Evaluation of the CMIP6 marine subtropical stratocumulus cloud albedo and its controlling factors

Abstract. The cloud albedo in the marine subtropical stratocumulus
regions plays a key role in regulating the regional energy budget. Based on
12 years of monthly data from multiple satellite datasets, the long-term,
monthly and seasonal cycle of averaged cloud albedo in five stratocumulus
regions were investigated to intercompare the atmosphere-only simulations
between phases 5 and 6 of the Coupled Model Intercomparison Project (AMIP5
and AMIP6). Statistical results showed that the long-term regressed cloud
albedos were underestimated in most AMIP6 models compared with the
satellite-driven cloud albedos, and the AMIP6 models produced a similar
spread as AMIP5 over all regions. The monthly averaged values and seasonal
cycle of cloud albedo of AMIP6 ensemble mean showed a better correlation
with the satellite-driven observations than that of the AMIP5 ensemble mean.
However, the AMIP6 model still failed to reproduce the values and amplitude
in some regions. By employing the Modern-Era Retrospective Analysis for
Research and Applications Version 2 (MERRA-2) data, this study estimated the relative
contributions of different aerosols and meteorological factors on the
long-term variation of marine stratocumulus cloud albedo under different
cloud liquid water path (LWP) conditions. The multiple regression models can
explain ∼ 65 % of the changes in the cloud albedo. Under
the monthly mean LWP ≤ 65 g m−2, dust and black carbon dominantly
contributed to the changes in the cloud albedo, while dust and sulfur
dioxide aerosol contributed the most under the condition of 65 g m−2 < LWP ≤ 120 g m−2. These results suggest that the
parameterization of cloud–aerosol interactions is crucial for accurately
simulating the cloud albedo in climate models.



CERES and MODIS
Estimating the cloud albedo is required multiple atmospheric variables such as the top-of-atmosphere (TOA) downward, upward (all-sky) shortwave fluxes, cloud liquid water path (LWP) and cloud fractions. In this study, the TOA downward and upward shortwave fluxes were obtained from the Clouds and the Earth's Radiant Energy System (CERES, Wielicki et al., 100 1996) Single Scanner Footprint (SSF) monthly Ed4A datasets. The LWP and cloud fractions were also obtained from the Moderate Resolution Imaging Spectro-Radiometer (MODIS;Platnick et al., 2003) collection 6.1 level 3 monthly products during the period from 2003 to 2014, i.e., MODIS MYD08_M3 (Aqua) and MOD08_M3 (Terra) products, respectively. The spatial resolutions of these products are 1.0°× 1.0°. The CERES TOA shortwave fluxes were converted from broadband (0.2-5.0μm) radiances by applying empirical angular distribution models to correct the instrument's incomplete spectral 105 response (Loeb et al., 2001). Then, the real-time fluxes were aggregated to produce 24-hour mean fluxes from empirical diurnal albedo models that create meteorology conditions at the over-flight time (Loeb et al., 2018). In addition, the cloud fraction, a fraction of MODIS cloudy pixels to the total pixels at each grid box, is determined based on daytime scenes entirely and represents all cloud phases (Platnick et al., 2003). As the CERES and MODIS instruments are both carried onboard Aqua (cross the equator time: 1130) and Terra (pass the equator time: 1030) satellites in polar orbits, we averaged 110 the Aqua and Terra products to obtain the observed combined all-sky albedo, cloud fraction, LWP and cloud albedo as in the works of the Engström et al. (2015) and Bender et al. (2017).

115
This study identified ten climate models that provide both CMIP5 and CMIP6 outputs and implemented the intercomparison of performance for the regressed cloud albedo during the period from 2003 to 2008. Furthermore, this study evaluated the cloud albedo for twenty-eight CMIP6/AMIP outputs during the period from 2003 to 2014. Tables 1-2 show the characteristics of CMIP5 and CMIP6 models. Note that there is a considerable discrepancy in the total cloud fractions between the CMIP models and MODIS observations, which is caused by different definitions, cloud overlap algorithms and 120 different threshold assumptions for cloud formation (Engström et al., 2015). Moreover, the total cloud fractions in the climate models are usually calculated based on daytime and nighttime cloud fractions while only daytime cloud fractions are considered to evaluate the observed cloud fractions. As used in Engström et al. (2015), this study also employed the total cloud fractions as there are no available MODIS simulator outputs for CMIP6.
Goddard Earth Observing System Model, version 5.12.4 (Gelaro et al., 2017). The aerosol reanalysis has been produced by a global data assimilation system that combines satellite-and ground-based observed aerosols with meteorological conditions.
Here, the mass mixing ratios of different aerosol types and air density at different levels from 3-hourly aerosol product 130 (inst3_3d_aer_Nv) and meteorological data from monthly atmosphere product (instM_3d_asm_Np) are collected to represent the monthly regional aerosol and meteorological conditions. The outputs of MERRA-2 reanalysis were used during the common period from 2003 to 2014 with satellite observations record. As selected in McCoy et al. (2017) and Li et al. (2018), the impacts of different aerosol types on marine stratocumulus cloud albedo were evaluated based on the mass concentrations of hydrophilic black carbon (BC), hydrophilic organic carbon (OC), sulfate aerosol (SO4), sulfur dioxide 135 (SO2), the smallest particles dust (DU, i.e., 0.1-1 μm size) and sea salt (SS, 0.03-0.1 μm size) at the 910hPa level.
Furthermore, this study employed the relative humidity at 850hPa (RH850) and vertical velocity at 900hPa (omega900) factors to investigate the meteorological effects on the cloud albedo.

Methods
The planetary albedo (α) can be calculated mainly from the cloud fraction f (Bender et al., 2011) as expressed in Eq. (1): where, αcloud and αclear denote the albedo under cloudy-sky and clear-sky conditions, respectively. For a given region where the cloud and surface type are homogeneous (i.e., constant αcloud and αclear), namely, a change in α should be driven by a change in the cloud fraction f. The cloud albedo can be estimated by the derivative of Eq. (1) as in Eq. (2):
The selection of variables is a crucial step to build a multiple linear regression model for the monthly cloud albedo as a 155 function of meteorological factors and aerosol types under two different LWP scenarios (LWP < 60 g m -2 and 60 g m -2 < LWP < 120 g m -2 ). This study selected suitable variables based on correlation analysis. If the correlation between the cloud albedo and a candidate is significant at a 90% confidence level, the variable was considered as a predictor factor. where a and b are regression coefficients, c is a constant term, Mi represents the ith meteorological predictor, I is the number of meteorological predictor variables, Aj is the jth aerosols predictor, and J is the number of aerosol predictor variables.
The relative contributions of each predictor to the change in the cloud albedo (Huang and Yi, 1991) were evaluated using Eq. (4): where m is the number of the monthly samples, a is the number of predictors, and Tij is the product of the regression coefficients of each term (bj) and predictor variables (xij).
After removing the effect of meteorological factors, we further investigated the pure relationship between aerosols and the cloud albedo using the partial correlations between αcloud and log 10 A, as expressed in Eq. (5): 170 r αcloudlog 10 A⋅M = r αcloud ⋅ log 10 A -r αcloud⋅M r log 10 A⋅M 1-r αcloud⋅M 2 1-r log 10 A⋅M 2 (5) where r αcloud⋅ log 10 A , r αcloud⋅M and r log 10 A⋅M is the total correlation between each variable pair and r αcloudlog 10 A⋅M is the correlation between αcloud and log 10 A which eliminates the effects of meteorological factors M. More details on the partial correlation are described in Jiang et al. (2018) and Engström and Ekman (2010). further indicates the homogeneity of cloud and surface types over these regions. The regressed cloud albedo from the satellite ranged from 0.30 to 0.42 for the five stratocumulus regions, which is consistent with previous studies (Bender et al., 2011;Engström et al., 2014). As the values averaged over Aqua and Terra albedos and cloud fractions were used as the 185 observation in this study, the regressed cloud albedo values need to be within the range of the Aqua and Terra (Engström et al., 2014). Regarding the AMIP5 and AMIP6 models, a higher correlation (> 0.8) appeared for most models at the five regions, especially higher at the Australian and Canarian regions. At the Peruvian, Namibian and Californian regions, the https://doi.org/10.5194/acp-2020-1245 Preprint. Discussion started: 11 January 2021 c Author(s) 2021. CC BY 4.0 License.
correlations of the observation were relatively higher than those of most climate models while the observed correlation was approximately close to the median value of model simulations at the Australian and Canarian regions.

190
The correlation coefficients of the AMIP6 models between planetary albedo and cloud fraction showed a lower value than those of the AMIP5, indicating that the linear relationship between cloud fraction and planetary albedo in the AMIP6 models' simulations is not superior to that of AMIP5. While the AMIP6 simulations displayed a similar spread in the estimated cloud albedo for all regions, some AMIP6 models produced a lower correlation coefficient than those of the Over the Canarian regions, the correlation and cloud albedo of AMIP6-MEM showed good agreement with those of the satellite observation compared with those of the individual AMIP6 models, resulting from the offsetting effect between models.
The ACCESS-ESM1-5 and CAMS-CSM1-0 models produce a higher correlation coefficient and more accurate cloud albedo 215 at the Peruvian region. The MPI-ESM1-2-HR, NorESM2-LM, CESM2-FV2 and IPSL-CM6A-LR models simulated the correlation and the cloud albedo comparable to the observations at the Namibian, Californian, Australian and Canarian regions, respectively. In contrast, the EC-Earth3, FGOALS-f3-L, FGOALS-f3-L, INM-CM5-0 and CESM2 models simulated the worst results at the Peruvian, Namibian, Californian, Australian and Canarian regions, respectively. Overall, the AMIP6 models reproduced the cloud albedo and correlation well at the Australian region while a higher uncertainty in 220 model's simulations, i.e., a larger intermodal spread, at the Peruvian region (Engström et al., 2014).
https://doi.org/10.5194/acp-2020-1245 Preprint. Discussion started: 11 January 2021 c Author(s) 2021. CC BY 4.0 License. Engström et al. (2014) also found that CMIP5 models simulating a higher cloud cover have a tendency to produce a smaller cloud albedo value. Darker clouds can offset the contribution of the higher cloud cover to the planetary albedo, resulting in relatively a consistent model-driven planetary albedo. This is a presentation of the "too few, too bright" problem that persists in GCMs (Nam et al., 2012). To validate whether or not this problem has been improved in the AMIP6 models, we 225 compared the relationship between regressed cloud albedo and cloud fraction (See Fig. S1). The correlations driven by the 28 AMIP6 models were -0.28, 0.19, -0.11, -0.71 and 0.43 for the Peruvian, Namibian, Californian, Australian and Canarian regions, respectively. Compared with the results from the CMIP5 models (Engström et al., 2014), noticeable progress was found at the Namibian and Californian regions while a high negative correlation was simulated at the Australian region, indicating that the new generation models need to be further improved to resolve the longstanding problem. Further, Figure 5 shows the annual cycles of the cloud albedo estimated by the satellite and AMIP5/AMIP6 models for the five regions. The seasonal variation in the cloud albedo at each region takes a shape of single peak distribution. In terms of similarity among regions, the cloud albedo at all regions reached the maximum value during the boreal winter season, i.e., December to January in the Northern Hemisphere while June to July in the Southern Hemisphere. Many previous studies 280 (Lin et al., 2009;Wood, 2012;Dong et al., 2014) have demonstrated that the seasonal variations of marine cloud properties (e.g., cloud fraction, LWP and cloud thickness) are strongly affected by meteorological conditions. Employing a 19-month record of ground-based lidar/radar observations from the Atmospheric Radiation Measurement Program Azores site, for example, Dong et al. (2014) found that the seasonal variations of cloud thickness and LWP are closely related to the seasonal synoptic patterns (e.g., transport of water vapor, relative humidity, high/low pressure system). Furthermore, the influence of 285 aerosols loading is non-neglectable. While the aerosols act as CCN, the concentration of CCN can significantly influence the cloud albedo of low clouds (Twomey, 1974). On the other hand, absorbing aerosols near stratocumulus may enhance absorbing solar energy, resulting in an influence on the dynamical evolution of stratocumulus causing a change in the cloud albedo (Wilcox, 2010). The seasonal cycle of the cloud albedo at the Australia region showed the largest amplitude among The COT usually increases with an increase in cloud LWP, resulting in an increase in the cloud albedo (Wood, 2012). Gryspeerdt et al. (2019) also concluded that LWP is the main factor controlling liquid cloud albedo. Thus, this study investigated the seasonal variation of LWP and found that the change in LWP is strongly correlated with the change in cloud 295 albedo at the Peruvian, Australian and Canarian regions (see Fig. S4). For the Namibian region, however, many studies have shown that the continuous transportation of absorbing biomass burning aerosols from Africa to the region during the African biomass burning season from August to October (Das et al., 2017) can reside above the clouds, resulting in an increase in the cloud albedo by thickening the stratocumulus (Wilcox, 2010(Wilcox, , 2012. Zuidema et al. (2018) also found that the biomass burning aerosols generally exist in the boundary layer at the earlier time of the biomass burning seasons and are mainly 300 located at above the clouds in September to October, which is caused by the northwestward transportation of the biomass burning aerosols from the African continent. However, Fig. 5b shows that the peak of the cloud albedo occurred in July and then continuously decreased from August to October at the Namibian, indicating that the changes in the cloud albedo are

The impacts of different aerosol types and meteorological factors on cloud albedo changes
Cloud liquid water may affect the COT, which is subsequently influencing the cloud albedo (Wood, 2012). Bender et al. There is a considerable discrepancy in the results between the two groups. For the lower LWP bin (i.e., LWP ≤ 60 g m -2 ), the results showed that the regression coefficient related BC to the cloud albedo was positive while DU and OC-related 340 coefficients were negative, which indicates that the cloud albedo increase with increasing BC and decrease with increasing DU/OC. Fig. 6b also clearly shows that DU and BC have a larger contribution to the change in the cloud albedo compared with other predictors, e.g., omega900 and RH850. Under LWP > 60 g m -2 , the contribution of SO4 to the cloud albedo was the largest. In addition, DU, SO2, BC and RH850 also considerably contributed to the cloud albedo.
In addition to the effects of LWP, the difference in the relative contribution may be induced by the regional variability in 345 aerosol types. A smaller LWP mainly appeared at the Namibian and Canarian where the main aerosol types are DU and BC, while lower BC loadings were found at the regions with a larger LWP (Fig. S5) The previous studies (e.g., McCoy et al., 2017;Li et al., 2018) showed that SO4 plays a key role in modulating Nd. Although their results showed significant positive coefficients of SO4 with Nd, this study found an unexpected negative correlation of 375 SO4 with the cloud albedo. Such a result may be driven by the fact that the sulfate aerosol particles and dust are externally mixed. The previous studies showed that sulfate-covered dust can act as CCN, which may induce a decrease in the cloud albedo by enhancing the collision-coalescence progress of droplets (Levin et al., 1996;Rosenfeld et al., 2001).
The results of this study showed a weak dependency of the cloud albedo with omega 900. Under LWP ≤ 60g m -2 , the upward vertical velocity has a positive but weak effect on the cloud albedo by enhanced vapor supersaturations, which activates 380 aerosol particles more and subsequently increasing Nd (Twomey, 1959;Reutter et al., 2009). Under LWP > 60g m -2 , no significant correlation between the cloud albedo and omega900 was found. Note that the analysis of this study employed the average data at the monthly scale rather than raw satellite measurements at a pixel scale, which may make the cloud albedo less sensitive to omega900. In the study, the unexpected negative dependency of the cloud albedo with RH850 was found for the two datasets divided in this study. Drier free-troposphere humidity usually drives stronger entrainment of dry air, which 385 induces evaporating and raising lifted condensation level, resulting in a reduced cloud thickness (Wood, 2012;Eastman and Wood, 2018). However, Ackerman et al. (2004) showed that the sedimentation of cloud droplets reduces the entrainment It is also found from Figure 6 that changes in LWP can also cause an alteration of the relationship between aerosol and the 400 cloud albedo. To further investigate the influence of meteorological factors on the relationship, the partial correlations were calculated to eliminate the influence of meteorological parameters individually or simultaneously. If the partial correlation is similar to the total correlation, it means that the influence of meteorological factors on the relationship is limited. In contrast, the influence of meteorological factors on the relationship may be significant if the partial correlation and the total correlation are the opposite sign. Given three meteorological parameters (omega900, RH850 and LWP) considered in this 405 study, the total correlation and partial correlation between the cloud albedo and different aerosols for two sample groups are given in Table 3. The total and partial correlations are similar for BC, OC, SO4 under LWP ≤ 60 g m -2 , indicating that meteorological factors have little influence on the interactions between the aerosols and the cloud albedo. The correlation of SS was weak when eliminating the effects of meteorological factors. When the influence of LWP was eliminated, the correlation of DU becomes much weaker, indicating that the correlation of DU is sensitive to LWP. On the contrary, the 410 correlation of SO2 was stronger when the influence of LWP was eliminated. Under LWP > 60 g m -2 , the correlation of SO2 ranged from -0.11 to 0.25 by eliminating the influence of meteorological parameters, indicating the relationship between SO2 and the cloud albedo is extremely sensitive to the influences of meteorological factors. However, the impacts of meteorological factors on the aerosol-cloud interactions were insignificant for the other aerosol types.

415
The cloud albedo at the subtropical marine subtropical stratocumulus regions has a key role in regulating the regional energy budget. However, climate models have a lack of skill to properly capture the cloud properties for the regions. Therefore, the CMIP6 has more attention to improving some long-standing model biases, e.g., the low cloud simulation over tropical oceans and surface processes (Stouffer et al., 2016). Accordingly, considerable improvements in reproducing the observed https://doi.org/10.5194/acp-2020-1245 Preprint. Discussion started: 11 January 2021 c Author(s) 2021. CC BY 4.0 License.
confidence in climate predictions, it is necessary to systematically evaluate and compare the performance of CMIP5 and CMIP6 models and to further study the processes that contribute to the cloud albedo using the satellite-driven and reanalysis data. This study investigated the performances of CMIP6 models in reproducing the cloud albedo at the five marine Due to the limitations of polar-orbiting satellite observations, this study did not obtain a complete diurnal cycle of cloud 445 properties and radiation flux, which may induce a bias in the results of this study. The diurnal cycle of marine subtropical stratocumulus cloud albedo is usually significant due to the diurnal cycle of solar energy (Wood, 2012). The maximum cloud thickness usually occurs in the morning and gradually decreasing over the afternoon due to absorbing solar radiation in the cloud layer (Wood et al., 2002;Christensen et al., 2013). It is a challenge to evaluate how much of the cloud albedo bias contributes to the diurnal cycle of cloud albedo. Therefore, it is necessary to evaluate the diurnal cycle of cloud albedo in the 450 marine subtropical stratocumulus regions for reducing the uncertainties in cloud radiation interactions in GCMs. Note that the "too bright, too few" problem was improved at the Namibian and Californian regions in AMIP6. However, even if some models can simulate the cloud albedo more reasonably, it is questionable if other cloud properties can be captured (e.g., total https://doi.org/10.5194/acp-2020-1245 Preprint. Discussion started: 11 January 2021 c Author(s) 2021. CC BY 4.0 License.
cloud fraction), consequently resulting in significant biases in radiation (see Fig. S3). Therefore, we need to pay more attention to improving the calculation of total cloud fraction in the GCMs. Recently, some studies are devoted to improving 455 cloud overlap parameterization for accurately simulate the cloud fractions in GCMs (Li et al., 2018(Li et al., , 2019. Accordingly, it is also necessary to evaluate the improvement of cloud overlap scheme on cloud radiation interaction using long-term satellitedriven observations and reanalysis data.