Automated generation of machine learning-based atomistic potentials for extreme conditions

Neural Network (NN) interatomic potentials are a powerful tool for atomistic scale simulations, combining the generality and accuracy of ab-initio methods with costs approaching those of classical potentials. A robust training dataset covering many atomic configurations must be computed with ab-initio methods to train an accurate NN potential. Recently, active learning (AL) algorithms have demonstrated the ability to generate training datasets quickly and efficiently by selecting atomic configurations for which a NN potential has high uncertainty. This facilitates the generation of training datasets through a minimum number of ab-initio calculations with little or no human intervention. A LAMMPS interface for our NN potential, named ANI, has facilitated large-scale GPU-accelerated MD simulations using domain decomposition. Utilizing this interface, we validate an autonomously generated ANI aluminum potential using both static and dynamic simulated properties, including a partial phase diagram. Additionally, we present million-atom shock simulations of aluminum to illustrate robustness and demonstrate extensibility to the prediction of high-pressure phases.