Complex Learning: Machine learning approaches for transition metal complexes

Transition metal complexes (TMCs) are essential for a large range of different applications in renewable energies, energy storage, medicinal chemistry and more. Finding TMCs that have optimal properties for such applications is an expensive and time-consuming task as it traditionally requires experiments. Computer-based methods can accelerate this process by screening a large number of TMCs for promising candidates beforehand. Particularly machine learning approaches that leverage existing data have recently drawn a lot of attention. However, most research in this field focuses on small organic molecules which are easier to handle due to their simpler chemistry. In this thesis Kneiding presents four different and novel machine learning approaches, specifically designed for their application to TMCs. In particular they are: (1) a neural network leveraging quantum chemistry informed graph representations, (2) a genetic algorithm performing directional and multi-objective optimization, (3) a set of autocorrelation descriptors capturing bond-bond and atom-bond relationships, and (4) a generative neural network for the generation of novel ligands with specific properties. In addition to these methods Kneiding also presents two TMC datasets and discusses how they can be exploited for the computer-aided discovery of TMCs.