In the field of design optimization for complex multibody dynamic systems, optimization based on high-fidelity dynamic simulation models often incurs high computational costs. Surrogate models can significantly reduce computation time and improve optimization efficiency while maintaining acceptable prediction accuracy. However, single surrogate models exhibit limitations when applied to diverse engineering optimization problems. Constructing hybrid surrogate models by integrating multiple modeling techniques offers an effective solution. This study proposes a hybrid surrogate modeling approach that combines an Improved Sparrow Search Algorithm-optimized Backpropagation Neural Network (ISSA-BP) with Gaussian Process Regression (GPR)-based residual correction. First, the standard Sparrow Search Algorithm is enhanced using Tent chaotic mapping and a rooster-based search strategy to optimize the initial weights and biases of the BP neural network, thereby improving both prediction accuracy and convergence speed. Subsequently, a GPR model is employed to correct the prediction residuals of the ISSA-BP model, further enhancing overall predictive performance. The proposed method is validated using several benchmark test functions and a quarter vehicle–seat–occupant dynamic system. The results demonstrate that the ISSA-BP model outperforms the conventional BP neural network, and the residual-corrected ISSA-BP-GPR hybrid surrogate model achieves superior accuracy, predictive capability, and robustness compared to single-model approaches. This method effectively meets the demands of dynamic design optimization in complex multibody systems.