A Data-Efficient Framework for Automated Identification of Reaction Networks and Interpretable Rate Models

Mathematical models are central to reaction engineering, underpinning mechanism discovery, process optimization, and industrial-scale decision support. Nevertheless, developing mechanistic reaction rate models remains labor-intensive, expert-dependent, and prone to structural bias. This work proposes a two-stage automated framework that identifies the minimal global reaction network via sparse optimization and then generates and discriminates interpretable rate expressions using symbolic regression (SR) with model-based design of experiments. A novel substructure-decomposition strategy is introduced to constrain SR to mechanistically meaningful forms, thereby shrinking the search space and improving interpretability. The framework is evaluated on two reaction systems, methanol synthesis using syngas and an enzymatic network, demonstrating high accuracy, strong data efficiency, and robust physical consistency. Finally, the role of augmented intelligence, incorporating domain knowledge to further enhance fidelity, is also discussed. This study therefore opens a pathway to accelerate knowledge discovery and build physics-grounded digital twins for reaction engineering applications.