Sampling design shapes reliability in pedestrian wind prediction

Pedestrian-level wind assessment is increasingly needed during early urban design, but machine-learning surrogate models depend critically on how computational fluid dynamics (CFD) training data are sampled and how many cases, each representing one CFD simulation for a specific geometry and inflow combination, are available. This study quantifies how sampling topology and dataset scale shape surrogate reliability for pedestrian-level wind prediction around a T-shaped residential tower. An experimentally checked and numerically assessed CFD baseline was resampled using Aligned, Staggered, Unstructured, and Sunflower topologies and evaluated across common dataset scales from 200 to 5,000 cases. Four surrogate model families were compared under the same 80%/20% train–test split and evaluation protocol. The results show topology-dependent convergence despite identical dataset-scale levels: estimated transition ranges are approximately 2.1 × 10³ to 2.2 × 10³ cases for Sunflower, 2.5 × 10³ to 2.8 × 10³ for Staggered, 3.0 × 10³ to 3.2 × 10³ for Unstructured, and 3.5 × 10³ for Aligned sampling. At approximately 2,000 cases, the root-mean-square error reduction is 20.9%–30.4% when switching from Aligned to Sunflower sampling across model families. These findings support the joint consideration of sampling topology and dataset scale in the development of reliable, budget-aware surrogate models for pedestrian-level wind prediction.