Integrated multiobjective optimization of RFSSW parameters for AA2024-T3 using ANOVA machine learning and NSGA II.

Multi-objective process optimization is critical in intelligent manufacturing, especially where complex, nonlinear interactions among parameters can significantly affect product quality. This study demonstrates a data-driven approach to optimizing Refill Friction Stir Spot Welding (RFSSW) parameters for an AA2024-T3 aluminum alloy. First, a 3³ full-factorial design of experiments was conducted to generate training and validation data on three pivotal process variables: rotational speed, plunge depth, and welding time. Statistical analysis using ANOVA highlighted plunge depth as the most influential factor, alongside notable interaction effects among the parameters. To build predictive models of joint load capacity, six machine learning techniques (MLP, RBF, GPR, k-NN, SVR, and XGBoost) were evaluated via cross-validation. XGBoost delivered the most accurate predictions, reaching R² values up to 0.89 with the lowest MAE and RMSE. Model interpretation methods such as feature importance and SHAP confirmed the dominant role of plunge depth, as suggested by ANOVA. The crux of the study lies in a multi-objective optimization framework using the NSGA-II evolutionary algorithm, targeting maximum weld strength in two distinct shear-testing variants (pure shear and free shear). The procedure generated a Pareto frontier of optimal parameter sets, from which a maximin strategy selected a high-performing compromise solution. These findings underscore the value of combining statistical methods, advanced machine learning, and evolutionary optimization in refining solid-state joining processes. More broadly, this integrated methodology provides a robust template for intelligent manufacturing applications that require balancing multiple performance objectives under complex process conditions.