Supervised and Unsupervised Learning in Atomistic Simulations of Materials Under Reactive Conditions

Knowledge of the equation of state and chemical kinetics of materials under reactive conditions is needed for a wide number of research areas, including studies of planetary interiors, astrobiology, and candidate materials for hydrogen storage. The characterization of these systems under dynamic conditions is often nontrivial, especially for disordered or amorphous phases in which the underlying symmetry of the atomic geometries can be difficult to determine. Here, we focus specifically on two of our efforts to create computational tools to address these problems, namely: (1) the Chebyshev Interaction Model for Efficient Simulation (ChIMES), which is a machine-learned molecular dynamics potential based on linear combinations of many-body Chebyshev polynomials, and (2) the Scalar Graph Order Parameter (SGOP), which characterizes atomic structures based on their graph topologies. These methods will be discussed in the context of molecular dynamics simulations of chemical decomposition under high pressures, Δ-learning with semi-empirical quantum models to yield hybrid functional and coupled cluster accuracy with orders of magnitude improvement in computational cost, and the characterization of multi-element materials over a broad range of pressures and temperatures. Our methods provide a way to conduct computationally efficient and highly accurate simulations over varying conditions, where physical and chemical properties can be difficult to interrogate directly and there is historically a significant reliance on theoretical approaches for interpretation and validation of experimental results.