Project Details
Projekt
Sample-efficient physics-informed reinforcement learning for liquid-liquid separation process control
Subject Area
Chemical and Thermal Process Engineering
Term
since 2021
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 466656378
The separation of liquid-liquid dispersions is essential in many processes and is often performed in horizontal, continuously-operated gravity separators. Preventing flooding, i.e., significant entrainment of the disperse phase in the outlet of a coherent phase, is crucial for efficient separation. The occurrence of flooding, however, is difficult to predict, as the sedimentation, coalescence, and flow conditions in a separator are affected by material system, operating, and geometric parameters. Deriving an accurate mechanistic model of the separator dynamics is hindered by a poor understanding of how these parameters influence the phase separation. In phase 1 of the priority program SPP 2331, we demonstrated the suitability of physics-informed neural networks (PINNs) as dynamic process models in such a setting of insufficient mechanistic knowledge and limited data availability. To further investigate the separator dynamics and to acquire experimental data for model training, we have meticulously upgraded an experimental setup for a pilot-scale separator in our laboratory with advanced instrumentation, measurement and automation capabilities. This comprehensive approach allows us to generate first-of-its-kind dynamic data for this poorly understood process. For the continuation phase, we envision taking the modeling and experimental research to the next level and achieving model-based separator control through highly sample-efficient physics-informed reinforcement learning (RL). Building on our achievements of the first phase and harnessing recent advances in model-based RL, physics-informed learning, and transfer learning, we aim to operate the experimental separator in an automated RL loop that involves continuous fine-tuning of physics-informed models. This innovative concept aims to overcome the well-known sample inefficiency of standard RL approaches, a major obstacle for a wider application of RL in chemical engineering, by leveraging recent machine learning (ML) techniques that have, already on their own, shown a potential to increase sample efficiency. Chemical engineering (CE) frequently encounters the need to control a process with limited mechanistic knowledge and data. The proposed combination of state-of-the-art ML techniques with experiment-in-the-loop RL thus has the potential to significantly advance the state-of-the-art in CE.
DFG Programme
Priority Programmes
