Data-driven modelling of grain boundary chemistry

The segregation of solutes to grain-boundaries (GB) in alloys can strongly influence physical and mechanical properties of materials. Such properties are e.g. brittle vs. ductile fracture, phase-transformations, or electrical conductivity. The GB-segregation can be described by different thermodynamic models, where the central parameter is the segregation energy (E_Seg) of the solute, which determines the solute concentration at the GB. A reliable and generally applicable method to calculate the E_Seg is given by density functional theory (DFT). However, this method has the disadvantage of high computational costs. Through the application of machine learning (ML) methods one can greatly reduce the computational effort. For the ML approach, a set of DFT results are used to train a model for E_Seg. This allows to replace further DFT calculations by ML predictions and thereby generate a significant speedup of segregation calculations. It is assumed, that it is possible to train an ML model that can predict the E_Seg depending on the segregation site and solute element. To train a reliable model it is necessary to identify the appropriate ML method and to find a correct representation of the atomistic data (atom-position, solute-element). The choice of the ML method and representation of the data is the focus of this thesis. Three different ML methods will be compared from which the best one will be selected. The data will be represented using different descriptors derived either from the atomic or electronic structure of the segregation site. Further, a method to construct optimal data-sets for training will be tested. Using the different ML methods and descriptors various models were set up. By comparing cross-validation scores the optimal models were found. For single solute models an average error of below 50 meV was achieved, where the best descriptors were derived from the electronic density of state with errors below 30 meV corresponding to relative errors of 3 %. By applying an active learning procedure to segregation of Re in W, the amount of training points could be reduced from 230 to 50 data-points while retaining the accuracy. For multi solute models the average errors were below 80 meV, corresponding to relative errors of 3 %. The such found ML model was trained on data for segregation of 11 different transition metals in W, which was used to extrapolate to all 30 transition metals.