The Dichotomy of Wet and Dry Trends Over India by Aerosol Indirect Effects in CMIP5 Models

Aerosol‐cloud interactions, also known as aerosol indirect effect (AIE), substantially impact rainfall frequency and intensity. Here, we analyze NEX‐GDDP, a multimodel ensemble of high‐resolution (0.25°) historical simulations and future projections statistically downscaled from 21 CMIP5 models, to quantify the importance of AIE on extreme climate indices, specifically consecutive dry days (CDD), consecutive wet days (CWD), and simple daily intensity index (SDII). The 21 NEX‐GDDP CMIP5 models are classified into models with reliable (REM) and unreliable (UREM) monsoon climate simulated over India based on their simulations of the climate indices. The REM group is further decomposed based on whether the models represent only the direct (REMADE) or the direct and indirect (REMALL) aerosol effects. Compared to REMADE, including all aerosol effects significantly improves the model skills in simulating the observed historical trends of all three climate indices over India. Specifically, AIE enhances dry days and reduces wet days in India in the historical period, consistent with the observed changes. However, by the middle and end of the 21st century, there is a relative decrease in dry days and an increase in wet days and precipitation intensity. Moreover, the REMALL simulated future CWD and CDD changes are mostly opposite to those in REMADE, indicating the substantial role of AIE in the future projection of dry and wet climates. These findings underscore the crucial role of AIE in future projections of the Indian hydroclimate and motivate efforts to accurately represent AIE in climate models.

By the end of the 21st century, anthropogenic aerosol concentrations are expected to decrease drastically under the representative concentration pathway (RCP) scenarios, leading to an increase in Asian monsoon precipitation (Levy Ii et al., 2013;Samset et al., 2018).
Besides their direct influence on solar radiation, aerosols can impact clouds and precipitation by serving as cloud condensation nuclei and inducing aerosol indirect effect (AIE). The "Twomey" effect (Twomey, 1977), usually observed in warm clouds, refers to the reduction of cloud droplet size and increase in droplet number concentration so for the same liquid water path, aerosols increase cloud reflectivity and suppress drizzle and increase the fractional cloud cover, leading to prolonged cloud lifetime and less rainfall (Albrecht, 1989). However, aerosol-cloud microphysical interactions may also invigorate convection during the Indian monsoon season (Gayatri et al., 2017;Leena et al., 2021;Sarangi et al., 2017Sarangi et al., , 2018. The net aerosol forcing reduces the wet-day frequency and promotes the dry-day frequency over a large part of India . A recent study attributed the negative precipitation trends in India to AIE (Saha & Ghosh, 2019). However, state-of-the-art coupled ocean-atmosphere models in the Coupled Model Intercomparison Project Phase 5 (CMIP5) still underestimate aerosols' effects on precipitation (Ekman, 2014). The significant uncertainties in climate projections associated with aerosol effects likely come from multiple sources, including uncertainties in the spatiotemporal and vertical distribution of various aerosol species, specieswise forcings, and poor or simplified parameterizations (Buchard et al., 2016;Lohmann & Feichter, 2005;Quinn & Bates, 2005;Sperber et al., 2013). For example, simple log-linear relationships between aerosol mass and cloud droplet number concentration and complex modules in which the number of activated aerosols depends on the simulated aerosol size distribution and composition and the updraft velocity have been used to represent aerosol activation in the CMIP5 models. The diverse treatments of aerosol effects in climate models present an opportunity to compare historical simulations and future projections from models that include only ADE and models that include both AIE and ADE to better understand the role of AIE at the climate scale. Grouping the CMIP5 models this way, Lin et al. (2018) showed the AIE effect on the intensity and trends of extreme precipitation over southeast Asia.
Here, we identify downscaled CMIP5 models that skillfully simulate the precipitation and its variability  over India. We isolate the AIE impact on the historical and future multidecadal trends of dry and wet indices during the Indian monsoon season by comparing two composites of model simulations with only ADE versus all aerosol effects (ADE + AIE). Using this framework, we address three questions: (a) Does including AIE improve the model skill of simulating multidecadal trends of wet and dry climates? (b) What are the implications of AIE on wet and dry indices in the historical run? (c) How does the inclusion of AIE affect the future projections of multidecadal trends in dry and wet climates over India? This paper is organized into four sections. Section 2 describes the data, methodology, and approach to selecting models that reliably simulate the historical climate. Section 3 presents the role of AIE on historical runs and future climate projections. Results and conclusion are summarized in Section 4.

Climate Model and Observation Data
The National Aeronautics and Space Administration (NASA) and Earth Exchange Global Daily Downscaled Projections (NEX-GDDP; Thrasher et al., 2012), NEX-GDDP data set of 21 models (Table S1 in Supporting Information S1) for the historical  and future  periods is available at https://nex.nasa. gov/nex/resources/365/. At 0.25° × 0.25° spatial resolution, the NEX-GDDP data is developed by the Terrestrial Hydrology Research Group at Princeton University using the Bias-Correction Spatial Disaggregation (BCSD) method applied to the CMIP5 simulations (Taylor, 2001). The BCSD method uses quantile mapping to bias correct the CMIP5 historical simulations to match the observational benchmark Global Meteorological Forcing Data set (GMFD; Sheffield et al., 2006), and the same mapping is applied to the future projections after removing the long-term trends (Thrasher et al., 2012;Wood et al., 2004). However, the GMFD-derived precipitation can exhibit regional biases over India that can impact the NEX-GDDP data, leading to uncertainty in the historical trends and future projections over India. Therefore, we first evaluate the downscaled data with observations from Indian Meteorological Department's (IMD) high-resolution (0.25° × 0.25°) gridded gauge precipitation data for the period 1975-2005 to quantify the biases. The IMD data has been interpolated from around 6,955 rain stations (count as of 2013) spread across India with varying availability periods (Pai et al., 2014).

Wet and Dry Climate Indices
We have selected three extreme climate indices, as described by the Expert Team for Climate Change Detection and Indices (ETCCDI; Karl & Easterling, 1999;Zhang et al., 2011) and computed the indices from the NEX-GDDP data set over the entire India region for the Indian summer monsoon period (June through September). The three indices feature dry and wet extremes: (a) consecutive dry days (CDD) defined as the number of consecutive dry days when precipitation is <1 mm, (b) consecutive wet days (CWD) defined as the number of consecutive wet days when precipitation is greater than 1 mm, and (c) simple daily intensity index (SDII; in mm/ day), the ratio of total precipitation to the number of wet days when precipitation is greater than or equal to 1 mm. CDD and CWD are good indicators of drought characteristics, which are important to the agricultural sector. SDII represents the average precipitation intensity occurring on a seasonal time scale.

Evaluation Metrics
Using the Taylor diagrams, we concisely assess the ability of the NEX-GDDP models to simulate the three climate extreme indices (CDD, CWD, and SDII) over the Indian subcontinent. Further, based on the Taylor diagram, the E-Skill score (Equation 1) evaluates the model's capability to simulate the spatial distribution (Schneider et al., 2007).
Here, STDEV OBS and STDEV MOD are the standard deviations of observation and model simulation. R is the pattern correlation coefficient between observed and modeled values, while R o is the theoretical mean correlation, which is considered equal to one for calculating the E-score. The k-value used to control the relative weight of the spatial correlations is set to 4 for the large-scale spatial pattern (Yang et al., 2013). Lower E values indicate closer agreement between the model simulation and observation. Based on the overall evaluation (described in detail in the next section), we classify models that better simulate the extreme indices as reliable models (hereafter REM) while other models are grouped as unreliable models (hereafter UREM). To validate the robustness of our classification method, we compare the spatial patterns of the ensemble mean, biases, and trends of each model group against the IMD observation data sets. Thiel-Sen's slope estimator estimates the trends (Sen, 1968). Also, a nonparametric Mann-Kendal test is used to test the significance at 95% confidence level at each grid point (Kendall, 1948;Mann, 1945).
Following Rotstayn et al. (2015), we further segregate the REM group (Table S2 in Supporting Information S1) into models that incorporate only aerosol direct effect (hereafter REM ADE ) and models that represent both aerosol indirect and direct effects (hereafter REM ALL ). The difference between the ensemble means of these two groups is used to illustrate the significance of AIE on the spatial patterns of wet and dry extremes in the current and future climates. To understand the changes in the historical period, we calculate the changes in the first decade (1975)(1976)(1977)(1978)(1979)(1980)(1981)(1982)(1983)(1984)(1985) and the last decade (1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005) with respect to climatology and determine the probability distribution of the indices based on the changes in each grid point over India. Similarly, we repeat the same analysis for the future period under the RCP4.5 scenario. For quantitative analysis, a nonparametric Kernel Density Estimator (KDEs; Parzen, 1962) is used to estimate the probability density functions (PDFs). Data smoothing techniques are applied to remove uncertainty associated with the climate system (Gramacki, 2018;Silverman, 2018).

Identifying Reliable and Unreliable Models
As shown in Figures 1a-1c, the Taylor diagrams illustrate a wide range of correlation coefficients of simulated CDD, CWD, and SDII by the 21 downscaled models against the IMD data over India. A correlation coefficient threshold >0.65 (for all three indices) is used to identify the REM group (marked in red in Figure 1), and the rest of the models are grouped as UREM (blue points in Figure 1). We found 12 models (∼57%) among the 21 models performing relatively better, with a pattern correlation of 0.65-0.9, which is higher than the other nine models. Similarly, a wide scatter of standard deviation of simulated CDD, CWD, and SDII by the downscaled models against the IMD data can be found. Quantitatively, the UREM group has a slightly higher variability of 0.48-1.85 compared to the REM group (0.6-1.3) for all three indices (Figures 1a-1c). To test whether downscaling has impacted the selection of models in the REM and UREM groups, the same analysis was repeated using model output directly from the CMIP5 models without downscaling. The CMIP5-based results showed a similar correlation coefficient threshold (CC > 0.55 for all models in the REM group; Figure S1 in Supporting Information S1).
The JJAS composite average of the REM group for all three indices shows more spatial resemblance with IMD than the UREM group ( Figure S2 in Supporting Information S1). The REM group depicts ranges from ∼6 to 20 days, 10-25 days, and 10-25 mm/day for CDD, CWD, and SDII, respectively, over central India, while the UREM shows 8-30 days, 10-40 days, and 8-18 mm/day for CDD, CWD, and SDII, respectively. Further, the REM group has an E-score <0.5 for all indices (Figures 1d-1f). Moreover, the domain averaged biases of most models in the REM group for CDD (−4 to 4 days), CWD (−4 to 4 days), and SDII (−4 to 4 mm/day) is significantly lower than the UREM group, indicating a relatively better representation of the observed spatial variability (Figures 1d-1f). The correlation coefficient, bias, and E-score for each model and index can be found in Table  S3 of Supporting Information S1.
The spatial distributions of the multimodel ensemble mean bias of the REM and UREM groups are compared in Figure 2. The biases of the REM composite are near zero over most of India, particularly over the central Indian monsoon region, compared to the UREM composite. The latter simulates a higher number of wet days (>16 days) over most of India ( Figure 2d). Moreover, the SDII of UREM shows dry biases throughout India, unlike the relatively low bias values of the REM group (Figures 2e and 2f). The contrasting biases in CWD and SDII of the UREM group suggest a higher frequency of lighter rainfall events over India, which is not the case observed. Thus, the lower biases by the REM group in all three indices substantiate the robustness of the segregation scheme and may allow us to make inferences about the current and future wet/dry events over India with more confidence. Notably, downscaling has significantly reduced the CMIP5 model biases over most of the regions in both the REM and UREM groups ( Figure S4 in Supporting Information S1).    Figure 3a) along with decreases in wet days (CWD; Figure 3d) over most regions. The SDII trends are rather spatially incoherent, unlike CDD and CWD. There is a statistically significant positive trend over parts of central, northern, and northeastern India, while negative trends are found over the west coast, southeastern coast, and parts of the Indo-Gangetic Plain (IGP) (Figure 3g). The REM group reasonably reproduces the overall observed trend pattern for all three indices over most of India (Figures 3b, 3e, and 3h). Quantitatively, the REM composite shows an overall positive trend of 0.4 days/31 years for CDD and negative trends of −0.5 days/31 years and −0.2 mm/days/31 years for CWD and SDII, respectively. As expected, the UREM group does not capture the spatial distribution and magnitude of the observed trends over most of India (Figures 3c, 3f, and 3i), further supporting our model classification scheme. The spatially averaged trends of the UREM group are 1.0 days/31 years, 0.6 days/31 years, and −0.7 mm/day/31 years for CDD, CWD, and SDII, respectively. Analysis of the CMIP5 spatial trends of the REM and UREM groups without downscaling shows similar trends to their downscaled counterpart ( Figure S5 in Supporting Information S1), further supporting the use of the downscaled data for further analysis. Overall, the REM group has higher fidelity in simulating the dry and wet indices (mean, spatial variability, and historical trends) over India, in stark contrast to the UREM group. Hence, hereafter, we only use the REM group of models to understand the impact of AIE on the Indian dry and wet climate.

Decomposing the Historical Trends into Contributions by Aerosol Direct and Indirect Effects
Within the REM group, the 12 models are equally divided between models that include both AIE and ADE (REM ALL ) and models that include only ADE (REM ADE ) ( Table S2 in Supporting Information S1). The spatial distribution of the composite mean values of the indices in the REM ALL group is similar to the REM ADE group ( Figure S2 in Supporting Information S1). However, a clear distinction in the spatial distribution of trends between the two groups is evident (Figures 4a-4f). The REM ALL group composite depicts a significant increasing dry-day trend over northern, central, and western India (Figure 4a). The spatial average increase of 0.9 dry day/31 years is much higher than the REM ADE group composite (Figures 4a and 4b) and closer to the observed trend of 0.7 dry days/31 years (Figure 3a). This difference is primarily due to a reduction in the dry days in the REM ADE group over parts of the east coast, central, and northwestern India, which is inconsistent with the REM ALL group. Further, CWD and SDII from REM ALL display negative trends over northern and central India (Figures 4c and 4e). Contrary to observation and the REM ALL group, REM ADE broadly indicates an increase in wet days and average precipitation intensity (Figures 4d and 4f).
The changes between the first (1975)(1976)(1977)(1978)(1979)(1980)(1981)(1982)(1983)(1984)(1985) and last decade (1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005) of CDD and CWD also depict an increase in dry days for both the REM ALL group and IMD (Figures 5a and 5b). In IMD, the first decade is a slightly flatter curve skewed to the right before shifting to positive change (peak at ∼0.2-0.3 days) in the last decade ( Figure 5a). The first decade peaks between −0.4 and −0.3 days, while the last decade gets a notable shift toward positive values, ∼1 day for the REM ALL group (Figure 5b). This decadal shift from negative to positive changes contradicts the CDD change (the first decade having a change of ∼1 to ∼−0.5 days during the last decade) of the REM ADE group (Figure 5c). The CWD drastically changes from ∼1 day in the first decade to ∼−2 days in the last decade for the IMD data set (Figure 5d). While this transition is coherently followed by the REM ALL group (Figure 5e), REM ADE tends to deviate in the opposite direction (Figure 5f). Further, both REM ADE and REM ALL could not capture the changes in SDII between the decades (not shown).
Hence, the inclusion of AIE in the models produces a distinct increase in the dry conditions and reduces the wet days during 1975-2005 compared to other ADE-only models. CMIP5-based REM ALL group also depicts similar drying of the Indian landmass compared to REM ADE ( Figure S6 in Supporting Information S1). Our results align with previous studies that found a reduction in precipitation in simulations that include all forcings, i.e., greenhouse gases (GHGs) and ADE + AIE (Sanap et al., 2015). We further investigated the changes in CMIP5-simulated cloud properties over the monsoon season using the base CMIP5 models. We plotted the composites of cloud occurrence frequency (on the axes of cloud layer thickness and cloud top height) for the two categories and found significant changes in the basic cloud structure in REM ALL compared to REM ADE ( Figure S7 in Supporting Information S1). The inclusion of AIEs in the models leads to an increase in overall cloudiness by enhancing the lifetime/prevalence of low and midlevel clouds. Using high-resolution simulations, Sarangi et al. (2018) have also reported similar aerosol-induced cloud structure changes and estimated a reduction in surface temperature by 1-2°C during the Indian monsoon. Such aerosol-mediated cloud radiative effect on surface temperature can weaken the monsoon circulation (Bollasina et al., 2011;Guo et al., 2015;Zelinka et al., 2014). Hence, an increase in cloudiness and associated surface cooling can explain the simulated general drying of the Indian landmass in the REM ALL models (as evidenced by an increase in the CDD and a decrease in the CWD and SDII). Lin et al. (2018) have also found a similar reduction in extreme rainfall due to AIEs over India and China. Note that the trend pattern simulated by the REM ALL group compares well with the observed IMD trends, indicating more realistic rainfall trends when the AIE is incorporated.

Future Projections of Wet and Dry Indices
To investigate the future projections of CDD, CWD, and SDII over India, we consider the change between mid-century (2045-2075) and the historical period  under the RCP4.5 scenario. As shown in Figures 6a and 6b, the spatial pattern of CDD by the REM ALL group depicts a distinct decreasing-increasing-decreasing pattern from south to north while CDD is projected to increase over a large part of the Indian landmass in REM ADE , except for parts of the west coast and the Ladakh region of India. There is a high likelihood that CDD will be reduced by ∼1 day when models incorporate all aerosol effects, while CDD may increase by ∼0.5 days when models only account for ADE (Figure 6c). The spatial distribution of CWD by REM ALL is like a photographic negative (increase-decrease-increase) of CDD (Figures 6d and 6a). There is a statistically significant increase in wet days by REM ALL over the west coast and parts of the north and east coast of India during the mid-century (Figure 6d), whereas there is a significant decrease in wet days over most of India in REM ADE (Figure 6e). As a result, a distinct sign change in CWD from ∼−1.8 days to ∼+1 day by REM ADE and REM ALL is evident (Figure 6f). The precipitation intensity (SDII) shows an increase of ∼1.5 and ∼1 mm/ day from the REM ALL and REM ADE groups throughout India, respectively (Figures 6g and 6h). These changes are also evident in the PDF (Figure 6i). Investigating the change in all three indices by the end of the century (2070-2100) also revealed a similar contrast in the spatial distribution of changes between REM ADE and REM ALL ( Figure S8 in Supporting Information S1). Furthermore, CDD projected by REM ALL (REM ADE ) shows a net shift from negative to positive (positive to negative) change by the end of the century ( Figure S8c in Supporting Information S1) compared to the mid-century (Figure 4c). Apart from this, CWD ( Figures S3d-S3f in Supporting Information S1) and SDII ( Figures S3g-S3i in Supporting Information S1) increase significantly by the end of the century compared to the mid-century for both model groups.
The overall picture that emerges from the future projections of CDD, CWD, and SDII over India is that by the mid-21st century, the "reliable models" that incorporated aerosol-cloud interactions depict a reduction in dry days by ∼1 day, and simultaneously an increase in wet days by ∼1 day as well as the rainfall intensity by ∼1 mm/day. This is also evident from the time series of multimodel changes relative to the historical period (Figure 7). CDD consistently shows negative changes with a few intermittent peaks reaching zero till ∼2070 (Figure 7a; black line and black box plot). In contrast, CWD shows positive changes consistently (Figure 7b; black line and black box plot). The SDII changes are positive right from the beginning of the future simulation. Average intensity increases from ∼0.45 mm/day in 2005 to ∼1.3 mm/day by 2075 (Figure 7c; black line and black box plot). Further by the end of the century, dry days show large variability, but the changes are close to zero (Figure 7a), while wet days remain almost the same (change of ∼+1 day; Figure 7b). Regardless of the frequency of extreme events, the intensity of rainfall events increases by 2 mm/day (Figure 7c). Interestingly, the changes of CDD and CWD are opposite between REM ALL and REM ADE during both periods of analysis, suggesting the importance of including the AIE on the frequency of extreme indices. From observations and the historical runs of REM ALL , the relative drying up (increase in CDD and decreasing SDII and CWD) could be due to the precipitation suppression by aerosols as suggested by earlier studies. By mid-century, the reduction in aerosol emissions could revert the changes in general, hence a decrease in dry days and an increase in wet days and precipitation intensity. Furthermore, as the atmosphere becomes cleaner by the end of the 21st century, the dominant role of the GHG forcing over aerosol will be apparent in the frequency and intensity of climate extremes (frequent and stronger). Our result supports previous studies that a stringent reduction of aerosol emissions combined with the dominant forcing of GHGs in the future may intensify the mean and extreme precipitation (Allan Richard & Soden Brian, 2008;Min et al., 2011;Samset et al., 2018;Smith & Bond, 2014;Zhao et al., 2019). Figure 6. The multimodel ensemble mean changes of (a and b) consecutive dry days (CDD), (g and h) consecutive wet days (CWD), and (m and n) simple daily intensity index (SDII) in the mid-century (2045-2075) relative to  from REM ALL and REM ADE . Stippling indicates >95% statistical significance. The corresponding probability density function (PDF) of the multimodel ensemble mean changes of the indices from REM ALL (black) and REM ADE (red). The shaded portions indicate confidence limits from 5% to 95%, determined by Silverman's rule.  . The shaded portion depicts the standard deviation. The box whisker plots present the mean, maximum, and minimum change of the indices at one standard deviation.

Summary and Conclusion
This study contributes to the possible selection of best model groups from a multimodel ensemble for projecting extreme events over the Indian subcontinent. Using statistical metrics, we evaluate and identify a subset of 21 NEX-GDDP downscaled models that skillfully simulate a few extreme climate indices (CDD, CWD, and SDII).
The key findings are summarized as follows: 1. Using a single statistical metric of the pattern correlation between the observed and simulated extreme indices, CMIP models can be classified into "reliable" (REM) and "unreliable" (UREM) models over India. The climatology and spatial distribution of historical trends over India between 1975 and 2005 from the REM agree with the observations and display relatively smaller biases than the UREM composites. 2. Further classifying the "reliable models" into models that incorporate both AIE and ADE (REM ALL ) and models that include only ADE (REM ADE ), we find that models in the REM ALL group can reproduce the spatial pattern of the observed historical trends in the extreme indices better than models in the REM ADE group. The "all effects" models substantiated a net increasing CDD and decreasing CWD and SDII seen in the IMD observations. On the contrary, the ADE-only models simulated a substantial increase in all three indices from 1975 to 2005. 3. There are clear evidence that AIE plays a discernible role in India's wet and dry extremes. Models that include AIE simulate spatial patterns of trends in the extreme indices that are exactly opposite to the patterns simulated by models with only ADE throughout the Indian landmass. 4. Future projection of the indices depicts a different story compared to the historical changes. The REM ALL group projected a net reduction in CDD and amplification of CWD and SDII by the mid-21st century, highlighting a general wetting trend in the future.
A promotion of dry days and reduction in wet days and average precipitation intensity by the REM ALL models for the historical period suggest a critical role of aerosol-induced precipitation suppression. The stringent measures to reduce aerosol emissions by the mid-century will could cause a shift in precipitation from more drizzle events to more heavy precipitation events and wetter days. However, this process is not spatially uniform due to the large spatial variability of aerosol concentration and meteorology and the dependence of AIE and ADE on aerosol species, among other factors, so further analysis is required to establish the robustness of the changes in the trends of extremes. Further, probably because of the relatively weaker forcing by aerosols compared to GHGs by the end of the 21st century, all the indices are projected to intensify, indicating more frequent and intense drought events as well as prolonged heavy precipitation events, a general signature of the GHG effect. Further analysis is required to assess models' robustness in the future projection of the Indian summer monsoon.
The inherent biases in the observational data set (IMD) arising from the interpolation technique used and varying gauge numbers each year have potential implications on the model evaluation metrics used in this study. This remains a major limitation of the study. Interestingly, even after downscaling, model biases over the northeast region remain. Unlike other regions, rainfall over northeastern India is strongly modulated by the Pacific Decadal Oscillation (PDO) (Choudhury et al., 2019). Moreover, due to topography, there limited active rain gauges to provide robust estimates of observed rainfall (Pai et al., 2014). Hence, regional biases may limit interpretations of aerosol effects in this region. Notwithstanding these limitations, the results presented in this study highlight the importance of the complex processes involved in aerosol-cloud interactions and their feedback in modeling the climate extremes over India. Our results also substantiate the thwarting role of aerosols on the future projections of rainfall events and their intensity.