Integrating generative AI and climate modeling for urban heat island mitigation
Mo Wang, Ziheng Xiong, Shiqi Zhou, Jiayu Zhao, Chuanhao Sun, Yuan-Kai Wang, Lie Wang, Soon Keat Tan
Abstract
Conventional urban heat island (UHI) studies often rely on static urban morphology inputs and oversimplified design variables, limiting their ability to support dynamic, climate-responsive urban planning. To address this gap, this study proposes a novel framework that integrates a hybrid generative adversarial network (GAN) with the Urban Weather Generator (UWG) for high-fidelity 3D urban form generation and microclimate simulation. The proposed GAN architecture combines the geometric accuracy of Pix2pix with the style refinement capability of CycleGAN, achieving improved morphologicalrealism (SSIM = 0.754, R 2 = 0.834 against ground-truth data) and resolving key distortions that impede microclimate analysis. Applied Shenzhen Bay Super Headquarters as a case study, ten urban development plans were generated and evaluated for their thermal performance. Results revealed that plans exceeding a facade-to-site ratio of 5.0 and footprint density of 0.30 showed intensified nocturnal heat retention, with Plan V exhibiting a + 2.3 °C nighttime temperature increase. In contrast, Plan I, with lower morphological density, achieved a 1.8 °C reduction, demonstrating superior heat dissipation. These insights provide actionable guidelines for climate-responsive urban planning, advocating for lower-density layouts with optimized facade exposure and increased vegetative cover. The proposed framework offers a robust tool for planners and policymakers to assess and design urban forms that enhance climate resilience while reducing UHI intensity.