(475g) Discovering Interpretable Models from Data Using Machine Learning: Application to Nonlinear Parametric Models

Recent advances in machine learning, powered by the steady progress in hardware and software capabilities as well as the availability of more data, have opened up great opportunities in modeling in numerous domains. However, the lack of transparency of such black-box models is of concern in certain applications, such as in fault diagnosis and control, where the cost of a mistake could be potentially high. Therefore, to build one's confidence one would like models that can be interpreted and explained. Here we present a hybrid machine learning framework that incorporates fundamental physiochemical mechanisms in the discovery of such transparent models. We build on our earlier success with linear models and report our progress on nonlinear parametric models that we typically come across in chemical engineering. Further, we show a strategy to address model identification given sparse data. We demonstrate our system using several examples from process control and reaction kinetics.