Hybrid Modeling of Grain Boundary Chemistry

Machine learning is increasingly becoming an integral part of materials engineering and offers new approaches to explore the relationship between the internal structure and the properties of materials. In the project "GRAMMACS: Hybrid modeling of grain boundary segregation" we will develop computational methods for a highly relevant aspect of materials engineering, namely the segregation of alloying elements to the grain boundary (GB). We combine segregation databases, physical modeling concepts developed in part by the project team, and machine learning methods to achieve predictive modeling of segregation where current approaches are not applicable or are known to fail. We follow the hypothesis that the most effective approach will be a combination of physical modeling and machine learning, i.e., a hybrid approach. In general, in materials engineering, data is not available at a scale comparable to image recognition or process automation with billions of data points, where a black-box approach to machine learning can work. Rather, we are dealing with relatively precise data points that require significant experimental or computational effort. We will therefore combine quantum mechanical simulations and thermodynamic simulations with appropriate regression methods to make the best use of the data. With these methods we aim to achieve a more predictive simulation of GB segregation, which determines the chemical composition and structure of KG and subsequently the mechanical properties of metallic alloys. Great attention is paid to the influence of dissolved alloying elements on GB cohesion. It can be imagined that alloying elements can act like a glue that can prevent the GB from fracturing. Conversely, they can also break the bonds and promote fracture. However, the relevance of segregation goes beyond the consideration of cohesion. It is well known that KG segregations are a precursor to and strongly influence precipitations of new phases. Moreover, phase transformations in steels often start at the KG and segregation state plays a crucial role here as well. Therefore, control of GB segregation and the resulting chemistry is a crucial requirement for functional and mechanical design of metallic alloys.

Machine learning is increasingly becoming an integral part of materials engineering and offers new approaches to explore the relationship between the internal structure and the properties of materials. In the project "GRAMMACS: Hybrid modeling of grain boundary segregation" we developed computational methods for a highly relevant aspect of materials engineering, namely the segregation of alloying elements to the grain boundary (GB). We combined segregation databases, physical modeling concepts developed in part by the project team, and machine learning methods to achieve predictive modeling of segregation. With these methods new tools for the predictive simulation of GB segregation, which determines the chemical composition and structure of KG and subsequently the mechanical properties of metallic alloys could be obtained. In general, in materials engineering, data is not available at a scale comparable to image recognition or process automation with billions of data points, where a black-box approach to machine learning can work. Rather, we are dealing with relatively precise data points that require significant experimental or computational effort. We therefore, combined quantum mechanical simulations and thermodynamic simulations with appropriate regression methods to make the best use of the data. Great attention was laid on the influence of dissolved alloying elements on GB cohesion. It can be imagined that alloying elements can act like a glue that can prevent the GB from fracturing. Conversely, they can also break the bonds and promote fracture. We have created large datasets for GB segregation and cohesion based on quantum mechanical calculations and applied machine learning to replace the expensive calculations with computationally much cheaper algorithms. These algorithms also make predictions for which no data were generated and provide therefore universal segregation models. The materials under investigation included Al, Cu, Fe, Mg, Mo, Nb, Ni, Ta, Ti, and W with a particular focus on Fe and W due to their large relevance as structural materials. The generated data and segregation models are not only relevant for reducing grain boundary embrittlement. It is well known that KG segregations are a precursor to and strongly influence precipitations of new phases. Moreover, phase transformations in steels often start at the KG and the segregation state plays a crucial role here as well. Therefore, our hybrid models for control of GB segregation and the resulting chemistry are of great help for designing functional and mechanical properties of metallic alloys.