A generalized method for refining and selecting random crystal structures using graph theory

Random generation of crystal structures is a key to the success of predicting unknown crystals. In this work, we introduce a general method for refining and selecting random structures that relies on minimal prior information. The method establishes a quotient graph from the random structure using a near-neighbor finding algorithm, which subsequently guides the refinement of the initial structure. To validate this approach, we apply it to nine distinct systems, and the outcomes indicate that it effectively yields a great number of low-energy structures. This technique could be integrated into most structure prediction algorithms to generate more sound initial structures, thereby expediting the search for ground state structures.