Dynamic initialization by repeated insertion of data

Abstract

In this study we isolate and assess the impact of dynamic initialization by repeated insertion of data (RI), also known as ‘nudging’. A novel control experiment compares this iterative technique with an alternative intermittent analysis-forecast cycle (AF) framework within the same assimilation system. Both schemes are versions of the operational global data assimilation system for numerical weather prediction at the Meteorological Office. A secondary issue is the impact of divergence damping as an extra initialization measure. The effectiveness of RI for control of high-frequency gravity waves is demonstrated, together with its beneficial impact on the short-period forecast of surface pressure. Changes to analysed fields due to RI are only a small fraction of the changes due to model evolution in six hours. RI is found to enhance upper-level divergence in the equatorial west Pacific and the problem of spin-up of the global mean precipitation is virtually absent, except when divergence damping is imposed during assimilation and removed during the forecast. The main benefit from divergence damping is a more accurate tropical wind field. RI is found to degrade the short-period forecast of the wind field in the basic system studied, and the extra errors are largest in strong upper winds at mid latitudes. This fact is discussed in the context of an earlier idealized study of RI. Analysis constraints are a primary source of difference between various ‘nudging’ schemes used in numerical weather prediction. The sensitivity of the impact from RI to these is evaluated by relaxing constraints in the basic system. The result emerges that as the analysis step of a nudging scheme becomes better balanced, the scope for improvement by using RI for initialization becomes smaller. Also, if the better-balanced analysis step entails stronger enforcement of non-divergence on the wind increments, then the damaging effect of RI on the wind field through advection becomes larger and can outweigh the benefit through initialization. A scheme combining RI for mass data with an iterative AF method for wind data retains much of the benefit and avoids most of the problems in the scheme with RI of all data. This suggests a method of improving the contribution of an RI structure to the overall performance of an assimilation scheme based on the nudging approach.