Probabilistic climate change predictions from GCMs

1
Probabilistic climate change predictions from GCMs

• Three broad methodological categories in the
  literature at present
• Direct use of GCM ensembles without formal
  application of observational constraints
• Methods designed to be constrained by
  observations, and, as far as possible,
  model-independent
• Methods designed to give results dependent on
  both observational constraints and GCM ensemble
  distributions

2
Climate sensitivity distributions from different
methods
Results from reduced complexity models,
constrained by historical temperature changes
GCM ensembles constrained by present day
climatology
Unweighted GCM ensembles
3
Distributions for regional- or continental-scale
transient changes

• Published/submitted approaches currently
  include
• Curve-fitting to multimodel ensembles (Raisanen,
  2005)
• Bayesian methods applied to multimodel ensembles
  (Tebaldi et al 2005 Furrer et al, 2005)
• Scaling to infer transient response from
  perturbed physics ensembles of equilibrium
  climate change (Harris et al, 2005)
• Projecting uncertainty in the historical response
  into the future to get model-independent
  estimates of continental-scale warming (Stott et
  al, 2005)

4
Example Multi-model curve fit
Main caveats no weighting sensitive to outliers
(because the ensemble is small) the ensemble
was not designed with sampling of uncertainties
in mind
5
What Observations Should We Use?
Possible observational constraints for weighting

• A broadly-based global estimate of the likelihood
  of the model
• A subset of regional variables to weight
  predictions of regional change
• Some optimal selection of time-mean fields
• Historical trends in temperature, ocean heat
  content etc.
• Attributable warming
• Cooling following volcanic eruptions
• ENSO response
• Palaeoclimate change (LGM, 6K, )
• Verification of ensemble predictions on shorter
  time scales (NWP/seasonal/decadal)

6
Some issues in assigning variable weights to an
ensemble of models

• Estimates of relative likelihood need to account
  for
• Noise arising from internal variability in
  simulations and observations
• Other observational errors (measurement biases,
  etc)
• Model inadequacy an additional uncertainty in
  our predictions which arises from structural
  model deficiencies. It is potentially very
  important, but difficult to quantify and yet to
  be tackled in climate prediction

Knutti and Meehl, (2005)
e.g., If we think an ensemble of model variants
can predict climate change in the real world, it
should be able to predict changes in other models
with alternative structural assumptions
7
Summary

• Significant progress since the IPCC TAR, but much
  still to do.
• Results depend on choices such as
• what uncertainties are sampled in the chosen
  ensemble of models
• the constraining observations
• prior distributions for uncertain inputs
• whether or not, and how, to weight ensemble
  members
• what should control the range of outcomes (e.g.
  observational uncertainties, ensemble spread).
• No obvious basis for choosing a best method
  yet, but the sensitivity of results to
  experimental choices should be quantified as far
  as possible.