A Bayesian Inference-Based Approach to Empirical Training of Strongly Coupled Constituent Models

Abstract

Partitioned analysis enables numerical representation of complex systems through the coupling of smaller, simpler constituent models, each representing a different phenomenon, domain, scale, or functional component. Through this coupling, inputs and outputs of constituent models are exchanged in an iterative manner until a converged solution satisfies all constituents. In practical applications, numerical models may not be available for all constituents due to lack of understanding of the behavior of a constituent and the inability to conduct separate-effect experiments to investigate the behavior of the constituent in an isolated manner. In such cases, empirical representations of missing constituents have the opportunity to be inferred using integral-effect experiments, which capture the behavior of the system as a whole. Herein, we propose a Bayesian inference-based approach to estimate missing constituent models from available integral-effect experiments. Significance of this novel approach is demonstrated through the inference of a material plasticity constituent integrated with a finite element model to enable efficient multiscale elasto-plastic simulations.