acceptodds
Withdrawn from ICLR 2027

SGCR: Structure-Gradient Coupled Replay for Continual Crystal Property Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Crystal structures and property distributions can evolve continuously as new material families and chemical spaces are explored, making continual learning essential for crystal property prediction. We study this problem under limited historical memory and propose \modelname, a structure-gradient coupled framework that jointly considers structural relevance and optimization interference. \modelname first constructs a frozen structural reference space to index a domain-balanced random memory of historical crystals. For each incoming batch, it retrieves historical samples according to both their structural relevance and their conflict with the current prediction objective, enabling more informative use of limited memory. We further introduce a calibrated structural-novelty signal that adaptively modulates gradient correction when current and historical objectives disagree, allowing the model to balance adaptation to newly arriving crystal distributions with retention of previously learned knowledge. Extensive experiments across multiple datasets demonstrate that \modelname consistently improves predictive performance and effectively alleviates catastrophic forgetting.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Crystal Property Prediction, Continual Learning, AI for science

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uMdXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SGCR for continual crystal formation-energy prediction with limited historical memory. It combines structure-guided replay with gradient-conflict correction and evaluates the approach on three datasets and three backbones.

Critical strengths

The evaluation covers complementary distribution shifts and multiple architectures. The method combines targeted replay with historical coverage while preserving a fixed inference architecture. Reported accuracy improvements and novelty-control experiments provide promising initial evidence.

Critical weaknesses
  1. Unclear practical motivation. The restrictive 500-record memory budget is not sufficiently justified. It is unclear whether the method remains advantageous when simpler replay methods can retain more historical data.

  2. Insufficient explanation of the mechanism. Although the structural features are learned, the connection between structural similarity, novelty, and forgetting is not convincingly established. The paper needs stronger evidence that structure provides useful information beyond gradient conflict.

  3. Limited experimental scope. All experiments concern formation energy and short sequences of domains. This provides limited support for the broader applicability to continual crystal property prediction.

  4. Limited methodological novelty. The method appears to be an incremental combination of existing replay and gradient-correction ideas. The benefit of this complexity over simpler alternatives is not sufficiently demonstrated.

  5. Concerns about experimental reliability. Missing implementation details and uncertainty estimates make the results difficult to assess and reproduce. Figures 3 and 8 are also nearly identical despite reporting different backbones, which requires clarification.

Questions for authors
  1. What practical constraint motivates the small memory budget, and would larger-buffer random replay achieve comparable results?

  2. What evidence shows that structural similarity or novelty predicts forgetting beyond the gradient-conflict signal?

  3. Is there evidence that the method generalizes beyond formation energy or remains effective over longer domain sequences?

  4. Does the proposed correction outperform a well-tuned constant correction or a simpler adjustment of the replay weight?

  5. Can the authors explain Figures 3 and 8 and clarify the experimental configurations and variability across runs?

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

The paper proposes SGCR for continual crystal formation-energy prediction with limited historical memory. It combines a frozen structural representation, structure- and conflict-guided replay retrieval, and novelty-dependent gradient correction. Experiments cover three datasets and three GNN backbones, with comparisons against continual-learning baselines and analyses of forgetting, component contributions, and computational cost.

Critical strengths
  1. The evaluation covers three distinct sources of distribution shift, including crystal-system changes, database expansion, and successive chemical substitutions. This makes the setting relevant to how materials datasets develop.

  2. The results show promising improvements across several backbones. In particular, on WBM with CGCNN, SGCR reduces MAE from the strongest baseline’s 0.1185 to 0.0946, a substantial improvement under the stated memory budget.

Critical weaknesses
  1. The benefit of novelty-dependent correction is not established against a well-chosen constant. In Figure 4, fixed correction already outperforms full SGCR on WBM with ALIGNN. The fixed control uses only ρ=1\rho=1, although Figure 6 shows that correction strength strongly affects performance. The reverse and permuted controls are useful, but they do not establish whether novelty conditioning improves over a constant selected under the same validation protocol.

  2. For WBM with CGCNN, Table 1 reports 0.0946±0.01250.0946\pm0.0125, while the full model in Figures 4 and 5 is approximately 0.1081. This could reflect a single run versus a multi-run average, but the distinction is not explained. The ablations also lack uncertainty estimates, leaving the small component gains difficult to assess so that the ablations cannot be clearly connected to the main results.

Questions for authors
  1. Does novelty-dependent correction outperform a constant coefficient tuned with the same validation budget, using matched random seeds?

  2. Which seeds and configurations produced Figures 4–6, and how do they relate to Table 1? Could the authors report the variation in the paired differences for the key ablations.

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

This paper introduces SGCR: structure-gradient coupled replay for continual representation learning for crystal property prediction. The proposed framework considers structure relevance and conflict scores and bake them into the correction gradients when switching the crystal domains, therefore achieving state-of-the-art results on continual crystal learning tasks across different datasets (MP, JARVIS-DFT,WBM). The authors also did ablations on different settings of the gradient-correction framework and proved optimal settings.

Critical strengths

The quantitative results look very compelling, i.e. the good performances across three distinct tasks and comparisons with different baseline approaches, including both general and graph-based CL make the SGCR results very impressive. The hyperparameter analysis and ablations study further consolidate the framework of SGCR.

Critical weaknesses

My main concerns include the following:

  1. The presentation and the soundness of the SGCR methods: it might be only me (curious about other reviewers' opinions), i found the methodology part (section 4) very hard to follow, with specific questions detailed in the questions section. Also, the methodology figure (figure 2) is hard to interpret, and it has only one line in the captions.

  2. The motivations and significance of the proposed approach to the community: according to my familiarity to the field, the continual learning is not a big concern for the crystal representation learning, and in fact the related works and the adopted baselines of continual learning methods are traced back to continual learning community, instead of the crystal learning one. Applying them and comparing with the newly proposed approach here makes sense, but still as a researcher from the crystal learning field, i feel the significance of the research outcome in this work remains to be evaluated. I am happy to adjust my view based on others' opinions though.

Questions for authors

I may not fully capture the methodology part, so i set my confidence score to 2, with the following specific questions:

  1. for Figure 1 (c), how are the numbers presented? shouldn't the best approach give the lowest mae, why it's placed at 100? adding (c) and its explanation into your captions could be helpful.

  2. How each component of figure 2 relate to each other? please complete the captions of figure 2 to make it self-contained.

  3. line 144, as for the domain splits, why did the author choose three different classification in the three datasets (section 5.1)? what's the rationale behind lattices types, chronological, and different substitution types? they are quite distinct from each other and appear a bit synthetic for me. in practice people don't do this, right? this is related to my second point in the critical weaknesses part.

  4. as for section 4,

  • what's gradient conflicts?
  • line 173-179, what's learnable, what's fixed (the encoder backbone is fixed?)? why "hθ(X) changes as the predictor learns new domains." while fixing the parameters in the graph encoder?
  • according to 1, is there only one linear layer after the graph encoder?
  • how is the local priority pi applied in section 4.4?
  • how did the authors derive (8) from (7)? there is a gap for me. also what's the big A in the if conditions
  1. line 466, how did the authors conclude that "JARVIS favors relatively weak correction, whereas WBM achieves its lowest error at an intermediate strength."? i see the lowest mae occurs in intermediate alpha for jarvis while low alpha for WBM.

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

Related papers

Discussion (0)

Sign in to comment.