Offline prediction of microstructure and mechanical properties of 7075 aluminum alloy double-pulsed MIG welds using phase space reconstruction and an improved MLP

To address the strong nonlinearity and small-sample characteristics of the double-pulse metal inert gas (DP-MIG) welding process for 7075 ultra-high-strength aluminum alloy, a weld microstructure–property prediction method based on phase space reconstruction (PSR) and an improved multilayer perceptron (IMLP) is proposed. AA7075-T651 aluminum alloy was used as the base material, and ER5356 welding wire served as the filler material. The effects of welding current on joint microstructure, mechanical properties, and fracture behavior were systematically investigated, and 190 A was identified as the optimal welding current. Based on experimentally measured welding process parameters, rescaled range (R/S) analysis was employed to verify the nonlinear chaotic characteristics of the welding process. According to Takens’ embedding theorem, a 15-dimensional fused dynamic feature vector was constructed as the model input. The IMLP incorporates residual blocks, a scaled dot-product self-attention gating mechanism, and a triple-regularization strategy (dropout, L2 weight decay, and batch normalization), thereby effectively addressing gradient vanishing, equal-weight feature treatment, and overfitting. A total of 100 welded joints (100 experimental samples) were fabricated and characterized, and the dataset was partitioned by stratified random sampling into 80 training samples and 20 independent test samples (8:2 ratio); five-fold cross-validation was applied to the 80 training samples for hyperparameter tuning and model selection. On the 20-sample independent test set, the IMLP achieved a coefficient of determination ( R 2 ) of 0.860 and a mean absolute percentage error (MAPE) of 5.14%, outperforming Ridge regression with polynomial features (Ridge-Poly), improved particle swarm optimization–support vector regression (IPSO-SVR), and the standard multilayer perceptron (MLP). These results, obtained for the AA7075-T651/ER5356 DP-MIG welding configuration investigated here, demonstrate that the proposed PSR-IMLP method is a promising data-driven approach for small-sample, strongly nonlinear, multi-objective weld-property prediction; validation on additional alloys, filler materials, and welding processes is required before the approach can be considered generally applicable.