Forecasting skill limits of nested, limited-area models:: A perfect-model approach

The fundamental hypothesis underlying the use of limited-area models (LAMs) is their ability to generate meaningful small-scale features from low-resolution information, provided as initial conditions and at their lateral boundaries by a model or by objective analyses. This hypothesis has never been seriously challenged in spite of some reservations expressed by the scientific community. In order to study this hypothesis, a perfect-model approach is followed. A high-resolution large-domain LAM driven by global analyses is used to generate a "reference run.'' These fields are filtered afterward to remove small scales in order to mimic a low-resolution run. The same high-resolution LAM, but in a small-domain grid, is nested within these filtered fields and run for several days. Comparison of both runs over the same region allows for the estimation of the ability of the LAM to regenerate the removed small scales. Results show that the small-domain LAM recreates the right amount of small-scale variability but is incapable of reproducing it with the precision required by a root-mean-square (rms) measure of error. Some success is attained, however, during the first hours of integration. This suggests that LAMs are not very efficient in accurately downscaling information, even in a perfect-model context. On the other hand, when the initial conditions used in the small-domain LAM include the small-scale features that are still absent in the lateral boundary conditions, results improve dramatically. This suggests that lack of high-resolution information in the boundary conditions has a small impact on the performance. Results of this study also show that predictability timescales of different wavelengths exhibit a behavior similar to those of a global autonomous model.