Abstract

Data-driven modeling of physical systems is a powerful tool for decision-making, however interpreting the relationships learned in black-box, non-linear statistical and machine-learning predictors is fraught with challenges. The use of sensitivity analysis and explainability metrics often is prone to confirmation bias and can be highly variable. We present an approach to leveraging non-linear structural equation models to incorporate subject matter expertise into the data-driven predictive models to increase interpretability and generalizability. This approach is tested in the context of the discrepancy with an expensive computer model and two scientific applications focused on learning physical relationships for physical experimental design.