A cross-configuration multi-fidelity framework for aerodynamic modeling: DeepONet with adaptive gated fusion and transfer learning

Accurate prediction of aerodynamic coefficients is essential for guiding aircraft design optimization and performance evaluation. However, existing multi-fidelity surrogate models are typically constrained to narrow geometric training distributions and rely on simplified fusion strategies that lack adaptive weighting. This inability to intelligently fuse heterogeneous fidelity data and extract universal shape-performance relationships causes dramatic performance degradation when evaluating novel configurations, severely limiting their applicability in diverse cross-configuration design scenarios. To address these challenges, we propose a cross-configuration multi-fidelity aerodynamic prediction framework termed DeepONet-GF-TL (Deep Operator Network with Adaptive Gated Fusion and Transfer Learning). The method employs the DeepONet operator architecture as its backbone to decouple geometric and flow-condition feature extraction; integrates an adaptive gated fusion module that dynamically evaluates local data reliability to optimally weight heterogeneous fidelity information across varying flow regimes; introduces a residual skip connection that directly routes raw inputs to the prediction layer, explicitly learning targeted corrections for complex physical mechanisms inherently absent in low-fidelity data; and incorporates a dual-pathway attention mechanism for refined feature representation. On this basis, a two-stage transfer learning strategy with a full-freeze mechanism enables robust cross-configuration generalization on scarce high-fidelity samples while preventing catastrophic forgetting of universal shape-performance mappings. Comprehensive evaluations on 11 subsonic-to-transonic airfoils demonstrate that the proposed DeepONet-GF-TL achieves lift and drag coefficient prediction R² exceeding 0.995, reducing MAE by 1.7 to 4.6 times compared to three established multi-fidelity baselines(MFGPR, MFNN, and TLNN). Tests on the CACM three-dimensional civil transport aircraft model with 20-dimensional geometric variations confirm robust cross-configuration generalization by accurately predicting the highly nonlinear support interference increments. The proposed method achieves an interference R² exceeding 0.953 for lift and 0.949 for drag across 450 unseen configurations. Moreover, the method achieves 3 to 4 orders of magnitude speedup over Computational Fluid Dynamics (CFD). These results demonstrate that the proposed framework enables accurate and efficient aerodynamic coefficient prediction across diverse geometries, fidelity gap types, and flow conditions.