acceptodds
Under review as a conference paper at ICLR 2027

OMatG-flash: An All-Atom Flow Map with Reinforce Adjoint Matching for Scalable Materials Discovery

10 views

Abstract

The discovery of novel inorganic materials drives technological breakthroughs in critical fields such as computing and energy storage. Generative AI has promised to accelerate the materials discovery pipeline, but state-of-the-art flow and diffusion models remain bottlenecked by the cost of proposing candidate materials. To address this, we introduce OMatG-flash, an all-atom flow map for inorganic crystal structure prediction (CSP) and de novo generation (DNG). OMatG-flash is a Pareto-optimal inference engine for materials, sampling candidate materials with an order of magnitude fewer inference steps and less wall-clock time than existing flow and diffusion models while demonstrating benchmark performance on par with the state-of-the-art. To enable post-training fine-tuning we apply Reinforce Adjoint Matching to flow maps, further improving performance on the CSP task. OMatG-flash showcases the potential of flow maps to accelerate generation of high-quality candidate inorganic materials and demonstrates a step forward in sample throughput necessary for data-hungry materials discovery workflows.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Generative AI, Materials Discovery, Flow Map, Flow Matching, Reinforcement 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_yigrOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors introduce OMatG-Flash and OMatG-Flash-RAM, a version with post-training. Their model closely matches the recent SOTA on most metrics and, through their flow map based approach, is much faster, which makes this model Pareto optimal for the crystal generation task.

Critical strengths

Critical strengths include:

  • much faster than current SOTA for similar performance
  • introduce post-training for flow-map-based crystal generation
  • well-presented, clear writing, intuitive figures
  • overall interesting ideas and good results
Critical weaknesses

Critical weaknesses include:

  • As the authors note, the base model lacks geometric accuracy compared to Crystalite.
  • The authors discuss issues with match rates, which are legitimate, but then they say they’re going to use METRe, which I agree is a better metric. But table 1 then reports MR (an acronym that isn’t introduced), and Table 2 is below table 3. Overall, I just found the results section slightly confusing to read because of the combination of these things. Easy to fix, though.
  • The post-training objective isn’t ablated at all (and also can’t compensate for the base model’s RMSE gap to Crystallite).
Questions for authors

It would be helpful to see some ablations on the post-training objective and some more discussion on the stability of the post-trained checkpoints. If no ablations are possible, more discussion on this would be appreciated.

3 comments
AuthorsOpenReview ↗

Response to Reviewer yigr

We thank the reviewer for their feedback and provide some further information and clarification below.

Geometric accuracy

We agree that sacrificing geometric accuracy on as-generated structures is a weakness of OMatG-flash. We point the reviewer to our response to zMwX where we explore this issue further. Briefly, we find that, after force field relaxation, the RMSE gap between our model and selected baselines is narrowed substantially.

Clarity

We thank the reviewer for pointing out that we neglect to properly introduce “MR” as an evaluation metric. MR is meant to stand for the flawed one-to-one match rate, where the StructureMatcher algorithm is used to compare two index-matched crystals to determine if they match or not. We will revisit our discussion of match rates, and the results section as a whole, such that the flow of our results is more coherent. We also agree that table placement should be improved and will account for this in our preparation for a camera-ready version of this work.

Post-training ablation

We agree with the reviewer that a more thorough ablation and discussion of post-training stability and performance is warranted. We did observe that, for the CSP energy-minimization reward, collapse occurs almost always. Various choices of reward hyperparameters can delay the onset of this collapse, but for CSP it is quite difficult to eliminate entirely. We do, however, encourage the reviewer to see our exploration of an alternative reward for de novo generation (DNG) in our response to whKH as this may fit the reviewer’s interest in ablation of a different reward. Using a reward from the literature on reinforcement learning for DNG improvement, we post-train OMatG-flash and find that not only does the S.U.N. and M.S.U.N. of generated samples increase, but the stability of the RAM optimization under this reward is greatly improved. In fact, for this case, we do not observe any reward collapse as we observed in the CSP case. We argue that, therefore, reward choice has a critical role in determining the stability of the proposed post-training scheme and that RAM is not inherently unstable when applied to flow map post-training.

As for comparison to Crystalite's geometric accuracy, we find that after relaxation the RMSE gap to Crystalite decreases substantially and we point the reviewer to our response to zMwX for more discussion on this. We also note that energy minimization is only a proxy for improving the CSP model. The CSP task is a structure matching task and it is not always true that the lowest energy crystal polymorph is the reference structure which must be matched. We therefore study energetics in our response to whKH to understand how post-training impacts the energetics of crystals produced by OMatG-flash. We find that combining RAM post-training on minimizing energy with amortized flow map inference for k=10k=10 gives state-of-the-art performance in terms of generating low energy structures at a fraction of the cost of a conventional flow or diffusion model.

Reviewer_yigrOpenReview ↗

Response to Authors

I am satisfied with the authors response and commitment to improve areas of the manuscript I found fault with in my initial review. I think this is a solid contribution to the field, and the further discussion of the alternative reward for de novo generation seems very promising.

We thank the reviewer for their engagement and feedback! We believe that their critical assessment has helped us to enhance the quality and clarity of our work.

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

OMatG-flash combines a SOTA model's backbone with a flow map for CSP and DNG tasks. It further adapts RAM as a post-training scheme to improve CSP using energy rewards computed after relaxation. Experiments show competitive generation quality at substantially higher throughput.

Critical strengths

S1. Few-step generation achieves a competitive quality–speed trade-off. OMatG-flash generates crystals with fewer model calls while retaining competitive performance on established benchmarks. Its substantial improvement in sampling throughput demonstrates the practical potential of few-step generation in atomistic generative models.

Critical weaknesses

W1. The methodological contribution beyond combining existing approaches is limited. The model builds on Crystalite’s representation and backbone, an existing flow-map training objective, and RAM post training. The only methodological extension is applying RAM to the flow map’s velocity, which would be considered a marginal contribution.

W2. The benefit of energy-based post-training is not sufficiently established. RAM optimizes energy after relaxation, while the main CSP evaluation measures reference matching. Although relaxed relative energy decreases during training, the paper does not establish whether RAM improves energetic quality relative to baselines under matched relaxation budgets. RAM is also excluded from DNG due to reward-hacking concerns, without evaluating existing reward strategies(e.g., GRPO with the novelty or uniqueness of group rollout). This leaves the practical benefit and scope of the post-training contribution unclear.

Questions for authors

Q1(experimental request). How about comparing RAM, the base model, and strong baselines under matched relaxation budgets, reporting energy before and after relaxation?

Q2(experimental request). How about evaluating RAM on DNG with a diversity reward used in prior works (e.g., using GRPO for uniqueness/novelty together with MLIP energy) to assess whether reward hacking can be mitigated?

6 comments
AuthorsOpenReview ↗

Response to Reviewer whKH

We thank the reviewer for their constructive criticism and kindly address some of their points below.

CSP on matched relaxation budget

We agree that the energy minimization reward is not entirely consistent with the CSP task of matching observed structures. We compute energetics on the as-generated and relaxed structures produced by OMatG-flash as well as several baselines. For each model listed we compute the relative energy with respect to the lowest energy polymorph in the standard (non-polymorph split) MP-20 test set. We use this to respect the training data of the baseline models though we reiterate that MP-20-ps should be considered as the canonical data split for future work. We report the proportion of structures that are as low or lower in energy than the reference lowest energy polymorph. All entries are for $k=1$ and relaxation was carried out with the LBFGS optimization algorithm using a force tolerance of 0.01 eV / Å and up to 500 steps. The asterisk indicates that structures which recorded a relative energy of 10 eV / atom or more were discarded to obtain a stable average though, for all models, this was only a small fraction.

ModelΔE ≤ 0 (as generated) ↑ΔE ≤ 0 (relaxed) ↑Mean ΔE* (as generated) ↓Mean ΔE* (relaxed) ↓Median relaxation steps ↓
OMatG-flash1.66%33.64%0.40360.031050
OMatG-flash-RAM1.96%37.38%0.29890.020545
OMatG Linear-ODE0.74%36.82%0.18890.020545
Crystalite5.00%42.09%0.05450.009435

We highlight that the reinforced model excels with respect to the base model at exactly the task for which it was trained. Similarly, the average energy after relaxation is lower on average than the base model. While we cannot argue that the energetics are stronger than all competing baselines, our result demonstrates a clear advantage over the base flow map. We also report results for $k=10$ inference using the reinforced model. The time taken to run inference on OMatG-flash and OMatG-flash-RAM and rank the samples via MACE single-point calculations fits squarely under the Crystalite $k=1$ inference budget.

ModelkMean ΔE* (relaxed) ↓ΔE ≤ 0 ↑
OMatG-flash100.012739.80%
OMatG-flash-RAM100.009141.16%
OMatG Linear-ODE10.020536.82%
Crystalite10.009442.09%

Combining RAM post-training with the amortized inference gives state-of-the-art performance at a fraction of the inference-time cost.

DNG post-training

At the time of submission we believed typical DNG rewards were quite prone to hacking. Since then, Prakash et. al. (OMatGRPO) have proposed a DNG reward that mitigates hacking. It combines the energy above the LeMat-Bulk hull (using the UMA machine learned force field), a creativity term for structures unique within the batch and novel with respect to MP-20, and a compositional-coverage bonus. It also has guards against known forms of hull hacking. We fine-tuned our pretrained MP-20 generator with RAM on this reward. Each step generated 4 groups of 16 crystals with 16 sampling steps, and each crystal was relaxed for 50 UMA steps before scoring. We trained for 8,423 steps and kept the checkpoint with the best M.S.U.N. on LeMat-GenBench (step 6,499, 29.9% S.U.N. + M.S.U.N. vs. 21.0% for base OMatG-flash)

After submission, a bug was found in the standard DNG LeMat-GenBench evaluation. The hull energies came from a different MACE model than the one scoring the generated structures. With the fix, our stable and S.U.N. rates are lower than in the submission. OMatGRPO and Crystalite numbers are as reported by their authors.

MethodValid ↑Novel ↑Stable ↑S.U.N. ↑S.U.N. + M.S.U.N. ↑
As Reported Crystalite (Broken Hull)97.253.212.71.524.1
As Reported OMatGRPO (Broken Hull)69.861.61.00.345.8
Our Reproduction OMatGRPO (Fixed Hull)69.861.60.30.0446.3
OMatG-flash, NFE 1696.658.73.60.721.0
OMatG-flash, NFE 896.061.43.90.820.6
OMatG-flash, NFE 496.166.42.60.419.1
OMatG-flash-RAM, NFE 1696.060.93.81.031.4
OMatG-flash-RAM, NFE 895.963.33.00.529.8
OMatG-flash-RAM, NFE 495.268.42.40.426.4

Though OMatG-flash does not outperform OMatGRPO in S.U.N. + M.S.U.N., it is roughly 160× faster at NFE 16 than OMatGRPO. OMatG-flash therefore remains on the Pareto front of quality versus throughput. If one uses this extra time to perform a heuristic ranking algorithm we can comfortably achieve SOTA.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, but I have two remaining questions:

  1. Could the authors report uniqueness separately? S.U.N. + M.S.U.N. alone does not establish whether the diversity-aware reward effectively prevents mode collapse.

  2. Could the authors confirm that the generated samples' energies and reference hull energies are computed consistently using the same MLIP and energy conventions?

We thank the reviewer for their engagement. Below we address remaining questions and concerns.

  1. We are happy to report uniqueness alongside our other results. Ultimately, we find that uniqueness does not collapse after RAM post-training which is encouraging that the model is not hacking very few structures. We include a table of results below:
MethodValid ↑Unique ↑Novel ↑Stable ↑S.U.N. ↑S.U.N. + M.S.U.N. ↑
Crystalite (broken hull)97.295.853.212.71.524.1
OMatGRPO69.868.961.60.30.046.3
OMatG-flash, NFE 1696.695.858.73.60.721.0
OMatG-flash, NFE 896.095.361.43.90.820.6
OMatG-flash, NFE 496.195.666.42.60.419.1
OMatG-flash-RAM, NFE 1696.095.760.93.81.031.4
OMatG-flash-RAM, NFE 895.995.663.33.00.529.8
OMatG-flash-RAM, NFE 495.294.968.42.40.426.4
  1. Most of our computed results for the rebuttal (OMatGRPO reproduction and OMatG-flash) are obtained using a consistent MLIP and energy convention. The current version of the paper and the "As Reported" Crystalite result in our rebuttal still use the inconsistent scheme from a previous bugged LeMat-GenBench version which was recently corrected on September 15th. For a camera-ready version of the work we will update our results in the paper such that they are computed using the corrected scheme. Overall, what we have found is that using the corrected hull shifts SUN structures such that they become MSUN whereas SUN + MSUN remains roughly the same so ranking is mostly unaffected.
Reviewer_zMwXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OMatG-flash attacks the inference cost of generative models for inorganic crystals. Flow and diffusion models must integrate an ODE to produce each candidate, which caps throughput. The proposal is a two-time flow map that learns the solution operator directly, so a structure comes out of a few composed jumps. It is parameterized with Riemannian MeanFlow, trained on a tangency plus consistency loss, and built on Crystalite's backbone with one token per atom and a separate lattice token, giving up equivariance for speed. Unlike most flow maps it is trained from data rather than distilled from a teacher.

The second piece is post-training. Flow maps are deterministic, so DDPO, PPO, and GRPO do not apply and the usual ODE to SDE conversion does not transfer. The authors adapt Reinforce Adjoint Matching by exploiting the tangency condition, apparently its first use on flow maps, with negative MACE-MPA-0 energy after 50 relaxation steps as the reward. This is applied only to crystal structure prediction.

The speed claim is the contribution; the quality claim is competitive rather than best. On the polymorph split MP-20-ps the post-trained model leads unconditional flow and diffusion models on METRe (71.10 against Crystalite's 70.87) but its cRMSE of 0.221 is clearly worse than Crystalite's 0.174. On de novo generation at 16 function evaluations it is second to Crystalite (22.6 against 24.1 combined S.U.N. and M.S.U.N.), with higher novelty and lower stability. The single-sample MP-20 match rate sits below MCFlow and OMatG, and the stronger numbers come from drawing three or five samples, which the authors say plainly is not apples to apples, their argument being that five draws still cost less than one from a conventional model. They also concede that MaskGXT beats them outright on CSP, flag their high RMSE as error compounding under few-step inference, and decline to run RAM on de novo generation because S.U.N. is known to be hackable by RL.

Critical strengths
  1. The speed claim is correctly scoped and does not depend on measurement choices.

The paper claims Pareto-optimality rather than SOTA, which is the honest version of what it shows. More importantly the claim survives scrutiny of how throughput was measured, because the model reuses Crystalite's backbone, so per-evaluation cost is comparable to the baselines and the NFE ratio largely determines the wall-clock ratio. Going from hundreds of integration steps to 4 to 16 composed jumps is a hardware-independent statement, and that is the load-bearing one.

  1. Reinforce Adjoint Matching on flow maps is a real methodological contribution.

The motivation is correct rather than decorative: flow maps are deterministic, so policy-gradient methods that require a stochastic MDP do not apply, and the usual ODE to SDE conversion fails outside the instantaneous limit. Deriving the objective through the tangency condition is a clean way around this, it needs no noise injection that would corrupt the generative dynamics, and it preserves few-step inference. This is the part of the paper I would expect others to build on.

  1. The paper repeatedly declines free wins.

It advocates the polymorph split and METRe/cRMSE over the leaky MP-20 protocol, which costs it numbers. It states in the main text that MaskGXT beats it on CSP and Crystalite beats it on DNG. And it refuses to run RAM on de novo generation on the grounds that S.U.N. and M.S.U.N. are known to be reward-hackable, forgoing an easy headline. That pattern raises my confidence in the numbers it does report.

Critical weaknesses
  1. The paper's defense of its accuracy deficit is an assertion, and the experiment that would test it is nearly free.

RMSE at a single sample on MP-20 is 0.1172 against 0.0592 for MCFlow, roughly double, and cRMSE on MP-20-ps is 0.221 against 0.174 for Crystalite. The authors acknowledge this and argue that it does not matter because production inference is always followed by MLIP relaxation, so landing in the right basin of attraction is what counts. That argument is plausible and entirely untested. Post-relaxation RMSE for both their model and the baselines would settle it, and they already run MACE relaxations to compute RAM rewards, so the cost is marginal. Without it, the quality side of the Pareto claim rests on an assumption rather than a measurement.

The presentation compounds this. Table 1 places k=3 and k=5 rows alongside k=1 baselines in the same column under a bold-the-best rule, so the bolded 78.27 sits next to MCFlow's 70.38. The text disclaims the comparison, but the argument the authors actually want to make, that they win at matched wall-clock, requires a compute-matched table rather than a caveat.

  1. RAM is missing the obvious non-learned baseline.

The reward is MACE energy after relaxation, and the paper already samples k times at inference. So the natural control is best-of-k reranked by that same MLIP, which requires no training at all. The RAM gain is METRe 68.56 to 71.10 and cRMSE 0.250 to 0.221, an increment that reranking at k=3 or k=5 might well recover. Absent this comparison I cannot tell whether RAM teaches the model something or merely relocates inference-time selection into training. Since RAM is the paper's methodological novelty, this is the comparison that determines what the contribution is worth.

Questions for authors

1.What is post-relaxation RMSE for OMatG-flash and for the baselines? Your defense of the accuracy gap is that MLIP relaxation makes it irrelevant, and you already run MACE relaxations for RAM rewards. If relaxed structures from your model land as close to reference as relaxed structures from MCFlow or Crystalite, the argument is made. This is the answer most likely to move my score.

2.How does RAM compare against best-of-k reranked by MACE energy? You already sample k times, and the reward is that same MLIP, so reranking is a zero-training control for the 68.56 to 71.10 METRe gain. Without it I cannot tell whether RAM adds capability or moves inference-time selection into training.

3.What criterion selected the post-training checkpoint near step 160? The appendix curves diverge shortly after, with relative energy exceeding 15 eV/atom and match rate collapsing. If the stopping point was chosen using the metrics reported in Tables 1 and 2, the RAM numbers are optimistic, and a practitioner would need a selection rule that does not see test performance.

4.How were the throughput numbers in Figure 1 obtained? The NFE comparison is self-evidently in your favor, but the figure's axis is structures per second. Please state the hardware, batch size, and whether baselines were re-run or taken from their papers.

5.Can you give a compute-matched CSP table? Your actual claim is that several draws from OMatG-flash cost less than one from a conventional model and score higher. A table at equal wall-clock would show that directly, rather than placing k=5 rows next to k=1 baselines with a caveat in the text.

1 comment
AuthorsOpenReview ↗

Response to Reviewer zMwX

We thank the reviewer for their thoughtful assessment of our work. Below we respond to some of their questions and concerns raised by the reviewer.

Relaxations

We agree that our claim of high RMSE being more innocuous than it appears is best substantiated by numerical evidence. We compare OMatG-flash (NFE=16) with Crystalite and OMatG on the crystal structure prediction (CSP) task using one-to-one match rate. The configurations corresponding to Crystalite and OMatG are obtained using their published MP-20 CSP checkpoints. We relax each structure and each MP-20 reference for up to 500 steps using MACE with the medium-mpa-0 checkpoint to a tight force tolerance of 0.01 eV / Å so that all structures are minimized with respect to a shared potential energy surface. For reference, we include a column where our model and the baselines are scored using as-generated structures on the MACE-relaxed MP-20 targets.

ModelNFERMSE (relaxed)RMSE (generated)
OMatG-flash160.06950.1178
OMatG-flash-RAM160.05370.0886
OMatG Linear-ODE2090.04400.0727
Crystalite4000.02900.0356

Relaxation narrows the gap substantially between OMatG-flash and the Crystalite and OMatG baselines. It does not completely erase the error in structural accuracy incurred by our use of a flow map but we highlight that the discrepancy between OMatG-flash and baselines decreases dramatically for both the base model and the reinforced model.

RAM vs. reranked best-of-kk

Here we report the performance of OMatG-flash-RAM in comparison to the base OMatG-flash model for $k=1,3,5$. We use the polymorph split reference set and checkpoint for this experiment. In order to use the same 9046 structure pool for each configuration, we rank the $k$ OMatG-flash base model samples using MACE single point calculations.

ModelMETRecRMSE
OMatG-flash (k=1)68.40%0.2525
OMatG-flash (k=3)71.09%0.2189
OMatG-flash (k=5)71.78%0.2096
OMatG-flash-RAM71.10%0.2206

The reinforced model is on par with $k=3$ inference but appears to not exceed MACE-rerenked $k=5$ inference for the CSP task. We point out, however, that lower energy and CSP performance are not equivalent. Match rates and RMSE are fundamentally structure matching tasks and the energy minimization post-training objective, while enhancing performance on a matched $k=1$ budget for one-to-one match rates and METRe, is only a proxy for structure matching quality and has a greater impact on generating structures which relax to lower relative energies. We encourage the reviewer to refer to our response to whKH for further discussion on the energetics of structures produced by the OMatG-flash base model vs. OMatG-flash-RAM. We feel that these results reinforce that post-training is working as intended as we find that the post-trained model produces structures that relax to lower energies. Furthermore, in that response, we do explore post-training for the de novo generation (DNG) task and clearly demonstrate that, using a stability and novelty guided reward, OMatG-flash-RAM obtains convincing gains in S.U.N.+M.S.U.N. with respect to the base model.

Checkpoint selection

In order to choose an OMatG-flash-RAM checkpoint for production we used the one which achieved the lowest relative energy per atom (validation), relying on no test-set metrics. In Figure 3 we show how this evolved during post-training. The selected checkpoint corresponds to step 320 on the right of Figure3. As briefly discussed in Section 3.2, the reward is proportional to the negative energy as computed by MACE medium-mpa-0.

Throughput Computation

We kindly point the reviewer to section F.3 of our appendix where we provide clearer details on how the numbers in Figure 1 were obtained. For competing models we use the numbers as reported in Table 2 of the Crystalite paper (arxiv.org/abs/2604.02270). We carried out our throughput calculation by following the same recipe whereby we use the largest possible batch size which fits on a single NVIDIA H100 GPU.

Compute-matched CSP

We agree that a more rigorous, compute-matched CSP result could be given for OMatG-flash. Below we perform $k$ inference under a time budget set by Crystalite. We perform $k=1$ inference for Crystalite and use that time budget to run $k$ inference, rank the best match, and compute match rate on the top-ranked samples from OMatG-flash. We include ranking time in our accounting of total time for fairness and use a matched batch size of 512 for both models.

ModelkMatch rateRMSETotal time
OMatG-flash162984.92%0.07111028.6 s
RAM (relax)2984.11%0.06101028.6 s
Crystalite172.10%0.03411031.4 s

Comments on the paper

1 comment

Dear Reviewers,

Please carefully review the authors’ rebuttal and engage in further discussion with the authors to clarify any remaining concerns and questions. Your active participation during the discussion period is greatly appreciated.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.