Generative model for large-cell crystal structure

Abstract The advancement of next-generation technologies depends on the accelerated discovery of novel materials through efficient in silico design strategies. As search for simple compounds have been saturated, large-cell crystal structures (LCS) become promising candidates for next-generation technologies across diverse applications fields. However, designing stable LCS with crystallographic symmetry remains a major challenge for generative models. We present GenLCS, a two-step diffusion-based framework that generates LCS by separately modeling discrete structural components and atomic coordinates with considering both symmetry and interatomic interactions. GenLCS is trained on a new benchmark dataset, MOA-over20, and outperforms existing models in generating novel, symmetric, and stable structures. Moreover, it successfully generates structures beyond the training distribution, including those with over 120 atoms per unit cell. Finally, several stable ternary LCS are presented from the generated structures. GenLCS offers a robust solution for symmetry-aware generative modeling of complex crystal structures, advancing the capabilities of inverse design in materials discovery.