Background
Urban-microclimate CFD is physically credible but each scenario takes hours to days — making thousand-scale design-space exploration practically infeasible.
Existing ML surrogates either force urban geometry onto regular grids or treat spatial elements independently. But urban microclimate is governed by neighborhood interactions: canyon effects, wind corridors, vegetation shading.
Approach
- Urban blocks represented as graphs — nodes = buildings/roads/vegetation, edges = spatial proximity
- GATv2 (Graph Attention Network v2) with attention-weighted message passing
- Training data: 30 Seocho-gu blocks × 1,500 OpenFOAM CFD scenarios
- Three-stage transfer learning evaluated on 10 adjacent Gangnam-gu blocks
- Optuna Bayesian optimization sweeps tree spacing, LAI, green-roof coverage
- Budget constraint: 50 million KRW per block
Key Results
| Metric | Value | |---|---| | Zero-shot transfer R² | 0.845 | | Few-shot (10-block fine-tune) R² | 0.887 | | Full retraining R² | 0.962 | | Optimal tree spacing | 5–8 m | | Optimal LAI | 4–5 | | Block-average cooling | 0.23 °C | | Block-maximum cooling | 1.98 °C |
Implications
- In the cost of one CFD run, the GNN can compare thousands of scenarios.
- Adjacent districts with no training data still get usable predictions out-of-the-box (zero-shot R²=0.845).
- Recommendations move from qualitative ("plant trees") to quantitative ("5–8 m spacing, LAI 4–5").
Related solutions
- Urban Thermal Diagnosis (CFD · GNN) — core technology stack
- Spatial Optimization of Green Infrastructure — extend Bayesian → NSGA-II multi-objective