Comparison of regional characteristics of land precipitation climatology projected by an MRI-AGCM multi-cumulus scheme and multi-SST ensemble with CMIP5 multi-model ensemble projections

Ensembles of climate change projections created by general circulation models (GCMs) with high resolution are increasingly needed to develop adaptation strategies for regional climate change. The Meteorological Research Institute atmospheric GCM version 3.2 (MRI-AGCM3.2), which is listed in the Coupled Model Intercomparison Project phase 5 (CMIP5), has been typically run with resolutions of 60 km and 20 km. Ensembles of MRI-AGCM3.2 consist of members with multiple cumulus convection schemes and different patterns of future sea surface temperature, and are utilized together with their downscaled data; however, the limited size of the high-resolution ensemble may lead to undesirable biases and uncertainty in future climate projections that will limit its appropriateness and effectiveness for studies on climate change and impact assessments. In this study, to develop a comprehensive understanding of the regional precipitation simulated with MRI-AGCM3.2, we investigate how well MRI-AGCM3.2 simulates the present-day regional precipitation around the globe and compare the uncertainty in future precipitation changes and the change projection itself between MRI-AGCM3.2 and the CMIP5 multiple atmosphere–ocean coupled GCM (AOGCM) ensemble. MRI-AGCM3.2 reduces the bias of the regional mean precipitation obtained with the high-performing CMIP5 models, with a reduction of approximately 20% in the bias over the Tibetan Plateau through East Asia and Australia. When 26 global land regions are considered, MRI-AGCM3.2 simulates the spatial pattern and the regional mean realistically in more regions than the individual CMIP5 models. As for the future projections, in 20 of the 26 regions, the sign of annual precipitation change is identical between the 50th percentiles of the MRI-AGCM3.2 ensemble and the CMIP5 multi-model ensemble. In the other six regions around the tropical South Pacific, the differences in modeling with and without atmosphere–ocean coupling may affect the projections. The uncertainty in future changes in annual precipitation from MRI-AGCM3.2 partially overlaps the maximum–minimum uncertainty range from the full ensemble of the CMIP5 models in all regions. Moreover, on average over individual regions, the projections from MRI-AGCM3.2 spread over roughly 0.8 of the uncertainty range from the high-performing CMIP5 models compared to 0.4 of the range of the full ensemble.


Introduction
To establish mitigation and adaptation strategies for future climate change, studies on climate change and assessment of its impacts have been conducted worldwide. These recent studies have frequently used future climate projections simulated by multiple general circulation models (GCMs) in the framework of the Coupled Model Intercomparison Project (CMIP); e.g., the CMIP phase 5 (CMIP5; Taylor et al. 2012). While the projections from GCMs and these study results are provided with 100-km or coarser resolution, strategies must be developed for regional and local scales. Thus, there is a scale mismatch between these studies and stakeholders.
Dynamical and statistical downscaling techniques may be used to overcome the mismatch, yet confidence in the resulting scenarios is not always increased by the downscaling (Wilby and Dessai 2010;Hall 2014;Shepherd 2014). The downscaled climate information at a regional scale can include regional processes and the effects of complex terrain that GCMs cannot resolve. However, bias in the large-scale circulation simulated by GCMs leads to bias in the downscaled atmosphere from the GCM output (Hall 2014). Moreover, regional climate models (RCMs) cannot represent phenomena with large temporal or spatial scales beyond their boundaries nor phenomena that are not well-represented in GCMs, such as tropical cyclones. The spatial patterns of the regional precipitation climatology from downscaling experiments strongly depend on the pattern from GCMs (Šeparović et al. 2013) as do the regional future changes (Inatsu et al. 2015;Karmalkar 2018). Therefore, GCMs with sufficiently high horizontal resolution are required to simulate the climate realistically from global to regional scales. The added value obtained by using enhancedresolution GCMs for various climate variabilities (e.g., in reproducing large-scale circulation features like the El Niño-Southern Oscillation) has been reviewed by Haarsma et al. (2016).
The Meteorological Research Institute (MRI) of the Japan Meteorological Agency has been developing such a high-resolution GCM for more than 10 years (Mizuta et al. 2006). At present, the atmospheric GCM version 3.2 (MRI-AGCM3.2; Mizuta et al. 2012) is generally run with horizontal resolutions of 60 km and 20 km and has successfully simulated the global distribution of tropical cyclones ) and regional characteristics such as the seasonal march of the East Asian monsoon (Kusunoki 2018). MRI-AGCM3.2 has been widely recognized as one of the GCMs participating in CMIP5 (IPCC 2013). In addition, the climate dataset generated by MRI-AGCM3.2 has been utilized in various international programs, for instance the Coordinated Regional Downscaling Experiment (CORDEX; https://www.cordex. org, Gutowski Jr. et al. 2016) in the World Climate Research Program, and the CMIP6 project, the High Resolution Model Intercomparison Project (HighResMIP; Haarsma et al. 2016). In Japan, domestic impact assessments and adaptation measures are often based on the climate simulations by MRI-AGCM3.2.
Climate simulations by MRI-AGCM3.2 have been designed to reproduce well regional-scale climatology and extreme events with reasonable quality and then to investigate their future changes . Present-day climate simulations have been performed using an observed sea surface temperature (SST) dataset while future climate simulations have used various future change patterns of SST and sea ice estimated from the CMIP simulations. In general, prescription of observed SST in AGCM simulations can reduce SST bias and the consequent systematic bias in the present-day climate simulations relative to coupled atmosphereocean GCMs (AOGCMs). The feedback by the atmosphere-ocean coupling process, however, is not simulated in an AGCM, which can lead to distorted precipitation fields in some regions (He and Soden 2016;. As well as the ensemble with different SST changes, an ensemble with different cumulus convection schemes has been created with the 60-km resolution MRI-AGCM3.2 (MRI-AGCM3.2H) because the computational resources required are relatively small compared to the 20-km resolution MRI-AGCM3.2 (MRI-AGCM3.2S). Thus, the ensembles with MRI-AGCM3.2 can be used to assess future projections and their uncertainty arising from the differences in SST changes and model physics.
In addition to its good reproduction of tropical cyclones, the performance of MRI-AGCM has been investigated for climatology and extreme events on various spatial scales (e.g., Mizuta et al. 2012). Focusing on the assessment of simulated precipitation and its future change, Kusunoki (2017) indicated that MRI-AGCM3.2H performs well in simulating global precipitation, comparably to or better than atmospheric-MIP (AMIP) models in the CMIP5 archive. Kitoh and Endo (2016) and Mizuta et al. (2017) assessed global precipitation, including extreme weather events, using MRI-AGCM3.2S and MRI-AGCM3.2H, respectively. The high resolution of MRI-AGCM3.2 has allowed realistic simulations of regional precipitation (e.g., Endo et al. 2012Endo et al. , 2017Kusunoki and Mizuta 2013;Ose 2017Ose , 2019Surendran et al. 2019). For instance, Endo et al. (2012) reported that MRI-AGCM3.2 at both resolutions realistically simulates the present-day precipitation over several areas in Asia, and discussed the contributions of SST changes and model physics to the uncertainty in future changes of regional precipitation using MRI-AGCM3.2H. Okada et al. (2017) and Kusunoki (2018) evaluated the modeled seasonal march of the rain band that appears in the East Asian monsoon season. In addition, Okada et al. (2017) showed that the difference in the future SST patterns in the tropics affects the precipitation change over Japan using MRI-AGCM3.2S projections.
While the ability of MRI-AGCM to accurately represent precipitation climatology over the globe and specific regions has been demonstrated, a comprehensive understanding of its ability and the projections of future precipitation change in regions around the globe has yet to be established. Moreover, there has been insufficient discussion of the global and regional properties of the uncertainty in future projections with the MRI-AGCM3.2 ensemble. For impact assessments of climate change, information about the most severe situations is very useful. The MRI-AGCM3.2 ensemble has a limited size of about 10 members due to its high horizontal resolution. The limitations of a small ensemble and also a single model may lead to an undesirably biased future projection with undesirably small uncertainty. The CMIP provides an ensemble of opportunity, although it is not completely clear that it provides the full uncertainty in climate projections (Knutti 2010). Thus, a comparison with the uncertainty from the CMIP ensemble is helpful to examine the limitations of the MRI-AGCM3.2 ensemble and, furthermore, gives information about whether a single AGCM can cover the uncertainty from the CMIP ensemble.
In this study, we use a comparison with the CMIP5 ensemble to comprehensively evaluate the regional precipitation climatology over the globe simulated by MRI-AGCM3.2. The precipitation climatology is the output most frequently analyzed in impact studies. To investigate the extent to which the MRI-AGCM3.2 projections are adequate for discussing regional precipitation changes, as well as projections of severe situation, our evaluation is conducted with respect to two points: how well does the simulated present-day precipitation represent the observations over individual regions, and to what degree does the MRI-AGCM3.2 ensemble capture the uncertainty in the future changes of regional precipitation projected with the CMIP5 multi-model ensemble (MME). Our principal purpose is to identify the characteristics of future changes and their uncertainty at a regional scale. Thus, we analyze the CMIP5 AOGCM ensemble throughout our analysis, instead of AMIP-type model simulations that are identical with the MRI-AGCM3.2 simulations. Evaluations of the present-day precipitation are conducted to identify model performance in representing the observations. Then, based on this performance, we create various subsets of the CMIP5 models and compare the future projections with them. In addition, comparisons between MRI-AGCM3.2 and the CMIP5 AOGCMs can provide information on the advantage of forcing by observed SST for simulations at regional scale. Comparisons of the presentday precipitation between MRI-AGCM3.2 and AMIP models are given in the Supplementary Material.

Climatological dataset from MRI-AGCM3.2
The climatological dataset used in this study is simulated using two horizontal resolutions of MRI-AGCM3.2 as described above. MRI-AGCM3.2H has a horizontal resolution of TL319 (triangular truncation at 319 waves;~60 km) and 64 vertical layers with the top at 0.01 hPa, and MRI-AGCM3.2S has a resolution of TL959 (~20 km) with the same vertical structure as MRI-AGCM3.2H. Three different cumulus convection parameterization schemes are implemented in MRI-AGCM3.2H: the Kain-Fritsch (KF) scheme (Kain and Fritsch 1990), the Arakawa-Schubert (AS) scheme (Randall and Pan 1993), and the Yoshimura (YS) scheme (Yoshimura et al. 2015). MRI-AGCM3.2S was only run using the YS scheme.
Present-day climate simulations are performed using the UK Met Office Hadley Centre's sea ice and SST dataset (HadISST1.1; Rayner et al. 2003) for historical SST and sea ice concentration observations. Future climate simulations are based on the representative concentration pathway (RCP) 8.5 scenario. Using spatial patterns of changes in SST and sea ice under the RCP8.5 scenario from the CMIP5 models, cluster analysis of the patterns produced three clusters of models. The SST change patterns (C1-C3) corresponding to these three clusters and their mean (C0) were estimated. Details of the analysis and the method of creating the four patterns are described in Mizuta et al. (2014). The four spatial patterns were applied to the future climate simulations with both resolutions of MRI-AGCM3.2. In sum, there are three (one) ensemble members for the present-day climate simulation from MRI-AGCM3.2H (MRI-AGCM3.2S) and 12 (4) members for the future climate simulation. The spatial pattern of standard deviation of the four SST changes from C0 to C3 is shown in Fig. 1. Similar patterns were obtained for the annual mean and two seasonal means, June-August (JJA) and December-February (DJF). There were large deviations in highlatitude regions as well as in the northwest regions of the Pacific and the Atlantic.
We computed annual mean precipitation P ANN , boreal summer (JJA) mean precipitation P JJA , and boreal winter (DJF) mean precipitation P DJF from monthly data over 20 years (1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003) for the present-day climate and 20 years of the late twenty-first century (2080-2099) for the RCP8.5 future climate scenario. In order to scale the future change of precipitation, annual mean surface air temperature (hereafter referred to as temperature) was also estimated. The ensemble members used are summarized in Table 1. The MRI-AGCM3.2H ensemble members for present-day and future climate are labeled HP and HF, respectively, and the MRI-AGCM3.2S members SP and SF. The scheme used by each member is specified by adding the scheme name; for example, HPYS for the present-day simulation by HP with the YS scheme.
Regional  we mainly analyze precipitation over land because of the importance for impact studies. To confirm the fundamental simulation ability as a GCM, global precipitation is also evaluated. The assessment of the geographical distribution of precipitation is performed using the skill scores proposed by Taylor (2001). The score S is defined as where R is the spatial correlation coefficient between reference observation and simulation and σ is the standard deviation of the simulation spatially normalized by the observation. The maximum correlation attainable R 0 is assumed to be 1 following Kusunoki (2017). We evaluated the simulations of the present-day precipitation averaged over 20 years on each native grid spacing of MRI-AGCM3.2H/S and also the simulations with MRI-AGCM3.2S interpolated onto the spacing of MRI-AGCM3.2H. Future change was estimated on the native grid spacing of each model without the spatial interpolation.

CMIP5 multi-model ensemble
The comparisons with the CMIP5 models used the outcomes simulated by 50 AOGCMs for the present-day climate (CMIP5 Present_All ) and by 42 AOGCMs for the  RCP8.5 future climate (CMIP5 RCP8.5_All ) (Taylor et al. 2012). One ensemble member run for each model was used. Annual and seasonal mean precipitation over the same period as the MRI-AGCM3.2 dataset were computed from monthly mean data with the original horizontal resolution of each model. We then interpolated the 20-year averaged values onto the MRI-AGCM3.2H grid. The annual mean temperature used as the scaling of future precipitation change was also interpolated onto the MRI-AGCM3.2H grid for the 20-year averaged values. The models are listed in the Supplementary (Table S1).
In addition to the full ensemble of all model members, the following subsets of high performance models were produced: the subsets with high Taylor's skill scores for global precipitation (CMIP GhighS ) and for regional land precipitation (CMIP RhighS ), and the subsets with small mean absolute error (MAE) for global precipitation (CMIP GsmallE ). Each of the subsets has 12 members and is the same size as the set of HF members and the members of these subsets are included in both CMIP5 Presen-t_All and CMIP5 RCP8.5_All . Hence, we investigated the characteristics of the future projections with HF, relative to the projections with not only the CMIP5 full set but also the subsets with high performance.

Observational products for validation
Simulations of present-day precipitation were validated against multiple precipitation products because of the discrepancies in precipitation estimates among observational products reported by Sun et al. (2018). We used the following observational datasets. The products based on global gauge observations, which are provided over land with a grid spacing of 0.5°, are the Global Precipitation Climatology Centre (GPCC) full data reanalysis v.7.0 (Schneider et al. 2016 The MSWEP product is provided at high resolution and covers the whole of our target period and thus was adopted as the reference dataset for the validation. MSWEP has precipitation estimates generally lying toward the mean of those from other datasets ( Fig. 6 in Sun et al. 2018). The observations were interpolated onto the MRI-AGCM3.2H grid. When evaluating Taylor's score and the model bias in the present-day precipitation by SP without spatial interpolation, the MSWEP precipitation was interpolated onto the MRI-AGCM3.2S grid.

Results
3.1 Evaluation of the performance of the present-day simulations 3.1.1 Global distribution of the simulations of present-day precipitation Figure 3 shows the global distribution of present-day P ANN derived from observations and the bias in the model simulation. Compared with the mean of CMIP5 Present_All with a maximum bias of more than 1% (Fig. 3f), MRI-AGCM3.2 simulates P ANN with a bias of less than 0.5% over most areas from the Tibetan Plateau through East Asia and Australia ( Fig. 3b-e). The three members of HPs, HP ALL , have a similar pattern of bias in the global land precipitation (Fig. 3c-e). Differences in the horizontal resolution between HPYS and SPYS are not largely responsible for the global distribution of the bias (Fig. 3b, e). Each member of MRI-AGCM3.2 simulates more precipitation in the Pacific intertropical convergence zone than the observations, as pointed out by Kusunoki (2017), and also over the Indian Ocean with different patterns for the three members. An overestimate of P ANN is notable over the central North Pacific and the western tropical Pacific, which is not seen or is small in the mean of CMIP5 Presen-t_All . Additionally, there is a large difference in northern South America where MRI-AGCM3.2 overestimates the precipitation and the mean of CMIP5 Present_All underestimates it. The spatial pattern of the bias is similar between the means of CMIP5 Present_All , CMIP5 GhighS , and CMIP5 GsmallE (Fig. 3f−h).

Regional performance of the simulations of presentday precipitation
In order to evaluate how well each member reproduces land P ANN at regional scale, Taylor's skill scores and model biases in the regional mean with respect to MSWEP in each of the 26 regions shown in Fig. 2 are summarized in Figs. 4 and 5, respectively. We also consider the performance of CMIP5 GhighS (i.e., the subset exhibiting the highest 12 skill scores for global precipitation; Fig. 4) and CMIP5 RhighS (i.e., the subset for regional land precipitation; Fig. 5), as well as the performances of HP All and CMIP5 Present_All . We used the skill score as the main metric to select the subsets for comparison because the difference of the pattern from the observations results in a large bias. Furthermore, corrections of the spatial pattern The scores for all HP members exceed 0.8 in 16 regions, and in an additional five regions some HPs have a score > 0.8, but not in wet regions of AMZ (region number 7) and SEA (24), dry regions of WSA (9) and SAH (14), and a high-latitude region of CGI (2; Fig. 4). The scores in CGI (2) are less than 0.8 but are within the discrepancy between the observations, as in the case with Europe (11−13) and Asia (18−21). There are high scores > 0.9 in North America (3−4) and Europe (11,13), in addition to TIB (21), EAS (22), and SAU (26) where the bias is small in Fig. 3. The global precipitation is also represented with a score above 0.9. Such a high score > 0.9 is approximately equal to or greater than the highest score of CMIP5 Present_All . The CMIP5 GhighS ensemble includes members with scores lower than the 50th percentile of CMIP5 Present_All in all regions and also for global land (GLB Land in Fig. 4). That is to say, CMIP5-GhighS can represent the pattern of global P ANN realistically but does not do so well for the patterns of regional land P ANN . HP All , by contrast, represents both patterns closer to the observations because the average score of 0.94 over HP All for global P ANN exceeds any scores from CMIP5 Present_All (GLB in Fig. 4) and all scores of HP All for regional land P ANN are above the lowest score of CMIP5 GhighS , apart from the three regions WSA (9), SAS (23), and SEA (24). All HP members show a higher score than the top three scores of the 75th percentile or more in CMIP5 GhighS , in 11 regions. Interpolating P ANN onto the MRI-AGCM3.2H grid, SPYS shows a higher score than HPYS in 12 regions. In four regions including TIB (21), the score of SPYS exceeds all scores of HP All . Without the interpolation, the score decreases in all regions and the decrease is 0.05 on average over the regions. For seasonal precipitation, HP All has a score > 0.8 in 12 regions in JJA and 19 regions in DJF (Fig. S1). The scores in SAU (26) are greater than the highest score of CMIP5 Present_All in both the annual and the seasonal means. Figure 5 shows that some HP members represent P ANN with a bias comparable to the discrepancies between the observations in 12 regions; e.g., South America (7-8, 10), South and Southeast Asia (22-24), and Australia (25-26). In the three regions CGI (2), SSA (10), and CEU (12) in particular, all biases from HP All are within the discrepancy of the observations. In CGI and CEU, all skill scores are also within the discrepancy (Fig. 4), indicating that HP All represents precipitation quantitatively and spatially close to reality. As seen in Fig. 3, the bias in TIB (21) is more than 30% smaller in the mean of HP All than in the 50th percentiles of CMIP5 Present_All and also CMIP5 RhighS . Similarly, the bias is approximately 10% smaller in EAS (22) and SAS (23) and > 25% smaller in NAU (25). All biases from HP All are smaller than the largest bias from CMIP5 RhighS in all regions. Focusing on a quartile of CMIP5 RhighS indicating the three smallest biases, the bias is larger in HP All . However, we should note that the combination of three members selected from CMIP5 RhighS varies region-byregion, whereas the same HP members are evaluated consistently in every region.
Although HP ALL simulates well the spatial pattern of global and global land precipitation relative to CMIP5 Present_All (Fig. 4), the members overestimate the precipitation, in particular, the global land precipitation with a bias comparable to the 75th percentile of CMIP5 Present_All and CMIP5 GhighS . This is because areas of negative bias are more widely distributed in the mean of CMIP5 Present_All than in HP ALL (Fig. 3) and, consequently, the bias is offset globally in CMIP5 Present_All . MAE in both the global land and sea and global land precipitation from HP ALL is smaller or comparable to the smallest MAE from CMIP5 Present_All (not shown). Depending on the convection scheme used, P JJA simulated by HP All are within the spread of observations in 17 regions and P DJF in 15 regions (Fig. S2). In CEU (12), all biases in the annual and the seasonal means are comparable to the discrepancies between observations. SPYS simulates P ANN with less bias than HPYS in more than half the regions on the common MRI-AGCM3.2H grids, and also P JJA in 10 regions and P DJF in 16 regions. In addition, the bias in P ANN from SPYS is small regardless of the convection scheme used in HP All in seven regions including TIB (21). The skill score reduces in all regions when evaluating the SPYS precipitation on the native grid (Fig. 4). On the other hand, the bias in the regional mean precipitation is almost equal in cases with and without the interpolation except for WSA (9) with the complex terrains of the Andes, where the bias reduces by 10% in the evaluation on the native grid. This implies that there may be a difference in the spatial pattern of precipitation between MSWEP and SPYS.

Model rank for the performance over the regions
To gain an overview of the performance of each model, we ranked 55 models using MAE and skill score: three of HP All , the interpolated SPYS, 50 of the individual model members of CMIP5 Present_All , and the MME mean of CMIP5 Present_All . The rank of each model and the MME mean were computed as follows. First, in each region, each model was ranked from 1 to 55 according to its MAE or score, with rank 1 assigned to the model with the smallest MAE or the highest score. Secondly, for each model, the order/rank in each region was averaged over all the regions. A small overall rank corresponds to the model having a smaller MAE or a higher score in many regions. The x-axis in Fig. 6 denotes the rank from 1 to 55 based on the average rank and the y-axis denotes the rank averaged over the regions. HP All and SPYS always rank in the top 15 models for both metrics except for MAE in P ANN from HPKF. In particular, HPYS and SPYS are always in the top five models for the seasonal precipitation. For the skill score, individual members of HPs and SPYS are often ranked higher than the MME mean of CMIP5 Present_All . HPKF has the best average rank of 13.7 on the score for P ANN , and HPYS and SPYS follow HPKF with average ranks of 14.0 and 15.2, respectively. HPYS has the best rank on the score for P JJA , followed by SPYS and HPKF. Thus, HP All and SPYS can simulate the spatial pattern and the regional mean of the present-day precipitation close to the observations in more regions than can the CMIP5 models. The CMIP5 models change their rank depending on the metric, and their ranks are not always above those of HP All . Namely, prescribing observed SST for a GCM contributes to a more realistic spatial pattern and amount of regional precipitation climatology around the globe. The order of rank is almost the same when using the AMIP-type simulations in the CMIP5 (Fig. S3), and HP All is ranked around the MME mean of the AMIP simulations, especially for the skill score.
3.2 Future change projection for the RCP8.5 scenario 3.2.1 Global comparison of the uncertainty in future precipitation change Figure 7a-d illustrates the future change of P ANN averaged over the projections by the HF All ensemble and the MMEs of CMIP5 RCP8.5_All , CMIP5 GhighS , and CMIP5 GsmallE . As described in IPCC (2013), the mean of CMIP5 RCP8.5_All projects an increase of global land P ANN at the 90% confidence level in most regions of the middle-and high-latitude Northern Hemisphere and a decrease around the Mediterranean (Fig. 7b). The regions with a significant change extend farther in HF All than in CMIP5 RCP8.5_All (Fig. 7a), and the sign of the change coincides with the sign given by CMIP5 RCP8.5_All in most regions (Fig. 7e). In Australia and northern South America, the mean of HF All projects an increase at the 90% confidence level while that of CMIP5 RCP8.5_All projects an insignificant decrease (Fig. 7e). Figure 7h shows that the mean of HF All projects a noticeably wetter climate over the land than the mean of CMIP5 RCP8.5_All , especially from the Tibetan Plateau to East Asia, over northern South America, and the dry regions of Australia and the Sahara. The projections from the full ensemble of the CMIP5 models and its subsets are similar over the land even when the models are selected by their performance on present-day precipitation (Fig. 7c, d, i, j). Although the precipitation change over the ocean is beyond the scope of this study, we note that the change over the tropical Pacific is different in the means of the CMIP5 RCP8.5_All and the subsets. HF All projects a precipitation change with an El Niño-like pattern on the ensemble mean, relative to the CMIP5 RCP8.5_All projections (Fig. 7h). The pattern, however, is not always detected in the comparison with the high-performing subset projections (Fig. 7i, j). The changes in P JJA and P DJF show a similar pattern over the tropical Pacific (Figs. S4 and S5). In order to identify features of the uncertainty in future precipitation changes from HF All , we explored how the uncertainty spread of HF All compares with that from the CMIP5 RCP8.5 models. Previous studies assessed the uncertainty from subsets of ensemble members using Fig. 7 a-d Spatial pattern of annual mean precipitation change in 2080-2099 with respect to 1984-2003 for the RCP8.5 scenario, scaled by the globally annual averaged temperature change in each model (% K -1 ). Full ensembles of a the HF members (HF All ) and b the CMIP5 AOGCMs (CMIP5 RCP8.5_All ), and the subsets of CMIP5 RCP8.5_All c with the 12 highest skill scores for global precipitation (CMIP5 GhighS ) and d with the 12 smallest MAE in global precipitation (CMIP5 GsmallE ). Crosshatched (hatched) areas in (a-d) indicate areas where changes are insignificant at the 33% (10%) level. e-g Precipitation change with HF All but for the grids with the opposite sign of the change between the ensemble means of HF All and of e CMIP5 RCP8.5_All , f CMIP5 GhighS , and g CMIP5 GsmallE (% K -1 ). h-j Difference of the ensemble mean of precipitation changes from HF All with respect to that from (h) CMIP5 RCP8.5_All , i CMIP5 GhighS , and j CMIP5 GsmallE (% K -1 )  Ito et al. 2020a). Following this methodology and expressing precipitation changes for ensemble E as ΔP(E), we computed the fraction of the maximumminimum range of ΔP(HF All ) to the range of ΔP(E′), and refer to this as a fraction of the relative range (FR).
Here, we used three ensembles as the ensemble E′: all 42 members of CMIP5 RCP8.5_All , and the 12 members of each of CMIP5 GhighS and CMIP5 GsmallE . The value of FR for HF All relative to CMIP5 RCP8.5_All , FR All , exceeds 0.7 around North America, southern Africa, East Asia, and Australia (Fig. 8a). Grid cells where FR All lies between 0.7 and 1.3 (i.e., the spread of ΔP ANN (HF All ) is within ± 30% of the spread of ΔP ANN (CMIP5 RCP8.5_All )), make up 11% of the global land grid cells, and grid cells with values between 0.5 and 1.5 make up 35%. As for CMIP5 GhighS and CMIP5 GsmallE , the percentage of the grid cells with FR within 0.7-1.3 (0.5-1.5) increases up to approximately 45% (75%) (Fig.  8b, c). Therefore, selecting the high-performance CMIP5 models to represent the present-day precipitation more than doubles the area where the uncertainty ranges of HF All and the selected CMIP5 models are close. The fraction from the uncertainty of CMIP5 GhighS , FR GhighS , exceeds 0.7 over most land areas. A similar spatial pattern is seen in the fraction from CMIP5 GsmallE , FR GsmallE . Of the 12 members in each of CMIP5 GhighS and CMIP5 GsmallE , three members of CMIP5 GhighS are not in CMIP5 GsmallE . Thus, the three members make their own patterns, e.g., Australia and South America around the tropical Pacific where a different pattern of ΔP ANN appears (Fig. 7i, j).
The tendency for a closer spread between HF All and the high-performing subsets of the CMIP5 models is detected in the seasonal precipitation change (Fig. S6). As with the annual precipitation change, for the seasonal precipitation change, the percentage of the grid cells with FR GhighS and FR GsmallE within 0.5-1.5 increases to approximately 70% of all land grid cells from 40% in FR All .

Uncertainty in the projected change of regional mean precipitation
The projected future change of regional mean P ANN over land is shown in Fig. 9. Here, we analyze the projections by the MMEs of CMIP5 GhighS and CMIP5 RhighS , together with the HF All ensemble and the CMIP5 RCP8.5_All MME. Additionally, considering the effects of model resolution, the future changes projected with SF are also considered, which were computed from the regional mean on the native grid of MRI-AGCM3.2S.
There is a wide spread in the projections ΔP(CMIP5 RC-P8.5_All ) in low-latitude regions and a narrower spread in the middle-and high-latitude regions, as shown in IPCC (2013) with respect to both land and sea areas. The uncertainty from ΔP(HF All ) of the annual and seasonal precipitation presents a similar picture (Fig. S7). This may be due to the difference in the main precipitation systems; i.e., the major systems are sub-grid scale (produced by cumulus convection schemes) at low latitudes and grid scale at middle and high latitudes.
The maximum-minimum uncertainty range from HF All partly overlaps all the range from CMIP5 RCP8.5_All , CMIP5 GhighS , and CMIP5 RhighS except for CMIP5 RhighS in CGI (2). The range from HF All is 0.4 of the range from CMIP5 RCP8.5_All on average over all 26 regions; the fraction increases to 0.8 (0.73) of the range from CMIP5-GhighS (CMIP5 RhighS ). In terms of the future change in the global average, specifically global land precipitation, the spread of the projections with HF All is equal to 0.74 of the spread from CMIP5 RCP8.5_All and to 0.95 of that from CMIP5 GhighS . Thus, the future change projected with HF All roughly has an uncertainty very close to the uncertainty in the projections with the high-performing CMIP5 models, as described above. In the individual areas, the projections with HF All have a wider spread than the range from CMIP5 GhighS and CMIP5 RhighS in SAF (17) and AMZ (7), and a comparable spread in NEB (8), WAS (19), EAS (22), SEA (24), and SAU (26). On the other hand, there is a relatively small spread of the HF All projections in NAS (18) and TIB (21). The projections with the SF ensemble lie within the uncertainty from HF All in most regions. Given that the SF ensemble is the ensemble of only four patterns of SST change, the spread is small but the individual regional pattern of spread resembles the spread from HF All .
Over the Tibetan Plateau and East Asia, where a significant increase is projected by the ensemble means of both HF All and CMIP5 RCP8.5_All (Fig. 7a, b), the rate of increase is 0.5-2% larger in HF All on average over TIB (21) and EAS (22). The rate in TIB is an additional 1% larger in the mean of the SF ensemble than that of HF All , which is a marked difference compared with other regions. The projected future change has the opposite sign of future change between the 50th percentiles of ΔP(HF All ) and ΔP(CMIP5 RCP8.5_All ) in six regions around the tropical South Pacific: Central and South America (6 −9) and Australia (25−26). In particular, in CAM (6) and NEB (8), more than three HF All members project an increase and different three members project a decrease so that the future change is insignificant. On the other hand, although there is a large uncertainty in the projections with not only CMIP5 RCP8.5_All but also CMIP5-GhighS and CMIP5 RhighS , more than 75% of HF All project an increase in AMZ (7), NAU (25), and SAU (26) and a Ito et al. Progress in Earth and Planetary Science (2020)   decrease in WSA (9). In particular, all members of HF All project the opposite sign to the 50th percentiles of ΔP(CMIP5 RCP8.5_All ) in WSA and NAU. In NAU, the 50th percentile of ΔP(HF All ) is 5.3% while that of ΔP(CMIP5 RCP8.5_All ) is − 0.7%. Of the six regions with opposite signs in different ensembles, the mean of HP All has a score > 0.8 in NEB (8) and Australia (Fig. 4) and a bias comparable to the discrepancies between observations in AMZ (7), NEB (8), and Australia (Fig. 5). This suggests that forcing with the observed SST reduces the SST bias in HP All compared with the CMIP5 coupled simulations, and consequently HP All performs well in representing the present-day precipitation. On the other hand, with regard to future simulations, the differences between the atmosphere-only model and the coupled model could cause the opposite sign in the future precipitation changes in the regions. A relatively large positive precipitation change in HF All is consistent with Kitoh and Arakawa (1999), who have demonstrated that the lack of SST and cloud-radiation feedback in AGCMs brings about an overestimate of precipitation in warm-SST regions.
For the global mean, HF All projects a wetter climate than CMIP5 RCP8.5_All , especially over the land; the projections with more than half the members of HF All exceed the maximum changes of CMIP5 RCP8.5_All and CMIP5 GhighS . DeAngelis et al. (2015) have indicated that CMIP5 models that underestimate the sensitivity of shortwave absorption to an increase in atmospheric water vapor tend to overestimate the precipitation increase. From their results, the CMIP5 model developed at MRI shows a lower sensitivity of shortwave absorption than the other CMIP5 models. The radiation scheme in that model is almost the same as the scheme used in MRI-AGCM3.2 for computing the sensitivity. Thus, the sensitivity could to some extent contribute to the greater increase of the precipitation by HF All than by CMIP5 RCP8.5_All . Similar to P ANN , the projections of HF All spread over 0.35 and 0.46 of the maximum-minimum uncertainty range from ΔP JJA (CMIP5 RCP8.5_All ) and ΔP DJF (CMIP5 RCP8.5_All ) on Fig. 9 Uncertainty estimates for future change projections of regional land mean precipitation in 2080-2099 with respect to 1984-2003 for the RCP8.5 scenario (% K -1 ). Change rate is scaled by the regional land annual temperature increment. Gray plots indicate CMIP5 RCP8.5_All . Yellow and green plots indicate CMIP5 GhighS and CMIP5 RhighS , respectively. Red plots indicate HF All . Blue plots and dark blue bar indicate SPYS and the ensemble mean, respectively Ito et al. Progress in Earth and Planetary Science (2020)  average over all 26 regions, respectively (Fig. S7). The fraction increases to 0.78 and 0.96 of the range from ΔP JJA (C-MIP5 GhighS ) and ΔP DJF (CMIP5 GhighS ), respectively.

Contribution of cumulus schemes and SST forcing to the uncertainty
The uncertainty from HF All is estimated from the future precipitation changes by 12 ensemble members corresponding to combinations of three different cumulus convection schemes and four SST change patterns. In order to discuss the sources of uncertainty in the HF All projections, we created two subsets, one with different convection schemes but the same C0 SST-change pattern (HF Conv ; three members) and the other with the same YS scheme but different SST change patterns (HF SST ; four members). The FR values were then estimated for the maximum-minimum uncertainty range from the projections with HF Conv and HF SST relative to the range with HF All (FR Conv and FR SST , respectively). FR Conv (FR SST ) denotes the ratio of HF Conv (HF SST ) covering the uncertainty from HF All . Figure 10 shows global distributions of FR Conv and FR SST . The area where either HF Conv or HF SST captures over 0.6 of the uncertainty from HF All is only approximately 40% of the global land, implying that the ensemble consisting of both multiple SST patterns and multiple convection schemes is needed to cover most of the uncertainty from the high-performing CMIP5 models. FR Conv is larger than FR SST in Africa and northern South America (Fig. 10a). On the other hand, a relatively large FR SST appears around Australia and Siberia (Fig. 10b). Around Asia, Endo et al. (2012) have assessed the contribution of each factor to the uncertainty by analysis of variance on the same MRI-AGCM3.2H dataset. Their results indicate that the uncertainty in ΔP ANN , as well as in ΔP JJA and ΔP DJF , is mainly caused by differences in the cumulus convection schemes in South and Southeast Asia, which is consistent with the relatively large FR Conv seen in Fig. 10a.
To investigate the atmospheric fields leading to the spreads of the projections of ΔP ANN , Fig. 11 shows the standard deviation in future change for three atmospheric circulation variables: sea level pressure (ΔSLP), temperature at 850 hPa (ΔT850), and geopotential height at 500 hPa (ΔZ500). Because of the common SST change pattern, the deviation in ΔT850 from HF Conv is globally small except for the western part of the tropical Pacific (Fig. 11b). On the other hand, HF Conv shows a relatively large deviation in ΔZ500 over the tropical and subtropical regions of 30°S−30°N (Fig. 11c). That is to say, the response of ΔZ500 over the regions depends on the cumulus convection scheme, which may relate to the difference in the precipitation changes between the different cumulus schemes. Therefore, FR Conv is large around Africa and northern South America over 30°S −30°N. The dependence of ΔZ500 on the schemes over 30°S−30°N is seen regardless of the patterns of SST change; the KF (AS) scheme gives a more positive (negative) anomaly in ΔZ500 than the YS scheme (Fig. S8a).
HF SST shows a roughly similar distribution of the deviation in ΔT850 to that in the SST change (Figs. 1a and 11e). Around Australia where FR SST is large, ΔSLP from HF SST shows a large deviation (Fig. 11d). The difference of ΔSLP between members with different SST change patterns responds to the prescribed patterns of El Niño/ La Niña-like SST change over the Pacific (Fig. S8b). The atmospheric response to the El Niño/Southern Oscillation (ENSO) expands over Australia and brings about a relatively large uncertainty in the precipitation changes with HF SST .
With reference to Mizuta et al. (2014), we divided CMIP5 models into three clusters to obtain future change patterns of SST and sea ice (C1-C3) for future climate simulations with MRI-AGCM3.2. The ΔSLP were then averaged over the models in each cluster and over all models to derive four spatial patterns of ΔSLP corresponding to the pattern from each member of HF SST . The spatial pattern of the deviation in ΔSLP from the clusters of the CMIP5 models (Fig. S9) resembles the pattern from the HF SST projections over the latitudinal band of 0°-30°S (Fig. 11d). That is, there is a similarity in the atmospheric response to the SST change between the ensembles of MRI-AGCM3.2 and the CMIP5 models, implying that the difference in the precipitation change around Australia between the ensembles arises from a quantitative difference in the response to a similar change of atmospheric circulation. The uncertainty could also be associated with differences in the convection schemes, especially in the northern regions of Australia at latitudes north of 30°S (Fig. 11c). These might lead to large uncertainty in NAU (25 in Fig. 9). A large FR SST around Siberia is partially affected by teleconnections such as the Pacific-North America patterns in response to the same ENSO-like change (Horel and Wallace 1981), as is also the case around North America and Asia (Fig. 11d). Urabe and Maeda (2014) and Imada et al. (2017) pointed out the seasonal variations in the appearance of the teleconnection pattern around Asia. Future study is required focusing on the effects of the seasonal change of the teleconnection pattern.

Enhanced model resolution
In order to examine the impact of the horizontal resolution of models on the simulations of P ANN , we summarized the relationship between the following four indices and the complexity in topography in the 26 regions: the extent of the change in skill score and model bias in the regional mean from HPYS to SPYS (Fig. 12a,  b), the difference in the ensemble mean of ΔP ANN (Fig.  12c), and the ratio of the maximum-minimum uncertainty range from ΔP ANN projected by SFYS to the range from HF SST (Fig. 12d). The indices were estimated on the native grid of each model. The future change is the projected change in the regional mean. The ensemble of SFYS and HF SST consists of the four members with the same YS scheme and different four patterns of SST change. The complexity in topography is defined as the standard deviation of the altitude represented by MRI-AGCM3.2S.
The impact of the enhanced horizontal resolution is a significant decrease in model bias. The relationship between the complexity and the decrease in the bias is statistically significant at the 99% confidence level. The difference in the ensemble means of future changes is positively correlated with the complexity, although the relationship is insignificant. This is also the case for the uncertainty range. Consequently, simulations by the high-resolution model may project a greater future precipitation increase around complex terrain regions, with a reduction of the model bias in the present-day precipitation. Furthermore, there is a possibility that the uncertainty in the future projections is greater in regions with more complex topography.
As seen in Fig. 4, the skill score is smaller across all regions in SPYS than in HPYS except for NEU (11) and NAU (25), regardless of the complexity. It seems that the high-resolution model cannot simulate the precipitation pattern better than the lowresolution one. However, Fig. 4 shows a large discrepancy in the score for the observations, with a value of 0.4 in TIB (21) and WSA (9), regions with high complexity in the topography. Thus, we cannot conclude that the high-resolution model performs poorly in representing the spatial pattern. Further validation is needed using observations with a resolution of a few tens of kilometers that is high enough to evaluate the simulations with MRI-AGCM3.2S.
Note that the difference in the simulations between various model resolutions derives from not only the represented topography but also the precipitation systems themselves that the model is able to resolve. According to Ito et al. (2020b), focusing on the areas around Japan that are included in EAS (22), the projection of the regional precipitation changes when considering the effect of typhoons that low-resolution models cannot represent.
The impact of the model resolution is more effective in the simulations of extreme indices; e.g., a seasonal maximum of 1-day precipitation total. Similar to the need for projections of average climate change, this is important for the impact assessment of climate change.

Summary
We have shown that the simulations by MRI-AGCM3.2 reproduce well the present-day precipitation climatology in regions around the globe and that the signs of future changes at the 50th percentile are identical to the CMIP5 multi-AOGCM ensemble in almost all regions. The projections of future precipitation change spread over roughly 0.4 of the uncertainty range from the full ensemble of the CMIP5 AOGCMs and, furthermore, 0.8 of the range from the high-performing CMIP5 AOGCMs.
The precipitation climatology simulated by MRI-AGCM3.2 is closer to observations at both global and regional scales in terms of the spatial pattern and amount. Compared with the means of the CMIP5 ensemble and the high-performing CMIP5 ensembles, the model bias is reduced by approximately 20% for regional averages from the Tibetan Plateau through East Asia and Australia. Overall, MRI-AGCM3.2 has a skill score of more than 0.8 in almost all the 26 global land regions and the model bias for at least one of the cumulus convection schemes is within the discrepancy between observations in about half of the regions.
When ranking the score and absolute error for the simulated land precipitation in individual regions, MRI-AGCM3.2 ranks in the top 15 of all 55 model runs including all 50 of the CMIP5 ensemble members and their mean (except for HPKF for the error in the annual precipitation). MRI-AGCM3.2 is ranked almost equal to or higher than the CMIP5 ensemble mean for the skill scores in the annual and the Fig. 12 Relationship to the complexity of regional topography for differences of the SP(SF) simulations from those by the HP(HF) simulations. a Skill score, b model bias, c future change of the regional mean, and d its uncertainty. R is the correlation coefficient. Asterisk denotes the 99% confidence level Ito et al. Progress in Earth and Planetary Science (2020)  seasonal mean and the biases in the seasonal mean. Therefore, MRI-AGCM3.2 can represent the spatial pattern and the regional mean of precipitation realistically in more regions than the CMIP5 ensemble members.
The spatial distribution of projected change in global precipitation of the ensemble means of MRI-AGCM3.2 and the CMIP5 models is similar. MRI-AGCM3.2 projects a greater increase of regional-mean precipitation than the CMIP5 ensemble mean, from 0.5 to 2% greater around the Tibetan Plateau and East Asia and even more in Australia where the model bias is small in MRI-AGCM3.2. This is also the case for the comparison with the high-performing CMIP5 ensemble.
The sign of future change for annual precipitation is the same at the 50th percentile in 20 out of 26 regions and the opposite in the other six regions around the tropical South Pacific; e.g., Australia and South America. In these six regions, the skill score for the present-day precipitation exceeds 0.8 in three regions and the bias is comparable to the difference between the observations in four regions. Thus, we expect that the use of the prescribed SST in the MRI-AGCM3.2 simulations results in a small bias in the present-day regional precipitation, while the difference between models with or without atmosphereocean coupling has an influence on the future projections. Although the relationship between the ability to reproduce the present-day climate and the ability to provide reliable future predictions is still unclear (e.g., Smith and Chandler 2010;He and Soden 2016), high performance in simulations of the present-day precipitation by MRI-AGCM3.2 is an important metric for assessing the accuracy of the model itself.
The uncertainty in the MRI-AGCM3.2 projections partly overlaps the maximum-minimum range of the uncertainty in the projections with the CMIP5 model ensemble in all regions. By considering the uncertainty range from MRI-AGCM3.2 as a fraction of the range from the CMIP5 models, the percentage of grid cells with the fraction between 0.7 and 1.3 (between 0.5 and 1.5) is approximately 10% (35%). The percentage increases to 45% (75%) when the comparison is made with respect to the CMIP5 models that perform well for precipitation simulations in the present-day, suggesting that the uncertainty is closer to that from the high-performing CMIP5 ensemble than that from the full ensemble. Also, regarding the uncertainty in future changes in the regional mean precipitation, the range from the MRI-AGCM3.2 projections increases from 0.4 of the range from all the CMIP5 models to around 0.8 of that from the high-performing CMIP5 models on average over individual regions. That is to say, the MRI-AGCM3.2 dataset provides the regional information on the projections with an uncertainty comparable to the uncertainty from the highperforming CMIP5 models around the globe, together with well-represented precipitation in the present-day climate.
The difference of projections by MRI-AGCM3.2 relative to the CMIP5 model ensemble may be caused partly by the difference between an atmospheric model and the atmosphere-ocean coupled models of the CMIP5 AOGCMs, as seen in regions around the tropical Pacific. MRI has developed a new model system that assimilates observed SST in such coupled models, which enables the evaluation of atmosphere-ocean coupled effects; e.g., SST and cloud-radiation feedback (Kitoh and Arakawa 1999). Comparisons of projections from such a model will facilitate investigation of the causes of the differences identified here.
The characteristics of the precipitation climatology summarized in this study are highly useful not only for studies of global and regional climatology using the MRI-AGCM3.2 dataset but also for decisions on adaptation strategies. As already discussed on the CORDEX phase 1 in Ito et al. (2020a), the use of a GCM dataset to discuss regional climate changes requires a fundamental evaluation of the projections by the GCM as the regional scale, as we have done here. This will be applied, for example, with HighResMIP and CORDEX phase 2.
Additional file 1: Table S1. CMIP5 models used in this study. (Retrieved from https://cmip.llnl.gov/cmip5/availability.html on May 23, 2019). Figure S1. As in Fig. 4 but for (a) JJA mean precipitation and (b) DJF mean precipitation. Figure S2. As in Fig. 5 but for (a) JJA mean precipitation and (b) DJF mean precipitation (%). Figure S3. As in Fig. 6 but for the AMIP-type simulations instead of the CMIP5 simulations. Black number corresponds to the number in the list of models at the bottom. Gray number denotes the rank in Fig. 6 only for the model of which the name coincides with that in Table S1. Figure S4. As in Fig. 7 but for JJA mean precipitation (% K -1 ). White in (a-d) and hatching in (e-j) indicate the grid cells where the models simulate no precipitation in the presentday climate. Figure S5. As in Fig. 7 but for DJF mean precipitation (% K -1 ). White in (a-d) and hatching in (e-j) indicate the grid cells where the models simulate no precipitation in the present-day climate. Figure S6. As in Fig. 8 but for (a) JJA mean precipitation and (b) DJF mean precipitation. Hatching indicates the grid cells where the models simulate no precipitation in the present-day climate. Figure S7. As in Fig. 9 but for (a) JJA mean precipitation and (b) DJF mean precipitation (% K -1 ). Figure  S8. Future change anomaly for each of 12 HF ensemble members from their mean. (a) Future change of geopotential height at 500 hPa and (b) that of sea level pressure. Members on each row are the member with the same SST-change pattern for the future climate simulation. Members on each column are the member with the same cumulus convection scheme. The anomaly is normalized by the annual variation of individual variables in the present-day climate. Crosshatching (Hatching) indicates the 99% (90%) confidence level. Figure S9. Standard deviation in future changes of sea level pressure using three clusters of the CMIP5 AOGCMs Ito et al. Progress in Earth and Planetary Science (2020)