New paper: Transfer learning accelerates reinforcement learning for process synthesis

Our paper "Accelerating process synthesis with reinforcement learning: Transfer learning from multi-fidelity simulations and variational autoencoders" by Qinghe Gao, Haoyu Yang, Maximilian F. Theisen and Artur M. Schweidtmann was published in Computers & Chemical Engineering.

Reinforcement learning agents can learn to build process flowsheets step by step by interacting with a process simulator. A major obstacle is the large number of simulations the agent needs, which makes training in rigorous simulators slow and computationally expensive.

The paper studies two transfer learning strategies to speed up learning: (i) transferring knowledge from shortcut process simulators to rigorous simulators, and (ii) transferring knowledge from process variational autoencoders. Suitable transfer learning improved both learning efficiency and the final design scores. Transfer can also slow learning when the pre-training and fine-tuning tasks differ strongly in decision range or reward function, so the pre-training data should match the complexity of the target task.

The work is part of our research on reinforcement learning for process design.

Paper: Q. Gao, H. Yang, M. F. Theisen, A. M. Schweidtmann (2025). Accelerating process synthesis with reinforcement learning: Transfer learning from multi-fidelity simulations and variational autoencoders. Computers & Chemical Engineering, 201, 109192. https://doi.org/10.1016/j.compchemeng.2025.109192