E VALUATION OF GCM SIMULATION MEAN FIELDS (1990-2001)

To evaluate and compare the mean fields of the observed (reanalysis) and GCM-simulated data during the period of concurrent data availability (1990-2001), Taylor diagrams (see Section 3.5) are constructed on a seasonal basis for each of the variables described in Chapter 5.

The similarity between the observed and GCM-simulated mean fields is

characterized in Figure 6.1. As the figure shows, the mean Z500 fields from both models are very similar to the mean fields from the reanalysis data, resulting in a high pattern correlation (>0.95 in each season for both models). The variability in the mean Z500 field is slightly underestimated by HadCM3 during winter, spring, and autumn and CGCM2 during winter, summer, and autumn, resulting in slight displacement of the points from the origin (o) (i.e., normalized standard deviation <1, Figure 6.1). Geopotential heights closer to the surface (Z850) are also simulated well, although pattern correlations are slightly lower than those for Z500 (still >0.95 in each season for both models). The mean HadCM3 Z850 pattern exhibits less spatial variability than observed in all seasons, while the mean CGCM2 Z850 pattern slightly overestimates the spatial variability during spring.

Mean 850-500 hPa layer thickness fields from both models are highly correlated with the observed mean field, although the variability in the field in underestimated (during all seasons by CGCM2 and during winter by HadCM3).

The variability in the mean Q850 field from CGCM2 matches observed variability well. However, while pattern correlations between simulated and observed mean fields are generally high (r>0.90 during winter r>0.80 during spring and autumn), during the summer the correlation between the two fields is only 0.57 (Figure 6.1c). Seasonal

correlations for the HadCM3 humidity variable, RH850, are lowest during winter (0.83) and highest during summer (0.93), with slight overestimation of the spatial variability in each season.

Correlations are much lower for the mean SLP field, especially for HadCM3 during autumn (r=0.41; Figure 6.1d). For CGCM2, correlations are high (>0.90) during summer and autumn, but lower during winter (0.72) and spring (0.65), although the ratio of modeled to observed standard deviation in HadCM3 is close to one in all seasons (Figure 6.1). CGCM2 overestimates the variability in the mean pattern during all seasons, while HadCM3 overestimates the variability during spring and summer, but underestimates the variability during winter. Although the analysis of HadCM3 included in Covey et al. (2003) and included in the IPCC TAR (IPCC 2001) encompasses global data over a longer simulation period, the HadCM3 SLP evaluation summarized in Figure 6.1 strongly resembles the results from these previous global analyses.

The geostrophic flow variables generally show similar agreement as the SLP fields from which they are derived. The correlations between the mean fields of the southerly component of the geostrophic flow (GEOS), are lower for both models during winter and autumn. The variability in the mean GEOS field is overestimated by

HadCM3 in all seasons, with the largest overestimation in summer and autumn (Figure 6.1c, d). CGCM2 generally captures the spatial variability in the mean SLP field better than HadCM3, although the spatial variability is also largely overestimated during summer (Figure 6.1c). The mean fields of the westerly component of the geostrophic flow (GEOW) from CGCM2 exhibit low correlations with the observed field during spring (r=0.47), but greater than 0.70 during the other seasons. During all seasons,

CGCM2 overestimates the spatial variability in the mean field (Figure 6.1). HadCM3 exhibits much higher correlations (>0.80 during winter, spring, and autumn and >0.65 during summer) and the variability in the GEOW field is better simulated, but still overestimated during summer (Figure 6.1c). The mean field of the strength of the resultant geostrophic flow (GEOWS) is well simulated by both models with high correlations (>0.90 during winter, spring, and autumn, and greater than 0.75 during summer) and only slight differences in variability in each season.

The poorest agreement between observed and simulated mean fields is associated with the geostrophic vorticity variables (GEOZS, GEOZW, and GEOZTOT). The mean southerly shear vorticity field (GEOZS) from both models exhibits relatively low

correlation with the observed field (r ranges from 0.24 to 0.44 for CGCM2 and from 0.39 to 0.55 for HadCM3). CGCM2 underestimates the variability in the field during each season (Figure 6.1), and by nearly 50% during spring. Conversely, HadCM3

overestimates the spatial variability in the mean GEOZS field by approximately 100%.

Correlations between the mean HadCM3 GEOZW field and the observed mean GEOZW field are generally in the 0.60-0.70 range, but are slightly higher during spring (0.77).

However, during all seasons, the variability in this field is overestimated. During spring and summer, the mean GEOZW field is more than twice as variable as the observed field.

Correlations between the modeled and observed fields are slightly lower for CGCM2.

While CGCM2 also overestimates the spatial variability in the field during spring and summer, the magnitude of the differences is smaller, resulting in a smaller RMS error (i.e., shorter distance to the origin). Pattern correlations between modeled and observed GEOZT fields are slightly higher for HadCM3 than CGCM2, although values for both

models are low relative to most of the other variables tested. However, HadCM3 also largely overestimates the spatial variability in the mean pattern during spring and summer, with more than twice the level of spatial variability as observed. CGCM2 exhibits similar behavior, although the level of overestimation is much lower.

The results presented here have broad implications for the downscaling work presented in Chapter 8. The Taylor diagrams presented in this section imply:

1) For both models, the variables exhibiting the closest agreement with reanalysis data (in terms of the 1990-2001 mean fields) are the large-scale circulation data (i.e., the geopotential height variables) and 850-500 hPa layer thickness.

2) The mean field of 850-hPa specific humidity is substantially more accurate than the 850-hPa relative humidity field (excepting summer)

3) The derived geostrophic flow and vorticity variables exhibit the lowest level of agreement with observations, with particularly poor correspondence during the summer.

Coupled with the statistical associations presented in Chapter 5, these results suggest that difficulties associated with precipitation downscaling will likely be exacerbated by poor agreement between GCM simulations and observations of the key predictor variables.

Conversely, the variables exhibiting the strongest associations with the surface temperature variables are among the most well-simulated in terms of the 1990-2001 large-scale mean fields.

Figure 6.1. Taylor diagram for 1990-2001 seasonal means of GCM-simulated variables relative to reanalysis data: a) winter (DJF), b) spring (MAM), c) summer (JJA), and d) autumn (SON). Symbols are as follows: HadCM3 Z500 (o), CGCM2 Z500 ( ), HadCM3 Z850 (o), CGCM Z850 ( ), HadCM3 THICK (o), CGCM2 THICK ( ), HadCM3 RH850

(o), CGCM2 Q850 ( ), HadCM3 SLP (∗), CGCM2 SLP (◊), HadCM3 GEOS (+), CGCM2 GEOS (x), HadCM3 GEOW (+), CGCM2 GEOW (x), HadCM3 GEOWS (+), CGCM2 GEOWS (x), HadCM3 GEOZS (∗), CGCM2 GEOZS (◊), HadCM3 GEOZW (∗), CGCM2 GEOZW (◊), HadCM3 GEOZT (∗), and CGCM2 GEOZT (◊).