Differential Evolution-Optimized Multi-Output Support Vector Regression-Based Prediction of Weld Bead Morphology in Wire-Fed Laser-Arc Directed Energy Deposition of 2319 Aluminum Alloy

Wire-fed laser-arc directed energy deposition (Wire-fed LA-DED) technology improves production speed while maintaining high quality and is particularly suited for manufacturing large, complex aluminum or titanium alloy components. The geometry of the weld bead (height and width) is influenced by multiple intricate parameters and variables during the manufacturing process. Accurately predicting the weld bead shape enables precise control over the surface flatness of the part, helping to prevent defects such as lack of fusion. This significantly reduces dimensional redundancy, enhances printing efficiency, and optimizes material usage. In this study, a quadratic regression prediction model for weld bead geometry was developed using the response surface methodology (RSM), with predictions generated through several machine learning models. These models included the backpropagation neural network (BPNN), support vector regression (SVR), multi-output support vector regression (MOSVR), extreme learning machine (ELM), and a differential evolution-optimized MOSVR (DE-MOSVR) model. Grid search and cross-validation techniques were utilized to identify the optimal parameters for each model to achieve the best predictive performance. A comparison of these models was conducted, followed by an evaluation of their generalization capabilities using an additional 20 sets of test data. The most accurate predictive model was selected based on a comprehensive assessment. The results showed that the DE-MOSVR model outperformed the others, achieving mean squared error, root mean squared error, mean absolute error, and R ² values for width (height) predictions of 0.0411 (0.0041), 0.2028 (0.0639), 0.1671 (0.0550), and 0.9434 (0.9433), respectively. It demonstrated the smallest deviation in the validation set, with mean deviations of 1.97% and 1.68%, respectively. The model we developed was validated through the production of prototype parts, providing valuable reference and guidance for predicting and modeling weld bead morphology in the Wire-fed LA-DED process.