Machine learning-aided phase and mechanical properties prediction in multi-principal element alloys

Multi -principal element alloys (MPEAs) have gained significant attention recently owing to their distinct microstructural characteristics and enhanced mechanical properties. These alloys have the potential to replace traditional alloys by providing a range of tailored mechanical properties, corrosion resistance, and radiation tolerance, all achievable through complex compositions. This study utilizes various machine learning (ML) techniques to predict phases of MPEAs based on two data sets, incorporating statistical measures and interactions of features. The first dataset consisted of 3D transition MPEAs, while in the second dataset, the 3D transition MPEAs alloyed with refractory elements were added to the first dataset. Valence electron concentration, mixing enthalpy, mixing entropy, melting temperature, and electronegativity were the most important features, with phase detection accuracy ranging between 68.0% to 97.7%. Next, to predict the mechanical properties, we developed a novel two -layered Bayesian shrinkage method to stack the prediction power and model -based clustering of Support Vector Machine Regression (SVMreg), Random Forest Regression (RFreg) and a Mixture of Linear Regression with Experts. This was done to predict the mechanical properties of MPEAs more accurately, including microhardness, ultimate and yield strength, elastic modulus, and elongation. The proposed Bayesian machine deals with the multicollinearity of the base predictive models and shrinks the predictions to find the optimum predictive candidates. The proposed Bayesian method outperforms the base ML methods in predicting the mechanical properties of MPEAs. Subsequently, we conducted thorough validation of the model on four distinct MPEAs through experimental assessments.