acceptodds
Under review as a conference paper at ICLR 2027

Semigroup Riemannian Transport Flow for Few-Step Crystal Generation

Abstract

Deep generative models are the default engine for inorganic crystal design, but each structure costs 50–5000 neural function evaluations, which dominates the cost of screening campaigns, closed-loop laboratories and inverse problems that need many samples per query. The obstacle is what the model learns: flow matching predicts an instantaneous velocity, whereas one large step needs the finite-horizon velocity—the average of that velocity over the interval the step covers. Supervising it is the hard part: distillation needs a converged teacher, and derivative-based self-consistency objectives differentiate through the model, requiring a Jacobian-vector product (JVP). The finite-horizon velocity is instead characterized by a composition law—one long transport step equals two composed shorter steps—together with a boundary condition that pins the field down as the step length goes to zero, and on the flat product manifold used for crystal generation the composition law holds exactly. It can be imposed with ordinary forward evaluations of a frozen copy of the model, with no teacher and no derivative through the model. This yields Semigroup Riemannian Transport Flow (SRTF): a semigroup consistency loss supplies the composition law, a standard flow-matching loss supplies the boundary condition, and the field is parameterized as a local velocity plus a correction proportional to the step length, so the two can be read and tested separately. We prove that the objective's population fixed point recovers the transport coefficients to leading order, so the correction is identified rather than merely fitted, and on trained networks the measured and predicted residual coefficients agree within 3%. On the four standard crystal-structure-prediction benchmarks SRTF exceeds the single-sample match rate of the published baselines considered here with only two to five sampling steps, 10–25 times fewer than the fastest of them, at flow-matching training memory, whereas our JVP implementation trains at only a third of the batch size on the same GPU memory.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
crystal generation, generative models, materials discovery, few-step generation, flow matching, Riemannian transport

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1M5DOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method for generating crystal structures with fewer sampling steps. Instead of predicting only the instantaneous velocity, the author proposes to learn an additional correction for larger sampling steps using a consistency-based objective. This can help avoid expensive derivative calculations and does not require a pre-trained teacher model. Their experiments supports this. Experiments on four crystal structure prediction benchmarks show improved match rates with only 2–5 sampling steps.

Critical strengths

The paper addresses an important practical problem of reducing the computational cost of crystal generation. The proposed method is relatively efficient and comes with theoretical analysis. The experiment results look quite convincing and encouraging.

Critical weaknesses

Although SRTF achieves higher match rates with fewer steps, the generated structures often have larger reconstruction errors. For example, on MP-20, the 5-step SRTF has an RMSE of 0.141, compared with 0.057 for FlowMM. This raises concerns about the trade-off between generation speed and structural accuracy. Since accurate crystal structures are important for practical applications, the paper could better demonstrate whether the few-step results are sufficiently reliable.

Questions for authors
  1. Could the authors provide a more detailed comparison of the trade-off between sampling speed, match rate, and structural accuracy, especially in the 2–5 step setting?
  2. Have the authors evaluated the physical stability or validity of the generated crystals beyond structural matching? Would the performance advantages remain after structural relaxation?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 1M5D

Thank you for the encouraging review and the precise questions. Table and appendix numbers refer to the revised PDF; the code is in the supplementary material. RMSE below is StructureMatcher's normalized RMS displacement.

W1 / Q1. Speed, match rate and precision: the five-step setting first (MP-20, full test split, 3 seeds, same backbone and prior draws; Table 18):

  • Against flow matching at the same five steps, SRTF raises MR₁ from 63.05% to 67.02% (+3.97 [3.49, 4.44]). At two steps the margin is larger: 64.3% vs. 47.3%.
  • On the structures both match, it is more precise at that budget: ΔRMSE −0.066 [−0.068, −0.064].
  • After CHGNet relaxation, the five-step advantage remains: 65.0% vs. 60.0% (fixed 2000-structure test subset).
  • Against recent crystal generators at equal budget, SRTF has lower RMSE than CrystalFlow, OMatG and CrysBFN at every budget, e.g. 0.142 vs. 0.232 / 0.194 at five NFEs on MP-20 (Table 2).

Against a much larger budget. Against flow matching at 50 steps, five-step SRTF is less precise on the structures both match (0.130 vs. 0.058), almost entirely through fractional-coordinate error. Lattice error is already at the 50-step level (1.37% vs. 1.31% in lengths, 0.56° vs. 0.54° in angles), while fractional-coordinate RMSE is 0.065 vs. 0.032 and falls to 0.035 by twenty steps; MPTS-52 shows the same pattern. Three remedies are available:

  • More steps. On identical structures, SRTF is more precise than 50-step flow matching from 20 steps on MPTS-52 (0.157 vs. 0.167) and from 10 on Perov-5 (0.104 vs. 0.129). On MP-20 it is within about 0.006 at 20 steps (0.067 vs. 0.060) and ahead at 50. At these 10–20-step operating points, generation takes 2.4–4.7× less time than flow matching at 50 steps.
  • A precision-oriented sampler. Selecting on RMSE gives MP-20 at five steps 66.8 / 0.129, and Perov-5 at two steps 51.6 / 0.084, on par with DiffCSP's 1000-step 52.0 / 0.076.
  • Relaxation (Q2 below).

Operating points:

intended useoperating point
fast candidate generation2–5 NFEs: MP-20 MR₁ 64–67% at 0.26–0.66 ms per structure (amortized)
recovery/precision balance10–20 NFEs: MR₁ 68.5 / 68.3%, RMSE 0.097 / 0.076, 1.3–2.6 ms (flow matching at 50: 68.1%, 0.070, 6.3 ms)
post-relaxation recovery10 NFEs: 67.2% after relaxation, vs. 67.4% for flow matching at 50 NFEs

Amortized generation time on MP-20 is 0.26 / 0.66 / 1.31 / 2.62 ms per structure at 2 / 5 / 10 / 20 steps, against 6.30 ms for flow matching at 50. The published FlowMM RMSE of 0.057 is averaged over a smaller matched set (MR₁ 61.4). Our same-backbone flow-matching model at 50 steps matches 68.1% at RMSE 0.070.

Q2. Geometric validity, and robustness across two MLIP relaxation models.

Protocol. We relaxed all samples with CHGNet (FIRE, positions + cell, fmax 0.05 eV/Å, ≤ 500 steps) on a fixed 2000-structure test subset, scoring against the DFT ground truth. Failed relaxations count as unmatched. MP-20 rows: mean of 3 seeds; MPTS-52: one seed (Table 19).

armMP-20 MR₁ pre → postMPTS-52 MR₁ pre → postMPTS-52 RMSE post
Flow matching, 5 steps62.8 → 60.017.7 → 16.60.208
SRTF, 5 steps67.0 → 65.023.4 → 21.90.183
SRTF, 10 steps68.6 → 67.224.4 → 23.70.163
SRTF, 20 steps68.0 → 67.225.5 → 24.90.150
Flow matching, 50 steps67.9 → 67.424.0 → 23.00.153
DiffCSP, 1000 steps57.5 → 57.418.1 → 18.20.127
  • At equal budget the advantage survives relaxation: +5.1 points on MP-20 and +5.3 on MPTS-52 at five steps.
  • Against 50 steps: SRTF at 10–20 steps performs comparably to 50-step flow matching on MP-20 (per-seed differences at ten steps −0.70 / +0.25 / −0.25), and is ahead on MPTS-52.
  • Precision on identical structures after relaxation (seed 42): at 20 steps, SRTF is level on MP-20 (0.0533 vs. 0.0530) and more precise on MPTS-52 (0.113 vs. 0.129).
  • Robustness across two MLIPs: a MACE-MP-0 check on 500 structures per arm corroborates the main conclusions (MP-20 at five steps: 63.2% vs. 58.6%; SRTF at 20 steps 68.0% vs. 66.8% for flow matching at 50).
  • Calibration: relaxing the DFT reference structures with the same protocol retains matches for 98.7–99.6% of them.

Geometric validity before relaxation:

  • No non-positive-definite cell occurs for any method or budget.
  • At the matched five-step budget, SRTF has fewer sub-0.5 Å contacts than flow matching (MP-20 0.9% vs. 2.1%; MPTS-52 9.0% vs. 10.3%).
  • At 20 steps, SRTF has fewer than flow matching at 50 on MPTS-52 (1.9% vs. 3.8%).

These are MLIP surrogates, not DFT, as stated in Appendix G.3.

These results address the precision–speed trade-off and show that the few-step advantage survives surrogate relaxation. We are happy to clarify further.

Thank you for the detailed response. I appreciate the more comprehensive analysis of the trade-off between sampling efficiency, match rate, and structural precision. The matched-budget comparisons and relaxation experiments substantially address my original concerns. In particular, the results showing that SRTF maintains its advantages after relaxation and achieves competitive precision with 10–20 sampling steps are convincing.

Thank you for the follow-up. We are glad that the matched-budget comparisons and the relaxation experiments address your concerns, and that the results on precision at few steps and on the advantage after relaxation are convincing. Your questions led us to add the speed–match-rate–precision analysis and the relaxation study, which strengthened the paper. Thank you again for your careful review.

Reviewer_aMHoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Semigroup Riemannian Transport Flow (SRTF), which learns a finite-horizon average velocity for few-step crystal structure generation. It supervises this velocity through semigroup consistency based on the composition law and parameterizes it as an instantaneous velocity plus a first-order correction in the time horizon. Empirically, it achieves a 10–25× reduction in the number of function evaluations (NFEs) relative to FlowMM's 50-NFE setting in Table 1.

Critical strengths

Empirical gains in sampling efficiency. The proposed method achieves competitive single-sample match rates with only 2–5 sampling steps, substantially reducing NFEs compared with the baselines in Table 1.

Critical weaknesses

Limited methodological novelty. SRTF is presented as a contribution in both its objective and its parameterization. However, the objective appears to be a direct application of the semigroup RMF proposed in Riemannian MeanFlow to a flat manifold. The only difference is the parameterization of the average velocity. SRTF would therefore be more appropriately framed as a parameterization method than as a framework that also contributes an objective. The proposed parameterization itself also appears incremental, and its performance gains seem small compared with the effect of changing the objective (Figure 5(a)).

Insufficient justification of the practical motivation. The introduction argues that per-structure sampling cost limits large-scale materials discovery. In practice, however, experimental validation is the bottleneck in crystal discovery. Even on the computational side, the bottleneck often lies in DFT-based structure relaxation and property evaluation rather than generative sampling.

Questions for authors
  1. The reported behavior of the JVP implementation seems substantially different from that reported in other papers. MeanFlow typically has modest memory overhead relative to flow matching because JVP computation requires only temporary memory. Moreover, MeanFlow and Terminal Velocity Matching achieve good performance with JVPs in Euclidean space without the reported convergence issue. To help explain this discrepancy, could you provide the exact baseline implementation and training setup, particularly the use of stop-gradient, the time-embedding strategy, and loss weighting?

  2. In Table 1, the single-sample match rate is higher at lower NFEs than at 20 NFEs on Perov-5 and Carbon-24. This trend is counterintuitive. Could you explain it?

  3. More recent CSP baselines, such as OMatG and Crystalite, appear to be missing. Could you include their results in the comparison?

  4. Could you justify the practical importance of few-step CSP models for materials discovery?

3 comments
AuthorsOpenReview ↗

Response to Reviewer aMHo

Thank you for an expert and detailed review. It led us to audit our JVP baseline, to sharpen the contribution statement, and to rewrite the method section so that every symbol is defined where it first appears. Table numbers refer to the revised PDF; the code, including the baseline, is in the supplementary material.

W1. Contribution. SRTF builds on the semigroup characterization and stop-gradient regression of Woo et al. (2026) (their Props. 2.3, 3.1), now credited in §1–2; we no longer list the objective as a contribution. We contribute:

  • A structured field w = a + τb that the objective identifies. The horizon enters only through a scalar, never the backbone, so the target reuses the online backbone pass. Prop. 1(iii) shows that the semigroup residual pins b to −½Dₜa at leading order, an analysis RMF does not provide.
  • Evidence on trained models. Switching off the transport branch costs over 40 points on MP-20 at one step (nothing for a chord-supervised control); b aligns with −½Dₜa (median cosine +0.65 vs. −0.11); the measured leading residual coefficient matches the prediction within 3%.
  • A crystal study against a corrected MeanFlow baseline (Q1) and recent generators (Q3); neither RMF paper evaluates crystals.

On Fig. 5a we agree: with supervision fixed, the decomposition adds +2.0 to +2.7 points at 1–2 steps (Table 13). Its role is identification, inspectability and a cheaper target, and the text now says so.

Q1. JVP baseline. You were right. MF-Riem-ME keeps MeanFlow's detached-target objective and computes the JVP target with torch.func.jvp without retaining its graph, followed by a separate online forward (gradients unchanged; JVP checked against finite differences). It uses MeanFlow's settings: adaptive weight with p = 1 and c = 10⁻³ (detached), positional embeddings of (t, t−r), and 25% of samples with r ≠ t (Appendix G.1). It trains at flow-matching memory (4.11 vs. 4.10 GB; SRTF 4.23), so we removed the memory, batch-size and non-convergence claims. The corrected comparison still favors SRTF (Tables 16–17; MR₁ %, same backbone and budget):

NFE 12520
SRTF (MP-20, 1000 ep, 3 seeds)52.9860.8864.2065.93
MF-Riem-ME30.0846.2554.3257.53
Δ [95% CI]+22.9 [21.3, 24.6]+14.6 [13.1, 16.2]+9.9 [8.5, 11.3]+8.4 [7.0, 9.8]
Δ, MPTS-52, 1000 ep+7.2+4.4+3.3+2.3

SRTF is also 1.72× faster per training step (79.3 vs. 136.7 ms), more stable across seeds (s.d. at most 0.8 vs. up to 4.1 points), and reaches the JVP arm's two-hour accuracy in about 40 minutes (Figure 11). The JVP arm does learn on MPTS-52, only more slowly.

Q2. MR₁ falling with more steps (Appendix G.6). The same-backbone flow-matching model falls further (Perov-5: 52.6 at three steps to 26.5 at twenty). The decline occurs at the loose tolerance stol = 0.5: at stol 0.3–0.4, SRTF's Perov-5 MR₁ is flat at 44–47% from three to fifty steps. The two-step peak consists of loosely placed structures (none within RMS 0.1 of its target, against 39% from five steps on), which further steps turn into precise matches or clear misses. At strict tolerance SRTF reproduces about twice as many Perov-5 structures as flow matching (44.1% vs. 21.1%). Carbon-24 shows the same shared decline, largely absent at strict tolerance, though this account applies only in part there. This analysis replaces "StructureMatcher artifact".

Q3. Recent baselines (Table 2; released checkpoints in our harness, samplers tuned on validation; MR₁ %, MP-20 / MPTS-52):

NFE 251020
CrystalFlow32.2 / 5.258.8 / 14.662.3 / 17.967.9 / 23.3
OMatG18.9 / 6.360.8 / 16.364.1 / 18.467.8 / 19.4
uFlowCSP51.3 / 9.8†53.3 / 11.4†54.1 / 11.8†54.5 / 12.2†
SRTF64.3 / 15.867.0 / 18.868.5 / 20.468.3 / 20.7

†Our retraining; no released checkpoint. SRTF has the highest MR₁ at 2, 5 and 10 NFEs, with fewer parameters (12.4M vs. 20.9M and 131M; uFlowCSP is an independently developed MeanFlow-based model) and lower RMSE than CrystalFlow and OMatG at every budget. At 20–50 NFEs these two catch up or pass it. Crystalite's published 400-step results are 66.09% / 31.56%: SRTF reaches 68.5% on MP-20 at ten NFEs, while Crystalite is higher on MPTS-52 (not run in our harness).

W2 / Q4. Motivation. We agree that experimental validation and DFT dominate a discovery campaign, and removed "dominates the cost". Sampling cost matters where sampling is repeated inside a loop: inverse problems such as structure solution from diffraction, generate-and-filter screening with cheap MLIP filters, and reward-guided sampling. Five-step generation takes about 0.66 ms per structure, against 6.3 ms for 50-step flow matching.

We hope the corrected baseline, the clarified contribution and the new comparisons resolve the soundness concerns.

Thank you for the detailed response and for correcting the JVP baseline, clarifying the connection to semigroup RMF, and adding comparisons with recent methods. These revisions address several of my earlier concerns, but my concern about the significance of the contribution remains.

Methodological contribution. Since the semigroup objective is inherited from prior work, the main methodological contribution lies in the structured parameterization and its analysis. While I acknowledge the benefits of computational reuse and interpretability, I still find the advance incremental. Under fixed supervision, the accuracy gains are relatively modest (2.0–2.7 percentage points at 1–2 NFEs). The synthetic identification experiment supports the proposed analysis but does not establish broader practical benefits of the parameterization, particularly beyond the studied setting.

Application impact. I recognize the improved low-NFE CSP results. However, the practical motivations involving diffraction-based structure solution, generate-and-filter screening, and reward-guided design are not directly supported by the presented experiments. Evaluations on tasks such as de novo materials generation or diffraction-conditioned structure solution would provide more convincing evidence of practical impact. Reduced sampling cost alone does not establish meaningful improvements in these discovery and design workflows.

Overall, while I appreciate the revisions and additional experiments, I remain unconvinced that the methodological and application contributions together constitute a sufficiently substantial advance. I will therefore keep my score unchanged.

AuthorsOpenReview ↗

Thank you; a clarification on the identification evidence and on scope

Thank you for engaging with the revision and for confirming that the baseline, the positioning relative to semigroup RMF and the comparisons with recent methods are now addressed. We respond briefly to the two remaining points.

Methodological contribution. We agree that the accuracy gain from the parameterization alone, under fixed semigroup supervision, is modest (+2.0 to +2.7 points at 1–2 NFEs, Table 13), and the revision states this. One clarification: the identification evidence is not only synthetic. On the trained MP-20 crystal models,

  • switching off the horizon head costs over 40 points at one step, against nothing for a chord-supervised control (Figure 5b);
  • b aligns with −½Dₜa (median cosine +0.65 vs. −0.11 for the control), with over 96% of states positively aligned for t ≥ 0.6, where few-step jumps are taken (Table 10);
  • the measured residual coefficient matches the predicted one within 3% (Table 11).

The synthetic manifold (Table 12) only removes the finite-difference estimator from this test. What the parameterization adds beyond accuracy is that the horizon correction becomes a separately identified and testable object with a cheaper target, rather than part of a monolithic map.

Application impact. We agree that our experiments establish few-step CSP, not end-to-end gains in diffraction-based structure solution, screening or reward-guided design. CSP on these four benchmarks is the standard evaluation for this model family (CDVAE, DiffCSP, FlowMM, CrystalFlow, OMatG), and the relaxation study shows that the few-step advantage survives a downstream relaxation step (65.0% vs. 60.0% at five steps). In the final version we will present the downstream workflows as motivation rather than as demonstrated benefits, and name de novo generation and diffraction-conditioned structure solution as the next evaluations.

We respect your assessment, and thank you again for a careful review that has substantially improved the paper.

Reviewer_9gHnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Semigroup Riemannian Transport Flow (SRTF) for few-step crystal generation. It parameterizes finite-horizon velocity as a local velocity plus a step-size-dependent correction, trained through flow matching and a consistency objective requiring one large transport step to agree with two smaller steps. This avoids pretrained teachers and Jacobian–vector products. Theoretical analysis establishes semigroup-based identification of the leading transport coefficients and a conditional sampling-error bound. Experiments on four crystal structure prediction benchmarks report strong single-sample match rates.

Critical strengths
  • The paper addresses an important practical bottleneck: the high sampling cost of crystal generation. The method is supported by a well-developed theoretical analysis and achieves strong results on four CSP benchmarks, reaching competitive single-sample match rates with only two to five network evaluations.
  • The method learns large transport steps through a simple consistency principle: one large step should agree with two smaller steps. This avoids both a pretrained teacher and JVP computation, enabling few-step generation while keeping training memory comparable to standard flow matching.
Critical weaknesses
  • The parameterization w=a+τbw=a+\tau b restricts the predicted velocity to vary linearly with step size. This is justified by a local Taylor expansion, but the neglected higher-order terms may matter for the large steps central to few-step generation. Since this restricted form generally cannot satisfy exact semigroup consistency, training must trade off consistency across different step sizes. It remains unclear whether this limits few-step accuracy. An ablation adding a quadratic term would help assess this limitation and justify the linear design.
  • The experimental comparison omits relevant efficiency-focused baselines, particularly CrysBFN [1] and CrystalFlow [2]. Both evaluate CSP and explicitly study sampling efficiency. Including their quality–NFE curves under a consistent evaluation protocol would better establish SRTF’s advantage over existing efficient crystal generators.

Reference

[1] Wu, Hanlin, et al. "A periodic bayesian flow for material generation." International Conference on Learning Representations. Vol. 2025. 2025.

[2] Luo, Xiaoshan, et al. "CrystalFlow: a flow-based generative model for crystalline materials." Nature Communications 16.1 (2025): 9267.

Questions for authors
  • Table 1 shows that no single NFE budget performs best across datasets and sample counts.- How should users select the step budget and hybrid sampler configuration for a new task, particularly when reference structures are unavailable for validation? Can a configuration selected on one dataset or budget transfer to others without substantial performance loss?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 9gHn

Thank you for the positive review, and for identifying the importance of higher-order horizon dependence. Table and appendix numbers refer to the revised PDF; the code is in the supplementary material.

W1. The linear τ-dependence. We added the controlled ablation you requested, and corrected an omission in the implementation description. The benchmark checkpoints use w⁽²⁾ = a + τb + τ²c, with an independent, zero-initialized τ² readout. The controlled supervision comparisons use w⁽¹⁾ = a + τb. The revised §3.2 introduces both, and Table 6 lists the variant used in each main comparison.

Results (matched, freshly trained runs; MP-20, 3 seeds, full test split, MR₁ %):

NFEw⁽¹⁾, hybridw⁽²⁾, hybridw⁽¹⁾, pure SRTF samplerw⁽²⁾, pure
156.25 ± 1.1356.64 ± 0.5356.25 ± 1.1356.64 ± 0.53
263.58 ± 1.0264.42 ± 0.4358.79 ± 0.6761.74 ± 0.50
566.24 ± 0.7667.21 ± 0.5066.29 ± 0.6867.31 ± 0.49
1067.64 ± 0.7068.52 ± 0.2367.23 ± 0.9567.73 ± 0.29
2068.13 ± 0.7468.42 ± 0.2767.35 ± 0.9667.82 ± 0.28
  • The first-order field captures most of the few-step benefit. The independent quadratic term adds finite-horizon flexibility: +0.3 to +1.0 points with the hybrid sampler, and 58.79% → 61.74% with the pure two-step sampler, where every step is a large jump. It also lowers seed variance.
  • The quadratic correction is non-negligible at large horizons. The ratio ‖τ²c‖/‖τb‖ is 0.52 at τ = 0.5 and 1.04 at τ = 1.
  • It preserves the leading-order identification of b. The quadratic head first enters the residual at O(τ³). The observed residual scaling is similar for both fields.
  • On MPTS-52 (one seed per arm), the two fields are within ±1 point.

Does the restricted form limit few-step accuracy? The ablation indicates it is not the bottleneck: freeing the next order changes MR₁ by at most 1.0 point with the hybrid sampler, at 0.45% extra parameters, and the measured semigroup residual keeps the same leading O(τ²) scaling for both fields (log–log slopes 1.88 and 1.80). Prop. 1(iii) explains why: what the restricted class can cancel is exactly the leading coefficient, which is what identifies b (Appendix G.4).

W2. CrysBFN and CrystalFlow. Both are released checkpoints run in our harness, with samplers tuned on validation (Table 2; CrysBFN in its caption):

  • CrysBFN at its ten trained steps: 60.5 on MP-20. SRTF scores 64.3 at two NFEs and 68.5 at ten.
  • CrystalFlow at 2 / 5 / 10 / 20 NFEs:
    • MP-20: 32.2 / 58.8 / 62.3 / 67.9, against SRTF's 64.3 / 67.0 / 68.5 / 68.3;
    • MPTS-52: 5.2 / 14.6 / 17.9 / 23.3, against 15.8 / 18.8 / 20.4 / 20.7.

SRTF leads at ten NFEs or fewer; we also added OMatG and uFlowCSP, and at two NFEs on MP-20 SRTF reaches 64.3% against 51.3% for the strongest evaluated baseline. CrystalFlow is level at twenty on MP-20 and ahead from twenty on MPTS-52. Table 2 now also reports RMSE, and SRTF's is lower than both baselines' at every budget (MP-20, 2 / 5 / 10 / 20 NFEs: 0.180 / 0.142 / 0.097 / 0.076 for SRTF, against 0.334 / 0.232 / 0.178 / 0.193 for CrystalFlow; at ten NFEs, 0.113 for CrysBFN). Appendix G.7 adds our retraining on Perov-5 and Carbon-24, where no checkpoints are released: SRTF leads at one and two steps, and the retrained CrystalFlow from five (Perov-5) and ten (Carbon-24) steps.

Q. Choosing the budget and sampler without references. Two separate tests, full test split, 3 seeds, Δ MR₁ against per-dataset tuning (Table 21):

  • Transfer. One configuration per budget, chosen on MP-20 validation and frozen, transfers to MPTS-52 within 0.23 points at the tested budgets of 2, 5, 10 and 20 NFEs. The frozen settings, (leading SRTF steps, coordinate anti-annealing), are (2, 0), (3, 1), (10, 2) and (2, 2), with no lattice anti-annealing.
  • Reference-free selection. Chosen per dataset from validation compositions only: each candidate configuration is ranked separately by its close-contact rate and its mean single-point MLIP energy, and the best average rank wins. No reference structures are used. It costs at most 0.96 points on MP-20 and MPTS-52, and at two NFEs it picks the tuned configuration exactly. Both procedures select a configuration within a given NFE budget; the budget itself remains a latency–quality choice.
  • Small, high-symmetry datasets. On Perov-5 and Carbon-24, the best 2–5-step configurations use strong anti-annealing that neither procedure selects. From ten steps on, every policy is within about 0.7 points (Perov-5) and 2.1 points (Carbon-24).

Practical guidance (Appendix G.5):

  • multi-element systems: 5–10 NFEs with a transferred or reference-free configuration;
  • when precision matters: 10–20 NFEs;
  • small, high-symmetry cells: tune anti-annealing on a few known structures.

We hope these additions answer your questions, and we are happy to clarify further.

Thank you for the detailed response. The controlled horizon-order ablation and comparisons with CrysBFN and CrystalFlow address my main concerns. I also appreciate the clarification of the parameterization used in the benchmark experiments and the practical guidance on sampler selection. I remain positive about the work and will maintain my original score.

Thank you for the follow-up and for your positive assessment. We are glad that our response addresses your concerns. Your request for the quadratic ablation led us to document the benchmark model's readout explicitly, which made the paper more precise. Thank you again for your careful review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.