Harnessing machine learning and deep learning methods to forecast whitecap fraction and sea-salt aerosol emissions in the ECMWF Integrated Forecast System (IFS-COMPO)

Within the Copernicus Atmosphere Monitoring Service (CAMS), ECMWF operates the Integrated Forecasting System with atmospheric composition extension (IFS-COMPO) to provide global forecasts and reanalysis of aerosols and trace gases. Emissions of sea-salt aerosols in IFS-COMPO are estimated by first computing the whitecap fraction, using a polynomial fit between a dataset of retrieved whitecap fraction from remote sensing and wind speed and sea-surface temperature (SST), and applying a shape function on the whitecap fraction to derive sea-salt aerosol emissions.

In the context of the Horizon Europe CAMAERA (CAMS AERosol Advancement) project, we apply a range of deep learning and machine learning algorithms to estimate whitecap fraction offline, using a two-year long dataset of whitecap fraction derived from remote sensing observations. Meteorological and oceanic predictors are used, including wind speed and direction, sea-surface temperature, significant wave height from wind- and total-sea, as well as the turbulent energy of breaking waves. The latter two parameters are provided by the wave model (WAM) that is included in IFS-COMPO. For some of the deep-learning and machine learning methods, the correlation and error of the estimated whitecap fraction are much improved as compared to the usual physical models used in the atmospheric composition and remote sensing communities. 

This work can be seen as a benchmark of machine learning/deep learning methods for the simulation of atmospheric composition processes. This expertise will be used for other processes such as desert dust emissions, in the CAMAERA project.