Parametric Analysis for Torque Prediction in Friction Stir Welding Using Machine Learning and Shapley Additive Explanations

Friction Stir Welding (FSW) has revolutionized modern manufacturing with its advantages, such as minimal heat-affected zones and improved material properties. Accurate torque prediction in FSW is crucial for weld quality, process efficiency, and energy conservation. Many researchers achieved models for torque based on experimental research, yet the models were limited to a specific type of material. In recent years, the use of machine learning techniques has increased in industry in general and in welding in particular. In this study, a machine learning model was prepared based on artificial neural networks, and Shapley-Additive Explanations were used to predict the rotational torque from 287 experiments that had been conducted in several previous studies. The achieved model has remarkable predictive performance, with an R-squared of 99.53% and low errors (MAE, MAPE, and RMSE). Moreover, a machine learning polynomial regression was examined for comparisons. A parametric importance analysis revealed that rotational speed, plate thickness, and tilt angle significantly affect torque predictions, while the rest of the variables had minimal importance.