acceptodds
Under review as a conference paper at ICLR 2027

REPAIRING CORRUPTED CRYSTAL STRUCTURES WITH GEOMETRY-GUIDED MULTI-BRANCH RELAXATION

Abstract

Crystal repair complements generation: recovering the intended structure from a corrupted input at fixed composition, using the geometry that survives corruption. Every atom is still there; which positions may be trusted is not known in advance. A shared collision projection reads short-range separations in the input and proposes where to start. Five CHGNet refinement routes relax that start under different degrees of freedom and tolerances. A frozen ExtraTrees selector reads observable geometric changes and candidate-level energy, force, and stress, then emits one structure. Clean structures and corruption labels stay hidden at inference and inform matching and subgroup analysis only after every prediction is fixed. On 2,520 corrupted test inputs, multi-branch repair matches 76.35% (1,924/2,520), against 44.01% (1,109/2,520) for relaxation-only and 29.96% (755/2,520) for the unchanged input, the share already inside the matcher tolerance; the 32.34-point gain over relaxation-only is bought with five relaxations against one. Match is highest for atom overlaps and lattice distortions, and lower for coordinate noise and coordination defects.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
artificial intelligence for science, crystal structure refinement, materials discovery, feedback-conditioned editing, reflection, structured repair plans

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_sxSiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The method repairs corrupted crystals using collision correction, five CHGNet relaxation routes, and a learned selector. Match rises from 44.01% for relaxation-only to 76.35%.

Critical strengths
  1. Collision correction alone raises Match to 65.08% at the same single-route budget (Table 7). This is a substantial improvement.

  2. The paper clearly reports failure cases, matching tolerances, and limitations.

Critical weaknesses
  1. Train/test separation by source crystal is unverified. The five-route method also uses more compute than the baseline, leaving generalization and selection gains uncertain. This does not prove leakage.

  2. The stress_weight of 0.1 is about 16 times the standard unit conversion. This may affect cell relaxation and the comparison with fixed-cell routes. I lean toward rejection until its effect is checked.

Questions for authors
  1. Can you verify disjoint source crystals and compare against a baseline with the same compute budget?

  2. Is stress_weight=0.1 intentional? Do the main results hold with the standard conversion?

Reviewer_vL1SOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies crystal structure repair with fixed atomic species and counts, without knowing which sites are corrupted. Geometric rules first correct atomic overlaps, followed by CHGNet-based relaxation under five configurations with different constraints and convergence settings. An ExtraTrees classifier selects the final structure using hand-engineered geometric and physical features. On four types of synthetic corruption derived from MP-20, the method reports a matching rate of 76.35%, compared with 44.01% for a single CHGNet relaxation.

Critical strengths

None

Critical weaknesses
  1. The practical motivation for the task is not sufficiently established. The paper does not explain how the chosen synthetic corruptions and recovery objective correspond to concrete needs in AI for science, such as experimental structure refinement or improvement of generated candidates. Without this connection, it is difficult to assess the scientific significance of the reported recovery gains.
  2. The methodological contribution appears limited. The approach combines geometric heuristics, existing CHGNet relaxation routines, and an ExtraTrees selector with hand-engineered features. The paper provides limited insight into what this combination contributes to learning or optimization beyond the specific repair setting. A clearer account of the new methodological or empirical understanding would be needed to support the overall contribution.
Questions for authors
  1. Which concrete application motivates this repair task, and how do the synthetic corruptions and recovery objective reflect the errors encountered in that application?
  2. What new methodological or empirical insight does the approach provide beyond combining geometric preprocessing, multiple relaxation configurations, and conventional candidate ranking?
Reviewer_8Np1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a 3 step pipeline for crystal structure repair, with an overlap correction heuristic followed by five independent CHGNet refinement routes and then the candidate selector which is learnt, which picks the best candidate by looking at the energy and geometric features, and it outperforms usual energy relaxation.

Critical strengths
  • Outperforms standard relaxation under the reported setup.
  • The paper is clear and easy to follow.
Critical weaknesses
  1. Practical motivation and synthetic corruptions. The practical motivation for the problem isn't too clear. As in, what real world situation is the paper trying to tackle? I can think of diffusion models proposing imperfect structures, or experimental data that has noise, etc., but the paper is not clear as to what the situation it applies to. And taking actual crystal structures and corrupting them artificially is a different problem than showing that this is applicable in any real situation where we might encounter such noisy data. Basically, applicability on a synthetic dataset is different from any real-world use case.

    Here it should also be noted that one of the corruption steps in their artificial data generation is pushing the atoms together, which their handwritten heuristic then undoes by pushing them apart, so success there is somewhat self-fulfilling; and it could be a contributing factor of why we see a large fraction of the additional matches (over relaxation-only) coming from the overlap family.

  2. Additional complexity and evaluation concerns. Additionally, from the ablations, we can also see that most of the heavy lifting is being done by the handwritten heuristic, the first step of this three-step process, followed by just a single relaxation route, at the same max 200-step relaxation budget. And without similar budget matched baselines, the rationale for the additional complexity and the extra computation cost (5 parallel runs followed by a selector) seems largely unmotivated.

    Crucially, the authors themselves say that the training and test set disjointness is unverified, so I'm not sure if we can actually draw any generalizable real world conclusions from this work as they don't rule out dataset contamination. The RMSD is also only reported for successful matches.

Questions for authors
  1. What real world situation is the paper trying to tackle? Is it intended for diffusion models proposing imperfect structures, experimental data that has noise, or another situation?
  2. How does success on artificially corrupted crystal structures establish applicability to a real-world use case?
  3. What justifies the additional complexity and computation of five runs followed by a selector, given the ablation results with a single relaxation route and the lack of similar budget matched baselines?
  4. Could the alignment between pushing atoms together in the artificial corruption and pushing them apart in the heuristic explain the large fraction of additional matches coming from the overlap family?
  5. Can the authors verify training and test set disjointness and rule out dataset contamination?
  6. Why is RMSD only reported for successful matches?
Reviewer_hoEVOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper focus on repairing crystal structure from a corrupted structure while given composition and atom count. The pipeline is: first using geometry rules to avoid atom collisions, and then running five relaxation branches, and finally using an ExtraTrees classifier trained to find out the target from candidates. The experiments shows 76.35% match rate compared with 44.01% for relaxation-only.

Critical strengths
  1. The problem is practically motivate and the method is straightforward which uses explicit geometric rule and constrained relaxation to solve it.

  2. The experient results show that it is useful with match rate from 44.01% to 65.08%.

  3. The paper honestly reports the influence of every component with ablations and limitations.

Critical weaknesses
  1. The method lacks novelty and the comparison in this paper isn't really fair. I think a single relaxation should run the same total number of steps as all five relaxations combined. Also, I want to see what the upper bound match rate would be if the selector were theoretically optimal and always picked the right candidate.

  2. The selector is only trained on 200 inputs, which is too few, so I doubt how reliable it is.

  3. The exact same crystal might appear in both the training set and the test set, just with different-looking corruptions. If that's the case, there is actually data leakage.

  4. Section 3.4 says class weights only rescale probabilities and don't change which candidate comes first. However, when training ExtraTrees, changing the weights changes how the trees split, so the top pick can change.

  5. Most of the performance gains actually come from the synthetic atom-overlap cases, so the experiments don't really prove that the main contributions.

Questions for authors
  1. Can you make sure the same material doesn't appear in both the training set and the test set?

  2. If you change the stress conversion back to the standard value, how much will the match rate change?

  3. I want to see what the upper bound match rate would be if the selector were theoretically optimal and always picked the right candidate.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.