Numerical model for non-penetrating laser welding of bipolar plates based on BP neural network

Laser welding technology has significant advantages in the precision machining and mass production of metal bipolar plates for fuel cells due to its characteristics of high precision and high efficiency. However, in the welding of ultra-thin metal bipolar plates, problems such as difficult-to-control weld quality and welding thermal deformation exist. To solve the above problems and optimize the welding process window more efficiently, a Gaussian surface-cylinder compound heat source model based on the non-penetrating welding of ultra-thin metal bipolar plates was constructed using COMSOL Multiphysics. A full factorial analysis scheme with four factors and five levels was constructed by adopting three parameters: penetration depth, weld width, and joint width. A BP neural network model was trained through JMP software combined with the actual weld pool morphology, making the modified simulation model closer to the actual welding weld pool. The test results show that when the modified laser heat source model is used for the welding process simulation, and the simulation results are compared with the actual bipolar plate welding test results, the relative errors between the simulation data and the test data are all within ±5%. This indicates that the model can well guide the actual engineering practice of the laser welding process for metal bipolar plates in the future.