Compound parametric metamodelling of large-eddy simulations for microscale atmospheric dispersion

In pollutant dispersion problems, mapping concentrations in the first tens or hundreds of meters from the source still remains a modelling challenge. Large-eddy simulations (LES) are able to represent time and space variability of turbulent atmospheric flow, which is of prime importance to assess public short-term exposure. However, they remain far from real time and subject to uncertainties, in particular to parametric uncertainties associated with the large-scale atmospheric forcing and the emission source position. In this work, we show that an efficient and accurate metamodel of the tracer concentration information provided by LES and encapsulating their associated uncertainties can be built using appropriate statistical tools combining machine learning and principal component analysis. We present a proof-of-concept study based on a simplified but representative flow configuration (two-dimensional flow around a surface-mounted cube) using the AVBP LES solver and testing a variety of metamodels (linear regression, Gaussian processes, random forest, gradient boosting, etc.). Results reinforce the idea that for sufficiently statistically-converged quantities of interest and for a sufficiently large LES data set, a compound surrogate model can succeed in synthesizing information from the LES in the whole computational domain (with a Q² predictivity coefficient above 90 %). Downstream of the obstacle, the Q² coefficient of all metamodels reaches excellent results over 90%. Upstream, the tracer concentration is subject to strong discontinuities; combining metamodels allows to guarantee a good predictivity coefficient over 75%.