A statistical physics approach to perform fast highly-resolved air quality simulations – A new step towards the meta-modelling of chemistry transport models

Abstract

A methodology rested on model-based machine learning using simple linear regressions and the parameterizations of the main physics and chemistry processes has been developed to perform highly-resolved air quality simulations. The training of the methodology is (i) completed over a 6-month period using the outputs of the chemical transport model CHIMERE, and (ii) then applied over the subsequent 6 months. Despite rough assumptions, this new methodology performs as well as the raw CHIMERE simulation for daily mean concentrations of the main criteria air pollutants (NO2, Ozone, PM10 and PM2.5) with correlations ranging from 0.75 to 0.83 for the particulate matter and up to 0.86 for the maximum ozone concentrations. Some improvements are investigated to expand this methodology to several other uses, but at this stage the method can be used for air quality forecasting, analysis of pollution episodes and mapping. This study also confirms that including a minimum set of selected physical parameterizations brings a high added value on machine learning processes.