Project Details
Projekt
Generalizing Automated Neural Flowsheet Synthesis to Novel Chemical Systems
Subject Area
Chemical and Thermal Process Engineering
Methods in Artificial Intelligence and Machine Learning
Methods in Artificial Intelligence and Machine Learning
Term
since 2021
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 466387255
Flowsheet synthesis is a key step in conceptual process design in chemical engineering (CE). In the first funding period of the priority program (PP), we developed and implemented a reinforcement learning (RL) scheme for automated but inventive and explorative flowsheet synthesis of steady-state chemical processes. The RL environment consists of a process simulator containing a priori physical knowledge, i.e. physicochemical property data and a set of general process unit models. Step by step, the RL agent builds process flowsheets, modifies them, and evaluates them in the process simulator to obtain feedback/reward with respect to some objective (e.g., net present value). The agent has no prior knowledge of CE and is trained to generate flowsheets solely through automated interaction with the simulator. It has proven capable of developing flowsheets for complex separation processes such as azeotropic or entrainer distillation with solvent selection and recycle streams to minimize substance loss. The RL setup is now on par with classical mathematical programming techniques for flowsheet design, such as superstructure reduction. However, there is great potential for the RL setup to transfer learned policies to novel, unseen chemical systems. In our continuation project for the second funding period of the PP, we will utilize systematic training to enable learning beyond a single task in one chemical system and to generalize to other chemical systems. We will build on and improve our methods from the first funding period, aiming at a highly generalized workflow: The (human) designer defines a feed stream of some chemical mixture. The developed generalized machine learning (ML) model will then suggest near-optimal flowsheets to separate the mixture, even if it was not included in the training process. To achieve this, the Burger group (CE) will focus on making the flowsheet simulator more flexible (more process units, multi-component mixtures, more detailed cost functions). The Grimm group (ML) will lead the development of generalized flowsheet data representations, scalable and flexible ML model architectures, and efficient training strategies. The training will be done primarily in the space of the thermodynamic properties that characterize the behavior of the chemical components in the process units. Close collaboration between the two groups is required at all stages of the project and has proven successful during the first funding period. The project is located in field F of the PP’s collaboration matrix and primarily in research area #6 creativity. There is great potential for collaboration with other projects, e.g., we can include property prediction from molecular structures into our process simulator. Furthermore, we will continue to share methods (RL, short-cut methods) with all projects focused on process design, optimization, or any other planning problems.
DFG Programme
Priority Programmes
