Incorporating Unmodeled Dynamics Into First-Principles Models Through Machine Learning

First-principles modeling of dynamical systems is a cornerstone of science and engineering and has enabled rapid development and improvement of key technologies such as chemical reactors, electrical circuits, and communication networks. In various disciplines, scientists structure the available domain knowledge into a system of differential equations. When designed, calibrated, and validated appropriately, these equations are used to analyze and predict the dynamics of the system. However, perfect knowledge is usually not accessible in real-world problems. The incorporated knowledge thus is a simplification of the real system and is limited by the underlying assumptions. This limits the extent to which the model reflects reality. The resulting lack of predictive power severely hampers the application potential of such models. Here we introduce a framework that incorporates machine learning into existing first-principles modeling. The machine learning model fills in the knowledge gaps of the first-principles model, capturing the unmodeled dynamics and thus improving the representativeness of the model. Moreover, we show that this approach lowers the data requirements, both in quantity and quality, and improves the generalization ability in comparison with a purely data-driven approach. This approach can be applied to any first-principles model with sufficient data available and has tremendous potential in many fields.