Machine Learning for Group Contribution Methods

Knowledge of the properties of mixtures is essential for the design and optimization of chemical engineering processes. However, since experimental data are unavailable for many substances and mixtures, predictive methods are often required. Physical group contribution methods, which model substances and mixtures based on structural groups as their building blocks, are widely used in industrial practice but suffer from limited flexibility and applicability, as their parameter sets are often incomplete. Flexible machine learning (ML) methods offer exciting new possibilities, but purely ML-based approaches generally lack physical interpretability, robustness, and consistency.

Therefore, I develop hybrid models that combine the advantages of physical group contribution methods with those of flexible ML methods. This makes it possible to improve prediction accuracy, extend the range of applicability, and simultaneously ensure thermodynamically consistent results. In this context, I investigate, among other things, how these hybrid methods can be efficiently trained on different types of data, which factors influence the quality of model predictions, and how the underlying structural groups can be optimized.

Teaching

• Thermodynamics II (SS 2026)
• Process Thermodynamics (WS 2026/27)
   

Office Hours
By appointment

Publications

-

Presentations

-

Posters

• H. Mennecke,N. Hayer, B. Schmid, H. Hasse, F. Jirasek: A Systematic Analysis of Machine-Learning-Enhanced UNIFAC Models (Poster), Thermodynamik-Kolloquium, Dortmund, 30.09.-02.10.2026.