Machine Learning for the Design and Control of Power2X Processes with Application to Methanol Synthesis

• Applicants: Professor Dr.-Ing. Achim Kienle; Professor Dr. Sebastian Sager; Dr.-Ing. Ronny Tobias Zimmermann
• Subject Area: Process Systems Engineering ; Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics ; Mathematics
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 466495488

The aims of this project are: 1) The development of new numerical methods which combine the strengths of traditional modeling and optimization approaches and data-driven machine learning (ML) and their application to 2) the development of a new methodology for the design and management of Power2Chemicals processes. The nonlinear dynamics due to strongly fluctuating feeds will be explicitly taken into account. Methanol synthesis will continue to be used as a theoretically demanding application example of high practical relevance. Significant innovations from a process engineering perspective compared to the 1st funding phase relate to a) the focus on CO2 as a carbon source for methanol synthesis, b) the consideration of transport limitations in the catalyst and the related additional degrees of freedom, which will be used for an integrated catalyst pellet and process design, c) the design of integrated, robust and load-flexible reactor concepts to increase the conversion by simultaneous separation of the by-product water, and d) the development of suitable process control concepts to compensate for unforeseen disturbances on the basis of the developed hybrid models. This implies new challenges for the mathematical methods and algorithms to be developed. These mainly concern the development, calibration, and optimization of hybrid models with multiple neural networks to describe the coupled reaction and transport processes in the catalyst and the consideration of discrete decision variables in the areas of process design, experimental planning, and process control through mixed-integer optimal control (MIOC). Our hybrid modeling methodology combines experimental data from a gradient-free kinetic reactor with available physico-chemical knowledge and efficient ML. A particular focus is on data-efficient training of the resulting hybrid models. We extend successful algorithmic approaches for training and experimental design. The resulting hybrid model is used for robust process design. Nominally optimal control profiles are determined using MIOC for characteristic feed curves. The robust process design is supported by robust control to compensate for unavoidable model errors and unavoidable model errors and unforeseen deviations from the nominal case. This is based on repetitive online optimization and may require further model reduction to meet real-time requirements. Finally, the developed concepts will be experimentally validated for a gradient-free reactor. The ambitious work program reflects the complementary expertise of the three applicants in the fields of experimental analysis, conceptual process design and control as well as efficient algorithms.

• DFG Programme: Priority Programmes
• Subproject of: SPP 2331: Machine Learning in Chemical Engineering. Knowledge Meets Data: Interpretability, Extrapolation, Reliability, Trust