Transcript Issues and Challenges

Credibility of Climate Model Projections of Future
Climate: Issues and Challenges
Linda O. Mearns
National Center for Atmospheric Research
SAMSI 2011-12 Program on Uncertainty Quantification
Pleasanton, CA, August 29, 2011
“Doubt is not a pleasant condition, but
certainty is an absurd one.”
-Voltaire
How can we best evaluate the
quality of climate models?
Different Purposes
• Reasons for establishing reliability/credibility
– for recommending what scenarios should be used for
impacts assessments,
– for selecting which global models should be used to
drive regional climate models,
– for differential weighting to provide better measures of
uncertainty (e.g, probabilistic methods)
• Mainly going to discuss this in terms of multimodel ensembles (however, there are important
limitations to the types of uncertainty that can
be represented in MMEs)
Reliability, confidence,
credibility
• What do we mean?
• For a long time, climate modelers/analysts
would make statements that if model
reproduced observations ‘well’ then we
had confidence in future projections – but
this really isn’t adequate – synonymous
with Gabbi’s naïve view
Average Changes in Temperature and Precipitation
over the Grid Boxes of the Lower 48 States
Three Climate Model 2xCO2 Experiments
Smith et al., 1989
Global Model Change in
Precipitation - Summer
Relationship to Uncertainty
•Historically (and even in the most recent IPCC
Reports) each climate model is given equal weight in
summarizing model results. Does this make sense,
given different model performances?
•Rapid new developments in how to differentially
weight climate model simulations in probabilistic
models of uncertainty of regional climate change
REA Method
• Summary measure of regional climate
change based on weighted average of
climate model responses
• Weights based on model ‘reliability‘
• Model Reliability
Criteria:
• Performance of AOGCM (validation)
• Model convergence (for climate change)
Giorgi and Mearns, 2002, 2003
REA Results for Temperature
A2 Scenario DJF
Summary
• REA changes differ from simple averaging
method
– by few tenths to 1 K for temperature
– by few tenths to 10% for precipitation
• Uncertainty range is narrower in the REA
method
• Overall reliability from model performance
was lower than that from model
convergence
• Therefore to improve reliability, must
reduce model biases
Tebaldi &
Knutti,
2007
Search for ‘correct’ performance
metrics for climate models –
where are we
• Relative ranking of models varies depending on
variable considered - points to difficulty of using
one grand performance index
• Importance of evaluating a broad spectrum of
climate processes and phenomena
• Remains largely unknown what aspects of
observed climate must be simulated well to
make reliable predictions about future climate
Gleckler et al., 2008,
JGR
Model Performance over
Alaska and Greenland
• RMSEs of seasonal cycles of temperature,
precipitation, sea level pressure
• Tendency of models with smaller errors to
simulate a larger greenhouse gas warming over
the arctic and greater increases in precipitation
• Choice of subset of models may narrow
uncertainty and obtain more robust estimates of
future climate change in Arctic
Walsh et al., 2008
Selection of ‘reliable’
scenarios in the Southwest
• Evaluation of CMIP3 models for winter
temperature and precipitation (using
modified Giorgi and Mearns REA method)
• Reproduction of 250 mb geopotential
height field (reflecting location of
subtropical jet stream)
• Two models (ECHAM5, HadCM3 score
best for these three variables)
Dominguez et al., 2010
Dominguez et al., 2010
SW Reliability Scores
Dominguez et al., 2010
Dominguez et al., 2010
Studies where selection
did not make a difference
• Pierce et al., 2009 - future average
temperature over the western US – 14
randomly selected GCMs produced results
indistinguishable from those produced by
subset of ‘best’ models.
• Knutti et al., 2010 – metric of precipitation
trend, 11 randomly selected GCMs,
produced same results as those from 11
‘best’ GCMs.
ENSEMBLES
Methodological Approach
Six metrics are identified based on ERA40-driven
runs
– F1: Large scale circulation and weather regimes
(MeteoFrance)
– F2: Temperature and precipitation meso-scale signal
(ICTP)
– F3: Pdf’s of daily precipitation and temperature (DMI,
UCLM,SHMI)
– F4: Temperature and precipitation extremes (KNMI, HC)
– F5: Temperature trends (MPI)
– F6: Temperature and precipitation annual cycle (CUNI)
6
n
i

f

w
R
i
i

1
The result
0,30,3
0,25
0,20,2
Wprod
Wredu
Wrank
0,15
0,10,1
0,05
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Christensen et al., 2010
The result
RCM
f1
f2
f3
f4
f5
f6
Wprod Wredu Wrank
C4I-RCA3.0
0,058 0,050 0,067 0,044 0,066 0,069 0,026 0,058 0,057
CHMI-ALADIN
0,071 0,058 0,067 0,070 0,060 0,069 0,054 0,066
CNRM-ALADIN
0,069 0,059 0,067 0,113 0,066 0,061 0,084 0,064 0,066
DMI-HIRHAM5
0,068 0,039 0,066 0,062 0,070 0,068 0,035 0,066 0,053
ETHZ-CLM
0,075 0,073 0,067 0,036 0,059 0,069 0,038 0,067 0,073
ICTP-RegCM3
0,073 0,112 0,065 0,066 0,069 0,067 0,112 0,075 0,073
KNMI-RACMO2
0,070 0,137 0,069 0,132 0,066 0,068 0,268 0,094 0,123
Met.No-HIRHAM
0,070 0,041 0,067 0,057 0,065 0,067 0,032 0,064 0,055
Meto-HC-HadRM3Q0
0,061 0,048 0,067 0,054 0,071 0,066 0,034 0,063 0,057
Meto-HC-HadRM3Q3
0,061 0,049 0,066 0,030 0,064 0,062 0,016 0,047 0,036
Meto-HC-HadRM3Q16
0,061 0,051 0,067 0,080 0,073 0,066 0,055 0,069 0,071
MPI-REMO
OURANOSMRCC4.2.3
0,068 0,072 0,066 0,038 0,069 0,069 0,039 0,068 0,063
SMHI-RCA3.0
0,057 0,053 0,067 0,054 0,067 0,069 0,035 0,063 0,062
UCL-PROMES
0,067 0,068 0,067 0,099 0,070 0,065 0,096 0,074 0,073
0,08
0,072 0,089 0,065 0,063 0,065 0,066 0,077 0,065 0,057
An application for 20202050
• Changes for European capitals 2021-2050
(Déqué, 2009; Déqué & Somot 2010)
• 17 RCMs in 5(7) GCMs
• Convert discrete data set into a continuous
PDFs of climate change variables.
– This is done using a Gaussian Kernel algorithm
applied to the discrete dataset with the aim to take
into account also the model specific weights
Temperature Pdf Climate Change
Pdf of daily temperature (°C) for DJF (left) and JJA (right) for the
1961-1990 (solid line) and 2021-2050 (dash line) with
ENSEMBLES weights (thick line) and for a single model
based on median Ranked Probability Score (thin line).
Deque & Somot 2010
To weight or not to weight
• Recommendation from IPCC Expert
Meeting on Evaluating MMEs (Knutti et al.
2010)
– Rankings or weighting could be used to select
subsets of models
– But useful to test statistical significance of the
difference between models and the subset vs.
the full ensemble to establish if subset is
meaningful
– Selection of metric is crucial - is it a truly
meaningful one from a process point of view?
More process-oriented
approaches
Process-oriented approaches
• Hall and Qu snow-albedo feedback
example
• What we are trying in NARCCAP
SNOW ALBEDO FEEDBACK
In the AR4 ensemble,
intermodel variations in snow
albedo feedback strength in
the seasonal cycle context are
highly correlated with snow
albedo feedback strength in
the climate change context
It’s possible to calculate an
observed value of the SAF
parameter in the seasonal cycle
context based on the ISCCP data
set (1984-2000) and the ERA40
reanalysis. This value falls near
the center of the model
distribution.
observational
estimate based
on ISCCP
95%
confidence
interval
It’s also possible to calculate an
estimate of the statistical error in
the observations, based on the
length of the ISCCP time series.
Comparison to the simulated
values shows that most models
fall outside the observed range.
However, the observed error
range may not be large enough
because of measurement error in
the observations.
(Hall and Qu, 2007)
What controls the strength of snow albedo feedback?
snow cover component
snow metamorphosis component
It turns out that the snow cover component is overwhelmingly responsible not
only for the overall strength of snow albedo feedback in any particular model,
but also the intermodel spread of the feedback.
Qu and Hall 2007a
Establishing Process-level
Differential Credibility of Regional
Scale Climate Simulations
Determining through in depth process-level analysis
of climate simulations of current (or past) climate,
the ability of the model to reproduce those aspects
of the climate system most responsible for the
particular regional climate; then analyzing the
model response to future forcing and determining
specifically how model errors in the current
simulation affect the model’s response to the future
forcing. Which model errors really matter?
Essentially it is a process-based integrated expert
judgment of to what degree the model’s response to
the future forcing is deemed credible.
The North American Regional Climate
Change Assessment Program (NARCCAP)
www.narccap.ucar.edu
•Explores multiple uncertainties in regional
and global climate model projections
4 global climate models x 6 regional climate models
• Develops multiple high resolution regional (50 km,
30 miles) climate scenarios for use in impacts and
adaptation assessments
•Evaluates regional model performance to establish
credibility of individual simulations for the future
•Participants: Iowa State, PNNL, LNNL, UC Santa Cruz, Ouranos
(Canada), UK Hadley Centre, NCAR
• Initiated in 2006, funded by NOAA-OGP, NSF, DOE, USEPA-ORD –
5-year program
NARCCAP Domain
Organization of Program
•
Phase I: 25-year simulations using NCEP-Reanalysis boundary
conditions (1980—2004)
•
Phase II: Climate Change Simulations
– Phase IIa: RCM runs (50 km res.) nested in AOGCMs current
and future
– Phase IIb: Time-slice experiments at 50 km res. (GFDL and
NCAR CAM3). For comparison with RCM runs.
•
Quantification of uncertainty at regional scales – probabilistic
approaches
•
Scenario formation and provision to impacts community led by
NCAR.
•
Opportunity for double nesting (over specific regions) to include
participation of other RCM groups (e.g., for NOAA OGP RISAs,
CEC, New York Climate and Health Project, U. Nebraska).
Process Credibility
Analysis of the Southwest
M. Bukovsky
What do we need to make
further progress?
• Many more in depth process oriented
studies that examine plausibility of process
change under future forcing - what errors
really matter and which don’t?
What is the danger of false certainty?
The End
How is this done?
• Using performance metrics – e.g., Gleckler
et al., 2008, Reichler and Kim, 2010
– For variously weighting schemes
– For selecting good (or bad) performers
– Example from ENSEMBLES program
(Christensen et al.)
– In probabilistic or selection approaches (does
selection or weighting really provide different
estimates? Usually not - but, example of
Walsh et al., Dominguez et al.