Hybrid Thermodynamic Models

• Applicant: Professor Dr.-Ing. Fabian Jirasek
• Subject Area: Chemical and Thermal Process Engineering
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 528649696

Predictive thermodynamic models for the properties of mixtures are fundamental for process design and optimization in chemical engineering. While physical models have been developed for decades and are established for multiple purposes in industrial practice, data-driven methods from machine learning (ML) offer new and promising opportunities to date. In the present project, these two concepts will be combined to develop hybrid thermodynamic models that unite the advantages of both worlds. Two general routes will be followed for this purpose, namely i) embedding ML algorithms into physical models and, inversely, ii) incorporating physical knowledge into ML algorithms. In the first route, ML methods will be developed and trained to predict the parameters of physical models for both mixtures of molecular components and electrolytes. The second route will incorporate physical laws, constraints, and boundary conditions in a priori extremely flexible deep neural networks. For this purpose, physics-informed neural networks (PINN) will be transferred to thermodynamics. Furthermore, novel hard-constraint methods will be developed, where, e.g., consistency criteria will be incorporated hard-wired into deep neural networks. Furthermore, methods for the design of experiments (DOE) for targeted augmentation of the data sets based on active learning strategies will be developed. Specifically, these methods will identify those unstudied data points whose measurement promises the most significant improvement of the developed hybrid thermodynamic models. This will introduce a new paradigm: whereas in the past, measurements were carried out if needed in practice, they will now be selected to bring the most information for modeling in the future. The DOE will be applied for planning fluid property measurements, which will then be carried out; the new data will subsequently be used for refining the developed hybrid models.

• DFG Programme: Independent Junior Research Groups