Impact of a spectral sampling technique for radiation on ECMWF weather forecasts

Radiation transfer computations are one of the most expensive components of atmospheric models integrations. Methods generally applied to reduce their computational cost, include less frequent update of the radiative fluxes or coarser spatial grids than those used for the rest of the model. In the operational configuration of the Integrated Forecast System (IFS) used at the European Centre for Medium Range Weather Forecasts (ECMWF), radiation accounts for ∼ 10% of the total computer time, and the radiative fluxes are computed with both a reduced temporal and spatial resolution. In this study, we show that these approximations have a negligible impact on the forecast errors when looking at the large-scale circulation but they lead to biases in the surface temperature. These are linked to inconsistencies in the representation of surface properties or to poorly sampled temporal evolution of the radiative fluxes. Here we explore an alternative method to alleviate such errors, based on a recently developed technique that estimates radiative fluxes from randomly-chosen subsets of spectral points designed to minimize the surface flux error. This approach introduces substantial uncorrelated noise in the radiative fluxes at the surface and heating rates, but it reduces the cost associated to the spectral integration. Therefore, it allows to update the radiative transfer computations more frequently and to perform them on a denser spatial grid. We evaluate how this approach applies to 5 days weather forecasts by testing different combinations of temporal and spectral sampling at high spatial resolution. We assess the impact of each radiation configuration on the forecast, by comparing the near-surface temperatures to observations from synoptic stations. The results show that a combination of spectrally sparse but temporally and spatially dense radiation calculations has the potential to reduce the forecast errors, particularly those associated with underresolved topography and stable nocturnal boundary layers, with affordable computational costs.