Abstract: This Addressing the high-dimensional nonlinear coupling problem between process parameters and weld performance in robotic arc welding, this study designed an orthogonal experimental scheme with four factors (welding speed, gas flow rate, welding current, and oscillation amplitude) at five levels. Penetration width and depth were selected as key weld performance indicators to construct a systematic experimental dataset. Based on this dataset, the performance of various data-driven predictive models was comparatively analyzed. Results indicate that the proposed hybrid model combining a Back-Propagation neural network with a Belief Rule Base (BP–BRB) achieves the highest prediction accuracy on the test set. The sum of mean squared errors (MSE) is significantly reduced from 1.661445 (achieved by the conventional BP model) to 1.042846, representing an accuracy improvement of approximately 37.23%, thereby effectively enhancing the modeling capability for the complex welding process. Furthermore, for multi-objective optimization of process parameters, the optimization performance of Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and their hybrid algorithm (PSO–GA) was compared. Experimental results demonstrate that the PSO–GA hybrid algorithm exhibits superior convergence speed and global search capability. Compared to the individual PSO and GA algorithms, its comprehensive optimization performance is improved by approximately 28.55% and 25.99%, respectively. This research provides a data-driven reference for the selection and multi-objective optimization of process parameters in robotic TIG welding.