An Adapted Similarity Kernel and Generalized Convex Hull for Molecular Crystal Structure Prediction

We adapted an existing approach to identifying stabilizable crystal structures from prediction setsthe Generalized Convex Hull (GCH)to improve its application to molecular crystal structures. This was achieved by modifying the Smooth Overlap of Atomic Positions (SOAP) kernel to define the similarity of molecular crystal structures in a more physically motivated way. The use of the adapted similarity kernel was assessed for several organic molecular crystal landscapes, demonstrating improved interpretability of the resulting machine-learned descriptors. We also demonstrate that the adapted kernel results in improved performance in predicting lattice energies using Gaussian process regression. Our overall findings highlight a sensitivity of similarity kernel-based landscape analysis methods to kernel construction, which should be considered when applying these methods.