GIDPC.Green Infrastructure · Disaster Prevention
Research & Technology
Building and EnvironmentUnder Review

CFD-Informed Graph Attention Network Surrogate for Urban Thermal Environment Prediction with Multi-Block Transfer Learning

A GATv2 surrogate that explicitly models neighborhood interactions — trained on 1,500 OpenFOAM scenarios — achieves zero-shot R² = 0.845 across districts, and Bayesian optimization identifies 5–8 m street-tree spacing as the top cooling strategy.

Authors
Junsuk Kang
Year
2026

Keywords

Urban heat islandGraph neural networkCFD surrogate modelTransfer learningGreen infrastructure optimisation

Background

Urban green infrastructure (UGI) is a critical UHI mitigation strategy, but evaluating its thermal benefit at the block scale requires CFD runs of hours to days. Existing ML surrogates (XGBoost, CNN) treat spatial elements independently — they cannot capture the neighborhood interactions central to urban microclimate.

Method

  • GATv2 (Graph Attention Network v2) — attention-weighted message passing aggregates neighbor information
  • Urban blocks as graphs: nodes = buildings/roads/vegetation/open spaces, edges = proximity/adjacency/wind-corridor alignment
  • Training data: 30 Seocho-gu blocks × 1,500 OpenFOAM CFD scenarios (Boussinesq)
  • Transfer-learning evaluation on 10 adjacent Gangnam-gu blocks (zero-shot, few-shot, full retrain)
  • Optuna Bayesian optimization sweeps UGI configurations (tree spacing, LAI, green-roof coverage) under a 50 M KRW per-block budget

Key Results

| Metric | Value | |---|---| | Zero-shot R² | 0.845 | | Few-shot (10-block fine-tune) R² | 0.887 | | Full retraining R² | 0.962 | | Optimal tree spacing / LAI | 5–8 m / 4–5 | | Block-average cooling | 0.23 °C | | Block-maximum cooling | 1.98 °C |

Significance

Physics-informed GNNs can serve as scalable, transferable tools for urban thermal planning — comparing thousands of scenarios in the time of a single CFD run, and generalizing to adjacent districts without retraining.

Research & Technology