminates both the iterative force evaluations of NEB-style methods and the sequential ODE/SDE integration of diffusion and flow matching models. Evaluated on the Transition1x and Halo8 datasets, the model generates physically consistent MEPs that capture energetic bottlenecks and support arbitrary-resolution sampling.

What carries the argument

SE(3)-equivariant generative model with Sinkhorn-weighted drifting fields that shifts distribution evolution into training for one-step pathway output.

Load-bearing premise

The training procedure with Sinkhorn-weighted drifting fields produces pathways that remain physically consistent and match electronic-structure results without requiring post-hoc force evaluations or iterative refinement on unseen reactions.

What would settle it

A direct comparison on held-out reactions where the model's generated pathways show large mismatches in energy barrier heights or geometries compared to nudged elastic band reference calculations.

Figures

Figures reproduced from arXiv: 2605.22990 by Philippe Schwaller, R\'emi Schlama.

Figure 1
Figure 1. Figure 1: Three paradigms for reaction pathway prediction. Top: Point-wise generators (e.g., diffusion) predict isolated TS geometries at fast inference but ignore the surrounding PES valley, often producing structures with steric clashes. Middle: Iterative chain-of-states methods (e.g., NEB) recover the full MEP but require thousands of force evaluations per reaction. Bottom: Drift-React generates the complete reac… view at source ↗
Figure 2
Figure 2. Figure 2: Training of an SE(3)-equivariant drifting model. The generator fθ transforms a linear interpolation prior pprior into a physically consistent reaction pathway, with reactant and product geometries pinned exactly via a sinusoidal envelope on the displacement output. During training, the empirical drifting field V (vectors) iteratively transports the generated pathways toward the ground-truth minimum energy … view at source ↗
Figure 3
Figure 3. Figure 3: Drift-React produces accurate, clash-free reaction pathways at orders-of-magnitude lower inference cost than baseline methods. Cumulative distributions over test reactions on Halo8 (solid lines) and Transition1x (dashed lines): (a) TS-RMSD. (b) Absolute activation barrier error |∆E ‡ TS|. (c) Mean inference time per reaction (log scale), error bars indicate one standard deviation. (d, e) Representative tes… view at source ↗
Figure 4
Figure 4. Figure 4: Drift-React advances the Pareto front of geometric accuracy and inference cost for point-wise TS prediction. (a) Mean TS-RMSD versus inference time on the Transition1x reactant– TS–product split. Drift-React (blue) and ReactOT (pink) jointly define the Pareto front: methods in the shaded region are dominated. Drift-React is approximately 10× faster than ReactOT while remaining within the geometric range of… view at source ↗
read the original abstract

Mapping reaction pathways and transition states (TS) is fundamental to chemistry but computationally expensive at scale. The minimum energy pathway (MEP) dictates reaction rates and mechanisms, yet recovering it via electronic-structure methods requires thousands of costly force evaluations. Recent generative models accelerate TS identification but rely on iterative inference and only predict isolated saddle-point snapshots, missing the continuous reaction trajectory. We introduce Drift-React, an $\mathrm{SE}(3)$-equivariant generative framework that predicts complete reaction pathways in a single forward pass from only reactant and product geometries. By shifting distribution evolution to training via a Sinkhorn-weighted drifting field, Drift-React eliminates both the iterative force evaluations of NEB-style methods and the sequential ODE/SDE integration of diffusion and flow matching models. Evaluated on the Transition1x and Halo8 datasets, our one-step model generates physically consistent MEPs that accurately capture energetic bottlenecks and enable arbitrary-resolution sampling along the reaction coordinate. For isolated TS prediction, Drift-React matches the sub-{\AA}ngstr\"om accuracy of state-of-the-art iterative models while delivering orders-of-magnitude acceleration, clearing a major computational bottleneck for large-scale reaction network exploration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Drift-React, an SE(3)-equivariant generative framework that predicts complete minimum-energy pathways (MEPs) for chemical reactions in a single forward pass from reactant and product geometries alone. By training with Sinkhorn-weighted drifting fields, the model eliminates iterative force evaluations (as in NEB) and sequential ODE/SDE integration (as in diffusion/flow models). On Transition1x and Halo8, it claims to produce physically consistent MEPs that capture energetic bottlenecks, support arbitrary-resolution sampling, and match sub-Ångström TS accuracy of state-of-the-art iterative models while providing orders-of-magnitude acceleration.

Significance. If the central claims hold, the work would substantially accelerate large-scale reaction-network exploration by removing the dominant computational bottlenecks of pathway generation. The one-step, non-iterative nature and ability to output full continuous trajectories (rather than isolated TS snapshots) would be particularly enabling for high-throughput mechanism studies.

major comments (2)
  1. [Abstract, §3] Abstract and §3 (training objective): the claim that Sinkhorn-weighted drifting fields alone enforce that generated trajectories lie on true MEPs (matching electronic-structure energies and geometries on held-out reactions, without post-hoc force evaluations or refinement) is load-bearing for the sub-Å accuracy and 'physically consistent' assertions, yet no explicit energy-regularization, force-matching, or MEP-projection term is described that would guarantee this property rather than statistical resemblance to the training distribution.
  2. [§4] §4 (evaluation on Transition1x/Halo8): the reported sub-Ångström TS accuracy and capture of energetic bottlenecks must be accompanied by direct comparisons of full MEP energies and forces against electronic-structure reference calculations on the same held-out reactions; without these, the 'no post-hoc force evaluations' claim cannot be verified.
minor comments (2)
  1. [§3] Notation: the precise definition of the drifting field and the Sinkhorn weighting should be stated explicitly with equations before the high-level description of the training procedure.
  2. [§3] The manuscript should clarify whether the model was trained with any auxiliary loss on energies or forces, or whether physical consistency emerges solely from the drifting-field objective.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We provide point-by-point responses below and indicate where revisions will be made.

read point-by-point responses

  1. Referee: [Abstract, §3] Abstract and §3 (training objective): the claim that Sinkhorn-weighted drifting fields alone enforce that generated trajectories lie on true MEPs (matching electronic-structure energies and geometries on held-out reactions, without post-hoc force evaluations or refinement) is load-bearing for the sub-Å accuracy and 'physically consistent' assertions, yet no explicit energy-regularization, force-matching, or MEP-projection term is described that would guarantee this property rather than statistical resemblance to the training distribution.

    Authors: The training data consists exclusively of MEPs obtained from electronic-structure calculations on Transition1x and Halo8. The Sinkhorn-weighted drifting field objective is formulated to learn the transport map that reproduces this distribution in a single forward pass. While no additional explicit force-matching or energy-regularization term appears in the loss (beyond the SE(3)-equivariant architecture and the weighted optimal-transport objective), the learned field produces trajectories whose geometries and implied energetics align with the reference MEPs, as quantified by the reported TS accuracies. We do not assert a strict mathematical guarantee of exact MEP membership for arbitrary unseen reactions, only statistical consistency with the training distribution. We will revise the abstract and §3 to state this distinction more precisely and avoid any implication of an enforcement mechanism beyond learned distribution matching. revision: partial

  2. Referee: [§4] §4 (evaluation on Transition1x/Halo8): the reported sub-Ångström TS accuracy and capture of energetic bottlenecks must be accompanied by direct comparisons of full MEP energies and forces against electronic-structure reference calculations on the same held-out reactions; without these, the 'no post-hoc force evaluations' claim cannot be verified.

    Authors: Section 4 reports sub-Ångström TS geometry accuracy and shows that generated paths capture energetic bottlenecks via energy profiles sampled along the reaction coordinate. The 'no post-hoc force evaluations' statement refers to inference: the model produces the full continuous MEP without requiring NEB-style iterations or additional force calls at generation time. We acknowledge that direct, quantitative force and energy comparisons on the complete generated MEPs versus fresh electronic-structure references for all held-out reactions are not presented. We will add these comparisons in the revised manuscript to strengthen verification of physical consistency. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; derivation is self-contained.

full rationale

The paper presents an SE(3)-equivariant generative model trained via Sinkhorn-weighted drifting fields on external datasets (Transition1x, Halo8) to produce reaction pathways from reactant/product inputs. Claims of one-step MEP generation, sub-Å accuracy on held-out reactions, and physical consistency are positioned as empirical outcomes of training and evaluation, not reductions of outputs to inputs by definition or via self-citation chains. No equations or steps are shown that rename fitted quantities as predictions or import uniqueness from author priors; the framework is independent of the target benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no details on free parameters, axioms, or invented entities are available. Full manuscript required to populate ledger.

pith-pipeline@v0.9.0 · 5740 in / 1103 out tokens · 19913 ms · 2026-05-25T05:18:18.220943+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

54 extracted references · 54 canonical work pages · 4 internal anchors

  1. [1]

    The Activated Complex in Chemical Reactions.The Journal of Chemical Physics, 3(2):107–115, February 1935

    Henry Eyring. The Activated Complex in Chemical Reactions.The Journal of Chemical Physics, 3(2):107–115, February 1935. ISSN 0021-9606. doi: 10.1063/1.1749604. URL https://doi.org/10.1063/1.1749604

    work page doi:10.1063/1.1749604 1935

  2. [2]

    Variational Transition State Theory.Ann

    D G Truhlar and B C Garrett. Variational Transition State Theory.Ann. Rev. Phys. Chem.,

    work page

  3. [3]

    URL https://www.annualreviews

    doi: 10.1146/annurev.pc.35.100184.001111. URL https://www.annualreviews. org/content/journals/10.1146/annurev.pc.35.100184.001111

    work page doi:10.1146/annurev.pc.35.100184.001111

  4. [4]

    Variational principles in a chemical reaction.International Journal of Quantum Chemistry, 20(S15):633–642, 1981

    Kenichi Fukui. Variational principles in a chemical reaction.International Journal of Quantum Chemistry, 20(S15):633–642, 1981. ISSN 1097-461X. doi: 10.1002/qua.560200866. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/qua.560200866

    work page doi:10.1002/qua.560200866 1981

  5. [5]

    How to Conceptualize Catalytic Cycles? The Energetic Span Model.Accounts of Chemical Research, 44(2):101–110, February 2011

    Sebastian Kozuch and Sason Shaik. How to Conceptualize Catalytic Cycles? The Energetic Span Model.Accounts of Chemical Research, 44(2):101–110, February 2011. ISSN 0001-4842. doi: 10.1021/ar1000956. URLhttps://doi.org/10.1021/ar1000956

    work page doi:10.1021/ar1000956 2011

  6. [6]

    Gert Kiss, Nihan Çelebi-Ölçüm, Rocco Moretti, David Baker, and K. N. Houk. Computational Enzyme Design.Angewandte Chemie International Edition, 52(22):5700–5725, 2013. ISSN 1521-3773. doi: 10.1002/anie.201204077. URL https://onlinelibrary.wiley.com/ doi/abs/10.1002/anie.201204077

    work page doi:10.1002/anie.201204077 2013

  7. [7]

    Vern L. Schramm. Enzymatic Transition States and Drug Design.Chemical Reviews, 118 (22):11194–11258, November 2018. ISSN 0009-2665, 1520-6890. doi: 10.1021/acs.chemrev. 8b00369. URLhttps://pubs.acs.org/doi/10.1021/acs.chemrev.8b00369

    work page doi:10.1021/acs.chemrev 2018

  8. [8]

    Uberuaga, and Hannes Jónsson

    Graeme Henkelman, Blas P. Uberuaga, and Hannes Jónsson. A climbing image nudged elastic band method for finding saddle points and minimum energy paths.The Journal of Chemical Physics, 113(22):9901–9904, December 2000. ISSN 0021-9606. doi: 10.1063/1.1329672. URL https://doi.org/10.1063/1.1329672

    work page doi:10.1063/1.1329672 2000

  9. [9]

    Peterson

    Per Lindgren, Georg Kastlunger, and Andrew A. Peterson. Scaled and Dynamic Optimizations of Nudged Elastic Bands.Journal of Chemical Theory and Computation, 15(11):5787–5793, November 2019. ISSN 1549-9618, 1549-9626. doi: 10.1021/acs.jctc.9b00633. URL http: //arxiv.org/abs/1906.10257

    work page doi:10.1021/acs.jctc.9b00633 2019

  10. [10]

    String Method for the Study of Rare Events

    Weinan E, Weiqing Ren, and Eric Vanden-Eijnden. String Method for the Study of Rare Events.Physical Review B, 66(5):052301, August 2002. ISSN 0163-1829, 1095-3795. doi: 10.1103/PhysRevB.66.052301. URLhttp://arxiv.org/abs/cond-mat/0205527

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1103/physrevb.66.052301 2002

  11. [11]

    Linear scaling electronic structure methods.Reviews of Modern Physics, 71(4):1085–1123, July 1999

    Stefan Goedecker. Linear scaling electronic structure methods.Reviews of Modern Physics, 71(4):1085–1123, July 1999. doi: 10.1103/RevModPhys.71.1085. URL https://link.aps. org/doi/10.1103/RevModPhys.71.1085

    work page doi:10.1103/revmodphys.71.1085 1999

  12. [12]

    Hirzel, David Duvenaud, Dougal Maclaurin, Martin A

    Rafael Gómez-Bombarelli, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, David Duvenaud, Dougal Maclaurin, Martin A. Blood-Forsythe, Hyun Sik Chae, Markus Einzinger, Dong-Gwang Ha, Tony Wu, Georgios Markopoulos, Soonok Jeon, Hosuk Kang, Hiroshi Miyazaki, Masaki Numata, Sunghan Kim, Wenliang Huang, Seong Ik Hong, Marc Baldo, Ryan P. Adams, and Alán Aspuru-...

    work page doi:10.1038/nmat4717 2016

  13. [14]

    Beaglehole, Miles J

    Isaac W. Beaglehole, Miles J. Pemberton, Elliot H. E. Farrar, and Matthew N. Grayson. Machine learning transition state geometries and applications in reaction property predic- tion, March 2025. URL https://chemrxiv.org/engage/chemrxiv/article-details/ 67cec10c6dde43c908420904

    work page 2025

  14. [15]

    ECTS: An ultra-fast diffusion model for exploring chemical reactions with equivariant consistency, March

    Mingyuan Xu, Bowen Li, Zhaojia Dong, Pavlo Dral, Tong Zhu, and Hongming Chen. ECTS: An ultra-fast diffusion model for exploring chemical reactions with equivariant consistency, March

    work page

  15. [16]

    URLhttps://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-f9vdp

    work page doi:10.26434/chemrxiv-2025-f9vdp 2025

  16. [17]

    Toney, Weiliang Luo, Johannes Kästner, Mathias Niepert, and Heather J

    Samir Darouich, Jacob W. Toney, Weiliang Luo, Johannes Kästner, Mathias Niepert, and Heather J. Kulik. Beyond the Training Domain: Robust Generative Transition State Models for Unseen Chemistry, January 2026. URLhttp://arxiv.org/abs/2601.16469

    work page arXiv 2026

  17. [18]

    Blau, and Teresa Head-Gordon

    Guanchen Wu, Chung-Yueh Yuan, Kareem Hegazy, Samuel M. Blau, and Teresa Head-Gordon. Machine-Learned Leftmost Hessian Eigenvectors for Robust Transition State Finding, March

    work page

  18. [19]

    URLhttp://arxiv.org/abs/2603.21323

    work page arXiv

  19. [20]

    Anstine, and Olexandr Isayev

    Filipp Nikitin, Dylan M. Anstine, and Olexandr Isayev. Right into the Saddle: Stereochemistry- Aware Generation of Molecular Transition States, April 2026. URLhttps://chemrxiv.org/ doi/full/10.26434/chemrxiv.15001681/v1

    work page doi:10.26434/chemrxiv.15001681/v1 2026

  20. [21]

    Chenru Duan, Yuanqi Du, Haojun Jia, and Heather J. Kulik. Accurate transi- tion state generation with an object-aware equivariant elementary reaction diffusion model.Nature Computational Science, 3(12):1045–1055, December 2023. ISSN 2662-

    work page 2023

  21. [22]

    URL https://www.nature.com/articles/ s43588-023-00563-7

    doi: 10.1038/s43588-023-00563-7. URL https://www.nature.com/articles/ s43588-023-00563-7

    work page doi:10.1038/s43588-023-00563-7

  22. [23]

    Diffusion-based generative AI for exploring transition states from 2D molecular graphs.Nature Communications, 15(1):341, January 2024

    Seonghwan Kim, Jeheon Woo, and Woo Youn Kim. Diffusion-based generative AI for exploring transition states from 2D molecular graphs.Nature Communications, 15(1):341, January 2024. ISSN 2041-1723. doi: 10.1038/s41467-023-44629-6. URL https://www.nature.com/ articles/s41467-023-44629-6

    work page doi:10.1038/s41467-023-44629-6 2024

  23. [24]

    FragmentFlow: Scalable Transition State Generation for Large Molecules, February 2026

    Ron Shprints, Peter Holderrieth, Juno Nam, Rafael Gómez-Bombarelli, and Tommi Jaakkola. FragmentFlow: Scalable Transition State Generation for Large Molecules, February 2026. URL http://arxiv.org/abs/2602.02310

    work page arXiv 2026

  24. [25]

    Generating transition states of chemical reactions via distance-geometry-based flow matching, November 2025

    Yufei Luo, Xiang Gu, and Jian Sun. Generating transition states of chemical reactions via distance-geometry-based flow matching, November 2025. URL http://arxiv.org/abs/ 2511.17229

    work page arXiv 2025

  25. [26]

    Kovar, and Esther Heid

    Leonard Galustian, Konstantin Mark, Johannes Karwounopoulos, Maximilian P.-P. Kovar, and Esther Heid. GoFlow: Efficient transition state geometry prediction with flow matching and E(3)-equivariant neural networks.Digital Discovery, 4(12):3492–3501, December 2025. ISSN 2635-098X. doi: 10.1039/D5DD00283D. URL https://pubs.rsc.org/en/content/ articlelanding/...

    work page doi:10.1039/d5dd00283d 2025

  26. [27]

    Adaptive Transition State Refinement with Learned Equilibrium Flows, July 2025

    Samir Darouich, Vinh Tong, Tanja Bien, Johannes Kästner, and Mathias Niepert. Adaptive Transition State Refinement with Learned Equilibrium Flows, July 2025. URL http://arxiv. org/abs/2507.16521

    work page arXiv 2025

  27. [28]

    Driving Reaction Trajectories via Latent Flow Matching, February 2026

    Yili Shen and Xiangliang Zhang. Driving Reaction Trajectories via Latent Flow Matching, February 2026. URLhttp://arxiv.org/abs/2602.10476. 11

    work page arXiv 2026

  28. [29]

    Gomes, Evangelos A

    Chenru Duan, Guan-Horng Liu, Yuanqi Du, Tianrong Chen, Qiyuan Zhao, Haojun Jia, Carla P. Gomes, Evangelos A. Theodorou, and Heather J. Kulik. Optimal transport for generating transition states in chemical reactions.Nature Machine Intelligence, 7(4):615–626, April 2025. ISSN 2522-5839. doi: 10.1038/s42256-025-01010-0. URL https://www.nature.com/ articles/s...

    work page doi:10.1038/s42256-025-01010-0 2025

  29. [30]

    Flow matching for reaction pathway generation, November

    Ping Tuo, Jiale Chen, and Ju Li. Flow matching for reaction pathway generation, November

    work page

  30. [31]

    URLhttp://arxiv.org/abs/2507.10530

    work page arXiv

  31. [32]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models, October 2022. URLhttp://arxiv.org/abs/2010.02502

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  32. [33]

    Krishnapriyan

    Sanjeev Raja, Martin Šípka, Michael Psenka, Tobias Kreiman, Michal Pavelka, and Aditi S. Krishnapriyan. Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional, April 2025. URL http://arxiv.org/abs/ 2504.18506

    work page arXiv 2025

  33. [34]

    Hare and Dean J

    Stephanie R. Hare and Dean J. Tantillo. Post-transition state bifurcations gain momentum – current state of the field.Pure and Applied Chemistry, 89(6):679–698, June 2017. ISSN 1365-

    work page 2017

  34. [35]

    URL https://www.degruyterbrill.com/document/ doi/10.1515/pac-2017-0104/html?lang=en

    doi: 10.1515/pac-2017-0104. URL https://www.degruyterbrill.com/document/ doi/10.1515/pac-2017-0104/html?lang=en

    work page doi:10.1515/pac-2017-0104 2017

  35. [36]

    NeuralNEB—neural networks can find reaction paths fast.Machine Learning: Science and Technology, 3(4):045022, December 2022

    Mathias Schreiner, Arghya Bhowmik, Tejs Vegge, Peter Bjørn Jørgensen, and Ole Winther. NeuralNEB—neural networks can find reaction paths fast.Machine Learning: Science and Technology, 3(4):045022, December 2022. ISSN 2632-2153. doi: 10.1088/2632-2153/aca23e. URLhttps://doi.org/10.1088/2632-2153/aca23e

    work page doi:10.1088/2632-2153/aca23e 2022

  36. [37]

    Transferable Learning of Reaction Pathways from Geometric Priors.The Journal of Physical Chemistry Letters, 16(45):11690–11699, November 2025

    Juno Nam, Miguel Steiner, Max Misterka, Soojung Yang, Avni Singhal, and Rafael Gómez- Bombarelli. Transferable Learning of Reaction Pathways from Geometric Priors.The Journal of Physical Chemistry Letters, 16(45):11690–11699, November 2025. doi: 10.1021/acs.jpclett. 5c02620. URLhttps://doi.org/10.1021/acs.jpclett.5c02620

    work page doi:10.1021/acs.jpclett 2025

  37. [38]

    Generative Modeling via Drifting

    Mingyang Deng, He Li, Tianhong Li, Yilun Du, and Kaiming He. Generative Modeling via Drifting, February 2026. URLhttp://arxiv.org/abs/2602.04770

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  38. [39]

    A new perspective on building efficient and expressive 3D equivariant graph neural networks, April 2023

    Weitao Du, Yuanqi Du, Limei Wang, Dieqiao Feng, Guifeng Wang, Shuiwang Ji, Carla Gomes, and Zhi-Ming Ma. A new perspective on building efficient and expressive 3D equivariant graph neural networks, April 2023. URLhttp://arxiv.org/abs/2304.04757

    work page arXiv 2023

  39. [40]

    Transition1x - a dataset for building generalizable reactive machine learning potentials.Scientific Data, 9(1):779, December 2022

    Mathias Schreiner, Arghya Bhowmik, Tejs Vegge, Jonas Busk, and Ole Winther. Transition1x - a dataset for building generalizable reactive machine learning potentials.Scientific Data, 9(1):779, December 2022. ISSN 2052-4463. doi: 10.1038/s41597-022-01870-w. URL https://www.nature.com/articles/s41597-022-01870-w

    work page doi:10.1038/s41597-022-01870-w 2022

  40. [41]

    Ucaxk, Sunwoo Kim, Juyong Lee, and Eunji Sim

    Minhyeok Lee, Jinyoung Jeong, Islambek Ashyrmamatov, Umit V . Ucaxk, Sunwoo Kim, Juyong Lee, and Eunji Sim. A dataset of chemical reaction pathways incorporating halogen chemistry. Scientific Data, 12(1):1655, October 2025. ISSN 2052-4463. doi: 10.1038/s41597-025-05944-3. URLhttps://www.nature.com/articles/s41597-025-05944-3

    work page doi:10.1038/s41597-025-05944-3 2025

  41. [42]

    Bernhard Schlegel

    H. Bernhard Schlegel. Geometry optimization.WIREs Computational Molecular Science, 1(5): 790–809, 2011. ISSN 1759-0884. doi: 10.1002/wcms.34. URL https://onlinelibrary. wiley.com/doi/abs/10.1002/wcms.34

    work page doi:10.1002/wcms.34 2011

  42. [43]

    The path of chemical reactions - the IRC approach.Accounts of Chemical Research, 14(12):363–368, December 1981

    Kenichi Fukui. The path of chemical reactions - the IRC approach.Accounts of Chemical Research, 14(12):363–368, December 1981. ISSN 0001-4842. doi: 10.1021/ar00072a001. URLhttps://doi.org/10.1021/ar00072a001

    work page doi:10.1021/ar00072a001 1981

  43. [44]

    Differential geometry of chemically reacting systems

    Akitomo Tachibana and Kenichi Fukui. Differential geometry of chemically reacting systems. Theoretica chimica acta, 49(4):321–347, December 1978. ISSN 1432-2234. doi: 10.1007/ BF00552483. URLhttps://doi.org/10.1007/BF00552483

    work page doi:10.1007/bf00552483 1978

  44. [45]

    W. Kabsch. A solution for the best rotation to relate two sets of vectors.Acta Crystallographica Section A, 32(5):922–923, 1976. ISSN 1600-5724. doi: 10.1107/S0567739476001873. URL https://onlinelibrary.wiley.com/doi/abs/10.1107/S0567739476001873. 12

    work page doi:10.1107/s0567739476001873 1976

  45. [46]

    The Journal of Chemical Physics 132(21), 214102 (2010)

    Søren Smidstrup, Andreas Pedersen, Kurt Stokbro, and Hannes Jónsson. Im- proved initial guess for minimum energy path calculations.The Journal of Chemical Physics, 140(21), June 2014. ISSN 0021-9606. doi: 10.1063/1. 4878664. URL https://pubs.aip.org/aip/jcp/article/140/21/214106/566679/ Improved-initial-guess-for-minimum-energy-path

    work page doi:10.1063/1 2014

  46. [47]

    Thompson, and Todd J

    Xiaolei Zhu, Keiran C. Thompson, and Todd J. Martínez. Geodesic interpolation for reaction pathways.The Journal of Chemical Physics, 150(16):164103, April 2019. ISSN 0021-9606, 1089-7690. doi: 10.1063/1.5090303. URL https://pubs.aip.org/jcp/article/150/ 16/164103/198363/Geodesic-interpolation-for-reaction-pathways

    work page doi:10.1063/1.5090303 2019

  47. [48]

    Vilhjálmur Ásgeirsson, Benedikt Orri Birgisson, Ragnar Bjornsson, Ute Becker, Frank Neese, Christoph Riplinger, and Hannes Jónsson. Nudged Elastic Band Method for Molecular Reactions Using Energy-Weighted Springs Combined with Eigenvector Following.Journal of Chemical Theory and Computation, 17(8):4929–4945, August 2021. ISSN 1549-9618. doi: 10.1021/acs. ...

    work page doi:10.1021/acs 2021

  48. [49]

    The atomic simulation environment—a Python library for working with atoms.Journal of Physics: Condensed Mat- ter, 29(27):273002, June 2017

    Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E Castelli, Rune Chris- tensen, Marcin Dułak, Jesper Friis, Michael N Groves, Bjørk Hammer, Cory Hargus, Eric D Hermes, Paul C Jennings, Peter Bjerre Jensen, James Kermode, John R Kitchin, Esben Leon- hard Kolsbjerg, Joseph Kubal, Kristen Kaasbjerg, Steen Lysgaard, Jón Bergmann Marons- son, ...

    work page doi:10.1088/1361-648x/aa680e 2017

  49. [50]

    Schütt, Oliver T

    Kristof T. Schütt, Oliver T. Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra, June 2021. URL http://arxiv. org/abs/2102.03150

    work page arXiv 2021

  50. [51]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization, January 2019. URLhttp://arxiv.org/abs/1711.05101

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  51. [52]

    Fukunaga and L

    K. Fukunaga and L. Hostetler. The estimation of the gradient of a density function, with applications in pattern recognition.IEEE Transactions on Information Theory, 21(1):32–40, January 1975. ISSN 1557-9654. doi: 10.1109/TIT.1975.1055330. URLhttps://ieeexplore. ieee.org/document/1055330

    work page doi:10.1109/tit.1975.1055330 1975

  52. [53]

    Mean shift, mode seeking, and clustering.IEEE Transactions on Pattern Analysis and Machine Intelligence, 17(8):790–799, August 1995

    Yizong Cheng. Mean shift, mode seeking, and clustering.IEEE Transactions on Pattern Analysis and Machine Intelligence, 17(8):790–799, August 1995. ISSN 1939-3539. doi: 10.1109/34.400568. URLhttps://ieeexplore.ieee.org/document/400568

    work page doi:10.1109/34.400568 1995

  53. [54]

    Comaniciu and P

    D. Comaniciu and P. Meer. Mean shift: A robust approach toward feature space analysis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):603–619, May 2002. ISSN 1939-3539. doi: 10.1109/34.1000236. URL https://ieeexplore.ieee.org/document/ 1000236. 13 A Baseline Implementation Details and Drift-React Configuration This appendix details t...

    work page doi:10.1109/34.1000236 2002

  54. [55]

    When the swarm is far from Y⋆, ˆZq may be much larger than ˆZp, leading to repulsion dominating attraction and the swarm drifting away from the target rather than toward it

    Decoupled normalization.The attractive and repulsive terms are normalized by indepen- dent partition functions, ˆZp and ˆZq respectively. When the swarm is far from Y⋆, ˆZq may be much larger than ˆZp, leading to repulsion dominating attraction and the swarm drifting away from the target rather than toward it. The multi-positive regime of Deng et al. [32]...

    work page