acceptodds
Under review as a conference paper at ICLR 2027

ASPIRE: Saddle-Point Discovery through Set Prediction and Physical Refinement

Abstract

Predicting thermally activated diffusion and defect evolution with event-driven models requires identifying atomic rearrangement mechanisms and their activation barriers. Discovering the associated saddle points is a major computational bottleneck: multiple rearrangements may originate from one metastable state, while costly local searches can fail or repeatedly converge to the same saddle. To address this challenge, we introduce ASPIRE (Atomistic Saddle-Point Inference with Refinement for Events), a framework that predicts a set of saddle candidates from a single initial atomic environment and refines them through Dimer searches on the original interatomic potential. The framework’s equivariant set predictor, Ev-Quiformer, integrates (i) geometry-conditioned scalar/vector event slots for generating multiple saddle-point proposals and (ii) a decoder that maps each slot to a full atomic displacement field by combining atom, slot, and anchor-relative vectors with invariant coefficients. We also contribute two datasets: (i) BccFe-4vacAV-4000, comprising 4,000 four-vacancy body-centered cubic iron configurations and 65,450 reference events grouped by initial state for set supervision and post-refinement evaluation; and (ii) BccFe-1to4Vac, comprising 5,372 configurations with one to four vacancies each. Theoretically, we establish conditions for proposal equivariance and relate geometric error to reference-event recovery under explicit assumptions on local refinement. Experimentally, ASPIRE achieves 77.20% reference-event coverage on this benchmark, compared with 75.73% for a conventional Dimer baseline, while requiring approximately half as many Dimer force evaluations. In a timing evaluation on 50 configurations, ASPIRE reduces wall time per configuration from 478.8 s to 176.3 s under the stated hardware settings.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for science, set prediction, saddle-point discovery

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_q1myOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a network architecture that takes an initial atomic conformation as input and outputs an adjusted atomic conformation, passing it through a filter to obtain a more reasonable initial search solution. This is then further fed into a subsequent Dimer search algorithm to obtain the final balanced conformation. The authors claim that their method achieves higher coverage and reduces the computational cost of the Dimer search.

Critical strengths
  1. This paper proposes an accelerated Dimer search scheme that preprocesses the initial condition by adjusting the positions of atoms to the corresponding saddle point configurations. This idea is correct.
Critical weaknesses
  1. This method lacks any novelty; many previous papers have already explored this idea, and the network architecture is also unoriginal. This paper simply uses a different dataset. For example, the following papers have already proposed similar ideas: Peterson, Acceleration of saddle-point searches with machine learning (2016) Xu & Delin, A general optimization framework for mapping local transition-state networks (2026)

  2. The experimental design is unfair; the number of input conformations is 96 vs. 80, and the number of steps is also different. How can you prove that your 77.20% and 75.73% and the 2.7x speedup are correct?

  3. Lack of multi-random seed experiments

Questions for authors
  1. The issue of inconsistent input conformation numbers needs a more detailed explanation, or additional experiments should be conducted. Furthermore, the analysis of convergence steps could be more detailed. The multi-random seed experiment needs to be supplemented.
3 comments

Reviewer 4 — q1my

W1. Related work, method positioning, and novelty

Response. We thank the reviewer for suggesting these two related studies. We have cited and discussed Peterson (2016) and MOTO (Xu and Delin, 2026) in the “Rare-event simulation, saddle search, and reuse” paragraph of Related Work in the revised manuscript. Below, we respond by examining the specific content of these studies.

The similarities between these studies and ASPIRE are largely confined to the highly abstract goal of improving saddle-search efficiency. Their specific problem settings, the tasks assigned to machine learning, and their optimization mechanisms differ substantially from ours. This limited commonality at an abstract level is therefore insufficient to establish that prior work has already proposed our saddle-configuration set prediction problem and the corresponding method.

First, Peterson (2016) primarily studies accelerating NEB-based saddle-point searches through a learned potential energy surface, which differs from the problem considered here. NEB requires predefined initial and final states. Our task is to discover multiple potential saddle points from a single metastable atomic configuration, without knowing the final states or the number of possible transition events. The ML model in Peterson’s work approximates the potential energy surface to reduce the cost of energy evaluations during subsequent numerical optimization; it does not learn to predict target saddle-point configurations. This also differs from the preprocessing of search starting points described in the review. Thus, although both approaches use ML to assist saddle-point searches, this work does not establish a precedent for the open-ended, ML-based saddle-configuration set prediction formulation adopted in ASPIRE.

Second, MOTO, proposed by Xu and Delin (2026), is closer to our work in addressing open-ended multi-saddle discovery, but its optimization strategy is also distinct. MOTO uses NSGA-II-based, multi-objective, non-learning online numerical optimization to optimize local perturbations applied to a reference stable configuration, so that the perturbed configurations have more favorable properties for subsequent searches. Its objectives primarily concern diversity, physical feasibility, and minimum-curvature-mode separability, aiming to improve the efficiency and coverage of subsequent minimum-mode-following searches. This is a non-learning optimization strategy, rather than an ML method trained with reference saddle configurations to directly predict target saddle geometries. MOTO improves open-ended saddle searches by optimizing perturbations; our model learns target configurations under reference-saddle supervision. The two approaches therefore use different optimization mechanisms.

Describing ASPIRE primarily as preprocessing or optimizing the initial configurations for Dimer searches does not fully capture the task performed by the ML model. In the computational workflow, the predicted configurations indeed serve as inputs to subsequent Dimer refinement. However, defining the method solely by this interface overlooks its direct prediction target. ASPIRE does not train a model to find perturbations that merely provide more favorable starting points for Dimer. It uses the complete atomic displacement fields of reference saddle points as supervision, directly learning a mapping from a single initial atomic configuration to multiple potential saddle-point configurations, with set matching supporting joint prediction of multiple events.

In other words, our objective is for the ML model to perform a substantial part of candidate saddle-configuration discovery, followed by Dimer refinement and validation on the original physical potential energy surface, rather than relying primarily on Dimer to discover the target saddles from optimized perturbations. This decomposition also underlies ASPIRE’s theoretical motivation: the local convergence properties of refinement connect ML-based candidate geometry prediction with the accuracy requirements of physical saddle-point calculations. Therefore, although these approaches may ultimately invoke numerical saddle-search algorithms, similar downstream interfaces do not imply identical problem definitions, learning objectives, or optimization mechanisms.

We understand the motivation for comparing the methods at the level of their overall search workflows. A description such as “preparing inputs for a subsequent numerical search” captures a method’s position in that workflow, but leaves its prediction target, supervision, and learning mechanism unspecified. At this level of abstraction, substantively different methods can share the same description; treating that common description as sufficient evidence that the contribution has already been made would risk overlooking methodological innovations within an established computational workflow.

We describe the novelty of our work at the following levels. We fully recognize the foundations provided by prior saddle-search research and existing ML architectures. We do not claim to have introduced set prediction, equivariant learning, or matching as basic ideas. Our contribution is to establish the ASPIRE proposal-and-refinement framework for the computationally expensive discovery of multiple unknown transition events from a single initial state in highly symmetric crystalline environments. This framework translates an open-ended physical search problem into a clearly defined multi-saddle configuration set prediction task. Building on this formulation, we identify a structural connection to established set prediction ideas and adapt them to the distinctive requirements of three-dimensional atomic geometry.

The cited saddle-search studies do not propose our saddle-configuration set prediction method, while the original studies introducing the relevant ML components do not establish a solution framework for this open-ended multi-saddle discovery problem, develop the corresponding task-specific adaptations, or provide the knowledge and experience gained from adapting and applying those components in this setting. Their existence demonstrates the scientific computing and ML foundations on which our work builds; it does not establish that the cross-domain connection, specific problem formulation, and task-specific adaptations developed here have already been explored.

Q1/W2. Number of evaluated input configurations

Response. We respond by describing the two evaluation settings. We evaluate both event coverage and force-evaluation cost on the full test set of 200 initial configurations. Force calls quantify the computational cost of physical refinement after candidate prediction, allowing us to focus on how the method works: how predictions close to saddle points reduce the cost of the subsequent physical search. It is also an experiment that can help distinguish the difference between how our method and the work that the reviewer mentioned above works. We conduct a larger-scale empirical evaluation of this specific mechanism across the full test set.

The 50-configuration wall-clock study serves a complementary purpose. Since these methods are not mutually exclusive and can be used together, this comparison primarily aims to identify their respective advantages and trade-offs, providing a basis for choosing or combining them. Such measurements require fixed hardware allocations, and evaluating multiple methods involves substantial runtime. We therefore supplement the full-set event-recovery and force-cost evaluation with wall-clock comparisons of multiple methods on a shared set of 50 configurations. This keeps both the input configurations and hardware resources consistent while providing evidence on practical runtime and complementary event discovery. The full-set recovery and force-cost results are reported in Section 5.3, and the 50-configuration wall-time and complementarity results are reported in Section 5.5 of the revised manuscript.

Q1/W3. Evaluation across random seeds

Response. We have completed three additional independent Random Dimer runs, varying the seeds that control random initialization of the search starts on the same 200 evaluation configurations, while keeping the search budget, Dimer cap, and reference catalogue unchanged. Full run-level results and seed-to-seed uncertainty estimates are reported in Section 5 of our common response, “Random Dimer Initialization Seeds and Uncertainty Intervals”. The summary statistics include the original run and the three additional runs. These statistics characterize the variability of the Random Dimer baseline under different search initializations. We retain the original run in the main comparisons.

Reviewer_9nrgOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ASPIRE discovers single-ended saddles for kinetic Monte Carlo on the fly. From one relaxed configuration, it predicts 96 approximate saddle geometries in a single forward pass. The predictor is an Equiformer encoder with DETR-style slots, trained with repeated Hungarian matching. Each proposed starting location from the model is then refined by a Dimer search on the true potential and an optional judge filters proposals before refinement.

The authors release a benchmark of 4000 four-vacancy bcc Fe configurations with 65450 reference events, generated by 80 random start Dimer searches per configuration. On 200 test states, ASPIRE reaches 77.20% coverage and 87.03% RWC, against 75.73% and 89.09% for the 80 random start Dimer searchers, with half the recorded force calls.

My expertise is generative modeling and ML evaluation, not atomistic simulation, so I focus on the learning formulation and the evaluation.

Critical strengths
  1. Sensible design and reusable benchmark. The paper follows a principled proposed-then-refine process. Every proposal is validated on the true potential, so prediction errors cost compute but cannot introduce unphysical events. The dataset and code are released with detailed appendices.

  2. Clear efficiency gain at small search budgets in Fig. 3. The comparison includes event reuse baselines such as KDB and k-ART style, and not only random search.

Critical weaknesses
  1. Recovery is at or below random search, so the contribution is mainly efficiency. However, the efficiency comparison is not matched. Coverage is +1.47 pts over random Dimer while RWC is −2.06 (Fig. 3). On the 50-state subset, ASPIRE is below random on both metrics (Fig. 5). No seeds or confidence intervals are reported. Also, random dimer is not an oracle here. Each reference catalogue is itself one 80-start random run and an independent run recovers only 75.73% of it. So the catalogue is incomplete but ASPIRE is trained to imitate it, and valid saddles outside it receive no credit. The cost gain is measured under unequal protocols. ASPIRE uses finer Dimer steps and 300 steps cap vs. 450 for random. The force count excludes inference and relaxation calls. ASPIRE additionally uses an A100 GPU. KDB-494 beats ASPIRE on coverage (82.64 vs. 75.68). Its slower wall time runtime can depend on the authors’ implementation and its small library.
  2. Very restricted setting and modest novelty. The setting is iron atoms only, four vacancies, and single potential. Train and test are drawn from the same generator. There is no transfer test to other number of vacancies. The model proposals are deterministic, so a missed event stays missed on every revisit of that state. The architecture combines known components: Equiformer, DETR, Group DETR, DAB-DETR.
Questions for authors
  1. Can you rescore all methods against a union catalogue of every valid saddle any method found? How many of ASPIRE’s unmatched outputs are new events?
  2. Report confidence intervals over test states and over seeds for the ASPIRE minus random dimer differences.
  3. Please report 1) results under matched step settings and caps, with all force calls and inference time counted 2) KDB with a library built from the full training set 3) the definition of KDB-112.
8 comments

Reviewer 3 — 9nrg

We thank the reviewer for the constructive feedback and these specific questions. We address the union-catalogue evaluation, uncertainty intervals, the search-cap settings and locations of the updated results, KDB with a library built from the full training set, and the definition of KDB-112 below.

Q1/W1. Can you rescore all methods against a union catalogue of every valid saddle any method found? How many of ASPIRE’s unmatched outputs are new events?

Response. We thank the reviewer for this valuable suggestion to evaluate all methods against a union catalogue. We agree that this is a particularly informative analysis: incorporating valid discoveries from all methods makes the evaluation more comprehensive and provides another perspective on the distinctive characteristics and value of ASPIRE, by showing which additional events it discovers and how those discoveries complement other methods. This analysis can therefore strengthen the evidence for ASPIRE’s specific contributions to event discovery beyond what its aggregate scores against a single reference catalogue capture. We present the catalogue construction and the resulting comparisons below.

We construct separate union catalogues for the 200-configuration search-budget experiment and the 50-configuration wall-clock evaluation subset. We retain events accepted by the original Dimer convergence and connectivity checks with at most 80 proposals/search starts per configuration and a 300-step translation cap. We preserve the 120-search reference catalogue, match discoveries using the same event-matching convention as the main evaluation, and jointly deduplicate reference-unmatched events across participating methods. The resulting catalogues contain 5,301 events for 200 configurations and 2,198 events for 50 configurations. The 200-configuration catalogue contains 3,719 reference events and 1,582 additional events; the 50-configuration catalogue contains 937 reference events and 1,261 additional events. Each experiment uses its own fixed catalogue, including across all search-budget curve points. RWC is evaluated at 800 K with a common prefactor, normalized within each configuration and then averaged across configurations.

We report coverage and RWC against these catalogues, together with the events found by ASPIRE but missed by each comparison method and their contributions to union RWC.

Full test set: 200 configurations

MethodUnion coverageUnion RWCEvents found by ASPIRE but missed by this methodRWC contribution of these additional events (percentage points)
ASPIRE81.48%93.23%——
Random Dimer56.39%80.56%1,70516.31

Wall-clock evaluation subset: 50 configurations

MethodUnion coverageUnion RWCEvents found by ASPIRE but missed by this methodRWC contribution of these additional events (percentage points)
ASPIRE49.97%85.13%——
Random Dimer34.57%71.73%42716.77
KeyBank-3252.41%51.22%40238.24
KeyBank-12868.01%63.85%27526.87
KDB-11269.03%83.64%19613.36
KDB-49477.34%83.86%15213.09
Union of all other methods95.12%96.45%332.14

We also count distinct events outside the 120-search reference catalogue after matching and cross-method deduplication:

ExperimentEvents in the 120-search reference catalogueASPIRE events outside the reference catalogueOf these, events missed by all other participating methods
200 configurations3,7191,3791,090
50 configurations93735027

These results quantify ASPIRE’s coverage of the common catalogue and its additional discoveries relative to other methods. ASPIRE can be used together with existing search and retrieval methods to improve event coverage through complementary discoveries. Combined RWC against the 120-search reference catalogue is reported in our general response on method performance and practical value.

In the revised manuscript, the physically validated beyond-reference counts are reported in Section 5.6, while the complementary recovery of reference events is reported in Section 5.5 and Table 3. The complete union-catalogue rescoring above is reported in Appendix H.4 and Table 8 of the revised manuscript. We thank the reviewer again for suggesting the union catalogue. This analysis adds quantitative evidence on common-catalogue coverage, additional discoveries, and events found exclusively by ASPIRE, providing a more complete account of its complementary value alongside existing methods.

Q2/W1. Uncertainty intervals

Response. We thank the reviewer for requesting uncertainty intervals. We agree that this deserves careful attention, since quantifying uncertainty would help assess the precision of the reported event-recovery and computational-cost estimates and the strength of the comparisons.

We have completed three additional independent runs of Random Dimer with different seeds for its random search initialization on the same 200 evaluation configurations. The search budget, Dimer translation cap, and reference catalogue remain fixed. Section 5 of our common response, “Random Dimer Initialization Seeds and Uncertainty Intervals,” reports the full results and the mean, sample standard deviation, and 95% t-confidence interval calculated across all four runs, including the original experiment. We thank the reviewer again for encouraging this analysis, which quantifies sensitivity to Random Dimer initialization and strengthens the empirical basis of our baseline comparison.

Q3(1)/W1. Search budgets, Dimer caps, and locations of the updated results

Response. We thank the reviewer for raising these questions. Specifying this information helps readers check the search resources and stopping conditions used in the comparisons and locate the corresponding evaluations. We summarize the changes, the settings used in the different experiments, and where the corresponding results can be found below.

For the controlled quality comparison between ASPIRE and Random Dimer, we align both methods to 80 proposals/search starts per configuration and a cap of 300 Dimer translation steps, and reevaluate the archived search results under these shared limits. The settings are described in “Comparison Experiment Settings”, and the updated event-coverage, RWC, and complementary-discovery results are reported in “Method Performance, Practical Value, and Complementary Event Discovery.” The GPU comparison provides ASPIRE and the non-ML baselines with one A100 GPU and K = 4 concurrent GPU-resident Dimer search slots, retaining the original CPU allocation, as described in “Comparison Experiment Settings.”

The updated event-recovery and force-cost results are reported in Section 5.3 and Figures 3–4; architectural ablations in Section 5.4 and Table 2; wall time and complementarity in Section 5.5, Figure 5, and Table 3; and beyond-reference discoveries in Section 5.6.

The union-catalogue construction and comparisons under the 80-start, 300-step limits are reported in Q1/W1 of this response and in Appendix H.4 and Table 8 of the revised manuscript. The reference-catalogue construction is specified separately from the proposal budget: only the test reference catalogue was extended from 80 to 120 searches per configuration; the training and model-selection reference data remain based on the original 80 searches. These reference-data settings are described in “Comparison Experiment Settings” above and in Sections 5.1–5.3 and Appendix H of the revised manuscript. We thank the reviewer again for drawing attention to the comparison settings and result locations. Reevaluation under aligned budgets and step caps strengthens the empirical basis for comparing reference-event recovery and computational cost, while the added descriptions of the settings and manuscript locations make the results easier to check.

Reviewer_mDr4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ASPIRE predicts multiple saddle-point candidates from a single atomic configuration using an equivariant set predictor, followed by Dimer refinement on the original interatomic potential. On four-vacancy BCC iron, it achieves comparable reference-event recovery with substantially fewer recorded search-force evaluations.

Critical strengths

The paper addresses an important bottleneck in event-driven atomistic simulation: discovering multiple distinct transitions without known products.

Refinement on the original potential provides a sound physical foundation. The grouped event dataset and evaluation of distinct-event recovery are useful contributions.

Ablations support the effectiveness of the spatial-anchor decoder.

Critical weaknesses

The methodological advance is incremental. Equivariant encoding, slot-based prediction, matching supervision, and proposal refinement are established ideas. The benefits over simpler learned multi-proposal baselines remain unclear.

Efficiency comparisons use different Dimer settings and hardware resources. The reported force-call reduction excludes parts of the workflow, limiting attribution of the speedup to the proposed model.

Validation is restricted to four-vacancy iron under one potential. Event coverage improves only slightly, while rate-weighted coverage decreases from 89.09% to 87.03%. Finite reference catalogues and the absence of trajectory-level validation leave the kinetic consequences unresolved.

Questions for authors

Does the advantage persist with matched Dimer settings, comparable resources, and complete cost accounting?

How does Ev-Quiformer compare with a simpler multi-head or sampled equivariant predictor under equal refinement budgets?

Which kinetically important events are missed, and how do these omissions affect KMC observables?

3 comments

Q1/W2. Does the advantage persist with matched Dimer settings, comparable resources, and complete cost accounting?

Response. We thank the reviewer for raising this valuable question about comparison fairness and complete cost accounting. We agree that these issues deserve careful examination, since matching proposal budgets, Dimer step caps, and hardware resources while including upfront costs would test whether the reported performance and efficiency gains persist under comparable conditions. We respond below with the aligned search settings, resource allocation, and cost analysis.

The detailed settings are presented in “C2. Comparison Experiment Settings” in our common response. For the controlled quality comparison, we align both methods to 80 proposals/search starts per initial configuration and a cap of 300 Dimer translation steps, and provide GPU access to non-ML baselines. ASPIRE’s physical refinement also uses the GPU.

In the existing K = 4 GPU timing measurements, ASPIRE saves approximately 87.34 seconds per configuration on average. Its measured wall time includes online candidate generation, model cold startup amortized over the evaluated configurations, and Dimer refinement. The aligned reference-recovery and force-cost comparison is reported in Section 5.3 and Figure 3, and the complete measured wall-time comparison is summarized in our common response and reported in Section 5.5 and Figure 5 of the revised manuscript.

For upfront costs, data generation took approximately 20 hours and model training approximately 33 hours. Assuming comparable average online costs and sequential execution of the pipeline applications, the measured savings would offset data generation alone after approximately 825 applications of the inference-and-refinement pipeline, or both costs after approximately 2,185 applications when counted as a fully sequential 53-hour investment. These applications can include repeated use on previously encountered configurations.

In on-the-fly KMC, conventional searches already required by the simulation can also supply training data. Training on a separate GPU can proceed while CPU simulation and data collection continue, and the trained model can be introduced at a subsequent KMC step. Thus, the 53-hour total describes fully sequential accounting, while these activities can overlap in this workflow. This reply can be read together with our response to Reviewer MUHz, Q2/W1, which provides the detailed cost and amortization analysis.

We thank the reviewer again for requesting this comparison. Reevaluation under aligned search budgets and computational resources, together with joint accounting of upfront and online costs, provides more complete empirical evidence for the reported practical computational benefits.

Q2/W1. How does Ev-Quiformer compare with a simpler multi-head or sampled equivariant predictor under equal refinement budgets?

Response. We thank the reviewer for this valuable suggestion to compare Ev-Quiformer with simpler equivariant predictors. We agree that this comparison deserves careful consideration, since evaluating alternatives under equal refinement budgets would test what gains our predictor provides in event coverage, rate-weighted coverage, and physical-search cost relative to simpler architectures. We discuss the following predictor alternatives and report the multi-head comparison below.

ApproachCandidate generation
Multi-head predictorShared encoder; one displacement field per output head
Randomly conditioned predictorSample different latent variables z for the same input X; predict fθ(X, z)
Iterative generative modelGenerate candidates from noise through multiple sampling steps

For the multi-head predictor, we have added a direct experimental comparison. Judge selection is disabled for both models. We evaluate both models on the same 50 test configurations, using the 120-search reference catalogue, a cap of 300 Dimer translation steps, and identical geometric and barrier-matching criteria:

MethodEvent coverage ↑RWC at 800 K ↑Mean force calls ↓
Ev-Quiformer78.94%92.84%44,157.42
Equivariant multi-head64.99%82.00%69,042.26

This comparison is now reported together with the other architectural ablations in Section 5.4 and Table 2 of the revised manuscript.

We appreciate the reviewer’s suggestion of a randomly conditioned predictor. The formulation fθ(X, z) is simple and intuitive: it explicitly conditions candidate generation on the initial configuration X and introduces randomness through z. This makes it a natural way to formulate the generation of multiple candidates from the same initial state. However, developing this formulation into an effective saddle-discovery model also involves several substantive learning and implementation problems.

An effective implementation needs to map variations in z to physically meaningful differences in which atoms participate, their displacement directions, and their cooperative motions. The reference saddle points form an unordered set, with no predefined labels assigning random conditions to particular events. The training mechanism therefore needs to learn these associations and use a finite sample budget to cover distinct physical events.

Our consideration of generative proposals also involves three aspects of this physical search problem.

First, sampling diversity does not necessarily translate into distinct events after refinement. Probability mass may concentrate on a few events, and geometrically different candidates may converge to the same saddle, leaving less-probable events missed under a fixed budget.

Second, repeated discoveries incur both generation and physical-refinement costs. Iterative generative models additionally require multiple network evaluations per candidate. These costs matter in our setting because repeated saddle searches are the computational bottleneck, and repeated recovery consumes resources without adding distinct events.

Third, the crystalline setting provides substantial physical structure. Lattice symmetry, local defect arrangements, and physical interactions constrain the possible atomic motions, and related collective rearrangement patterns can recur across configurations. This structure provides information for learning how an initial state gives rise to different saddle geometries.

These points can be read and discussed together with our response to Reviewer MUHz, Q5/W3.

On the other hand, our formulation is particularly intuitive for this task. Given the initial atomic configuration and interatomic potential, the associated saddle events are determined by the physical problem and can include multiple distinct rearrangements. Their reference geometries form an unordered set. This leads directly to our adaptation of deterministic set prediction and bipartite matching. The local-convergence motivation further supports learning sufficiently close candidate geometries, with physical refinement providing final accuracy and validation. Target configurations and diversity are modeled directly in our supervision strategy to enable the abilities of our method to reduce the iterative repeated costly discoveries and latent variable mapping modeling and learning.

We thank the reviewer again for suggesting this additional comparison. The multi-head experiment directly compares candidate event recovery and search cost under the same refinement budget, providing empirical support for the task-specific architectural adaptations in Ev-Quiformer.

Q3/W3. Which kinetically important events are missed, and how do these omissions affect KMC observables?

Response. We thank the reviewer for asking about missed events and their kinetic implications. This question deserves careful consideration: analyzing the rate weights of missed events helps assess the practical significance of event recovery for dominant escape channels, waiting times, and transition-selection probabilities. We further analyzed missed events and their kinetic importance across the full test set of 200 configurations, with both methods using 80 search starts per configuration and a cap of 300 Dimer translation steps. All results below use the 120-search reference catalogue, containing 3,719 reference events, with the same event-matching criteria as our main reference-catalogue evaluation. Events are grouped by their contribution to the total reference escape rate of their respective initial configurations at 800 K, assuming a common prefactor.

Per-event rate contributionReference eventsMissed by ASPIREMissed by Random Dimer
<1%1,564653789
1%–10%1,602140380
10%–50%5093562
≥50%4443
Total3,7198321,234

The updated ASPIRE misses fewer reference events and achieves higher rate-weighted coverage than Random Dimer. Against the 120-search reference catalogue, the mean RWC is 91.83% for ASPIRE and 86.24% for Random Dimer, with per-configuration medians of 98.32% and 92.56%, respectively. Of the 832 reference events missed by ASPIRE, 793 individually contribute less than 10% of their initial configuration’s total reference escape rate.

Missing events underestimates the total escape rate and consequently increases the waiting time assigned by KMC. For a given state, if the reference catalogue’s total escape rate is K and the recovered fraction is r, the mean waiting time increases from 1/K to 1/(rK), a factor of 1/r. For example, missing a channel carrying half the total rate doubles the mean waiting time. The selection probabilities of the remaining events are also renormalized, changing transition branching and potentially affecting diffusion and defect-cluster evolution. This motivates evaluating both event coverage and rate-weighted coverage.

ASPIRE’s higher RWC indicates more complete recovery of the reference-catalogue escape rates used in KMC, while complementary use with conventional searches can further mitigate omissions. ASPIRE recovers 615 reference events missed by Random Dimer, while Random Dimer contributes 213 reference events missed by ASPIRE. Combining their existing search results raises the mean reference RWC to 95.29% and recovers 43 of the 44 reference events contributing at least 50% of their respective initial configurations’ total reference escape rates. These results demonstrate the complementarity of the two methods and support their individual or combined use according to computational budgets and kinetic coverage requirements. The combined results correspond to the sum of both search budgets.

This analysis quantifies the rate weight of missed reference events and explains their effects on KMC waiting times and transition probabilities at a given state. Comparing KMC trajectories can trace the accumulated effects of these differences in waiting times and branching probabilities and thereby assess their impact on long-term observables such as diffusion coefficients and cluster evolution. Complementary event recovery quantifies how combining searches improves kinetic coverage. We thank the reviewer again for drawing attention to these kinetic implications. The question prompted us to add rate-stratified omission analyses and complementary-channel statistics, strengthening the connection between event-discovery results and kinetic behavior.

Reviewer_MUHzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a framework that accelerates saddle point discovery in atomistic systems. Saddle point discovery in such systems is challenging because multiple distinct atomic rearrangement events can occur from the same initial metastable configuration. The paper formulate the problem as set prediction and introduce an equivariant set predictor with 96 event slots to generate multiple saddle point proposals from single initial configuration. The model uses a sequential one to many matching objective to learn from an arbitrary number of reference events. The predicted saddle points can then be refined with Dimer method. They show with 4 vacancy configurations in bcc iron benchmark that their method achieves similar reference event coverage while reducing force evaluations by half, compared to conventional Dimer search method, and also reports significant reduction in wall time.

Critical strengths
  • The reduction in Dimer force evaluations and wall time is clearly shown.
  • The use of set predictor makes sense for their task of handling multiple possible events starting from a single configuration. Sequential one-to-many matching is a reasonable design to supervise multiple candidates predictions with an arbitrary number of reference events.
Critical weaknesses
  • While the improvement in computational efficiency is the main point of the paper, there is no discussion of the cost of generating the training dataset, training the model, and set predictor inference. It seems that the method requires a substantial number of reference events for training. These compute might be needed anyway if we just use Dimer for everything, but the benefit of model training could be marginal if we need to run Dimer for most of the configs to collect training data and can run inference on only the fraction of it.
  • The benchmark is done on only one system (4 vacancy configurations in bcc iron). I think the paper should show the method works for at least one more system to show the general applicability of the method. Also, no experiment was done to evaluate if the model can transfer across different vacancy counts or atomistic systems.
  • The motivation of the proposed methods is clear, but there's no ablation study to justify that they were critical for the framework. (e.g., equivariance, set predictor)
Questions for authors

I listed the questions in the order of importance.

  • Can you provide evaluation on one more benchmark?
  • Also it would be great if you can provide end-to-end cost analysis including computational cost of training data generation, model training, and model inference.
  • How does performance change as the number of training configurations or reference events (compared to the entire search space) is varied? I am curious how much training data is required to obtain the efficiency gains.
  • Was equivariance neccessary? Can't non-equivariant model learn equivariance from data (while making the inference faster)? Or does sample efficiency improve a lot with equivariance? Only asking this because it seems like it is proposed as one of the main method.
  • Curious why you used deterministic set predictor instead of some generative model-based methods to generate a distribution of events. In particular, how does this determinstic formulation affect event diversity?
  • Can your trained model generalize to configurations with different vacancy counts or other physical systems?
11 comments

Reviewer 1 — MUHz

We thank Reviewer MUHz for the careful evaluation and for recognizing the computational savings, the suitability of set prediction, and the rationale for sequential one-to-many matching. Your questions have helped us explain the scope, data requirements, computational costs, and design motivations of our work more fully. We respond to each concern below.

Q1/W2. Can you provide evaluation on another benchmark?

Response. We thank the reviewer for suggesting evaluation on an additional benchmark. We agree that this question deserves careful attention: additional benchmarks across material systems and defect settings would strengthen the empirical assessment of generalization and provide more comprehensive evidence of the method’s effectiveness. We also very much want to get such additional validation results.

Establishing a benchmark with the same or similarly demanding challenges requires collecting and validating reference saddle-point data, followed by model training, physical refinement, method comparisons, ablations, and computational cost evaluation. Completing this full workflow takes substantial time.

Establishing the benchmark and connecting data generation, the learning task, candidate prediction, physical refinement, and evaluation already form part of our work. Together, they provide a reusable, representative foundation for studying and comparing ML methods for this scientific problem. The multi-vacancy bcc Fe benchmark provides a representative atomistic setting for studying point-defect configurations and their local transition dynamics in metallic crystals. The highly symmetric bcc lattice and vacancy-induced local structural changes bring together geometric symmetry constraints and complex collective atomic motions. Configurations containing dispersed vacancies, small vacancy clusters coexisting with dispersed vacancies, multiple small vacancy clusters, and connected vacancy clusters cover a range of local environments relevant to vacancy migration, interactions, aggregation, and structural rearrangement. The rich migration and rearrangement mechanisms of small vacancy clusters are supported by previous studies (Rahman et al., 2023; Fan et al., 2011). These features provide a physically grounded setting for evaluating the discovery of multiple competing transition events from a single initial configuration within the local configurational and kinetic environments of point defects in metallic crystals.

Therefore, we are currently using this representative system to establish a methodological framework and conduct a comprehensive assessment, while also incorporating broader validation into our objectives.

Q2/W1. Can you provide an end-to-end cost analysis, including data generation, training, and inference?

Response. We thank the reviewer for asking for a complete cost analysis covering data generation, training, inference, and physical refinement. We agree that these costs warrant careful consideration, since accounting for both upfront and online costs would establish when accumulated search savings offset data generation and training, providing a firmer basis for assessing the computational benefit of deployment. We present the measured costs and the corresponding amortization estimates below.

In our experiments, data generation took approximately 20 hours, and model training took approximately 33 hours. In the existing K = 4 GPU timing measurements, Random Dimer and ASPIRE required an average of 216.18 and 128.84 seconds per configuration, respectively, saving approximately 87.34 seconds per configuration inference and subsequent physical refinement. The ASPIRE timing includes online candidate generation and model cold startup amortized over the evaluated configurations.

Assuming comparable average costs, the accumulated online savings would offset the 20-hour data-generation time after approximately 825 configuration inferences performed sequentially, each followed by physical refinement. If both data generation and training are counted as a sequential 53-hour upfront investment, the corresponding estimate is approximately 2,185 configuration inferences performed sequentially.

These counts refer to applications of the inference-and-refinement pipeline, including repeated use on previously encountered configurations. We use the 200 test configurations to assess generalization across vacancy topologies. Subsequent simulations can involve substantially more than 200 applications of the trained model, as discussed in Q3/W1.

Long-term materials evolution provides substantial demand for such repeated applications. At 800 K with the common prefactor of 10¹³ s⁻¹ used in our setting, the median single-channel characteristic waiting time among the reference events is approximately 2.57 ns. Accounting for competition among events, the median of the mean escape waiting times across the test configurations is approximately 66 ps. Nuclear-material service-life research concerns evolution over years to decades, corresponding to application timescales of approximately 10⁷–10⁹ seconds (Long-Term Materials Performance Research). This separation of timescales motivates repeated local event searches in long-term evolution modeling. In systems containing multiple active volumes (AVs) requiring searches, the model can serve different AVs and their subsequent configuration updates. Savings therefore accumulate across both spatial regions and successive updates, providing sustained opportunities to amortize the upfront costs.

In on-the-fly KMC, simulation, data collection, and model training can overlap. A simulation can proceed with conventional searches while accumulating initial configurations and physically validated saddle events. Model training can begin using the data already collected while simulation and data collection continue. After training, the model can be introduced at a subsequent KMC step, with its proposals refined and validated on the original potential energy surface.

In this workflow, searches already required by the simulation also supply training data. Training can begin before data collection is complete, and the simulation can continue while training proceeds. With training on a separate GPU and conventional simulation on CPUs, these activities can overlap. The 20-hour data-generation time and 33-hour training time therefore describe a fully sequential accounting scenario, while the overlapping workflow allows these activities to proceed concurrently. We report both the data-generation-only and fully sequential amortization estimates to make these costs explicit. We thank the reviewer again for this suggestion: complete cost accounting and amortization analysis provide a quantitative basis for assessing when the accumulated computational savings from repeated use can offset the upfront investment.

Comments on the paper

5 comments

Common response to all reviewers

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their time, careful reading, thoughtful comments, and constructive suggestions, and the Area Chair for overseeing the review process. We appreciate the recognition of the computational savings offered by our approach, the rationale and suitability of set prediction for this task. In response to the reviewers’ comments, we have revised the manuscript and carried out the following additional experiments and analyses:

  1. Research positioning and contributions: We have expanded the presentation of the open-ended multi-saddle discovery problem, ASPIRE’s solution framework, and the task-specific adaptations that enable Ev-Quiformer to predict sets of complete 3D atomic configurations. We have also added references and discussion of the relationship between these contributions and prior saddle-search and machine learning methods, as well as other deterministic mapping or generative sampling alternatives for this specific problem.

  2. Comparison settings and reference catalogues: We have aligned the controlled ASPIRE–Random Dimer comparison to a shared proposal budget and Dimer step cap, provided both ML and non-ML methods with GPU-assisted physical search, and expanded only the test reference catalogues to 120 searches per configuration while retaining 80 searches for training and model selection.

  3. Model design and additional validation: We have updated the event-slot design and added controlled architectural ablations, comparisons with a simpler equivariant multi-head predictor and an adapted generative proposal model, event-library scaling analysis, and analyses of union catalogues, complementary discoveries, and the kinetic importance of missed events.

  4. Computational cost and practical value: We have added measured GPU wall-clock comparisons and cost accounting for data generation, training, inference, retrieval, and physical refinement, together with amortization estimates and a discussion of repeated use in on-the-fly KMC.

Key updated results and protocol details are included in the main paper and appendices. Below, we summarize the shared concerns and corresponding revisions, followed by point-by-point responses to each reviewer, with references to the relevant sections, figures, and tables in the revised manuscript.

Common concerns and major revisions

C1. Motivation and Contribution — Reviewers mDr4, 9nrg, and q1my.

We thank the reviewers for their questions about the motivation and contributions. The question helps make the sources of our contribution and the rationale for our methodology clearer. We respond below by presenting our contributions: bringing a science application problem with distinctive application settings and challenges into machine learning research and translating it into a learning task; identifying cross-domain connections and transferring, adapting, and extending established ML ideas to meet the problem’s geometric and physical requirements; and contributing new knowledge and experience in adapting and applying machine-learning ideas and methods.

First, we establish a framework that translates a scientific problem with distinctive settings and demanding challenges, which has not yet been sufficiently explored using ML, into a learning task with an intuitive formulation and explicit objectives. Specifically, we study the use of machine learning for assisted, efficient, repeated open-ended multi-saddle discovery in highly symmetric body-centered cubic (bcc) Fe crystals, addressing a computational bottleneck in on-the-fly kinetic Monte Carlo (KMC) simulations. Starting from a single metastable atomic configuration, the task is to efficiently and accurately discover multiple distinct saddle points without prior knowledge of product states or possible transition events.

This problem imposes stringent physical accuracy requirements. Even with an inexpensive EAM potential, for which individual energy and force evaluations are relatively cheap, repeated Dimer searches, duplicate convergence, and the accumulated number of force evaluations create a substantial computational bottleneck.

To address this challenge, ASPIRE combines learned saddle-point proposals with physical refinement. This design is motivated by the local convergence conditions for minimum-mode refinement analyzed in Proposition 1 and empirically validated by our evaluations of proposal quality and search efficiency. This decomposition makes the open-ended, computationally expensive multi-saddle search problem naturally expressible as a configuration set prediction task: predicting multiple saddle-point displacement fields from a single 3D atomic configuration. It enables task-specific ML design that exploits the speed advantages of direct candidate prediction and its generalization to unseen configurations, while refinement on the original potential energy surface ensures the required final accuracy and physical validity.

Second, we identify a structural connection between this learning problem and established set prediction ideas, and adapt those ideas to its distinctive 3D geometric and physical requirements. Ev-Quiformer is designed around the structure of this task. Predicting an unordered set of possible saddle points from one initial configuration naturally motivates set prediction. However, unlike the object categories and low-dimensional bounding boxes typically predicted by DETR-style methods, each set element here is a complete 3D saddle-point candidate in a shared, high-dimensional atomic configuration space, involving complex collective atomic displacements and rotational equivariance. Existing set prediction methods therefore provide a suitable abstraction but require task-specific adaptations.

We use Equiformer to extract equivariant atomic representations and design conditioned scalar event slots, adapting conventional queries with scalar features, encoding event context while enabling equivariance, and an equivariant atomic displacement-field decoder that maps these representations to complete, equivariant atomic displacement fields. Together with sequential one-to-many Hungarian matching, these components enable set-supervised learning to jointly generate multiple candidate saddle-point configurations in a single forward pass.

To the best of our knowledge at the time of submission, this is the first work to adapt and use bipartite-matching-style supervision for set prediction to predict a set of complete 3D molecular or atomic configurations from a single complete 3D configuration in one forward pass. Our results also demonstrate the value of transferring this learning idea across domains. Our novelty concerns the identification of this insufficiently explored structural connection and the nontrivial adaptations needed for 3D atomic configuration prediction. Physical refinement on the original potential energy surface, distinct-event coverage, and search-cost evaluations demonstrate the practical value of this formulation and adaptation. Together, the problem formulation, methodological adaptations, and evaluations contribute new knowledge and experience in adapting and applying set-prediction methods to complete 3D configurations under geometric and physical constraints.

C2. Comparison Experiment Settings — Reviewers mDr4, 9nrg, and q1my.

We thank Reviewers mDr4, 9nrg, and q1my for raising questions about the fairness and consistency of the experimental comparisons. We agree that these issues warrant careful attention, since specifying and controlling the search budgets, hardware access, and reference-catalogue construction would strengthen the fairness and reproducibility of the comparisons and make the sources of the measured gains easier to assess. We describe the updated settings below.

Test reference catalogue. Only the reference catalogue for the test configurations has been extended from 80 to 120 searches per initial configuration, by adding 40 searches while keeping the test configurations unchanged. The reference data used for training and model selection remain based on the original 80 searches per configuration. The expanded test catalogue provides a more comprehensive reference for evaluating event recovery.

Proposal budget and Dimer cap. For the controlled quality comparison, we align ASPIRE and Random Dimer to 80 proposals/search starts per initial configuration and a cap of 300 Dimer translation steps, and reevaluate the archived search results under these shared limits. Each point on the ASPIRE recovery curve is obtained by applying a judge threshold; lowering the threshold admits more proposals, and the plotted range is limited to operating points with at most 80 proposals submitted to refinement per configuration. The Random Dimer curve analogously varies the number of random-start searches up to 80. Thus, 80 is the shared maximum physical-search budget, rather than a separately sampled ASPIRE subset.

GPU access. The updated GPU comparison provides both ASPIRE and the non-ML baselines with one A100 GPU and K = 4 concurrent GPU-resident Dimer search slots, retaining the original CPU allocation. GPU execution includes EAM energy and force evaluations, Dimer rotation, effective-force projection, and atomic-coordinate translation updates. Coordinates, forces, and search directions remain on the GPU throughout the search, reducing repeated transfers of atomic arrays between CPU and GPU. Searches can use an available GPU slot or fall back to CPU when all four slots are occupied. ASPIRE uses the GPU for both model inference and subsequent Dimer refinement, so GPU-assisted physical search is available to both ML and non-ML methods.

The updated comparison settings and reference-catalogue protocol are now stated in Sections 5.1--5.3 and Appendix H of the revised manuscript. The measured GPU wall-time comparison is reported in Section 5.5 and Figure 5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.