Rapid Pedestrian-Level Wind Prediction Across Urban Building Layouts Using a Geometry-Conditioned Multi-Scale Neural Operator

Convolutional and neural-operator surrogates enable rapid urban wind prediction, but their apparent accuracy can depend on physical-scale inconsistencies, related-layout leakage, and the mixture of idealized and realistic cases. The scientific question is whether a geometry-conditioned local–spectral model provides a reproducible improvement once these evaluation risks are controlled, and whether that improvement persists in realistic neighborhoods. We combine an audited UrbanTALES representation, grouped evaluation, and Geo-MSO to predict two-dimensional pedestrian-level velocity statistics and turbulent kinetic energy. The 538 cases comprise 224 idealized and 314 realistic configurations; inputs use fixed physical windows and targets are normalized by friction velocity. On 81 geometry-held-out cases, Geo-MSO achieved a wind-speed MAE of 0.1275 ± 0.0005 m s−1, with baseline-relative mean reductions of 2.15% versus U-Net and 6.62% versus FNO. However, among the 47 realistic test cases, it outperformed FNO in only 51.1%, with a median FNO-minus-Geo-MSO difference of approximately 0.00010 m s−1. Aggregate improvement therefore does not establish consistent realistic-domain superiority. All eight primary error diagnostics favored Geo-MSO over FNO across three seeds, whereas most U-Net paired intervals included zero. Unequal parameter counts and conditioning also prevent attribution to a single architectural component. Idealized pretraining accelerated target convergence without improving final realistic accuracy; narrow street-canyon jets remained underpredicted. Measured end-to-end inference took 0.114–1.747 s on an RTX 4090 D. The empirical contribution is a reproducible, morphology-qualified assessment of a combined model-and-evaluation pipeline, supporting preliminary screening rather than peak-wind certification or replacement of validated CFD.