acceptodds
Under review as a conference paper at ICLR 2027

OxiGen: Oxidation-State-Aware Crystal Generation

Abstract

Generative models have the potential to accelerate inorganic materials discovery by enabling inverse design, but generating experimentally realisable crystals remains challenging. Oxidation states are widely used to assess the compositional validity of crystals and guide inorganic materials discovery. While existing generative models for crystals can generate materials with charge-neutral oxidation-state assignments, they poorly reproduce the distributions of oxidation states observed in synthesised materials. To address this limitation, we propose OxiGen, an oxidation-state-aware crystal diffusion model that explicitly represents oxidation states during generation. OxiGen enforces global charge neutrality by construction using a structured output layer with exact inference over a finite-state automaton. Empirically, OxiGen substantially improves oxidation-state fidelity, generates the highest rate of stable, unique, and novel crystals among evaluated methods, and maintains high compositional validity even under property conditioning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Crystals, Materials, Diffusion Models, Generative Models, Structured Prediction

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_tFqmOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces OxiGen, a generative diffusion model for inorganic crystalline materials that explicitly represents oxidation states and enforces global charge neutrality by construction. This is achieved through a structured output layer that uses dynamic programming over a finite-state automaton. The authors show that OxiGen's approach substantially improves oxidation-state fidelity and maintains high compositional validity under band-gap conditioning. The rate of generated S.U.N. structures is also slightly increased through the structured sampling, which overall yields very competitive performance of OxiGen in the de novo generation task.

Critical strengths
  • The explicit representation of oxidation states and the structured output layer for exact charge-neutral inference is clearly motivated for generating inorganic crystalline materials, proven to be highly effective, and general enough so that it should be able to be incorporated in other generative models.
Critical weaknesses
  • The authors state that the design of the structured output layer is general enough that it can be incorporated into other generative models for inorganic crystalline materials. To increase the significance of this work, this should be shown explicitly for one example.
  • A comparison against using classifier-free guidance on global charge neutrality is missing.
Questions for authors
  • Within the MatterGen framework, it should be straightforward to use classifier-free guidance on global charge neutrality as an alternative approach to OxiGen's approach. How does the performance compare?
  • Could the authors show for one other generative model that their approach works there too?
  • For the results in Figure 3, what is the rate of (meta-)stability and uniqueness of the generated structures at each target band gap? I assume these metrics are also significantly higher than the ones of MatterGen?
3 comments
AuthorsOpenReview ↗

Response to Reviewer tFqm

Thank you for your helpful comments and suggestions. We address your questions below.

W2+Q1. How does the performance of OxiGen compare to classifier-free guidance on charge neutrality?

In our setting, all structures in the training dataset have charge-neutral oxidation-state assignments. Therefore, the distribution conditioned on charge neutrality would not differ from the unconditional distribution and so standard classifier-free guidance would not provide any conditioning signal. Other soft constraint/guidance approaches could potentially be adapted for charge neutrality, but these generally do not provide the same exact guarantee of charge neutrality as our structured output layer. If we misunderstood the reviewer's comment and another experiment was suggested please let us know and we would be happy to run it.

W1+Q2. Could the authors show for one other generative model that their approach works there too?

We agree that demonstrating our approach works for another generative model would provide stronger evidence for its generalisability. We have implemented our method for OMatG [1] and will report the results once the experiments are complete.

Q3. What is the rate of (meta-)stability and uniqueness of the generated structures at each target band gap?

As reported in Appendix Table 14, OxiGen has a higher S.U.N. rate than MatterGen at all band gap targets, with particularly large improvements at 3-7 eV. These improvements are driven by a large increase in stability, while novelty and uniqueness moderately decrease, consistent with the typical stability-novelty tradeoff.

Table: Band-gap-conditioned generation results for target band gaps of 1, 3, 5, and 7 eV. Results are shown for a single fine-tuning seed, with 1,024 generated structures at each target.

MethodTarget (eV)RMSE (eV) ↓Stable (%) ↑Unique (%) ↑Novel (%) ↑S.U.N. (%) ↑EhullE_{\mathrm{hull}} (eV/atom) ↓RMSD (Å) ↓
MatterGen11.11649.3299.0282.4233.590.18470.1961
MatterGen31.72445.9098.9383.9831.050.15300.2684
MatterGen52.18549.5194.8282.3231.540.15130.3016
MatterGen72.62255.3786.8284.1835.160.14560.3338
OxiGen11.13860.5596.4873.6335.840.14890.1813
OxiGen31.32271.4896.2971.9743.550.08770.1895
OxiGen51.48672.7591.1170.4142.970.08160.2303
OxiGen72.12568.9587.7975.9841.890.08780.2580

References

[1] Philipp Höllmer, Thomas Egg, Maya Martirossyan, Eric Fuemmeler, Zeren Shui, Amit Gupta, Pawan Prakash, Adrian Roitberg, Mingjie Liu, George Karypis, Mark Transtrum, Richard Hennig, Ellad B. Tadmor, and Stefano Martiniani. Open Materials Generation with Stochastic Interpolants. In Proceedings of the 42nd International Conference on Machine Learning, volume 267, pp. 23417–23450. PMLR, July 2025. URL https://proceedings.mlr.press/v267/hollmer25a.html.

AuthorsOpenReview ↗

Additional Experiments for OMatG

As requested, we have implemented our approach for OMatG [1], which uses discrete flow matching [2,3] rather than the masked discrete diffusion process [4] used by OxiGen. As with OxiGen, we replace the element representation with a species representation and apply the structured output layer to the denoiser predictions at each sampling step. The results are summarised in the table below. Similarly to OxiGen, introducing the species representation alone results in a small improvement in oxidation-state fidelity, while structured sampling leads to more substantial improvements, with compositional validity and oxidation-state frequency score closely matching the MP-20-OS reference values. The S.U.N. rate is largely consistent across OMatG variants and lower than for OxiGen. These results demonstrate the benefits of our approach for another generative model with a different discrete process. We will update the manuscript with these additional results. We thank the reviewer again for their suggestions and would be happy to answer any further questions.

Table: De novo generation results for OMatG with the species representation and structured sampling, compared with the baselines from Table 2. Results are reported as mean ± std over three independently seeded training runs for the OMatG species variants and five independently seeded training runs for the other generative models, with 10,240 generated structures per seed. MP-20-OS denotes the test set.

MethodComp. Val. (%) ↑fOS (%) ↑dOS ↓S.U.N. (%) ↑Ehull (eV/atom) ↓RMSD (Å) ↓
MP-20-OS91.7353.78––––
MatterGen81.18±0.9437.89±1.020.1537±0.003924.54±0.910.1799±0.00850.1747±0.0203
OxiGen90.65±0.9453.01±0.960.0911±0.002826.38±1.010.1575±0.00800.1679±0.0117
OMatG82.32±0.4843.08±0.350.1284±0.006019.03±0.780.1831±0.00300.5244±0.0278
OMatG species, unconstrained84.82±0.4145.75±0.780.1123±0.005919.06±0.230.1749±0.00430.5450±0.0078
OMatG species, constrained91.67±0.2156.22±0.370.0835±0.002219.31±0.440.1633±0.00370.5418±0.0097

References

[1] Philipp Höllmer, Thomas Egg, Maya Martirossyan, Eric Fuemmeler, Zeren Shui, Amit Gupta, Pawan Prakash, Adrian Roitberg, Mingjie Liu, George Karypis, Mark Transtrum, Richard Hennig, Ellad B. Tadmor, and Stefano Martiniani. Open Materials Generation with Stochastic Interpolants. In Proceedings of the 42nd International Conference on Machine Learning, volume 267, pp. 23417–23450. PMLR, July 2025. URL https://proceedings.mlr.press/v267/hollmer25a.html.

[2] Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky T. Q. Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete Flow Matching. In Advances in Neural Information Processing Systems, volume 37, pp. 133345–133385. Curran Associates, Inc., 2024. doi: 10.52202/079017-4239. URL https://proceedings.neurips.cc/paper_files/paper/2024/hash/f0d629a734b56a642701bba7bc8bb3ed-Abstract-Conference.html.

[3] Andrew Campbell, Jason Yim, Regina Barzilay, Tom Rainforth, and Tommi Jaakkola. Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design. In Proceedings of the 41st International Conference on Machine Learning, volume 235, pp. 5453–5512. PMLR, July 2024. URL https://proceedings.mlr.press/v235/campbell24a.html.

[4] Subham Sekhar Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin T. Chiu, Alexander Rush, and Volodymyr Kuleshov. Simple and Effective Masked Diffusion Language Models. In Advances in Neural Information Processing Systems, volume 37, pp. 130136–130184. Curran Associates, Inc., 2024. doi: 10.52202/079017-4135. URL https://proceedings.neurips.cc/paper_files/paper/2024/hash/eb0b13cc515724ab8015bc978fdde0ad-Abstract-Conference.html.

Thank you for your replies and clarifications, and for the additional results in the OMatG framework. My concerns have been addressed.

Reviewer_gaGWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • Goal: Inorganic crystal generative model with the goal of reproducing the correct distributions of oxidation states, with goal-directed optimization using classifier-free guidance
    • Authors do this by enforcing global charge neutrality
  • I think this is a really nice idea (we were in fact thinking of something similar in my team, but now the authors have done it perhaps we don't have to) tacking an important gap in inorganic crystal generation, and reflected in the high SUN of structures generated
  • From the ML-perspective, the novelty of the methods (diffusion models with CFD) is not so high, but the application domain is important
Critical strengths
  • Very polished work, tackling an well-defined and important gap in crystalline materials generation.
  • I also really like the evaluation using LeMat-GenBench, as it allows to compare to existing models very easily in a standardized way; I like the comparison to the selected models (I think the right models were chosen) as it helps to assess whether or not a reader would actually want to use it
  • Really nice, clear figures and plots
  • Good the authors included error bars for comparisons
Critical weaknesses
  • Cons:
    • ICSD data not publicly available, requires a license
  • To improve:
    • Documentation in code (not sure it is so easily reproducible from what is provided)
Questions for authors
  • Questions:
    • The fact that charge neutrality has to be enforced during sampling to improve the performance I see as a limitation, makes me wonder how much the model is actually learning during training? Why does it struggle to generalize?
    • Have you assess how far out of distn could you push the models? e.g., could the model still produce something reasonable for very different crystals from the ICSD than those used in training? I saw that a composition-disjoint split was used, but this isn't fully clear to me exactly how different were the held-out structures from training?
    • What are the failure modes observed on the held-out test structures? Are they different from the other methods analyzed (e.g., Crystalite, MatterGen) and if so how do they differ?
    • Is there a specific reason the ICSD was used over other public datasets (maybe datasets with reliable oxidation states do not exist in the public domain, but this is not obvious to me).
2 comments
AuthorsOpenReview ↗

Response to Reviewer gaGW

We thank the reviewer for their positive feedback and their appreciation of the problem as an important gap in inorganic crystal generation.

W1+Q4. Is there a specific reason the ICSD was used over other public datasets?

To our knowledge, the ICSD is the only available dataset containing manually curated oxidation-state labels. Public datasets such as the Materials Project do contain oxidation-state labels for some structures, but these are generally assigned automatically using methods such as pymatgen's BVAnalzyer or oxi_state_guesses, and may therefore contain labelling errors. We only use the ICSD dataset to train and evaluate our oxidation-state prediction model, which is then used to label the publicly available MP-20 dataset with oxidation states. The resulting MP-20-OS dataset is used for training the generative models and is available alongside our code.

Q1. Why does charge neutrality need to be enforced during generation?

A key failure mode that we observe in existing crystal generative models is that they can generate crystals with chemically plausible local coordination environments, but that have no plausible global oxidation-state assignment. This behaviour occurs across baselines with different network architectures, including graph neural networks and transformers, and persists even when we add the explicit oxidation-state representation. This suggests to us that these models learn local chemistry reasonably well, but struggle with global constraints such as charge neutrality. Our structured output layer provides a simple and effective solution to this problem by guaranteeing charge neutrality during generation. However, developing generative models that can better learn global chemistry rather than relying on explicit constraints is an interesting direction for future work.

Q3. What are the failure modes observed and how do they differ from the baseline models?

As discussed above, a key failure mode of the baseline models is the generation of crystals with implausible compositions requiring unusual oxidation states. For example, MatterGen generated Ca₂Ti₂O₇, which is similar to the known perovskite CaTiO₃ (mp-4019), but would require an implausible Ti⁵⁺ cation to charge balance. As shown in the paper, OxiGen substantially reduces these failures and generates oxidation states at a similar frequency to the training data.

However, enforcing charge neutrality does not guarantee chemical validity and we do observe some remaining failure modes. In particular, because oxidation states are generated jointly with the crystal structure, the resulting assignments may disagree with the true assignment inferred from the final structure. As reported in Appendix F.2 (Table 13), OxiGen achieves 100% charge neutrality, with 90.69% site-level and 72.47% crystal-level agreement between the generated oxidation states and those independently inferred from the final structures using the oxidation-state prediction model. In the 27.53% of structures where the oxidation-states disagree, we observe some chemically invalid assignments, such as assignments containing oxidation state zero in ionic compounds (1.2% of generated structures) and unusual mixed-valence assignments involving elements outside the established set of mixed-valence elements (3.9%). In most cases, a plausible alternative oxidation-state assignment exists, but these examples highlight a limitation of explicitly generating oxidation states. We have further highlighted this failure mode in Appendix F.2 of the revised version.

Q2. How far out of distribution can we push the model?

To clarify, the composition-disjoint ICSD split is used to train and evaluate the oxidation-state prediction model, whereas the generative models are trained on MP-20-OS using the original MP-20 splits. We agree that out-of-distribution generation is an important challenge in materials discovery, particularly when targeting properties or combinations of properties that are rare among known crystals. As shown in Figure 3, MatterGen's compositional validity decreases substantially when conditioning on higher, more out-of-distribution band-gap targets, whereas OxiGen maintains consistently high validity. OxiGen also achieves higher S.U.N. rates across all conditioning targets (Appendix Table 14), demonstrating that our approach performs particularly well in this challenging setting compared with the baseline. Conditioning on more challenging property targets or combinations of properties is something we plan to explore further in future work.

W2. Documentation and reproducibility

Thank you for your feedback. We have updated the documentation with clearer instructions on reproducing all our experiments including training and evaluating the oxidation-state prediction model, labelling MP-20 with oxidation states, training and evaluating OxiGen and the ablations, and fine-tuning OxiGen for conditional generation.

Thank you for the thorough response, both to me and the other reviewers. I think the work is solid and impactful. I will maintain my score at 3.

Reviewer_mBqrOpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

Generate neutral crystals composed of charged ions by tweaking MatterGen to manipulate the atom type and it's oxydation state (a,z) at each site (instead of just the atom type).

Critical strengths

Replacing each element aa by a species (a,z)(a,z) is a simple trick that loses nothing, and it turns charge neutrality into a single linear integer constraint.

Critical weaknesses

No critical weakness found

Questions for authors

No question for the authors

1 comment
AuthorsOpenReview ↗

Response to Reviewer mBqr

We thank the reviewer for the positive assessment of the paper's clarity and soundness. We agree that replacing elements aa with species (a,z)(a,z) is a simple change. However, as we show in our ablation study, this change alone does not improve performance as most generated oxidation-state assignments violate charge neutrality. Our main methodological contribution is therefore formulating charge neutrality as a structured inference problem and integrating exact constrained sampling into a crystal diffusion model. This guarantees charge neutrality during generation and substantially improves oxidation-state fidelity, addressing a significant limitation of existing baselines. We would be happy to answer any further questions the reviewer might have.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.