PROJECT DESCRIPTION
1. Results from Prior Research
Yuk Yung was supported by ATM-0529268, ATM-0840787 and ATM-0934303 to study greenhouse
gases and heating rates, resulting in 10 publications listed in the References.
Guang Zhang was supported by ATM-0601781, ATM-0832915, AGS1015964, and EaSM-1048995
to improve parameterization of convection in the NCAR climate model, especially CAM3 and CCSM3,
summarized in 15 papers listed in the References.
2. Introduction
2.1 The Difficulty of Predicting Global Warming. More than three decades ago a distinguished study
group was formed to assess the effect of CO2 on climate (Charney et al. 1979), with the following
conclusion on what is now known as the equilibrium climate sensitivity (ECS), “We estimate the most
probable global warming for a doubling of CO2 to be near 3°C with a probable error of ± 1.5°C.” Charney
et al. (1979, p.9) identifies clouds as a major source of uncertainty for estimating ECS, “How important
the overall cloud effects are is, however, an extremely difficult question to answer. The cloud distribution
is a product of the entire climate system, in which many other feedbacks are involved. Trustworthy
answers can be obtained only through comprehensive numerical modeling of the general circulations of
the atmosphere and oceans together with validation by comparison of the observed with the modelproduced cloud types and amounts. Unfortunately, cloud observations in sufficient detail for accurate
validation of models are not available at present.” To reflect the pioneers’ insight and wisdom that will
guide our proposed investigation, we emphasize the words in italics.
About a decade later, the first Intergovernmental Panel on Climate Change (IPCC, Houghton et al.
1990, also known as AR1) reports on the assessment of ECS, “near the Earth’s surface, the global average
warming lies between 1.5°C and 4.5°C, with a “best guess” of 2.5°C.” These results are based on 22
general circulation models (GCM). AR1 states that one of the key areas of uncertainty in climate
assessment is, “clouds: primarily cloud formation, dissipation, and radiative properties, which influence
the response of the atmosphere to greenhouse forcing.”
Cess et al. (1989) first recognize that differences in simulated
cloud radiative effect (CRE) could quantitatively explain the spread
in ECS estimates by the different models. Their results are elegantly
summarized in Fig. 1 as a correlation between ECS and CRE
(related to their CRF/G). While the influence of clouds on ECS is
expected, the compact relation in Fig. 1 reveals that cloud feedbacks
are the primary sources for the uncertainty in the model predictions
of ECS reported in AR1. More than two decades later the fifth IPCC
assessment report (AR5) states that ECS ranges from 2.1°C to 4.7°C
with a multi-model-mean ~3.4°C. The large uncertainty has
stubbornly resisted reduction for two decades since AR1. The
reason is simple. Despite the fundamental importance of the relation
Fig. 1. The global sensitivity parameter
summarized in Fig. 1, Cess et al. (1989) could not pinpoint which
as defined by Cess et al. (1989) for the 2
model has the correct CRE; the authors could not determine even
K SST perturbation experiments plotted
as a function of the ratio of the change
the sign of the correct CRE, as cloud feedback could be negative
in the climate sensitivity parameter λ to
according to two models in Fig. 1.
change in net cloud radiation budget for
Fig. 2 provides an overall summary of the essential components
19 GCMs. To convert λ to ECS,
that
control ECS in a typical GCM (Stephens 2005). The direct
-2
multiply by 4 Wm , so that the smallest
contribution to ECS due to radiative forcing by greenhouse gases
and the largest values of ECS are about
1.6 K and 4.5 K, respectively. Taken
(GHG, in this case, doubling of CO2) is about 1°C (black bar). The
from Cess et al. (1989).
warming induces a positive feedback by water vapor, which
increases by 7%/K according to the Clausius-Clapeyron relation. The cumulative effect is about 2°C (blue
bar). An additional positive feedback due to snow/ice and albedo feedback raises ECS to 2.5°C (red bar).
So far there is unanimity on these results from different climate models, as these feedbacks are primarily
thermodynamic. However, adding the cloud feedback introduces a major source of uncertainty, with ECS
1
ranging from 2°C (low green bar) to
66
Fig. 2. ECS of a typical
*
climate model as different
5.5°C (high green bar) and a mean
55
Cloud
feedbacks are systematically
44
value of 3.5°C (middle green bar). If
Feedback
added in the model. Different
Range
we compare the warming in a high ∆TT 33
treatments of cloud processes
22
ECS model (favored model, see later
in the model produce a large
*
11
discussion) to the direct effect of GHG
spread in predicted ECS.
Taken from Stephens (2005).
00
forcing (black bar), the amplification
DirectGHG
GHG
Snow/Ice
Direct
Snow/Ice Albedo
Forcing
forcing
is more than a factor of 5!
Albedo
Water
Water
Vapor
vapor
The above discussion explains
Cloud
CloudFeedback
feedback
Feedback
feedback
why it is so hard to get all the
feedbacks right. Even though the fundamental physics is solid and dependable, modeling all the feedbacks
is like building a house of cards. The vulnerability of this enterprise was noticed by a critical mind
(Lindzen, 1990), “Yet in the major numerical models, all feedbacks between warming and water are
positive. In the absence of these positive feedbacks, the same models would yield warming (due to
doubling of CO2) only one-half to one-fifth of those cited above. When it is recognized that at least some
of these feedbacks may be negative rather than positive, it is easy to see that the actual response to a
doubling of CO2 may be much less.” While the specific possible negative feedback mechanisms proposed
by Lindzen (1990) and Lindzen et al. (2001) were refuted later (Betts 1990, Lin et al. 2002, and Su et al.
2008), it is a sobering fact that we have made virtually no progress in reducing the uncertainties in
predicting ECS for 23 years from AR1 in 1990 to AR5 in 2013. This persistent uncertainty in predicting
ECS impedes policy decisions. A new strategy is urgently needed.
2.2 A Possible Connection
Fig. 3. Scatterplot of ECS versus
between ECS and the Presentpresent-day simulated zonal mean
RH over ocean from May to
Day Mean State. A recent paper
August in the dry (left) and moist
by Fasullo and Trenberth
(right) zones. Mean observed
(2012a) shows that subtropical
values (vertical black lines) and
relative humidity (RH) is
their range (gray) are based on
strongly correlated to ECS. Fig.
values from AIRS retrievals and
the MERRA and ERA Interim
3 shows a scatterplot of ECS
reanalyses. Dots denote individual
versus present-day simulated
model runs from coupled 20th
zonal mean RH over ocean from
century runs. This figure is taken
May to August in the dry (D)
from Fasullo and Trenberth
(2012a). The points representing 4
and moist (M) zones. The areas
NCAR models are from Fasullo
(D and M) are defined in their
and Trenberth (2012b).
Fig. 3: D is a dry region centered
at 15S and 400 hPa, M a wet region at 700 hPa above the equator, representing the downwelling and
upwelling regions of the Hadley Circulation, respectively. Mean observed values (vertical black lines)
and their range (gray) are based on values from Atmospheric Infrared Sounder (AIRS) retrievals, the
Modern-ERA Retrospective Analysis for Research and Applications (MERRA), and ERA Interim
Reanalysis data. Dots denote individual model runs from coupled 20th century runs. Names of the CMIP3
models are listed in their Fig. 4 and Table S1. The correlations with ECS for each region are indicated by
the least-squares regression lines. Observations and model values are based on climatologies from 1980 to
1999, except for AIRS, which is based on 2002–2007 data. The significance of this paper is threefold: (a)
there is a compact anti-correlation (correlation) between ECS and RH in the dry (moist) region, (b) from
intercomparison between model and data in region D, models with high ECS are favored by the RH
observations, and most importantly (c) there is a connection between ECS and the present-day mean state.
We will briefly comment on (c). This is an intriguing result, implying that a precise determination of
the mean state of the present-day atmosphere could lead to an accurate prediction of ECS a century later!
Referring to Fig. 3, note that in the dry region (D), the correlation between ECS and RH is –0.81. Models
with high ECS such as N (MIROC 3.2 Highres, the symbols are shown in their Fig.3), P (UKMO
HADGEM 1), and J (CCCMA CGCM3.1 – T47) are favored by the data, whereas models with low ECS
2
such as A (NCAR PCM-1), B (INMCM 3.0), and C (IAP FOALS 1-0g) are ruled out. In the moist region
(M), the correlation is less good (0.65). We will defer a discussion of its significance and the performance
of successive generations of NCAR models labeled PCM1, CAM3, CAM4 and CAM5 to § 4.
This paper points to a new path to solve the problem that seems intractable as discussed in § 2.1.
However, climatologist Shell (2012) has an insightful commentary on this work, “Thus, determining the
climate sensitivity from observations requires two nontrivial steps. First, the relation between short-term
(seasonal, interannual, or decadal) and long-term feedbacks needs to be determined using models. Second,
modeled short-term feedbacks must be compared to observations to test the models. … However, shortterm and long-term cloud feedbacks do not appear to be correlated in models (Dessler 2010). Thus, the
first step is not only difficult but may even be impossible, if short-term observations are not a guide to
long-term changes. Regarding the second step, one reason for the lack of success in evaluating cloud
feedbacks is the difficulty of obtaining cloud properties from either satellite or surface-based
observations.” We will propose a research effort to address these issues as follows.
2.3 A Challenge and Opportunity Presented. A major challenge presented by the Fasullo and Trenberth
(2012a) paper is to find a physical mechanism that connects the present-day mean state of the atmosphere
to a future ECS, which is based on a perturbation of this mean state. The authors clearly recognize the
problem, “the reasons for the simulated variability of RH remain poorly understood, and constraints
linking present-day variations to future climate have not yet been thoroughly explored.” It is obvious that
RH and cloud feedback are both intimately tied to the Hadley Circulation of the atmosphere, but the
detailed mechanism that couples the two is not understood. The fundamental theory must be elucidated if
we are going to have any confidence in applying it to predict ECS.
This work presents an opportunity to differentiate the models’ estimates of ECS based on model
performance in simulating present-day climate, as present-day simulations bear imprints of future climate
change. However, the mechanisms for the connection between present-day climate and ECS are not
explained and the primary model physics responsible for the connection is not known. There is an entire
new world that waits to be explored.
In addition to exploring the influence of global warming on ECS, we will also explore the influence
of global warming on extreme weather (e.g. drought and flooding). Most observational studies (Adler et
al. 2003, Gu et al. 2007, Adler et al. 2008, Li et al. 2011) and climate models (Allen and Ingram 2002,
Held and Soden 2006, Stephens and Ellis 2008) suggest that global precipitation is increasing more
slowly than the total mass of water vapor in response to global warming. Previous studies (Chou and
Neelin, 2004, Chou et al. 2009, Li et al. 2011, Durack et al. 2012, Chou et al. 2013) that separated the
data into wet versus dry regions have found that the precipitation has a tendency to increase in the wet
areas and to decrease in dry areas, which is referred to as the rich-get-richer mechanism. The previous
global studies of precipitation provide a background for the investigation of the precipitation and drought
at the regional scale. In this proposal, we will explore extreme weather from the observations and model
simulations. The challenge and opportunity will be addressed in § 3.
3. Objectives
We have two broad objectives in this proposal. The first is to identify reliable global datasets that relate
cloud changes to circulation changes. The second objective is constraining model-predicted climate
sensitivity through quantification of circulation-cloud feedback using satellite observations and reanalysis
data. The two objectives are not independent. A fundamental theory (as opposed to an empirical relation)
will provide greater confidence in our ability to predict ECS. The intercomparison between data and
model provides a rigorous test of the underlying theory.
3.1 A Working Hypothesis. In the absence of a complete theory, a working hypothesis is sketched in Figs
4a and 4b. Fig. 4a shows a schematic NE–SW cross section over the northeastern Pacific or Atlantic from
40°N to the Equator showing the characteristic increase of cloud top height and boundary-layer inversion
height with increasing sea surface temperatures and decreasing subsidence. The deep tropical cumulus
clouds denote the ascending branch of the Hadley Circulation, whereas shallow cumuli and stratocumulus
clouds can be found in the descending branch. This figure illustrates the crucial role that convection and
the Hadley Circulation play in the formation of various types of clouds.
3
Fig. 4a. Schematic NE–SW cross section over the northeastern
Pacific or Atlantic from 40°N to the Equator showing the
characteristic increase of cloud top height and boundary-layer
inversion height with increasing sea surface temperatures and
decreasing subsidence. The deep tropical cumulus clouds denote
the ascending branch of the Hadley Circulation, shallow cumuli
and stratocumulus clouds can be found in the descending branch.
Adapted from Emanuel (1994).
0° Deep convective
Clouds
Trade
Cumulus
40°N
Stratus
Stratocumulus
Cumulus
Fig. 4b. A simple two-box view of the tropical circulation. Present-day is
represented in blue and the increased CO2 condition is represented in red.
Straight arrows indicate atmospheric circulation. Curly arrows indicate
outgoing longwave radiation to space. Figure adapted from Kelly and
Randall (2001).
Fig. 4c. The change of zonal-mean mass-weighted
stream function range at 500 hPa from the presentday to the end of the 21st century projected in the
RCP4.5 scenario versus TOA net CRE change
averaged between 45°S to 40°N for the 15 CMIP5
models. The solid line is the least squares linear fit to
the data with correlation coefficient marked. Taken
from Su et al. (2014).
What happens to the circulation if we perturb the system by doubling CO2 in the atmosphere? Fig. 4b,
adapted from Kelly and Randall (2001), presents a simple two-box view of the tropical circulation. The
Cloudy Sky box represents the deep convective region on the left side of Fig. 4a; the Clear Sky box
represents the region between clouds to the right of the deep convective region. Present-day is represented
in blue and the increased CO2 condition is represented in red. Straight arrows indicate atmospheric
circulation. Curly arrows indicate outgoing longwave radiation to space. Neglecting the energy transport
between the tropics and extra-tropics, the primary energy balance averaged over the tropics is between
radiative cooling (QR) in clear sky and latent heating in cloudy sky. They correspond to the heat source
and sink, respectively, of the Carnot engine that produces the work to drive the atmospheric motion with
ascent in the cloudy region and descent in the clear region. When atmospheric CO2 concentration
increases, heat loss to space is reduced due to enhanced greenhouse effect. Thus, the latent heating in the
convective region has to decrease, too. A weakened tropical circulation (red arrows) is expected. Such a
simple thermodynamic argument is valid on the large-scale spatial averages and is applicable to both the
Hadley and Walker Circulations. The slowing down of the atmospheric circulation when CO2 is increased
is now considered a secure and robust result (see, e.g., Held and Soden 2006, Stephens and Ellis 2008).
The fact that it has taken such a long time for the scientific community to reach this consensus suggests
that this conclusion is neither obvious nor trivial.
Our working hypothesis consists of three components: (a) clouds are formed by large-scale
circulation, (b) global warming slows down the large-scale circulation, and (c) on average cloud forcing
is proportional to circulation change. Note that (a) is well-known (see Fig. 4a), and (b) is generally
accepted (see Fig. 4b). Part (c) is new and forms the heart of our working hypothesis. We base it on two
reasons. First, clouds are produced by motion (convection). In the absence of motion the atmospheric
state would be defined by radiative equilibrium (see e.g., Goody and Yung 1989), and there would be no
clouds. Second, based on preliminary work by this team analyzing the cloud forcing in 15 CMIP5 models
4
(Su et al. 2014), we derived an approximately linear
relation as summarized in Fig. 4c, where we demonstrated
that the largest changes in cloud radiative forcing (CRE)
occurs in the models with the largest changes in
circulation as measured by the mass streamfunction at
500 hPa.
Our working hypothesis offers an intuitive
explanation for the relation between ECS and RH in Fig.
3, based on the following argument. RH is a good proxy
for circulation. ECS depends sensitively on change in
cloud forcing (Fig. 1) by our working hypothesis.
Therefore, the relationship between ECS in the future and
Fig. 5. Schematic of proposed technical approach.
Results of inter-comparisons between modeling and
RH in the present-day is intimately linked via the
data will provide understanding of the importance of
circulation. How robust is this relationship? It will
different atmospheric processes, leading to a stateobviously break down when the perturbed state evolves
of-the-art Benchmark Model that will be used for
sufficiently far from the mean state to cause a regime
climate prediction with quantitative characterization
of uncertainties. The Benchmark Model will be key
change. What are the limits within which this relationship
to climate assessment, adaptation and science
holds, and thereby useful for improving our prediction of
informing decisions.
ECS? What are the critical cloud data that address the
concerns of Shell (2012)? A plan for research addressing these issues follows.
3.2 Achieving Climate Prediction. A schematic of the strategic plan along with manpower allocation is
presented in Fig. 5. Acronyms for investigators can be found in § 6. It has four components. At the
foundation are the global datasets from observations or assimilation of observations, as summarized in
Table 1. CMIP5 model outputs will be analyzed and critically compared to the observed data. Diagnostic
studies will be carried out to reveal why the models differ from each other and from the observed data.
Results of intercomparisons between modeling and data will provide understanding of the importance of
different atmospheric processes, leading to a state-of-the-art Benchmark Model that will be used for
climate prediction with quantitative characterization of uncertainties. The Benchmark Model will be key
to climate assessment,
Table 1: A list of data products to be used in the proposed project. All data except
adaptation and the
GRACE (marked in yellow) are available at Caltech/JPL.
science that should
Source
Instruments
Key Variables
Resolution
Period
CloudSat CALIPSO
TOA, Surface Rad. Flux,
07/2006-present
inform decisions. The
CCCM
~20 km
CERES MODIS
L/IWP, PW, Cloud profile
details
will
be
ARM,
CERES,
1983-present
GEWEX SRB
Surface Rad. Flux
1ºx1º
ERBE, ISCCP
presented in § 4.
In-situ
and
SEAFLUX
LH & SH
0.5ºx0.5º
4. Proposed Research
satellites
1998-present
Cloud Fraction, Cloud Top
ISCCP
Multi. Satellites
280 km
1982-present
4.1
Analyzing
Temp/Pressure
L/IWC, Cloud (Opt Dep,
Observed Datasets and
1.3x3.5km,
CloudSat
Cloud Radar
Frac, & Rad. Flux), Heating
7/2006-present
0.5km
CMIP5 Outputs
Rates, Precip.
CALIPSO
Lidar
Cloud Extinction, IWC
5km, 60m
8/2006-present
Our approach involves
AIRS, AMSR-E
H O, T, Cloud fraction,
50 km, 2km
Aqua
9/2002-present
a combined analysis of
CERES
SST, PW, Rad. flux, Precip
0.25ºx0.25º
Aqua
AMSU-A
Temperature
1ºx1º
9/2002-present
observational datasets
300-400 km,
Aura
MLS
Upper Trop. H O, T, IWC
8/2004-present
3 km
and climate model
Terra
MISR
Cloud Motion Vector
0.5ºx0.5º
3/2002-present
simulations. As cloud
MODIS
Cloud (Top T/pressure,
Terra/Aqua
5km, TOA
1999-present
CERES
Opt Dep, Rad. Flux)
feedback
plays
a
TRMM
PR, TMI
Precipitation, PW
0.25ºx0.25º
dominant
role
in
1998-present
GPCP
Blended data
Precipitation
2.5ºx2.5º
determining the spread
1979-present
GPCC
Gauge
Precipitation
2.5ºx2.5º
1986-present
of climate sensitivity,
GPS
CHAMP, COSMIC
Temperature
5ºx5º
4/2001-12/2011
we will focus on the
GRACE
GRACE
Underground Water
150,000 km
2002-present
SBUV
Nimbus-7, NOAA
Reflectivity
1.5ºx1.5º
10/1978-present
parameterizations that
MERRA
Re-analysis
T, H O, winds
0.5ºx0.67º
1979-present
bear a strong relevance
NCEP/NCAR
Re-analysis
Atmospheric winds
2.5ºx2.5º
1948-present
ERA-Interim
Re-analysis
Atmospheric winds
1.5ºx1.5º
1989-present
to cloud feedback. To
2
2
2
2
5
link current climate states to climate sensitivity through the quantification of cloud feedback, we will
analyze CMIP5 present-day simulations and future projections to identify coherent spatial or temporal
variabilities in cloud-related variables.
4.1.1 Satellite and Reanalysis Data We will integrate a number of satellite and in-situ datasets (see Table
1), many of which have been developed by Collaborator H. Su. The datasets include TOA and surface
radiative fluxes from CALIPSO-CloudSat-CERES-MODIS (CCCM) merged product, CERES on Terra
and Aqua satellites, CloudSat 2B-FLXHR, and the GEWEX Surface Radiation Budget (SRB) project;
surface latent and sensible heat fluxes from the SEAFLUX project; and the ISCCP cloud product. The
cloud profile measurements such as the CloudSat/CALIPSO combined cloud fraction (2B-GEOPROFLIDAR), cloud optical depth (2B-TAU), LWC and IWC (2B-CWC-RVOD) will also be used. In addition,
we will examine water vapor and relative humidity measurements from Aqua AIRS and Aura MLS.
Precipitation from TRMM, GPCP and CloudSat, and atmospheric state variables (e.g. ω500) from
NCEP/NCAR, MERRA and ERA-interim reanalysis will be used. Table 1 lists the relevant datasets. The
variables listed in Table 1 have close counterparts in the CMIP5 archive. We will use datasets for clouds
that are hitherto not widely used, e.g., Herman et al. (2013) and Kahn et al. (2013).
Fig. 6. a: EOF1 of ISCCP explains 22.8% of the variance in the data. b: PC1 of ISCCP and the normalized ENSO index are
plotted in blue and red respectively. c: Regression of SST PC1 on ISCCP. d: EOF1 of TOMS explains 6.9% of the variance in
the data. e: PC1 of LER and the normalized ENSO index are plotted in blue and red respectively. f: Regression of SST PC1 on
LER. g: EOF4 of SST. h: PC4 of SST with a clear upward trend. i: Regression of SST PC4 on ISCCP.
We provide examples of cloud data to argue that criticisms such as Shell’s (2012) § 2.2 may be too
pessimistic. Figs. 6a and 6b show the first EOF and associated PC, respectively, for deseasonalized
ISCCP data (Rossow and Schiffer 1991,1999). The normalized ENSO index is over-plotted on top of the
PCs. Visual comparison of the first PCs with the ENSO index shows strong association, and the spatial
pattern shows the east-west anomalies typical of ENSO. What drives the cloud variability seen in these
6
figures? The obvious driver is SST. We
performed an EOF analysis of SST; the first EOF
explains 27.5% of the variance. We then
regressed the ISCCP clouds with the first PC of
SST. The results are shown in Fig. 6c. Its
striking resemblance to Fig. 6a confirms SST as
the main driver of cloud changes. Figs. 6d, 6e
and Xf present the similar results using data from
340 nm Lambertian equivalent reflectivity (LER)
of the Earth (Herman et al. 2013). The major
contributors to LER at this wavelength are
Rayleigh scattering and clouds. The former does
not change, and we can regard LER as a proxy
for cloud reflectivity. Note the agreement
between the results of these two independent sets
of data. The percent variance explained by the
ENSO mode of ISCCP and LER data is 22.82%
and 6.92%, respectively, suggesting that LER
may be a “noisier” dataset. Is there a decadal
trend in the ISCCP data associated with a trend
with SST? Figs. 6g and 6h show the 4th EOF
Fig. 7. The multi-model-mean zonal-mean vertical pressure
velocity (ω), cloud fraction (CF), relative humidity (RH) and
and
associated
PC,
respectively,
for
cloud radiative effect (CRE) changes from the present-day to
deseasonalized SST data. This EOF accounts for
the end of the 21st century projected in the RCP4.5 scenario.
4.34% of variance; it resembles the Pacific
(a) Changes of ω in color shadings and the present-day
Decadal Oscillation (PDO). The PC clearly has a
climatological ω in contours, (b) changes of CF in color
shadings and RH in contours, and (c) changes of area-weighted
trend. Fig. 6i presents the regression of ISCCP
top-of-atmosphere (TOA) net, longwave and shortwave CREs.
clouds on this PC. This pattern suggests the
The changes refer to the differences of each variable between
response of clouds to SST warming. Comparing
the averages for 2074-2098 in the RCP4.5 scenario and for
these results with Su et al. (2014), we see that
1980-2004. Taken from Su et al. (2014).
this imprint on cloud cover from the long-term
warming trend in SST is approximately consistent with climate model projections of cloud changes. LER
is an example of an under-utilized dataset for studying cloud change.
4.1.2 Analysis of Mean States in CMIP5 Models. Su et al. (2014) examined the correlations of zonal mean
cloud fraction, relative humidity and the Hadley circulation strength to ECS. For the proposed study, we
will conduct further analysis of mean state variables (25-year averages) not explored in Su et al. (2014).
These include water vapor, precipitation, IWC/LWC and IWP/LWP and cloud radiative fluxes and
heating rates. We will examine full 3-dimensional structures, in addition to zonal means.
Water vapor is intimately linked to clouds, and its variations combined with temperature determine
the variations of relative humidity. While lower-tropospheric water vapor changes are strongly governed
by surface temperature changes, upper tropospheric (UT) water vapor is influenced by convective
transport of water vapor and detrainment of cloud ice, thus exhibiting stronger sensitivity to surface
temperature change, nearly 3 times that for lower tropospheric water vapor (Soden and Fu 1995, Su et al.
2006). Because of the uncertainties in the model representations of convective processes, UT water vapor
simulations show a large spread among climate models and between the models and observations (Jiang
et al. 2012). The water vapor radiative kernels (Soden et al. 2008) indicate that TOA longwave fluxes are
most sensitive to the UT water vapor. Thus, UT is the most important area for the positive water vapor
feedback. Our preliminary analysis found that there is a systematic difference between the high and low
sensitivity models in that the warmer models have greater increases of UT water vapor under global
warming, consistent with a stronger and deeper penetration of tropical convection shown in Fig. 7.
Whether the systematic differences in the UT water vapor exist in present-day simulations and how the
7
present-day water vapor simulations relate to climate sensitivity are subjects of our proposed research.
Once we establish the linkage of present-day water vapor distributions to ECS, we will use satellite
observations from Aqua AIRS and Aura MLS to differentiate the models’ performances and thus place a
constraint on ECS. Meanwhile, we will compare the satellite data with reanalysis datasets and document
the needed improvements in water vapor profiles for reanalysis data.
4.1.3 Analysis of CRE in CMIP5 Models. Based on preliminary work (Su et al. 2014), we have
demonstrated that the detailed structures of the zonal mean vertical velocity change are useful for
understanding the dynamic control of cloud changes. To quantify the magnitude of the meridionally
varying structures in the Hadley Circulation change, we apply Empirical Orthogonal Function (EOF)
analysis to the 15 models’ zonal mean vertical velocity change at 500 hPa (ω500) within 45°S-40°N (using
the full vertical velocity profile yields similar results). In this EOF analysis, the spatial pattern of interest
is the change of zonal mean ω500 as a function of latitude. Each model is equivalent to a “time slice” in the
conventional temporal series for EOF analysis. The Principal Components (PCs) of the leading EOF
mode represent the magnitudes of the dominant structure of the Hadley Circulation change for all models.
By so doing, we can objectively determine the magnitude of the meridionally varying structures of the
Hadley Circulation change for each model and avoid biases by using local indices. A similar method was
used in a recent study to identify the SST warming patterns and amplitudes in a number of model
simulations (Tokinaga et al. 2012).
The resulting first EOF mode explains 57% of the variance across the models. The corresponding
meridional structure of the first EOF mode is shown in Fig. 8. It is strikingly similar to the multi-modelmean ω500 change (Fig. 7), justifying the representativeness of the multi-model-mean in the preceding
discussions. Combined with climatological ω500 (the dashed line in Fig. 8a), the first EOF mode of ω500
change clearly defines the “strengthening” and “weakening” segments of the Hadley Circulation change,
marked in Fig. 8a. The PCs of the first EOF mode have values between 0.026 and 2.0 (Fig. 8b), two
orders of magnitude difference among the models, suggesting a tremendous disagreement among the
models in terms of dynamic response to the doubling of CO2 and other forcings in the projected climate
change scenario, as found in earlier studies (e.g., Bony et al. 2013). This suggests that the inter-model
differences in the magnitude of the Hadley Circulation change could explain about 50% of the spread in
model simulated CREs.
Fig. 8. The linkage between the meridional
structure of the Hadley Circulation change
and TOA CRE change. (a) Meridional
structure of the first EOF mode of the 500
hPa vertical pressure velocity (ω500) change
across the 15 models superimposed with
climatological ω500. The amplitude of the first
EOF mode is enlarged by 10 times. “S” and
“W” denote the “strengthening” and
“weakening” segments of the Hadley
Circulation change, respectively. (b) The
Principal Component of the first EOF mode
for the Hadley Circulation change versus
TOA net CRE change averaged between
45°S and 40°N for the 15 models. The solid
lines are the least squares linear fit to the
data, with correlation coefficients marked.
Taken from Su et al. (2014).
8
Fig. 9. Model performance metrics that capture the spatial variations of zonal-mean cloud fraction and relative humidity
associated with the Hadley Circulation. (a) regression slopes (αCF) of model zonal-mean cloud fraction profiles onto the joint
CloudSat/CALIPSO cloud fraction versus the regression slopes (αRH) of model zonal-mean relative humidity profiles onto the
combined AIRS/MLS relative humidity between the surface and 100 hPa from 45°S to 40°N. (b) similar to (a), except for the
spatial correlations between the model and observed cloud fraction (ρCF) and relative humidity (ρRH) profiles. The red and blue
colors denote high and low sensitivity models, respectively, with the red and blue solid circles corresponding to the ensemble
composites for the high and low sensitivity models, respectively. The black solid circles denote observations. Taken from Su et
al. (2014).
To quantify the models’ performance in representing the coherent spatial variations of cloud fraction
and relative humidity associated with the Hadley Circulation, we compute the regression slopes of the
model zonal mean cloud fraction and relative humidity (αCF and αRH) onto the observed profiles as well as
the spatial correlations between the model and observed profiles (ρCF and ρRH) between the surface and
100 hPa within 45°S to 40°N (Fig. 9). These metrics would yield a value of one when the models and
observations match perfectly (black solid circles, Fig. 9). Although cloud fraction and relative humidity
are two highly correlated quantities, the observed cloud fraction and relative humidity measurements are
taken from independent instruments and thus can serve as two observational constraints. We find that the
models that are closer to the observations tend to have ECS higher than the multi-model-mean (labeled in
red), while the low sensitivity models (labeled in blue) are generally farther away from the observations.
The high and low sensitivity model composites are clearly separated in terms of the deviation from the
observations with the high sensitivity composite being much closer. Because of the strong correlation
between cloud fraction and relative humidity, the regression slopes and spatial correlations are distributed
approximately diagonally: low (high) αCF and ρCF are usually associated with low (high) αRH and ρRH.
Considering the observational data uncertainty of about 25% for cloud fraction and relative humidity, we
may crudely define the “best estimates” as those models with αCF and αRH between 0.75 and 1.25, or with
ρCF and ρRH between 0.75 and 1.0. With these cut-off values, we find that only 6 high sensitivity models
plus the high sensitivity composite fit into the area of “best estimates”.
We caution that the limited model samples examined here do not necessarily provide robust statistics
of ECS. However, the tendency of better-performing models to have higher ECS is in general agreement
with earlier studies (Fasullo and Trenberth 2012a, Klein et al. 2013, Tung et al. 2008). These preliminary
analyses must be confirmed by more extensive work proposed here.
4.2 Benchmark Model
4.2.1 The CAM5 Model. The Community Atmosphere Model version 5 (CAM5) is the atmospheric
component of the Community Earth System Model, version 1 (CESM1). It has a finite volume dynamic
core with a horizontal resolution of 1 degree and 30 vertical levels. In the CAM5, the large-scale cloud
and precipitation processes are parameterized with a prognostic two-moment bulk cloud microphysics
9
scheme of Morrison and Gettelman (2008). The moist boundary layer is parameterized based on
Bretherton and Park (2009), and shallow convection is based on Park and Bretherton (2009). The ZhangMcFarlane (Z-M, Zhang and McFarlane 1995) convection scheme with a dilution approximation for the
calculation of convective available potential energy (Neale et al. 2008) is used for deep convection. The
radiation scheme uses the Rapid Radiative Transfer Method for GCMs (RRTMG) described by Iacono et
al. (2008). The Z-M scheme bypasses all the details of the formation mechanisms of cloud water and
precipitation inside the convective updrafts, and the conversion of cloud water to rainwater is determined
through a tunable parameter. The rainwater is removed immediately from the updrafts either as surface
precipitation or through evaporation in the atmosphere.
We recently developed a
computationally efficient, physically
based
convective
microphysics
parameterization and incorporated it
into the Z-M scheme in the CAM5
(Song and Zhang 2011, Song et al.
2012). In addition to conventional
thermodynamic
processes,
the
scheme now considers many
microphysical processes (Fig. 10 for
a schematic). They include droplet
activation and ice nucleation by
aerosols, autoconversion of cloud
water/ice to rain/snow, accretion of
cloud water by rain, accretion of
cloud water, cloud ice, and rain by
snow,
homogeneous
and
heterogeneous freezing of rain to
Fig. 10. Schematic of convective parameterization including microphysics
form snow, Bergeron-Findeisen
process, fallout of rain and snow, condensation/deposition, self-collection of rain drops, and selfaggregation of snow. The inclusion of a convective cloud microphysics scheme allows for a more
realistic interaction between convection and large-scale clouds through hydrometeor exchange, which is
important for both cloud water/ice and water vapor distribution in the upper troposphere.
4.2.2 Attribution analysis of decadal variability of the atmosphere and surface at global and regional
scales. The decadal variability of climate system has profound impacts on our predictions of global and
regional climate change. However, the relative contributions of various dynamical and thermodynamic
processes to the total temperature anomalies remain largely unexplained. We propose to quantify the
contribution of radiative and non-radiative feedbacks to the temperature decadal variability using the
coupled atmosphere-surface climate feedback-responses analysis method (CFRAM). The CFRAM is a
novel climate feedback analysis method formulated by Lu and Cai (2009). The traditional climate
feedback analysis methods (e.g., partial radiative perturbation method, Wetherald and Manabe 1988)
focus on the feedback parameterizations (the ratio of the radiative perturbation at the TOA due to a
specific radiative feedback to the total change of the surface temperature) and cannot provide a direct
estimate of temperature change caused by any individual feedback. In contrast, the CFRAM considers
temperature changes both in the atmosphere and at the surface in the atmosphere-surface column as
climate response. In addition, the CFRAM considers the impacts of not only the radiative feedback but
also the non-radiative feedback, such as atmospheric transport and convection. We have applied the
CFRAM in the NCAR CCSM3.0 to understand the role of climate feedbacks in the global and regional
warming pattern formation and the land-sea warming contrast in response to a doubling of CO2 (Song et
al. 2013, Song et al. 2014). It was found that the CFRAM analysis is highly accurate in quantifying the
role of climate feedbacks in the spatial distribution of global and regional climate change simulated by
the CAM3 coupled to a slab ocean model. Fig. 11 shows an example from Song et al. (2014) of
10
Fig. 11: Total surface temperature changes in
low and high latitudes due to 2 x CO2 in the
NCAR CAM3 coupled to a slab ocean model
and attribution to direct CO2 forcing and
each feedback process using the CFRAM
analysis. The radiative feedbacks include
water vapor (q), cloud shortwave (cld_sw)
and longwave (cld_lw), and albedo. nonradiative feedbacks include dynamic
transport (dyn_adv), atmospheric convection
(conv), large-scale condensation (lg_cond),
PBL diffusion, surface sensible (shf) and
latent (lhf) heat fluxes, and oceanic
adjustment.
equilibrium surface temperature changes in the tropics and high latitudes due to doubling of CO2, and
their decomposition into contributions from different feedback processes. To attribute the decadal
variability of atmospheric and surface temperature to different physical, dynamical and cloud radiative
feedback processes, we will apply the CFRAM to the NCAR CAM5/CESM1 simulation to estimate the
model's climate feedbacks. We will identify contrasting periods of decadal variation and apply the
CFRAM analysis to them to understand what processes contribute the most to the decadal surface and
atmospheric changes. Su et al. (2014) show that CMIP5 models have a large spread in climate sensitivity.
High sensitivity models are in better agreement with observations. As can be seen from Fig. 10, a unique
advantage of CFAM is that it can attribute the temperature change due to doubling of CO2 to individual
processes such as cloud feedbacks. Thus, this work can help to identify factors responsible for the spread
of climate sensitivity as well.
4.2.3 Sensitivity of simulated decadal variability in the NCAR CAM5/CESM1 to convective
parameterization. Representation of convection in GCMs is one of the most difficult tasks in climate
modeling and prediction. Convection not only affects the cloud amount and distribution, its associated
hydrological cycle and condensational heating also affect the water vapor distribution and large-scale
circulation. It is well known that the simulations of the climatology, diurnal cycle, intraseasonal
variability (MJO), and ENSO in climate models are sensitive to the model convection parameterization.
However, the sensitivity of simulated decadal variability or climate prediction to convection
parameterization is much less known. We will investigate the sensitivity of simulated decadal variability
in the NCAR CAM5/CESM1 to convection parameterization. Many of the GCMs in IPCC AR4 have
systematic biases in the simulation of Intertropical Convergence Zone (ITCZ), El Nino and Madden
Julian Oscillations (MJO) (Dai 2006, Lin et al. 2006), and the situation did not seem to have improved
much in IPCC AR5. Zhang (2002) developed an improved convection parameterization scheme, and
simulations using it showed marked improvement in both the climatology and variability of tropical
circulation in the NCAR CAM3 and CCSM3 (Zhang and Mu 2005, for MJO; Li and Zhang 2008, for El
Nino, and Song and Zhang 2009, Zhang and Song 2010 for ITCZ). How sensitive is the simulated
decadal climate variability and climate sensitivity in CESM1 to convection parameterization? We will
address this issue by performing additional simulations with the improved Z-M convection
parameterization.
The observational study of Solomon et al. (2010) shows that from 1980 to about 2000 stratospheric
water vapor concentrations increased by about 1 part per million by volume (ppmv), followed by a drop
of about 0.4 ppmv thereafter. The radiation model calculation suggested that changes of that magnitude
could slow the rate of global surface temperature increase by 25% after 2000 and enhance the rate by
30% from 1980 to 2000. Recently Tselioudis et al. (2010) analyzed the satellite observations of tropical
weather States derived from ISCCP cloud property retrievals and found that the deep convection weather
State increased in frequency from 1983 to about 2000 and remained at a nearly constant level after that.
This indicates that the observed decadal changes in tropical deep convection could explain in part the
stratospheric water vapor variability patterns. In convection schemes, the entrainment rate of
environmental air into convective updrafts is an important factor that can affect the cloud top height and
11
therefore the transport of water vapor into the upper troposphere and stratosphere. Recently several
entrainment rate parameterizations have been proposed (see the review of De Rooy et al, 2013); we will
investigate the sensitivity of simulated decadal variability to different entrainment formulations in the ZM convection scheme in CAM5.
Song et al. (2012) show that convective microphysics parameterization in the NCAR CAM5
significantly improves the cloud ice and liquid water simulation in convective updrafts to bring them in
line with CloudSat observations. Since amount of cloud ice and liquid water in convective updrafts
affects the detrainment of hydrometeors into stratiform or anvil clouds, the direct effect of convective
microphysics parameterization in CAM5 is to alter the large-scale cloud ice and liquid water budgets.
This strongly affects the cloud distribution and cloud radiative forcing, as noted by Gettelman et al.
(2013). Thus, it likely will affect the decadal climate variability. As we test the sensitivity test of decadal
variability to entrainment in convective parameterization, we will also investigate its sensitivity to
convective microphysics. We will perform multi-decadal CAM5 simulations with and without convective
microphysics and use the output to analyze the sensitivity of decadal variability to convective
microphysics parameterization.
The model results will be compared to observations. In the moist region (M) in Fig. 3, the correlation
is less good (0.65). As Fasullo and Trenberth (2013a) did not explain, our guess about the weaker
correlation between RH and ECS in the moist regions is that models simulate RH very poorly. We can
study the impact of convective parameterization in the successive generations of NCAR models marked
by color symbols in Fig. 3. Note that the NCAR model, which used to have lower ECS than the GFDL
model (see Fig. 1 of Stephens 2005), gradually reduces the differences with the GFDL model in later
versions. Significant as it is, this progress by itself is not sufficient to improve climate prediction. AIRS
retrievals of RH in the just released Version 6 for 2002 - 2013 are much improved over the 2002-2007
data available to Fasullo and Trenberth (2012a). Comparison to increasingly refined data such as RH data
from AIRS provides the ultimate test of model paramerterizations.
4.2.4 New CAM5 and CAM5/CESM1 runs
In addition to existing CMIP5 model output, we will carry out a number of runs with the CAM5 and the
coupled CESM1 models to provide tests of some of the ideas and hypothesis discussed earlier.
Run 1a. CAM5 with SST fixed to pre-industrial climatology for 1×CO2; Run 1b is same as Run 1a with
4×CO2
Runs 2a and 2b. Repeat Run 1a and R1b using varying model physical parameterizations.
The above runs will be carried out with ensembles of 5. We will examine the circulation response to
direct CO2 forcing in these runs. The experiments with varying model physical parameterizations will
form an ensemble to examine the model spread in circulation response originated from model physics.
Run 3a. CAM5 150 yr from with 1×CO2, coupled to a slab ocean; Run 3b is same as 3a with 4×CO2.
Run 4a and 4b. Repeat Run 3a and 3b with model physical parameterizations varying as in Runs 2a and
2b.
We will examine the cloud feedbacks in these runs. The circulation responses derived from Run 2a and 2b
will be correlated with cloud feedbacks in Run 4a and 4b to prove our hypothesis that direct circulation
response to CO2 forcing is a primary driver for inter-model difference in cloud feedbacks.
Run 5a. CAM5/CESM1 with observed SST from 1980-present.
Run 5b. CAM5/CESM1 with varying model parameterizations and compare to observations.
These runs will be compared carefully against state-of-the-art observations. We will identify the optimal
model physical parameterizations that produce the best match to the observations. The climate sensitivity
with the optimal combinations of model parameterizations will be derived from Run 4a and 4b.
4.3 Investigation of Extreme Weather From Observations and Models. Precipitation is of utmost
importance to life on Earth and directly relevant to the societal impact of hydrological cycles on water
sources, agriculture, energy production, and recreation. Constraining model-simulated precipitation is
practically important and may also serve to constrain climate sensitivity. Here, we propose to seek the
connection between precipitation and ECS so as to reveal the interactions between water cycle
(precipitation) and energy cycle (temperature). Previous studies showed that precipitation changes are
12
spatially inhomogeneous when the globe warms (e.g., Neelin et al. 2006, Randall et al. 2007). The
patterns of precipitation changes tend to exhibit a rich-get-richer phenomenon (Chou and Neelin, 2004),
i.e., precipitation would increase over the climatological convective zones but decrease at the margins of
the convective zones. Thus, it makes sense to construct “climate-sensitive precipitation indices” based on
regional features, rather than global-mean.
To reveal extreme weather, we will first identify extreme events (e.g., drought and flooding) using the
precipitation data. The strength and duration of each event will be monitored using the precipitation data
from multiple measurements (GPCP, GPCC, and TRMM). The intercomparison among different data sets
of precipitation will be conducted to discover the most robust characteristics of extreme events. In
addition to the observational studies, we will also analyze the simulated precipitation from numerical
models (e.g., NCAR CAM5/CESM1). We will first analyze the characteristics of simulated extreme
events based on the precipitation from models. Then we will compare the characteristics of drought and
flooding between simulations and observations. Especially, we will investigate how good numerical
models are
in simulating
strengths
durations
of extreme
weather. Numerical
models will help us
Using
GRACE and
otherand
sensors
to detect
groundwater
storage changes
improve our understanding and forecasting of extreme weather.
High la tude precipita on increase
Alaska glaciers mel ng
Greenland ice sheet mel ng
Upper Midwestern U.S. flooding
California groundwater deple on
China groundwater deple on
Aral Sea shrinking
Tibetan Plateau snow increase
Middle Eastern groundwater deple on
Indian groundwater deple on
North African groundwater deple on
Southeastern U.S. drought
High Plains groundwater deple on
Indian monsoon
Mekong drought
Orinoco
o flods
Peruvian glaciers mel ng
Amazon drought
Southern Africa groundwater deple on
NW Australia groundwater deple on
Argen na drought
Guaranii aquifer deple on
Patagonia glaciers mel ng
Antarc c ice sheet mel ng
cm/yr
water)
for)
california)
ucchm.org)
Trend map courtesy of Felix Landerer, JPL
Fig. 12 Trend in ground water (cm/yr) measured by GRACE in the recent decade since launch. Negative values
(red) indicates drought. Based on unpublished JPL data from M. Watkins (2013, private communication).
We will also explore the intensity and duration of drought and flooding by examining some
atmospheric variables (e.g., temperature, water vapor, cloud fractions, vertical velocity, and liquid
water/ice water), which are potentially related to extreme events. NCEP2 air temperature and vertical
velocity, AIRS water vapor, cloud fractions and liquid water/ice water data from ISCCP, CloudSat, and
CALIPSO will be analyzed in the proposed research. We will analyze the statistical characteristics of
these atmospheric variables and compare them to the observational characteristics of extreme events to
see if there is any relation between them. Investigation of these atmospheric variables will enable a better
understanding of the favorable conditions for extreme events. We will also explore the variability of
GRACE (Tapley et al. 2004) underground water during the extreme events (drought/flooding). These
results can help us evaluate how the extreme weather influences the underground water, which is
important for surface subsidence. As shown in Fig. 12, the trends in the GRACE underground water
13
reveal that since 2002 large fractions of the globe, including southeastern parts of USA are stressed by
drought.
Finally, the recycling rate of atmospheric moisture, an important physical quantity related to the
hydrological cycle and hence extreme weather, will be investigated. The recycling rate of atmospheric
moisture is defined as the ratio of monthly precipitation (P) to monthly mean column water vapor (W)
(Chahine et al. 1997). In other words, the recycling rate of atmospheric moisture provides a simple
parameter to explore the comparison between precipitation and column water vapor, which was
extensively discussed in previous studies (Allen and Ingram 2002, Adler et al. 2003, 2008, Held and
Soden, 2006, Wentz et al. 2007, Gu et al. 2007, Stephens and Ellis 2008). The recycling rate of
atmospheric moisture can help us characterize the intensity of the atmospheric branch in the hydrological
cycle. The proposed investigations of the recycling rate and the related hydrological cycle will provide a
background for studies of extreme weather. Based on multiple observational data sets and numerical
simulations, we will examine the temporal variation of the recycling rate for extreme events. The
recycling rate will tell us the environment of atmospheric water during drought and flooding events. The
comparison of recycling rate between observations and simulations will further provide information about
how good models are in simulating the water cycling in the atmosphere.
5. Broader Implications and Relevance to NSF goals
The goals of the NSF solicitation are to improve modeling capabilities to:
(1) Achieve comprehensive, reliable global and regional predictions of decadal climate variability and
change … (3) Maximize the utility of available observational and model data for impact,
vulnerability/resilience, and risk assessments … (4) Effectively translate climate predictions and
associated uncertainties into the scientific basis for policy and management decisions … Our proposed
work will address objectives (1), (3) and (4).
The broader implications of our study include a deeper understanding of the cloud feedback processes,
and on that basis a number of physically consistent metrics that are crucial to climate sensitivity can be
derived and tested against observations. This will lead, through an improved simulation capability, to a
reduction of uncertainties associated with global model projections of climate change, especially in
extreme weather, with profound implications for adaptation and the science that should inform decisions.
According to Busby (2008). “While some of these purported threats-abrupt climate change and sea-level
rise-have been overstated by advocates, several concerns, mostly related to the effects of extreme weather
events on the United States and its strategic interests overseas, are sufficient enough that they already
constitute security threats.”
The broader educational impacts of the proposed research will be realized through promotion of
teaching, and training of graduate students, postdoctoral fellows and other junior scientists, and broad
dissemination of results through presentations, peer-reviewed publications and via the Internet. In
addition, the methodology and results from this study will be incorporated in graduate-level courses.
14