Machine learning accelerated design of high strength and lightweight high-entropy alloys

Abstract

High-entropy alloys (HEAs) with multiple principal elements demonstrate impressive mechanical properties, such as high hardness and strength, and have potential in various industrial applications. However, their vast compositional space makes the design of HEAs with desired properties via experimental methods infeasible. In this study, a data driven approach was used to accelerate the design of lightweight and high-strength HEAs. A HEA dataset that consists of 11 elements was created by meticulously curating the required HEA data from two reliable established datasets. By using this dataset, an XGBoost regressor machine learning model was developed to rapidly predict the tensile yield strength (YS) of HEAs across the huge compositional space. The model demonstrated good predictive performance, with a mean absolute error of 75.785 MPa on the unseen experimental HEA data, highlighting its ability to generalize to new data. Matminer, an open-source Python library designed for materials informatics, was used to generate composition-based descriptors to replace the atomic percentage features that restricted the exploration of wider compositional space of HEAs. A sequential feature selection technique and permutation feature importance were also used to select relevant features and reveal the importance of each input feature to the tensile YS prediction. The developed model was used in a conditional random search to explore a generated dataset of 10000 HEAs, aiming to design HEAs with desired properties. The top-performing designed HEA demonstrated a lower calculated density of 4.795 g/𝑐𝑚3 and a relatively high tensile YS of 1235.39MPa compared to the HEAs in the existing dataset.