Colloquia

Summary

Designing chemical separation processes is time-consuming and heavily dependent on heuristics and numerical simulations. While mixtures with similar compositions often require similar process designs, each case is typically developed from scratch. This work explores the use of reinforcement learning (RL) to accelerate process design by training a virtual agent to generate effective processes for mixtures with fixed but variable components.

The proposed framework consists of two stages: (1) a topology step, where the sequence of separation units is selected, and (2) a parameter step, where design and operational parameters are assigned. A flexible process simulation tool was developed for integration with the RL environment. In the topology step, configurations are generated via Deep Q-Learning, with a Deep Q-Network (DQN) estimating the value of adding specific unit operations. The agent incrementally constructs the process by selecting operations expected to yield the greatest improvement, as evaluated by a numerical process simulator. The parameter step applies both Bayesian optimization and evolutionary algorithms.

A case study on a quaternary hydrocarbon mixture with varying compositions showed that the proposed design approach can produce designs comparable to heuristic-based solutions in both structure and performance. For mixtures with similar compositions, the agent consistently proposed the same topologies while adapting design and operational parameters. In its current form, the agent's actions are limited to selling a stream or placing a distillation column, but the flowsheeting framework is extendible to additional unit operations. Although this study used the Antoine equation to estimate vapor pressures for phase-equilibrium calculations, the framework can incorporate alternative thermodynamic models to handle non-ideal mixtures.