GeoABC Exposes the Isotropic Blind Spot in Neural CFD Surrogates

The arXiv preprint (abs/2606.09963, June 2026) introduces GeoABC, a geometry-conditioned anisotropic boundary correction module that injects direction-aware modulation into neural operator layers. Unlike standard Fourier or DeepONet backbones that treat near-wall regions uniformly, GeoABC explicitly separates tangential propagation from normal-wall constraints using boundary geometry as a structural prior rather than mere input features. Experiments on 2D airfoil and 3D car datasets report an average 38% drop in near-boundary relative L2 error, yet the work remains a preprint with no peer review and provides no dataset sizes or training split details. This gap echoes limitations seen in the foundational Fourier Neural Operator paper (Li et al., 2020), where boundary handling was similarly isotropic and produced comparable near-wall discrepancies on incompressible flows. A further overlooked connection appears in classic boundary-layer theory (Schlichting & Gersten), whose anisotropy the model now operationalizes but without testing on unsteady or high-Reynolds regimes that dominate real aerospace certification. Consequently, while GeoABC narrows the structural near-wall gap, its gains may shrink on geometries or flow conditions absent from the two evaluation suites.