Molecules play a key role in almost all areas of life due to their diverse properties. These originate from the individual spatial arrangement (structure) of electrons and atomic nuclei. The relationship between molecular structure and molecular properties is at the core of theoretical chemistry. As the corresponding calculations are time-consuming and resource-intensive, we examine AI (artificial intelligence) algorithms and apply them to chemical problems. AI algorithms allow us to reliably predict properties of molecules for which data is not (yet) available.
We are particularly interested in the inverse problem: Which molecules (structures) exhibit a set of desired properties? If questions of this kind could be answered systematically, the virtual design of molecules would become possible. An example is a catalyst for the reduction of carbon dioxide, a much-discussed chemical reaction with regard to climate change.
Synthesis planning is a cornerstone of chemical research and its industrial application. Two aspects play an important role: time and selectivity. On the one hand, it is necessary that the conversion of reactants to desired products occurs within an economically viable period of time. On the other hand, the formation of unwanted byproducts is to be prevented. We are concerned with the design of tailored electrophiles and nucleophiles, which undergo selective reactions on well-defined time scales. For this purpose, we develop and study methods of machine learning and quantum chemistry, which we apply to problems related to chemical energy conversion. We verify the reliability of our methods by means of Mayr's reactivity scales.
Machine learning is usually associated with large data sets, which are readily available for pioneering applications such as image recognition, spam filters, etc. In the sciences, however, data generation is expensive, difficult, and time-consuming. Therefore, it is desirable to build powerful prediction models with as few data points as possible. To realize such a data-economic scenario, we develop and study active-learning algorithms. We reached an important milestone with the release of the batchwise variance-based sampling (BVS) method developed in collaboration with the group of Markus Reiher (ETH Zurich).