Projections from 15 global climate models and 2 reanalysis products (National Center for Environmental Prediction (NCEP)/National Center for Atmospheric and Climate Research (NCAR) reanalysis (NNR) and European Centre for Medium-Range Weather Forecasts (ECMWF) 40-year reanalysis (ERA40)) were utilized to project changes in the net precipitation (P − E) over the Southern Ocean and Antarctica during the 21st century. Three time periods, 1979–2000, 2046–2055, and 2091–2100, of data were compared. The P − E was related to a classification of synoptic circulation patterns obtained using a neural network algorithm known as self-organizing maps (SOMs). SOM classification was successfully used as a quality control tool to assess the simulated atmospheric circulation and model performance in P − E. The models predicted an increase of Antarctic P − E that averages 0.42 ± 0.01 mm year−1 for the coming hundred years based on the difference between 1979–2000 and 2091–2100. P − E changes of individual models ranged from 0.02 to 0.71 mm year−1. P − E integrated over the entire Antarctic ice sheet was forecast to increase more quickly from the end of the twentieth century until 2046–2055 than from 2046–2055 until 2091–2100. Contributions to the predicted change in P − E were evaluated for both thermodynamic and dynamic processes. The projected change in Antarctic P − E was primarily due to thermodynamic changes rather than circulation changes. The dynamic component of P − E change, associated with the circulation, was important at subcontinental scales, especially over the coastal regions. The role of dynamic changes was maintained until the end of the 21st century. Intermodel variation in predicted P − E changes and differences between models and reanalyses in the twentieth-century simulations severely restrict the reliability of these projections and highlight the need for improved polar simulations in climate models.
 The net precipitation (P − E) over Antarctica is an important factor in the mass balance of the continental ice sheet through snow accumulation and ablation and thereby has a major impact on global sea level variations. Surface sublimation and blowing-snow processes, which also help to determine the local mass balance of the ice sheet, are estimated to have limited contributions to the statistics of the surface mass balance at scales larger than 100 km [Frezzotti et al., 2004; Genthon, 2004]. Recent climate model projections under increasing greenhouse gas concentrations predict a warmer atmosphere containing more water vapor. At the Earth's surface, this suggests increased precipitation (P) and evaporation (E), but projections of regional changes, and especially changes in P − E, leave significant uncertainties. When the atmosphere becomes more humid, it transports moisture more efficiently toward polar regions, which may have thermal impacts that cause increased melt and breakup of ice sheets, with substantial changes on timescales of a century [Hansen, 2005]. On the other hand, increased accumulation on the Antarctic ice sheet because of a net increase in precipitation is suggested by modeling studies with increasing carbon dioxide concentration [Wild et al., 2003; Huybrechts et al., 2004; Gregory and Huybrechts, 2006]. However, Hansen  argues that model computations do not realistically incorporate processes that are implicated in the potential for accelerating ice sheet disintegration. Recent observations confirm the importance of these ice discharge processes [Alley et al., 2005; Velicogna and Wahr, 2006].
 When moist air flows poleward, first over the coastal region and then across the Antarctic continent, it loses kinetic energy with increasing altitude and water vapor because of condensation. How far into the Antarctic continent an air parcel can flow depends on its speed and its height while the air parcel is still over the ocean. Moisture transport can be decomposed into mean and eddy components by applying Reynolds averaging [e.g., Bromwich et al., 1995]. The mean component corresponds to the stationary moisture transport, while the eddy component is the temporally varying component and includes moisture transported by synoptic activity. The mean and eddy components are roughly proportional over the coastal regions, but the eddy component dominates the inland precipitation. The eddy component exhibits a close relation with elevation and is a positive contribution to transporting moisture from the ocean toward the pole [Cullather et al., 1998]. Because of the differences in the characteristics of the two moisture convergence components, two distinctly different precipitation regimes can be expected over Antarctica.
 Several earlier studies have attempted to quantify the current atmospheric water balance in the polar regions [e.g., Bromwich et al., 1995; Cullather et al., 1998; Zou et al., 2004] using observational (radiosondes- and satellite-derived) and model analysis/reanalysis data. These studies advocated the practice of estimating P − E from the horizontal moisture convergence fields over polar regions, where it is otherwise difficult to measure because of sparse and sometimes poor-quality data.
 The Intergovernmental Panel on Climate Change (IPCC) is preparing the fourth assessment report (AR4) of the scientific, technical and socioeconomic understanding of anthropogenic climate change and its consequences. One key element in the report is the analysis of a set of projections of Global Climate System Models (GCSMs). This study is a contribution to that effort, focusing on a description of changes in the circulation and moisture content in the Antarctic atmosphere as simulated by a set of GCSM projections.
 In an earlier study, Raphael and Holland  found that circulation modes of a set of IPCC models were not able to simulate the middle- to high-latitude temperature gradient, resulting in overly weak poleward heat and moisture fluxes. Miller et al.  studied the response of the Southern Hemisphere annular modes, Antarctic Oscillation (AAO) or Southern Annular Mode (SAM), and El Niño Southern Oscillation (ENSO), simulated by a set of IPCC models to increasing concentrations of greenhouse gases and aerosols and depleted ozone. GCSM simulations have shown that the AAO circulation index is enhanced because of increasing greenhouse gas concentrations and ozone depletion [see, e.g., Shindell and Schmidt, 2004]. Miller et al.  suggested that the models were underestimating the feedback coupling stratospheric changes to surface annular variations. On the basis of these earlier studies, one would expect that the IPCC models will simulate changes in temperature, and probably in moisture too, in the correct direction, but with smaller magnitudes than in reality. P − E is also likely to be affected by the changes in AAO and ENSO [Bromwich et al., 2000; Genthon et al., 2003].
 Synoptic scale analysis of IPCC model output by Lynch et al.  found increased cyclone activity and stronger zonal winds during the 21st century in the Antarctic. The coherence of precipitation and temperature patterns with the sea-level pressure (SLP) climatology suggests that changes in these fields are, at least to a certain extent, related to changes in circulation [Lynch et al., 2006]. These findings support the importance of the moisture processes at synoptic scales in the total P − E signal suggested by Noone and Simmonds  and van Lipzig and van den Broeke .
Krinner et al.  simulated Antarctic surface mass balance with a high-resolution (60 km) atmospheric model for two time periods, 1981–2000 and 2081–2100. The boundary conditions (sea surface temperature and sea ice concentration) were taken from the IPCC GCSM model projections. The simulation revealed increasing P − E over the ice sheet during the 21st century. The P − E change was largely due to increased precipitation because of the higher moisture holding capacity of warmer air, while changes in snowmelt and surface turbulent latent heat flux were small. The changed atmospheric dynamics regionally modulated the thermodynamic signal, especially off the coast of Antarctica.
 A recent study of observational data by Turner et al.  reveals significant warming of the Antarctic winter troposphere, with no evident changes in the atmospheric circulation. This suggests that in situ effects, like changes in cloud amounts, greenhouse gas concentration, and other local thermodynamic processes, may play an important role.
 In order to study the synoptic-scale variability of the net precipitation, daily-model-generated sea-level pressure (SLP) was analyzed using a classification technique based on a neural-network-algorithm-denoted self-organizing maps (SOM) [Kohonen, 2001]. This method has proven to be a robust one, allowing successful identification of outliers and the extraction of useful information from multimodel data sets having large mutual scatter [e.g., Lynch et al., 2006; Cassano et al., 2006a, 2006b; Hewitson and Crane, 2006; Hope et al., 2006].
2. Data and Methods
2.1. Net Precipitation
 The output of 15 IPCC models (Table 1) was obtained from a repository maintained by the Program for Climate Model Diagnosis and Intercomparison (PCMDI) at the Lawrence Livermore National Laboratory, USA. To aid data analysis, the collection procedure of IPCC model output defines the data format to have uniform grid specifications and variable names. Another part of this common specification requires output to be interpolated to nine constant pressure levels. The horizontal resolution, on the other hand, varies between the models.
Table 1. Models Used in This Study
National Center for Atmospheric Research
Canadian Centre for Climate Modelling and Analysis
Météo-France/Centre National de Recherches Météorologiques
CSIRO Atmospheric Research
University of Bonn and KMA
Max Planck Institute for Meteorology
LASG/Institute of Atmospheric Physics
US Department of Commerce/NOAA/Geophysical Fluid Dynamics Laboratory
US Department of Commerce/NOAA/Geophysical Fluid Dynamics Laboratory
NASA/Goddard Institute for Space Studies
NASA/Goddard Institute for Space Studies
Institut Pierre Simon Laplace
Center for Climate System Research (The University of Tokyo), National Institute for Environmental Studies, and Frontier Research Center for Global Change (JAMSTEC)
Center for Climate System Research (The University of Tokyo), National Institute for Environmental Studies, and Frontier Research Center for Global Change (JAMSTEC)
Meteorological Research Institute
 Daily fields of precipitation (P) and surface latent heat flux (λE) were obtained covering three time slices: 1979–2000, 2046–2055, and 2091–2100. The first time slice was from the twentieth-century experiment (20C3M), while the latter two were from the 720-ppm CO2 stabilization experiment (SRES A1B) [Nakicenovic and Swart, 2000]. SRES A1B, the “balanced energy sources” scenario, was chosen from the group of scenarios because it produces medium-level carbon emissions. According to SRES A1B, carbon emissions will increase until around 2050 and then start to decrease. P − E can be estimated from the IPCC model output by subtracting E from P. E is computed from the surface latent heat flux λE, where λ is the latent heat of vaporization if Ts exceeds the freezing point and the latent heat of sublimation otherwise [Deas and Lowney, 2000]. Additionally, P − E was computed from the moisture transport convergence (as was done for example by Bromwich et al., 1995). However, because of the coarse vertical resolution (a result of the vertical interpolation of the model output, see previous paragraph) and the daily averaging, P − E calculated from the moisture transport convergence resulted in values that were approximately 60%–80% lower than the P − E values calculated from the model P and λE. The low vertical resolution suggests that the convergence method was less reliable, confirmed by the fact that the P − E calculated from P and λE was closer to the estimates of earlier studies (see section 3.2). Hence the latter technique was used for this analysis.
 After the computation of P − E on the model grid, the fields of P − E were interpolated to a Lambert equal-area scalable earth (EASE) grid projection (developed at the National Snow and Ice Data Center in Boulder, CO; see http://nsidc.org/data/ease/) of 42 × 42 points and 200-km grid spacing. This grid is centered on the pole, with an analysis domain extending from the pole to 51°S latitude at 0°, 90°E, 180°, and 90°W longitude.
 The mean sea-level pressure fields from the European Centre for Medium-Range Weather Forecasts (ECMWF) 40-year reanalysis (ERA40) [Uppala et al., 2005] and from the National Center for Environmental Prediction (NCEP)/National Center for Atmospheric and Climate Research (NCAR) reanalysis (NNR) [Kalnay et al., 1996] for 1991–2000 were interpolated to EASE grid as well. These data sets were used to test the distribution of circulation patterns in the IPCC models for the 1991–2000 period.
2.2. Synoptic Classification Algorithm
 The SOM algorithm utilizes an unsupervised learning process to produce a low-dimensional representation of the training data samples while preserving the topological properties of the input space. As a result, multidimensional, large data sets are reduced to more easily interpreted forms seen as arrays of nodes or maps. Hence the SOM technique is especially suitable for visualizing high-dimensional data. Unlike other similar clustering techniques, the SOM technique does not need a priori decisions on data structure but is trained when processing the data itself. The SOM algorithm places more nodes in the areas of high-input-data density and thus attempts to preserve the probability density function of the input data. Kohonen  published a detailed description of the SOM algorithm, and a paper by Hewitson and Crane  provides more information on the application of SOMs to climate data. Several other studies have also applied SOMs for atmospheric studies [see, e.g., Cavazos, 1999; Malmgren and Winter, 1999; Cavazos, 2000; Ambroise et al., 2000; Crane and Hewitson, 2003; Reusch et al., 2005; Cassano et al., 2006a, 2006b; Crimmins, 2006; Hope et al., 2006; Hope, 2006].
 In earlier studies, we have found that the SOM technique is quite suitable for the creation of synoptic climatologies of the polar regions [Cassano et al., 2006a, 2006b; Lynch et al., 2006]. In these studies SOMs consist of a two-dimensional array of nodes, where each node represents a circulation pattern. In this analysis, the SOM was configured to produce a 7 × 5 array of circulation patterns (Figure 1), a size that has been found suitable for synoptic climatologies since it compactly displays the major circulation patterns and storm tracks [Hewitson and Crane., 2002; Cassano et al., 2006b; Lynch et al., 2006]. In Figure 1, the SOM nodes are presented in a regular array, which do not correspond to the quantitative measure of similarity between the nodes. The Sammon mapping scheme [Sammon, 1969] is one approach that computes distances between the nodes and shows how they are related in the two-dimensional space and is discussed in more detail in section 2.3.
 For this study, anomalies from the domain-averaged daily SLP for the time periods 1991–2000, 2046–2055, and 2091–2100 were retrieved from the output of 15 IPCC models (Table 1). Mean daily SLP values over the common grid domain were calculated from daily SLP fields, and anomaly SLP values were computed at each grid point as the domain mean SLP was subtracted from the grid point SLP. The SOM based on SLP anomalies was found to better represent the different circulation regimes than an SOM trained using daily SLP data. Training an SOM with SLP anomalies also has an advantage that it can represent horizontal pressure gradients, which drive atmospheric circulation, while neglecting the magnitude of the SLP. This is a useful property when analyzing data including all seasons because then the SOM classification is affected more by the intensity of cyclones than by the seasonal cycle of pressure. As in Lynch et al. , SLP anomaly values over the Antarctic continent were considered unreliable and masked from the SOM input data. Using the SOM algorithm, a first-guess map was created and the map was then refined by utilizing competitive learning of the training process. For a detailed description of our analysis procedure, see Cassano et al. [2006b].
 After the SOM map has been created, selected records of daily anomaly SLP are classified to the SOM and the frequency distribution (how many records are classified to each node) is determined. Furthermore, two distributions can be subtracted and frequency changes analyzed. A relevant question then is how to determine the significance of frequency changes that are detected.
 One approach is to compare distributions derived from data with the random process distributions. If an observed distribution differs markedly (say at the 95% confidence level) from the random process, it is probably statistically significant. Here we have n data records and N SOM nodes. A data record is either mapped to a node with the probability p = 1 / N or not (repeated n times), so we can assume the process to be binomial. The expected number of mapped fields per node is n / N and the expected frequency per node is 1 / N. The variance of the binomial distribution is np(1 − p). The binomial distribution approaches the normal distribution rapidly when n increases, and a reasonable approximation for the 95% confidence limits of the number of records per node is np ± 1.96 . Similarly, 95% confidence limits for the node frequency would be p ± 1.96. If node frequency exceeds this limit, we deem it to be statistically significant.
 Having mapped two data sets with number of records n1 and n2 to an SOM, r1 = n1p1 records were mapped to a node from the first data set and r2 = n2p2 to the same node from the second data set. It is assumed that these data sets result from two random, independent, binomial processes with estimators of node frequency variances p1(1 − p1) / n1 and p2(1 − p2) / n2, respectively. The hypothesis to be tested is that the difference of node frequencies is zero, that is, we expect no frequency change. Now the random variable (p2 − p1)/, which is at least approximately normal, can be used as a test statistic. If the value of the test statistics exceeds 1.96, we reject the null hypothesis and deem the frequency change to be statistically significant.
2.3. Climatology of Synoptic Patterns
 The resulting Antarctic climatology constructed using the SOM technique is presented in Figure 1. It retains information on high-frequency processes, which are not apparent in monthly or seasonal mean fields [Barry and Perry, 2001]. All major circulation patterns are presented in the climatology. Low-pressure and high-pressure patterns are clearly separated, and patterns that are similar to each other are clustered together on the map. A set of circumpolar troughs populate mainly the three rightmost columns on the map. High-pressure patterns are located in the left part of the SOM, and a variety of weak circulation patterns fall between cyclonic and high-pressure nodes.
 As noted above, the nodes of the SOM do not have equal distances to the neighboring nodes in the data space because the SOM algorithm approximates the probability density function of input data and locates more nodes where data are dense. Sammon mapping [Sammon, 1969] was applied to compute a distortion surface showing Euclidean distances between the nodes, and it can be used to identify groups of similar nodes. Sammon mapping approximates local geometric relations of SOM nodes in two dimensions. The curvature in locations of SOM nodes (Figure 2) is caused by the shape of the probability density function of the input SLP data set.
 In Figure 2, Sammon mapping of the SLP anomaly SOM shows, for example, that the nodes in row 4 are more distant from each other than the nodes in row 0. This reflects expectations because row 4 contains synoptic patterns ranging from high-pressure to strong cyclones. Row 0, on the other hand, contains varying types of low-pressure patterns only. The SOM has so much curvature that the upper right corner looks unordered (Figure 2). For instance, node (5,2) is closer to node (4,0) than to node (5,1). To emphasize the unordered locations of nodes in the upper right corner, part of the Sammon map is shaded in dark gray.
3. Twentieth Century: Models and Reanalyses
 Any time slice of SLP anomaly fields can be projected to the SOM, and frequencies of occurrence per SOM circulation node can be determined. A frequency distribution based on the data of all 15 models for 1991–2000 is presented in Figure 3. If the data were uniformly projected to the nodes, every node frequency would be 1 / 35 = 2.86%. Because the SOM was created using this data set, the future time slices node frequencies are relatively close to each other.
 In Figure 3, nodes in the leftmost column, lower left, and upper right corners have higher frequencies, while nodes elsewhere have frequencies less than 2.86%. This difference is partly due to the fact that all time slices were used to create the SOM, and the nodes over-represented (under-represented) in Figure 3 are nodes which are less (more) frequent during the later two decades. On the other hand, the SOM algorithm does not try to create N nodes that all occur equally frequently. Instead, the algorithm seeks to define nodes that span the range of the data space. It is possible that a very infrequent pattern is also very different from all of the other patterns in a data set. This pattern would be represented by one (or more) node in the final SOM, but would have a low frequency of occurrence.
 Frequency distributions of the SOM nodes can be used to assess the ability of the individual models to simulate present-day Antarctic circulation by comparing model node frequency distributions to those calculated from two reanalysis products, ERA40 and NNR, for 1991–2000 (Figure 4). The two distributions from the reanalyses look rather similar to each other, signifying similarities in circulation patterns of both reanalyses, and rather different from the multimodel ensemble represented in Figure 3.
 Circulation types located in the two uppermost rows and in the lowest row along columns 3, 4, and 5 in the SOM occur more frequently in the reanalysis data than in the 15-model ensemble (Figure 3). These nodes represent different versions of the circumpolar trough and some three-wave patterns. On the other hand, the model ensemble includes more days of both high pressure (such as nodes (0,2), (0,3), (0,4), and (1,4)) and some manifestations of a circumpolar trough (node (6,1)).
 One striking difference is the frequency of node (0,4), which for the model ensemble has the highest frequency of 5.4% (Figure 3), while it appears in the reanalysis data relatively seldom (1.1% to 1.4% of the time). This anomalous node frequency is predominantly due to IPSL-CM4, GISS-AOM, and CNRM-CM3 models. An anomalous node frequency due to the CNRM-CM3 model in an SLP-based SOM was also identified by Lynch et al. , who focused on two seasons (DJF and JJA), while in this study data from every month were used.
 Frequency distributions of individual models were compared with the reanalysis distributions in order to assess the models' ability to simulate the present-day Antarctic circulation. A cumulative sum of reanalysis and model frequencies is plotted in Figure 5. The reanalysis lines (bold solid) are similar, with a high correlation (R2) of 0.998 (see Table 2). This correlation is a measure of the similarity of frequency distributions and furthermore the similarity of the underlying synoptic-scale circulations.
Table 2. Correlation Squared (R2) of Cumulative Sums of Frequencies of Occurrence of Circulation Types for 1991–2000 Between the Named Model and the ERA40 Reanalysis
 Models deviating most from the reanalyses have higher frequencies in the high-pressure part of the SOM (columns 0 and 1) than the reanalyses. This is apparent because their cumulative frequencies increase faster up to node number 14 (see dotted and dashed lines in Figure 5). The frequencies of occurrence of node numbers 16 to 35 increase very little compared with the reanalyses ones, indicating less frequent occurrence of cyclonic circulation patterns. This suggests that these models may have a systematic bias toward low cyclonicity.
 In Table 2, the 15 models have been ranked according to R2 with the ERA40. Almost all models correlate relatively well with the reanalysis, GFDL-CM2.1 correlating even better with ERA40 than NNR and IPSL-CM4 model has the lowest R2. Eight models have R2 better than 0.95 and three models have R2 better than between 0.90 and 0.95. On the basis of the ranking in Table 2 and the consistency of the modeled mean P − E with estimates from the literature (see next section), we selected five models to be used in evaluating changes in P − E over Antarctica during the 21st century.
3.2. Net Precipitation
 Mean values of twentieth-century Antarctic P − E have been published in several studies based on different methods and data. In a synthesis of studies until that time, Ohmura et al.  listed observationally derived mean P − E ranges from 147 to 184 mm year−1. According to Bromwich , P − E based on multiyear glaciological data suggest rates of 151–156 mm year−1, and a later analysis using the ECMWF (European Centre for Medium-Range Weather Forecasts) operational analysis resulted in similar values [Bromwich et al., 1995]. In later studies, Cullather et al.  obtained an average P − E of 151 mm year−1 using a variety of observational data sets and the ECMWF operational analysis, while Vaughan et al.  based their study on surface observations, resulting in a value of 166 mm year−1. This last value was replicated in a modeling study by Wild et al. . Bromwich et al.  obtained results based on regional climate model (Polar MM5) experiments of 181 ± 16 mm year−1. In a regional climate modeling study, van de Berg et al.  obtained a mean surface mass balance of 178 mm year−1 integrated over the Antarctica for 1980–2002. van de Berg et al.  simulated the Antarctic surface mass balance using the same model as van de Berg et al. , but now calibrated with observations. They estimated mean rates of 171 mm year−1 over the grounded ice sheet and of 186 mm year−1 for all Antarctic ice sheets in 1980–2004. Using a high-resolution atmospheric model, Krinner et al.  estimated the average surface mass balance for 1981–2100 to be 151 mm year−1. Monaghan et al.  combined model simulations with observations primarily from ice cores and estimated a 50-year mean (1955–2005) P − E of 182 mm year−1 over the grounded ice sheet. As is evident from the range of values quoted here, mass balance estimates remain dependent on both data source and analytical technique. Furthermore, all of the estimates quoted span different time periods according to data availability. The estimates above, however, seem to agree that the long-term mean value is in the range of 150–190 mm year−1.
 The mean values of P − E over Antarctica for individual models, two averages of subsets of the models, and the two reanalysis products are presented in Table 3. Values are based on the points of the common intercomparison grid located higher than 1 m above the sea level (Figure 6). Values were found to be rather insensitive to threshold levels of altitude up to 100 m. The common topography data set applied in these calculations was sampled from the TerrainBase database of the National Geophysical Data Center, USA. The models in Table 3 are ordered according to the ranking in Table 2. Mean P − E of the models for 1979–2000 varies significantly from 123 up to 269 mm year−1. The highest, and most unlikely, value of P − E is obtained from the MRI-CGCM3.2.2 simulation, despite relatively good agreement between this model and the reanalysis near-surface circulation. This model has the highest mean precipitation rate P over Antarctica of all 15 models and relatively small E, resulting in high P − E. CGCM3.1(T63) and MRI-CGCM3.2.2 are the only models applying flux corrections of heat and water [Dai, 2006], which may have biased the fluxes of MRI-CGCM3.2.2 over the Antarctic ice sheet. Hines et al.  found that even a relatively high-resolution regional climate model (MM4) may inadequately represent the moist physics over Antarctica, although over the surrounding ocean the model results can be quite realistic. This finding emphasized the need for the development of model physical parameterizations adapted for high southern latitudes. Average P − E over the Southern Hemisphere region, excluding Antarctica, and global average P − E of MRI-CGCM3.2.2 are relatively close to the corresponding averages of the other models.
Table 3. Mean Net Precipitation (P − E) and its 95% Confidence Limit (mm/year) Integrated Over the Antarctic for 1979–2000 in the Third Column; Fourth, Fifth, and Sixth Columns Show the Difference of Mean P − E Between Five Time Periods and its 95% Confidence
 For the further study of projected P − E trends in the 21st century, five models were chosen, first, according to their correlation with the ERA40 circulation and, second, based on their ability to simulate a mean P − E for the period 1979–2000 that is consistent with the range of observations cited above. The five models that were selected in this way are: GFDL-CM2.1, MIROC3.2(hires), CGCM3.1(T63), CCSM3, and ECHAM5/MPI-OM. R2 of these models with ERA40 sea-level pressure is better than 0.98 at the 97.5% confidence level, and their 1979–2000 mean moisture mass balance is 171 mm year−1, which is consistent with estimates from the literature.
Bromwich et al.  note that the reanalysis products, and therefore probably GCMs in general, underestimate precipitation, especially over high elevations. This is due in part to the coarse spatial resolution of most global models. It is reasonable to assume that an overly smooth coastal escarpment in a coarse resolution model is causing low-pressure systems to precipitate less than they should and it also affects the location of high precipitation [Bromwich et al., 1995]. For example, if the Antarctic Peninsula is not well resolved in a model, it produces too little lee cyclogenesis [Turner et al., 1998]. In this study, the GCSMs, on average, have an even coarser resolution (typically 2–3°) than the reanalysis. The model with the highest resolution is MIROC3.2(hires), which also has the highest projected P − E changes (Table 3).
4. Twenty-first Century Trends in Net Precipitation
 Many earlier studies based on observations or models have reported increasing trends in the Antarctic P − E in the twentieth century [e.g., Cullather et al., 1998; Bromwich et al., 2004]. Conversely, in recent papers by van de Berg et al. , Monaghan et al. , and van den Broeke et al. , no significant change in Antarctic accumulation during the last decades of the twentieth century has been found. Nevertheless, significant increases have been reported in simulations with increasing greenhouse gas concentrations: Wild et al.  reported that, in a simulation where CO2 concentration in the atmosphere doubles, the annual net accumulation over Antarctica increases by 22 mm year−1 compared with the present climate. Huybrechts et al.  compared results from ice sheet model driven by a climate scenario where CO2 concentration doubles in 60 years. They found an associated increase of 15%–20% in mean Antarctic P.
 Decadal P − E mean values of individual models were studied in order to identify if they reveal any systematic changes across the suite of models. Ten models simulated significant changes (at the 95% confidence level) between the P − E mean values of periods 1979–1989 and 1990–2000. Seven models, CSIRO-Mk3.0, MIROC3.2(hires), CGCM3.1(T63), MRI-CGCM3.2.2, FGOALS-g1.0, GFDL-CM2.0, and GISS-AOM simulated significantly increased mean P − E in 1990–2000 from the mean of 1979–1989 (Table 3, 4th column). On the contrary, mean P − E of GFDL-CM2.1, CCSM3, and MIROC3.2(medres) was significantly smaller during the last decade of the twentieth century than the decade earlier. Additionally, five models did not reveal any significant change in mean P − E. In summary, the models do not display any uniform or systematic change in decadal mean P − E over the Antarctic ice sheet in the end of the twentieth century.
 All but 2 of the 15 models in this study project an increase of P − E over Antarctica (Table 3) for both 21st-century time slices. The increasing P − E reflects warmer air temperatures and associated higher atmospheric moisture and hence an increase of P. Evaporation increases as well because of warmer air, but less than precipitation. Two models, ECHO-G and GFDL-2.1, project a small decrease of P − E after the middle of the 21st century. The P − E trend is defined to be the P − E difference (having unit of mm year−1) between two time slices divided by the minimum number of years between the time slices resulting in a unit of mm year−2.
 On average, the P − E is projected to increase 0.29 mm year−2 by 2100 (26.5 mm year−1/90 year), which is comparable to the results obtained by Wild et al.  in the CO2 doubling experiment and Huybrechts et al. , where the P − E increase was larger because of the higher CO2 increase rate. Model projections chosen for this study are based on SRES A1B scenario, where the atmospheric CO2 concentration approximately doubles by 2100.
 The annual 26.5-mm increase (0.29 mm year−2) over 90 years is an order of magnitude smaller than the significant upward trend of 2.4 mm year−2 from ECMWF operational analyses for the time period of 1985–95 found by Cullather et al.  and 1/5 the size of the P trend of 1.3–1.7 mm year−2 for 1979–99 found by Bromwich et al. . Precipitation trends from Bromwich et al.  however, have large error bounds; for example, the P trend based on ERA15 reanalysis is 1.35 ± 1.10 mm year−2 with 95% uncertainty limits. Additionally, as mentioned earlier, van de Berg et al. , Monaghan et al. , and van den Broeke et al.  suggest that there has been no trend in Antarctic P − E since the early 1980s and the models studied here did not reveal any systematic change in P − E during the late twentieth century (see above). Krinner et al.  simulated Antarctic P − E for two time periods of 1981–2000 and 2081–2100 obtaining an increase of 32 mm year−1, which results in an upward trend of 0.4 mm year−2 over 80 years, which matches the upward trend of the five-model ensemble ((21.4 + 16.3) mm year−1/90 year = 0.42 mm year−2, Table 3). Their atmospheric model had 60-km resolution over Antarctica, which is significantly higher than GCSMs analyzed here and is expected to simulate higher precipitation rates than coarser resolution models.
 The 21st-century trends in P − E have significant variability across the models, with the maximum upward 45-year trend being 0.71 mm year−2 (32.1 mm year−1/45 year, see Table 3, MIROC3.2(hires)) and the maximum downward trend being 0.13 mm year−2 (−5.7 mm year−1/45 year, GFDL-CM2.1). Additionally, six models project stronger P − E increase before the middle of the 21st century, while five models project stronger increases after the middle of the century. The 15-model ensemble shows a slightly weaker P − E trend after the middle of the century compared with the one for the coming 50 years, but given the high variation between models, the confidence of this conclusion is limited.
4.1. Trends Associated With Circulation
 In section 3.2, a set of five models were chosen that simulate present-day circulation realistically and provide reasonable estimates for mean P − E to study the projected changes in the Antarctic circulation and associated P − E. As the frequencies of occurrence of each circulation type for all 15 models in Figure 3 for 1991–2000 were computed, corresponding frequencies for the set of five models and for the three time slices 1991–2000, 2046–2055, and 2091–2100 can also be calculated. Changes of the frequencies between the three time slices (Figure 7) signify changes in circulation over time. The frequency distribution of the selected five models for 1991–2000 (Figure 7a) resembles frequency distributions of the reanalyses (Figure 4) and therefore emphasizes the skill of the selected models to simulate present-day circulation.
 Before the middle of the 21st century, magnitudes of the relative changes are rather substantial from a decrease of −1.0% (node (3,4)) to an increase of 2.5% (node (5,4)) (Figure 7b). Most of the nodes with decreasing frequency of occurrence in the left part of the SOM are associated with anticyclonic and weak circulation patterns; node frequencies of these patterns continue to decrease at a statistically significant rate after the middle of the century. The significance of the frequency change was defined using the formula given in the last paragraph of section 2.2.
 Patterns with increasing frequency are located either in the upper middle part of the SOM or in the lower right corner and are associated with a strong circumpolar trough (cf. Figure 1). When a node frequency increases by 1%, this means that the circulation pattern associated with the node will occur almost 4 days more often (0.01 × 365 days) in a year. Because node frequencies are typically about 3%, this is a very significant relative change. The occurrence of the pattern will increase from 11 to 15 days per year. The change in the circulation patterns can be further highlighted by adding changes of increasing patterns together. By summing the frequencies of the eight significantly increasing patterns from Figure 7b, the frequency change is 8.5%, indicating that the associated strong cyclonic patterns will occur 1 month longer in a year during time period 2046–2055 than in 1991–2000. Similar significant changes occur between time period 2046–2055 and 2091–2100 (Figure 7c).
 Similar analysis was carried out here, and positive and negative phases of the AAO index and Multivariate ENSO Index (MEI) were mapped to the SOM using ERA40 SLP for 1991–2000 (Figure 8). The nodes where positive phases of AAO (and negative phases of MEI) are more frequent are mainly associated with the cyclonic circulation patterns. This relationship is more evident for AAO than for MEI. As noted earlier, these circulation patterns are predicted to become more frequent during the 21st century (Figures 7b and 7c) signifying the increase of positive AAO index circulations. This result is in agreement with the findings of Miller et al. , who found a positive annular trend from a set of GCSM projections in the 21st century. This trend was a response to increasing concentrations of greenhouse gases and tropospheric sulfate aerosols.
4.1.2. Net Precipitation
 Using information of daily SLP anomaly field classification to each circulation node, daily P − E fields were associated with the nodes of the SOM and the average daily P − E of each node was computed. The anomalies of these P − E patterns were calculated as the difference between the average P − E for all of the samples that map to a particular node and the average P − E for all of samples. In Figure 9, ratios between P − E anomalies of the five-model subset for the 35 circulation patterns identified in the SOM analysis and the average P − E for all the samples are illustrated. Representation of ratios rather than P − E anomalies provides more details over Antarctica.
 From Figure 9, it can be seen that nodes associated with circulation patterns that are projected to become less frequent during the 21st century are associated with negative P − E anomalies over the Southern Ocean (lower left part of the figure). These negative anomalies extend around the Antarctic continent. The circulation patterns that are forecast to have an increasing frequency of occurrence during the 21st century (located either in the lower right or in the upper middle part of the figure) have P − E patterns, with sectors of negative and positive anomalies extending toward the Antarctic continent from the Southern Ocean.
 The average daily P − E over the Antarctic ice sheet for each node in the SOM for the period 1991–2000 is shown in Figure 10a. The largest average daily P − E values are found in nodes (2,0), (3,0), (5,0), and (4,4) of the SOM, which correspond to three-wave circulation patterns with strong low-pressure systems around Antarctica. Correspondingly, the high-pressure nodes in the lower left part of the SOM and some nodes in the middle right part ((5,2) and (5,3)), representing circumpolar troughs, have relatively low average P − E. From this analysis, it appears that average daily P − E over the Antarctic ice sheet is dependent on the SOM-derived circulation patterns over the surrounding ocean according to the IPCC model output.
 The average annual P − E over the ice sheet can be calculated from Figures 7a and 10a as:
where fi is the frequency of occurrence of node i (shown in Figure 7a for the time period 1991–2000), pi is the node-averaged daily P − E for node i (shown in Figure 10a for the time period 1991–2000), and N is the number of nodes in the SOM (35).
 Changes in the node-averaged daily P − E between 2046–2055 and 1991–2000 and 2091–2100 and 2046–2055 are shown in Figures 10b and 10c. As can be seen from this figure, the average daily P − E is forecast to decrease before the middle of the 21st century in two nodes only: (0,1) and (1,2). After the middle of the 21st century P − E is forecast to decrease in nodes in the lower left part of the SOM, which also show a significant frequency decrease. The increase of P − E is large in nodes with significant frequency increase (see Figures 7b and 7c), and some weak circulation nodes (such as (0,2), (0,0), and (1,3)) with significantly decreasing frequency show relatively strongly increasing P − E. These nodes represent weak circulation patterns. Generally, the nodes experiencing significant decrease in frequency of occurrence show decrease or only small increase in averaged daily P − E.
 The changes presented in Figure 10 represent changes in precipitation and evaporation for each circulation pattern (node) and as such represent thermodynamic changes predicted by the IPCC models independent of any dynamic (circulation) changes. Following equation (1) the average annual P − E over the ice sheet can be calculated for future time periods using values from Figures 7 and 10 as:
where Δfi is the change in frequency of occurrence of node i between the two time periods of interest (2046–2055 and 1991–2000 for Figure 7b and 2091–2100 and 2046–2055 for Figure 7c), Δpi is the change in node-averaged daily P − E for node i between the two time periods of interest (Figures 10b and 10c). This expression can be expanded to
 In this expression, fipi denotes the contribution to the annual P − E of node i at the initial time and the remaining three terms represent contributions to the change in annual P − E between the two time periods. fiΔpi represents the contribution to the change in annual P − E due to changes in the node-averaged P − E and is referred to here as the thermodynamic change, Δfipi represents the contribution to the change in annual P − E due to the change in frequency of occurrence of node i and is referred to here as the dynamic change, and ΔfiΔpi represents the contribution to the change in annual P − E due to the change in the node frequency of occurrence acting on the change in node-averaged P − E and is referred to here as the dynamic/thermodynamic change since changes to P − E due to both types of processes are indistinguishable.
 The P − E change components were calculated between 1991–2000 and 2046–2055, using the approach based on (2), and shown in Figure 11. Figure 11a shows that P − E change per node is similar to the changes in frequency of occurrence (Figure 7b), with a negative change of P − E associated with high-pressure and weak circulation patterns and a positive change of P − E associated with a strong circumpolar trough and three-wave patterns. The thermodynamic change of P − E shown in Figure 11b is positive for most nodes, consistent with the expectation that under a warmer climate P − E will increase over the Antarctic continent.
 As indicated by (2), the sign of the dynamic change is the same as the sign of the change in frequency of occurrence of each node. On a node basis, the magnitude of the dynamic change is similar to the magnitude of the net change in P − E (Figure 11a).
 The dynamic change contributes most positively to the net change (positive values in Figure 11c) in the nodes of the most remarkable P − E change (such as (3,0), (4,0), (5,4), and (6,4)). The term describing both a thermodynamic and dynamic change (Figure 11d) is relatively small compared with the other two components. This indicates the validity of (2), being able to separate thermodynamic and dynamic components of P − E change.
 The total P − E change can be estimated as a sum of the three partitions over all nodes (Table 4). This results in 18.8 mm year−1 (91% of the total trend) for the thermodynamic change, 1.2 mm year−1 (6%) for the dynamic change, and 0.7 mm year−1 (3%) for the combined thermodynamic/dynamic term. The net effect of the dynamic term is small because the sum of differences of node frequencies of two periods is zero and because internodal P − E differences are relatively small compared with the temporal P − E changes per node. This results in the circulation-related decreases indicated in the left and upper right portion of the SOM, nearly canceling the circulation-related increases in the upper center and lower right portions of the SOM (Figure 10c).
Table 4. Sum of P − E Change Components (mm year−1) of the Five-Model Subset From Equation (2) Integrated Over the Entire Antarctic Ice Sheet and Subregions of the Peninsula, West Antarctica, and East Antarcticaa
1991–2000 vs. 2046–2055
2046–2055 vs. 2091–2100
Values in parenthesis give the ratios between the components and the net change.
Antarctic Ice Sheet
Table 4 shows the trend components of the three Antarctic subregions defined, as shown in Figure 6. The total P − E change over the Peninsula is projected to be almost four times larger than the total P − E change over the entire Antarctic ice sheet before the middle of the 21st century. Additionally, the effect of the dynamic term is rather significant, being about 31% of the total P − E change. Over West Antarctica, the total P − E change is about 40% larger than the total P − E change over the entire Antarctic ice sheet. The ratio between the dynamic component and the total P − E change is 18%. The combined thermodynamic/dynamic term is also relatively large over West Antarctica. With the dynamic term, these two circulation-related terms correspond to 25% of the total P − E change. The components of P − E change have rather different values over East Antarctica versus the Peninsula and West Antarctica. During the first part of the twentieth century, the thermodynamic term completely dominates the total P − E change in East Antarctica.
P − E change over the entire Antarctic ice sheet after the middle of the 21st century is smaller than the P − E change before the middle of the 21st century (Table 4 and Figure 12). P − E change of all nodes results in 13.2 mm year−1 (81%) of the total trend for the thermodynamic change, 2.0 mm year−1 (12%) for the dynamic change, and 1.1 mm year−1 (7%) for the third term (Table 4). The thermodynamic change is less dominant than before the middle of the twentieth century in the net P − E change. The dynamic term and the third term are slightly bigger than before the middle of the twentieth century.
 The increase of P − E in the Peninsula after the middle of the 21st century is 47% smaller than before. The role of the thermodynamic term decreases to 44% of the total P − E change. In West Antarctica, P − E change after the middle of the 21st century is only slightly smaller than before the century and 62% of the predicted P − E change is due to the thermodynamic term. In East Antarctica, P − E change after the middle of the 21st century is smaller than before the middle of the 21st century. There the thermodynamic term remains the dominant component contributing significantly to the net P − E change.
 In Figure 13, P − E change in Antarctica and the Southern Ocean is presented. Before the middle of the 21st century, an increase of P − E over the Southern Ocean and Antarctica (see Figure 13, upper left panel) is apparent. This upward trend is especially strong in a sector north of Ross Sea, in the Antarctic Peninsula, and parts of West Antarctica. As noted by Lynch et al. , surface-air temperature over the Antarctic Peninsula will warm considerably and will be associated with increasing P − E according to this set of models. The thermodynamic change is dominating the increase of P − E (Figure 13, upper right panel). The dynamic change is positive over the Southern Ocean, especially along the coast of West Antarctica and the Peninsula (see Figure 13, lower left panel). These are the regions where intense cyclones increase precipitation. The contrast between the small regions eastward and westward of the Antarctic Peninsula reveals that the Peninsula is resolved in the simulations. It should be emphasized that, over these regions of highest positive dynamic change, the thermodynamic change is still dominating (note the different color scales in Figure 13). Marshall  reports that increasing westerlies are associated with the observed warming trend in the Peninsula, which according to this study is associated with increasing P − E. The dynamic change is negative over large parts of the Antarctic ice sheet near the coast, but positive over the adjacent ocean. This indicates the coastal and inland precipitation regions are experiencing different P − E change because of circulation.
 In the midlatitudes (Figure 13, compare corners of the panels), the P − E trend is downward. Midlatitude high-pressure systems become more frequent (see Figures 1 and 7) and along with warming [Lynch et al., 2006], this causes more evaporation over the ocean which decreases P − E. It is reasonable to assume that at least part of the increased moisture which evaporated into the atmosphere in the midlatitudes is transported toward Antarctica by low-pressure systems [see, e.g., Meehl, 1987].
 After the middle of the 21st century, P − E continues to increase in Antarctica and adjacent seas (Figure 14, upper left panel), but now the importance of changing circulation has slightly increased (Figure 14, lower left panel) when compared with the thermodynamic change (Figure 14, upper right panel) than before the middle of the 21st century. Where the thermodynamic component is high, already dominating circulation modes transport increasing amounts of moisture, which seems to occur especially over the Southern Ocean. The dynamic component continues to be important over the sector north of Ross Sea, the Peninsula, West Antarctica, and some coastal regions.
 The models used in this study were ranked based on how well their circulation patterns correlated with the reanalysis, with the aim to select a group of models which best simulate the present-day climate. The inference is that these would also give the most reliable projections for the circulation of the next hundred years. The circulation ranking was carried out using the SOM approach, which validated the model data, revealing differences not apparent in the ensemble mean and variance. In addition to assessing the modeled circulation based on reanalyses, literature-based estimates of P − E in Antarctica were utilized when evaluating the reliability of the models. As was the case with the circulation, the predicted values of P − E values varied significantly between models. Circulation and P − E were not related as systematically as surface-air temperature [see Lynch et al., 2006]. Some models that simulated a synoptic climatology over the Antarctic coastal regions and surrounding oceans that were rather similar to the reanalyses produced estimates of P − E over the Antarctic ice sheet that differed substantially from observationally derived values reported in literature. These biases may be due to many factors including inadequate representation of the moisture physics in the models, an ill-resolved surface energy balance, and too-coarse spatial resolution. Calculation of the atmospheric convergence of moisture has the potential to be a more reliable method to estimate P − E. This, however, would require data in the original model resolution both in time and space, which is unavailable for all the models in this comparison at the moment.
 Several models simulated both a reasonable circulation and estimates of P − E. Hence we selected a subset of five models for the subsequent analysis of circulation-related changes in P − E over Antarctica. All 15 models revealed an increase of P − E in the 2046–2055 time period relative to the present-day, and all but two predicted increases for the end of the 21st century. This increasing P − E is associated with increasing atmospheric moisture content and increasing precipitation. The magnitude of P − E trend varied significantly depending on the model, and an adequate estimate of current day mean P − E was not necessarily reflected in a similarly good estimate of current day multidecadal trends.
 The majority of the circulation patterns extracted from the data were associated with upward P − E trends in Antarctica. The changing frequencies of the circulation types modify the regional patterns of P − E, and circulation patterns forecast to become more common especially after the middle of the 21st century are associated with higher P − E. Increasing snow accumulation does not necessarily mean increased mass of the Antarctic ice sheet as a recent study by Velicogna and Wahr  suggests. They note that the Antarctic ice sheet is losing mass because of the near-coastal discharge of ice, although it is growing in thickness in the interior. Velicogna and Wahr  estimated the rate of mass loss to be 152 ± 80 km3/year in 2002–2005, mostly from the West Antarctic Ice Sheet. By using the ice sheet mask of our study (area = 13.52 × 106 km2) and ice density of 900 kg m−3, we obtain a rate of 5–15 mm year−2 water equivalent. This rate is much more than the projected increase of P − E (0.42 ± 0.01 mm year−2) based on the IPCC model projections utilized in this study. The rate, however, is based on 3 years of data and is likely to contain a significant amount of interannual variability. Longer term mass balance estimates by Rignot and Thomas  suggest that the East Antarctic ice sheet grows 20 ± 21 Gt/year and show the West Antarctic ice sheet loss of 44 ± 13 Gt/year. By combining these two rates and calculating similarly as above, a more moderate rate of mass loss (1.6 ± 2.1 mm year−2) is obtained.
 The upward P − E trends estimated in this study are two orders of magnitude smaller than the observed interannual P − E variability of 20 mm year−2 and one order of magnitude smaller than the observed interdecadal variability [Monaghan et al., 2006] over the Antarctic ice sheet. Using a regional meteorological model calibrated with observations, van den Broeke et al.  found an annual accumulation increase which was sixty times smaller (about 0.15 mm year−2) than the standard deviation of the annual accumulation for 1980–2004 over the grounded Antarctic ice sheet. They suggested that the interannual variability overwhelms the possible long-term increase in P − E.
 In this study, synoptic-scale circulation and associated changes in the net precipitation P − E were analyzed during the 21st century based on data from 15 GCSM simulations. The 15 models were able to simulate the Antarctic circulation reasonably well as an ensemble, but they displayed remarkable intermodel differences as well. Some models simulated rather poorly the present-day circulation, especially the three models which mainly populated a small part of the SOM [see also Lynch et al., 2006].
 Using the SOM approach, we identified regionally dependent changes of P − E associated with the circulation. P − E was generally increasing everywhere in the Antarctic. Most apparent changes were decrease of P − E in the midlatitudes, increase over the Southern Ocean, and increaseof P − E in West Antarctica and the Antarctic Peninsula, which is experiencing considerable warming [Lynch et al., 2006; Turner et al., 2005]. Over the interior of Antarctica, a small increase in P − E can be seen. The interannual variability is, however, so large than in these models that it easily obscures multidecadal P − E changes.
 Three terms were derived from P − E change: thermodynamic change, dynamic change, and a term where changes in circulation act on the changes in thermodynamics. The thermodynamic term is dominating the net P − E change integrated over the entire Antarctic ice sheet and increases P − E of almost every circulation pattern. The dynamic change, on the other hand, led to decreases in P − E for anticyclonic circulation patterns and increases in P − E for patterns with a circumpolar trough. The dynamic term reveals relatively strong differences in changes between the coastal and inland P − E. The circulation-related P − E change was particularly important at the subcontinent scale and was responsible for up to 56% of the net change. The Antarctic Peninsula, West Antarctica, and East Antarctica experience quite different P − E changes, with the dynamic component contributing positively to the net P − E change in West Antarctica and in the Peninsula, but being close to zero in East Antarctica.
 After the middle of the 21st century, P − E continues to increase at smaller rates in the Peninsula, West Antarctica, and East Antarctica. The magnitude of the thermodynamic change was smaller especially in the Peninsula, while the magnitude of the dynamic change remained almost the same. This signifies remarkable, continuing changes in circulation through the 21st century. This rearrangement of P − E due to the changing atmospheric circulation may have quite fundamental implications in the evolution of the Antarctic ice sheet. The amount of P − E change, the dominance of the thermodynamic component, and the local importance of the dynamic change, especially over the coastal regions, are in accordance with the findings of Krinner et al. , who modeled the Antarctic surface mass balance with a significantly higher resolution atmospheric model than GCSMs utilized in this study.
 It is important to keep in mind that intermodel variation in simulated P − E was quite significant as were differences between models and reanalyses in the twentieth century. This intermodel variability suggests a high level of uncertainty in the results, which severely restricts the reliability of these projections.
 We acknowledge the international modeling groups for providing their output for analysis, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) for collecting and archiving the model output, the JSC/CLIVAR Working Group on Coupled Modelling (WGCM) and their Coupled Model Intercomparison Project (CMIP) and Climate Simulation Panel for organizing the model data analysis activity, and the IPCC WG1 TSU for technical support. The IPCC Data Archive at Lawrence Livermore National Laboratory is supported by the Office of Science, U.S. Department of Energy. This research used data provided by the Community Climate System Model project (.www.ccsm.ucar.edu), supported by the Directorate for Geosciences of the National Science Foundation and the Office of Biological and Environmental Research of the U.S. Department of Energy. The Victorian Partnership for Advanced Computing, Melbourne, Australia is gratefully acknowledged for providing computational facilities enabling this study. This paper was much improved by comments from three anonymous reviewers. We thank Dr. Phil Reid (University of Tasmania) for useful discussions. This work was supported by ARC grant FF0348550, NSF grant OPP0229645, and NOAA grant NA03OAR4320179 (PI Richard Seager).