Place: 10 Years ICAMS - International Symposium, Ruhr-Universität Bochum, Germany
Machine learning (ML) methods can be used to bypass computationally expensive simulations in materials science by learning from previous results. This allows to speed-up screening procedures and is therefore a highly promising approach for the discovery of novel materials. However, for an effective application, descriptors have to be constructed which are informed by domain knowledge of the material and are also in a suitable input format for the ML algorithm. The present work summarizes different approaches of the construction of descriptors that are based on the local atomic structure, chemistry and magnetic configuration of the material of interest. Different ML methods are then selected and applied on the generated feature space to predict different physical quantities, including the structural stability of transparent conductors, intermetallic phases and magnetically disordered iron.