Machine learning assisted prediction of organic salt structure properties

We demonstrate a machine learning-based approach which predicts the properties of crystal structures following relaxation based on the unrelaxed structure. Use of crystal graph singular values reduces the number of features required to describe a crystal by more than an order of magnitude compared to the full crystal graph representation. We construct machine learning models using the crystal graph singular value representations in order to predict the volume, enthalpy per atom, and metal versus semiconductor/insulator phase of DFT-relaxed organic salt crystals based on randomly generated unrelaxed crystal structures. Initial base models are trained to relate 89,949 randomly generated structures of salts formed by varying ratios of 1,3,5-triazine and HCl with the corresponding volumes, enthalpies per atom, and phase of the DFT-relaxed structures. We further demonstrate that the base model is able to be extended to related chemical systems (isomers, pyridine, thiophene and piperidine) with the inclusion of 2000 to 10,000 crystal structures from the additional system. After training a single model with a large number of data points, extension can be done at significantly lower cost. The constructed machine learning models can be used to rapidly screen large sets of randomly generated organic salt crystal structures and efficiently downselect the structures most likely to be experimentally realizable. The models can be used as a stand-alone crystal structure predictor, but may serve CSP efforts best as a filtering step in more sophisticated workflows.