Process Optimization and Microstructure in High-Speed Coaxial Dual-Laser Welding of SUS301 Thin Sheets Using an...

To predict weld geometry and clarify structure-property relationships in high-speed coaxial dual-laser butt welding of 1 mm-thick SUS301 stainless steel sheets, an SSA-BP neural network model was established to describe the nonlinear correlation between welding parameters and weld morphology. The model related continuous laser power, welding speed, pulse frequency, and pulse width to weld width and penetration depth. To improve the transparency of model validation, conventional BP and SSA-BP models were compared using the same independent test set, and five-fold cross-validation was performed using the original experimental samples. On the independent test set, the SSA-BP model achieved an overall correlation coefficient of R = 0.960, with RMSE values of 0.0561 mm and 0.0439 mm for weld width and penetration depth, respectively. Compared with the conventional BP model, SSA-BP reduced the overall RMSE, MAE, and MAPE by 25.9%, 36.4%, and 29.6%, respectively. The five-fold cross-validation further indicated stable prediction performance under different data partitions. Based on the predicted and experimentally measured weld geometry, candidate parameter sets were screened according to the weld aspect ratio (Φ = h/w). Within the present experimental window, joints with Φ = 0.82-0.84 showed more stable weld formation and relatively higher ultimate tensile strength (1211.4-1264.8 MPa) than two representative joints outside this interval (796.0 MPa at Φ = 0.63 and 1061.1 MPa at Φ = 0.88). Therefore, this interval should be regarded as a favorable empirical range under the present welding conditions rather than a universal optimum. Fractographic observations of a representative high-strength joint showed abundant dimples and tear ridges, indicating ductile fracture characteristics. EBSD analysis further revealed a graded microstructure from the weld center to the base metal. The weld center and fusion line-adjacent regions exhibited relatively high fractions of high-angle grain boundaries (66.2-70.6%), while phase distribution, GND density, and KAM maps indicated a gradual phase transition and localized but non-continuous strain concentration features across the joint. These results indicate that the present approach provides an effective route for weld geometry prediction and for linking morphology screening with tensile response and microstructural heterogeneity in SUS301 thin sheet welding.