Initialization with the data assimilation method

Abstract

The initial conditions generated by a data assimilation technique with a general circulation model were evaluated. Two types of dynamic initialization, i.e., “the forward adjustment” and “the forward‐backward adjustment”, were tested using the real data produced by the NMC objective analysis. The degree of balance in the initial condition was examined in terms of the smoothness in the development of the rate of precipitation, angular momentum, and kinetic energy during the starting period of the prediction. The quality of the initialization was, however, appraised by the performance of the subsequent prediction. For the forward adjustment, various types of schemes of data insertion were studied with regard to the amount of shock produced and the degree of faithfulness of the injected data. In dealing with forward‐backward processes, one difficult problem is how to design a reversible algorithm. An approximate method is presented for maintaining the reversibility in a system which includes kinetic energy dissipation, radiation, and the moist heating. In addition, the question is discussed as to which physical processes of the model must be included for the forward‐backward initialization. It is found that the dynamic initialization using the forward‐backward data assimilation produces predictions that are, in some ways, better than those from the static or conventional initialization. However, for the forward data assimilation the predictions are consistently worse than those from the static initialization.

Outlook for Improved Numerical Weather Prediction Using Satellite Data with a Special Emphasis on the Hydrological Variables,

“…Because the moisture and vertical velocity fields both spin up in these methods, good initial precipitation rates can be obtained (Miyakoda et al, 1978;Fiorino and Warner, 1981).…”

“…The first group of dynamic methods involve forward and backward time integrations cycling around the current observational time with data inserted *repeatedly until convergence is obtained (Nitta and Hovermale, 1969;Miyakoda and Moyer, 1968;Hayden, 1973;Temperton, 1976;Morel et al, 1971;Miyakoda et al, 1978). Methods of this sort were analyzed by Talagrand (1981b).…”

“…Methods of this sort were analyzed by Talagrand (1981b). When the model contains irreversible processes, principally precipitation, the forward-backward integration is not well defined (Miyakoda et al, 1978).…”

“…The current four-dimensional data assimilation scheme at GFDL is a compromise between the two above dynamic methods (Miyakoda et al, 1976;Miyakoda et al, 1978;Stern and Ploshay, 1983). The current model state at the main synoptic times (0000 GMT or 1200 GMT) is used as first guess fields for 01…”

“…However, more recent experiments with the GFI)L model using conventional NMC initial conditions (Miyakoda et al, 1978;Sirutis et al, 1980) show that it takes 2.5 to 3 days of model integration for the tropical rainfall to appear ("spin up"). This experiment included convective precipitation; the …”