The segregation of solutes to grain boundaries (GB) strongly influences a range of materials properties related to, for example, brittleness or phase transformations. This process is controlled by the GB segregation energy. In this work we are interested in the site-specific segregation energies, which can be calculated to a high accuracy using DFT. However, such calculations are very tedious and expensive to perform.

Here we apply machine learning (ML) methods to reduce the computational costs of comprehensively mapping segregation to different GBs. In contrast to the common practice of using structure-related parameters, we propose to use features derived from the electronic density of states as input for the ML algorithms. This means that information on the local electronic structure at the segregation site is also included. For this purpose, we represent the electronic structure via the moments of the projected density of states. We have applied these new features to solute segregation in tungsten and compared them with other commonly used structure-related parameters. The results show a clear improvement over regression relying on purely structure-related features.