Machine learning assisted prediction of organic salt structure properties

This paper presents a machine learning approach for predicting the properties of crystal structures after relaxation, based on their unrelaxed structures. Crystal graph singular values significantly reduce the number of features needed to describe a crystal. Machine learning models, using these singular values, predict volume, enthalpy per atom, and phase (metal vs. semiconductor/insulator) of DFT-relaxed organic salt crystals. Initial models were trained on 89,949 structures of 1,3,5-triazine and HCl salts. The model's ability to extend to related chemical systems (isomers, pyridine, thiophene, piperidine) was demonstrated by adding 2000-10,000 structures from each system. These models enable rapid screening of large sets of organic salt crystal structures to identify experimentally realizable candidates, serving as a filtering step in crystal structure prediction (CSP) workflows.

Ethan P. Shapera, Dejan-Krešimir Bučar, Rohit P. Prasankumar, Christoph Heil