Project Details
Projekt
Learning Groups for Group Contribution Methods (LEGO)
Applicants
Professorin Dr. Sophie Fellenz; Professor Dr.-Ing. Fabian Jirasek; Professor Dr. Marius Kloft
Subject Area
Chemical and Thermal Process Engineering
Methods in Artificial Intelligence and Machine Learning
Methods in Artificial Intelligence and Machine Learning
Term
since 2024
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 553345933
Knowledge of the thermodynamic properties of pure substances and mixtures is critical in chemical engineering, e.g., for process design and optimization. Since experimental data are scarce, prediction methods are indispensable. Most physical prediction methods are group contribution methods (GCMs), which model substances based on structural groups as building blocks. Although GCMs are established in all process simulation software packages, they still have severe limitations hampering accuracy and applicability. This project will tackle the primary limitations of existing GCMs by combining them with machine learning (ML) approaches: Suboptimal structural group definitions, rigidity in ignoring the molecular context, and incomplete parameterizations. First, we propose employing reinforcement learning (RL) to rigorously optimize the structural groups on which GCMs rely. Second, we will use representation learning (RepL) to capture the molecular context of structural groups, enhancing the flexibility and accuracy of parameterizations. Lastly, we will develop advanced matrix completion methods considering the learned group representations to predict missing interaction parameters, further extending the applicability of GCMs. By fusing advanced ML techniques with the established physical framework of GCMs, this project aims to forge powerful hybrid models that substantially advance the prediction of thermodynamic properties. By retaining the GCMs framework, the hybrid models will be thermodynamically consistent and readily implementable into established process simulation software, fostering applicability of and trust in the new models. However, the impact of this project will go beyond advancing GCMs. The project will, e.g., provide insights into the relevance of particular structural groups for thermodynamic property prediction, which will be relevant for thermodynamic model developments in general.
DFG Programme
Priority Programmes
