Machine Learning Models for Predicting Mechanical Properties in Friction Stir Welding of Al Alloys
Bhiksha Gugulothu, K. Srividya, S. Vijayakumar, Itha Veeranjaneyulu, Shanmugam Revathi, M. Ramya
Abstract
Friction Stir Welding (FSW) has developed as an extremely reliable solid-state joining technique for Al alloys, which possesses superior mechanical properties and minimal defects compared to conventional fusion welding. Accurate prediction of welding output parameters such as Ultimate Tensile Strength (UTS), Elongation (%), Hardness, and Wear Rate is crucial for ensuring weld quality and optimizing process conditions. In this study, four machine learning (ML) approaches were employed, such as Backpropagation Neural Network (BP), Extreme Gradient Boosting (XGB), Support Vector Machine (SVM), and Principal Component Analysis (PCA) to predict the output responses with the help of a dataset derived from experimental values. Initial correlation analysis highlighted D/d ratio and tilt angle as critical parameters, moderately correlated with UTS, elongation, and wear rate, but weld speed was inversely related to hardness. In the model performance, SVM achieved the highest prediction accuracy of more than 99.5% for UTS and elongation, followed by BP and XGB methods. PCA demonstrated stable performance, particularly for hardness prediction and BP achieved ~99.44% accuracy for wear rate. Error analysis demonstrated that SVM exhibited the lowest and most stable percentage errors, particularly for UTS and elongation, while PCA showed the least deviation for hardness predictions. The Mean Squared Error (MSE) values for the models were as follows: BP 0.6280, XGB 1.1681, SVM 0.4838, and PCA 0.9590. Overall, the comparative study validates the effectiveness of integrating ML algorithms for accurate welding output prediction, offering potential for real-time process monitoring and optimization in advanced manufacturing environments.