Accelerating Crystal Structure Prediction Using a Machine-Learned Force-Field for Drug-like Molecular Crystals

Molecular crystal structure prediction requires a very large amount of computational resources to find usefully-accurate structures. Recent progress in using equivariant neural networks (machine-learned force fields") trained on electronic structure calculations can drastically speed up these calculations needed for finding low-energy crystal polymorphs. Making molecular crystal structure prediction tractable in silico would revolutionize the analysis of the manufacturability of drug compounds during late-stage drug discovery. Here, we report on progress we have made in training a machine-learned force field and using it in molecular crystal structure prediction.