Hybrid Thermodynamic Models (DFG Emmy Noether Junior Research Group)

Press release

Combining machine learning and physical knowledge: junior research group develops models for industrial production (in German)

Scope

Information on the thermodynamic properties of mixtures is the basis for designing and optimizing chemical engineering processes – for example in the chemical, pharmaceutical and biotechnology industries, or when shifting production routes from fossil to renewable feedstocks. Since the vast number of possible components, mixtures and state points cannot be studied exhaustively in experiments, models are required that provide reliable predictions where no experimental data are available. Established physical models, such as equations of state and Gᴱ models, are often limited in this respect because parameters are missing for many components and mixtures. Machine learning (ML) offers powerful new tools, but suffers from the comparatively small data sets in thermodynamics and is a priori agnostic of physical laws. The goal of this junior research group, funded by the German Research Foundation (DFG) within the Emmy Noether Programme, is therefore to develop hybrid models that combine the strengths of both approaches. Two routes are pursued: First, ML algorithms are embedded in physical models to predict their parameters; using end-to-end training directly on phase equilibrium data, group-contribution methods such as UNIFAC as well as models for electrolyte solutions will be completed. Second, physical knowledge – such as the Gibbs-Duhem equation or permutation invariance – is incorporated into deep neural networks as soft and hard constraints, so that their predictions are thermodynamically consistent. Furthermore, design-of-experiment strategies based on active learning are developed to identify the experiments that are most informative for training the models. All models are available via ml-prop.mv.rptu.de.

Funding

German Research Foundation (DFG), Emmy Noether Programme · Duration: 2023 - 2029 · Funding: approx. EUR 2 million · Lead: Prof. Dr.-Ing. Fabian Jirasek