Searching for optimal process routes: A reinforcement learning approach

Abstract

Developing optimisation tools is a key target in supporting computer-aided process design as the complexity of the designed space grows beyond conventional unit operations. A process design problem can be formulated as a search of an optimal processing route in the thermodynamic state space, going from feedstock to products. This paper describes a design architecture that enables reinforcement learning agent to use trial-and-error to narrow its search to the most promising routes, rather than exhaustively enumerating solutions. In each iteration, the agent employs previously collected data to guide the search for new trajectories. This is successfully demonstrated in a hydrogen production process using both conventional and intensified process design principles. The agent outperformed standard nonlinear optimisation methods in competitive computational time. Limitations and future work are discussed.