Modelling and simulation of atmospheric dispersion are essential to ensure the safety of emissions emitted into the air by the authorized operation of industrial facilities and to estimate the health consequences of accidents that could affect these facilities. Over the past twenty years, physical dispersion models have undergone significant improvements in order to take into account the details of topography and land use that make real industrial environments complex. Although 3D models have seen their use increase, they have very significant calculation times, which hinders their use in multi-parametric studies and the assessment of uncertainties that require a large number of calculations. It would therefore be desirable to obtain the very precise results of current models or similar results in a much shorter time. Recently, we have developed a strategy consisting of reducing the dimension of distribution maps of an atmospheric pollutant obtained using a reference 3D physical model for different meteorological conditions, then having these maps learned by an artificial intelligence (AI) model which is then used to predict maps in other meteorological situations. The postdoctoral project will focus on complementing the research started by evaluating the performance of dimension reduction and model substitution methods already explored and by studying other methods. Applications will concern, in particular, the simulation of concentrations around an industrial production site that emits gaseous emissions into the atmosphere. The developments will aim to obtain an operational meta-modelling tool.