Prediction of phases and mechanical properties of magnesium-based high- entropy alloys using machine learning: A Data Paper

The data is of HEA of magnesium composed of Mg, Al, Cu, Mn and Zn. The first five columns represent percentages by weight of the Mg matrix and alloying elements. These were followed by ultimate tensile strength (UTS) and yield strength (Yield) in MPa. The next column presents Young's Modulus (E) in GPa. The columns that followed represented number of elements (Num_of_Elem), calculated alloy density (Density_calc) and diffusion and thermodynamic features. The features were: Enthalpy of mixing (dHmix), entropy of mixing (dSmix), change of Gibbs free energy of the mix (dGmix), melting temperature (Tm), number of parameters (n_Para), Atomic size difference (Atom_Size_Diff), electronegativity difference (Elect_Diff), valence electron concentration (VEC), alloy synthesis route (Synthesis_Route), hot or cold working processing (Hot_Cold_Working), homogenization temperature (Homogenization_Temp), homogenization time in hours (Homogenization_Time), annealing temperature in degrees Celcius (Annealing_Temp), Annealing time in hours (Annealing_Time), if quenching was done (Quenching), multiphases present in microstructure (Multiphase), intermetallic structure (IM_Structure), Phases, crystal structure (Structure).

The data underwent a thorough cleaning process, was scrutinized for any missing values, and subsequently encoded and transformed to suit machine learning applications. The selection of pertinent features was accomplished through techniques such as backward elimination, forward selection, and regularization. Feature engineering was employed to devise novel features like calculated density (Density_calc), entropy change upon mixing (dSmix), atomic size disparity (Atom_Size_Diff), electronegativity variance (Elect_Diff), and valence electron concentration (VEC). The iterative testing phase assessed the influence of these newly engineered features on the performance of the model, which contributed to the refinement of the results.

The finalized dataset comprised 60 instances and 29 attributes. Tests for multicollinearity confirmed that the variance inflation factor (VIF) values remained within acceptable limits.