Composition optimization of cobalt-free Fe-Cr-Ni-Al/Ti multi-principal element alloys for strength-ductility trade-off based on machine learning

Multi-principal element alloys (MPEAs) represent a novel class of advanced structural materials. However, due to the vast composition space, traditional experimental methods and first-principles calculation are not efficient for exploring this space. In recent years, machine learning (ML) has emerged as a promising tool for alloy design, offering an effective approach to develop high-strength and high-toughness MPEAs. In this work, we proposed a random-exhaustive feature selection (REFS) method to select an appropriate feature subset from a high-dimensional feature for model training. Moreover, we compared the metrics of models trained by various ML algorithms using either composition input or parameter input strategies. The results indicated that the models trained by using the Gradient Boosting Regression (GBR) algorithm with the parameter input strategy exhibited the high R2 values of yield strength (0.947), ultimate tensile strength (0.922), and fracture elongation (0.844). Finally, we conducted a two-step composition design for cobalt-free Fe-Cr-Ni-Al/Ti MPEAs. The optimized models designed Al5(Fe10Cr35Ni55)95 and Al2Ti1(Fe10Cr35Ni55)97 MPEAs, which demonstrated the significant improvements of 56.42%, 29.45%, and 5.82% in yield strength (YS), ultimate tensile strength (UTS), and fracture elongation (FE), respectively, compared to equiatomic FeCrNi MPEA. In addition, by using the association rule mining (ARM) on the model predictions, the compositional dependence of mechanical properties and the possibility of data-driven alloy design were revealed.