A series of seasonal (12-week) simulations are computed using the Weather Research and Forecasting (WRF) model, as applied to a roughly 2600 km x 2800 km nested domain covering the Amazon Basin and adjacent coastline. The simulations are computed using 4-km grid spacing, which is adequate to resolve at least the broader structures of most mesoscale precipitating systems. Simulations are carried out for both austral summer (DJF) and austral fall (MAM), as averaged over the four years spanning 2005-2008. Precipitation statistics are computed and evaluated against rainfall retrievals from the 13-year Tropical Rainfall Measuring Mission (TRMM), as well as compared to a series of 12-km companion simulations using the Betts-Miller convective parameterization.
As compared to the 12-km simulations, the convective-resolving runs show a significantly improved overall mean rain rate, with -15% error versus the TRMM retrievals for the 4 km runs, compared to roughly 80% error at 12 km. The rain-rate probability distribution function (PDF), the timing and diurnal evolution of precipitation, and the overall seasonal dependence are also significantly improved. Both the high- and low-resolution cases capture the distinct modes of convection along the coastline and in the basin interior. However, the 4-km case shows significantly better statistics for both modes.
While significantly improved compared to the 12-km simulations, the 4-km case still shows important errors relative to the TRMM retrievals. Comparison of the rain rate PDFs shows that the 4-km runs produce too few rain events overall (4.5% of events in the model, versus 7.5% in the retrieval) but produce a disproportionate number of events at high rain rates. The relatively good agreement found in overall mean rain rate is thus in part due to canceling errors. The model also produces errors in the diurnal evolution, with peak precipitation occurring several hours too early, and with the amplitude of the diurnal cycle damped relative to the TRMM data.
Dividing the simulation domain into two subdomains shows that the model performs better for the propagating mode along the coast, as compared to the local mode in the basin interior. The problems with the local mode are apparently tied to the behavior of the stratus layer following peak convection, which in turn affects the net radiative input the following day.