Abstract
Mathematical models are paramount to the field of reaction engineering, facilitating reaction mechanism discovery, process optimisation, and informed decision making in academic and industrial settings. Nevertheless, the development of precise mechanistic reaction rate models remains experimentally intensive, requires expert knowledge, and is susceptible to the introduction of structural bias. Similarly, the identification of a suitable reaction network that depicts all chemical transformations remains a non-trivial task, with existing techniques often being ill-suited for large and complex systems, hence limiting their scalability and implementation within chemical and biochemical applications. This work develops a two-stage autonomous framework that exploits non-linear sparse optimisation to identify the minimum size global reaction network representative of the system under study, and subsequently proposes and discriminates between interpretable rate equations developed through symbolic regression (SR). The generated SR expressions are constrained to a mechanistically meaningful form through the use of a novel substructure decomposition strategy, largely reducing the search space that must be explored and increasing model interpretability. The framework is evaluated on two case studies; the first, depicting a catalytic methanol synthesis reaction network, and the second, an enzymatic kinetic resolution network. The methodology exhibited high levels of accuracy, scalability, data efficiency, and robust network identification consistent with known physics. Lastly, the potential of augmented intelligence, is discussed as a method to enhance fidelity. Therefore, this work represents a key step toward autonomous process modelling and digitalisation in reaction engineering, providing a foundation for accelerated design and development of chemical and biochemical processes.