Comparison between laser welding modelling approaches: from low fidelity numerical simulation to high fidelity and Physics informed neural networks

This paper presents a systematic comparison of three modelling frameworks for predicting weld bead geometry in laser beam welding: a low-fidelity (LF) thermal finite element model, a high-fidelity (HF) coupled thermo-fluid model, and a physics-informed neural network (PINN). The approaches are implemented, calibrated, and validated against experimental cross-sections from overlap welding of 6082-T6 aluminum alloy sheets. The influence of model architecture, physical assumptions, and mesh resolution on predictive accuracy and computational cost is discussed. Results show that while the HF model captures complex phenomena such as keyhole dynamics and surface deformation, both the LF model and the PINN achieve comparable accuracy in predicting weld bead dimensions at significantly lower computational cost. For PINNs specifically, a non-monotonic dependence of training performance on network architecture is identified, with an instability threshold linked to the total number of trainable parameters. These findings provide practical guidelines for model selection in laser welding process optimization.