Project 3: Thin-film morphology generation from flow matching

This project aims at generating thin-film morphologies for organic electronic materials using machine learning. The primary objective is to model complex, amorphous thin-film architectures accurately and efficiently through deep generative methods. These models operate at a coarse-grained level to simplify the process, later refining the generated structures to an atomistic resolution via backmapping techniques.

We will exploit flow matching with recently proposed regularization strategies that straighten the diffusion trajectories. This will allow accurate inference with few diffusion time steps. In addition, we will combine diffusion with the novel free-form flow approach that enables better incorporation of physical and chemical prior knowledge into the learning process. The free-form flow’s enhanced freedom in network architectures and training algorithms will also facilitate joint representation learning across resolutions. 

Team

Prof. Dr. Ullrich Köthe

Principal Investigator (IWR, Heidelberg University)

Phone: +49 6221 5414834

More Information

Prof. Dr. Tristan Bereau

Principal Investigator (ITP, Heidelberg University)

Phone: +49-6221-54-9448

More Information