Probabilistic Climate Change Predictions

1
Probabilistic Climate Change Predictions Murphy,
Collins, Sexton, Webb, Harris, Barnett, Jones
2
Motivation

• probabilistic climate predictions are required
  for quantitative impact and risk assessments
• need a probability distribution for any climate
  variable at any location
• implies a need for large ensembles of GCM
  simulations

3
Sources of uncertainty

• future changes in greenhouse gas and aerosol
  emissions
• effects of natural variability
• modelling of Earth system processes

4
CO2 in IPCC EMISSIONS SCENARIOS
A1FI A2 B1 B2
5
Ensemble of initial conditions
HadCM2 GHG
6
Ensemble of climate models
7
PROBABILISTIC CLIMATE PREDICTIONS
required position
current position
100
Probability
Probability
0 20 40 60
0 20 40 60
2080s SE England winter rainfall
2080s SE England winter rainfall
8
Modelling uncertainties
X CLIMATE MODELS
ATMOSPHERE CLOUDS
Probability
LAND AND VEGETATION
CARBON CYCLE
OCEAN CIRCULATION
ETC.
0 20 40 Increase
in winter rainfall
9
Types of parameterisation uncertainty
Consider a simple parameterisation C
?.f(G) C (e.g.) grid box mean cloud cover, soil
moisture, turbulent momentum flux f(G) some
function of the GCM primary grid box variables
(T,q,u,v,p) ? tunable parameter
Parameter uncertainty C ??.f(G) Structural
uncertainty C ?.f?(G) Stochastic uncertainty
C ?.f(G) ?, where ? varies with time to
represent the effect of variations in sub-grid
scale organisation on the grid scale solution.
10
Long term aim

• Probabilistic predictions for specific 21st
  century periods
• Range of emissions scenarios and initial
  conditions
• Cover stochastic, parameter and structural
  uncertainties
• Cover uncertainties arising from atmosphere,
  ocean, chemistry and ecosystem processes
• Enormous ensemble of transient simulations using
  an Earth system model
• Too expensive !

11
A first large ensemble climate prediction

• Focus on equilibrium response to doubled CO2
• Focus on uncertainties in physical atmospheric
  processes
• Consider parameter uncertainties
• Can use the atmospheric GCM coupled to a simple
  mixed layer ocean
• Cheaper model -gt larger ensemble

12
Experimental design

• Choose a subset of HadAM3 parameters known to
  influence key physical atmospheric processes
• Identify plausible maximum and minimum values
• Perturb parameters one at a time to create 53
  versions of the model
• Make 53 predictions of the equilibrium response
  to a doubling of CO2

13
Parameter Perturbations

• Large Scale Cloud
• Ice fall speed
• Critical relative humidity for formation
• Cloud droplet to rain conversion rate and
  threshold
• Cloud fraction calculation

• Boundary layer
• Turbulent mixing coefficients
  stability-dependence, neutral mixing length
• Roughness length over sea Charnock constant,
  free convective value

• Convection
• Entrainment rate
• Intensity of mass flux
• Shape of cloud (anvils) ()
• Cloud water seen by radiation ()

• Dynamics
• Diffusion order and e-folding time
• Gravity wave drag surface and trapped lee wave
  constants
• Gravity wave drag start level

• Radiation
• Ice particle size/shape
• Cloud overlap assumptions
• Water vapour continuum absorption ()

• Land surface processes
• Root depths
• Forest roughness lengths
• Surface-canopy coupling
• CO2 dependence of stomatal conductance ()

• Sea ice
• Albedo dependence on temperature
• Ocean-ice heat transfer

14
Results from a first large ensemble climate
prediction

• Equilibrium response to doubled CO2 from 53
  versions of HadAM3 coupled to mixed layer ocean
• Ensemble of changes in 20 year mean fields
• Component of variance due to natural variability
  estimated from a 600 year experiment with
  standard model version sliced into 20 year chunks

15
Ensemble mean change divided by ensemble standard
deviation of changes for 20 year mean fields in
DJF
Ensemble mean change is highly robust for surface
temperature but varies widely with location for
precipitation and circulation
16
Ensemble Standard Deviation of changes in 20 year
mean fields in DJF
Compare the red and blue curves to see the
average contribution of natural variability to
the variance of changes at individual grid points
Compare the red and black curves to see the
extent to which the ensemble variance can be
reconstructed by scaling the change pattern of
one member up and down according to values of
climate sensitivity
17
(No Transcript)
18
Changes in regionally averaged annual
precipitation ()
Effect of natural variability
Effects of natural variability and modelling
uncertainties
S Europe
Amazonia
19

• A Climate Prediction Index
• Measures how well model reproduces relevant
  observed properties
• Use to identify plausible model versions and for
  likelihood-based weighting of their predictions

20
Is the CPI relevant to climate change prediction ?
Good predictor
Bad predictor
Bad skill
Good skill

• Evaluating mean present-day climate is a
  NECESSARY BUT NOT SUFFICIENT condition for a good
  prediction.

21
The next stage getting to probabilistic
predictions

• Ensembles where several parameters are perturbed
  simultaneously, taking a range of values.
• Need to sample parameter space in a comprehensive
  and unbiased manner.
• Generate an ensemble 100 members that span
  parameter space in an unbiased, cost-effective
  way.

22
Selecting multiple parameter perturbations

• Current ensemble of single parameter
  perturbations allows us to predict reliability of
  models with multiple parameter perturbations
  using a linear assumption.
• Choose a set of plausible model versions which
  achieve maximum coverage of parmeter space

23
Parameter values in the estimated top 50 model
versions
24
Predicting response of multiple parameter
perturbations

• Making linear assumption, we can also estimate
  the response of model versions containing
  multiple parameter perturbations.
• Can produce pdfs now, subject to this assumption.

25
Model-independent pdf from Gregory et al 2002
The black and red pdfs are obtained by combining
the impact of individual parameter perturbations
using linear statistical estimation The required
assumptions have been verified by actually
running a few model versions with multiple
parameter perturbations.
Murphy, Sexton, Barnett, Jones, Webb, 2003
26
Pdfs of regional changes

• Linear assumption unreliable
• Wait for the ensemble with multiple perturbations
  to run
• In progress

27
Probabilistic predictions for 21st century

• Need to scale 2xCO2 pdfs to obtain time dependent
  pdfs for response to a given emissions scenario
• Calibrate scaling relationships for a small
  subset of ensemble members
• Simulate response to aerosol forcing using
  atmosphere/mixed layer ocean
• Simulate pre-industrial to 2100 using full
  coupled model
• Develop scaling relationships
• Apply scaling to large ensemble of 2xCO2
  responses to infer probabilistic predictions of
  21st century changes
• Extend to regional scale using the RCM

28
Future Work Consider additional sources of
uncertainty

• structural parameterisation uncertainty
• stochastic parameterisation uncertainty
• uncertainties from additional Earth System
  modules

29
Structural uncertainty could be large
Parameter uncertainty
Structural uncertainty (multi-GCM ensemble)

• Coordinated multi-GCM ensemble predictions
• Make structural perturbations to a single GCM

30
Change in cloud radiative forcing for different
ensembles of models. Less variety in QUMP
ensemble.
31
Why consider stochastic uncertainties ?

• Additional source of uncertainty which must be
  accounted for
• A cheap way of realising the benefits of high
  resolution ?

300km GCM
Mean sea level pressure in winter
100km GCM
Observed
32
Uncertainty in additional Earth System modules

• Over the next few years we plan to address
• ocean physics
• terrestrial carbon cycle
• sulphate aerosol forcing

33
Matrix of uncertainties

Structural Parameter Stochastic Atmospheric
physics No Yes
No Ocean physics
No Yes
No Terrestrial carbon cycle No
Yes No Ocean carbon cycle
No No
No Sulphur cycle
No Yes No Other
atmospheric chemistry No No
No
Yes In the plan No Not yet in the plan
34
Ensemble prediction system for seasonal-centennial
probabilistic forecasts

• Why ?
• Improve decadal predictions by sampling
  modelling uncertainties in ensemble design.
• Improve centennial predictions by using forecast
  skill as an additional test of model reliability.
• Use verification of decadal forecast
  probabilities to build confidence in
  unfalsifiable centennial forecast probabilities

35
Worldwide approach needed...

• Climateprediction.net
• ENSEMBLES FP6
• IPCC AR4
• QUEST
• Etc.