(105c) Discovering First-Principles-Based Models Using Machine Learning for Physicochemical Systems

The outstanding progress of machine learning (ML) in applications such as computer vision, game playing, and natural language processing in recent years has many scientists and engineers excited about the potential opportunities in traditional science and engineering domains. However, there is an essential difference between applying ML in these applications versus physiochemical (and biological) domains. The latter is governed by fundamental laws of physics and chemistry (and biology), which is typically not the case in the former. While purely data-driven machine learning has its immediate uses, the lack of interpretability and explainability of black-box models is a major concern in many applications such as process control and safety where the cost of a mistake can be potentially quite high. Thus, the longterm success of artificial intelligence (AI) in scientific and engineering domains, we believe, would depend on leveraging first-principles knowledge effectively. This requires the careful integration of symbolic AI (i.e., knowledge-based reasoning framework) with numeric AI (i.e., machine learning). Here, we present a data-driven hybrid-AI modeling framework that exploits a priori fundamental knowledge using symbolic AI, while leveraging the statistical advantages of machine learning. Such an approach enables the end-user to gain insights about the underlying mechanism, and can be used to generate causal insights. We demonstrate the efficacy of our approach on applications in reaction kinetics, transport phenomena, and control theory, across various length scales in domains ranging from biological systems to process systems.