Evaluation of multidecadal variability in CMIP5 surface solar radiation and inferred underestimation of aerosol direct effects over Europe, China, Japan, and India

Authors


Corresponding author: R. J. Allen, Department of Earth Sciences, University of California Riverside, Riverside, CA 92521, USA. (rjallen@ucr.edu)

Abstract

[1] Observations from the Global Energy Balance Archive indicate regional decreases in all sky surface solar radiation from ∼1950s to 1980s, followed by an increase during the 1990s. These periods are popularly called dimming and brightening, respectively. Removal of the radiative effects of cloud cover variability from all sky surface solar radiation results in a quantity called “clear sky proxy” radiation, in which multidecadal trends can be seen more distinctly, suggesting aerosol radiative forcing as a likely cause. Prior work has shown climate models from the Coupled Model Intercomparison Project 3 (CMIP3) generally underestimate the magnitude of these trends, particularly over China and India. Here we perform a similar analysis with 173 simulations from 42 climate models participating in the new CMIP5. Results show negligible improvement over CMIP3, as CMIP5 dimming trends over four regions—Europe, China, India, and Japan—are all underestimated. This bias is largest for both India and China, where the multimodel mean yields a decrease in clear sky proxy radiation of −1.3±0.3 and −1.2±0.2 W m−2decade−1, respectively, compared to observed decreases of −6.5±0.9 and −8.2±1.3 W m−2decade−1. Similar underestimation of the observed dimming over Japan exists, with the CMIP5 mean dimming ∼20% as large as observed. Moreover, not a single simulation reproduces the magnitude of the observed dimming trend for these three regions. Relative to dimming, CMIP5 models better simulate the observed brightening, but significant underestimation exists for both China and Japan. Overall, no individual model performs particularly well for all four regions. Model biases do not appear to be related to the use of prescribed versus prognostic aerosols or to aerosol indirect effects. However, models exhibit significant correlations between clear sky proxy radiation and several aerosol-related fields, most notably aerosol optical depth (AOD) and absorption AOD. This suggests model underestimation of the observed trends is related to underestimation of aerosol direct radiative forcing and/or deficient aerosol emission inventories.

1 Introduction

[2] Solar radiation incident upon the surface of the Earth plays a critical role in the climate system, driving surface temperatures, large-scale atmospheric circulation, and the hydrological cycle, while also of extreme importance to the biosphere. Measurements in many regions throughout the world have shown large multidecadal swings in all sky surface solar radiation, with decreases throughout the 1950s–1980s (“dimming”) and increases during the 1990s (“brightening”) [e.g., Ohmura and Lang, 1989; Gilgen et al., 1998; Stanhill and Cohen, 2001; Liepert, 2002; Wild et al., 2005; Wild, 2009a]. These variations in surface solar radiation are also consistent with independent observations of sunshine duration [e.g., Ohmura and Lang, 1989; Kaiser and Qian, 2002; Sanchez-Lorenzo et al., 2007, 2008; Stanhill and Cohen, 2008], diurnal temperature range [e.g., Liu et al., 2004a; Wild et al., 2007; Makowski et al., 2008, 2009], and pan evaporation [e.g., Peterson et al., 1995; Liu et al., 2004b; Roderick et al., 2007]. More recently, satellite measurements also support the existence of the brightening [Pinker et al., 2005; Hatzianastassiou et al., 2005].

[3] There are several mechanisms that can contribute to multidecadal variability in all sky surface solar radiation. Clouds, for example, have a significant impact on all sky surface solar radiation, with decreased cloud cover normally consistent with increased surface solar radiation. Atmospheric aerosols can also alter all sky surface solar radiation via scattering and absorption, both resulting in a decrease of surface solar radiation. In addition, aerosol indirect effects, such as enhanced cloud albedo [Twomey, 1977] and longer cloud lifetime [Albrecht, 1989], can reduce the amount of solar radiation reaching the surface.

[4] Most studies of dimming/brightening have been based on all sky surface solar radiation [e.g., Wild, 2009aand references within]. Because the data are for all sky conditions, clouds need to be considered as a contributing factor to dimming/brightening. Unfortunately, few clear sky observations exist on a widespread basis, making it difficult to quantify dimming/brightening trends under cloud-free conditions [e.g., Wild et al., 2005, 2009]. Norris and Wild [2007], however, developed a quantity called cloud cover radiative effect (CCRE), defined as the change in downward shortwave radiation produced by a change in cloud cover, to quantify the impact of cloud cover anomalies on surface solar radiation. After removing CCRE from all sky observations, a quantity called residual flux—which we refer to as clear sky proxy radiation—remains, allowing the radiative effects of long-term changes in anthropogenic aerosol to be more clearly distinguished from natural weather and climate variability.

[5] Using clear sky proxy, and the few clear sky observations that exist, changes in cloud cover do not appear to be the dominant factor in driving dimming/brightening trends over Europe and East Asia [Wild et al., 2005; Qian et al., 2006; Norris and Wild, 2007; Ruckstuhl et al., 2008; Norris and Wild, 2009; Wild et al., 2009]. 2009]. Over Europe, for example, Norris and Wild [2007] found dimming/brightening trends of similar magnitude using both all sky and clear sky proxy. Moreover, changes in cloud cover were inconsistent with the dimming/brightening, with slight decreases in cloud cover during dimming and slight increases during brightening. Similarly, Norris and Wild [2009] found similar Chinese dimming trends based on both all sky and clear sky proxy. Ruckstuhl et al. [2010] found ambiguous trends in surface solar radiation under all overcast conditions at 15 Swiss and eight northern German sites, with slightly negative but nonsignificant trends from 1981 to 2005. Under overcast conditions with thick clouds, a significant increasing trend was found, but about 9 times smaller than the increase under clear skies. These studies suggests aerosol direct effects account for most of the observed surface solar radiation changes. This conclusion is supported by examination of changes in aerosol optical depth (AOD), which agree well with observed dimming/brightening trends over several regions, including Europe and China [Streets et al., 2006, 2008; Ruckstuhl et al., 2008; Streets et al., 2009]. The dominant role of aerosol direct effects on dimming/brightening trends is also supported by modeling studies. For example, Kvalevåg and Myhre [2007] used a multistream radiative transfer model and found aerosols were the major contributor to global mean dimming, particularly aerosol direct effects. Furthermore, Romanou et al. [2007] conducted sensitivity experiments with the Goddard Institute for Space Studies (GISS)-ER climate model and found much of the 20th century dimming trend over the Northern Hemisphere—especially over central Africa, Europe, and South and East Asia—is a result of the tropospheric direct aerosol effect. We note, however, that some have found a significant contribution of aerosol indirect effects on decadal variations in Surface Solar Radiation (SSR). Ohmura [2009], for example, found that the aerosol direct and indirect effects played about an equal weight in changing global solar radiation at five sites in Europe and Japan.

[6] Prior studies have shown global climate models generally simulate the observed dimming/brightening qualitatively but underestimate the corresponding magnitude over several regions. Using models from the Coupled Model Intercomparsion Project 3 (CMIP3), Ruckstuhl and Norris [2009] found large discrepancies in the sign and magnitude of modeled and observed dimming and brightening trends over Europe. Although most models displayed a transition from decreasing to increasing surface solar radiation, the timing of the reversal varied by over two decades. However, nearly all of the models produced a transition that was consistent with their corresponding aerosol emission inventory. Poor model performance was therefore attributed to incorrect aerosol emission histories. Similar results were obtained by Folini and Wild [2011] with the ECHAM5-HAM climate model, with premature termination of the European dimming attributed to deficient aerosol emissions.

[7] Dwyer et al. [2010] conducted a similar evaluation of CMIP3 dimming/brightening using clear sky proxy over China and Japan. Although all of the models exhibited significant dimming trends over China from 1961 to 1989, the largest model trend was less than half the magnitude and significantly different from the observed trend (based on clear sky proxy) of −8.6 W m−2 decade−1. No systematic relationship was found between the magnitude of the dimming and aerosol indirect effects. However, models that included black carbon (BC) aerosols produced stronger decreasing trends than those models without BC. Over Japan, observations yielded negligible dimming from 1971 to 1989, but significant brightening of 5.3 W m−2 decade−1from 1990 to 1999. Models, however, significantly underestimated this brightening trend, with the largest model trend about half of the observed trend at 2.3 W m−2 decade−1.

[8] Similar CMIP3 biases of dimming and brightening trends were found by Wild and Schmucki [2011]. Based on all sky surface solar radiation, about half of the models were able to reproduce the observed decadal variation in a qualitative way; this improved under clear sky conditions, where more models yielded a convex-shaped polynomial, in agreement with the observed (all sky) time series. Qualitative agreement was generally better for dimming over China and India, as opposed to Europe and Japan. All models, however, showed much smaller dimming/brightening amplitudes as compared to observations. Over India, for example, observations (based on all sky radiation) showed a significant dimming trend from 1964 to 2000 of −6.9 W m−2 decade−1. The corresponding trend in the models ranged from −4.0 W m−2 decade−1 to 0.45 W m−2 decade−1, with a multimodel mean of −1.2 W m−2 decade−1.

[9] As previously mentioned, multidecadal variations in surface solar radiation are consistent with several independent observations, including diurnal temperature range (DTR) [e.g., Liu et al., 2004a; Wild et al., 2007; Makowski et al., 2008, 2009]. DTR is an excellent proxy for changes in surface solar radiation because solar flux is only present during daylight and therefore primarily affects the daily maximum temperature. The nighttime minimum temperature, however, is primarily governed by thermal radiative effects. Thus, subtracting the minimum temperature from the maximum temperature (i.e., DTR) results in a quantity primarily governed by solar radiative effects. Poor simulation of the decadal variations in DTR by CMIP3 models [Wild, 2009b] provides further evidence for deficient simulation of dimming/brightening trends. However, other factors (e.g., atmospheric circulation) may contribute to decadal DTR variations [Easterling et al., 1997].

[10] Additional indirect evidence for model underestimation of dimming/brightening trends is their inability to reproduce the hemispheric asymmetry of trends in near-surface air temperature [Wild, 2012]. In the Northern Hemisphere, where aerosol emissions are largest, a lack of warming is observed during the dimming time period, whereas strong warming is observed during the brightening time period. In the Southern Hemisphere, where aerosol emissions are weaker, a steady warming throughout the dimming and brightening periods is observed. CMIP3 models are unable to reproduce this hemispheric asymmetry [Wild, 2012].

[11] In support of the upcoming Intergovernmental Panel on Climate Change fifth assessment report, the next model intercomparison project, CMIP5 [Taylor et al., 2012], is now underway. Similar to CMIP3, CMIP5 is comprised of historical simulations of the 20th century using coupled atmosphere-ocean global climate models, which include an interactive representation of the atmosphere, ocean, land, and sea ice. Models also performed future projections of climate change, based on scenarios of future emissions called Representative Concentration Pathways (RCPs) [Moss et al., 2010]. Notable improvements in some CMIP5 models include interactive prognostic aerosols, chemistry, and dynamical vegetation. A new class of models, called Earth System Models, also include biogeochemical coupling that accounts for carbon fluxes between ocean, land, and atmosphere reservoirs. Given such advancements, particularly those related to the representation of aerosols, it is of interest to evaluate the ability of CMIP5 models to simulate the observed dimming and brightening trends. Models that exhibit the observed magnitude and timing of dimming/brightening likely have more realistic aerosol radiative forcing, better regional-scale climate simulations and projections, and more accurate estimates of climate sensitivity.

[12] This paper is organized as follows: Section 2 describes the observational data and models used and provides a description of how clear sky proxy anomalies are calculated (with more details provided in section A). Section 3 compares observed and modeled all sky and clear sky proxy trends over four regions, including Europe, China, India, and Japan. Seasonal dimming and brightening trends are also discussed. Section 4 evaluates dimming/brightening in a suite of “perturbed physics” versions of the GISS climate models, and section 5 presents the first multimodel comparison of dimming/brightening with aerosol properties. Finally, section 6 offers an interpretation of the CMIP5 dimming/brightening trends, and conclusions follow in section 7.

2 Data and Methods

2.1 Observations

[13] Monthly mean values of all sky surface solar radiation (direct + diffuse) were obtained from the Global Energy Balance Archive (GEBA) [Ohmura et al., 1989; Gilgen and Ohmura, 1999] for Europe, China, India, and Japan. These are the only regions of the world with more than a handful of GEBA stations with multidecadal records. We set to missing any monthly values flagged as suspicious by GEBA quality control. For inclusion in our analysis, we required stations to have at least 1 monthly value in 2005 or later, and most have data through 2007. We required stations in China and Japan to have at least 1 monthly value in 1964 or earlier, and most have data back to 1961. For Europe and India, we required stations to have data only in 1971 or earlier because we lack cloud data prior to 1971 in these regions. We excluded stations with data gaps exceeding 24 months or missing more than 20% of the monthly values. We relaxed the latter threshold to 30% for India to increase the number of qualifying stations, and none of those had data gaps exceeding 13 months. We excluded stations poleward of 65°N because they have zero or near-zero solar radiation during winter, which would create problems for our subsequent calculations of relative changes. These steps will ensure nearly full sampling of the longest possible brightening and dimming periods. The number of stations meeting the preceding criteria is 38 for Europe, 6 for China, 6 for India, and 33 for Japan.

[14] For additional quality control, we compared time series of all sky radiation anomalies with time series of cloud cover anomalies from nearby weather stations (usually co-located with the GEBA station). Variations in cloud cover dominate month-to-month anomalies in all sky radiation, and it is suspicious if they are only weakly correlated. We converted station cloud cover anomalies to “cloud cover radiative effect” anomalies using the procedure described below and calculated the correlation with the GEBA station all sky radiation anomalies. Following Norris and Wild [2007], any GEBA station in Europe or Japan with a correlation less than 0.60 was excluded from the analysis. This eliminated seven stations in Europe and two stations in Japan. We relaxed the correlation threshold to 0.40 and 0.35 for GEBA stations in China and India, respectively, due to the much smaller number of stations available. Sensitivity tests indicate that the magnitude of regional mean trends changes very little with correlation threshold until very few stations are left. At the end of the selection procedure, we had 31 stations for Europe, 6 stations for eastern China, 6 stations for India, and 31 stations for Japan. Figure 1 shows the locations of the GEBA stations.

Figure 1.

Location of the GEBA stations (black crosses), cloud stations (red dots), and International Satellite Cloud Climatology Project (ISCCP) equal-area grid boxes (rectangles) for each of the four regions.

[15] We calculated monthly anomalies in all sky surface solar radiation for each GEBA station and then aggregated station anomalies to grid box anomalies according to the median, rather than the mean, to mitigate the potential effect of outliers. The source of our grid boxes was the equal-area grid from the International Satellite Cloud Climatology Project (ISCCP) [Rossow and Schiffer, 1999]. Regional mean anomalies were computed by aggregating grid box anomalies according to the median of 24 grid boxes over Europe, 6 grid boxes over China, 6 grid boxes over India, and 15 grid boxes over Japan. Figure 1 shows the locations of the ISCCP grid boxes. At least half of the grid boxes were required to have data in order to calculate a regional mean, and most of the time, all grid boxes contributed. We found close agreement between regional anomaly time series of GEBA all sky radiation and downward shortwave radiation flux at the surface from the ISCCP Flux Data set [Zhang et al., 2004] on interannual timescales when cloud variability dominates aerosol variability. Aerosol changes dominate cloud changes on interdecadal timescales; however, so our analysis did not use the ISCCP Flux Data set for trend analysis because that data set used model-derived tropospheric aerosol prior to 2000 and had no temporal variations thereafter.

[16] Monthly mean values of total cloud cover were obtained from several sources. The NDP026-D [Hahn and Warren, 2003] provided visual reports of total cloud cover during 1971–1996 from weather stations in Europe and India. Unfortunately, cloud cover for the 1960s was not publicly accessible for most countries in Europe and India. The NDP-039 database for China [Shiyan et al., 1997] provided monthly mean station cloud cover through 1993, and the Japan Meteorological Agency website (http://www.data.jma.go.jp/obd/stats/data/en/smp/index.html) provided monthly mean station cloud cover through the present. We employed ISCCP monthly mean cloud amount in equal-area grid boxes during 1983–2009 [Rossow and Schiffer, 1999] to extend the cloud records for Europe, China, and India past the 1990s. Some apparent artifacts exist in the ISCCP cloud record, and we applied an approach similar to that described in Clement et al. [2009] to mitigate them. Specifically, we removed via linear regression all cloud variability related to changes in satellite view angle and all cloud variability common to the entire area viewed by a satellite. These corrections eliminated the discrepancy between station and ISCCP cloud cover in the late 1980s noted by Norris and Wild [2007]. Station cloud cover anomalies were aggregated to grid box anomalies according to the median. Cloud cover anomaly time series from surface and satellite data sets exhibited close agreement, as was previously demonstrated in Norris and Wild [2007, 2009]. We aggregated anomalies from temporally overlapping data sets to produce a single cloud cover time series for each grid box.

[17] As was done in Norris and Wild [2007, 2009], the radiative impact of surface observed cloud cover anomalies was estimated by multiplying the monthly grid box cloud cover anomalies (CC) by the ratio of long-term mean SW cloud effect (inline image) divided by the long-term mean cloud cover (inline image). We call the resulting values SW cloud cover radiative effect anomalies (CCRE). The difference between conventional cloud radiative effect and cloud cover radiative effect is that the former includes radiative impacts of all changes in cloud properties, whereas the latter includes only those radiative impacts that vary linearly with cloud cover anomalies. Section A provides a more detailed explanation.

[18] Since multidecadal clear sky flux records are not available for most GEBA stations and ISCCP clear sky flux is unsuitable for trend analysis, we must estimate clear sky flux anomalies from other parameters. We do so by applying linear regression to GEBA all sky radiation anomalies (dependent parameter) and estimated SW CCRE anomalies (independent parameter) and then calculating residuals from the best fit line. This is done separately for each grid box and calendar month. We refer to the resulting quantity as “clear sky proxy” anomalies, which were called “residual flux” anomalies in Norris and Wild [2007, 2009], Ruckstuhl and Norris [2009], and Dwyer et al. [2010]. Clear sky proxy anomalies include both clear sky solar radiation anomalies and the effects of changes in cloud albedo that are uncorrelated with cloud cover. We also note that clear sky proxy anomalies are related to the true clear sky flux for the portion of the sky that is cloud free (i.e., inline image). For example, detrended correlations between the CMIP5 ensemble mean clear sky proxy time series and SWclr×(1−CC¯) range from 0.94 for China to 0.99 for Europe. Based on individual models, the corresponding (mean) correlation ranges from 0.53 for China to 0.80 for Europe. Thus, clear sky proxy radiation includes cloud masking of aerosol direct effects on surface solar radiation. The use of clear sky proxy anomalies enables GEBA trends to be more distinctly seen because interannual “noise” from cloud cover variability is removed.

[19] One important question is whether GEBA station measurements accurately represent temporal variability of regional mean radiation flux. The close agreement between GEBA and ISCCP for interannual variability in all sky solar flux [Norris and Wild, 2007, 2009] answers this in the affirmative, at least when more than 20 stations are available. We unfortunately have only six stations for China in the present work since records for the other stations end in 2000. Nonetheless, it appears that six stations may be sufficient to characterize regional multidecadal trends; as will be subsequently shown, the magnitude of dimming and brightening trends averaged over six stations in China are within 25% of those averaged over 23 stations in Norris and Wild [2009]. Another important question is whether low-frequency variability in radiation flux is too spatially heterogeneous to be characterized by a single regional average. Our results suggest that a single regional average is appropriate. For their respective dimming periods, we find that 21 out of 24 European grid boxes, 5 out of 6 China grid boxes, 6 out of 6 India grid boxes, and 12 out of 15 Japan grid boxes exhibit decreasing solar flux (not shown). For their respective brightening periods, we find that 21 out of 24 European grid boxes, 5 out of 6 China grid boxes, and 15 out of 15 Japan grid boxes exhibit increasing solar flux (not shown).

2.2 Models

[20] Monthly downwelling all sky and clear sky surface solar radiation, as well as monthly total clover cover, from the historical simulation of the 20th century were downloaded for all available models from the CMIP5 data portal (http://cmip-pcmdi.llnl.gov/cmip5/data_portal.html). This resulted in a total of 42 climate models and 173 realizations, as displayed in Table 1. To extend the historical simulations beyond their nominal ending year of 2005, we primary use RCP4.5 since most models performed this experiment. In situations where a model lacked a RCP4.5 simulations, we use either the historical extended experiment—in which the historical simulation is continued through 2012—or RCP8.5. Table 1 lists which experiment was used to extend each model beyond 2005. In a few situations, a model, or a realization of a model, lacked post-2005 data. For example, Community Earth System Model Fast Chemistry (CESM1-FASTCHEM) has no archived experiments extending past 2005. Of the five GFDL-CM3 realizations, only one has data past 2005. We still use these models in our analysis, but the corresponding trends end in 2005, as opposed to 2007. Although aerosol emissions are different between the various RCPs, these differences are negligible over the short, 2 year time period. We therefore do not discriminate between those models that were extended through 2007 using RCP4.5, RCP8.5, or historical extended.

Table 1. CMIP5 Models and Number of Simulations (Runs) Useda
InstitutionModelNo.4.5HistExt8.5Aerosols
  1. a

    An “X” indicates which experiment—RCP4.5, Historical Extended or RCP8.5—was used to extend the historical simulation past 2005. Models with archived aerosol data, including aerosol optical depth (AOD) at 550 and 870 nm, absorption aerosol optical depth (AAOD) at 550 nm, as well as column loads (load) and emissions (emis) of individual aerosol species, are also indicated. If a model lacked data for some of its runs, then the available runs are indicated (e.g., R1). NOIE, No Aerosol Indirect Effects.

  2. b

    Commonwealth Scientific and Industrial Research Organization;

  3. c

    College of Global Change and Earth System Science;

  4. d

    Centre National de Recherches Meteorologiques;

  5. e

    Centre Europeen de Recherche et Formation Avancees en Calcul Scientifique;

  6. f

    Queensland Climate Change Centre of Excellence;

  7. g

    Institute of Atmospheric Physics;

  8. h

    Korea Meteorological Administration;

  9. i

    Japan Agency for Marine-Earth Science and Technology;

  10. j

    Atmosphere and Ocean Research Institute (University of Tokyo);

  11. k

    National Institute for Environmental Studies

CSIROband Bureau of MeteorologyACCESS1.01X  No AAOD
 ACCESS1.31X  No AAOD
Beijing Climate CenterBCC-CSM1.13 X  
 BCC-CSM1.1(m)3 X  
GCESSc, Beijing Normal UniversityBNU-ESM1X  AOD 550
Canadian Centre for Climate Modelling and AnalysisCanESM25 X Load and emis
National Center for Atmospheric ResearchCCSM46X   
Community Earth System Model ContributorsCESM1(BGC)1X   
 CESM1(CAM5)3X  AOD 550
 CESM1(CAM5.1, FV2)4    
 CESM1(FASTCHEM)3    
 CESM1(WACCM)4R2-R4   
CNRMd/ CERFACSeCNRM-CM510 X  
CSIROb, Industrial Research Organization, and QCCCEfCSIRO-Mk3.6.010X  X
LASG, IAPg, Chinese Academy of Sciences, and CESSFGOALS-g25 X  
LASG, IAPg, Chinese Academy of SciencesFGOALS-s23R2 R3AOD 550
The First Institute of Oceanography, SOA, ChinaFIO-ESM3    
NOAA Geophysical Fluid Dynamics LaboratoryGFDL-CM35R1  No BC/OA emis
 GFDL-ESM2G1X  No emis
 GFDL-ESM2M1X  No emis
NASA Goddard Institute for Space StudiesGISS-E2-H p1-p3 and NOIE5 eachp2 & p3p1 X (p2 and p3); No AOD 870
 GISS-E2-H-CC1 X  
 GISS-E2-R p1-p3 and NOIE5 eachp2 & p3p1 X (p2 and p3); No AOD 870
 GISS-E2-R-CC1 X  
National Institute of Meteorological Research/KMAhHadGEM2-AO1X   
Met Office Hadley CentreHadCM310X   
 HadGEM2-CC3R1 R2-R3No AAOD
 HadGEM2-ES4X  No AAOD
Institute for Numerical MathematicsINM-CM41X   
Institut Pierre-Simon LaplaceIPSL-CM5A-LR6R1-R4  No emis/AOD 870
 IPSL-CM5A-MR3R1  No emis/AOD 870
 IPSL-CM5B-LR1X  No emis/AOD 870
JAMESTi, AORIj, and NIESkMIROC-ESM3R1  X
 MIROC-ESM-CHEM1X  X
AORIj, NIESk, and JAMESTiMIROC4h3X  No AAOD/AOD 870/SO2 emis
 MIROC55 X X
Max Planck Institute for MeteorologyMPI-ESM-LR3X   
 MPI-ESM-MR3X   
 MPI-ESM-P2    
Meteorological Research InstituteMRI-CGCM3 p1-p25 total R1-R3 (p1) X
Norwegian Climate CentreNorESM1-M3 X No AOD 870
 NorESM1-ME1X  No AOD 870

[21] In addition to the radiation and cloud variables, we also downloaded several monthly aerosol-related fields. These include ambient (“wetted”) aerosol optical depth (AOD) at both 550 and 870 nm, ambient absorption aerosol optical depth (AAOD) at 550 nm, and ambient fine aerosol optical depth (FAOD) at 550 nm (particles with wet diameter less than 1 μm). AOD fields do not include stratospheric aerosols if these are prescribed within the model but includes other possible background aerosol types. Additional aerosol fields include column loads (dry masses) and emission rates of individual anthropogenic aerosol species. Column loads of the following aerosol species, which depend on the model, were acquired: black carbon (BC), primary organic matter (POA), secondary organic matter (SOA), organic matter (OA), and sulfate (SO4). Models that resolved both POA and SOA fields were summed to yield OA. GFDL-CM3, GISS-E2-H, and GISS-E2-R models also included column load of ammonium (NH4); GFDL-CM3 also included column load of nitrate (NO3). Downloaded emission rates include sulfur dioxide (SO2), POA, OA, and BC. CSIRO-Mk3-6-0, in addition to both GISS and both Norwegian models, also include direct emission of sulfate. We note that not all models archived aerosol data (e.g., CNRM-CM5), and therefore, not all models are included in our comparison of dimming/brightening and aerosol fields.

[22] Each CMIP5 variable is interpolated to the same ISCCP equal area grid boxes (Figure 1) used for the observations for each region. Interpolation is performed using the National Center for Atmosphere Research Command Language (NCL) function “linint2_points,” which uses bilinear interpolation to interpolate from a rectilinear grid to an unstructured grid. Regional mean anomalies were computed by aggregating grid box anomalies. CMIP5 cloud cover radiative effect anomalies, and the corresponding clear sky proxy anomalies, are calculated in a similar manner as with the observations (section 2.1). Anomalies are computed for the entire time period for each region (e.g., 1961–2007 for China). In addition to all sky and clear sky proxy anomalies, we also calculate the corresponding relative anomalies. Relative all sky anomalies are obtained by dividing the monthly all sky anomalies by the corresponding long-term monthly mean all sky radiation. Relative clear sky proxy anomalies are obtained by dividing the monthly clear sky proxy anomalies by the long-term monthly mean difference between all sky radiation and cloud cover radiative effect.

2.3 CMIP5 Emission Inventory

[23] Unlike CMIP3, models participating in CMIP5 use the same aerosol emission inventory. Anthropogenic aerosol emissions, which covers the historical period (1850–2000) in decadal increments at a horizontal resolution of 0.5° in latitude and longitude, are from Lamarque et al. [2010]. Sulfur dioxide emissions in Lamarque et al. [2010] are based on a bottom-up mass balance method from Smith et al. [2011]. Emissions of black carbon and organic carbon come from an update of Bond et al. [2007] and Junker and Liousse[2008]. Also included in Lamarque et al. [2010] are biomass burning emissions, which are based on a combination of three data sets: the Global Inventory for Chemistry-Climate inventory [Mieville et al., 2010] for the 1900–1950 time period; the Reanalysis of Tropospheric chemical composition inventory [Schultz et al., 2008] for the 1960–1990 time period; and the Global Fire Emissions Database version 2 (GFEDv2) inventory [van der Werf et al., 2006] for the 2000 estimate. As previously mentioned, post-2000 emissions come from the Representative Concentration Pathways (RCPs) [Moss et al., 2010], which are based on a range of projections of future population growth, technological development, and societal responses. A total of four RCPs exist, RCP2.6, 4.5, 6.0, and 8.5, where the number denotes the radiative forcing in 2100 relative to 1850 [van Vuuren et al., 2011]. The four pathways are harmonized to match historical emissions, and each other, in 2000, with harmonization maintained through 2005. Tropospheric and stratospheric ozone data, from 1850 to 2099, comes from Cionni et al. [2011]. Throughout this paper, we refer to this collection of emission inventories as the “CMIP5 emissions inventory” or “CMIP5 emissions.”

[24] We note that three models modified CMIP5 emissions. GISS-E2-R and GISS-E2-H scaled CMIP5 biomass burning emissions of BC and OA by 1.4 [Shindell et al., 2012, 2013]. CSIRO-Mk3-6-0 scaled all BC emissions by 1.25 and all OA emissions by 1.5 [Rotstayn et al., 2012].

[25] CMIP5 protocol also specifies an aerosol data set for models with prescribed (offline) aerosols. Offline aerosol fields come from a CAM3.5 simulation with a bulk aerosol model driven by CCSM3 sea surface temperatures and the 1850–2000 CMIP5 emissions. Thus, aerosol fields for models with prescribed aerosols are based on the same CMIP5 emissions, but such models lack aerosol-meteorology feedbacks. We evaluate the impact of prescribed versus prognostic aerosols on simulated dimming and brightening trends using the GISS climate model in section 4.

3 Results

[26] A simple way to characterize the long-term variations in surface radiation flux is through linear trends. Observed and modeled trends are estimated using least squares regression over all months. Uncertainty estimates account for the effects of autocorrelation by using the effective sample size, n(1−r1)(1+r1)−1, where r1 is the lag-1 autocorrelation and nis the number of months in the time series [Wilks, 1995]. All time series are first detrended before estimating the effective sample size. Trend significance is assessed using a two-tailed Student's t-distribution.

[27] Figure 2 shows a legend listing each model and the associated line color, line style, and symbol. This legend applies to all subsequent figures, unless otherwise noted. Depending on the figure, we plot the model-mean (average over all realizations for each model), the CMIP5 ensemble mean (average over all model-means), or the individual realizations of each model. Because the CMIP5 ensemble mean is based on the average of the model-means, those models with more realizations (e.g., CSIRO-Mk3-6-0 possesses 10, whereas ACCESS1-0 possesses only one) will carry similar weight in the CMIP5 ensemble mean.

Figure 2.

Legend showing each CMIP5 model and the corresponding color, line style, and symbol used. Observations (OBS) are in black. A few models, such as GISS, include perturbed physics runs, which are distinguished with a “p2” or “ p3.” This legend applies for all figures unless otherwise noted.

3.1 Europe

[28] Figure 3 shows the time series of all sky and clear sky proxy annual mean anomalies over Europe for both observations and models. Observed all sky and clear sky proxy anomalies both show similar multidecadal variability, with a decrease from 1971 through the mid-1980s, followed by a recovery. Compared to all sky anomalies, clear sky proxy anomalies possess less interannual variability due to the removal of cloud cover radiative effects. This is particularly noticeable for the models and allows the radiative effects of aerosols to be more clearly distinguished from natural weather and climate variability. The effects of major volcanic eruptions, including El Chichon in 1982 and Pinatubo in 1991, are also more clearly seen as negative deviations for clear sky proxy anomalies.

Figure 3.

European (top) all sky, (middle) clear sky proxy, and (bottom) relative clear sky proxy (left) annual mean anomaly time series and (right) scatter plots of the 1971–1986 dimming trend versus the 1987–2007 brightening trend. The left panels show the model-mean time series for each model, with the CMIP5 ensemble mean in thick red; scatter plots show each model-realization. Observations are in black. Error bars show the 95% confidence interval of the trend, accounting for autocorrelation. All sky and clear sky proxy units are W m−2, with trend units of W m−2 decade−1. Relative clear sky proxy units are %, with trend units of % decade−1.

[29] Since solar dimming over Europe occurs prior to ∼1990, we choose the 1971–1986 and 1987–2007 time periods for trend calculations (results are not sensitive to the exact choice of transition year—see below). The right-hand panels of Figure 3 show the corresponding 1971–1986 trend versus the 1987–2007 trend. Observed all sky trends exhibit a decrease (dimming) of −3.3±3.3 W m−2decade−1over 1971–1986 and a significant increase (brightening) of 3.5±1.9 W m−2decade−1over 1987–2007. Consistent with the CMIP5 all sky time series, CMIP5 all sky trends possess large scatter, with individual model realizations ranging from −4.0 to decade−1 during the dimming time period and from −2.4 to 8.0 W m−2decade−1during the brightening time period. Table 2shows that the CMIP5 multimodel mean yields a nonsignificant increase in all sky radiation of 0.4±1.0 W m−2decade−1during the dimming time period, and a significant increase of 2.0±0.8 W m−2decade−1during the brightening time period. Thus, the CMIP5 mean incorrectly simulates brightening during the observed dimming time period. However, due to the confounding influence of cloud cover variability on all sky radiation, interpretation of these modeled versus observed all sky trends is difficult.

Table 2. Observed and CMIP5 Model Mean Annual Mean Trends in All Sky and Clear Sky Proxy Radiation During the Dimming (D) and Brightening (B) Time Period for Each Regiona
Region All SkyClear Sky ProxyRel All SkyRel Clear Sky Proxy
OBSCMIP5OBSCMIP5OBSCMIP5OBSCMIP5
  1. a

    Also included are the corresponding trends in relative all sky and relative clear sky proxy radiation. The 95% uncertainty estimates are also included, accounting for autocorrelation. Units are W m−2decade−1for absolute trends and % decade−1for relative trends.

EuropeD–3.3±3.30.4±1.0–3.7±1.30.0±0.8–1.7±2.30.2±0.4–1.8±0.8–0.0±0.2
 B3.5±1.92.0±0.82.7±0.81.3±0.73.1±1.41.4±0.41.7±0.50.6±0.3
ChinaD–8.5±1.5–1.3±0.2–8.2±1.3–1.2±0.2–5.5±1.1–0.7±0.1–4.0±0.7–0.5±0.1
 B2.6±2.2–0.8±0.42.3±1.8–0.5±0.81.6±1.5–0.4±0.21.0±1.0–0.2±0.3
IndiaD–7.5±1.1–1.5±0.2–6.5±0.9–1.3±0.3–3.5±0.5–0.7±0.1–2.6±0.3–0.5±0.1
JapanD–5.8±2.2–0.7±0.3–4.6±2.3–0.9±0.2–4.5±1.7–0.4±0.2–2.4±1.2–0.4±0.1
 B4.4±2.0–0.2±0.35.5±1.2–0.4±0.43.0±1.2–0.1±0.22.3±0.5–0.2±0.2

[30] Based on clear sky proxy anomalies, observations show dimming and brightening trends of −3.7±1.3 and 2.7±0.8 W m−2 decade−1, respectively. These values are very similar to those obtained based on all sky anomalies, but clear sky proxy trends possess less uncertainty, and both dimming and brightening trends are significant. Similarly, CMIP5 clear sky proxy trends also feature reduced scatter relative to the all sky trends, which facilities comparison of modeled versus observed dimming/brightening. CMIP5 multimodel mean clear sky proxy trends are 0.0±0.8 W m−2 decade−1during the observed dimming time period and 1.3±0.7 W m−2 decade−1during the observed brightening time period. This suggests models as a whole significantly underestimate the decrease in surface solar radiation over 1971–1986. Similarly, Table 3show that no individual model-mean possess a decrease larger than −1.0 W m−2 decade−1. In terms of individual model realizations, none reproduce the magnitude of the observed dimming, with the largest decrease in clear sky proxy radiation about half that observed, at −1.5 W m−2 decade−1in GISS-E2-R_NOIE. However, several individual model realization's 95% confidence range overlaps with the corresponding 95% confidence range in observed dimming. These results are similar to the aforementioned CMIP3 studies of Ruckstuhl and Norris [2009] and Wild and Schmucki [2011], in addition to Haywood et al. [2011], based on the HadGEM-2 climate model forced with CMIP5 emissions.

Table 3. CMIP5 Model Mean Annual Mean Trends in Clear Sky Proxy Anomalies Over the Dimming (D) and Brightening (B) Time Period for Each Regiona
ModelEuropeChinaIndiaJapan
DBDBDDB
  1. a

    The 95% uncertainty estimates are also included, accounting for autocorrelation. The corresponding observed trends are listed in the bottom row. Units are W m−2 decade−1.

ACCESS1–01.3±1.32.4±0.9–1.7±0.6–1.2±1.2–2.0±0.4–0.5±0.6–0.6±0.7
ACCESS1–30.0±1.43.0±1.1–1.4±0.5–2.5±1.2–1.9±0.5–0.8±0.7–0.1±0.8
BCC–CSM1–10.8±0.90.6±0.8–0.8±0.40.8±0.9–0.9±0.4–0.7±0.60.5±0.7
BCC–CSM1–1–m0.3±0.81.0±0.6–0.7±0.30.1±0.7–1.1±0.4–0.7±0.50.3±0.6
BNU–ESM0.9±1.61.0±1.0–1.4±0.7–0.6±1.5–1.5±0.6–0.5±1.0–0.1±0.8
CanESM20.1±0.91.4±0.5–1.3±0.30.4±0.7–1.4±0.3–1.0±0.40.1±0.4
CCSM4–0.3±0.81.2±0.8–1.2±0.3–0.0±0.8–1.2±0.4–0.7±0.3–0.3±0.5
CESM1–BGC0.3±1.21.1±0.7–1.1±0.51.0±1.0–1.1±0.6–1.0±0.60.6±0.8
CESM1–CAM50.2±0.81.5±0.7–1.3±0.4–1.2±1.0–1.9±0.5–0.8±0.6–1.1±0.8
CESM1–CAM5–1–FV20.1±0.82.3±0.9–1.6±0.4–0.3±0.9–1.8±0.5–1.3±0.5–0.2±0.7
CESM1–FASTCHEM–0.2±0.60.7±0.7–0.8±0.31.6±1.0–0.7±0.5–0.7±0.50.3±0.6
CESM1–WACCM0.1±0.81.7±0.8–0.8±0.4–0.4±0.7–1.2±0.5–0.7±0.4–0.1±0.5
CNRM–CM50.7±2.24.4±1.5–1.8±0.3–1.9±1.2–2.4±0.2–0.8±0.3–1.2±0.4
CSIRO–Mk3–6–00.3±0.71.6±0.6–1.8±0.3–1.2±0.8–2.4±0.3–1.2±0.4–0.7±0.5
FGOALS–g2–0.3±0.50.3±0.2–0.2±0.30.2±0.6–0.4±0.2–0.3±0.30.2±0.3
FGOALS–s2–0.4±0.50.5±0.4–1.6±0.30.2±0.8–2.0±0.3–1.4±0.4–0.2±0.4
FIO–ESM0.1±0.8–0.4±0.6–0.1±0.40.7±0.8–0.1±0.4–0.4±0.5–0.2±0.5
GFDL–CM3–0.7±1.02.2±1.0–2.5±0.4–1.7±1.7–3.5±0.4–1.4±0.4–1.1±0.6
GFDL–ESM2G0.5±1.1–0.4±0.9–1.4±0.6–0.4±1.3–0.9±0.5–1.1±0.80.1±0.9
GFDL–ESM2M0.0±1.20.9±1.0–1.5±0.7–1.2±1.4–0.9±0.8–0.7±0.8–1.1±1.1
GISS–E2–H–1.0±0.60.7±0.4–1.5±0.20.5±0.7–1.3±0.5–1.1±0.3–0.4±0.4
GISS–E2–H_p2–0.5±0.61.1±0.5–1.6±0.3–0.8±0.9–1.3±0.4–1.4±0.3–0.4±0.4
GISS–E2–H_p3–0.6±0.81.6±0.6–1.7±0.3–0.7±0.8–1.4±0.4–1.4±0.3–0.4±0.4
GISS–E2–H_NOIE–1.0±0.50.5±0.5–1.4±0.30.3±0.8–1.2±0.4–1.4±0.3–0.2±0.5
GISS–E2–H–CC–0.7±0.80.8±0.8–1.5±0.50.1±1.2–0.6±0.7–1.2±0.6–1.0±0.6
GISS–E2–H–R–0.8±0.60.7±0.4–1.5±0.30.2±0.7–1.4±0.5–1.3±0.3–0.1±0.4
GISS–E2–H–R_p2–0.2±0.60.8±0.5–1.4±0.3–0.7±0.8–1.3±0.4–1.2±0.3–0.7±0.5
GISS–E2–H–R_p3–0.1±0.71.4±0.7–1.6±0.3–0.9±0.9–1.4±0.4–1.1±0.4–0.5±0.5
GISS–E2–H–R_NOIE–0.8±0.60.7±0.5–1.4±0.3–0.2±0.8–1.1±0.5–1.2±0.3–0.4±0.5
GISS–E2–H–R–CC–0.5±0.81.0±0.7–1.6±0.5–0.5±1.3–1.3±0.8–1.2±0.5–0.4±0.6
HadCM3–0.4±0.50.4±0.3–0.7±0.3–0.2±0.5–0.4±0.3–0.9±0.3–0.7±0.3
HadGEM2–AO–0.2±1.11.9±1.0–2.0±0.5–0.9±1.0–2.0±0.5–1.2±0.5–0.8±0.7
HadGEM2–CC0.1±0.82.2±0.7–1.6±0.3–0.9±0.8–1.5±0.3–0.6±0.4–0.6±0.4
HadGEM2–ES0.1±0.81.9±0.7–1.5±0.3–0.6±0.8–1.7±0.3–0.8±0.4–0.5±0.3
INM–CM4–0.5±0.6–0.1±0.5–0.4±0.4–0.2±1.2–0.4±0.4–0.4±0.6–0.4±0.7
IPSL–CM5A–LR–0.1±0.81.1±0.6–1.0±0.3–0.8±0.6–1.4±0.2–0.5±0.3–0.3±0.8
IPSL–CM5A–MR0.4±0.81.1±0.5–1.2±0.3–0.8±0.8–0.9±0.3–0.9±0.4–0.8±0.4
IPSL–CM5B–LR2.8±1.40.9±1.2–1.1±0.5–0.2±1.4–1.3±0.5–1.5±0.7–0.6±0.8
MIROC–ESM0.6±0.72.6±0.9–1.5±0.4–0.7±0.9–1.2±0.4–1.5±0.4–0.9±0.6
MIROC–ESM–CHEM0.7±1.12.7±0.8–1.6±0.5–1.3±1.1–0.9±0.4–1.1±0.5–0.5±0.7
MIROC4h0.2±0.71.6±0.7–1.4±0.3–0.8±0.7–1.5±0.3–0.8±0.3–0.4±0.4
MIROC50.7±1.03.8±0.9–1.7±0.3–1.4±0.8–1.5±0.3–1.0±0.3–0.9±0.4
MPI–ESM–LR–0.8±0.70.1±0.5–0.6±0.3–0.2±0.8–0.4±0.3–0.5±0.5–0.1±0.6
MPI–ESM–MR–0.5±0.70.2±0.4–0.3±0.30.2±0.8–0.8±0.4–0.4±0.6–0.8±0.6
MPI–ESM–P0.1±0.9–0.7±0.6–0.2±0.5–0.3±1.0–0.4±0.5–0.2±0.7–0.3±0.9
MRI–CGCM30.4±0.71.0±0.6–0.5±0.3–0.1±0.8–1.1±0.2–0.7±0.4–0.6±0.5
MRI–CGCM3_p20.3±0.81.5±0.9–0.2±0.40.5±1.0–1.2±0.2–0.6±0.5–0.3±0.7
NorESM1–M0.4±1.02.2±0.7–1.3±0.3–0.2±0.8–2.0±0.4–0.7±0.4–0.6±0.5
NorESM1–ME–0.6±1.31.0±0.9–1.2±0.6–0.6±1.3–1.8±0.7–0.6±0.70.2±0.7
OBS–3.7±1.32.7±0.8–8.2±1.32.3±1.8–6.5±0.9–4.5±2.35.5±1.2

[31] The observed increase in European clear sky proxy radiation over 1987–2007 is also underestimated by the CMIP5 mean, but this is not significant at the 95% confidence level. In fact, some individual models simulate the magnitude of the observed brightening, with CNRM-CM5 (blue circle) and MIROC5 (gray diamond) actually overestimating the brightening. The 1987–2007 model mean clear sky proxy trend for these two models is 4.4±1.5 and 3.8±0.9 W m−2 decade−1, respectively (Table 3).

[32] Changing the choice of transition year results in small differences in dimming/brightening and does not significantly alter any of our observation-model comparisons. We note, however, that underestimation of the dimming in the CMIP5 mean is exacerbated with the choice of a later transition year. For example, observed clear sky proxy trends from 1971 to 1989/1990 to 2007 are similar to those reported above at −2.7±1.1 and 3.3±1.0 W m−2 decade−1, respectively. However, the corresponding CMIP5 mean trends are 0.4±0.6 and 1.5±1.0 W m−2 decade−1. Thus, the model mean simulates brightening, as opposed to the observed dimming, over the alternate 1971–1989 time period.

[33] The bottom panels in Figure 3 shows that similar results are obtained with relative clear sky proxy anomalies. Note that the relative clear sky proxy anomalies exhibit less interannual variability than the corresponding absolute values. Based on observations, the standard deviation of the annual mean clear sky proxy (after detrending both time periods and averaging the corresponding standard deviations for dimming and brightening) is 1.4 W m−2, compared to 0.8% for relative clear sky proxy. Models show a similar decrease of the standard deviation −1.0 W m−2for clear sky proxy versus 0.6% for relative clear sky proxy (based on the mean standard deviation across all model-runs). The larger variability in nonnormalized clear sky proxy appears to be related to seasonal differences in interannual variability that get modified depending on whether there is weighting by insolation (original clear sky proxy) or not (relative clear sky proxy). The largest variability in clear sky proxy occurs during May-June-July (MJJ), when insolation is largest. The standard deviation of observed clear sky proxy during MJJ is 3.1 W m−2 versus 0.7 W m−2 during November-December-January (NDJ). However, seasonal differences in the variability of relative clear sky proxy are reduced −1.0% during MJJ and 1.4% during NDJ. Thus, the larger interannual variability in clear sky proxy, as opposed to relative clear sky proxy, appears to be related to the availability of solar insolation such that more solar radiation is associated with more variability in clear sky proxy radiation.

[34] Similarly, the uncertainty in the observed and modeled relative trends is generally reduced (as are the relative trends themselves). Observations show a significant decrease in relative clear sky proxy radiation of −1.8±0.8% decade−1from 1971 to 1986, followed by significant recovery of 1.7±0.5% decade−1from 1987 to 2007. The corresponding trends in the CMIP5 ensemble mean are 0.0±0.2 and 0.6±0.3% decade−1, respectively. Thus, both the relative dimming and brightening trends in the CMIP5 ensemble mean are significantly less than those observed. Similar to absolute clear sky proxy trends, however, the 95% confidence interval of relative dimming for several individual model realizations overlaps with the observed range, and the relative brightening trends in both CNRM-CM5 and MIROC5 are larger than observed.

[35] As previously mentioned, European all sky and clear sky proxy radiation trends are similar for observations, as well as for CMIP5 (Table 2). This implies most of the dimming and brightening is not caused by clouds, leaving aerosol direct effects as the likely cause. Similarity between all sky and clear sky proxy radiation trends exists not only for Europe but for the other three regions as well (Table 2). This further supports the central role of aerosol direct effects in driving multidecadal variability in surface solar radiation. A subsequent analysis will rigorously quantify the relative role of aerosol effects on dimming and brightening trends in CMIP5 models.

3.2 China

[36] A similar analysis is presented in Figure 4 for China. Observations show a large decrease in both all sky and clear sky proxy radiation from 1961 to 1989 of −8.5±1.5 and −8.2±1.3 W m−2 decade−1, respectively, in line with previous studies as summarized in Wild [2009a]. Following this significant decrease is a much weaker recovery, most of which occurs during the early 1990s, which is proceeded by a relatively stable period through 2007. The corresponding 1990–2007 observed trends in all sky radiation are 2.6±2.2 W m−2 decade−1and a significant 2.3±1.8 W m−2 decade−1for clear sky proxy radiation. Although a recent analysis by Tang et al. [2011] suggests this increase in the early 1990s may be spurious, several other proxies related to surface solar radiation show a transition from dimming to brightening. For example, Qi and Wang [2012] found changes in observed extreme temperatures and diurnal temperature range over some regions in China after 1990 that are consistent with increased surface solar radiation. Similar results are obtained by Liu et al. [2004a] and Wang et al.[2012]. Thus, we still refer to this latter time period from 1990 to 2007 as the “brightening” time period.

Figure 4.

As in Figure 3, but for China. Scatter plots show the 1961–1989 dimming trend versus the 1990–2007 brightening trend.

[37] Similar to Europe, CMIP5 all sky trends over the dimming and brightening time periods exhibit large scatter, ranging from −3.4 to 1.6 W m−2 decade−1during 1961–1989 and from −4.8 to 4.1 W m−2 decade−1during 1990–2007. Most model-realizations correctly simulate the observed decrease in all sky radiation during the dimming time period, but all significantly underestimate the magnitude of the decrease. The CMIP5 multimodel mean yields a significant decrease in all sky radiation of −1.3±0.2 W m−2 decade−1during the dimming time period, and a significant decrease of −0.8±0.4 W m−2 decade−1during the brightening time period (Table 2). Similar results are obtained based on clear sky proxy, with a significant decrease of −1.2±0.2 W m−2 decade−1during 1961–1989 and a nonsignificant decrease of −0.5±0.8 W m−2 decade−1from 1990 to 2007. Thus, the CMIP5 ensemble mean simulates the observed dimming over 1961–1989 but significantly less than observed. The ensemble mean also simulates continued dimming during the 1990–2007 time period, in disagreement with observations.

[38] The tendency of individual model realizations is again more clearly discerned using clear sky proxy radiation. Most model realizations yield decreased clear sky proxy radiation during both time periods, with a larger decrease during the 1961–1989 time period. The largest decrease of a single realization during this time period (and therefore, closest to observations) is −2.8 W m−2 decade−1by the GFDL-CM3 (red X). Table 3 show that this model also produces the largest model-mean decrease of −2.5±0.4 W m−2 decade−1. However, this is about 30% of the observed decrease and significantly less. GFDL-CM3, as well as other models that yield the largest dimming trends (e.g., CNRM-CM5, CSIRO-Mk3-6-0, and HadGEM2-AO), also tend to yield continued dimming during 1990–2007, when observations suggest a recovery. CESM1-FASTCHEM yields the largest brightening trend over 1990–2007 of 1.6 W m−2 decade−1, as well as the largest model-mean brightening of 1.6±1.0 W m−2 decade−1(Table 3). However, CESM1-FASTCHEM simulates weak dimming during 1961–1989 of −0.8±0.3 W m−2 decade−1.

[39] The significant underestimation of the observed dimming from 1961 to 1989 is even more pronounced in relative clear sky proxy radiation (bottom panels of Figure 4). Observations yield a decrease of −4.0±0.7% decade−1, whereas the CMIP5 ensemble mean yields −0.5±0.1% decade−1, ∼14% as large as observed. These results are similar to those based on CMIP3 [Dwyer et al., 2010; Wild and Schmucki, 2011], and with more recent simulations using HadGEM-2 [Haywood et al., 2011]. Consistent with Europe, we also note larger interannual variability in clear sky proxy, compared to relative clear sky proxy. The standard deviation of the observed annual mean time series (after detrending both dimming and brightening time periods separately and averaging the corresponding standard deviations) is 2.6 W m−2versus 1.3%, respectively. Again, the larger variability in the nonnormalized clear sky proxy time series is related to large variability during MJJ (4.7 W m−2) as opposed to NDJ (2.5 W m−2). Significant summertime inflation of the variability does not occur in the relative clear sky proxy time series (1.5% for MJJ versus 2.1% for NDJ). Similar conclusions are reached based on the models.

3.3 India

[40] Results for India are displayed in Figure 5. Unlike the other regions, observations show that India does not exhibit a recovery in all sky or clear sky proxy radiation, as found in previous studies. Both quantities show large decreases throughout 1971–2007, with all sky radiation decreasing by −7.5±1.1 W m−2 decade−1and clear sky proxy radiation decreasing by −6.5±0.9 W m−2 decade−1. Once again, all sky and clear sky proxy trends are very similar, suggesting most of the dimming signal is caused by aerosol direct effects, as opposed to clouds. Table 2 shows the corresponding CMIP5 ensemble mean trends are also negative at −1.5±0.2 and −1.3±0.3 W m−2 decade−1, respectively. These trends, however, are significantly less than observed.

Figure 5.

As in Figure 3, but for India. Scatter plots show the 1971–2007 dimming trend on both axes (since India lacks a “brightening” time period).

[41] Based on individual model realizations, clear sky proxy trends are all negative except for three (FGOALS-g2, FIO-ESM, and MPI-ESM-LR). Table 3, however, shows that all model-mean clear sky proxy trends are negative. Similar to China, GFDL-CM3 yields the largest decrease in clear sky proxy radiation over India. The largest decrease of any GFDL-CM3 realization is −3.7 W m−2 decade−1, and the model mean is similar at −3.5±0.4 W m−2 decade−1. Similarly, Table 3 shows that CSIRO-Mk3-6-0 and CNRM-CM5 yield relatively large model-mean dimming, in excess of −2.0 W m−2 decade−1. However, these values are less than half that observed and significantly less. Significant underestimation of the dimming trend is even more apparent based on relative clear sky proxy trends. The observed decrease in relative clear sky proxy radiation is −2.6±0.3 W m−2 decade−1. The corresponding CMIP5 mean trend is about 20% as large at −0.5±0.1 W m−2 decade−1, with the largest simulated relative trend (again from GFDL-CM3) of −1.3±0.3 W m−2 decade−1.

3.4 Japan

[42] The last region we consider, Japan, is shown in Figure 6. Time series of both all sky and clear sky proxy radiation show a large decrease during the first 10–15 years, followed by a relatively stable period from the mid-1970s to ∼1990, which in turn is followed by a recovery. Using 1984/1985 as the transition year, observed all sky radiation shows a significant decrease from 1961 to 1984 of −5.8±2.2 W m−2 decade−1, most of which occurs during the first decade. This is followed by a significant increase from 1985 to 2007 of 4.4±2.0 W m−2 decade−1. Similar results are obtained with clear sky proxy anomalies, with dimming of −4.6±2.3 W m−2 decade−1and brightening of 5.5±1.2 W m−2 decade−1. We note that these observed trends are quite large compared to the relatively small changes in Japanese emissions (to be discussed below). This suggests multidecadal variation in surface solar radiation over Japan may be related to transport of aerosol from China, which is broadly consistent with the midlatitude location of both countries and the prevalence of westerly winds, especially during winter [Igarashi et al., 2006; Takashima et al., 2009; Verma et al., 2011].

Figure 6.

As in Figure 3, but for Japan. Scatter plots show the 1961–1984 dimming trend versus the 1985–2007 brightening trend.

[43] The corresponding CMIP5 ensemble mean dimming and brightening trends based on clear sky proxy radiation are −0.9±0.2 and −0.4±0.4 W m−2 decade−1, respectively. Thus, similar to China, the CMIP5 ensemble simulates the observed dimming but significantly underestimates the magnitude and continues to simulate dimming during the brightening time period. The scatter plots in Figure 6 show no model realization is able to reproduce the magnitude of the observed dimming and brightening trend over Japan—not only for clear sky proxy radiation, but for all sky radiation too. GISS-E2-H_p3 simulates the largest dimming of any realization at −2.0 W m−2 decade−1, about 45% as large as observed. MIROC-ESM and IPSL-CM5B-LR simulate the largest model-mean dimming at −1.5±0.9 W m−2 decade−1(Table 3). CESM-CAM5-1-FV2 yields the largest brightening of any realization at 1.0 W m−2 decade−1, about 20% as large as observed. Moreover, Table 3shows that only eight models possess positive model-mean 1985–2007 clear sky proxy trends, with CESM1-BGC yielding the largest at 0.6±0.8 W m−2 decade−1. Although the CMIP3 analysis of Dwyer et al. [2010] did not find dimming over Japan (a shorter time period from 1971 to 1989 was used), underestimation of the brightening is similar to that reported here.

[44] As with the other regions, changing the transition year between dimming and brightening yields similar results. For example, the observed trends in clear sky proxy radiation from 1961 to 1989 and 1990 to 2007 are −3.7±1.6 and 5.2±2.0 W m−2 decade−1, respectively. The corresponding trends based on the CMIP5 mean are −0.8±0.2 and 0.0±0.7 W m−2 decade−1. Thus, over this alternate brightening time period from 1990 to 2007, the CMIP5 mean no longer simulates dimming, which is in closer agreement to the observations. However, the CMIP5 mean is still unable to simulate the observed brightening, and the largest brightening trend of any model realization remains ∼30% as large as that observed.

3.5 Seasonal Dimming and Brightening

[45] Table 4 shows observed and CMIP5 mean trends in clear sky proxy radiation (all sky radiation is not discussed in this section) over the dimming and brightening time periods for two seasons, May-June-July (MJJ) and November-December-January (NDJ). These two seasons approximately correspond to the summer and winter solstice in the Northern Hemisphere (NH), respectively. Observed seasonal trends are similar to those based on the annual mean, with decreases in clear sky proxy radiation during each region's dimming time period and increases during each region's brightening time period. For most regions, dimming and brightening trends are largest during MJJ, when solar insolation is strongest, and smallest during NDJ, when solar insolation is weakest. Over China, for example, clear sky proxy radiation decreases by −10.6±2.2 W m−2 decade−1from 1961 to 1989 during MJJ, while only decreasing by about half this rate during NDJ at −6.0±1.3 W m−2 decade−1. Similarly, the increase in clear sky proxy radiation over China from 1990 to 2007 is larger during MJJ at 4.8±3.8 W m−2 decade−1as compared to NDJ at 0.5±1.8 W m−2 decade−1. The CMIP5 mean yields similar results, with the decrease in MJJ clear sky proxy radiation about twice as large as that during NDJ (−1.7±0.3 versus −0.7±0.3 W m−2 decade−1). Furthermore, although models simulate continued Chinese dimming from 1990 to 2007, they simulate larger dimming during MJJ relative to NDJ.

Table 4. Observed and CMIP5 Model Mean Seasonal Mean Trends in Absolute and Relative Clear Sky Proxy Radiation During the Dimming (D) and Brightening (B) Time Period for Each Regiona
Region Clear Sky ProxyRel Clear Sky Proxy
MJJNDJMJJNDJ
OBSCMIP5OBSCMIP5OBSCMIP5OBSCMIP5
  1. a

    Seasons include May-June-July (MJJ) and November-December-January (NDJ). The 95% uncertainty estimates are also included, accounting for autocorrelation. Units are W m−2 decade−1.

EuropeD–5.7±3.30.1±0.7–0.9±1.00.0±0.3–1.9±1.10.0±0.2–1.6±1.80.0±0.5
 B4.1±2.12.6±0.70.7±0.50.3±0.31.2±0.70.8±0.22.0±1.00.5±0.6
ChinaD–10.6±2.2–1.7±0.3–6.0±1.3–0.7±0.3–3.4±0.7–0.5±0.1–4.9±1.2–0.5±0.2
 B4.8±3.8–1.0±0.60.5±1.8–0.1±1.11.6±1.2–0.3±0.20.6±1.6–0.1±0.8
IndiaD–6.4±1.9–1.3±0.5–7.1±1.1–1.3±0.4–2.1±0.6–0.4±0.1–3.5±0.5–0.6±0.2
JapanD–1.5±4.0–1.1±0.4–6.6±4.1–0.6±0.3–0.5±1.3–0.3±0.1–4.5±2.7–0.4±0.2
 B6.9±2.5–0.6±0.53.1±1.4–0.2±0.62.2±0.8–0.2±0.12.2±1.0–0.2±0.4

[46] A similar seasonal dependency holds for Europe, where observed dimming and brightening are both larger during MJJ as compared to NDJ. The CMIP5 mean also yields significantly larger brightening during MJJ relative to NDJ. However, Table 4shows that this seasonal dependency—for both China and Europe—does not exist when considering the corresponding trends in relative clear sky proxy radiation. The seasonal differences in relative clear sky proxy trends for both regions are much smaller and not significantly different. This highlights the importance of solar insolation in driving seasonal differences in dimming and brightening and also suggests minimal seasonal differences in aerosol load trends.

[47] Exceptions to this seasonal dependency on dimming/brightening include India, where the decrease in clear sky proxy from 1971 to 2007 is somewhat weaker during MJJ, relative to NDJ, but not significantly so. The CMIP5 mean also shows nonsignificant seasonal differences. This could be due to India's lower latitude, with less seasonal contrast in solar insolation, or due to the strong monsoon during the summer, which would favor enhanced aerosol removal via wet deposition and less aerosol during MJJ [e.g., Ramanathan et al., 2007; Chung et al., 2010]. Assuming Indian emission trends lack significant seasonal variability (a good assumption outside of heavy biomass burning regions and also implied by the similar seasonal relative trends for Europe and China), this favors smaller increases in aerosol load over India during MJJ. We favor this latter interpretation, since observations show MJJ also exhibits a weaker decrease (and significantly different) in relative clear sky proxy radiation as compared to NDJ (−2.1±0.6 versus −3.5±0.5 W m−2 decade−1; Table 4). This weaker decrease in relative clear sky proxy radiation during MJJ is not due to seasonal differences in solar insolation and is therefore likely related to smaller increases in aerosol load during MJJ due to the monsoon. Models also reproduce this seasonal dependence in relative dimming, but the difference is not significant at −0.4±0.1 W m−2 decade−1for MJJ versus −0.6±0.2 W m−2 decade−1for NDJ.

[48] We also note an additional exception over Japan, where the observed decrease in clear sky proxy radiation is weaker during MJJ as compared to NDJ, but not significantly so. The corresponding trends in relative clear sky proxy radiation show a similar but statistically significant difference, where the decrease in MJJ relative clear sky proxy radiation is −0.5±1.3 W m−2 decade−1versus −4.5±2.7 W m−2 decade−1for NDJ. Similarly, the CMIP5 mean also shows weaker relative dimming during MJJ, but the difference is not significant. Besides the aforementioned relative clear sky proxy trends for India, this is the only instance of significantly different seasonal relative clear sky proxy trends. The larger relative dimming during NDJ over Japan suggests a larger increase in aerosol load during winter. Recall that we have suggested part of the dimming over Japan may be a result of emissions from China. Eastward transport of Chinese aerosol is favored during the winter months, when zonal winds are more westerly, as found by several studies [Igarashi et al., 2006; Takashima et al., 2009; Verma et al., 2011]. Thus, we argue that the larger decrease in NDJ relative dimming over Japan is consistent with the notion that part of the Japanese dimming is related to advection of Chinese aerosol (to be discussed in section 6.3). Furthermore, because CMIP5 models significantly underestimate the magnitude of Chinese dimming, it is not surprising they also significantly underestimate the magnitude of Japanese dimming.

4 GISS Models

[49] Although most CMIP5 models use the same aerosol emission inventory, their representation of aerosol processes and estimation of aerosol radiative forcing are not the same. One clear distinction is the use of prescribed aerosols, where precomputed transient aerosol fields are read in offline, versus the more physically realistic prognostic aerosol approach, where aerosol fields are calculated online as a function of atmospheric state and transient emission inventories. Another relevant distinction involves if, and how, each model represents aerosol indirect effects. To evaluate the effects of prescribed versus prognostic aerosols on simulated dimming and brightening trends, we investigate a suite of “perturbed physics” versions of the GISS-E2 climate model [Schmidt et al., 2006; Shindell et al., 2013]. Since clear sky proxy anomalies contain the effects of changes in cloud albedo that are uncorrelated with cloud cover, we also investigate the importance of the first (cloud albedo) aerosol indirect effect. We note that other differences exist between the prescribed and prognostic GISS model runs (to be discussed), so any dimming/brightening differences between the two may not be completely due to the use of prescribed versus prognostic aerosols.

[50] Figure 7 shows scatter plots of clear sky proxy radiation over the dimming and brightening time period for each region based on several versions of the GISS model. The standard GISS models, denoted as GISS-E2-H and GISS-E2-R are the same, except E2-H uses the Hybrid Coordinate Ocean Model and E2-R uses the Russell ocean model [Hansen et al., 2007]. Both use prescribed aerosols (and ozone) using emissions as described in Koch et al. [2011], which are generally similar to the CMIP5 emissions [Lamarque et al., 2010]. One exception, however, is a larger global peak of SO2emissions (75 versus 60 Tg S per year), about a decade later than CMIP5 (1990 versus 1980). Furthermore, CMIP5 global BC biomass burning emissions are about 25% smaller at the end of the century. The standard GISS models parameterize the aerosol indirect effect according to Hansen et al.[2005], which is based on the empirical effects of aerosols on cloud droplet number concentration for low-level, warm stratiform clouds only, and includes both the cloud albedo effect [Twomey, 1977] and the cloud cover [Albrecht, 1989] effect. However, a scaling factor is chosen such that the cloud cover effect is the dominant effect, which results in a total aerosol aerosol indirect effect of −1 W m−2based on aerosol changes between 1850 and 2000.

Figure 7.

Scatter plots of annual mean (left) clear sky proxy and (right) all sky trends during the dimming versus brightening time period for each of the four regions. Panels show the model-mean for each GISS model, with the letters “H” and “R” referring to the ocean model, and “p2,” “p3,” and “NOIE” referring to the aerosol representation (see text for further description). Models with larger dimming and brightening trends better agree with the observations. Uncertainty estimates are as in Figure 3. Note that this figure uses colors/symbols for each GISS model as displayed in the top right panel.

[51] The GISS p2 and p3 models use prognostic aerosols based on an online aerosol chemistry and transport model [Koch et al., 2011], which includes processes such as uptake and rainout of soluble species, production of sulfate within clouds, removal of species by dry deposition using a resistance in-series scheme, as well as gravitational settling. Furthermore, chemistry and aerosols are coupled together, such that oxidant changes affect sulfate oxidation, and the aerosols affect photochemistry. Both GISS p2 and p3 use the CMIP5 [Lamarque et al., 2010] emissions inventory. GISS p2 uses the same aerosol indirect effect parameterization as the standard model, while p3 parameterizes the aerosol cloud albedo effect following Menon et al.[2010]. This parameterization is based on a prognostic equation to calculate cloud droplet number concentration. Unlike the aforementioned Hansen et al. [2005] parameterization, no scaling factor is assumed. Thus, comparing the standard GISS model with GISS p2 yields an approximation of the prescribed versus prognostic signal, while comparing GISS p2 versus p3 yields the sensitivity to two different parameterizations of aerosol indirect effects. GISS NOIE models are analogous to the standard model (and therefore, use prescribed aerosols), except there is no aerosol indirect effect. Comparing the standard GISS model with GISS NOIE therefore yields the indirect effect signal based on the parameterization of Hansen et al.[2005]. Unlike previous scatterplots, Figure 7 shows the model-mean trend (average of the five realizations).

[52] In Europe, clear sky proxy trends based on the standard GISS models are very similar to GISS NOIE, suggesting aerosol indirect effects do not significantly contribute to the European dimming/brightening. Note that both the standard model and GISS NOIE also yield the largest dimming, in better agreement to observations (Table 3). However, GISS p2 and p3 yield more brightening. The larger dimming and weaker brightening in the standard models may be partially due to the aforementioned differences in SO2 emissions, since the global SO2peak is larger and occurs a decade later (in 1990), than the CMIP5 emissions GISS p2 and p3 use. The larger brightening in GISS p3, compared to p2, suggests the Menon et al. [2010] aerosol indirect effect parameterization results in more brightening than the Hansen et al. [2005] parameterization. In China, the opposite occurs, with GISS p2 and p3 yielding the least brightening (continued dimming, actually) during the 1990–2007 brightening time period. GISS NOIE, compared to the standard model, shows a bit less dimming during 1961–1989, which suggests aerosol indirect effects contribute a small amount (∼0.1 W m−2 decade−1) to the dimming trend over China seen in clear sky proxy. Over India, comparison of GISS with GISS NOIE shows aerosol indirect effects again contribute a small amount to the dimming trend based on clear sky proxy (∼0.1–0.3 W m−2 decade−1). Differences between the standard GISS models and p2 are small, suggesting little impact of prescribed versus prognostic aerosols. GISS p3 models, however, yield the largest dimming (larger than p2), which suggests the Menon et al. [2010] indirect effect parameterization yields more dimming than the Hansen et al. [2005] parameterization. Over Japan, the standard GISS models (particularly E2-R) yield the least dimming during the brightening time period, similar to China, and in better agreement to observations.

[53] Figure 7 also includes the corresponding analysis based on all sky surface solar radiation. As expected, based on the all sky trends, indirect effects have a larger effect, but only over Europe and China. This effect, however, is still relatively small at <1 W m−2and generally not significant. Interestingly, most of the enhanced European dimming due to indirect effects is offset when prognostic aerosols are used (enhanced brightening is more robust). Again, however, the enhanced dimming in the standard models may be partially due to the larger global SO2 emissions peak, which occurs a decade later than the CMIP5 emissions GISS p2 and p3 use. Over China, indirect effects result in more dimming of about −1 W m−2, which is the only case of a significant aerosol indirect effect. However, indirect effects for China also lead to more dimming during the brightening time period. Indirect effects based on all sky radiation appear to be small over both India and Japan. Similar to the clear sky proxy analysis, the all sky analysis shows no single GISS model performs best overall.

[54] To summarize, Figure 7 shows that the model-mean dimming/brightening trends are very similar for the different versions of the GISS model, with no significant differences. Similarly, no single version of the model routinely simulates dimming/brightening in best agreement to the observations (Figures 3456 and Table 2). This suggests the use of prescribed versus prognostic aerosols, as well as inclusion of aerosol indirect effects, have a relatively small impact on the corresponding dimming and brightening trends in clear sky proxy radiation. Thus, it is unlikely that the poor simulation of dimming/brightening trends by the CMIP5 mean is caused by models with prescribed (simplified) aerosols. Analysis of the suite of GISS models also suggests that the poor simulation is likely not due to models that lack aerosol indirect effects. However, additional analysis is required to support this statement, due to the differing parameterizations of aerosol indirect effects in the models. The small role of indirect effects in driving dimming/brightening is consistent with observational studies, at least over Europe [Ruckstuhl et al., 2010]. Additional observational studies are required to better understand the role of indirect aerosol effects on dimming/brightening.

5 CMIP5 Aerosol-Clear Sky Proxy Relationships

[55] Sheridan and Ogren [1999], and more recently Papadimas et al. [2012], observed that aerosol direct radiative effects at the surface are linearly proportional to the aerosol optical depth (the higher the AOD, the less solar radiation reaches the surface). Xia et al. [2007] found that an exponential relationship is more suitable than a linear relationship over a larger range of AOD. This is based on the Beer-Lambert law, I=I0exp[−AOD], where I is the intensity of solar radiation at the surface and I0 is the intensity of radiation at the top of the atmosphere. We therefore relate CMIP5 trends in clear sky proxy radiation to trends in exp[−AOD]. Note that an increase in AOD is associated with a decrease in exp[−AOD] and vice versa.

[56] Figure 8 shows scatterplots of the trend in clear sky proxy radiation versus the trend in exp[−AOD 550 nm] for each of the four regions. Scatterplots show values for both the dimming (smaller symbol) and brightening (larger symbol) time period using individual model realizations (e.g., five for GFDL-CM3). All four regions exhibit significant positive correlations, ranging from 0.77 over Europe to 0.35 over India. Significance of correlations is estimated from a Student's t-test, with the t value calculated according to inline image, where n is the number of trends. Thus, much of the variability in clear sky proxy trends is accounted for by aerosol optical depth trends, with more brightening/less dimming associated with a stronger decrease in AOD 550 nm (increase in exp[−AOD 550 nm]), and less brightening/more dimming associated with a stronger increase in AOD 550 nm (decrease in exp[−AOD 550 nm]). This relationship is strongest over Europe, where nearly all models yield a decrease in AOD 550 nm, as well as a corresponding brightening, during both time periods. Both brightening and the decrease in AOD 550 nm are largest during the 1987–2007, when observations show significant brightening exists (Figure 3). Most models, however, possess brightening trends during the 1971–1986 time period (except for GISS), in disagreement to observations. However, these models also possess decreases in AOD 550 nm during this time period. This suggests the incorrect dimming trend in most models over Europe during 1971–1986 is due to a premature decrease in AOD, which in turn is likely caused by a premature decrease in aerosol emissions (to be discussed in section 6).

Figure 8.

Scatter plots of annual mean clear sky proxy trends versus exp[−AOD 550 nm] trends during the dimming (smaller symbols) and brightening (larger symbols) periods for each of the four regions. The correlation coefficient and the corresponding significance level is included in each panel. Only models with archived aerosol data are included (Table 1). Model colors and symbols follow Figure 2.

[57] Similar results exist for China and Japan, where most models simulate an increase in AOD 550 nm over both time periods and a corresponding decrease in clear sky proxy radiation. For China, two outliers—the dark blue and orange diamonds—are prominent. These points correspond to both MRI-CGCM3 models, which simulate a small decrease/increase in clear sky proxy radiation over the 1990–2007 brightening time period and possess a decrease in AOD 550 nm. This appears to be caused by a relatively large increase in AOD 550 nm associated with Pinatubo, resulting in a negative AOD 550 nm trend from 1990 to 2007 (not shown). Compared to China, more models exhibit a decrease in AOD over Japan during its brightening time period (including the aforementioned MRI-CGCM3 models), but most continue to simulate dimming, in disagreement with the observed brightening.

[58] Over India, where only dimming is observed, all models show an increase in AOD 550 nm from 1971 to 2007. The 20 GISS realizations (p2 and p3) show relatively large increases in AOD 550 nm but small dimming. Reasons why are not clear, but removing the 20 GISS realizations (five realizations for GISS-E2-H and E2-R for p2 and p3) results in a significant increase in the correlation, from 0.35 to 0.69. We note that the GISS models also simulate relatively large increases in AOD during the dimming time period for China and Japan but generally lack correspondingly large dimming trends. The relatively large increases in GISS AOD may be related to its high sensitivity to aerosol water uptake in regions of high relative humidity, which causes strong nonlinearities in optical properties [Shindell et al., 2012, 2013].

[59] Figure 9 shows a similar analysis but based on trends in column load of sulfate. The correlation between clear sky proxy trend and trend in sulfate load is significantly negative over all four regions, ranging from −0.77 over Europe to −0.25 over Japan. Models with larger increases in sulfate load therefore possess larger decreases in clear sky proxy radiation and vice versa. Table 5shows that the corresponding correlations between clear sky proxy trend and the trend in other aerosol loads—including BC and OA—are also negative (except over Europe) but generally weaker than those based on SO4.

Table 5. CMIP5 Correlations Between the Trend in Annual Mean Clear Sky Proxy Radiation and Several Aerosol Fields, for Both the Dimming and Brightening Time Perioda
Aerosol FieldClear Sky ProxyClear Sky
EuropeChinaIndiaJapanEuropeChinaIndiaJapan
  1. a

    Also included are the corresponding correlations using clear sky radiation trends. All correlations are significant at the 99% confidence level, unless otherwise indicated.

exp(–AAOD 550 nm)–0.650.500.800.45–0.760.630.880.65
exp(–AOD 550 nm)0.770.550.350.430.950.800.870.81
exp(–fine AOD 550 nm)0.800.540.450.440.960.850.930.88
exp(–AOD 870 nm)0.570.670.900.400.890.830.970.66
BC Load0.64–0.25–0.2595–0.290.79–0.25–0.01<90–0.39
OA Load0.46–0.43–0.30–0.09<900.56–0.55–0.480.1895
SO4 Load–0.77–0.42–0.68–0.25–0.80–0.42–0.48–0.26
BC Emissions0.78–0.520.13<90–0.560.89–0.77–0.67–0.73
OA Emissions0.74–0.620.0<90–0.560.86–0.78–0.77–0.72
SO2 Emissions–0.750.15900.340.40–0.840.270.26950.44
Figure 9.

As in Figure 8, but for annual mean clear sky proxy trends versus column sulfate load trends.

[60] Scatterplots of the clear sky proxy trend versus the trend in exp[−absorption AOD 550 nm] are shown in Figure 10. All regions, except for Europe, possess significant positive correlations, with India and China featuring correlations of 0.80 and 0.50, respectively. For China, India, and Japan, both time periods feature an increase in AAOD 550 nm (decrease in exp[−AAOD 550 nm]), and a decrease in clear sky proxy radiation, for nearly every model realization. Models that simulate the largest dimming trends, such as CSIRO-Mk3-6-0 (purple X) and GFDL-CM3 (red X) also feature the largest increase in AAOD 550 nm. This is particularly apparent for China and India and suggests absorbing aerosols are important drivers of dimming over these two regions. Moreover, it suggests model underestimation of the dimming—particularly the CMIP5 mean—may be related to deficient representation of absorbing aerosol.

Figure 10.

As in Figure 8, but for annual mean clear sky proxy trends versus exp[−AAOD 550 nm] trends.

[61] Correlations between clear sky proxy trend and the trend in other aerosol fields are listed in Table 5. Also, included are the corresponding correlations based on clear sky radiation trends. Note that the relationships become stronger when trends in clear sky radiation are used. For example, the correlation between clear sky radiation trend and exp[−AOD 550 nm] trend ranges from 0.80 in China to 0.95 in Europe. These results are similar to Zubler et al. [2011], who found consistent patterns of dimming and brightening in the time series of AOD and clear sky radiation in Europe, based on simulations with the regional climate model Consortium for Small-scale Modelling in Climate Mode. An improved relationship is also found for exp[−AAOD 550 nm], where the correlation with clear sky radiation trend varies from 0.65 in Japan to 0.88 in India. Note, however, that the opposite exp[−AAOD 550 nm] relationship in Europe still occurs using clear sky radiation. We elaborate on this issue in section 6.1.

6 Interpretation of CMIP5 Dimming/Brightening

[62] There are several reasons why CMIP5 models may underestimate the magnitude of the observed dimming and brightening. Because aerosol direct effects appear to be the primary drivers of dimming and brightening, possible sources of error include (1) aerosol (SO2, OA, and BC) emissions do not increase/decrease enough; (2) modeled aerosol loads (which involve simulation of transport, aging, and deposition) do not increase/decrease enough; (3) models neglect important aerosol species (e.g., ammonium and nitrate); and (4) models are not sensitive enough to aerosols (i.e., deficient scattering and absorption of solar radiation). Similarly, incorrect model simulation of the sign of observed changes in surface solar radiation (e.g., dimming over Europe and brightening over China and Japan) is likely related to lack of, or incorrect timing of, a transition from increasing to decreasing aerosol emissions. We also note that the observed dimming/brightening trends—due to possible urbanization effects [Alpert et al., 2005; Alpert and Kishcha, 2008; Wild, 2009a]—may represent an upper bound, particularly during the dimming time period.

6.1 Europe

[63] For Europe, CMIP5 clear sky proxy trends are primarily due to changes in sulfur dioxide emissions. Figure 11 shows that CMIP5 European SO2 emissions drop drastically from ∼320 to 250 ng m−2 s−1from 1971 to 1986 (during the observed dimming time period) and then from ∼250 to 50 ng m−2 s−1from 1987 to 2008 (during the observed brightening time period). These changes are much larger than the corresponding changes for the other three regions. Moreover, Figure 9 shows the European correlation between sulfate load trends and clear sky proxy trends is −0.77, and Figure 10 shows the correlation between exp[−AOD 550 nm] trends and clear sky proxy trends is 0.77. Similarly, the European correlation between SO2 emission trends and clear sky proxy trends is −0.75 (Table 5). These relationships are much stronger than those for the other regions, and all suggest more European brightening/less dimming is associated with larger decreases in sulfate aerosol.

Figure 11.

CMIP5 ensemble mean annual mean aerosol emission time series for black carbon (BC), organic matter (OA), and sulfur dioxide (SO2) for each region. Units are ng m−2 s−1.

[64] Trends in other aerosol species (BC and OA) are less important for CMIP5 European dimming/brightening, and any effects of these aerosols on clear sky proxy trends appear to be swamped by the large sulfate trends. In fact, Table 5shows that BC load and OA load trends both possess a positive correlation with clear sky proxy trends of 0.64 and 0.46, respectively. Similarly, Figure 10 shows a negative correlation between exp[−AAOD 550 nm] trends and clear sky proxy trends of −0.65. These relationships are opposite of those based on sulfate and opposite the corresponding relationships for the other regions. Evidence that the positive correlation between BC/OA load trends and clear sky proxy trends is actually due to changes in sulfate load comes from the fact that the correlation between trends of BC and sulfate load is significantly negative at −0.50. Similarly, a negative correlation exists between trends of OA and sulfate load, where r=−0.38. Thus, models that have larger increases in sulfate load have smaller increases in BC/OA load but larger decreases in clear sky proxy radiation (and vice versa).

[65] CMIP5 underestimation of the European dimming (Figure 3) appears to be largely due to the emission inventories, which show decreasing aerosol emissions (all of them) over the entire 1971–1986 dimming time period. Similar underestimation of European dimming using ECHAM5-HAM was found by Folini and Wild [2011], where simulated dimming ended around 1970, as compared to the observed termination in the mid-1980s. Their study used the Japanese National Institute for Environmental Studies (NIES) aerosol emissions data, which shows decreases in both BC and OA from 1970 onward but a later decrease (1980s) in SO2emissions. Unlike our CMIP5 results, BC/OA emissions were found to be important contributors to the simulated European dimming in Folini and Wild [2011], particularly from 1950 to 1970. Although reasons for the early termination of the dimming in ECHAM5-HAM are more ambiguous than those for CMIP5 (since NIES SO2emissions continue increasing through the 1980s), a premature decrease in BC/OA may contribute. This emphasizes significant differences exist between emissions data sets (CMIP5 SO2emissions begin decreasing about a decade before NIES), which will have an impact on the simulated dimming and brightening trends. Similar conclusions were reached based on CMIP3 emission inventories [Ruckstuhl and Norris, 2009].

6.2 China and India

[66] Similar to Europe, Figure 9 shows sulfate load trends also correlate well with clear sky proxy trends over China and India, at −0.42 and −0.68, respectively. Unlike Europe, however, the best aerosol-clear sky proxy relationship for China and India occurs with exp[−AAOD], where a significant positive correlation exists at 0.50 and 0.80, respectively (Figure 10). Correlations between BC load trends and clear sky proxy trends are also significantly negative but weaker, at −0.25 and −0.25, respectively. Because BC load trends—as opposed to scattering aerosols like OA or SO2—are the main contributors to trends in exp[−AAOD] (all regions possess a correlation between trends of BC load and exp[−AAOD] of at least 0.61), the most important aspect of BC is not necessarily the amount but how much it absorbs. Models that feature more solar absorption by BC (and therefore possess a larger AAOD, and in turn, larger trends in AAOD) result in larger dimming. This implies the significant model underestimation of the dimming over China and India may be due to underestimation of BC solar absorption—a bias found in several other models [Sato et al., 2003; Ramanathan and Carmichael, 2008; Koch et al., 2009; Chung et al., 2012]. We also note that deficient BC emissions may contribute, since larger increases in BC would result in larger increases in AAOD. Support for deficient BC emissions is consistent with the large uncertainties, which are about a factor of 2, with uncertainty ranges of 4.3–22 Tg yr−1(for 1996) [Bond et al., 2004, 2007]. Similarly, most models underestimate BC surface concentrations [Koch et al., 2009; Menon et al., 2010] and aerosol optical depth [Chung et al., 2012].

[67] Underestimation of emissions by CMIP5 is further supported by Huneeus et al. [2012], who estimated aerosol emissions by assimilating daily total and fine mode MODIS AOD 550 nm into an aerosol model of intermediate complexity. Relative to the a priori emissions, they found the estimated emission flux for all anthropogenic species increased, ranging from 13% for SO2 to 45% for BC. More importantly, their estimated emissions are generally larger than CMIP5 emissions and suggests a possible underestimation by CMIP5. For example, global 2002 emission fluxes of BC and SO2 were estimated to be 15±13.5 and 83±25.5 Tg yr−1, compared to the CMIP5 values of 8 and 54 Tg yr−1, respectively. Thus, relative to Huneeus et al. [2012], CMIP5 SO2 emissions are ∼30% lower, and BC emissions are ∼50% lower. Similar underestimation of CMIP5 emissions also occurred over two of our regions, Europe and Asia (India was not specified).

[68] The GFDL-CM3 model yields the largest dimming trends over both China and India, in best agreement with observations. This is interesting given the fact GFDL-CM3 includes both ammonium (NH4) and nitrate (NO3). Besides inclusion of NH4 in the GISS models, these two aerosol species are not included in any of the other models. Both of these species—particularly NH4—increase over the dimming time period for both regions (not shown) and therefore contribute to the increase in AOD and the dimming signal. However, GFDL-CM3 does not have an exceptionally large increase in AOD over China (Figure 8). GISS has a relatively large increase in AOD over both China and India but lacks exceptionally large dimming. Moreover, CSIRO-Mk3-6-0 contains only three anthropogenic aerosol species—BC, OA, and SO4—yet simulates the second largest dimming over both China and India (based on models with aerosol data). This suggests the inclusion of NH4and NO3in GFDL-CM3 is not the reason for its relatively large dimming.

[69] Figure 10 shows GFDL-CM3 and CSIRO-Mk3-6-0 both possess the largest increase in AAOD over China and India during dimming. In contrast, GISS models do not possess a particularly large increase in AAOD. This seems to suggests the larger dimming in GFDL-CM3 and CSIRO-Mk3-6-0 is due to the larger increase in AAOD, and the weaker dimming in the GISS models—despite relatively large AOD trends—is due to smaller increases in AAOD. For CSIRO-Mk3-6-0, this is partially due to its relatively large trends in BC emissions and BC load, which is consistent with the fact CSIRO-Mk3-6-0 increased CMIP5 BC emissions by 25% [Rotstayn et al., 2012]. GFDL-CM3, however, does not possess particularly large increases in BC load, which suggests GFDL-CM3 simulates larger solar absorption by BC aerosols, which results in a larger increase in AAOD and more surface dimming. We note that both GFDL-CM3 and CSIRO-Mk3-6-0 treat BC as internally mixed with sulfate aerosol [Donner and et al., 2011; Rotstayn et al., 2012], which enhances BC solar absorption by up to a factor of 2 [e.g., Bond and Bergstrom, 2006; Koch et al., 2009]. However, internal mixing of BC does not guarantee larger dimming trends. For example, CESM1-CAM5 accounts for mixing of BC with other aerosol species, while also including a representation of aerosol sizes (unlike most other models where the size distribution is prescribed and only aerosol mass is computed) [Liu et al., 2012]. CESM1-CAM5, however, does not possess particularly large dimming trends for China or India (Table 3).

[70] A recent analysis of 10 models participating in the Atmospheric Chemistry and Climate Model Intercomparison Project (ACCMIP)—eight of which also participated in CMIP5—supports the notion that deficient AOD, particularly AAOD, contributes to CMIP5 underestimation of dimming over both China and India [Shindell et al., 2012, 2013]. ACCMIP models generally reproduce present-day climatological AOD relatively well, with global normalized mean biases of ∼±20% relative to AeroNet, with somewhat larger biases relative to satellites. However, most models underestimate AOD over East Asia (which includes China) by ∼40%, based on both AeroNet and satellite observations. Exceptions to underestimation of East Asian AOD include two models—one of which is GFDL-AM3—the atmospheric component of GFDL-CM3. More importantly, ACCMIP models tend to perform worse in simulating observed present-day AAOD. Strong AAOD underestimation by most models, of more than ∼50%, occurs over several regions including East Asia, South and Southeast Asia, South America, and Southern Hemisphere Africa. Not surprisingly, these are regions with large BC emissions. Again, however, the GFDL-AM3 yields the smallest overall underestimate of present-day AAOD, with relatively small AAOD underestimation over East Asia and South/Southeast Asia. Thus, GFDL-AM3 (and the CMIP5 coupled model, GFDL-CM3) has more AAOD than most models, in better agreement to observations. These results are consistent with our analysis showing the GFDL-CM3 yields the largest dimming trends over China and India, as well as the largest increases in AAOD, and with our hypothesis that the large observed dimming trends over China and India are primarily caused by absorbing aerosols. Furthermore, the significant ACCMIP underestimation of present-day AAOD in regions with high BC emissions helps to explain why all models significantly underestimate the magnitude of Chinese and Indian dimming.

[71] Model underestimation of the brightening over China (most models show continued dimming; Figure 4) appears to be due to CMIP5 emission inventories. Figure 11 shows that all aerosol emissions over China continue increasing through the brightening time period, particularly BC and OA. We note, however, that the observed brightening in China primarily occurs during the early 1990s, followed by a relatively stable period. The early 1990s also corresponds to a general replacement of the instruments in China. Several aerosol emission inventories for China show emissions (SO2 and BC/OA) peak in the mid-1990s and then decrease until the early 2000s [Streets et al., 2008; Lu et al., 2011; Qin and Xie, 2012]. This decrease does not exist in CMIP5 emissions (although SO2emissions remain flat from 1990 to 2000), which could be due to the relatively coarse, decadal resolution of CMIP5 emissions. We note, however, that after the early 2000s, all of these emission data sets show increased Chinese emissions, which seemingly opposes the lack of continued dimming in the observed clear sky proxy anomalies (Figure 4). We note a small, nonsignificant decrease in relative clear sky proxy from 2000 to 2007 of −1.3±3.0 W m−2 decade−1(Figure 4), which would fit the renewed increase in emissions.

6.3 Japan

[72] Model underestimation of brightening over Japan (most show continued dimming) is likely related to emission inventories, with little decrease in any of the aerosol emissions (Figure 11). The relatively large dimming/brightening trends, coupled with the relatively small changes in emissions over Japan, suggest part of the dimming/brightening over Japan may be related to advection of Chinese aerosols, as discussed in section 3.5. We note that differences do exist in the time series for both regions; most notably, Japanese dimming occurs primarily during the 1960s and is then followed by relatively stable values of surface solar radiation, whereas Chinese dimming is more uniform over the dimming time period. Although Japanese emissions do increase rapidly from 1960 to 1970 (particularly SO2)—consistent with the observed decrease in clear sky proxy radiation—SO2emissions rapidly decrease in the subsequent decade. Advection of Chinese aerosol may help to explain the relatively stable values of Japanese surface solar radiation after the initial dimming (Figure 6), with reductions in Japanese emissions being offset by increases in Chinese emissions. Note that Japanese BC emissions continue increasing, which also likely offsets some of the decrease in local SO2emissions. Because models underestimate dimming and brightening in China, it is not surprising that they show a similar bias over Japan. Moreover, reasons for model underestimation of Chinese dimming likely help explain model underestimation of Japanese dimming. For example, Figure 10 shows the correlation between clear sky proxy trends and exp[−AAOD] trends is also significantly positive for Japan at 0.45. Thus, model underestimation of dimming over Japan is again possibly due to underestimation of increases in AAOD, due to deficient BC solar absorption and/or deficient increases in BC emissions.

7 Conclusions

[73] This study has investigated multidecadal variability in surface solar radiation in observations and CMIP5 models over four regions—Europe, China, India, and Japan. To remove the effects of cloud cover variability on all sky surface solar radiation, we focus on a quantity called “clear sky proxy radiation,” which is obtained by removing cloud cover radiative effects from all sky radiation. Similar to prior work based on CMIP3 models [Ruckstuhl and Norris, 2009; Dwyer et al., 2010; Wild and Schmucki, 2011], we find that models generally simulate the observed reduction in downwelling surface radiation (dimming) and the observed recovery (brightening). However, this is not always the case, and significant underestimation of the trends exists for all regions except Europe. For example, over China from 1961 to 1989 and Japan from 1961 to 1984, the largest reduction in clear sky proxy radiation of any model realization is 35% and 45% as large as observed, respectively (Figures 4 and 6). Over India from 1971 to 2007, the largest reduction in clear sky proxy radiation of any model realization is ∼55% as large as observed (Figure 5). Similarly, brightening over Japan from 1985 to 2007 is significantly underestimated, with the largest model realization ∼20% as large as observed. No individual model performs particularly well overall, nor does an individual model perform best for all four regions. Corresponding biases in the CMIP5 mean are even larger (Table 2).

[74] Trends in relative clear sky proxy radiation (Table 2) are generally smaller than those based on absolute values but yield very similar conclusions. Seasonal trends yield similar dimming and brightening as the annual mean (Table 4), but most regions experience larger dimming/brightening during summer, as compared to winter, based on absolute amounts. However, this seasonal dependency does not exist for relative clear sky proxy seasonal trends, which highlights the importance of solar insolation in driving seasonal differences in dimming/brightening, particularly for China and Europe (Table 4). Over Japan, relative clear sky proxy trends during dimming are larger during winter, which we suggest is consistent with Chinese aerosols driving part of the Japanese dimming signal. Models generally produce similar seasonal conclusions.

[75] In some situations, poor model performance is likely due to deficient CMIP5 emissions. This includes underestimation of the European dimming (Figure 3), since CMIP5 emissions are all decreasing over the dimming time period (Figure 11). Model underestimation of the brightening over China (most models show continued dimming; Figure 4) also appears to be due to CMIP5 emission inventories, since all aerosol emissions over China continue increasing through the brightening time period, particularly BC and OA. This is in contrast to several other emission inventories, which show a decrease in Chinese emissions from mid-1990s to early 2000s [Streets et al., 2008; Lu et al., 2011; Qin and Xie, 2012]. Such a relatively short-term decline may not be resolved in the decadal CMIP5 inventory. Similarly, model underestimation of the brightening over Japan is also likely related to CMIP5 emissions, with little decrease in any of the aerosol emissions. Possible underestimation of emissions by the CMIP5 inventory—particularly for Europe and Asia—is also suggested by Huneeus et al.[2012].

[76] Reasons for model underestimation of the dimming over China, Japan, and India are less certain but may be due to underestimation of BC solar absorption and/or underestimation of the increase in BC emissions. All three regions possess relatively large intermodel correlations between exp[−AAOD 550 nm] trends and clear sky proxy trends, at 0.45 for Japan, 0.50 for China, and 0.80 for India. This relationship is the strongest of any of the aerosol-clear sky proxy relationships evaluated for both India and Japan. Trends in exp[−AAOD 550 nm] also correlate well with BC load trends, ranging from 0.61 over India to 0.75 over Japan. Support for deficient BC forcing comes from the fact most climate models underestimate BC solar absorption [Sato et al., 2003; Ramanathan and Carmichael, 2008; Koch et al., 2009; Chung et al., 2012] as well as BC surface concentrations [Koch et al., 2009; Menon et al., 2010] and aerosol optical depth [Chung et al., 2012]. Similarly, ACCMIP models, eight of which are included in CMIP5, strongly underestimate present-day AAOD over four regions, including China and India [Shindell et al., 2012, 2013]. The CMIP5 models that simulate the largest dimming over China and India (GFDL-CM3 and CSIRO-Mk3-6-0) possess the largest increase in AAOD. Interestingly, GFDL-AM3 exhibits the smallest present-day AAOD underestimation bias of all ACCMIP models, suggesting the CMIP5 coupled model, GFDL-CM3, possess larger aerosol absorption of solar radiation than most models, in better agreement to observations [Shindell et al., 2012, 2013]. Underestimation of AAOD is likely due to a combination of deficient BC optical properties, and underestimation of BC emissions, since total BC emission uncertainties are about a factor of 2, with uncertainty ranges of 4.3–22 Tg yr−1(for 1996) [Bond et al., 2004, 2007]. Additional work is necessary to better understand why models significantly underestimate Chinese, Indian, and Japanese dimming and the corresponding effects on simulated climate in these regions. Furthermore, better quantification of possible urbanization effects on the observational record is required.

Appendix A

[77] We may consider all sky solar flux (SWall) for each grid box as the average of solar flux under overcast conditions (SWovc) and clear sky conditions (SWclr) with weighting by the presence or absence of fractional cloud cover (CC).

display math(A1)

We can create separate equations for the long-term mean and anomalies where inline imageindicates values from the same calendar month averaged over many years, and ()indicates anomalies from the long-term monthly means.

display math(A2)
display math(A3)

We will neglect the higher order terms in equation (A3) hereafter. Equation (A2) may be transformed into

display math(A4)

where CRE is surface cloud radiative effect as conventionally defined. Equation (A4) may be substituted into equation (A3) to obtain

display math(A5)

Note that the first term on the right side of equation (A5) is what we have defined as cloud cover radiative effect (CCRE).

display math(A6)

By subtracting the clear sky flux anomaly from both sides of equation (A5), we obtain a relationship between anomalies in CRE and CCRE.

display math(A7)

An anomaly in CRE is thus the linear sum of anomalies in CCRE, overcast flux weighted by mean cloud cover and clear sky flux weighted by mean cloud cover. Overcast flux anomalies predominantly result from changes in cloud optical thickness. This factor is part of CRE but not CCRE. Clear sky flux anomalies affect CRE because they change the reference state from which cloudy-sky fluxes are compared.

[78] GEBA data provide monthly values of SWallbut not SWclr. The cloud data sets provide monthly values of CC. ISCCP provides long-term mean SWall and SWclr but not monthly values sufficiently reliable for trend analysis. Since no multidecadal record of SWclr is available on a widespread basis, we must instead infer it from other parameters. We can rewrite equation (A5) as

display math(A8)

where

display math(A9)

If no variations in overcast flux occur, then what we have defined as clear sky proxy flux (CLRproxy) is the same as clear sky flux weighted by the mean cloud-free sky fraction. Otherwise, changes in overcast flux will be included in clear sky proxy flux. Ruckstuhl et al. [2010], however, provide evidence that trends in overcast flux are much smaller than trends in clear sky flux, at least for Europe. Although we could calculate CLRproxyvia subtraction in equation (A8), we instead obtain it via linear regression in each grid box and calendar month as a residual from the best-fit line between SW'all and CCRE'. Any variability in SWovcthat linearly correlates with variability in CCRE will thereby be removed from CLRproxy.

Acknowledgments

[79] This study was funded by NSF award AGS-1131976, SIO/UCSD. We acknowledge the World Climate Research Programme's Working Group on Coupled Modeling, which is responsible for CMIP, and we thank the climate modeling groups (listed in Table 1of this paper) for producing and making available their model output. For CMIP, the U.S. Department of Energy's Program for Climate Model Diagnosis and Intercomparison provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. We thank three anonymous reviewers for their helpful suggestions.