Estimating ensemble size requirements of AGCM simulations

Abstract

SummaryThis study investigates the statistical methods for determining the minimum sample size necessary for an ensemble set generated with an atmospheric general circulation model. Due to the limits imposed by computational cost, an improved and a priori estimation of ensemble size is highly desirable. In this context, the methodology shown here is an important step for defining the number of integrations required in a numerical experiment. We show that the global distribution of ensemble size has a spatial and seasonal dependence. In addition, the ensemble size is dependent on the variable being analyzed and the geographic region of interest. For example, we show that a relatively large number of integrations are required to simulate the seasonal mean air temperature at 925 hPa and the sea-level pressure at mid to high latitudes. The seasonal mean precipitation, however, can be well represented with relatively few integrations at high latitudes, but it requires a large ensemble size at the tropics, particularly over the monsoon regions. These latitudinal differences in the number of integrations are associated with the internal variability in the model. Furthermore, differences among the variable fields partly arise due to the distinct shapes of the associated property distributions. Both Gaussian and nonparametric statistics are considered here. The Wilcoxon Rank Sum test reveals that the air temperature, sea-level pressure and precipitation do not follow a Gaussian distribution in some regions of the globe. Thus, we suggest the nonparametric approach be used whenever the normal assumption is violated or cannot be assessed. This study is based on the determination of the sampling size of ensemble simulations using the statistical Gaussian method of Wehner (2000). We extend this previous by considering a nonparametric approach.

The impact of the subtropical South Atlantic SST on South American precipitation

“…Previous studies such as Chang et al (2000), Saravanan and Chang (2000) and Barreiro et al (2002) have had good results for precipitation and atmospheric circulation with a relatively small ensemble size performed with the NCAR CCM3 model forced with SSTs in the tropical Atlantic and South Atlantic Ocean, involving similar decades as used in this work. In addition, this ensemble size is in agreement with the estimate of Taschetto and England (2008).…”

“…Apparently, the rainfall reproducibility in DJF over most of Brazil is lower than 0.3 (Fig. 3a), indicating that the internal variance accounts for approximately 70% of the total variance (Taschetto and Wainer, 2008). This reveals that the SACZ region over South America is dominated by internal variance, which explains the difficulty in predicting this phenomenon.…”

“…According to Shukla (1993) the tropical atmospheric patterns are strongly determined by the underlying SSTs, especially when a large-scale seasonal simulation is run for a long period. Shukla (1998) showed that two integrations with different initial atmospheric conditions respond very similarly to the tropical Pacific SSTs.…”

“…This first evaluation is very important to verify the model performance, before speculating on the mechanisms by which the SST anomalies impact the climate. Only the main results of the statistical treatment are shown here, a detailed discussion can be found in Taschetto and Wainer (2008).…”

Towards predictive understanding of regional climate change

“…The following factors increase demands on computational resources: First, internal variability has a very strong imprint on climate trends even on time scales as long as several decades and spatial scales as large as continents 81 . This calls for large ensemble simulations 96 .…”