Wind model validation
At infrared.city, we build AI models that predict urban wind speeds fast enough for real-time planning workflows, without waiting days for traditional CFD simulations. Today, we're sharing results from our latest model, which we tested across a wide range of urban geometries.
The new model achieves an average Index of Agreement of 0.988 and a mean absolute error of just 0.31 m/s, tested against high-fidelity CFD simulation.
Performance at a glance
| Metric | Value |
|---|---|
| Mean Absolute Error (MAE) | 0.31 m/s |
| Index of Agreement (IA) | 0.988 |
Benchmarked against high-fidelity CFD simulation across diverse urban scenarios.
IA of 1.0 = perfect match.
What we're predicting
The model takes an encoded urban geometry and outputs a full wind speed field at pedestrian level. No location-specific tuning required: the model generalises across urban typologies globally, from dense street grids to sprawling mid-rise districts.
How we validated it
We compared predictions against CFD simulations across diverse urban scenarios. Each case was evaluated on two metrics:
- MAE, mean absolute wind speed error in m/s
- Index of Agreement, measures how closely the predicted spatial pattern matches the reference simulation, capturing both magnitude and distribution accuracy
Speed
Traditional CFD simulations can take hours to set up and run. Our model delivers a comparable wind pattern in seconds, making it practical for iterative design, early-stage masterplanning, and large-scale urban analysis.
What comes next
We're continuing to push accuracy further and expand validation coverage. If you're working on wind comfort assessments, pedestrian safety, or urban microclimate analysis, we'd love to talk.