University of Twente Student Theses


Physics-informed reinforcement learning for process design.

Uijthof, E.M.T. (2023) Physics-informed reinforcement learning for process design.

[img] PDF
Abstract:Recently, reinforcement learning was used to generate sensible flowsheet designs for several applications, including distillation trains and other simplified reaction and separation problems. However, because of a lack of engineering knowledge, these learning agents are unable to discover general physics-related concepts and transfer learned information to new processes, requiring retraining prior to every unseen use-case. This study set out to improve the ability of an agent to learn underlying chemical and physical processes in a simplified distillation separation problem, increasing its applicability across an wider variety of use-cases. To increase the generality of the model, two additions were made to an existing reinforcement learning infrastructure. First, the learning agent was trained on multiple feed-fraction use-cases instead of only on one, switching between feed-fractions at a set frequency. Second, information regarding the processes in the process units and chemical properties of the components in the streams was supplied to the agent in the form of the feed-composition and the boiling points of the components. Passing the engineering knowledge to the agent was achieved by concatenating the property values at the end of the flowsheet fingerprint that described the state of the flowsheet. It was found that, for both implementations, the average model performance was worse than an established benchmark. Yet, on some occasions, the integrated models outperformed the benchmarks and generated flowsheets with a higher model scores. However, currently results are not consistent enough for this type of model to be of value in the flowsheet design process. As direct flowsheet fingerprint concatenation was deemed unreliable, future studies should focus on changing the way engineering knowledge is supplied to the reinforcement learning agent. Other information integration techniques such as reward shaping have shown promising results in other fields and could provide a solution here. When an adequate methodology for property integration is found, the efficiency of designing chemical processes could be increased greatly, thereby, contributing to the development of sustainable solutions in the field of chemical engineering.
Item Type:Essay (Master)
Faculty:TNW: Science and Technology
Subject:35 chemistry, 50 technical science in general, 58 process technology
Programme:Chemical Engineering MSc (60437)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page