Predictive Modelling of Ultimate Tensile Strength in Similar and Dissimilar AA2050-AA7475 Friction Stir Welds Using Machine Learning

Abstract Aerospace-grade heat-treatable aluminium alloys are often joined using solid-state welding processes such as friction stir welding (FSW). However, limited research has examined dissimilar AA2050-AA7475 joints. In this study, various machine learning (ML) models are developed to predict the ultimate tensile strength (UTS) of similar (AA2050–AA2050, AA7475–AA7475) and dissimilar (AA2050–AA7475) FSW joints. A dataset of 108 experimental trials was generated using six process parameters: tilt angle, shoulder diameter, rotational speed, pin diameter, tool offset, and traverse speed. Five ML models, such as Artificial Neural Network (ANN), Support Vector Regressor (SVR), Random Forest Regression (RFR), XGBoost, and Gaussian Process Regression (GPR), were evaluated using pooled out-of-fold cross-validation. XGBoost achieved the highest prediction accuracy (R² = 0.761), while GPR (R² = 0.718) was selected for its ability to model continuous variations and estimate prediction uncertainty. Joint type was the most influential factor affecting UTS. Optimal performance was observed with a tool offset of 0.5–0.75 mm, rotational speed of 700–900 rpm for dissimilar joints, and traverse speed of 180–220 mm/min for similar joints. GPR validation showed a MAPE of 2.1%, with all predictions within ±5% of experimental values, confirming its reliability.