CompGen: A Conditional Generation Framework for Inverse Composition Design of Catalytic Surfaces

Shuizhou Chen, Chenghan Sun, ZhiyuanLiu, Andi Han, Ichigaku Takigawa, Quan QIAN

Additional Submission Instructions: For the camera-ready version, please include the author names and affiliations, funding disclosures, and acknowledgements.

Track: Track 1: Original Research/Position/Education/Attention Track

Keywords: Inverse Materials Design, De Novo Generation, Diffusion Model

Abstract: Generating adsorption configurations, that is, how small atoms or molecules bind to complex catalyst surfaces, remains underexplored in inverse materials design. We present CompGen, a conditional generative framework that reformulates 3D structure prediction as a 2D shell-wise composition task centered on the adsorption site. CompGen uses a Chemically Informed Autoencoder (CIAE) to embed sparse compositions into a continuous, periodic table aware latent space learned with a multi-stage pretraining process. A conditional diffusion model then samples in this latent space under multi-physical conditions, including adsorbate identity, adsorption energy, and relevant elements, enabling inverse composition design of catalytic surfaces. Pretrained on a subset of Open Catalyst 2020, CompGen is fine-tuned to more complex high-entropy alloy (HEA) surfaces and achieves strong fine-tuned performance. Extensive experiments show robust zero-shot and few-shot behavior, highlighting CompGen’s effectiveness for data-efficient, domain-transferable inverse design of catalytic surfaces.

Submission Number: 163