Abstract

Large-eddy simulation (LES) is essential for urban microclimate studies, resolving complex,
transient turbulent structures. However, LES remains computationally expensive, particularly
for real-time operations and early-stage iterative design. To address this, we developed
a generative surrogate model using Conditional Flow Matching (CFM) with optimal transport
to reconstruct and generate plausible instantaneous multi-variable urban microclimate
snapshots. The model learns a time-dependent vector field to transport Gaussian noise into
physically realistic flow fields, conditioned on geometrical information and steady-state wind
patterns. To mitigate exorbitant 3D training costs, we implemented a patch-based generation
strategy using overlapping 64 × 64 × 64 voxel patches. Shared-noise initialization and
a variance-preserving weighted averaging scheme are employed during ODE integration to
ensure spatial continuity across full urban domains. Our methodology produces diverse 3D
multi-channel (u, v,w, T) snapshots, robustly reproducing input steady-state fields with a
Normalized RSME of 2.68 % for wind and 0.22 % for temperature and capturing turbulence
stochasticity, maintaining standard deviation differences of only 0.30 m/s and 0.33 ◦C compared
to CityFFD simulations. Validation against first-order, second-order statistics, and
local wind profiles confirms that our generative approach effectively preserves high-frequency
fluctuations and statistic globally and locally. This reconstruction method provides an efficient
path to LES-quality instantaneous fields at significantly reduced costs, bypassing the
out-of-distribution (OOD) generalization limits of deterministic AI models across diverse
urban geometries.