Development of a Predictive Model for Urban Cooling Blue Infrastructure using Integrated CFD and U-NET CNN

The objectives of this study are to develop a predictive modeling framework for evaluating the cooling performance of fog-based blue infrastructure using an integrated CFD and U-Net CNN approach. Urban heat stress has intensified with rapid urbanization and climate change, creating demand for localized and effective cooling strategies beyond conventional grey infrastructure. Fog-cooling systems offer rapid and flexible cooling through evaporation, latent heat flux, and induced air mixing, yet their performance is high-fidelity CFD simulations. To address this challenge, this study constructed a high-resolution CFD model for Gwacheon Sports Park and simulated 33 parametric scenarios by systematically varying air temperature (25-35°C), relative humidity (50, 70, 90%), and wind conditions. The resulting three-dimensional cooling fields were transformed into structured tensors and used to train an Attention U-Net surrogate model designed to efficiently capture both global and localized cooling patterns. This study reveals that fog cooling reduces pedestrian-level air temperature by approximately 0.5-0.75°C and area averaged temperature by 0.37-0.53°C, with stronger cooling occurring under higher ambient temperature. This study finds that thermal comfort consistently improves, with UTCI reductions of 0.46-0.73°C across all meteorological conditions, indicating meaningful mitigation of outdoor heat stress even under high humidity. This study also demonstrates that the Attention U-Net model accurately reproduces CFD-derived cooling fields while reducing computation time to below 0.1s, enabling rapid scenario screening for microclimate-sensitive urban design. However, this study should address several remaining limitations in future research, including the incorporation of transient CFD physics, expansion of the training dataset to broader environmental conditions, and evaluation of multi-source fog configurations to enhance the robustness, generalizability, and operational relevance of the proposed hybrid modeling framework.