acceptodds
Under review as a conference paper at ICLR 2027

SuperFlow: Two-Stage Joint Flow Matching for Superconductor Doping

Abstract

Finding effective doping strategies is a difficult but valuable step in superconductor development: both the substitution choice and its doping rate matter, and an unfavorable change can lower the superconducting critical temperature . Locating an optimum experimentally requires synthesizing and characterizing multiple compositions along a substitution trajectory, with substantial time and resource costs. We introduce a benchmark for parent-conditioned optimal-doping prediction. The benchmark combines literature-labeled ICSD structures with ICSD and Materials Project data from 3DSC. We use a large language model to extract experimental values from source papers and correct selected 3DSC labels based on the literature. The benchmark measures doping-rate error and whether a proposed change moves toward an observed optimum. We further propose SuperFlow, a two-stage joint flow-matching framework that predicts doping rate together with crystal structure for a specified substitution axis. A first flow generates a candidate crystal. A goal head reads this result and guides a second joint refinement. SuperFlow achieves the best moving-parent and large-move doping-rate errors and direction accuracies among the evaluated methods on both datasets. Moving-parent doping-rate MAE is 0.11633 on ICSD and 0.08847 on MP. Selected case studies show that predicted doping rate changes are accompanied by local structural changes toward dataset targets. The benchmark and method provide a basis for prioritizing experiments and, with further validation, supporting industrial searches for optimal doping.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Superconductors, Optimal Doping, Disorder Flow Matching, Guided Generation

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 4 reviews, above 27% of the others.
Reviewer_FJUnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies parent-conditioned optimal doping for superconductors. Given a parent crystal and a host–dopant substitution axis, the goal is to predict the doping rate that reaches the highest recorded Tc in that substitution series, together with the doped crystal structure. The authors build a benchmark from ICSD and Materials Project data in 3DSC. They use GPT-5.4 to extract Tc labels from the original papers and correct some 3DSC labels. They then propose SuperFlow, a two-stage flow-matching model. Stage I generates a candidate crystal from the parent. In Stage II, a goal head predicts whether to move, which direction, and how far, and a second integration pulls the candidate toward that goal. A finite-step bound is given for the second pass. On both benchmarks, SuperFlow gets the lowest doping-rate error and the best direction accuracy on moving and large-move parents.

Critical strengths
  1. The task is useful and clearly defined. Predicting how much to dope from a known parent is closer to real experimental decisions than plain Tc regression. The evaluation separates parents that need to move from those already at the optimum, and it also checks movement direction, which makes sense.
  2. The ablations are consistent. Across both datasets, the second pass improves over One-round SuperFlow on moving and large-move parents, so the two-stage design does seem to help.
Critical weaknesses
  1. The baselines are weak. Direct-x is close to random and seems under-trained. Simple baselines like "keep the parent doping rate" and "training-set average for the same host–dopant pair" are missing.
  2. The LLM-extracted labels are only shown through ten examples, with no quantitative accuracy check.
  3. The theory adds little. Prop. 5.1 only shows that the second pass gets close to the predicted goal, not to the true optimum.
  4. All results use a single seed, and the MP test set is small. The gains may not be significant.
  5. The paper is hard to follow. The notation is heavy, and the main text has a lot of implementation detail but little intuition. Readers outside materials science will struggle to understand the method.
Questions for authors
  1. What are the results of (a) always predicting the parent doping rate x0 and (b) predicting the training-set mean optimal doping rate for the same host–dopant pair?
  2. How does Direct-x perform with a training budget comparable to SuperFlow's?
  3. Can you report the extraction accuracy of the LLM labels on a random sample checked by hand?
  4. All results come from a single seed, and the MP test set is small (81 trajectories, 86 large moves). Can you report variance over several seeds?
  5. On ICSD, the full model has a higher all-parent error than One-round SuperFlow (0.105 vs. 0.089). Does the model move parents that are already optimal? How sensitive is this to the movement threshold?
Reviewer_DRHUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces SuperFlow, a two-stage joint flow-matching framework for optimal-doping prediction. Given a parent crystal and a specified substitution condition, the model predicts a doping rate and corresponding crystal structure. A structure-aware correction module is introduced to guide a second generation stage, improving doping-rate prediction.

Critical strengths
  • Doping optimization is an important problem. The proposed benchmark provides a useful setting for evaluating predictions.
  • The presentation is clear, and the paper is easy to follow.
Critical weaknesses

See the questions.

Questions for authors
  • How are available substitution rules and substitution sites determined?
  • Are the generated structures physically reasonable, if not is there any method to show structure validity and Tc related property performance computationally?
  • The performance of Tc predictor is not reported.
Reviewer_pNLtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a benchmark and a two-stage flow-matching framework, SuperFlow, for parent-conditioned superconductor doping optimization. Given an existing crystal and a specified host–dopant substitution axis, the task is to predict the doping fraction corresponding to the highest experimentally recorded critical temperature (T_c) within the substitution series.

The proposed method builds on the crystal representation and graph backbone of DMFlow. The first stage generates an initial crystal proposal, while a structure-aware correction (SAC) module predicts a revised doping goal based on the proposal and parent information. A second flow integration then jointly refines the doping fraction and crystal geometry toward this goal. The paper also derives a finite-step bound on the refined composition's distance to the predicted goal.

The experiments demonstrate improvements in doping-rate error and direction accuracy on moving-parent and large-move subsets of both ICSD and MP benchmarks. However, the full method does not consistently improve all-parent performance, and the extent to which the second integration contributes beyond the learned doping goal remains unclear.

Critical strengths
  1. Practical and well-motivated problem. The benchmark addresses a concrete experimental decision: whether to increase, decrease, or preserve the dopant fraction of an existing superconductor along a specified substitution axis. This formulation is potentially useful for prioritizing experimental exploration.

  2. Literature-informed benchmark construction. The authors revisit selected superconductivity labels using source-paper evidence and evaluate their approach on two datasets with different structural provenance. The explicit distinction between moving parents and parents already at an observed optimum is valuable.

  3. Promising experimental results. SuperFlow achieves the lowest doping-rate error and highest direction accuracy among the evaluated methods on moving-parent and large-move subsets of both datasets. The improvements over the one-round variant suggest that the refinement procedure provides useful information, particularly when substantial doping adjustments are needed.

Critical weaknesses
  1. The contribution of the learned correction and second integration is not sufficiently disentangled. The One-round SuperFlow ablation removes both the SAC goal prediction and the second integration, making it difficult to determine which component accounts for the improvement. Since SAC already predicts a target doping fraction (x^\dagger), a more direct comparison would evaluate this prediction before the second flow and compare it with the refined endpoint. This would clarify whether the second integration improves the primary doping recommendation or mainly generates a corresponding crystal structure. The existing Parent-only Direct-x baseline does not resolve this question because it uses a different prediction head and training procedure.

  2. The trade-off between improving and preserving the parent is insufficiently evaluated. On ICSD, the full model has higher all-parent xMAE than One-round SuperFlow (0.105 versus 0.089), despite improving moving-parent performance. This matters because 420 of the 1,002 ICSD test parents are already at an observed optimum. The results should separately quantify performance on these already-optimal parents. A keep-the-parent baseline would also provide a useful reference for evaluating whether the model achieves an appropriate balance between changing and preserving the input composition. Moreover, uncertainty estimates obtained by resampling complete substitution trajectories would help establish the consistency of the reported improvements.

  3. Benchmark reliability and generalization remain uncertain. A substitution series can contain as few as three measured doping fractions, meaning that the highest recorded (T_c) may not reliably identify the underlying physical optimum. Small differences in reported (T_c) can also change the optimal-doping label. Although the authors group linked records and source structures across splits, the overlap of material families and substitution axes between partitions is not quantified. Furthermore, the structural pretraining corpus contains ICSD structures, and its potential overlap with downstream test structures or close structural variants should be clarified. These issues affect whether the results demonstrate generalization to unfamiliar materials or interpolation within previously represented families.

  4. Direction accuracy does not fully capture successful doping optimization. A prediction can move in the correct direction while overshooting the observed optimum, potentially ending farther from it than the original parent. Although the paper acknowledges this limitation in Appendix F, an additional metric measuring the fraction of predictions that actually reduce the distance to the observed optimum would better characterize practical usefulness.

  5. The structural benefits are supported primarily by selected examples. The Cu–O and Fe–As case studies illustrate changes in local structural descriptors, but it is unclear whether joint structural refinement consistently produces more accurate or physically meaningful geometries. Aggregate evaluation against experimentally resolved target structures, together with a composition-only modification baseline that retains the parent geometry, would strengthen this claim. Results should distinguish experimentally resolved structures from artificially doped structures. The theoretical bound guarantees tracking of the model-predicted doping goal under its assumptions, but does not establish the accuracy of that goal, structural fidelity, or improvement in experimental (T_c).

Questions for authors
  1. Can the authors report the doping-rate error and direction accuracy of the SAC-predicted goal (x^\dagger) directly, before the second integration? How do these compare with the final output? A second-pass ablation without SAC guidance would further help isolate the contribution of the refinement mechanism.

  2. What are the xMAE and false-movement rates for parents whose initial doping fractions are already optimal? How does SuperFlow compare with keeping the parent unchanged, particularly on ICSD? Can the authors provide trajectory-level bootstrap confidence intervals for the reported improvements?

  3. How much overlap exists between training and test sets in terms of material families and substitution axes? Does structural pretraining contain downstream test structures or close variants? How would the results change when restricting evaluation to densely sampled trajectories or those with clearly identified (T_c) maxima?

  4. For what fraction of moving parents does the predicted doping fraction actually reduce the distance to the observed optimum relative to the input parent? How frequently does the model overshoot the optimum enough to produce a worse result?

  5. Can the authors provide aggregate lattice and local-geometry errors against experimentally resolved target structures, compared with retaining the parent geometry after changing its composition? This would help establish whether joint structural generation adds value beyond the doping-rate prediction itself.

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

The paper proposes a two-stage flow-matching model to predict the optimum doping fraction and structure given a parent crystal and a substitution axis. Their two-stage process involves an initial flow model which changes the composition and structure, followed by a learned SAC (structure-aware correction) head, and a second guided flow pass using the same flow backbone.

Critical strengths
  1. They construct literature-informed optimum doping benchmarks (ICSD and MP) and supplement ICSD with literature-derived labels and corrections. This benchmark itself is a useful contribution.

  2. They report improved moving-parent MAE, including large-move cases, for their two-stage model compared with the single-stage model, which omits SAC and the second flow pass. The two-stage model also improves all-parent MAE on MP. These results support the usefulness of the combined SAC-and-refinement stage.

Critical weaknesses
  1. The existing single-stage comparison removes both the SAC head and the additional flow pass. Thus, one cannot isolate how much of the improvement comes from supervised SAC versus the additional refinement of the second flow stage. The main quantitative metric does not take geometric errors into account. The second stage also updates the geometry, and Figure 4 shows improvements in selected local structural descriptors for two cases. Aggregate structural evaluation across test pairs with valid alignment would help establish whether these improvements generalize.

  2. On ICSD, One-round SuperFlow outperforms the two-round model on the all-parent benchmark. A more explicit characterization of the improvement-versus-preservation tradeoff is therefore necessary. Reporting errors separately on already-optimal parent crystals in both datasets would make this tradeoff clearer.

I think there are some promising directions here. However, quantitatively dissecting the individual benefits of SAC versus the second flow, as well as reporting both doping errors and the frequency of unnecessary doping changes on already-optimal parents, is needed to change my current evaluation. In its current form, I lean toward a weak reject.

Questions for authors
  1. How does SAC's feasible target doping fraction compare with the final doping prediction under the same evaluation metrics? This comparison is necessary to isolate the individual benefits of SAC and the second flow pass.

  2. Can you report aggregate structural evaluation across test pairs with valid alignment to establish whether the improvements in the two cases shown in Figure 4 generalize?

  3. Can you report doping errors separately on already-optimal parent crystals in both datasets, together with the frequency of unnecessary doping changes, to clarify the improvement-versus-preservation tradeoff?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.