Examining Uncertainty Quantification Techniques for Machine Learned Interatomic Potentials

Machine learned interatomic potentials offer a fast and accurate approximation of energy and forces for atomistic simulations. When deploying ML IAPs in forward atomistic simulations that can leverage a hybrid physics-based and data-driven approach, knowledge of the uncertainty in the surrogate model prediction enables the ability to dynamically switch between traditional and ML approaches, ensuring preserved accuracy. In this talk, we will explore and examine several different methods of building in uncertainty quantification in machine learned interatomic potentials using probabilistic neural network potentials. Monte Carlo Dropout, mixture density networks, and other techniques will be discussed and applied to ML IAPs.