acceptodds
Under review as a conference paper at ICLR 2027

Physics-Routed MoE with Crystallographic Priors for Long-Tailed Space-Group Recognition

Abstract

Determining a crystal's space group from a powder X-ray diffraction (PXRD) pattern is a 230-way fine-grained classification problem characterized by severe class imbalance and substantial similarity among diffraction patterns. Many rare space groups have fewer than 20 training examples and are easily confused with frequent groups in the same crystal system. We present Phys-MoE, a Mixture-of-Experts classifier that incorporates crystallographic priors into routing and representation learning. Phys-MoE consists of three components: a hierarchical router that masks space groups inconsistent with the predicted crystal system, a rule-biased gating mechanism that encourages expert specialization across symmetry blocks, and a rule-semantic contrastive objective based on the Jaccard similarity between symmetry-operation sets. Experiments on CCDC, CoREMOF, and an inorganic distribution-shift benchmark demonstrate the effectiveness and generalizability of Phys-MoE. On CCDC, Phys-MoE achieves 81.70% overall Top-1 accuracy and 26.19% accuracy for space groups with fewer than 20 training examples, establishing a new state of the art for long-tailed space-group recognition from PXRD patterns. Ablation studies demonstrate that the three crystallographic components work jointly to improve recognition, while error and oracle analyses identify crystal-system prediction as a remaining bottleneck for rare classes. These results show that incorporating crystallographic structure into expert routing and representation learning provides an effective approach to accurate and reliable space-group recognition under long-tailed data distributions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
powder X-ray diffraction, space-group recognition, long-tailed recognition, mixture of experts, crystallographic priors, hierarchical routing, contrastive learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MJKROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper uses crystallographic priors for long-tailed space-group recognition. The results look promising, and the ablations support the main design choices. I have a few questions below.

Critical strengths

see Questions

Critical weaknesses

see Questions

Questions for authors
  1. Could you compare RSCL with a same-class-only contrastive loss and a similarity based only on crystal systems? Keep the backbone and other losses the same, so we can see what the symmetry-operation relations actually add.

  2. The expert allocation is based on the CCDC distribution. Pls try initializing B to zero, with the same eight experts and losses. Does the chosen allocation help, or can the gate learn equally well without it?

  3. Table 17 has just 42 CCDC extreme-tail test samples from 42 classes. Is this also the set used for the main results? If yes, pls report the mean and std over the five training seeds for this bucket. With only one sample per class, resampling within each class won’t capture the uncertainty in the test set.

  4. How are the symmetry-operation sets standardized before computing their overlap? The setting and origin choice matter here. Also, I couldn’t find the value of δ or whether the features in Eq. (11) are L2-normalized.

  5. I found the two evaluation settings a bit confusing. Table 1 trains a separate model on each dataset, while Table 12 tests frozen CCDC-trained models.

2 comments

Thank you for suggesting the contrastive and initialization controls and for pointing out where the evaluation protocols needed clarification.

Q1. Same-class, crystal-system, and operation-set contrastive objectives

We evaluated all three configurations on the CCDC test set with the same backbone, HR, RGE, other losses, data splits, training seeds, and folds. The same-class control uses positives from the same space group. The crystal-system control also uses positives from other space groups in the same crystal system. Both controls assign a weight of 1 to each non-self term in the denominator. Full RSCL uses operation-set similarity to select cross-class positives and weights the denominator terms by 1−S1-S.

Contrastive objectiveOverall Top-1Tail Top-1Extreme-Tail Top-1
Same-class positives only79.35%41.40%23.72%
Crystal-system relations only79.82%42.15%23.51%
Full operation-set RSCL81.70%46.50%26.19%

Full RSCL improves Extreme-Tail Top-1 by 2.47 percentage points over the same-class control and 2.68 points over the crystal-system control.

Q2. Zero initialization of BB

We varied the initial value of BB while keeping eight experts, the losses, data splits, training seeds, and folds fixed. In both variants, BB is updated during training. The table reports five-seed means on the CCDC test set.

Initialization of BBOverall Top-1Tail Top-1Extreme-Tail Top-1
All zeros80.78%44.62%23.81%
Rule-based initialization (±0.5)81.70%46.50%26.19%

Rule-based initialization improves Overall, Tail, and Extreme-Tail Top-1 by 0.92, 1.88, and 2.38 percentage points, respectively.

Q3. Extreme-Tail test examples and variation across seeds

The Phys-MoE row in submitted Table 2 used validation scores. The updated test evaluation uses the same 42 CCDC Extreme-Tail examples as Table 17. These examples cover 37 space groups, so we will correct the Classes entry from 42 to 37 in both CCDC rows. The CCDC benchmark contains examples for 225 of the 230 space groups.

The original per-seed validation records were unavailable. For the test-set update, we retrained Phys-MoE with five seeds and evaluated each run on these 42 examples.

Training seedClass-balanced Top-1
125.68%
225.90%
326.13%
426.35%
526.80%
Mean26.17%
Sample SD0.43 percentage points

Within-group resampling gives a fixed contribution from the 35 space groups with one example each. The other two groups have three and four examples, respectively, and their draws can vary across replicates.

Q4. Operation-set standardization, threshold, and feature normalization

We use the fractional-coordinate operations listed for each space group in Section 2.3 of International Tables for Crystallography, Volume A (2016), in the setting and origin of the corresponding entry. We compare the tabulated representatives by matching their linear and fractional translation parts. The default threshold is δ=0.65\delta=0.65. We L2-normalize the projected features before computing the inner products in Equation (11). IUCr space-group tables

Q5. Evaluation protocols in Tables 1 and 12

For Table 1, we train and evaluate a separate model on each dataset. For Table 12, we train each mask configuration on CCDC and freeze its parameters before evaluating the inference routes on CCDC, CoREMOF, and Inorganic. The Extreme-Tail bucket in Table 12 is defined by the CCDC training frequencies.

The additional experiments are sufficient and address my concerns.

Reviewer_2wG3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Phys-MoE, a physics-informed mixture-of-experts classifier for identifying space groups from powder X-ray diffraction (PXRD) patterns. The primary motivation is to improve classification performance under severe class imbalance, particularly for space groups with very few training examples. Their numerical results support improved classification accuracy for rare classes.

Critical strengths
  1. A particular strength is the use of known crystallographic relationships to guide classification rather than treating the 230 space groups as unrelated categories. The proposed framework achieves encouraging improvements over existing approaches, especially for rare space groups.

  2. The experimental evaluation is generally thorough. In addition to comparisons across three datasets, the authors provide component ablations, incremental model construction, routing analyses, error diagnostics, and robustness experiments involving controlled perturbations. These analyses help clarify the contributions of different components.

Critical weaknesses

No major weaknesses; my suggestions generally aim to strengthen the methodology and exposition.

Questions for authors
  1. The proposed framework uses eight expert classifiers, with a fixed initial allocation across Tail, Transition, and Head experts. Appendix C clarifies the allocation and the initialization of the expert-bias matrix. However, the rationale for choosing 8 experts is not fully established. Since the experts share the same encoder and have identical linear classification architectures, with several also using the same loss function, it would be useful to understand how much additional performance is gained by increasing the number of experts. I recommend a sensitivity analysis comparing a few choices for the number of experts (maybe 2,4,8 and 12). This would clarify whether the performance improvements arise from meaningful expert specialization or primarily from the increased capacity of the mixture model.

  2. The rule-semantic contrastive learning component uses a threshold δ\delta to determine whether observations from different space groups are treated as positive pairs. This threshold seems potentially important, particularly for rare space groups, where borrowing information from related groups is a central motivation for the method. I recommend reporting the sensitivity of overall and Tail/Extreme-Tail accuracy to a few reasonable choices of δ\delta. It would also be helpful to summarize the distribution of the off-diagonal entries of the similarity matrix and indicate how the number of cross-class positive pairs changes with the threshold. These results would help assess the robustness of the method and provide a clearer justification for the selected threshold.

  3. I would appreciate some additional explanation of the formulation in Equation (11). The denominator uses weights ωij=1−S(yi,yj)\omega_{ij}=1-S(y_i,y_j). For observations belonging to the same space group, the similarity is one, so their denominator weights become zero, even though these observations are included as positive pairs in the numerator. Consequently, the ratio inside the logarithm is not necessarily bounded above by 1. If this is intentional, some clarification would help interpret this quantity.

  4. The improvements for Extreme-Tail space groups are particularly encouraging and represent an important contribution of the paper. Given the very small number of training examples for these groups, I suggest reporting the number of space groups and test observations in each frequency bucket. The authors appropriately report bootstrap confidence intervals for the comparisons. A brief clarification of the resampling procedure, particularly how class-balanced accuracy and variation across training seeds are handled, would be helpful.

3 comments

Thank you for suggesting the expert-count and threshold comparisons and for asking us to clarify the test-set counts and confidence intervals.

Q1. Expert count and model capacity

We compared 2, 4, 8, and 12 experts on the CCDC test set, using the same backbone, data splits, training seeds, and folds.

ExpertsTotal parametersOverall Top-1Tail Top-1Extreme-Tail Top-1
267.41 M80.88%43.75%23.81%
467.52 M81.31%45.12%25.24%
867.76 M81.70%46.50%26.19%
1268.00 M81.57%45.95%25.71%

Eight experts achieved the highest Top-1 on all three metrics and used 0.24 M fewer parameters than twelve. Expert count and classifier-head capacity vary together in this comparison.

Q2. Threshold sensitivity and cross-class positives

The submitted model uses δ=0.65\delta=0.65. For the threshold comparison, we held the data splits, seeds, and folds fixed and evaluated each setting on the CCDC test set. Samples from distinct space groups qualify as positives when S>δS>\delta.

Threshold δ\deltaUnordered space-group pairs with S>δS>\deltaOverall Top-1Tail Top-1Extreme-Tail Top-1
0.602,15682.00%47.20%28.57%
0.651,74281.70%46.50%26.19%
0.701,42881.20%45.30%23.81%

Of the 26,335 unordered pairs of distinct space groups, 91.81% have S≤0.60S\leq 0.60, 1.57% fall in (0.60,0.65](0.60,0.65], 1.19% in (0.65,0.70](0.65,0.70], 2.10% in (0.70,0.75](0.70,0.75], and 3.33% have S>0.75S>0.75. At δ=0.75\delta=0.75, 876 pairs qualify. Within a batch, a qualifying group pair (g,h)(g,h) contributes ngnhn_g n_h unordered positive example pairs, where ngn_g and nhn_h are the two group counts in that batch. Among the thresholds tested for accuracy, 0.60 achieved the highest scores on all three metrics.

Q3. Denominator in Equation (11)

Yes, the weighting is intentional. The factor 1−S1-S reduces the contribution of pairs with similar symmetry operations to the denominator. A same-space-group positive has S=1S=1, so its score enters the numerator and has zero weight in the denominator. The ratio divides the positive-pair score by the weighted sum of pair scores plus ϵ\epsilon. When the ratio exceeds 1, the corresponding negative-log term is negative.

Q4. Frequency buckets, training seeds, and confidence intervals

The baseline rows in submitted Table 2 were evaluated on the CCDC test set. The Phys-MoE row used validation scores, which affected the reported gains and confidence intervals. The table below gives the test-set results from the new benchmark runs. Gains and intervals are relative to the best test-set baseline in each bucket. The four ablation tables report the full-model scores from their corresponding CCDC test-set experiments.

CCDC frequency bucketSpace groupsTest examplesPhys-MoE Top-1Best baseline Top-1Gain (pp)95% CI (pp)
Many-shot322,23281.05%80.40%+0.65[−1.68, 3.18]
Medium-shot7664766.44%61.01%+5.43[1.02, 10.04]
Tail8029246.40%39.13%+7.27[3.21, 11.53]
Extreme-Tail374226.17%10.00%+16.17[13.12, 18.95]

Class-balanced Top-1 gives each space group equal weight within its bucket. In each bootstrap replicate, we sample examples with replacement within each true space group, separately for each seed. Phys-MoE and the baseline use the same sampled indices within a seed. We average the paired gains across the five seeds and take the 2.5th and 97.5th percentiles of the replicate gains. Across the five new training runs, the sample standard deviation of Extreme-Tail Top-1 was 0.43 percentage points.

Thank you for the additional experiments and explanations. These address most of my first three questions. However, I am concerned that Table 2 compares validation results for Phys-MoE with test results for the baselines. Have the other comparisons been checked for this issue?

I also have some concerns about the Extreme-Tail confidence interval. With 42 test examples across 37 groups, at least 32 groups must have only one example. Resampling within these groups will always return the same observation, so the bootstrap misses their test-sampling uncertainty. The small variation across training seeds does not address this problem. I would appreciate a reconsideration of the interval or a clear discussion of this limitation.

Thank you for following up. We checked the other comparisons and confirm that they use test-set results. In the submitted Table 2, the Phys-MoE, Gain, and CI rows used validation results, while the baseline rows used test results. We provided the corrected test-set results in our first response: the Extreme-Tail accuracy is 26.17%, with a gain of 16.17 percentage points over the strongest baseline.

To assess test-sampling uncertainty for unseen structures, we perform n-fold evaluation within each seed and regenerate the data partitions across random seeds. Different available structures can therefore enter the test set across evaluations. All methods share the same partition within each seed and fold.

These measures remain constrained by the available structures. The CCDC source dataset contains 225 space groups and over one million samples, but some groups have very few samples. For example, P4bm has only six. Repeated partitioning across seeds and folds inevitably reuses this limited pool. For singleton groups within a given split, within-class bootstrap resampling always returns the same observation. This step therefore cannot capture performance variation on other unseen structures from those groups. The reported [13.12, 18.95] percentage-point paired-bootstrap interval for the 42-example evaluation is conditional on the available samples and realized partitions. We acknowledge this limitation and will state it in the revised manuscript.

To increase test coverage, we also evaluated Phys-MoE on 92 examples across the same 37 Extreme-Tail space groups. The class-balanced Top-1 accuracy is 26.24%, compared with 26.17% in the 42-example evaluation. The reported 95% bootstrap confidence interval for the gain over XRDecoupler is [11.89, 20.67] percentage points. In this expanded set, 33 groups have at least two test examples. The remaining four groups have only one test example each after allocating training data because so few structures are available. These four groups retain the singleton resampling limitation described above.

Reviewer_E3wrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies long-tailed space-group classification from simulated powder X-ray diffraction (PXRD) patterns. The paper evaluates the method on CCDC, CoREMOF, and Inorganic datasets, reporting 81.70% overall Top-1 accuracy and 26.19% class-balanced Extreme-Tail accuracy on CCDC. The central claim is that incorporating crystallographic label relations into both prediction and representation learning improves recognition under severe class imbalance.

Critical strengths

The method incorporates task-relevant structure through a coherent learning design. The space-group-to-crystal-system mapping provides a direct and meaningful prior for this classification problem. Using this structure for conditional prediction and expert weighting is well motivated. The operation-level relationships offer a potentially useful mechanism for sharing information with poorly represented classes.

Critical weaknesses

The most distinctive crystallographic component is incompletely specified, and the current evidence does not sufficiently isolate the value of its particular semantics. The symmetry-operation similarity matrix determines both the positive pairs and denominator weights in RSCL. Section 3 refers readers to Appendix C for the operation-set representation, but Appendix C.1 largely repeats the Jaccard definition. It does not specify the finite operation representatives, rotation/translation encoding, treatment of equivalent translations, settings and origins, or operation-equality rule. Finally, the existing ablations establish a benefit from the complete modules, but do not isolate the information provided by the particular operation-set similarity. Removing RSCL removes the entire contrastive objective.

Questions for authors
  1. Please specify how the finite operation sets are constructed, including settings/origins, rotation and translation representation.

  2. Are the CCDC Top-5 result in Table 1 and the crystal-system accuracy in Table 4 computed from the same checkpoints, seeds, and test examples? Please specify whether Top-5 uses masked or unmasked scores and how it is computed when the selected crystal system contains fewer than five space groups

1 comment

Thank you for asking how we construct the symmetry-operation sets and how we aggregate the CCDC metrics.

Weakness. Contribution of RSCL

Table 3 removes the full RSCL objective. To examine the role of operation-set similarity, we ran two additional controls on the CCDC test set while holding the backbone, HR, RGE, other losses, data splits, training seeds, and folds fixed. Extreme-Tail Top-1 was 26.19% with full RSCL, 23.72% with same-class positives only, and 23.51% with crystal-system relations only. Full RSCL improved on the two controls by 2.47 and 2.68 percentage points, respectively. The comparison tests cross-class positive selection and denominator weighting together.

Q1. Construction of the symmetry-operation sets

For each space group gg, we construct the finite set Og\mathcal{O}_g from the operation representatives listed in Section 2.3 of International Tables for Crystallography, Volume A (Aroyo, 2016). We use the setting and origin of the corresponding entry. Each fractional-coordinate triplet specifies an affine map x↦Rx+tx\mapsto Rx+t, where RR is the linear part and tt is the fractional translation. We compare the listed representatives by matching both RR and tt in their tabulated form. As defined in Section 3, the similarity is

{|\mathcal{O}_g\cup\mathcal{O}_{g'}|}.$$ [IUCr space-group tables](https://it.iucr.org/Ac/contents/) · [Explanation of the tabulated operations](https://it.iucr.org/Ac/explanationAc.pdf) **Q2. Space-group Top-5 in Table 1 and crystal-system Top-1 in Table 4** Both tables use the same CCDC test examples. Table 1 averages masked space-group Top-5 over five training seeds. Table 4 uses one checkpoint from the Table 1 runs across all folds. Its mean fold-level crystal-system Top-1 is 91.89%. For a matched comparison, we computed both metrics using the same checkpoint and test examples within each fold of each seed. We rank space-group scores after masking, within the predicted crystal system. If that system contains fewer than five space groups, we use all of its candidates. | Training seed | Masked space-group Top-5 (%) | Crystal-system Top-1 (%) | |---|---:|---:| | 1 | 91.22 | 92.50 | | 2 | 93.06 | 94.21 | | 3 | 93.93 | 94.99 | | 4 | 94.80 | 95.80 | | 5 | 96.64 | 97.29 | | **Mean ± sample SD** | **93.93 ± 2.01** | **94.96 ± 1.79** | Each seed row averages the five folds equally. The last row gives the mean and sample standard deviation across the five seed-level scores. In this matched comparison, masked space-group Top-5 is below crystal-system Top-1 for every seed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.