Robust increase of Indian monsoon rainfall and its variability under future warming in CMIP-6 models

The Indian summer monsoon is an integral part of the global climate system. As its seasonal rainfall plays a crucial role in India’s agriculture and shapes many other aspects of life, it affects the livelihood of a fifth of the world’s population. It is therefore highly relevant to assess its change under potential future climate change. Global climate models within the Coupled Model Intercomparison Project Phase 5 (CMIP-5) indicated a consistent increase in monsoon rainfall and its variability under global warming. Since the range of the results of CMIP-5 was still large and the confidence in the models was limited due 5 to partly poor representation of observed rainfall, the updates within the latest generation of climate models in CMIP-6 are of interest. Here, we analyse 32 models of the latest CMIP-6 exercise with regard to their annual mean monsoon rainfall and its variability. All of these models show a substantial increase in June-to-September (JJAS) mean rainfall under unabated climate change (SSP5-8.5) and most do also for the other three Shared Socioeconomic Pathways analyzed (SSP1-2.6, SSP24.5, SSP3-7.0). Moreover, the simulation ensemble indicates a linear dependence of rainfall on global mean temperature with 10 high agreement between the models and independent of the SSP; the multi-model mean for JJAS projects an increase of 0.33 mm/day and 5.3% per degree of global warming. This is significantly higher than in the CMIP-5 projections. Most models project that the increase will contribute to the precipitation especially in the Himalaya region and to the northeast of the Bay of Bengal, as well as the west coast of India. Interannual variability is found to be increasing in the higher-warming scenarios by almost all models. The CMIP-6 simulations largely confirm the findings from CMIP-5 models, but show an increased 15 robustness across models with reduced uncertainties and updated magnitudes towards a stronger increase in monsoon rainfall.

select the land area with longitude 67.5°0'0"E -98°0'0"E and latitude 6°0'0"N-36°0'0"N, comprising India and neighboring regions. The land area is obtained by using land-sea-masks for each model that are based on the percentage of the grid cells occupied by land (see Fig. 3 for each model). The resolution strongly differs between the models ranging over land from about 100 km to 500 km (See Table 2). Mean rainfall is obtained by averaging the monthly rainfall data from June-September over the region of interest. The WFDE5 dataset of precipitation over land (Cucchi et al., 2020) is used to evaluate the models capacity 130 of representing the Indian monsoon. This dataset has been generated on the basis of ERA5 reanalysis data and has undergone bias-adjustment methods following Weedon et al. (2010Weedon et al. ( , 2011. It is provided at 0.5°spatial resolution. For calculating the change in interannual variability, we apply the Singular Spectrum Analysis Method (Golyandina and Zhigljavsky, 2013) with a window size of 20 years to extract the nonlinear trend.

Model comparison
To evaluate the models' quantitative capacities of capturing the Indian monsoon rainfall, we compare their projected seasonal mean rainfall with WFDE5 reanalysis data over land (Cucchi et al., 2020) for two 30-year-periods in the past . We choose these periods to obtain a model evaluation for a historic period as well as for a period close to present.
The seasonal mean rainfall from the reanalysis data is 6.1 mm/day with a standard deviation of 0.5 mm/day for 1900-1930 Several models underestimate the seasonal mean rainfall, especially the models of the Canadian Centre for Climate Modeling 145 and Analysis (CanESM5-CanOE, CanESM5) which capture just about half of the reanalysis rainfall amount. All models that underestimate the rainfall for 1900-1930 show rainfall means below the lower threshold in 1985-2015, too. GFDL-CM4 for 1900-1930 and GISS-E2-1-G for 1985-2015 capture the seasonal rainfall quantitatively best. The other models that are closest to the reanalysis mean overlap for both periods, e.g. CNRM-CM6-1, NorESM2-MM and FGOALS-f3-L. For the two chosen time periods, models that capture, over-or underestimate the mean rainfall within twice the standard deviation mostly have 150 the same tendency for both periods. The multi-model mean for 1900-1930 is 5.6 ± 1.1 mm/day and 5.7 ± 1.1 mm/day for 1985-2015.
In order to identify models with a potentially realistic representation of the Indian monsoon rainfall, we also analyze the spatial precipitation distribution for 1985-2015. We choose this period since it is closer to present time and therefore closer to the simulated time period in the future. As reference data set we use WFDE5 reanalysis data again. The distribution is 155 dominated by rainfall over the Western Ghats, the Himalaya region, the west coast of the Bay of Bengal, the northeast of India and the north of Myanmar partly even exceeding 20 mm/day averaged over JJAS and the 30 year period . The east of central India reaches rainfall values above 10 mm/day (Fig. 2). The spatial rainfall pattern for the CMIP-6 models in 1985-2015 is shown in Fig.3. Models that captured the rainfall quantitatively well mostly simulate a spatial pattern close to the reference distribution e.g. NorESM2-MM, CNRM-CM6-1, FGOALS-f3-L. FIO-ESM-2-0 overestimates the rainfall in the Himalaya 160 region. The models with the tendency to underestimate the rainfall as ACCESS-CM2, CanESM5-CanOE, CanESM5 mostly are not able to capture the spatial pattern. Especially the southwest coast of India and the Himalaya region are not reproduced according to the reanalysis data by most of these models. An exemption for the models with low rainfall values are the models of EC-Earth-Consortium (EC-Earth3, EC-Earth3-Veg) which simulate a pattern very close to the reference distribution.
Presenting and discussing the results of this study, we decided to focus on the models within mean plus/minus twice the 165 standard deviation which also deliver a reasonable spatial rainfall pattern. Nevertheless, we will provide information for all 32 models.

Trend in Indian summer monsoon mean rainfall for the end of the 21st century
In order to determine the long-term trend in Indian monsoon rainfall, we first analyze the temporal time series between 1850-2100 for all models under SSP5-8.5 (Fig. 4). All available models show a clear positive long-term trend. The models exceed the  To analyze the change in mean seasonal rainfall until the end of the 21st century, we calculate the difference between the periods 2070-2100 and 1985-2015 for the four SSPs. In the stronger forced scenarios (SSP3-7.0 and SSP5-8.5), all models 175 project an increase of precipitation. In the scenarios with less forcing (SSP1-2.6 and SSP2-4.5), the clear majority of models project an increasing trend, too. The only models to project a decrease are the models of the National Center for Atmospheric Research (CESM2-WACCM in SSP1-2.5 and SSP2-4.5 and CESM2 in SSP2-4.5). On average over all models an increase of 24,3% is projected under SSP5-8.5 (Fig. 6) and of +18,6% in SSP3-7.0 (Appendix Fig. A1), of +11,9% in SSP2-4.5 (Appendix Fig. B1) and of +9.7% in SSP1-2.6 (Appendix Fig. C1). CanESM5 and CanESM5-CanOE show the maximum relative increase 180 in all scenarios by the end of the 21st century. But as shown in Fig. 1 and Fig. 3, they clearly underestimate the rainfall and do not capture a realistic pattern of the rainfall distribution. CESM2-WACCM shows the minimal increase of 7.8% under SSP5-8.5. This model was able to capture the mean rainfall in 1985-2015 within twice the standard deviation and is able to capture a reasonable pattern of the rainfall. Focusing on the models that captured the mean rainfall in 1985-2015 within twice the standard deviation (upper panel in Fig. 6), the relative increase is 17.4% under SSP5-8.5, i.e. slightly less than the average 185 over all models. Also in the other scenarios the trend is less for these models compared to the average over all models. In summary, a robust increase of seasonal rainfall between 1985-2015 and 2070-2100 can be derived under global warming.  Most models project that this increase will contribute to the precipitation especially in the Himalaya region and to the northeast of the Bay of Bengal, as well as the Western Ghats (Fig. 7). Individual models indicate decreasing rainfall along the southwest coast of India and around Myanmar. 190 Furthermore, we analyzed the dependence of rainfall on global mean temperature (GMT, Fig. 8). The simulation ensemble indicates a linear dependence of rainfall on GMT, with a high agreement between models and independent of the scenarios. The multi-model mean indicates an increase of 0.33 mm/day ranging from 0.11 mm/day to 0.54 mm/day. The relative dependence is 5.3% per degree of global warming ranging from 1.7%/K to 13.4%/K for SSP5-8.5 across models. Considering only the more realistic models, the projected mean change is 6.1%/K for SSP5-8.5.    outside of the range in CMIP-6 partly underestimated (13 models) and party overestimated (3 models) the observed mean in 225 1985-2015. Modeling centers whose models underestimated the rainfall within two standard deviations in our study mostly underestimated the rainfall already in CMIP-5. Some models with realistic patterns in CMIP-6 are updates from CMIP-5 that already revealed a pattern relatively similar to reanalysis data, e.g. NorESM2-MM. As in CMIP-5, models with the tendency to underestimate the rainfall in the evaluation period are mostly not capable either of capturing the spatial rainfall pattern in CMIP-6. But there are also various models that improved their capacity in capturing the Indian monsoon, such as the models    Table 1. Grey dashed lines indicate the slope (the hydrological sensitivity) for SSP5-8.5 some of them are based on overlapping model components, the models cannot be regarded as independent from each other.
The results have to be interpreted against this background. is not directly comparable to ours. An intensification of the Indian monsoon rainfall also has been found in other studies using CMIP-5 (Lee and Wang, 2014;Mei et al., 2015;Sharmila et al., 2015;Varghese et al., 2020). There is a widespread agreement that a reason for the intensification of the South Asian monsoon rainfall is an increase in moisture convergence (Singh et al., 2019). This enhanced thermodynamic effect dominates over the dynamic effect which refers to the decreasing 250 monsoon circulation. In this way, the thermodynamic effect determines the positive sign of the change in monsoon rainfall whereas the dynamic effect would lead to a decrease in rainfall (Sooraj et al., 2015).
We found that the monsoon rainfall is linear dependent on the GMT. The projected increase in rainfall is 0.33 mm/day per degree of global warming. The agreement between models and the independence of the scenario is remarkable. The median dependence of relative change in precipitation on GMT taking into account all models has increased from 3.2%/K in CMIP-255 5 to 5.3%/K in CMIP-6. Considering only the models with a more realistic representation of the monsoon, the increase is even more noticeably from 2.3%/K in CMIP-5 to 6.1%/K in CMIP-6. It also has to be mentioned that the range of projected CMIP-6 still have a tendency to underestimate the amount of precipitation in this period. This was already observed in  where all of the models out of the range underestimated the rainfall. In our analysis, we focused on the models with the more realistic representation of the Indian monsoon. We found that all models show an increase in mean summer monsoon rainfall under SSP5-8.5 and SSP3-7.0 by the end of the 21st century. An increase also was found in SSP2-4.5 and SSP1-2.6 by all 290 models apart from two models in SSP2-4.5 and one model in SSP1-2.6. Under SSP5-8.5, the models exceed the envelope of the baselines variability on average in 2045 in . An multi-model mean increase of rainfall of 24,3% is projected under SSP5-8.5 and of +18,6% in SSP3-7.0, of +11,9% in SSP2-4.5 and of +9.7% in SSP1-2.6. The majority of models project that that the increase will contribute to the precipitation especially in the Himalaya region, the northeast of the Bay of Bengal and to the west coast of India. Besides, the simulation ensemble indicates a linear dependence of rainfall on global mean temperature 295 independent of the SSP; the multi-model mean for JJAS projects an increase of 0.33mm/day and 5.3% per degree of global warming. Furthermore, under SSP5-8.5 a majority of 28 out of 32 models project an increase in interannual variability by the end of the 21st century after removing the trend with Singular Spectrum Analysis.
We have seen in this study that low resolution models did not capture the spatial pattern of the monsoon rainfall in historic periods well. Small scale topography and its atmosphere feedback influence the rainfall to a relevant extent. Thus, the ongoing 300 effort to improve the resolution of the individual CMIP models should be continued. Since other rainfall features such as extremes and the variability of rainfall on a subseasonal scale are beyond the scope of this study, they need to be analyzed in further studies owing to their high relevance e.g. for high-risk flooding events.
The projected increase in summer monsoon rainfall in combination with the projected longterm increase in interannual variability will be accompanied by an increased number of extremely wet years and potentially more high rainfall events 305 (Turner and Slingo, 2009;Sharmila et al., 2015). While crops need water especially in the initial growing period, high rainfall events during other growing states can harm the plants Revadekar and Preethi (2012). Thus, the projected development might have serious consequences for the agriculture in India and neighbouring regions. Since the change differs from the decreasing tendency in the second half of the 20th century, the development of adaptation strategies for the 21st century is required.
Appendix A: Appendix A: Change in Indian mean summer monsoon rainfall Table 1. Overview of data availability for the 32 models used in the study (precipitation/temperature). Only those models are selected for which data for historic period and SSP5-8.5 was available at the time of the study.