A Study on the Prediction Performance of the Temperature Field in the GMAW Process According to the Collocation Point Distribution of the Physics-Informed Neural Network (PINN)

In gas metal arc welding (GMAW), temperature field prediction is essential to prevent the deterioration of weld quality caused by distortion or cracking. However, conventional temperature prediction approaches are limited by high computational costs or dependence on large datasets. Recently, physics-informed neural networks (PINNs), which utilize governing equations without data, have gained attention as an alternative. Despite their potential, studies on PINN training strategies specifically designed for temperature prediction in GMAW remain limited. This study investigates the impact of collocation point distribution, a key factor in PINN training, on the accuracy of temperature field prediction in GMAW. Based on the characteristics of the welding process, collocation points were distributed according to three strategies: uniform distribution, concentration along the weld seam, and concentration along both the weld seam and the heat source. The prediction performance of each strategy was quantitatively evaluated through comparison with the results obtained from the finite element method (FEM). Among the models, the one with collocation points concentrated around the weld seam and heat source achieved the highest prediction accuracy, with an R² value of 0.89. Moreover, the results were analyzed with the training loss and the spatial distribution of collocation points.