Rapid prediction of wind field in street canyons with infrastructure elements using neural operators

Rapid and accurate prediction of wind fields is critical for optimizing the wind environment in street canyons. However, most previous studies on rapid wind field prediction have focused on idealized geometric models and neglected the effect of infrastructure elements (e.g., envelope features, street furniture, and barriers), which limits their application in wind environment optimization. This study used the Fourier Neural Operator (FNO) and Deep Operator Network (DeepONet) models to rapidly predict wind fields in street canyons with different infrastructure elements. The results revealed that different element placements affect airflow patterns: the windward placement forms upper and lower recirculation zones, the leeward placement weakens the primary vortex, and the base placement disturbs the near-ground flow. Both models achieved high prediction accuracy and the relative L2 errors of DeepONet (5.56%–6.37%) were slightly lower than those of FNO (6.08%–7.12%). They effectively captured complex flow features, including primary recirculation zones, vortex structures, and wake regions. Errors aggregated in regions with sharp velocity gradients and flow separation around elements. Both models predicted leeward cases more accurately than windward ones. Moreover, the trained models could predict wind fields for unseen cases in approximately 0.02 s while maintaining relative L2 errors below 7.2%. The FNO model is suitable for early-stage design exploration because it only requires geometric information, whereas DeepONet is more appropriate for detailed design validation as it requires boundary velocity data. This study provides efficient methods for rapid assessment and design optimization of wind field in street canyons.