Multi-scale aspects of incremental 4D-Var at ECMWF
Recent developments in the ECMWF operational 4D-Var system have focused on extracting more information from observations in the analysis to increase the accuracy of the operational forecast initial conditions. This includes improved observation operators and observation coverage as well as improvements to the analysis system, which we will focus on here. A common theme is that the additional observation information is at smaller scales or more non-linear and to extract this information requires increased resolution to some aspects of the analysis system, which in turn requires optimizations to the system to offset the increased computational costs. For example, the wavelet B formulation allows geographically varying scale separation of the background errors, where errors of the day (from Ensemble of Data Assimilations, EDA) dominates the small scales and climatology is gradually mixed in at larger scales. To improve the covariances, especially at the smaller scales, the number of EDA members was increased from 25 to 50, while keeping computational cost and quality of the individual members the same through careful optimizations. Incremental 4D-Var also allows us to gradually add smaller scales at each minimization, and with more outer loops (from current 3 to 4 or 5) more information is extracted from observations with non-linear observation operators such as rainy radiances. To afford more outer loops, we have made recent developments to the operational data assimilation system (Continuous Data Assimilation) that allows us to do more work earlier and include more recent observations while still delivering the analysis at the same time in operational context.
Dr. Elias Valur Holm ECMWF