Graph-Based Generative Machine Learning for Optimal Molecular Design

点击翻译按钮获取中文摘要

The objective of this project is to develop a platform for the automated design of molecules in Chemical Engineering. Designing molecules boils down to identifying molecules that optimize application-specific properties such as solubility or ignition delay. The platform connects molecular modeling with graph-based Machine Learning methods by representing molecules as molecular graphs with atoms as nodes and chemical bonds as edges. The platform has three components: The first component is a graph neural network tool for the automated and chemically interpretable prediction of physicochemical properties of molecules and chemical mixtures. The second component involves the generation of molecules using graph-based generative Machine Learning. The third component combines the results of the first two for the design of molecules with desired properties, which depends on both the generation of novel molecules as well as the prediction of their physicochemical properties. In particular, the platform will advance and accelerate the improvement of chemical processes through the automated discovery of novel working fluids, solvents, reactants, and catalysts with optimal properties. The platform will be made available as open-source in the Git-environment of the SPP 2331 and can be used, adapted, and actively extended by researchers. To this end, documentation, tutorials, and case studies from Chemical Engineering will be created for the individual components of the platform. At the same time, the platform will enable data-driven molecular modeling in other projects within the SPP 2331 and will be scaled up through these application use cases.