Uncertainty-Quantification-Driven Autonomous Workflows for Upscaling of Complex Materials Properties

Many engineering models for require parametric or functional inputs that can in principle be obtained from lower scale simulations. For example, transport coefficients for radiation-induced defects computed from molecular dynamics can be used to inform kinetic Monte Carlo models, that can themselves inform cluster-dynamics simulation of microstructural evolution. However, in many cases, the number of lower-scale calculations required to obtain these higher-scale properties can be very large, which can lead to extremely long times-to-solution, especially when human intervention is needed at any step of the process. We demonstrate how tailored uncertainty-quantification approaches can be used to autonomously drive the execution of upscaling workflows at large computational scales. I will show how information can be systematically upscaled into different representations in order to develop reliable reduced-order models from simulation data.