Topology-based crystal structure generator

Crystal structure prediction is a new dynamically developing field. Two main types of crystal structure prediction methods exist: (1) based on global optimization and (2) based on data mining. It seems promising to hybridize data mining and global optimization techniques. The former generally involve no empirical information and are truly predictive, the latter rely on the databases of existing crystal structures, and are relatively fast, but prone to error, because the databases are far from complete. Furthermore, the theorist's dream is to be as little dependent on empirical data as possible and be always capable of predicting new structures, not contained in the databases. Here we present an approach to generate an infinite number of crystal structures from a finite set of idealized periodic nets. The resulting structures are highly ordered, possess nontrivial symmetries, and often low energy. Topologically generated structures can be used for initializing evolutionary crystal structure prediction calculations, or on their own - as an extended data mining approach. The efficiency of the proposed approach in both scenarios is confirmed by a series of tests, which we also present here. As an additional enhancement to evolutionary algorithm we introduce a technique for adjusting fractions of variation operators on the fly. Tests show significant performance improvements due to both of these developments. (C) 2018 Elsevier B.V. All rights reserved.