Data-driven prediction of joint strength in friction stir welding: A machine learning and feature sensitivity analysis

Friction stir welding (FSW) is a solid-state joining process characterized by a non-linear and complex relationship between the process parameters and joint strength. In the present work, the prediction of FSW joint strength is done through a comprehensive data-driven framework along with parameter sensitivity analysis. Different machine learning (ML) models have been applied along with hyper-tuning parameter optimization to predict the FSW joint strength. The data-driven framework consists of 275 experimental datasets systematically collected from the peer-reviewed literature. The effect of FSW parameters (such as joint type, plate thickness, pin height, tool tilt angle, and tool shoulder to pin diameter ratio) on the joint strength has been examined. Results revealed that the highest accuracy was yielded by the Random Forest Regression and Artificial Neural Network models, followed by the K -Nearest Neighbors and Support Vector Regression; thereby highlighting the efficacy of ML in modeling multifactorial welding processes. Furthermore, SHapley Additive exPlanation analysis has been performed to obtain the feature importance and their impact on joint strength prediction. The tool rotational speed and plate thickness were the most influential parameters affecting the joint strength, as per the analysis. Also, the contribution of tool traverse speed and pin height was moderate but process-dependent. This integrated approach not only enhances the interpretability of ML models but also provides actionable insights for process optimization; particularly in dissimilar welding scenarios where parameter sensitivity is delicate due to material incompatibilities. It was also found that the tool rotational speed and plate thickness were the dominant variables on which the joint strength depends. Process-dependent effects exhibited by the tool traverse speed and pin height were linked to heat input and material flow. Dissimilar joints displayed higher sensitivity to parameter variations as compared to the similar joints. By bridging empirical experimentation with computational intelligence, this present research work contributes to the development of intelligent welding systems capable of real-time parameter tuning for next-generation manufacturing technologies.