LoreX: A Low-Energy Region Explorer Boosts Efficient Crystal Structure Prediction
Chuan-Nan Li, Han-Pu Liang, Siyuan Xu, Haochen Wang, B. Zhao, Jingxiu Yang, Xie Zhang, Zijing Lin, Su‐Huai Wei
Abstract
Machine learning has boosted the remarkable development of crystal structure prediction (CSP), greatly accelerating modern materials design. However, slow location of the low-energy regions on the potential energy surface (PES) is still a key bottleneck for the overall search efficiency. Here, we develop a low-energy region explorer (LoreX) to rapidly locate low-energy regions on the PES. This achievement stems from graph-deep-learning-based PES slicing, which classifies structures into different prototypes to divide and conquer the PES. The accuracy and efficiency of LoreX are validated on 27 typical compounds, showing that it correctly locates low-energy regions with only 100 selected samples. The powerful capability of LoreX is demonstrated in solving two challenging problems: discovering new boron allotropes and identifying the puzzling crystal structures of the ordered vacancy compound CuIn 5 Se 8 . This study establishes a new method for rapid PES exploration and offers a highly efficient and generally applicable approach to accelerating CSP.