Does the Atmospheric Global Model MRI-AGCM3.2 Perform Better than CMIP6 Atmospheric Models in Simulating Precipitation?


 The performance of the Meteorological Research Institute-Atmospheric General Circulation model version 3.2 (MRI-AGCM3.2) in simulating precipitation is compared with that of global atmospheric models registerred to the sixth phase of the Coupled Model Intercomparison Project (CMIP6). The MRI-AGCM3.2 with the grid size of 20-km and 60-km and 36 CMIP6 models are forced with observed sea surface temperature for 20-year period from 1995 to 2014. The horizontal resolution of the MRI-AGCM3.2 is relatively finer than CMIP6 models. As for global domain, the reproducibility of MRI-AGCM3.2 models are better than or equal to CMIP6 models in simulating geographical distribution of annual precipitation and intense precipitation events. Models with higher horizontal resolution tend to be better than those with lower resolution in simulating global precipitation. As for East Asia, the performance of MRI-AGCM3.2 models are better than or equal to CMIP6 models in simulating summertime monthly precipitation and the seasonal march in the Japanese rainy season, and extreme precipitation events. Higher horizontal resolution models also tend to perform better than lower resolution models in simulating precipitation over East Asia. The advantage of models with higher horizontal resolution over those with lower resolution in reproducing precipitation is more evident over East Asia than over the globe.


Introduction 42
The performance to simulate current-day climatology by Atmospheric General 43 Circulation Models (AGCMs) is usually assessed by specifying the observed sea surface 44 temperature (SST) as a underlying boundary condition. This sort of simulation is called 45 an Atmosphere Model Intercomparison (AMIP)-type experiment. Lau et al. (1996), 46 Lau and Yang (1996), Liang et al. (2001), Kusunoki et al. (2001) and Kusunoki (2018a) 47 analyzed AMIP-type experiments by AGCMs and reported that simulated precipitation 48 in summer is smaller than observations over East Asia based on AMIP-type experiments. 49 Also, Kang et al. (2002) and Kusunoki 2018a) indicated that most AGCMs do not 50 reproduce the northward marching of summertime rainy band over East Asia. 51 However, Kusunoki et al. (2006), Kitoh and Kusunoki (2008) and Kusunoki 2018a) 52 revealed that AGCMs with higher horizontal resolution perform better than those with 53 lower horizontal resolution with respect to summer precipitation over East Asia. In the 54 case of simulating heavy rainfall events, Kusunoki et al. (2006) and Randall et al. 55 (2007) indicated the advantage of AGCMs with higher horizontal resolution over those 56 with lower horizontal resolution. 57 We have been developing a high horizontal resolution global atmospheric model 58 The pentad mean and monthly mean dataset of the Climate prediction center Merged 157 Analysis of Precipitation (CMAP) provided by Xie and Arkin (1997) are also selected 158 for 20 years through 1995 to 2014. The grid spacing is 2.5 degree. 159 Table 3 summarizes the features of observational data for verification. 160 161 4 Global precipitation 162

Geographical distribution 163
The global distributions of annal precipitation (PAV, Table 3) are compared in Fig.1. 164 In the GPCP 1dd observation (Fig. 1a), precipitation is large over the Indian Ocean, 165 over the tropical area of the Pacific Ocean and the Atlantic Ocean, and over the Amazon. 166 Similar feature also appears in the GPCP data of 2.5 degree grid interval (Figs. 1b). 167 Precipitation by the CMAP of 2.5 degree grid interval (Fig. 1c The 20-km model (SPD,Fig. 1d ) and the 60-km model (HPD,Fig. 1e) tend to 171 overestimate precipitation over the maritime continent and the South Pacific objective skill measures based on the GPCP 1dd (green circle). The location of green 180 circle means perfect simulation. Figure 2a shows the bias and root mean square error 181 (RMSE) of simulations. All models show positive bias partly due to the overestimation 182 of precipitation over the Pacific Ocean (Figs. 1 and S1). Black circles in Fig. 2 indicates 183 the multi-model ensemble (MME) average skill which is based on the 2-dimensional 184 global spatial distribution of precipitation constructed with the MME mean of CMIP6 185 models (Fig. 1f). Black squares in Fig. 2 display the average of the skill of each CMIP6 186 models (AVM). In the case of linear skill measures such as bias (Fig. 2a,horizontal 187 axis), the MME average is identical to the AVM. The biases of the 20-km model (red S) 188 and the 60-km model (purple H) are slightly larger than the MME average (black circle) 189 and the AVM (black square) of CMIP6 models (Fig. 2a). 190 The RMSE (Fig. 2a,vertical axis) of the 20-km and 60-km models are relatively 191 smaller than those of CMIP6 models. Since RMSE is a nonlinear skill measure, the 192 MME average (black circle) differs from the AVM (black square). The MME average 193 of CMIP6 models (black circle) is almost higher than the RMSEs of all individual 194 CMIP6 modes (black characters). This advantage of MME average is consistent with 195 previous studies such as Lambert and Boer (2001), Gleckler et al. (2008), Reichler and 196 Kim (2008, Kusunoki and Arakawa (2015) and Kusunoki (2018a). 197 To show the uncertainty of observation, the GPCP data of 2.5 degree grid interval 198 (green square) and the CMAP data of 2.5 degree grid interval (green diamond) are also 199 plotted as well as the GPCP1dd (green circle). The uncertainty (spread) among 200 observations (three green marks) are smaller than the magnitudes of bias and RMSE by 201 models. 202 deviation. The radial distance of one means perfect simulation. The angle from the 207 y-axis implies the spatial correlation coefficient. The perfect simulation coincides with 208 the location of green circle. The spatial correlation coefficients of the 20-km (red S) and 209 60-km (purple H) models are relatively larger than those of individual CMIP6 models. 210 The spatial correlation coefficient of the MME average (black circle) of CMIP6 models 211 is almost higher than any other models. The symbols of all the models are plotted outer 212 area of the radius one quadrant. It means that spatial variability of all simulations is 213 overestimated. 214 In summary, Fig. 2 indicates that the reproducibility of the 20-km and 60-km models 215 are equivalent to or better than CMIP6 models in simulating global distribution of PAV. 216 This is similar to the result of previous studies on CMIP5 models (Kusunoki 2017 ;Fig. 217 1) which reported the advantage of the 60-model over CMIP5 atmospheric models in 218 simulating global distribution of PAV. 219 220 4.3 Extreme precipitation events 221 Table 3 shows the definition of extreme precipitation indices used for verification 222 based on Frich et al. (2002). The maximum 5-day precipitation total (R5d) is often used 223 to define heavy precipitation events leading to water related disaster such as inundation 224 and landslide. The maximum 1-day precipitation total (R1d) is widely used to define extreme precipitation events happening once a year. On the other hand, maximum 226 consecutive dry days (CDD) is an index estimating the possibility of dry condition 227 ought. PAV is also included in Table 3 for comparison. 228 Figure 3 compares the reproducibility of the 20-km and 60-km models with those of 229 CMIP6 models as to four extreme indices. The spatial correlation coefficient of global 230 distribution of extreme precipitation events is selected as skill measure. As for PAV, the 231 spatial correlation coefficients of the 20-km and 60-km models are almost the same as 232 that of the best CMIP6 model and the MME average (black circle). In the case of R5d, 233 the skill of 60-km models (purple lines) are better than the AVM of CMIP6 models 234 The correlation coefficient between grid spacings and spatial correlations coefficient for shows how model performance depends upon on grid spacing for all four extreme 249 events (Table 3). All four negative correlation suggests the advantage of higher 250 horizontal resolution models over low resolution models, but statistical significance 251 above 99% level is only recognized for PAV and R5d. 252 253 5 Precipitation over East Asia 254

Geographical distribution 255
The rainy season over Japan (the Baiu) starts in the middle of May and terminates in 256 the end of July. Figure  The MME average of CMIP6 models also underestimates precipitation of the Baiu 266 rain band (Fig. 6f). The spatial coefficient C of the best CMIP6 model is high as 0.837, 267 but precipitation is still underestimated (Fig. 6g). The worst CMIP6 model shows 268 erroneous excessive precipitation to the south of 25°N (Fig. 6h). 269 Figure 7 shows the objective skill scores of models for June precipitation over East 270 Asia (Fig. 6). Most models underestimate precipitation (Fig. 7a,horizontal axis). In 13 terms of RMSE (Fig. 7a,vertical axis), the RMSE of the 20-km and 60-km models are 272 relatively smaller than CMIP6 models and is smaller than the MME average of CMIP6 273 models (black circle). In the Taylor diagram (Fig. 7b), almost entire models are 274 displayed inside the radius one quadrant. This mans the underestimation of spatial 275 variability. The spatial correlation coefficients of the 20-km and 60-km models are 276 relatively larger than those of CMIP6 models. 277 In the case of July which corresponds to the later half of the Baiu season, the 278 reproducibility of the 20-km and 60-km models are also higher than CMIP6 models in 279 simulating July precipitation over East Asia region with respect to bias, spatial 280 variability and spatial correlation coefficient (Figs. S2-S3). Models with higher 281 horizontal resolution tend to perform better than those with lower resolution in 282 simulating monthly precipitation in warmer season over East Asia (Fig. S4). 283 The superiority of the 20-km and 60-km models over CMIP6 models in simulating 284 summer precipitation is similar to the results by Kusunoki (2018a)  show similar northward migration of the Baiu (Figs. 8b-c). In the 20-km model (Fig. 8d), 293 precipitation of the Baiu is underestimated and northward migration is not clear. The 294 60-km model (Fig. 8e) seems to simulate larger precipitation in the Baiu season than the 295 20-km model (Fig. 8d). The MME average CMIP8 models apparently underestimate 296 precipitation and the location of the Baiu is shifted to north as compared to observation 297 (black contour). Even the best CMIP6 model underestimate precipitation amount of the 298 Baiu (Fig. 8g). The worst model shows the erroneous location of Baiu (Fig. 8h). 299 In terms of objective skill scores (Fig. 9), most models underestimate precipitation 300 (Fig. 9a,horizontal axis). The RMSEs of the 20-km and 60-Km model are relatively 301 less than those of individual CMIP6 models and the MME average of CMIP6 model 302 (Fig. 9a,vertical axis). Most models underestimate spatial variability (Fig. 9b). The 303 spatial correlation coefficients of the 20-km and 60-km models are relatively higher 304 than those of individual CMIP6 models and the MME average of CMIP6 model (Fig. 305 9b). Models with higher resolution tend to perform better than those with lower 306 horizontal resolution in simulating the seasonal march of rainy season over Japan (Fig. 307 10). The correlation coefficient between grid spacings and spatial correlations 308 coefficient for seasonal march of the Baiu is -0.520 which is larger than the 99% 309 statistical significance level. 310 In summary, the reproducibility of the 20-km and 60-km models is better than 311 CMIP6 models in simulating the seasonal march of the Japanese rainy season. This is 312 similar to the results by Kusunoki (2018a) with respect to CMIP5 models. 313 314 5.3 Extreme precipitation events 315 Figure 11 compares the performance of models in simulating extreme precipitation 316 events over East Asia. The skills of 20-km and 60-km models are equivalent to or better than those of CMIP6 models for all four indices. The advantage of the 20-km and 318 60-km models (Fig. 11) are more evident than global case (Fig. 3). 319 Figure 12 depicts how the model skill depends on the horizontal resolution of models 320 in the case of spatial correlation coefficient for four extreme precipitation indices over 321 East Asia. Large negative correlation coefficients for all four indices imply the 322 advantage of higher horizontal resolution models over lower resolution models. 323 Magnitude of statistical correlation coefficients and significance levels of all four 324 indices over East Asia (Fig. 12) are larger than those over the globe (Fig. 5). This 325 suggests that the advantage of higher horizontal resolution models is much more evident 326 over East Asia than over the globe. is selected for 20 years from 1981 to 2000. In Fig. 12, the performances of CMIP6 344 models are better than those of CMIP5 models in terms of AVM and MME. CMIP6 was 345 accomplished about 10 years after CMIP5. Figure 12 suggests the climate models were 346 apparently improved during this last decade due to continuous and great efforts by 347 climate modelling scientists. The 20-km and 60-km models perform relatively better 348 than CMIP5 and CMIP6 models in simulating the spatial pattern of global annual 349 precipitation. 350 As for monthly precipitation over East Asia, also the performances of CMIP6 models 351 are better than those of CMIP5 models (Fig. S5). The 20-km and 60-km models perform 352 relatively better than CMIP5 and CMIP6 models especially in warm season (Fig. S5). 353 The performance of 20-km and 60-km models are better than or equal to CMIP5 and 354 CMIP6 models in simulating R1d over East Asia (Fig. S6). The 20-km and 60-km 355 models perform relatively better than CMIP5 and CMIP6 models in simulating the 356 seasonal march of Japanese rainy season (Fig. S7). 357 358

Conclusions 359
Our results are summarized as follows. 360 1. The performance of MRI-AGCM3.2 models is higher than or equal to CMIP6 361 atmospheric models with respect to the geographical distribution of annual 362 precipitation and intense precipitation over the globe.
2. Models with higher horizontal resolution perform better than those with lower 364 resolution in simulating global precipitation. 365 3. The reproducibility of MRI-AGCM3.2 models is higher than or comparable to 366 CMIP6 atmospheric models as to monthly precipitation, the seasonal march of 367 Japanese rainy season and extreme precipitation events over East Asia. 368 4. Models with higher horizontal resolution perform better than those with lower 369 resolution in simulating precipitation over East Asia. 370 5. The advantage of higher horizontal resolution models over lower resolution models is 371 more evident over East Asia than over the globe.      Fig. 1 The global distributions of climatological annual precipitation PAV (mm day -1 ). 594 a-c Observations (Table 2). d SPD. Averaged period is 20 years from 1995 through 595  (Table 2). 609 Colored letters denote the MRI-AGCM3.2 models. Red S shows the 20-km model.   (Table 3). 643 For the 60-km model, only the first member is chosen out of four ensemble simulation. 644 The skill measure is the spatial correlation coefficient C verified against the GPCP 645 1ddv1.3 observation for the global distribution of extreme indices. Horizontal lines 646 show statistical significance levels. Scatter plot in the case of PAV is displayed in Fig. 4. 647 Larger negative correlation coefficient means that the advantage of higher resolution 648 model over low resolution model is much more evident.  2000. Skill measure is spatial correlation coefficient against the GPCP 1ddv1.3 695 observation. For details of CMIP5 models, see Table 5 in Kusunoki and Mizuta (2021). 696

Fig. S1
The biases of annual precipitation PAV (mm day -1 ). a Reference observation 706