A coupled reinforcement learning and IDAES process modeling framework for automated conceptual design of energy and chemical systems

Abstract

Computer-aided process engineering and conceptual design in energy and chemical engineering has played a critical role for decades. Conventional computer-aided process and system design generally starts with process flowsheets that have been developed through experience, which often relies heavily on subject matter expertise. These widely applied approaches require significant human effort, either providing the initially drafted flowsheet, alternative connections, or a set of well-defined heuristics. These requirements not only limit the flexibility of the flowsheet design process, but also make the system design highly reliant on the engineer's experiences and expertise. In this study, a novel reinforcement learning (RL) based automated system for conceptual design is introduced and demonstrated on the Institute for the Design of Advanced Energy System (IDAES) Integrated Platform. IDAES is an open-source platform with extensible libraries of dynamic unit operations and thermophysical property models. It provides the capability of optimizing energy and chemical process flowsheets with state-of-the-art solvers and solution techniques. The RL approach provides a generic tool for identifying process configurations and significantly decreases the dependence on human intelligence for energy and chemical systems conceptual design. An artificial intelligence (AI) agent performs the conceptual design by automatically deciding which process-units are necessary for the desired system, picking the process-units from the candidate process-units pool, connecting them together, and optimizing the operation of the system for the user-defined system performance targets solely according to the reward system, while the reward system can incorporate user's experiences and knowledge to advance the training process. The AI agent automatically interacts with the physics-based system-level modeling and simulation toolset IDAES to guarantee the system design is physically consistent. This study showcases the application of the RL-IDAES framework through two demonstration cases. These cases prove the framework's capability of designing and optimizing complicated systems with high flexibility at affordable computing costs. To illustrate, designing the hydrodealkylation of toluene system from 14 candidate process-units yielded 123 feasible designs within 20 hours on a standard PC. Moreover, the framework's versatility is demonstrated by the ability to transfer a trained RL model to different training cases, thus enhancing the overall performance of the reinforcement learning process.