ECMWF Newsletter #163

Continuous long-window data assimilation

Elías Hólm
Simon Lang
Peter Lean
Massimo Bonavita

 

In the next upgrade of ECMWF’s Integrated Forecasting System to IFS Cycle 47r1, the continuous data assimilation (Co-DA) introduced in IFS Cycle 46r1 will be extended to connect the early-delivery assimilation (DA) and long-window assimilation (LWDA) into one assimilation cycle, referred to as continuous long-window data assimilation (Co-LWDA).

In continuous data assimilation (Lean et al. in Newsletter No. 158), newly arrived observations are added to the successive minimisations of the early-delivery stream. In the current setup of the IFS data assimilation cycle, the DA analysis is only used to initialise the forecasts, and it does not feed into the subsequent LWDA analysis update (see the figure). The starting point for both the DA and LWDA analysis is a first guess provided by a short-range forecast from the previous LWDA analysis. That same short-range forecast also serves as the background, which together with its estimated errors enters into the data assimilation calculations.

However, since the DA analysis has already run by the time the LWDA analysis starts, the LWDA analysis can be started from a more accurate first guess provided by the DA analysis, while the background and its error estimate remain the same. This connects DA and LWDA into one assimilation cycle, with LWDA just providing the final four minimisations of an eight-outer-loop assimilation cycle and adding the observations available in the remaining 5 hours of the 12-hour assimilation window. In this framework, the LWDA analysis can be viewed as a time extension of the DA analysis. The Co-DA concept has thus been extended to include DA and LWDA in one assimilation cycle, and we refer to this as Co-LWDA.

The experiments for this change show all the common features seen when we increase the number of outer loops. The fit of the short-range forecasts to independent observations improves noticeably for observations such as water vapour, cloud and precipitation sensitive radiances. These are the observations that need more outer loops to extract their full information content. This improvement is significant and similar in structure to the improvement seen from going from four to five outer loops. There is also a significant increase in the number of observations used in LWDA, particularly infrared radiances, which are very sensitive to the correct position of clouds for their quality control. Improvement is also seen against SATOB, which measures wind from the movements of clouds and water vapour. Together this shows that clouds, water vapour and winds are more accurate in the short-range forecast. Beyond the short range, the effect will largely be to increase significantly the benefits that will arise when we are able to implement increased inner-loop resolution.

The connection of DA and LWDA into a single Co-LWDA assimilation cycle is a step towards further optimisation of the assimilation system. Co-LWDA enables more accurate configurations at the same cost as before, and the run-time of subsequent outer loops can be spread out in time to best fit observation availability and computer resources. This further flexibility will be used for further optimisation in future IFS upgrades.

For further details, watch out for the article on ‘Continuous data assimilation for global numerical weather prediction’ by Lean et al., submitted to the Quarterly Journal of the Royal Meteorological Society.

Continuous long-window data assimilation setup. In the current setup (left) of the IFS data assimilation cycle, the 8-hour early-delivery analysis (DA) is only used to initialise the forecast (FCST), and it does not feed into the subsequent 12-hour long-window analysis (LWDA). The starting point for both analyses, the first guess (FG), is provided by the short-range forecast from the previous LWDA analysis. In IFS Cycle 47r1 (right), the DA analysis provides a more accurate first-guess trajectory (FG) as a starting point for the LWDA analysis, while the common background and its error estimate remain the same. This connects the two into one assimilation cycle, continuous long-window data assimilation (Co-LWDA).