Interpretable Machine Learning Design and Discovery of Complex-composition Metallic Materials

High-entropy alloy (HEA), also known as Complex Concentrated Alloy (CCA) or Multi-principal Elements Alloys (MPEA), is a type of metallic material with multiple principal elements. Compared to conventional alloys, HEA engenders vast opportunities for designing new materials with desirable structural and functional properties such as mechanical properties, thermal/electric conductivity, corrosion resistance, super-conductivity, radiation absorption, and hydrogen storage. However, the HEA compositional space is exceedingly large due to the nature of possessing multiple principal elements. Therefore, fundamental challenges arise in efficiently exploiting compositions with exceptional features. HEA compositions and processing methods control the formation of HEA phases, including solid-solution phases (SS) and intermetallic phases (IM). Phases of HEA determine the properties and need to be carefully designed. In this dissertation, we will discuss Machine Learning (ML) based HEA phase prediction methods, interpretation of HEA phase formation, HEA properties prediction, high-throughput HEA design methodology, and experimental synthesis of HEA with desired properties.
As HEA phase prediction methods have evolved from single physic-based parameters to first-principles calculations and ML approaches, the accuracy and capability of the HEA phase prediction methods are continuously improving. However, the prediction of HEA phases, especially the IM, is still underdeveloped due to the expensive computing power needed for first-principles calculations, or the lack of appropriate ML features and the limited dataset needed for ML approaches. To address these issues, we developed novel ML models with detailed phase classification and high accuracies, where nine phase categories can be predicted with accuracies close to 90 %. This model utilizes an innovative set of phenomenological ML features mined from binary phase diagrams and the feature engineering technique. 86 new HEAs were synthesized to validate the model's accuracy. The HEA phase formation interpretation has significant scientific importance. ML normally provide accurate predictions without disclosing the science behind it. Therefore, we identified and interpreted the key scientific factors controlling IM formation, guiding the HEA design and in-depth studies on phase formation.
The prediction of HEA properties includes melting temperature, density, cost, and mechanical properties (e.g., hardness, tensile/compression yield strength, and fracture strain). Various prediction methods will be introduced for different properties. Based on the phase and property prediction models, a comprehensive high-throughput HEA design method is developed to search for compositions with desired phase and properties. Through the use of this method, a series of HEAs have been designed specifically for marine environments, demonstrating the effectiveness of the method.