Inpainting Crystal Structure Generations with Score-Based Denoising

Abstract

Searching for the optimal atomic position of additive atoms in a given host structure is crucial in designing materials with intercalation chemistry for energy storage. In this study, we present an application of the SE(3)-equivariant diffusion model for such conditional crystal structure predictions using inpainting methods. The model, built upon the \verb|e3nn| framework, was pre-trained on the Materials Project structure database via denoising score matching. By solving the reverse stochastic differential equation using the predictor-corrector method, the model is capable of \textit{de novo} crystal generation as well as conditional generation -- finding atomic sites of additive atoms within a host structure. We benchmarked the model performance on the WBM dataset and showcased examples of ion intercalation in different \ce{MnO2} polymorphs. This efficient, probabilistic site-finding tool offers the potential for accelerating the materials discovery.