Accelerated Structure Prediction of Organic Molecular Crystals

Organic molecular solids form key components of many common goods including medicines, fertilizers, and paints. Organic molecular crystals have been proposed as high-performance semiconductors and optoelectronics. The properties of the organic molecular crystals depend strongly on both the choice of constituent molecules and the crystal structure. Controlling and predicting the crystal structure often proves to be a main impediment toward application. Current approaches to crystal structure prediction rely on generating large numbers of structures. This is computationally expensive because it often requires performing calculations on 10,000's or 100,000's of structures, most of which are not experimentally realizable. Here we accelerate the approach by constructing machine learning models to predict the properties of relaxed organic molecular crystal structures using only knowledge of the generated unrelaxed crystal structures. We demonstrate and validate our approach on organic salts formed from small ring molecules. The constructed models are able to both interpolate within the same chemical systems and provide limited extrapolative predictions to unseen chemical compositions.