Automated generation of mechanistic models for chemical process digital twins using reinforcement learning part I: Conceptual framework and equation generation

• Proposes a methodology for automated development of mechanistic reactor models • A reinforcement learning-agent transforms general balance equations into models • Systems of equations are represented in object-oriented postfix notation • The workflow is demonstrated in in silico and Taylor-Couette reactor case studies • Modeling is accelerated significantly compared to exhaustive search approaches Deriving versatile and robust mechanistic models from experimental data is a key challenge in engineering and natural sciences. This is especially true in chemical reaction engineering, where reactor manufacturers and operators increasingly pursue the development and maintenance of digital twins that rely on frequent model updates and ask for automation of this modeling process. In this work, we propose an automated workflow that generates accurate mechanistic reactor models from experimental concentration data of a given reactor. At the core of this workflow, a reinforcement learning agent assembles an interpretable reactor model by iteratively simplifying general differential balance equations and fitting the resulting candidate model to experimental data. We demonstrate the performance of our workflow in two case studies. An in silico case study shows that the workflow correctly reconstructs the model underlying a synthetic data set, is robust against noise in the input data, and has favorable scaling properties. The agent accelerates the model derivation process significantly compared to an exhaustive enumerative search. Secondly, an experimental case study is conducted employing a Taylor-Couette prototype reactor. A liquid-phase esterification reaction of (2-bromophenyl)methanol and acetic anhydride was used as a test system. Based on the experimental data, the workflow derives meaningful mechanistic models, with the most accurate model showing a normalized root mean squared error of 2.4%. Future work encompasses the integration of automated experiments into the workflow and the transfer of our workflow to process units beyond chemical reactors.