Automated generation of mechanistic models for chemical process digital twins using reinforcement learning part II: Compartmentalization and learning-based recalibration

Abstract

Developing predictive models is central to building digital twins for chemical processes, which have a variety of applications in their development and operation. Mechanistic models are highly interpretable and have a larger domain of validity compared to data-driven models, but require significant time and expert knowledge to construct. In this contribution, a workflow for automated mechanistic model generation is extended to handle systems comprised of interdependent, spatially distributed phenomena. The search for accurate models is performed by hierarchically connected reinforcement learning agents. Different ways to incorporate human expertise in model generation are explored, and an ontology is introduced to manage expert and modeling knowledge. The extended workflow is shown to reliably find accurate models of chemical systems, exemplified on a phase transfer catalysis reaction and a Taylor-Couette reactor. For the latter, its non-ideal flow patterns were predicted within a deviation of 5 %, and automatically generated compartmentalization results were found to have comparable physical interpretations to bespoke models from literature. Additionally, the reinforcement learning agents were able to accurately recalibrate models up to twice as fast when drawing upon pre-training under a different operation condition. By generalizing all parts of the automated modeling procedures, we enable the efficient (re-)use of knowledge previously confined to the human modeler. We envision that in the future, the role of experts can be shifted from actively constructing each model iteration to curating knowledge and working collaboratively with autonomous agents.