ADMS Urban is a non-linear static model whose input data $p$ varies one simulated hour after the other. The model computes a high-dimensional concentration vector $y= ℳ(p)$ which can contain 10⁵ concentrations. A full-year simulation of $NO_2$ concentrations can take dozens of days of computations, which greatly limits the range of methods that can be applied to the model, especially for uncertainty quantification. This work proposes a method to replace ADMS Urban with a so-called emulator, i.e., a close approximation of ADMS Urban whose computational cost is negligible. First, the output concentration field $y$ is projected on a few modes of a proper orthogonal decomposition $[Ψ_1 ... Ψ_N]$, so that $y \simeq ∑^N_{j=1} \alpha_j Ψ_j$ where $\apha_j$ is the projection coefficient on $j$-th mode and $N$ smaller than 10. Then, the reduced model $f(p) = Ψ^T ℳ(p)$ is replaced by a statistical emulator $\hat{f}$ so that $\hat{f}(p) \simeq f(p)$ and the computational cost of $\hat{f}(p)$ is negligible.