Neural Network Approaches to Inverse Problems in Chemical Kinetics

Abstract

Chemical kinetics modeling of multicomponent reaction systems has traditionally relied on deterministic and semi-empirical approaches based on the law of mass action and the Langmuir–Hinshelwood formalism. However, such models often require assumptions regarding reaction mechanisms that are difficult to verify experimentally. In contrast, conventional black-box neural networks require large training datasets and frequently exhibit poor extrapolation performance. This review analyzes modern neural network approaches for solving inverse problems in chemical kinetics, including physics-informed neural networks (PINNs), neural ordinary differential equations (Neural ODEs), chemical reaction neural networks (CRNNs), and hybrid constrained architectures. Particular attention is given to their applicability, limitations, and interpretability in small-data conditions. The integration of physicochemical constraints, including mass conservation, stoichiometry, and thermodynamic relationships, directly into neural network architectures is shown to improve the reliability and physical consistency of kinetic models. Scientific machine learning (SciML) approaches are identified as a promising direction for kinetic parameter estimation, reaction network reconstruction, and chemical reactor scale-up.