Mutual information for explainable deep learning of multiscale systems

Timely completion of design cycles for complex systems ranging from consumer electronics to hypersonic vehicles relies on rapid simulation-based prototyping. The latter typically involves high-dimensional spaces possibly correlated control variables (CVs) and quantities interest (QoIs) with non-Gaussian multimodal distributions. We develop a model-agnostic, moment-independent global sensitivity analysis (GSA) that differential mutual information rank the effects CVs QoIs. data requirements this information-theoretic approach GSA are met by replacing computationally intensive components physics-based model deep neural network surrogate. Subsequently, is used explain surrogate predictions, surrogate-driven deployed as an uncertainty quantification emulator close loops. Viewed method interrogating surrogate, framework compatible wide variety black-box models. demonstrate provides useful distinguishable rankings via validation step applications in energy storage. Consequently, our “outer loop” accelerated product identifying most least sensitive input directions performing subsequent optimization over appropriately reduced parameter subspaces.