Abstract: In the aerodynamic shape design of lifting-body vehicles, integrating aerodynamic performance evaluation into the design loop enables real-time, performance-driven fine-tuning of the vehicle geometry, thereby facilitating efficient iterative optimization. However, if the aerodynamic analysis step remains computationally expensive, it can still hinder the overall efficiency of subsequent shape updates. Addressing this challenge requires a surrogate model that balances high predictive accuracy with computational speed, effectively replacing conventional computational fluid dynamics (CFD) methods in the aerodynamic evaluation stage. This paper presents a surrogate model for aerodynamic force prediction based on the DeepONet architecture. The model is trained to predict aerodynamic forces acting on lifting-body configurations parameterized by a set of geometric descriptors. A key innovation of the proposed approach lies in the incorporation of a geometry-parameter encoding layer, which preprocesses the geometric descriptors to align their latent representation with the dimensionality of the operating-condition parameters. This alignment ensures that both geometric and operational factors are jointly accounted for in the prediction. Furthermore, the branch network of DeepONet is leveraged to inject global information from the entire data space into the training of individual data subspaces, enhancing model generalization. Numerical results demonstrate that the DeepONet-based surrogate model achieves mean prediction errors of less than 3% for primary aerodynamic forces, indicating high predictive fidelity and significant improvement in accuracy compared to non-neural-network approaches such as Kriging and Radial Basis Function (RBF) interpolation. Compared to general deep neural network architectures, the proposed DeepONet-based model reduces the overall average prediction error by 16.4%, underscoring its superior performance in aerodynamic force estimation for lifting-body vehicles.