(Co-)Crystal Structure Prediction with Machine Learned Potentials

Molecular crystals are a class of materials that are held together by weak van der Waals interactions that have applications in pharmaceuticals, organic electronics, and energetic materials. Energetic materials (EMs) include explosives, propellants, and pyrotechnics. Because experimental screening methods can be costly, and in some cases hazardous, computational screening methods such as crystal structure prediction are crucial to the development of the field. Co-crystals are of particular interest in materials discovery and development due to the possibility of enhancing desirable properties from one or more of the co-formers while discarding less desirable properties. Due to the enormous search space inherent in co-crystal structure prediction, it is imperative to develop search algorithms capable of both exploration of the potential energy surface and exploitation of low energy regions. The random structure generator, Genarris, and the genetic algorithm, GAtor, have been shown to successfully predict the structure of single component crystal structures, including several EMs. These two algorithms have been adapted to search for co-crystals. Another component of co-crystal and crystal structure prediction is the computational expense to optimize and rank large systems. This is a rate limiting factor in co-crystal structure prediction. Hence, there has been interest in developing and using machine learned interatomic potentials, such as AIMNet, in crystal structure prediction workflows. We will benchmark the performance of AIMNet in comparison to density functional theory as well as use a trained AIMNet potential to perform CSP for a target of the most recent CCDC blind test.