Bayesian neural networks for molecular discovery
Designing and creating new molecules and materials with specific tailored properties can lead to huge progress in medicine, solar cells, catalysts, and many other scientific areas. But identifying which compounds have the properties we desire is not easy. While quantum mechanical computations can determine many properties in a matter of minutes or hours, the number of possible molecules is so huge that search by trial and error is futile. Based on databases of known compounds and their properties, deep neural networks have proven extremely efficient in predicting properties of new compounds. These neural networks can guide our search, but have no notion of uncertainty and can give misleading results that are difficult to diagnose. To be used efficiently, the neural networks need to know what they don’t know. In this project we will develop methods for uncertainty quantification in deep neural networks aimed at the search for new and exciting materials and molecules.