acceptodds
Under review as a conference paper at ICLR 2027

When Reaction Geometry Is Periodic: Understanding and Learning Transition States

Abstract

Learning transition-state (TS) geometries directly from reaction endpoints offers a route to accelerating reaction-pathway discovery, yet periodic systems pose distinct geometric challenges. Equivalent lattice representations alter endpoint displacements, local interactions can cross unit-cell boundaries, and reaction motion can be highly localized within extended structures. We characterize these effects across three periodic reaction benchmarks and introduce SurfReact-TS, a curated benchmark for constrained periodic surface reactions. Guided by this analysis, we develop PeriTS, a geometry-aware flow-matching framework that incorporates PBC-consistent endpoint conditioning, periodic interaction modeling, coordinate constraints, and reaction-focused learning. Across the three benchmarks, PeriTS or its matched One-shot counterpart achieves the lowest mean RMSD among the evaluated methods. Controlled studies further show that the value of iterative refinement depends strongly on the reaction regime, while increasing candidate breadth provides only modest gains in the settings studied. Selected-case DFT calculations provide complementary energy and force diagnostics. Overall, our results establish periodic TS prediction as a geometry-dependent learning problem in which representation, model design, and inference strategy should be adapted to the underlying reaction system.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for chemistry, Transition-state prediction

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_qzDsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PeriTS, a flow-matching framework for generating transition states (TSs) in periodic systems. The authors analyze difficulties specific to periodic structures, such as the minimum image convention, and address them with a periodic reaction prior, periodic reaction conditioning, and a training objective focused on mobile atoms only. The method is benchmarked on one newly curated and two publicly available datasets and compared against state-of-the-art models developed for TS generation of small organic molecules. The paper further compares multi-step and single-step inference, as well as the generation of multiple TS candidates per reaction. Evaluation is based on RMSD over all atoms and over mobile atoms only, complemented by a DFT-based energetic evaluation on a subset of 13 structures. While these insights are valuable, the paper in its current form has substantial issues in rigor and completeness, including missing citations, missing promised results, and inconsistencies in the text.

Critical strengths
  1. The paper adapts established techniques for periodic systems, such as the minimum image convention, to the task of TS generation.
  2. The method is compared against state-of-the-art baselines across multiple datasets.
  3. The paper contributes a newly curated dataset for periodic TS generation.
Critical weaknesses
  1. Limited novelty. Core components such as the minimum image convention and periodic neighbor lists are well-established techniques, but they are presented as novel contributions. The paper should more clearly distinguish what is new from what is standard practice.

  2. Low task difficulty and limited improvement. The midpoint RMSD on the benchmark datasets is already very low, so the mapping from the prior to the TS is comparatively easy, and the model improves only modestly over this starting point. On LiTraj in particular, the midpoint RMSD is so low that the models appear to learn mostly noise, which may also explain why the one-shot variant outperforms the multi-step baseline.

  3. Insufficient energetic evaluation. A DFT evaluation on 13 structures does not support statistically meaningful conclusions and should be expanded substantially.

  4. Missing methodological details. It is never explained how yty_t is obtained during training. The hyperparameters introduced in Appendix C3 are neither explained nor connected to the loss function, and no pseudo-algorithm is provided. Dataset sizes for the external benchmarks are also not reported. Together, these omissions limit reproducibility.

  5. Missing promised results. The comparison to SaddleFlow announced in Appendix D1 is missing.

  6. Presentation and referencing. The manuscript contains clear signs of unfinished or unreviewed text. Appendix E5 includes unfilled template placeholders in place of the actual computational settings (e.g., “[DFT code and version] with [exchange–correlation functional], [pseudopotential], an energy cutoff of [XX] eV”), so the DFT setup underlying the energetic evaluation is never specified. Several passages are also duplicated verbatim (e.g., Appendix C2 and C3; Tables 2 and 8). Such artifacts suggest the text was generated or assembled from templates without careful review. Moreover, the paper cites only 16 works, which suggests an incomplete treatment of related work. For example, AEFM is introduced in Appendix D2 without a citation.

  7. Multi-candidate ablation. No information is given on how the additional heads are trained. If they are trained like the original head on a frozen shared backbone, they would be expected to collapse to nearly identical predictions rather than provide a meaningful set of TS candidates.

Questions for authors
  1. How is yty_t obtained during training, and how do the hyperparameters in Appendix C3 enter the loss function?
  2. Why does TS-DFM produce predictions that are worse than the starting midpoint? An analysis of this behavior would be informative.
  3. Which starting point does React-OT use? If it uses the standard Cartesian midpoint, could you report results with the periodic midpoint for a fair comparison?
  4. How were the additional heads in the multi-candidate ablation trained? Could you report the pairwise RMSD between candidates to verify whether they collapse? Have you considered independent deep ensembles or inference from slightly noised midpoints as alternatives for generating diverse candidates?
Reviewer_woUJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a flow-matching-based generative model, PeriTS, for predicting periodic transition states, which also works for surface reactions since the input can specify a mobility mask and the dimensionality of PBC. The authors further introduce SurfReact-TS, a benchmark containing 330 constrained Pt-surface reaction triplets involving C/H adsorbates. The paper conducted experiments on SurfReact-TS, LiTraj, and MP20Bat. PeriTS achieves the lowest error on SurfReact-TS and MB20Bat, whereas the one-shot variant performs the best on LiTraj. The paper then conducted an ablation study that demonstrates that the PBC midpoint and periodic graph are helpful for learning periodic TS.

Critical strengths
  • The extension of the previous transition-state generative model to periodic systems, where PBC is handled explicitly, is straightforward and grounded.
  • Empirical results show that the proposed model outperforms the baseline.
Critical weaknesses
  • The new benchmark SurfReact-TS is claimed as one of the central contributions. However, there is very little detail on this benchmark. All we have is Appendix A.1. Without more detail, it is hard to gauge the actual contribution of this new benchmark and its validity.
  • The paper seems incomplete. The training losses are not specified at all. And in Appendix C.3, some auxiliary-loss weights are mentioned, but the definition of the auxiliary loss is never given. Also, some parts of the paper are unfinished, such as Appendix E.5.
Questions for authors
  • Why does one-shot perform better on LiTraj only? And it seems that all the method performs similar as the PBC midpoint which is strange. This seems to suggest that the model learns nothing on this benchmark.
  • As mentioned in Appendix C.4, on MP20Bat, the one-shot model uses a capped nearest-image graph, whereas PeriTS uses an uncapped multi-image graph. In the same paragraph, it says that for MB20Bat, a multi-image graph is needed. Shouldn't you also run the one-shot variant using the multi-image setup for a fair comparison?
  • It seems that overall the flow matching does not provide too much benefit, especially if one accounts for the inference cost as well (the paper uses NFE=6 by default). Why is that?
Reviewer_A7q9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies transition-state geometry prediction in periodic systems. It examines periodic endpoint alignment, cross-boundary neighborhoods, and localized atomic motion, then uses these observations to build PeriTS. Results across surface reactions, lithium trajectories, and battery structures show that the benefit of iterative prediction depends on the dataset. The paper also reports limited DFT diagnostics for predicted surface-reaction structures.

Critical strengths
  1. The geometric analysis leads to changes that can be tested directly. In particular, replacing the periodic midpoint worsens prediction in the main ablation. This makes a useful case for treating periodic endpoint equivalence explicitly.
  2. The one-shot comparison is informative, and the paper reports where it wins. Extra prediction steps help on some datasets but do not consistently help on LiTraj. This is more useful than presenting refinement as uniformly beneficial.

The paper acknowledges concurrent periodic flow prediction, including SaddleFlow, and distinguishes its task from potential-assisted search such as CatTSunami. The geometric analysis and its tests are the clearest contributions.

Critical weaknesses
  1. The reported DFT calculations have an unfinished methods paragraph. Appendix E.5, page 15, contains literal placeholders including “[DFT code and version],” “[exchange–correlation functional],” “[pseudopotential],” “[XX] eV,” and “[k-point setting],” followed by an instruction to add further settings. The actual settings are needed to assess the reported calculations. This is unfinished manuscript text; it does not establish AI authorship.
  2. The main results establish better geometric prediction more clearly than better transition-state search. The single-point energy and force diagnostics in Section 5.5 are limited, and the paper correctly acknowledges that they do not certify a saddle or its connectivity. It remains unclear whether the improved geometries make subsequent saddle searches more reliable or less expensive. That limits the practical conclusion that can be drawn from the RMSD gains.

I lean toward rejection. The geometric results are useful, but the DFT report needs to be completed and the practical claims should stay within what the validation shows.

Questions for authors
  1. What DFT settings were actually used for the reported diagnostic calculations? Were predictions and reference structures evaluated with the same setup?
  2. For the cases already studied, is there evidence that the predictions converge to the intended saddle under refinement, or improve convergence relative to the one-shot model? If this was not tested, please make that limit clear in the main conclusions.
Reviewer_xWqTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies transition-state geometry prediction in periodic systems and introduces the SurfReact-TS benchmark. It proposes PeriTS, which combines periodic endpoint conditioning, periodic message passing, coordinate constraints, and iterative flow prediction. Experiments compare it with a one-shot counterpart and adapted baselines across three datasets, supplemented by selected DFT calculations.

Critical strengths
  1. Periodic transition-state prediction is a relevant problem. The analysis of endpoint imaging, cross-boundary neighbours, and localised reaction motion provides a useful motivation for the model.
  2. The reported geometric improvements are promising. On SurfReact-TS, PeriTS reduces mean all-atom RMSD from 0.1174 to 0.1022 angstrom relative to the one-shot model (Table 1). The paper also acknowledges that refinement does not help uniformly across datasets.
Critical weaknesses
  1. Evidence for practical transition-state search remains limited. The DFT study covers only 13 single-point evaluations and does not establish saddle-point character, endpoint connectivity, or reduced optimisation cost. Appendix E.5 also contains placeholders for the DFT code, functional, cutoff, and other settings, preventing reproduction of these results.
  2. The added value of iterative refinement is not cleanly isolated on MP20Bat. The primary PeriTS model uses an uncapped multi-image graph, while One-shot uses a capped nearest-image graph (Appendix C.4). The small difference between them could therefore partly reflect the graph configuration. Matched configurations and repeated-seed uncertainty are needed to assess this improvement.
Questions for authors
  1. Can you provide the missing DFT settings and evidence that the predicted geometries improve subsequent saddle-point searches?
  2. How do PeriTS and One-shot compare on MP20Bat with identical graph settings and multiple training seeds?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.