Machine-Learnt Fragment-Based Energies for Crystal Structure Prediction

Crystal structure prediction involves a search of a complex configurational space for local minima corresponding to stable crystal structures, which can be performed efficiently using atom-atom force fields for the assessment of intermolecular interactions. However, for challenging systems, the limitations in the accuracy of force fields prevents a reliable assessment of the relative thermodynamic stability of potential structures. Here we present a method to rapidly improve force field lattice energies by correcting two-body interactions with a higher level of theory in a fragment-based approach, and predicting these corrections with machine learning. We find corrected lattice energies with commonly used density functionals and second order perturbation theory (MP2) all significantly improve the ranking of experimentally known polymorphs where the rigid molecule model is applicable. The relative lattice energies of known polymorphs are also found to systematically improve towards experimentally determined values and more comprehensive energy models when using MP2 corrections, despite remaining at the force field geometry. Predicting two-body interactions with atom-centered symmetry functions in a Gaussian process is found to give highly accurate results with as little as 10-20% of the training data, reducing the cost of the energy correction by up to an order of magnitude. The machine learning approach opens up the possibility of using fragment-based methods to a greater degree in crystal structure prediction, providing alternative energy models where standard approaches are insufficient.