As we move toward a more circular chemical industry, many processes need to be redesigned or newly developed. To support this shift, we’re creating computer-aided methods for conceptualizing chemical and biotechnological processes. These methods integrate modeling, simulation, and optimization in a tightly linked approach. Advances in machine learning have opened up new opportunities for computer-aided conceptual design.
We’re currently working on the following projects:
Reinforcement Learning-Based Process Design
In partnership with the Professorship for Bioinformatics, we have been at the forefront of using machine learning to create complete chemical processes from the ground up. We have built robust process simulators that can evaluate a wide range of potential process designs. A reinforcement learning agent suggests economically optimal processes within the simulator, trained through self-play using an AlphaZero-based approach. This method allows the agent to generate complex flowsheets, including those with recycle loops and azeotropic distillation sequences. It can even handle multiple chemical systems with a single agent, something not seen before in reinforcement learning-based process design. Beyond process engineering, the developed algorithms are adaptable to any type of planning problem. The project is funded by the DFG under the SPP 2331 Machine Learning in Chemical Engineering.
