(389h) Equivariant Latent Diffusion for the Generative Design of Oxygen Evolution Reaction Catalysts

The development of highly efficient and stable electrocatalysts is a key challenge in advancing electrochemical water splitting, which plays a central role in renewable energy storage and hydrogen production. Oxide–based electrocatalysts have attracted considerable attention due to their excellent oxidative stability and performance. However, their optimization remains difficult because of the vast and complex structural design space that involves diverse compositions and intricate surface–adsorbate interactions. With the growing application of machine learning (ML) in this domain, the Open Catalyst Project (OCP) has released large-scale DFT datasets—OC20 [1] and OC22 [2]—to support data-driven catalyst research. These datasets contain millions of bulk–surface–adsorbate configurations along with associated energy and force labels, and have become widely used benchmarks for ML model training and evaluation. Based on these datasets, state-of-the-art models such as GemNet [3], eSCN [4], EScAIP [5], and EquiformerV2 [6] have achieved high accuracy in predicting total energy and atomic forces. In particular, EquiformerV2, an E(n)-equivariant transformer, has demonstrated strong performance in modeling 3D atomic structures by preserving symmetry and distance information, establishing new benchmarks on OC20/OC22. These models are primarily optimized for property prediction tasks and are not directly applicable to atomic structure generation. However, recent advances in backbone denoising models for crystalline systems suggest that such architecture may serve as effective foundations for generative tasks, enabling the exploration of atomic-scale design spaces beyond regression [7, 8].

In this work, we propose a conditional generative model that aims to generate full structures of oxygen evolution reaction (OER) catalysts—encompassing bulk, surface, and adsorbate components—conditioned on the type of adsorbate. The model combines a pretrained EquiformerV2 backbone with a score-based diffusion head. The backbone is used in a frozen or partially fine-tuned configuration to embed structural features into a latent space. A low–dimensional condition embedding, derived from the input adsorbate type, is concatenated with the latent features and passed to the diffusion head, which predicts denoising directions from perturbed coordinates using a score-based generative modeling approach. The model is trained primarily on the OC20 dataset, which primarily includes metallic catalyst systems, and further augmented with OC22 to enhance the generation capacity for OER structures. Generated samples are evaluated based on geometric validity metrics such as interatomic distances and atomic overlap, along with diversity analysis and qualitative visualization. We expect the model to achieve a geometric validity rate above 80%, defined by physically plausible interatomic distances and minimal atomic overlap. On a per-condition basis, the generator is anticipated to yield 10-30 structurally diverse and chemically reasonable configurations, enabling broader exploration of the catalyst design space. As a future extension, the model can be integrated with downstream property predictors—including overpotential predictor and Pourbaix diagram estimators—to move beyond structure generation toward fully autonomous catalyst design. This study explores the feasibility and scalability of conditional generative models in complex heterogeneous catalytic systems and lays the groundwork for future dual-purpose architectures that unify structure generation and property prediction within a single framework.

References

1. Chanussot, L., et al., Open catalyst 2020 (OC20) dataset and community challenges.Acs Catalysis, 2021. 11(10): p. 6059-6072.
2. Tran, R., et al., The Open Catalyst 2022 (OC22) dataset and challenges for oxide electrocatalysts.ACS Catalysis, 2023. 13(5): p. 3066-3084.
3. Gasteiger, J., et al., GemNet-OC: developing graph neural networks for large and diverse molecular simulation datasets.arXiv preprint arXiv:2204.02782, 2022.
4. Passaro, S. and C.L. Zitnick. Reducing SO (3) convolutions to SO (2) for efficient equivariant GNNs. in International conference on machine learning. 2023. PMLR.
5. Qu, E. and A. Krishnapriyan, The importance of being scalable: Improving the speed and accuracy of neural network interatomic potentials across chemical domains.Advances in Neural Information Processing Systems, 2024. 37: p. 139030-139053.
6. Liao, Y.-L., et al., Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations.arXiv preprint arXiv:2306.12059, 2023.
7. Zeni, C., et al., A generative model for inorganic materials design.Nature, 2025: p. 1-3.
8. Takahara, I., K. Shibata, and T. Mizoguchi, Generative Inverse Design of Crystal Structures via Diffusion Models with Transformers.arXiv preprint arXiv:2406.09263, 2024.