A CycleGAN-Pix2pix framework for multi-objective 3D urban morphology optimization: enhancing thermal performance in high-density areas

Abstract

Enhancing understanding of urban thermal environments is crucial for reducing energy consumption, improving resident comfort, and mitigating urban heat island effects. Previous studies rarely addressed the systematic optimization of local-scale three-dimensional (3D) urban spaces for thermal improvements. This study developed an integrated CycleGAN-Pix2pix-based (CP-GAN) model chain to automatically generate 3D urban morphology coupled with Local Climate Zones (LCZ)and perform performance assessment as well as morphology optimization. The approach was applied to six typical sites in Hong Kong to optimize four variables of urban thermal environments: summer solar radiation (SSR), Universal Thermal Climate Index (UTCI), optimal UTCI area (UTCIA), and floor area ratio variance (FARV). The results showed that (1) CP-GAN model achieved a 0.028 higher average SSIM compared to the previous Pix2pix model, validating its efficiency. (2) The iterated optimal Pareto solutions significantly improved thermal performance compared to the initial morphology, particularly in samples characterized by compact mid-rise buildings, achieving the best optimization results. (3) Four primary optimization strategies were identified: increasing building heights in a suitable interval, reducing building volumes and their variability, expanding green spaces, and arranging building layout rationally. This integrated framework supports sustainable urban design and regeneration, contributing to more livable and resilient cities.