A comparative study of machine learning and response surface methodologies for optimizing wear parameters of ECAP-processed ZX30 alloy.

Magnesium, valued for its lightweight, recyclability, and biocompatibility, faces challenges like its poor wear behavior and mechanical properties that limit its adaptation for a multitude of applications. In this study, various statistical analyses, and machine learning (ML) techniques were employed to optimize equal channel angular pressing (ECAP) process parameters for improving the wear behavior of Mg-3wt.% Zn-0.7 wt% Ca alloy. ECAP was conducted up to four passes via route Bc at 250 °C. Wear testing of both as-annealed (AA) and ECAP-processed alloys was performed using the dry ball-on-flat wear method under varying loads, speeds, and time. One pass (1P) and 4Bc-ECAP resulted in a notable uniform grain refinement of 86 % and 91 %, respectively, compared to the AA. X-ray diffraction (XRD) analysis confirmed a refined structure attributed to extensive dynamic recrystallization. Mechanical wear testing revealed a significant reduction in volume loss (VL), up to 56 % and 28.5 % after 1P and 4Bc samples, respectively, compared to the AA sample, supported by the observed texture intensity. The coefficient of friction (COF) stabilizes at 0.30-0.45, indicating low friction characteristics. Next, by adjusting wear load and speed through design of experiments (DOE), the wear output responses, VL and COF, were experimentally investigated. The output responses were predicted in the second step using ML, 3D response surface plots, and statistical analysis of variance (ANOVA). According to the regression model, the minimal VL was attained at a 5 N applied load. Also, the wear speed and VL at different passes are inversely proportional. On the other hand, the optimal COF was obtained at applied load about 2-3 N and 250 mm/s at different passes. The wear process variables were then optimized using different optimization techniques namely, genetic algorithm (GA), hybrid DOE-GA, and multi-objective genetic algorithm (MOGA) approaches.