Prediction of the crystal structure of avadomide using machine learning potentials and crystal engineering

Abstract

Today, machine learning (ML) and crystal structure prediction (CSP) are principal tools in computational materials discovery. In this study, CSP of the pharmaceutical compound avadomide is presented, utilizing a workflow that combines the evolutionary algorithm USPEX, machine learning potentials in moment tensor potential (MTP) formulation, and crystal engineering concepts. Our study begins with the optimization of stable molecular conformations of avadomide using DFT, followed by the generation of likely crystal structures with the evolutionary algorithm USPEX. The optimization and ranking of structures based on ML potentials—trained on the subset of DFT data—were further refined through iterative improvements via active learning. The ML potentials employed in this study are currently constrained to local atomic environments and primarily captured short-range interatomic interactions. To address this limitation, the top-ranked structures were analyzed using crystal engineering concepts and found to contain synthons similar to those in experimental structures of related compounds in the CSD database. Using the synthon approach, two potential crystal structures were proposed for avadomide.