(735ae) Accelerating Simulation of Materials with Machine Learning

A new generation of advanced materials will be required for energy sustainability and quantum technology, and computation aids materials search and investigation. Molecular dynamics (MD) simulation, in particular, allows us to calculate transport properties and investigate dynamic behavior, which are difficult to probe experimentally. Challenges of MD simulation include the trade-off between accuracy and computational expense, analyzing simulation data, and uncertainty in calculated properties. My research uses machine learning to address these challenges. The trade-off between accuracy and computational expense motivates the ML potential, commonly parameterized by neural networks (NN), because they can reach DFT level accuracy at orders of magnitude faster. One system of interest is liquid alloys, and I previously trained NN potentials for liquid Ni-Al-W and integrated them with MD simulations. The training process required an iterative build-up of the training dataset with human feedback, because NNs commonly extrapolate outside their training regime, motivating an uncertainty quantification method. We develop and show that the delta method for NN potentials identifies data-points where the model extrapolates. We also describe different sources of uncertainty and difficulty of obtaining uncertainties of physical properties, especially derivatives and equilibrium properties. We obtain uncertainties of equilibrium lattice constants and bulk moduli, and show that Frequentist and Bayesian regression uncertainty are specific to model choice while Gaussian process uncertainty is model general/agnostic. My work will allow robust, automated investigation of dynamics properties at scale.