Urban microclimate: a generative model reproduces the wind and temperature fields of an LES simulation in a matter of seconds

Guided by the geometry of buildings, a flow-matching model generates realistic 3D turbulent fields – with an error of less than 3 per cent in mean wind speed – to help cities adapt to climate change.

Finely predicting wind and temperature between buildings is useful for designing neighborhoods adapted to heat waves and gusts. Large scale simulations (LES) know how to do this, but their cost of calculation makes them impracticable in an iterative design approach, and conventional learning models deliver only an average prediction that ignores the random nature of turbulence.

Generate rather than predict

Researchers propose a "conditional flow matching" generic model which, from the geometry of buildings and the average flow, produces instantaneous three-dimensional fields of speed and temperature within seconds. To bypass the GPU memory limit, the domain is cut into overlapping blocks and share the same noise initialization, which preserves the continuity of flow structures.

Compared to the reference LES data, the model recovers average statistics with a normalized quadratic error of 2.99% for wind and 1.77% for temperature, and turbulence magnitudes with approximately 7% to 9% error. An application to local gust forecasting illustrates the value of the approach for resilient urban design.