A Machine Learning Framework for the Prediction of Grain Boundary Segregation in Chemically Complex Environments

The discovery of complex concentrated alloys has unveiled materials with diverse atomic environments, prompting the exploration of solute segregation beyond dilute alloys. Data-driven methods offer promising for modeling segregation in such chemically complex environments, and are employed in this study to understand segregation behavior of a refractory complex concentrated alloy, NbMoTaW. A flexible methodology is developed that uses composable computational modules, with different arrangements of these modules employed to obtain site availabilities at absolute zero and the corresponding density of states beyond the dilute limit, resulting in an extremely large dataset containing 10 million data points. The artificial neural network developed here can rely solely on descriptions of local atomic environments to predict behavior at the dilute limit with very small errors, while the addition of negative segregation instance classification allows any solute concentration from zero up to the equiatomic concentration for ternary or quaternary alloys to be modeled at room temperature. The machine learning model thus achieves a significant speed advantage over traditional atomistic simulations, being four orders of magnitude faster, while only experiencing a minimal reduction in accuracy. This efficiency presents a powerful tool for rapid microstructural and interfacial design in unseen domains. Scientifically, our approach reveals a transition in the segregation behavior of Mo from unfavorable in simple systems to favorable in complex environments. Additionally, increasing solute concentration was observed to cause anti-segregation sites to begin to fill, challenging conventional understanding and highlighting the complexity of segregation dynamics in chemically complex environments.