On the convergence of reanalysis produced by different data assimilation streams: a case study with NASA-GMAO MERRA-2 reanalysis system

Due to the enormous computational costs of producing reanalyses, it has become common practice to split the reanalysis production into multiple, parallel computational streams that usually include some year-long periods of overlap to account for the necessary spin-up integration period of each stream and foster continuity between successive streams. A fundamental assumption underlying the use of parallel NWP-DAS computational streams is that they will eventually converge to a common atmospheric state in a relatively short period of time. This important expectation of convergence has very limited investigation, and this study presents a more in depth analysis on this question.

 

Differences between reanalyses valid at the same times but produced by different streams were determined using overlapping periods from NASA’s GMAO MERRA-2 reanalysis. By the end of the overlapped periods, the monthly mean of the differences between the streams are small, as desired. However, the standard deviations of these differences were shown to be a large fraction (40 to 60%) of the estimated standard deviations of typical analysis errors. It was also shown that the streams do not appear to be converging as it was expected in terms of standard deviations of the differences, since little variation is seen between the beginning and the end of the spin-up period. This occurs although the overlapping streams use the same observations, numerical weather prediction and data assimilation models, and assimilation algorithm, differing only in the background information applied at the beginning of each overlapping period.