Training Accurate and Physically Meaningful Machine Learning Force Fields for Water and Understanding Their Transferability
Tristan Maxson, University of Alabama
Water is an important material relevant to biology, chemistry, and catalysis among others, however, the accurate simulation of water is challenging because it requires long molecular dynamics simulations to equilibrate the system due to the complex interactions between the molecules. The MB-Pol potential has been developed to reproduce CCSD(T) quality results, but at a lower cost than full quantum mechanical simulations and allows to reproduce the ground truth for physical properties (e.g. density, compressibility, radial distribution function, etc.) which are typically inaccessible using quantum mechanical calculations due to long equilibration times (10 to 100 ns). Machine learning force fields (MLFFs) serve as an alternative to accelerate atomistic simulations by learning the complex set of interactions from existing data. We show that the Allegro extension of NequIP can accurately learn from the liquid phase of H2O. The transferability of MLFF models trained is also explored to determine what information is crucial to the resulting MLFF. Importantly, we find that our MLFFs can describe physically meaningful many-body interactions in gas phase that are not directly trained on and allow even faster simulations than MB-pol. In addition, the methods used in this work for developing training sets are scalable on HPC systems and demonstrate the importance of validation via metrics beyond simple energy, force, and stress errors.
Abstract Author(s): Tristan Maxson