Abstract Deep learning offers new methodologies and tools for meteorological modeling, yet its application in urban canopy models remains nascent. This study develops an Urban Canopy Neural Network (UCNN), a deep learning‐based urban canopy model designed for complex building morphologies. When training the UCNN, the Town Energy Balance (TEB) model was used to simulate land‐atmosphere exchange fluxes of diverse building morphologies. We designed 5,000 unique combinations of six key building parameters (roughness length, building fraction, green space fraction, mean building height, ratio of vertical‐to‐horizontal surface area, and albedo) in the TEB model within their respective reasonable ranges. These simulations were driven by meteorological data from 10 urban flux tower sites across globally diverse regions, generating a total of 14,400,000 data samples. To evaluate generalization, the UCNN was tested against an independent set of observational flux data from 11 additional global sites with diverse morphologies. The UCNN demonstrated excellent generalization, achieving Mean Absolute Errors of 2.18, 7.89, 16.31, and 7.31 for upwelling shortwave radiation, upwelling longwave radiation, sensible, and latent heat fluxes, respectively. This performance is comparable to mainstream models in international intercomparison projects. Furthermore, the UCNN achieves these results with superior computational efficiency, running approximately five times faster than the original TEB model. This study presents an effective deep learning pathway to address complex urban morphologies.
Meng et al. (Sun,) studied this question.