Machine Learning-Based Fatigue Life Prediction using Weld Geometry of Lap GMAW Joints

Automotive chassis components, typically fabricated using Gas Metal Arc Welding (GMAW) in lap joint configurations, play a critical role in supporting the vehicle body, transmitting power, and maintaining stability during operation. Due to exposure to vibration and cyclic loads, these components require superior fatigue durability. Since the fatigue durability of lap joints is primarily governed by weld geometry, even slight variations necessitate repeated fatigue testing, which involves considerable time and cost. Therefore, developing a reliable model capable of quantitatively predicting fatigue life based on weld geometry has emerged as an urgent research need.Recently, data-driven predictive models using machine learning have attracted significant attention. In this study, a non-neural network-based machine learning approach was employed to predict the fatigue life of lap joints using weld geometry information as input features. Lap joint welds with varying geometries were fabricated under different welding conditions and evaluated through fatigue testing. Subsequently, multiple regression models were developed and compared using Mean Absolute Percentage Error (MAPE) and the coefficient of determination (R²). The analysis revealed that linear regression models exhibited limited accuracy due to structural simplicity, whereas Support Vector Machine (SVM) models with nonlinear kernels showed superior performance. Among ensemble methods, the Bagged Tree model yielded stable predictions, while the Boosted Tree model suffered from error sensitivity. Within Gaussian Process Regression (GPR) models, the Exponential kernel achieved the highest accuracy, with an R² of 0.9980 and a MAPE of 0.53%, confirming its effectiveness for fatigue life prediction in lap joint welds.