RANDOM ERROR GROWTH IN NMC GLOBAL FORECASTS

The three-dimensional structure of random error growth in the National Meteorological Center's Medium-Range Forecast Model is investigated i n an effort to identify the sources of error growth. The random error growth is partitioned into two types: external error growth, which is due to model deficiencies, and internal error growth, which is the sel f-growth of errors in the initial conditions. Forecasts from winter 19 87, summer 1990, and winter 1992 are compared to assess seasonal varia tions in regional error growth as well as forecast model improvement. The following is found: In the tropics, large external error growth at the 200-mb level is closely associated with deep convection. There is evidence of significant model improvements in the tropics at the 850- mb level between 1987 and 1992. The spatial structure of the external error growth in the midlatitudes suggests that the representation of o rography in the model, especially over Antarctica and the Rockies, is a significant source of errors. Internal error growth in the midlatitu des is greater over the Atlantic and European regions than over the Pa cific region and appears to be associated with blocking phenomena, esp ecially over the North Atlantic and Europe. The Northern Hemisphere ex hibits a seasonal cycle in the magnitude of error growth, but the Sout hern Hemisphere does not. The results for the external and internal er ror growth rates were obtained using a parameterization of the correla tion between forecasts and the verifying analyses. The parameterizatio n is based on the assumption that linear random error growth is caused primarily by model deficiencies, and the validity of this assumption is examined. The results suggest that, in the tropics, significant inc reases in forecast skill may be obtainable through both model and anal ysis improvement. In the midlatitudes, however, there is less potentia l for increases in forecast skill through model improvement, and decre asing the analysis error becomes more important. The parameterization yields results that are physically meaningful and in agreement with pr evious predictability studies, and that provide quantitative estimates of the spatial and temporal distribution of the sources of forecast e rrors.