Cloud tuning in a coupled climate model: Impact on 20th century warming



[1] Climate models incorporate a number of adjustable parameters in their cloud formulations. They arise from uncertainties in cloud processes. These parameters are tuned to achieve a desired radiation balance and to best reproduce the observed climate. A given radiation balance can be achieved by multiple combinations of parameters. We investigate the impact of cloud tuning in the CMIP5 GFDL CM3 coupled climate model by constructing two alternate configurations. They achieve the desired radiation balance using different, but plausible, combinations of parameters. The present-day climate is nearly indistinguishable among all configurations. However, the magnitude of the aerosol indirect effects differs by as much as 1.2 Wm  − 2, resulting in significantly different temperature evolution over the 20th century.

1 Introduction

[2] Despite decades of efforts, clouds remain one of the largest source of uncertainties in climate predictions from general circulation models (GCMs). Globally, clouds cool the planet by − 17.1 Wm  − 2 [Loeb et al. 2009]. This cooling results from a partial cancelation between two opposing contributions: cooling in the shortwave ( − 46.6 Wm  − 2) and warming in the longwave (29.5 Wm  − 2). To put these numbers in perspective, the radiative impact of the increase in long-lived greenhouse gases since 1750 is estimated to be 2.63  ±  0.26 Wm  − 2 [Forster et al., 2007, Table 2.12]. It should therefore not come as a surprise that uncertainties in the representation of clouds can have considerable impact on the simulated climate. For example, nearly a quarter century ago, Mitchell et al. [1989] showed that replacing one cloud parametrization with another one could significantly impact the predicted warming from a doubling of CO 2.

[3] Uncertainties that arise from interactions between aerosols and clouds have received considerable attention due to their potential to offset a portion of the warming from greenhouse gases. These interactions are usually categorized into first indirect effect (“cloud albedo effect”; Twomey [1974]) and second indirect effect (“cloud lifetime effect”; Albrecht [1989]). Modeling studies have shown large spreads in the magnitudes of these effects [e.g., Quaas et al., 2009].

[4] CM3 [Donner et al., 2011] is the first Geophysical Fluid Dynamics Laboratory (GFDL) coupled climate model to represent indirect effects. As in other models, the representation in CM3 is fraught with uncertainties. In particular, adjustable cloud parameters used for the purpose of tuning the model radiation can also have a significant impact on aerosol effects [Golaz et al. 2011]. We extend this previous study by specifically investigating the impact that cloud tuning choices in CM3 have on the simulated 20th century warming.

2 Method

[5] We start with the standard CMIP5 GFDL CM3 model and construct two alternate configurations by adjusting parameters associated with cloud processes illustrated in Figure 1. They include the lower bound on velocity variance for activation (σw,min), the autoconversion threshold radius (rthresh), and the erosion time scales. Cloud condensation nuclei (CCN) are activated using a subgrid distribution of vertical velocity whose variance is controlled by turbulence mixing with a minimum imposed value σw,min [Golaz et al. 2011]. Increasing σw,min leads to more CCN being activated and more reflective clouds. Autoconversion converts cloud water to rain. The conversion occurs once the mean cloud droplet radius exceeds rthresh. Larger rthresh delays the formation of rain and increases cloudiness. Erosion dissipates clouds by lateral mixing with the environment. Reducing erosion timescales increases cloudiness.

Figure 1.

Illustration of key cloud processes with adjustable parameters in the GFDL CM3 model: activation of cloud condensation nuclei (CCN) to form cloud droplets, autoconversion of cloud water to rain, and erosion by lateral mixing.

[6] Our approach resembles the creation of “parallel Worlds” by Mauritsen et al. [2012] but for different parameters. We select the autoconversion threshold radius as primary parameter because of its strong control on the magnitude of the indirect effect [Golaz et al. 2011]. The default in CM3 is 8.2  μm. We select alternate values from other GFDL models: 6.0  μm (HiRAM; Zhao et al. [2009]) and 10.6  μm (AM2; GFDL Global Atmosphere Model Development Team [2004]). We denote these configurations as CM3w and CM3c, respectively. This range is similar to some sensitivity experiments with CAM5 by Wang et al. [2012].

[7] Altering the autoconversion in CM3 has a considerable impact on cloudiness and radiation balance. In order to perform coupled climate integrations, the radiation balance must be restored. We target a radiation balance within the range of 0.50  ±  0.43 Wm  − 2 for the period 2001–2010 [Loeb et al. 2012]. Table 1 summarizes the cloud parameter choices. For CM3w, the reduced cloudiness from the more efficient autoconversion is offset by slowing cloud erosion. We set the inverse erosion time scales to 1.0e − 6 s  − 1, corresponding to the original value suggested by Tiedtke [1993]. For CM3c, the increased cloudiness from the less efficient autoconversion is offset by reducing the lower bound on the variance for activation from 0.7 to 0.2 ms  − 1.

Table 1. Summary of Model Configurationsa
 rthreshErosion ( × 10 − 6s − 1)σw,minTOA rad.
 (μm)MainConv.Turb.(ms − 1)(Wm − 2)
  • a Adjusted cloud parameters include autoconversion threshold (rthresh), erosion inverse time scales (main value and enhanced values under convective and turbulent conditions), and lower bound on the variance for activation (σw,min). Also listed is the top-of-atmosphere net radiation for the period 2001–2010 using fixed sea surface temperature (HadISST; Rayner et al. [2003]).

[8] There is a long history of using autoconversion to tune climate models [e.g., Rotstayn, 2000]. Values as low as 4.5  μm have been reported, even though they are not supported by observations [e.g., Gerber, 1996; Boers et al., 2006]. Pawlowska and Brenguier [2003] analyzed flight segments whose lengths were comparable to GCM grid box sizes and found that precipitation in stratocumulus clouds forms when the maximum mean volume droplet radius exceeds 10  μm. Attempts to explain reduced values in large-scale models have invoked the neglect of in-cloud subgrid-scale variability [Pincus and Klein, 2000; Larson et al. 2001) and the dispersion effect (Rotstayn and Liu, 2005].

[9] The tuning choices made here are justifiable based on the current state-of-the-art. However, some choices might be more desirable than others. CM3c is the configuration with the preferable autoconversion setting. In CM3, the lower bound on the vertical velocity variance for CCN activation is larger than in other studies (e.g., 0.2 m s  − 1 in Ghan et al. [2001] and 0.3 m s  − 1 in Storelvmo et al. [2006]). Furthermore, the frequency of occurrence of this lower bound is 98% [Golaz et al. 2011]. It could be argued that the reduced value of 0.2 m s  − 1 in CM3c is preferable. Erosion time scales are poorly constrained, and both CM3c and CM3w are equally plausible.

3 Results

3.1 Atmosphere-only Experiments

[10] We first discuss results from a set of atmosphere-only experiments with climatological boundary conditions. For each configuration, we perform three integrations using present-day SSTs but different forcings: (1) PD: present-day emissions (primary aerosols and short-lived gases) and long-lived greenhouse gas (GHG) concentrations, (2) PI: preindustrial emissions and GHG concentrations, (3) PI nonGHG: preindustrial emissions and present-day GHG concentrations. All three integrations use dynamic potential vegetation (no anthropogenic landuse effect) and neglect forcing from explosive volcanoes. This set of integrations allows us to calculate the adjusted radiative forcing (AF) with relatively inexpensive simulations (15 years following a 1 year spin-up in our case).

[11] We compute the total adjusted forcing (AF tot) and the non-GHG forcing (AF nonGHG). We also estimate the GHG forcing as a residual of the two (AF GHG = AFtot − AFnonGHG). By construction, AF tot incorporates fast responses to all anthropogenic forcings, whereas AF nonGHG isolates effects of aerosols (direct and indirect), and short-lived gases. We also compute the Cess climate sensitivity (λ) by uniformly increasing SSTs by 2 K [Cess et al., 1989]:

display math(1)

where ΔF is the difference in net top-of-the-atmosphere radiation flux. We define the temperature change metric, math formula. math formula can be regarded as a first-order correlate of the possible temperature change from preindustrial to present-day conditions.

[12] Table 2 lists values of adjusted forcings, Cess sensitivity, and temperature change metric. These results confirm the previous finding that the autoconversion threshold strongly modulates the magnitude of the aerosol indirect effects [Golaz et al. 2011]. In CM3c, autoconversion is the least efficient, and the indirect effects offset the majority of the radiative impact from greenhouse gases with a total AF of 0.34 Wm  − 2. In CM3w, autoconversion is the most efficient. The magnitude of the indirect effect is substantially reduced, resulting in a total AF of 1.54 Wm  − 2. Since the Cess sensitivity is unaffected, we expect the predicted warming from preindustrial to present-day to be dominated by changes in AF.

Table 2. Summary of Adjusted Radiative Forcings (Total, Non-GHG, GHG), Cess Climate Sensitivity λ, and Temperature Change Metric math formula.
 AF totAF nonGHGAF GHGλmath formula
 W m  − 2W m  − 2W m  − 2K/ (W m − 2)K
CM30.99 − 1.602.590.670.67
CM3w1.54 − 1.012.550.651.00
CM3c0.34 − 2.282.620.660.22

3.2 Coupled Experiments

[13] We perform coupled integrations for CM3w and CM3c starting from the initial conditions of the first CMIP5 CM3 historical ensemble member. Control simulations are first run for 100 years to let the shorter time scales in the system adjust to the modified cloud tunings. We then create an ensemble of five historical members (1860–2005) starting every 50 years from the control simulations. The control simulations are each run for 500 years. All forcings for the control and historical simulations are identical to the CMIP5 CM3 integrations. The set of CM3w and CM3c simulations represent a total of nearly 2500 years.

[14] A summary of the present-day (1981–2000) global mean climatology performance is presented in Figure 2 in the form of target diagrams. The left panel measures the performance of the three model configurations against observations. For each metric, models tend to be clustered together. The right panel measures the difference between alternate configurations CM3c, CM3w, and the reference CM3. Comparing the two panels, it is clear that the distance between models is much smaller than the distance between models and observations. In other words, the overall impact of the alternate cloud tuning is small when measured against present-day climatology.

Figure 2.

Target diagrams [Jolliff et al. 2009] illustrating present-day coupled model performance. Model results are annual mean ensemble averages for the period 1981–2000. Vertical axes represent normalized biases (B*), and horizontal axes normalized unbiased root mean square difference multiplied by the sign of the model and reference standard deviation difference (RMSD* ′ ). The left panel compares model with observations, and the right one alternate configurations (CM3c, CM3w) with CM3. Fields shown are sea-level pressure (SLP), 500 hPa geopotential height (z500), 200 and 850 hPa zonal wind (U200, U850), sea surface temperature (SST), land air surface temperature (TAS), precipitation (Precip), and shortwave and longwave cloud forcing (SWCF, LWCF). Respective data sources are ERA40, HadISST, CRU, and CERES-EBAF [Uppala et al., 2005; Rayner et al., 2003; Brohan et al., 2006; Loeb et al., 2009].

[15] However, there is a substantial difference in the temporal evolution from preindustrial to present-day as illustrated in Figure 3 by global mean surface air temperature anomalies. CM3 tends to underpredict the warming due to a strong cooling from aerosol effects [Donner et al., 2011; Levy et al., 2013]. CM3w (largest AF tot, weakest indirect effect) is warmer than CM3 and follows observations more closely. CM3c (smallest AF tot, strongest indirect effects) is colder than CM3 with almost no discernible warming except towards the end of the simulated period. The spread between CM3w and CM3c maximizes during the second half of the 20th century when indirect effects are largest because sulfate emissions reach their peak.

Figure 3.

Time evolution of global mean surface air temperature anomalies. Color lines represent the CMIP5 GFDL CM3 model (green) and the two alternate configurations, CM3w (red) and CM3c (blue). Each line is a five-member ensemble average. Anomalies are computed with respect to 1881–1920. Model drift is removed by subtracting from each ensemble member the linear trend of the corresponding period in the control simulation. Also shown are observations from NOAA NCDC [Vose et al., 2012], NASA GISS [Hansen et al. 2010], and HadCRUT3 [Brohan et al. 2006]. A 5 year running mean is applied to model results and observations. Letters above the horizontal axis mark major volcanic eruptions: Krakatoa (K), Santa María (M), Agung (A), El Chichón (C), and Pinatubo (P).

[16] The CM3 temperature increases by 0.22°C between 1881–1920 and 1981–2000. This number includes a correction for control simulation drift and as a result is smaller than the 0.32°C warming reported by Donner et al. [2011]. For the alternate configurations, the temperature change is 0.57°C for CM3w and − 0.01°C for CM3c. CM3w is closest to observations, which report warmings of 0.59°C (NOAA NCDC), 0.53°C (NASA GISS), and 0.56°C (HadCRUT3).

[17] Figure 4 depicts the time evolution of the 700 m ocean heat content anomalies. CM3w warms the most and better matches observations. Furthermore, it is the only configuration with a net preindustrial to present-day increase in ocean heat content.

Figure 4.

Time evolution of global mean upper 700 m ocean heat content anomaly. Color lines represent the CMIP5 GFDL CM3 model (green) and the two alternate configurations, CM3w (red) and CM3c (blue). Each line is a five-member ensemble average. Anomalies are computed with respect to 1860–1880. Model drift is removed by subtracting from each ensemble member the linear trend of the corresponding period in the control simulation. Observations from [Levitus, 2012] are shown with offsets to match model averages over the period (1955–2005). Letters above the horizontal axis mark major volcanic eruptions: Krakatoa (K), Santa María (M), Agung (A), El Chichón (C), and Pinatubo (P).

[18] Volcanic eruptions are clearly noticeable in Figures 3 and 4. Explosive volcanoes do not impact aerosol concentrations in CM3. Instead, an identical times series of volcanic optical properties is imposed in all configurations [Donner et al., 2011]. However, the response to individual volcanoes appears to be generally larger in CM3c than CM3w. A detailed analysis of the volcanic response is beyond the scope of this work.

4 Discussion

[19] We have shown that there is sufficient ambiguity in the CM3 adjustable cloud parameters to construct alternate configurations (CM3w, CM3c) that achieve the desired radiation balance. These configurations exhibit only modest differences in their present-day climatology. Indeed, one would be hard pressed to select the “better” configuration solely based on present-day metrics such as those in Figure 2. However, CM3w and CM3c differ significantly in the magnitude of their indirect effects. As a result, their predictions of the 20th century warming are strongly affected (Figures 3 and 4).

[20] CM3w predicts the most realistic 20th century warming. However, this is achieved with a small and less desirable threshold radius of 6.0  μm for the onset of precipitation. Conversely, CM3c uses a more desirable value of 10.6  μm but produces a very unrealistic 20th century temperature evolution. This might indicate the presence of compensating model errors. Recent advances in the use of satellite observations to evaluate warm rain processes [Suzuki et al., 2011; Wang et al., 2012] might help understand the nature of these compensating errors.

[21] CM3 was not explicitly tuned to match the 20th temperature record. However, our findings indicate that uncertainties in cloud processes permit a large range of solutions for the predicted warming. We do not believe this to be a peculiarity of the CM3 model. Indeed, numerous previous studies have documented a strong sensitivity of the radiative forcing from aerosol indirect effects to details of warm rain cloud processes [e.g., Rotstayn, 2000; Menon et al., 2002; Posselt and Lohmann, 2009; Wang et al., 2012]. Furthermore, in order to predict a realistic evolution of the 20th century, models must balance radiative forcing and climate sensitivity, resulting in a well-documented inverse correlation between forcing and sensitivity [Schwartz et al. 2007; Kiehl, 2007; Andrews et al. 2012]. This inverse correlation is consistent with an intercomparison-driven model selection process in which “climate models’ ability to simulate the 20th century temperature increase with fidelity has become something of a show-stopper as a model unable to reproduce the 20th century would probably not see publication” [Mauritsen et al., 2012].


[22] We are grateful to Leo Donner and the entire Global Atmospheric Model Development Team for innumerable stimulating discussions about AM3/CM3. We acknowledge Tom Delworth and Rong Zhang for a thought provoking analysis of CM3 results. Figure 1 was drafted by Cathy Raphael.