Efficient design of lightweight AlCrFeNiTi-based high-entropy alloys via computational thermodynamics and interpretable machine learning

High-entropy alloys generally exhibit superior comprehensive properties. However, due to the variety of elements involved, advancing their research using traditional experimental design methods involves tremendous effort and a low success rate. In this study, computational thermodynamics, machine learning, and first-principles methods are combined to efficiently design AlCrFeNiTi-based HEAs. Initially, using computational thermodynamics, a composition-performance database was established. Subsequently, various machine learning models and intelligent optimization algorithms were employed to select HEAs with low density and high strength. The CatBoost algorithm predicted the hardness with a Mean Percentage Absolute Error (MPAE) of 11.68 %, and the density with an MPAE of only 0.28 %. Furthermore, preliminary assessments of the target alloy's properties were conducted using first-principle calculations of special quasi-random structures. Explainable machine learning methods facilitated an understanding of the impact of composition, revealing that the content of Al had the most significant effect on the alloy's mechanical properties, while Ti exhibited a complex, non-linear relationship with hardness and density. The compositional design indicates that Fe0·28Cr0·16Ni0·18Al0·18Ti0.2, with a density of 5.88 g/cm³, is a potential high-modulus, lightweight HEA. This approach offers a feasible means for the efficient design and comprehension of design principles in various HEAs.