Deep learning of chemical reactions

Chemical reactions are all around us, forming the basis of life. For example, we digest food and breathe air, convert energy stored in molecules to muscle movement, or transfer and process information using reactions. A chemical reaction transforms one or more educts into one or more products. Using quantum-mechanical calculations, reactions can be studied in detail, for example how the atoms in the educts iteratively move until the product is formed. However, such calculations are prohibitively slow to use on a larger scale. The project aims to adapt generative machine learning models to predict the atomic configurations along reaction pathways in the gas phase, as well as for organo- and biocatalytic reactions, i.e., reactions that are catalyzed by either small organic molecules or enzymes. Here, architectures prominent in other fields of research, such as diffusion models for image generation, will be adapted for chemical research questions. The trained models can then be used in larger machine learning pipelines to predict reaction properties, helping to understand, model, and optimize diverse chemical reactions on the computer.