Chapter 3: Hybrid Model Construction

Hybrid modelling combines data-driven and mechanistic modelling, providing a cost-effective solution to modelling complex chemical and biochemical reaction kinetics when working with an incomplete understanding of the underlying mechanisms. At this chapter’s core is an illustrative case study demonstrating how to build a hybrid model for dynamic simulation. Two prominent structures for combining data-driven and mechanistic models are introduced, namely the discrepancy and embedded structures. For both designs, a step-by-step explanation is given for estimating the parameters from process data and correlating them with the current state and operating conditions. A procedure for robust uncertainty estimation and propagation is also explained. Throughout, potential over-parameterisation and over-fitting pitfalls are highlighted, and the nuances of building either structure are illustrated and compared in depth. The case study culminates in a comparison of the accuracy and uncertainty of the two models, tying back to decisions made during parameter estimation. While such conclusions are case-specific, reasoning model performance in terms of the nonlinearity expected of the data-driven component provides a valuable frame for understanding the challenges of hybrid modelling that arise from imperfect data and an incomplete process understanding. This cements the necessary theoretical background for later chapters.