Design and experimental validation of a low-temperature toughness-oriented welding wire for 690 MPa steel via BP neural network modeling

To meet the stringent requirements for strength and toughness of welded joints in high-strength steel structures operating under extreme low-temperature conditions, this study proposes an intelligent alloy design method for 690 MPa-class low-temperature high-toughness welding wire based on a BP neural network. A dataset comprising 104 sets of welding wire compositions and corresponding impact energy values at −50 °C was collected. A predictive model was constructed using alloying elements as input variables and low-temperature impact toughness as the output target. The model was optimized through training with a Bayesian regularization algorithm. Based on multiple performance metrics, the most accurate model structure was selected to predict the optimal composition range for the welding wire. Experimental wires were fabricated according to the predicted compositions and subjected to welding trials using the gas metal arc welding process. The results demonstrated that the deposited metal primarily consisted of acicular ferrite with refined grain structures, exhibiting excellent mechanical properties. The average impact energy at −50 °C reached 87 J, closely matching the predicted value. Further analyses using TEM, EBSD, and XRD revealed a synergistic enhancement mechanism involving microstructural refinement and carbide strengthening. This work indicates that the proposed method offers promising potential for the intelligent design of welding wires with well-balanced high strength and toughness.