Selection of Trajectories to Improve Thermal Fields During the Electric Arc Welding Process Using Hybrid Model CFD-FNN

Abstract

Effective thermal management is essential in welding processes to maintain structural integrity and material quality, especially in high-precision industrial applications. This study examines the thermal behavior of an AISI 1080 steel plate containing 100 blind holes filled using robotic electric arc welding. Temperature measurements, recorded with eight strategically positioned thermocouples, monitored the thermal evolution throughout the robotic welding process. The experimental results validated a computational heat transfer model developed with ANSYS Fluent software to simulate and predict temperature distribution achieving a mean absolute percentage error (MAPE) below 4.53%. A feedforward neural network was trained with simulation-generated data to optimize welding sequences. The optimization focuses on minimizing the area under the thermal history curves, reducing temperature gradients, and mitigating overheating risks. Integrating CFD simulations and neural networks introduces a hybrid methodology combining precise numerical modeling with advanced predictive capabilities. The hybrid CFD-FNN results reached a determination coefficient (R2) of 0.93 and an MAPE of 3.5% highlighting the potential of this approach to predict the thermal behavior in multipoint welding processes. This model generated optimized welding trajectories improving the uniformity of the temperature field, reducing thermal gradients and minimizing temperature peaks, thus aiding in preventing overheating. This framework represents a significant advancement in welding technologies, demonstrating the effective application of deep learning techniques in optimizing complex industrial processes.