Machine Learning-Based Design of Critical Raw Material (CRM)-Free Multi-Principal Element Alloys (MPEAs)

Abstract

This thesis presents a comprehensive, multi-pronged computational strategy for the accelerated and sustainable design of multi-principal element alloys (MPEAs), addressing the challenge of navigating an enormous compositional space (~10100 combinations) with limited high-quality experimental data. A machine learning (ML) framework was developed using exclusively experimental data from a consistent synthesis route (melting and casting), avoiding synthetic data augmentation to ensure model robustness and generalizability. Benchmarking revealed that while synthetic augmentation may boost accuracy, it undermines reliability in imbalanced datasets. The resulting framework demonstrated superior generalizability for phase prediction, guiding pre-experimental alloy design. For mechanical property prediction, an open-source toolkit, MAST-ML, was employed to assess the robustness and limitations of standard ML pipelines in this complex materials domain. Insights from this evaluation led to the development of a novel ML–metaheuristic optimization framework that simultaneously optimized yield strength, ultimate tensile strength, and elongation, addressing the longstanding strength–ductility trade-off in MPEAs using ML for the first time.