2D-Trained Models Transfer to 3D Aircraft Flows with Under 0.7% Drag Error

Researchers at an undisclosed U.S. institution trained both linear POD and nonlinear Least Volume autoencoders on two-dimensional airfoil and channel flows. The resulting 28-41 dimensional latent spaces were applied zero-shot to extruded wings and full blended-wing-body geometries. POD reconstructions kept integrated forces within 0.1% lift and 0.7% drag of high-fidelity CFD; LVAE performed slightly worse on transfer despite better in-domain accuracy. A follow-on dual-encoder fine-tuning step using only five 3D cases reduced reconstruction MSE by an additional factor of five. The approach directly addresses the data scarcity that currently forces designers to run hundreds of expensive 3D RANS simulations during early configuration trades. Because the source data are cheap 2D slices, the method lowers the barrier to exploring novel planforms that would otherwise remain computationally inaccessible. The counter-intuitive superiority of linear POD on transfer suggests that the dominant coherent structures in external aerodynamics are still largely linear even when local separation is nonlinear. Next steps include embedding the transferred bases inside gradient-based shape optimization loops and testing whether the same 2D-to-3D mapping holds for unsteady maneuvers and transonic buffet. If validated, the technique could cut early-phase CFD budgets by an order of magnitude while preserving force accuracy required for certification.