Machine learned thermodynamically consistent material models with uncertainty quantification
Please login to view abstract download link
Current constitutive models cannot accurately describe the large variety of novel composites and alloys. Data efficient experimental calibration of new models will be necessary to enable engineering applications of these materials. Uncertainty quantification will also be necessary to identify regimes where learned models are valid and can provide feedback for an optimal experimental design algorithm. In this talk, we demonstrate a machine learning framework for modeling the equation of state (EOS) of materials under strong shocks. In particular, we focus on learning thermodynamically consistent EOS's such that the specific entropy is a convex function of the specific volume and internal energy. This guarantees the stability of shock simulations of materials using our learned EOS's. Using variational inference with normalizing flows and a convexity preserving prior, we obtain UQ aware EOS models. We demonstrate the framework towards learning the EOS for a metal matrix composite from high fidelity shock simulations.
