Hazardous atmospheric dispersion in urban areas: A Deep Learning approach for emergency pollution forecast

Today, Computational Fluid Dynamics approaches have a high level of spatial/temporal accuracy in modelling atmospheric transport and dispersion in very complex environments. Several numerical models require, however, heavy computational resources and prolonged simulation time up to several days. This time constraint is specifically crucial for intervention planning in case of accidental or malevolent toxic releases in a city. In this paper, we propose to use synthetic data generated by a realistic 3-D transport/dispersion simulator, to train a learning framework called MCxM. The latter relies on a sequence of masking and correction operations to progressively apply the spatial constraints and underlying physics of transport and dispersion. The learning phase uses the urban geometry of the French city Grenoble. We then test the effectiveness of the trained MCxM in a different French city: Paris. The results show that the MCxM's forecasts are virtually instantaneous and generalize successfully to unseen conditions.