Manufacturing Machine Learned Interatomic Potentials for Shock Physics

The ability to reliably generate efficient interatomic potentials capable of accurately simulating novel materials under non-equilibrium conditions would be a major advancement for materials science. Machine learning has facilitated the generation of such potentials in a variety of individual studies with excellent agreement to both experiment and other theoretical methods. To generate a neural network potential, quantum calculations determine energies and forces on a training set of atomic configurations generated through active learning. This dataset is then used to train a Neural Network (NN) to predict an energy as a function of atomic coordinates for a given material. These NNs are useful for performing MD simulations and computing physical properties. This presentation will focus on the consistency and reproducibility of this methodology by generating potentials for a variety of materials utilizing the same protocol. Generated potentials will be validated through the computation of bulk properties and simulated phase diagrams. Finally, shock simulations were performed with these potentials to demonstrate their ability to perform under non-equilibrium conditions.