Degree

Master of Science in Engineering (MSE)

Department

Chemical Engineering

Document Type

Thesis

Abstract

Distillation columns are the most used separation technique in all chemical fields. The design of column internals presents a hybrid action space problem, where trayed internal parameters are continuous and packed parameters are discrete; Hybrid meaning the action space consists of both discrete and continuous design parameters. Hybrid action spaces have presented a challenge in the reinforcement learning space, since the most robust algorithms are designed to handle specific action types. This research develops and validates a hierarchical multi-agent reinforcement learning framework that divides the action spaces into different agents. The Controller agent selects between internal types, the Soft Actor- Critic agent optimizes continuous parameters, and the Deep Q-Network agent handles discrete choices. The agents are integrated with AspenPlus software for dynamic modeling. Extensive experimental validation of the framework was done on three binary separations and over 10 random seed values for each separation. All 30 training runs resulted in the discovery of configurations with optimal hydraulic performance. Statistical analysis verified that each agent/algorithm showed legitimate learning throughout their respective training phases. Policy generalization performance revealed that the Benzene-Toluene system had a narrow action space that made it hard for the Deep Q-Network agent to converge to the optimal packing type. This can be important for future development and training of algorithms. Training required only 10.7 minutes on average, proving to be feasible for industrial implementation. This research addresses a gap found in existing research of reinforcement learning integration for process design in current research. These research groups used reinforcement learning to take process descriptions or chemical components and specify the equipment required to complete the separation. This work addresses this gap along with the hybrid action space challenge, developing a multi-agent reinforcement learning framework that can be used on generated flowsheets to perform detailed design on the equipment.

Date

2-6-2026

DOI

https://proquest.com/docview/3347818834

First Committee Chair

Dhan Lord Fortela

First Committee Member

Ashley Mikolajczyk

Second Committee Member

Magdy Bayoumi

Share

COinS