Machine learning discovery of a new cobalt free multi-principal-element alloy with excellent mechanical properties

Abstract

In the present study, the machine learning (ML) method was utilized to construct a composition–structure–property model incorporating physical features. To enhance the predictive accuracy, the volume fraction of the two phase microstructure was merged into the dataset serving as the physical constraint for the input variables. The physical features, the chemical composition and the temperature difference between the initial and final melting temperatures were selected as the input and output variables, respectively. To deal with the small sample data, the generalized regression neural network (GRNN) was selected and applied with optimization algorithms e.g., fruit fly optimization algorithm (FOA) and particle swarm optimization (PSO). The performance of the GRNN, FOA-GRNN and PSO-GRNN models were compared. As a result, the PSO-GRNN model was the most promising model and could be utilized to search for new multi-principal elements alloy (MPEAs) with targeted properties. Based on the ML results, a novel Fe2.5Ni2.5CrAl MPEA was designed and synthesized for experimental characterization. The DSC analysis shows that the developed alloy possesses narrower melting range and the predicted value is in excellent agreement with experiments with a relative error below 10%. The designed alloy possesses a typical dual-phase structure (FCC+BCC/B2) and exhibits exceptional mechanical properties with superior plasticity at the cast condition. This property improvement is due to solid solution strengthening and nanoparticles strengthening effects. Our proposed alloy can be a promising choice for selected high performance applications.