The reliable joining of thermoplastic composites is a critical requirement in modern manufacturing, where achieving zero-leakage joints is essential. For this application, ultrasonic welding is a highly efficient technology. Traditionally, standard heuristic methods and static experimental designs are used to optimize machine parameters. However, the process exhibits high stochastic variability due to complex, nonlinear thermomechanical interactions, which significantly influence the final seal quality and the reliability of the entire production system. This paper presents a practical prediction-optimization framework using a hybrid stacking ensemble neural network to process welding data. To improve the accuracy and stability of the manufacturing process, the predictive model is integrated with a Monte Carlo simulation. Evaluation showed that the proposed framework achieved the best performance among the evaluated benchmark models, with a coefficient of determination R² = 0.8523 and a mean absolute error MAE = 0.7224. The proposed framework identifies candidate optimized machine parameters in a simulation-based workflow and defines stable operating conditions for subsequent experimental validation, providing a bounded data-driven approach for minimizing leakage in ultrasonic welding.