Do sophisticated parameterizations of aerosol-cloud interactions in CMIP5 models improve the representation of recent observed temperature trends?

Authors

  • Annica M. L. Ekman

    1. Department of Meteorology, Stockholm University, Stockholm, Sweden
    2. The Bolin Centre of Climate Research, Stockholm University, Stockholm, Sweden
    Search for more papers by this author

  • This article was corrected on 3 FEB 2015. See the end of the full text for details.

Correspondence to:

A. M. L. Ekman,

annica@misu.su.se)

Abstract

Model output from the Coupled Model Intercomparison Project phase 5 (CMIP5) archive was compared with the observed latitudinal distribution of surface temperature trends between the years 1965 and 2004. By comparing model simulations that only consider changes in greenhouse gas forcing (GHG) with simulations that also consider the time evolution of anthropogenic aerosol emissions (GHGAERO), the influence of aerosol forcing on modeled surface temperature trends, and the dependence of the forcing on the model representation of aerosols and aerosol indirect effects, was evaluated. One group of models include sophisticated parameterizations of aerosol activation into cloud droplets; viz., the cloud droplet number concentration (CDNC) is a function of the modeled supersaturation as well as the aerosol concentration. In these models, the temperature trend bias was reduced in GHGAERO compared to GHG in more regions than in the other models. The ratio between high- and low-latitude warming also improved compared to observations. In a second group of models, the CDNC is diagnosed using an empirical relationship between the CDNC and the aerosol concentration. In this group, the temperature trend bias was reduced in more regions than in the model group where no aerosol indirect effects are considered. No clear difference could be found between models that include an explicit aerosol module and the ones that utilize prescribed aerosol. There was also no clear difference between models that include aerosol effects on the precipitation formation rate and the ones that do not. The results indicate that the best representation of recent observed surface temperature trends is obtained if the modeled CDNC is a function of both the aerosol concentration and the supersaturation.

1 Introduction

Anthropogenic aerosol direct (clear-sky scattering of solar radiation) and indirect (impact on cloud microphysical and radiative properties) effects are associated with large uncertainties in model projections of past and future climate [Forster et al., 2007]. There are many reasons for this uncertainty, but important factors are the large spatial and temporal variability of the aerosol population (i.e., aerosol size, composition, and number) as well as the complex interactions between aerosols and clouds [cf., e.g., Lohmann and Feichter, 2005]. Early generations of general circulation models (GCMs) disregarded temporal variability in aerosol forcing. However, GCMs have evolved rapidly during the last decades and a majority of the GCMs participating in the Coupled Model Intercomparison Project phase 5 (CMIP5, http://cmip-pcmdi.llnl.gov/cmip5/) include time-evolving concentrations (and subsequent direct radiative effects) of what are believed to be the most important aerosol components—dust, sea salt, nitrate, sulfate, particulate organic matter, and black carbon—either through explicit aerosol modules or through prescribed time-varying aerosol mass concentration fields.

Many of the CMIP5 GCMs also consider aerosol indirect effects, at least the effect of changing aerosol concentrations on cloud albedo [Twomey, 1977]. However, the actual model parameterizations of this effect vary substantially (cf. Table 1). Some models assume a simplified log-linear relationship between the aerosol mass (or parts of the aerosol mass) and the cloud droplet number concentration (CDNC) [e.g., as in Quaas and Boucher, 2005], whereas some models include more complex descriptions where the number of activated aerosols is dependent on the modeled aerosol size distribution and composition as well as the updraft velocity (e.g., using a parameterization such as [Abdul-Razzak and Ghan, 2000]). Including explicit aerosol modules to predict the characteristics of the atmospheric aerosol population at each time step and coupling these aerosol modules to sophisticated aerosol activation schemes is computationally very demanding, and the question arises: Do the more complex descriptions of aerosols and aerosol-cloud interactions improve the modeled climate?

Table 1. Short Description of the CMIP5 Models Used in the Studya
Model Horizontal     
(Ensemble Run #ModelingResolution CloudTwomeybAlbrechtc 
in GHG/GHGAERO)Center(lat×lon)AerosolMicrophysicsEffectEffectReference
  1. a

    The modeling center institute abbreviations are taken from http://cmip-pcmdi.llnl.gov/cmip5/docs/CMIP5_modeling_groups.pdf. The models are organized according to their complexity in terms of aerosol and aerosol-cloud parameterizations. Models with only one ensemble member (i.e., not included in the main part of the analysis) are marked by italic font.

  2. b

    The number of CDNC is dependent on the aerosol concentration and/or composition.

  3. c

    The precipitation formation (autoconversion) is dependent on the number of cloud droplets.

  4. d

    The number of CDNC is dependent on the calculated aerosol concentration and composition and the water vapor saturation as in Abdul-Razzak and Ghan [2000].

  5. e

    The number of CDNC is dependent on the calculated aerosol concentration and composition and the water vapor saturation as in Ming et al. [2006].

  6. f

    Diagnostic-A: Based on an empirical function between all total hydrophilic aerosol and CDNC. In GISS-E2-H/R, CDNC is semiprognostic: the activation is determined through an empirical scheme, but there are sinks for CDNC that includes droplet loss through freezing. In HadGEM2-ES, dust aerosols are not contributing to CDNC.

  7. g

    Diagnostic-S: Based on an empirical function between sulfate and CDNC.

CESM1-CAM5-1-FV2NSF-96×144OnlineTwo-momentPrognostic,dYesMeehl et al. [2013]
(2/4)DOE-      
 NCAR      
GFDL-CM3NOAA-90×144OnlineOne-momentPrognosticeYesDonner et al. [2011]
(3/5)GFDL      
MIROC-ESMMIROC64×128OnlineTwo-momentPrognosticdYesWatanabe et al. [2011]
(3/3)       
NorESM1-MNCC96×144OnlineTwo-momentPrognosticdYesKirkevåg et al. [2013]
(1/3)       
FGOALS-g2LASG-60×128PrescribedTwo-momentPrognosticdYesLi et al. [2013]
(1/5)CESS      
CSIRO-MK3-6-0CSIRO-96×192OnlineOne-momentDiagnostic-AfYesRotstayn et al. [2012]
(5/10)QCCCE     
GISS-E2-HNASA-90×144PrescribedOne-momentDiagnostic-AfYesMenon et al. [2010]
(5/5)GISS     
GISS-E2-RNASA-90×144PrescribedOne-momentDiagnostic-AfYesMenon et al. [2010]
(5/6)GISS      
HadGEM2-ESMOHC145×192OnlineOne-momentDiagnostic-AfYesBellouin et al. [2011], Collins et al. [2011]
(4/4)      
CNRM-CM5CNRM-128×256PrescribedOne-momentDiagnostic-AfNoVoldoire et al. [2012]
(6/10)CERFACS      
IPSL-CM5A-LRIPSL96×96PrescribedOne-momentDiagnostic-AfNoDufresne et al. [2013]
(3/6)       
CanESM2CCCMA64×128OnlineOne-momentDiagnostic-SgNoGillett et al. [2012]; Arora et al. [2011]
(5/5)       
BCC-CSM1-1BCC64×128PrescribedOne-momentNoNoWu et al. [2008]
(1/3)       
BNU-ESMGCESS64×128PrescribedOne-momentNoNohttp://esg.bnu.edu.cn
(1/1)       
CCSM4NCAR192×288PrescribedOne-momentNoNoGent et al. [2011]
(3/6)       

The subset of CMIP5 models analyzed in the present study treats aerosols and aerosol indirect effects with different degrees of complexity. By categorizing the models with respect to their description of the relation between the aerosol population and the CDNC, the effect of different representations of aerosol indirect effects could be evaluated. In one group of models, the activation of aerosols into cloud droplets is treated explicitly so that the modeled CDNC is dependent on both the supersaturation as well as the aerosol concentration and composition. In another group of models, the CDNC is determined diagnostically by empirical relations between the aerosol concentration and the CDNC. A third type of model includes no effect of changing aerosol concentrations at all on the modeled CDNC. A comparison was also made between models with similar complexity regarding aerosol effects on CDNC but with different representations of the aerosol population (explicit aerosol module versus prescribed aerosol fields) and aerosol effects on the precipitation formation rate (cloud lifetime or Albrecht effect).

The strongest fingerprint from changing aerosol concentrations should be seen in net radiative fluxes and surface temperatures. Wilcox et al. [2013] also showed that the time evolution of the historical global mean temperature trend was better represented in CMIP5 models that include aerosol indirect effects compared to models with only a representation of the direct effect. In the present study, the impact of anthropogenic aerosol forcing on the spatial distribution of surface temperature trends was examined. More specifically, the mean global (70°S to 70°N), tropical (25°S to 25°N), and Northern Hemisphere (NH) midlatitude (30°N to 60°N) and NH high-latitude (60°N to 70°N) surface temperature trends in CMIP5 model simulations that included or excluded temporal changes in anthropogenic aerosol concentrations were compared with observations. In other words, in contrast to Wilcox et al. [2013], the latitudinal distribution of the temperature trend, rather than the global average, is investigated in more detail.

As the anthropogenic aerosol forcing varies strongly with latitude, a reasonable treatment of anthropogenic aerosol effects on surface temperature should imply that not only the global bias is reduced compared to observations but also the surface temperature trends within different regions are better represented than if only greenhouse gas forcing is considered. In the present study, a model was considered to improve if the simulation with changing anthropogenic aerosol concentrations meant that the model was closer to observations both in terms of average trends for the different regions, as well as in terms of the ratio between high- and low-latitude trends, compared to the simulation with only greenhouse gas forcing. It is assumed that the ratio between middle- and high-latitude temperature trends and tropical temperature trends will further demonstrate the ability of a model to capture meridional variations in temperature trends.

It should be noted that improving the temperature bias may not be the main motivation for including complex aerosol modules and aerosol-cloud parameterizations in GCMs. Understanding physical processes in the atmosphere as well their feedback on climate change may be at least as important. The paper is organized as follows. In section 2, the model output and temperature observations are described. Section 3 compares observed temperature trends for the years 1965 to 2004 with the simulated temperature trends over the same time period, with and without anthropogenic aerosol effects, respectively. Section 4 presents a discussion on whether the simulations are improved or not when a change in aerosol radiative forcing is considered and if models of similar complexity regarding their representation of aerosols and aerosol indirect effects show similar results. In section 5, the results are summarized and conclusions are drawn.

2 Models and Observations

The focus lies on the Northern Hemisphere and the time period between 1965 and 2004. During this time period and over this hemisphere, there has been a rapid change in the magnitude and spatial distribution of anthropogenic aerosol particle and precursor emissions with a significant decrease over Europe and North America and a substantial increase over large parts of Asia [Lamarque et al., 2010]. At the same time, the spatial coverage of surface temperature observations has been relatively good. The emphasis of the evaluation was not on examining how well the reference version of each model per se reproduced the observed surface temperature trends but rather on whether each model ensemble improved or not (i.e., if an individual model ensemble mean and spread was closer to observations) when changes in anthropogenic aerosol forcing were included compared to when only greenhouse gas forcing was considered.

Temperature trends for the time period 1965 to 2004 were calculated from both observational data and model output. The observed surface temperature trends were obtained from HadCRUT4 [Morice et al., 2012]. This data set is based on station, buoy, and ship observations and also provides uncertainty information, retrieved from an ensemble of 100 simulations sampling the estimated observational uncertainty. For the observed temperature trend calculations, the 5th and 95th percentiles of the ensemble provided by HadCRUT4 were utilized to calculate a range of possible temperature trends (cf. Figures 1 and 2).

Figure 1.

(a) Global (70°S to 70°N), (b) tropical (25°S to 25°N), and (c) midlatitude (30°N to 60°N) and (d) high-latitude (60°N to 70°N) temperature trends (K decade−1) between 1965 and 2004 in GHG (left marker) and GHGAERO (right marker) simulations. The observed average value (cf. Table 2) is marked by the gray dashed line and the gray shaded region depicts the 5th to 95th percentile ranges of the HadCRUT4 ensemble (cf. section 2). The whiskers represent the maximum and minimum of each model ensemble. Models with no representation of aerosol indirect effects are marked with a “+”, and models with prognostic aerosol activation (cf. Table 1) are marked with a “o”.

Figure 2.

Ratio between (a) midlatitude, (b) high-latitude, and tropical temperature trends between 1965 and 2004 in GHG (left marker) and GHGAERO (right marker) simulations. The observed average value (cf. Table 2) is marked by the gray dashed line, and the gray shaded region depicts the 5th to 95th percentile ranges of the HadCRUT4 ensemble (cf. section 2). The whiskers the maximum and minimum of each model ensemble. Models with no representation of aerosol indirect effects are marked with a “+”, and models with prognostic aerosol activation (cf. Table 1) are marked with a “o”.

Coupled ocean-atmosphere model output was downloaded from the CMIP5 archive (cmip-pcmdi.llnl.gov/cmip5). In order to compare the difference in modeled temperature trends with and without aerosols, the “Historical” (all natural and anthropogenic forcings, hereafter referred to as “GHGAERO”) and “HistoricalGHG” (long-lived greenhouse gas changes only, hereafter referred to as “GHG”) simulations from CMIP5 were compared. Details on the different CMIP5 model experiments can be found on cmip-pcmdi.llnl.gov/cmip5/. In the GHG experiment, only greenhouse gas concentrations are allowed to evolve in time while emissions of aerosols and other forcing variables are set to pre-industrial values.

Table 1 summarizes the main characteristics (with respect to their treatment of aerosols and aerosol direct and indirect effects, according to the cited references) of the 11 models that provide more than one simulation (ensemble member) for both the GHGAERO and GHG scenarios. The number of ensemble members and the horizontal grid resolution are also given in Table 1. Note that for each individual model, the number of simulations provided for GHG may be different to that provided for GHGAERO. The 11 models in Table 1 with multiple ensemble members are the ones that will be evaluated in the main part of the analysis of this study. However, for completeness, four models with only one ensemble member are also included in the tables (presented with italic font) and in the summarizing discussion of the results (section 4).

Unfortunately, there is only one model with multiple ensemble members that excludes any type of indirect aerosol effects (CCSM4). In CCSM4, the effective droplet radius at surface is prescribed to one value over polluted areas (most land surfaces) and one value over pristine areas (such as over the ocean, sea ice, and snow covered land). For the present study, it would certainly have been useful with a number of models such as CCSM4, i.e., models that only include aerosol direct effects and that have more than one ensemble member. However, this was not available from the CMIP5 archive. By including models with only one ensemble member in the discussion and tables, two more models (BCC-CSM1-1 and BNU-ESM) with no aerosol indirect effects could be added to support the overall conclusions.

In the present study, it will be assumed that the difference between GHGAERO and GHG is mainly due to anthropogenic aerosol forcing. It should be noted that the GHGAERO simulations include, in addition to changes in anthropogenic aerosol forcing, changes in natural (e.g., volcanic) aerosols and other anthropogenic forcing parameters such as land-use and ozone concentrations. However, the radiative forcing due to changes in anthropogenic aerosol emissions is in general much larger than the forcing due to changes in tropospheric ozone, land-use change, and natural aerosols [cf., e.g., Forster et al., 2007]. Four models that have reported results for simulations including only anthropogenic aerosol forcing to the CMIP5 archive (CanESM2, CSIRO-Mk3-6-0, GISS-E2-H, and GISS-E2-R) were also investigated in more detail. In these simulations, aerosol emissions (or concentrations) are allowed to vary in time whereas all other forcing variables are set to preindustrial values. The temperature trends in the aerosol forcing-only simulations can be compared with the residual between GHGAERO and GHG to evaluate the assumption that the residual is mainly governed by changes in anthropogenic aerosol emissions. The correlation coefficients of the global, tropical, and midlatitude and high-latitude temperature trends (for definitions of the latitude bands, cf. section 3.2) between the only anthropogenic aerosol-forcing simulations and the residual between GHGAERO and GHG are high (r2=0.8) supporting the assumption that the residual between GHGAERO and GHG can be used as a proxy for temperature trends caused by anthropogenic aerosol forcing. In addition, Wilcox et al. [2013] showed that for the twentieth century time evolution of global-mean near-surface temperature, the sum of the temperature anomalies due to greenhouse gases, natural, and anthropogenic aerosol only forcing gave a good reproduction of the anomaly from an all forcing run. M. Stolpe et al. (Aerosol forcing, 20th century warming and the climate feedback parameter, submitted to Geophysical Research Letters, 2013) also showed that the total historical forcing during the twentieth century and the sum of the CO2 and aerosol forcing are well correlated (r2=0.7).

Temporal variations in the natural aerosol burden could also affect the residual between GHGAERO and GHG. The eleven models included in the main part of the evaluation (i.e., the ones with more than one ensemble member for both GHG and GHGAERO) all provide “HistoricalNat” simulations to the CMIP5 archive, i.e., simulations where only natural forcing is applied. Only three of these models display a negative global average temperature trend due to natural forcing during the period 1965 to 2004. The most likely reason for this overall positive temperature trend in the models is that during the beginning of the time period of study, the simulated surface temperatures are still “recovering” from the Mount Agung eruption in 1963. When plotting the temperature trends due to natural forcing versus the temperature trends of the residual between GHGAERO and GHG, the correlation is poor (r2=0.04). This result is consistent with Wilcox et al. [2013], who showed that the contribution to temperature trends from natural factors for the period 1965–2004 was very small compared to that from greenhouse gases and anthropogenic aerosols, and confirms that the main difference between GHGAERO and GHG is due to anthropogenic aerosols and not due to natural forcing from, e.g., volcanic aerosols.

Four of the models in Table 1 simulate aerosol chemistry and microphysics online, whereas the other models utilize precalculated aerosol fields, typically provided as 10 year averages and derived with the same atmospheric model as the fully coupled ocean-atmosphere GCM, but with prescribed historical sea surface temperatures as a lower boundary.

Three models have prognostic cloud droplet number concentrations (CDNC) (MIROC-ESM, CESM1-CAM5-1-FV2, and GFDL-CM3), where the calculated CDNC is dependent on aerosol composition, size, and number concentration as well as on water vapor saturation. In MIROC-ESM and CESM1-CAM5-1-FV2, the cloud microphysics module predicts both the number and mass of cloud droplets and ice crystals (two-moment scheme). Aerosols can also influence the ice crystal number through heterogeneous freezing.

The other models (except CCSM4) utilize empirical relationships to diagnose the relation between the aerosol mass concentration and CDNC. In other words, there are three models (HadGEM2-ES, CSIRO-Mk3-6-0, and CanESM2) that simulate aerosol chemistry and physics online, but they still determine the CDNC through an empirical formula. In four models (IPSL-CM5A-LR, CNRM-CM5, CanESM2, and CCSM4), the precipitation formation rate (autoconversion) is not dependent on the CDNC; i.e., the aerosol effect on cloud lifetime is not included [Albrecht, 1989].

3 Results

3.1 Observed and Modeled Temperature Trends Between 1965 and 2004

Temperature trends were calculated using linear regression over the years 1965 to 2004 for both model output and observations. Time series of simulated global average surface temperature trends and the impact of different forcings can be found in, e.g., Wilcox et al. [2013]. The simulated temperature trend for the GHGAERO simulations, i.e., the “best guess” from the respective model centers, was compared with observations (Table 2 and right markers for all model output presented in Figures 1 and 2).

Table 2. Observed and Modeled (GHGAERO Simulations) Global (70°S to 70°N), Tropical (25°S to 25°N), Midlatitude (30°N to 60°N) and High-Latitude (60°N to 70°N) Average Temperature Trends (K decade−1), and Midlatitude and High-Latitude Amplification Factors (Compared to the Tropics)a
Model/Obs    MidlatitudeHigh-latitude
1965–2004GlobalTropicalMidlatitudeHigh-LatitudeAmplificationAmplification
  1. a

    All values are area weighted. Models with only one ensemble member are marked with italic font. Modeled averages that are lower than observed averages are marked with boldface font.

HadCRUT40.1400.1290.2300.3151.792.45
Multimodel average0.1880.1630.2820.4691.782.97
CESM1-CAM5-1-FV20.1540.1370.2350.2171.711.58
GFDL-CM30.1850.1620.2700.5011.693.10
MIROC-ESM0.1650.1350.2660.4091.983.04
NorESM1-M0.1270.1160.2040.1471.761.27
FGOALS-g20.1710.1680.2640.3551.572.11
CSIRO-MK3-6-00.1600.1250.3020.4512.413.61
GISS-E2-H0.1530.1510.1800.4361.202.89
GISS-E2-R0.1570.1590.2240.4491.412.83
HadGEM2-ES0.1670.1310.3110.5402.374.11
CNRM-CM50.1710.1420.2730.5071.923.58
IPSL-CM5A-LR0.2760.2670.3560.4511.331.69
CanESM20.2450.2220.3630.5851.632.63
BCC-CSM1-10.2180.1510.3320.6222.194.11
BNU-ESM0.2450.1690.4360.5812.593.45
CCSM40.2220.1660.3240.6091.953.66

Global trend. According to the HadCRUT4 data set, the global mean surface temperature trend between 1965 and 2004 was approximately 0.140 K decade−1(Table 2). The ensemble mean global mean temperature trends in the GHGAERO simulations is in all models higher compared to the observations. However, 8 of 11 models have trends whose uncertainty falls within that of the observations (Figure 1a). The multimodel average values in Table 2 show that the overall larger warming in the models mainly arises from a larger temperature increase at NH high latitudes, as this is where the bias compared to observations is the highest (almost 50% as a multimodel mean compared to around 25% in the tropics and at midlatitudes). It is only in CESM-CAM5-1-FV2 that the high-latitude ensemble mean temperature trend is lower than the mean in the observations.

Figure 1a shows that the ensemble range varies quite substantially between the different models. There seems to be no direct relation between the number of ensemble members and the ensemble spread, neither is it necessary so that the models with a more sophisticated aerosol activation parameterization (CESM-CAM5-1-FV2, GFDL-CM3, and MIROC-ESM) have a larger spread than the other models or that the GHGAERO simulations display a larger range than the GHG simulations.

Tropical trend. In the tropics, nine models provide ensemble values that fall within the observed uncertainty range (Figure 1b), but their ensemble means are in general higher than observations. The ensemble ranges of the two models with the largest biases, IPSL-CM5A-LR and CanESM2 (bias of 107 and 72%, respectively), do not fall within the uncertainty of the observations. It is likely that at least part of the overestimate by IPSL-CM5A-LR is due to an incomplete treatment of the radiative effects of volcanic aerosols [cf. Santer et al., 2013]. The large positive bias in CanESM2 may be related to the fact that the CDNC in this model is only a function of sulfate and not the complete aerosol population (Table 1). A large part of the increase in the aerosol loading over Asia (and thus CDNC) is for example caused by carbonaceous aerosols.

NH midlatitude trend. At midlatitudes, six models provide ensemble values that fall within the uncertainty of the observations (MIROC-ESM, CESM-CAM5-1-FV2, GFDL-CM3, GISS-E2-H, GISS-E2-R, and CNRM-CM5). Five models clearly overestimate the observed temperature trend in the GHGAERO simulations (HadGEM2-ES, CSIRO-Mk3-6-0, IPSL-CM5A-LR, CanESM2, and CCSM4).

NH high-latitude trend. At high-latitudes, only ensemble members of four models fall within the observed temperature range (MIROC-ESM, CESM-CAM5-1-FV2, GFDL-CM3, and CSIRO-Mk3-6-0). The rest of the models predict a temperature trend that is larger than the observations, with an overestimate up to 93% (CCSM4).

NH middle- and high-latitude amplification. The models performing well in the tropics do not necessarily perform equally well at middle or high latitudes. According to the observations, the ratio between the midlatitude and tropical temperature trend is approximately 1.79 (Table 2) and a majority of the models (8 out of 11) have ensemble members that can represent a ratio that falls within the uncertainty of the observations (Figure 2). The observed ratio between the high-latitude and tropical temperature trend is calculated to be 2.45. The multimodel average is an overestimate relative to the observations. Most models (9 out of 11) also overestimate the ratio, but five models display ensemble members that are within the observed uncertainty (MIROC-ESM, GFDL-CM3, GISS-E2-H, GISS-E2-R, and CanESM2).

3.2 The Influence of Anthropogenic Aerosols on Modeled Temperature Trends Between 1965 and 2004

In this subsection, the temperature trends in GHG (left markers for all model output presented in Figures 1 and 2) and GHGAERO (right markers for all model output presented in Figures 1 and 2) are compared and contrasted with observations.

Global anthropogenic aerosol influence. Table 3 shows the anthropogenic aerosol influence (obtained by subtracting the GHGAERO from the GHG simulation) on the global mean temperature trend between 1965 and 2004. All models, except CCSM4, agree on the sign of the global mean temperature change due to aerosols. Three models, however, predict a temperature trend less than 0.01 K decade−1(MIROC-ESM, GISS-E2-R, and CanESM2). CanESM2 and CCSM4 both display a large warming due to aerosols at high latitudes, whereas MIROC-ESM and GISS-E2-R mainly show a warming over the Southern Hemisphere. The inclusion of changing aerosol concentrations also reduces the global temperature bias compared to observations in all models except CCSM4, which is the only model without a representation of aerosol indirect effects (Figure 1a). In five models, the inclusion of aerosol direct and indirect effects brings the model range of temperature trends into agreement with the observed uncertainty. (CESM1-CAM5-1-FV2, CNRM-CM5, GFDL-CM3, GISS-E2-H, and HadGEM2-ES).

Table 3. Area-Weighted Difference in the Simulated Net Surface Radiative Flux Rad (W m−2Decade−1×10−2) and Surface Temperature Temp (K Decade−1×10−2) Trends (1965 to 2004) Between the GHGAERO and GHG Simulations (Taken to be Mostly Driven by Anthropogenic Aerosol Forcing, cf. Section 2)a
 GlobalTropicalMidlatitudeHigh-latitude
 (70°S to 70°N)(25°S to 25°N)(30°N to 60°N)(60°N to 70°N)
  1. a

    Positive trends are marked by boldface font. Models with only one ensemble member are marked with italic font.

Model        
1965-2004RadTempRadTempRadTempRadTemp
CESM1-CAM5-1-FV2−28.2−7.8−38.6−5.5−10.4−7.2−20.6−17.5
GFDL-CM3−29.0−11.8−51.7−12.9−14.0−16.5−31.9−25.6
MIROC-ESM−8.1−0.6−13.31.4−1.9−1.6−10.0−2.2
NorESM1-M-11.9-3.2-21.7-4.4-17.4-4.3-0.3-24.2
FGOALS-g20.1-3.3-8.3-2.810.0-2.0-5.6-13.1
CSIRO-Mk3-6-0-22.4-4.1-44.0-7.09.24.29.86.8
GISS-E2-H-14.1-4.7-19.0-3.6-27.9-10.25.9-8.3
GISS-E2-R-11.0-0.9-17.2-0.8-18.0-2.610.10.8
HADGEM2-ES-10.9-8.3-18.1-8.6-7.0-9.5-4.4-10.4
CNRM-CM5-20.9-4.5-39.0-3.73.5-3.91.6-0.8
IPSL-CM5A-LR-10.0-1.5-17.4-0.8-26.8-7.9-0.3-3.0
CanESM26.3-0.4-0.1-0.518.74.619.88.2
BCC-CSM1-14.87.2-2.94.52.92.525.15.3
BNU-ESM-14.82.4-32.06.82.29.72.42.4
CCSM4-0.12.6-4.67.5-1.70.032.119.2

Tropical anthropogenic aerosol influence. Focusing on the tropics, all models agree on that the change in aerosol emissions between 1965 and 2004 (i.e., the difference between GHGAERO and GHG) has caused a decrease in the net radiative flux at the surface (Table 3). Despite this, two models (MIROC-ESM and CCSM4) still predict a change in the overall tropical surface temperatures that is positive or close to zero. The zero change in surface temperature due to aerosols in MIROC-ESM is due to a slight general warming in the tropical SH, whereas the NH tropics experience a slight cooling (cf. the difference in zonally averaged net surface radiative flux trends between GHGAERO and GHG displayed in Figure 3), i.e., in agreement with the majority of models. In CCSM4, the temperature change is close to zero in most parts of the tropics (Figure 3) but becomes positive around 18°N and further north. In all models except CCSM4, the temperature change in the tropics due to aerosols also reduces the bias compared to observations (Figure 1b). In CCSM4, the bias is unchanged. In 5 of the 11 models (CESM1-CAM5-1-FV2, GFDL-CM3, HadGEM2-ES, GISS-E2-H, and CNRM-CM5), including aerosol effects is necessary in order to have ensemble members that are within the uncertainty of the observed tropical temperature trend.

Figure 3.

Zonally averaged net surface radiative flux (gray shading, right label) and surface temperature (colored shading, left label) trend (1965–2004) difference between GHGAERO and GHG simulations (taken to be mostly driven by anthropogenic aerosol forcing, cf. section 2). Full lines denote ensemble average and shading ensemble spread. Models with no representation of aerosol indirect effects are marked with a “+”, and models with prognostic aerosol activation (cf. Table 1) are marked with a “o”.

NH midlatitude anthropogenic aerosol influence. At midlatitudes, three models (CSIRO-Mk3-6-0, CNRM-CM5, and CanESM2) display a positive net surface radiative flux due to changes in aerosol concentrations (Table 3 and Figure 3). CSIRO-Mk3-6-0 and CanESM2 also display a warming of midlatitude surface temperatures, whereas CNRM-CM5 displays a cooling, most likely due to advection of cooler air from the tropics. The rest of the models display a negative net surface radiative flux change due to aerosols. This feature typically leads to a subsequent surface cooling, but in CCSM4, the mean midlatitude surface temperature change is still positive. In this model, the net surface radiative flux change over land is positive whereas the flux change over the ocean is negative. The land surface temperature is more sensitive to radiative flux changes than the ocean temperature, and thus, the mean surface temperature change in CCSM4 is positive. Other models, e.g., CESM-CAM5-1-FV2, also display a clear contrast between midlatitude ocean and land changes in radiative fluxes. However, in contrast to CCSM4, these models display a clear surface cooling in the tropics, which most likely acts to cool the overall midlatitude surface temperature.

In all models that display a negative temperature trend due to aerosols, the temperature bias of the ensemble mean for each model is reduced (Figure 1c), except in GISS-E2-H where the midlatitude ensemble mean bias is slightly worse and also changes sign when aerosol effects are included. None of the models that display a positive temperature trend due to aerosols (CSIRO-Mk3-6-0, CanESM2, and CCSM4) have a reduced bias.

Figure 4 shows that the sign of the mean midlatitude net surface radiative flux is a subtle balance between a strong positive forcing over Europe and north America (caused by a decrease in aerosol optical depth, AOD, and CDNC) and a strong negative forcing over northeast Asia and the north Pacific Ocean (primarily caused by an increase in AOD and CDNC). The three models that generate a mean positive midlatitude net surface radiative flux (CanESM2, CSIRO-Mk3-6-0, and CNRM-CM5) are thus compared to the other models at midlatitudes less influenced by changes in aerosol emissions over Asia than emission changes over Europe and North America. It is worth noting that these three models all include the Twomey effect but exclude the Albrecht effect (Table 1). However, it is not clear if this is the reason why the models are more influenced by changes in aerosol emissions over Europe than over Asia. It is also worth noting that all models with a fully prognostic treatment of the CDNC (MIROC-ESM, CESM1-CAM5-1-FV2, and GFDL-CM3) display relatively strong overall midlatitude cooling compared to the other models. A prominent feature in these three models, although not unique, is a strong cooling over the North Pacific Ocean (40–60°N, cf. Figure 5).

Figure 4.

Difference in net surface radiative flux trend (1965 to 2004) between GHGAERO and GHG simulations (taken to be mostly driven by anthropogenic aerosol forcing, cf. section 2) [Wm−2 decade−1]. Trends significant at 5% level are displayed in full color, while nonsignificant trends are masked with a gray filter. Models with no representation of aerosol indirect effects are marked with a “+”, and models with prognostic aerosol activation (cf. Table 1) are marked with a “o”.

Figure 5.

Difference in net surface temperature trend (1965 to 2004) between GHGAERO and GHG simulations (taken to be mostly driven by anthropogenic aerosol forcing, cf. section 2) [K decade−1]. Trends significant at 5% level are displayed in full color, while nonsignificant trends are masked with a gray filter. Models with no representation of aerosol indirect effects are marked with a “+”, and models with prognostic aerosol activation (cf. Table 1) are marked with a “o”.

NH high-latitude anthropogenic aerosol influence. At high latitudes, over half of the models (6 out of 11) display a positive net surface radiative flux change due to changes in aerosol emissions (Table 3). However, even with a positive surface radiative flux change, one of the models (CNRM-CM5) displays a net surface cooling, which most likely is due to advection of cooler air from the midlatitudes. Five models (MIROC-ESM, CESM1-CAM5-1-FV2, GFDL-CM3, HadGEM2-ES, and IPSL-CM5A-LR) display a negative net surface radiative flux change and subsequent cooling above 60°N. It is noteworthy that once again all models with a prognostic treatment of the CDNC display a negative net surface radiative flux due to aerosols and subsequent cooling. In general, the bias compared to observations is reduced in all the models with a negative temperature trend due to aerosols. Despite an improvement in the temperature bias when including aerosol effects, only four models (MIROC-ESM, CESM1-CAM5-1-FV2, GFDL-CM3, and CSIRO-Mk3-6-0) have ensemble members that are within the uncertainty of the observed polar temperature trend.

Anthropogenic aerosol influence on NH middle- and high-latitude amplification. Figure 2 shows that when including aerosol effects, the ratio between midlatitude and tropical temperature trends increases in most of the models (7 out of 11). Four models (MIROC-ESM, CESM1-CAM5-1-FV2, GFDL-CM3, and CanESM2,) display an improvement of the ensemble mean bias compared to observations when aerosol effects are included. In five models (HadGEM2-ES, GISS-E2-H, GISS-E2-R, CNRM-CM5, and CCSM4), the inclusion of aerosol effects results in a somewhat worse bias of the ensemble mean compared to the ensemble where only changes in greenhouse gas concentrations are considered, but there are still members of the model ensemble that are within the observed uncertainty.

For high latitudes, the ratio between the high-latitude and tropical temperature trend also increases in most of the models (8 out of 11) when including aerosol effects (Figure 2). Only two models (MIROC-ESM and CanESM2) display a reduced ensemble mean bias. In five models (CESM1-CAM5-1-FV2, GFDL-CM3, GISS-E2-H, and GISS-E2-R), the ensemble mean bias is not reduced when changes in aerosol concentrations are considered, but the ensembles have members that are within the observed uncertainty. Note that it is not certain that a model which shows a reduced bias in the ratio between high-latitude or midlatitude and tropical warming also shows a reduced bias in the temperature trend itself (cf., e.g., CanESM2).

4 Discussion

Table 4 (and Table S1) summarizes the effect of including temporally varying anthropogenic aerosol emissions in the different models, i.e., the difference when going from the GHG to the GHGAERO simulation ensemble. It is clear that the observed temperature trends are better represented in the GHGAERO ensemble than in the GHG ensemble in the models which include a prognostic relation between the aerosol and CDNC (MIROC-ESM, CESM1-CAM5-1-FV2, and GFDL-CM3), as most of the surface temperature trend statistics are improved in GHGAERO compared to GHG. In addition, the amplification factors are in general better represented in the models with prognostic activation when anthropogenic aerosols are considered, which means that it is not only the overall magnitude of the bias compared to observations that is reduced but also the meridional distribution of the temperature trend is better represented. The temperature trend bias in the models with prognostic aerosol activation but only one ensemble member (NorESM1-M and FGOALS-g2) is also in general reduced when temporal variations in aerosol forcing are considered. However, NorESM1-M simulates a very large cooling due to anthropogenic aerosols at midlatitudes and, in particular, high latitudes (not shown).

Table 4. Comparison Between the GHGAERO and GHG Ensemblesa
ModelGBGRTBTRMBMRHBHRMAMARHAHARinline image
  1. a

    If the inclusion of aerosol effects improves the bias (GB = global bias, TB = tropical bias, MB = midlatitude bias, HB = high-latitude bias, MA = midlatitude/tropical ratio bias, HA = high-latitude/tropical ratio bias), the box is marked with a “ +”. A worsening is marked with a “−”. Similarly, if the model ensemble covers the observed range (GR = global range, TR = tropical range, MR = midlatitude range, HR = high-latitude range, MAR = midlatitude/tropical ratio range, HAR = high-latitude/tropical ratio range), the box is markedwith a “ +”, etc. The last column shows the total number of “ +” minus the total number of “−”. Models with only one ensemble member are marked with italic font.

CESM1-CAM5-1-FV2+++++++++6
GFDL-CM3+++++++++++10
MIROC-ESM++++++++++++12
NorESM1-M+++++++2
FGOALS-g2+++++++++6
HADGEM2-ES+++++++2
CSIRO-Mk3-6-0+++++++2
GISS-E2-H+++++++2
GISS-E2-R++++++++4
IPSL-CM5A-LR+++++++2
CNRM-CM5++++++++4
CanESM2++++++0
BCC-CSM1-1+++−6
BNU-ESM+−10
CCSM4+++−6

The magnitudes of the zonal mean net surface radiative flux change and surface temperature change due to anthropogenic aerosols are quite different in the three models with prognostic aerosol activation (Figure 3). MIROC-ESM displays one of the lowest differences between GHGAERO and GHG over the Northern Hemisphere, whereas GFDL-CM3 has one of the highest. When comparing the latitudinal distribution of the trends, there are however a few common features; the three models with prognostic CDNC all show a surface cooling due to aerosols in the Northern Hemisphere that increases with increasing latitude, while the net surface radiative flux change is the largest in the tropics.

Comparing the spatial distribution of the surface radiative flux perturbation (Figure 4), it is also clear that the three models with prognostic CDNC in general show a larger contrast between the strong negative perturbation region over South and Eastern Asia and the strong positive perturbation region over Europe compared to the other models. In addition, the zonal variability in the net surface radiative perturbation change due to aerosols at midlatitudes and high latitudes is more pronounced in the three models with prognostic CDNC. The pattern of surface temperature change due to aerosols also displays some similarity between the three models with prognostic CDNC; there is a relatively large cooling over the northernmost Pacific and northeast Siberia, a prominent cooling over Northeast America, a warming over the north Atlantic and a slight (but generally nonsignificant) warming over southeastern U.S. (Figure 5). None of the other models show all these features simultaneously.

Table 4 also shows that the four models with a diagnostic relation between aerosols and CDNC benefit less than the models with prognostic activation from including changing anthropogenic aerosol concentrations, but they still benefit more than the model without any aerosol indirect effects at all (CCSM4). This result is also corroborated when models with one ensemble member and no aerosol indirect effects (BCC-CSM1-1 and BNU-ESM) are considered. It is not simply so that the models with aerosol indirect effects always cool more, or warm less, than the models with no indirect effects. CanESM2 and CSIRO-Mk3-6-0, for example, simulate a larger warming than BCC-CSM1-1 due to changing anthropogenic aerosol concentrations at midlatitudes and a larger warming than both BCC-CSM1-1 and BNU-ESM at high latitudes. CanESM2 and CSIRO-Mk3-6-0 also display a positive surface net radiative flux perturbation due to aerosols at midlatitudes, whereas CCSM4 display a negative surface net radiative flux perturbation. However, for the models that include a diagnostic relation between aerosols and CDNC, the four models that improve most significantly when including aerosols (HadGEM2-ES, CNRM-CM5, GISS-E2-H, and GISS-E2-R) all show (similarly to the models with prognostic CDNC) a general cooling due to aerosols over the Northern Hemisphere (Figure 3), a strong negative surface radiative flux perturbation over South and Eastern Asia (Figure 4), and a larger negative temperature trend due to aerosols over the northernmost Pacific and northeast Siberia than the other models (Figure 5).

The two versions of the GISS model, GISS-E2-H and GISS-E2-R, differ only by the ocean model applied, whereas the descriptions of aerosols and aerosol indirect effects are identical in the two models. Both models show a similar distribution of the net surface radiative flux change between GHGAERO and GHG (Figure 4), whereas the spatial distribution of the surface temperature change is more different (Figure 5). This result is expected as the regional distribution of the sea surface temperature change may be strongly affected by the representation of the ocean boundary. However, Figures 1, 2, and 3, as well as the overall summary presented in Table 4, show that the latitudinal distribution of the surface temperature change due to changing aerosol concentrations is rather similar in the two versions of the GISS model. It is only at high latitudes that the temperature response differs notably; GISS-E2-R simulates a slight warming due to changing anthropogenic aerosol concentrations, whereas GISS-E2-H simulates a clear cooling.

For the models that consider aerosol indirect effects on clouds, there is no clear difference between the models that include an explicit online aerosol module and the ones that utilize prescribed aerosol fields (cf. Table 1). In contrast, one of the models with prescribed aerosol fields (CNRM-CM5) performs at least as well as the other models. In addition, although only one ensemble member is available from NorESM1-M and FGOALS-g2, FGOALS-g2 benefits more than NorESM1-M from including changes in aerosol emissions, although FGOALS-g2 utilizes prescribed aerosol fields. There is also no clear difference between the group of models that include the Albrecht effect and the models that do not, despite the fact that the Albrecht effect and the Twomey effect should be of approximately the same magnitude [Lohmann and Feichter, 2005]. However, a larger number of models would be necessary in order to improve the statistics and draw more firm conclusions.

It is difficult to say why the models with prognostic CDNC improve in more regions than the other models when aerosol effects are included. A key feature appears to be the strong negative radiative flux perturbation over South and Eastern Asia, extending out over the northern Pacific Ocean. The stronger cooling over the north Pacific in the models with prognostic CDNC may be due to that the hydrophobic aerosols emitted over Asia are not activated immediately, but they are transported and aged over time so that they eventually become efficient cloud condensation nuclei. Another reason may be that the supersaturation changes with a changing temperature and that this affects the number of activated aerosols in the models with prognostic CDNC.

5 Summary and Conclusions

In this study, the influence of anthropogenic aerosol direct and indirect effects on simulated surface temperature trends between 1965 and 2004 in a subset of CMIP5 models was examined. The models have different representations of aerosols, ranging from prescribed aerosol fields to fully explicit aerosol modules. The models also have different representations of aerosol indirect effects, ranging from no representation of aerosol indirect effects at all, over empirical relations between aerosol mass and cloud droplet number concentration (CDNC), to fully interactive two-moment aerosol-cloud microphysics schemes. By categorizing the models with respect to their description of the aerosol population and the relation between the aerosol and CDNC, the effect of different representations of the aerosol population and aerosol indirect effects was evaluated. Simulations considering time-varying GHG and anthropogenic aerosol concentrations were compared with simulations that only considered GHG concentrations that varied in time. Both sets of simulations were evaluated against observed surface temperature trends obtained from HadCRUT4.

When time-varying anthropogenic aerosol concentrations were considered, the simulated temperature trend bias was reduced (compared to the simulation with only time-varying GHG forcing) in more regions in the models with a more sophisticated parameterization of aerosol activation into cloud droplets (i.e., models where the cloud droplet number concentration is a function of the modeled supersaturation as well as the aerosol concentration and composition) than in the other models. The bias of the modeled ratio between the high- and low-latitude temperature trends was also reduced in more regions in the models with a more sophisticated parameterization of aerosol activation than in the other models. In the model group that include no aerosol indirect effects at all, the temperature trend bias was reduced in the least number of regions. In fact, in these models, the temperature trend bias was in general larger in the simulations considering changing anthropogenic aerosol concentrations over time than in the simulation with only time-varying GHG forcing. In the models with a simplistic description of aerosol indirect effects, i.e., where the CDNC is an empirical function of aerosol mass, the bias compared to observations was in general reduced when time-varying aerosol concentrations were considered but not in as many regions as in the models where the number of CDNC is a function both the aerosol concentration and the supersaturation.

No clear difference could be found between the models that include an explicit aerosol module and the ones that utilize prescribed aerosol fields. For example, the GISS-E2-H model (which utilizes prescribed aerosol fields) obtained a reduced bias in as many regions as the HADGEM2-ES model (which utilizes an explicit aerosol module) when comparing the simulation with time-varying anthropogenic aerosol concentrations with the simulation forced by only GHG concentrations that varied in time. Both models belong to the group of intermediate complexity regarding the representation of aerosol indirect effects; i.e., the CDNC is a function of the aerosol mass.

It is not clear exactly why the models that include a more sophisticated parameterization of aerosol activation into cloud droplets improved in more regions when anthropogenic aerosol effects were included compared to the other models. The zonal average of the anthropogenic aerosol radiative forcing and temperature response was quite different among these three models. One model (MIROC-ESM) displayed one of the lowest anthropogenic aerosol impacts on surface temperature trends, whereas another model (GFDL-CM3) displayed one of the highest. A key feature, however, appeared to be a relatively strong cooling at midlatitude and high latitudes in particular over the northern Pacific Ocean. The reason behind this cooling is unknown and should be examined in future studies.

To summarize, the study indicates that in order to simulate the latitudinal distribution of recent temperature trends in agreement with observations, it is better to include a simple empirical relation between the CDNC and the aerosol concentration than no relation at all. However, the best representation is obtained if the modeled CDNC is a function of both the aerosol concentration and the supersaturation. The results also indicate that it may be sufficient to use prescribed aerosol fields instead of an explicit aerosol module. In addition, for the representation of observed surface temperature trends, there is no clear benefit from using a simplified parameterization of the aerosol effect on precipitation formation (cloud lifetime or Albrecht effect) compared to no relation at all between the aerosol concentration and the precipitation formation rate.

Appendix A

Table A1 displays similar statistics as Table 4, but the bias has to be improved by at least 20% when comparing GHGAERO with GHG.

Table A1. Same as Table 4, But the Bias Has to be Improved by at Least 20% (GB = Global Bias, TB = Tropical Bias, MB = Midlatitude Bias, HB = High-Latitude Bias, MA = Midlatitude/Tropical Ratio Bias, HA = High-Latitude/Tropical Ratio Bias)a
ModelGBGRTBTRMBMRHBHRMAMARHAHARinline image
  1. a

    A bias reduction less than 20% is marked with a “0”. Only the result for models with more than one ensemble member is displayed.

MIROC-ESM0+0+0+0+0+++7
CESM1-CAM5-1-FV2++++++-+0+5
GFDL-CM3+++++++0+++9
HADGEM2-ES+++++++2
CSIRO-Mk3-6-0+++++−2
GISS-E2-H+++++++2
GISS-E2-R0+0+0+++1
IPSL-CM5A-LR00+0−7
CNRM-CM5++++0+0+2
CanESM2000+0+−4
CCSM4+0−8

Acknowledgments

I would like to thank Andrew Gettelman and two anonymous reviewers for their comments that significantly helped to improve the manuscript. The World Climate Research Program's Working Group on Coupled Modeling, which is responsible for CMIP, is acknowledged, and the climate modeling groups (listed in Table 1 of this paper) are thanked for producing and making their model output available. 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. The author also thanks the Met Office Hadley Centre and the Climatic Research Unit at the University of East Anglia for providing the HadCRUT4 data set and Urs Beyerle at ETH-Zurich for CMIP5 data downloading support. This work was partly funded by Vinnova-VINNMER grant 2010-00996.

Erratum

  1. In the originally published version of this article there were typos on pages 821 and 830, as well as in Table 1. These typos have since been corrected and this version may be considered the authoritative version of record.

Ancillary

Advertisement