In the course of the digitalization of the process industry, there is currently an intensive discussion on how machine learning (ML) methods can be used in process engineering. In preliminary work at LTD, the potential of data-driven ML methods for predicting thermodynamic properties of mixtures has already been demonstrated. The concept of so-called recommender systems, which are established in online retail and streaming service providers, was thereby successfully transferred to the field of thermodynamics. Also in other fields of process engineering, ML methods open up new possibilities; however, the transfer from their classical application areas to problems in process engineering is not trivial. Of particular importance is the question of how data-driven ML approaches can be usefully combined with fundamental thermodynamic knowledge to create hybrid models. Within the scope of the research project, such novel hybrid models for the prediction and description of fluid properties shall be developed. Hybridization will be investigated on several levels: i) Generation and use of hybrid data sets to train the models, e.g. experimental data, molecular simulations, predictions of physical models. ii) Enforcing physical constraints in ML models, e.g. by Physics-Informed Neural Networks. iii) Merging physical theories and models with flexible ML algorithms.

[back]