Investigations on grain boundary segregation with physical models and data-driven approaches

The topic of solute grain boundary (GB) segregation is of great relevance for materials engineering, since even a small amount of solute content in a material can lead to enrichment at the GB, affecting the material's properties. In this work, the topic of GB segregation is approached using theoretical, as well as experimental methods. Where those two approaches agree and disagree with each other is discussed and machine learning models are built that utilize GB segregation databases, filled with experimental segregation data. At the start, a purely physical approach to the topic of GB is presented. This is followed by a hybrid approach of combining Bayesian inference with physical modeling to investigate GB segregation. To conclude this thesis, a purely data-driven approach for predicting GB segregation is explored. In the first part, the hard-coating material TiBN is investigated on a purely physical basis. The way B is incorporated into the fcc-TiN matrix is explored, with a special interest in B segregation on fcc-TiN GBs. In addition, the solubility of B in fcc-TiN and h-TiB2 is evaluated. The computational methods used for this investigation are ab-intio calculations, as well as the CALPHAD approach. Experimentally, various spectroscopic and microscopic methods are used to explore the phase composition and nano-composite structure of two TiBN coatings created by physical vapor deposition (PVD). In the second part, a physics-based forward model and a data-driven inverse model are explored for the FeP system. P GB segregation in ¿-Fe has been intensively studied in the past, especially using Auger-electron spectroscopy (AES). AES segregation data from literature is taken and compared to enrichment predictions from ab-initio and semi-empirical calculations in combination with thermodynamic models. In addition, Bayesian inference is used to optimize the parameters of the thermodynamic models. The effects of each parameter on the agreement are discussed. Lastly, the influence of interstitial segregation sites is detailed, suggesting that P also segregates interstitially to ¿-Fe GBs. In the third part, a purely data-driven approach for predicting GB segregation is investigated. For this reason, AES segregation data from literature are collected and compiled into one database, which is subsequently used to train simple machine learning (ML) models. The predictive capabilities of these ML models are first tested on multiple solute elements separately. For P in $\alpha-$Fe, the models are able to accurately predict the influence of aging temperature and bulk composition on the GB segregation, as well as site-competition effects. Finally, a universal model for ¿-Fe is created, which is able to predict the GB enrichment for a variety of elements with an accuracy comparable to the average measurement error.