An exploratory interpolation-based explainable machine learning approach for analyzing tensile strength in friction stir welding

Friction stir welding (FSW) parameters greatly influence the mechanical performance of welded joints; however, a data-driven analysis of these relationships is often limited by a small number of experimental studies. This study utilized an exploratory machine learning framework designed to explore tensile strength behavior in tube plate FSW joints using a structured factorial data set composed of 27 experimental observations. The parameters of interest in this study include tool rotational speed, pin clearance and shoulder diameter. To visualize parameter interactions across the experimental bounded design space, 500 synthetic samples were created for interpolation based analysis. Through cross-validation, polynomial regression, random forest, and XGBoost models were compared. XGBoost had a significantly lower prediction error than the other two models under the assumptions made in this model. Model interpretability was conducted using SHAP analysis to assess the relative influence of the process parameters on tensile strength. Results showed that the most dominant process parameters to predict the tensile strength behavior were rotational speed and shoulder diameter. Therefore, this framework provides an interpretable data-driven method for exploring parameter interactions in friction stir welding, which can facilitate future experimental investigations.