Radiative transfer is frequently cited as the slowest part of an atmospheric model; while perhaps a little unfair (radiation accounting for only 3.5% of the cost of ECMWF's highest resolution operational model), there is no shortage of ideas in the literature for speeding up radiation schemes. In this talk I will describe a flexible tool "ecCKD" for generating gas optics models using the correlated k-distribution (CKD) method, and in particular explore the potential to use the "full-spectrum correlated-k" (FSCK) method in NWP. Via the use of one band for the entire thermal infrared and one for the entire near-infrared, FSCK enables the number of pseudo-monochromatic spectral intervals to be drastically reduced, from over 100 in each of the shortwave and longwave spectra (in the current operational gas optics scheme at ECMWF) to around 25, with a corresponding factor-of-4 speed-up in the radiation scheme. Training against 50 line-by-line test profiles and 34 greenhouse gas scenarios ensures that the resultings gas-optics models are accurate for the full range of terrestrial conditions, plus climate conditions from the last glacial maximum up to 8x pre-industrial CO2 concentrations. Care must be taken to ensure there are sufficient spectral intervals to represent the spectrum of cloud absorption and scattering, as well as surface albedo variations. The reduced number of spectral intervals leads to increased noise in the stochastic McICA solver, but recent optimizations of the Tripleclouds solver make it as fast as McICA but free from stochastic noise. Finally, the new gas optics scheme is demonstrated online in forecasts by the ECMWF model.
