GAN Predictability for Urban Environmental Performance: Learnability Mechanisms, Structural Consistency, and Efficiency Bounds

Generative adversarial networks (GANs) can rapidly predict urban environmental performance. However, most existing studies focus on a single target and lack cross-performance comparisons under unified conditions. Under unified urban-form inputs and training settings, this study employs the conditional adversarial model pix2pix to predict four targets—the Universal Thermal Climate Index (UTCI), annual global solar radiation (Rad), sunshine duration (SolarH), and near-surface wind speed (Wind)—and establishes a closed-loop evaluation framework spanning pixel, structural/region-level, cross-task synergy, complexity, and efficiency. The results show that (1) the overall ranking in accuracy and structural consistency is SolarH ≈ Rad > UTCI > Wind; (2) per-epoch times are similar, whereas convergence epochs differ markedly, indicating that total time is primarily governed by convergence difficulty; (3) structurally, Rad/SolarH perform better on hot-region overlap and edge alignment, whereas Wind exhibits larger errors at corners and canyons; (4) in terms of learnability, texture variation explains errors far better than edge count; and (5) cross-task synergy is higher in low-value regions than in high-value regions, with Wind clearly decoupled from the other targets. The distinctive contribution lies in a unified, reproducible evaluation framework, together with learnability mechanisms and applicability bounds, providing fast and reliable evidence for performance-oriented planning and design.