Interpretable ensemble learning for tensile strength prediction and welding process window optimization of friction stir-welded 7N01 aluminum alloy

This study aims to achieve high-accuracy prediction of the ultimate tensile strength (UTS) of friction stir-welded A7N01P-T4 aluminum alloy joints, elucidate the effects of welding parameters on UTS, and optimize the welding process window. Rotational speed, welding speed, and shoulder plunge depth were used as input variables, with UTS as the output. Machine learning models were developed using 27 experimental samples and evaluated by leave-one-out cross-validation. Comparison of decision tree, K-nearest neighbors, and Gaussian process regression identifies decision tree as the best baseline model, with an R 2 of 0.905. An ensemble learning strategy was subsequently introduced to construct a gradient boosting decision tree, increasing R 2 to 0.966, and grid search optimization of its key hyperparameters further improves R 2 to 0.978. SHapley Additive exPlanations analysis shows that the importance of the welding parameters follows the order of welding speed, rotational speed, and shoulder plunge depth, and the identified parameter effects agree well with the experimental results and observed microstructural evolution. The optimized model predicts fully feasible ranges of all the welding parameters for a joint efficiency of at least 90%, and the shoulder plunge depth acts as a key parameter for further refining the high-strength welding process window. These results demonstrate that the proposed approach not only improves the accuracy of UTS prediction but also reveals welding parameter effects and enables optimization of the welding process window, providing a quantitative basis for selecting FSW parameters.