This preprint presents a multi-regime surrogate modeling framework for aerodynamic CFD based on a Mixture-of-Experts (MoE) architecture. Aerodynamic flows often exhibit regime-dependent behavior driven by laminar–transitional–turbulent evolution, flow separation, and stall, which makes the input–output mapping inherently piecewise and difficult to approximate with monolithic surrogates such as MLPs, GPR, SVR, or random forests. To address this, we construct a synthetic CFD generator that emulates nonlinear aerodynamic responses across camber, thickness, Reynolds number, angle of attack, turbulence proxies, and mesh-fidelity factors, and we train an MoE surrogate on this controlled multi-regime dataset. The MoE achieves test errors roughly two orders of magnitude lower than conventional surrogates and eliminates the error spikes typically observed near stall. Gating visualizations show that individual experts specialize in attached, transitional, and stall-dominated flow regimes, providing physically interpretable decomposition of the aerodynamic domain. We also define an expert-disagreement–based uncertainty measure that naturally concentrates on regime boundaries and sparsely sampled regions, making the proposed MoE surrogate compatible with active-learning loops around high-fidelity CFD solvers (e.g. RANS/LES). The overall framework demonstrates that MoE is a robust, interpretable, and scalable surrogate modeling approach for complex aerodynamic regimes. This v2 update adds extended ablation studies on the number of experts and regime-regularization strength, detailed gating-map visualizations in the (α,Re)(\alpha, Re)(α,Re) plane, and regime-wise error analysis on the held-out test set. The updated results further clarify how the MoE specializes across aerodynamic regimes and how the regime-aware regularization improves stability and interpretability of the surrogate.