Minimizing the training dataset for efficient crystal structure prediction by graph neural network

Abstract Machine learning has boosted the remarkable development of crystal structure prediction (CSP), greatly accelerating modern materials design.The state-of-the-art methods, however, rely on a giant training dataset, and there is a trend to keep increasing the dataset, as triggered by the great success of artificial intelligence models.Here, we show that a giant training dataset is not really necessary.Only around 100 carefully selected training samples for a target composition is already sufficient for highly accurate CSP, which is demonstrated in our newly developed CSP framework named SCCOP.This achievement stems from: (i) graph-deep-learning-based potential energy surface (PES) slicing, which classifies structures into different prototypes; and (ii) a multi-start optimization algorithm to divide-and-conquer the PES. The accuracy and efficiency of our framework is validated on 27 typical compounds.The powerful capability of SCCOP is further demonstrated in two challenging problems: searching for the lowest-energy structures in boron allotropes and determining the structure of ordered vacancy compound CuIn5Se8.Our method discovers several new boron allotropes, and identifies the puzzling crystal structures of the ordered vacancy compound CuIn5Se8.This study establishes a new paradigm for CSP and offers a highly efficient and generally applicable CSP framework.