Transcript Haywood_LSAR_2012

Quantifying uncertainty in
simulations of past, present and
future climate
Alan M. Haywood
School of Earth & Environment, University of Leeds
What inspires me...
Combining geological data and models
Rationale
• Evidence from observations (palaeo), climate models
and basic understanding indicates that climate can
change and anthropogenic climate change is real
• Society needs scientific evidence in order to act
• Consider types of actions
– Mitigation - reduce emissions of greenhouse gases
– Adaptation – learn to live with climate change that
we are already committed to
– If it gets too bad what about Geoengineerg?
Mitigation
• United Nations Framework Convention on Climate
Change (UNFCC) remit to limit greenhouse gas
concentrations to avoid “dangerous” climate change
• Need to know what a “safe” level of greenhouse gas
concentrations is:
– Sensitivity of the global climate system to different
levels of GHGs (climate sensitivity)
– Relationship between GHG emissions and
concentrations (carbon cycle feedback)
– Risk of dangerous/abrupt/rapid/irreversible events
e.g. shutdown of Atlantic Meridional Circulation,
melting of Greenland, death of Amazon rainforest...
Adaptation
• Society needs to adapt to some level of climate change
that is inevitable
• Adaptation decisions are in the hands of many different
bodies e.g. governments, water companies, energy
companies, large and small commercial businesses,
farmers, individuals.
• Requires:
– Information about climate change at regional and
local scales
– Multivariate information; temperature, precipitation,
winds, fluxes, etc.
– Information about extreme events; storms, droughts,
heatwaves etc.
Uncertainties
Global mean projections
from different models
using the same GHG
concentrations are
different
Global mean carbon cycle
feedbacks from different
models using the same GHG
emissions are different
Figure 10.5
Source: IPCC Fourth Assessment Report
Uncertainties
Regional patterns of change from
different models are different
Figure 10.9
Source: IPCC Fourth Assessment Report
Uncertainties
Models differ in their projection
of dangerous events
Figure 10.15
Source: IPCC Fourth Assessment Report
Why do uncertainties exist?
• Models have “errors” i.e. when simulating present-day
climate and climate change, there is a mismatch
between the model and the observations
• Differences in model formulation can lead to
differences in climate change feedbacks
How to deal with uncertainties?
Continue to improve models until
global and regional projections
converge
Climate change already happened
by the time models converge
Use techniques other than
comprehensive climate modelling
Cannot extrapolate from noisy (possibly
non-existent) time series
Is convergence a useful indicator of
reliability?
Simple models cannot provide all the
information
Climate change may not be linear
Combine information from climate
models, observations (+ palaeo)
and understanding to quantify the
uncertainty in projections
Risk-based approached used in
other disciplines where scientific
uncertainties exist
Climate Change
Projection/Retrodiction
• In the presence of uncertainties in climate model
projections adopt a probabilistic approach
• Sources of uncertainty:
– Initial conditions, natural variability
– Boundary conditions, emissions/concentrations of
greenhouse gases and other forcing agents
– Model errors and uncertainty, different models
giving different projections
• Probabilistic climate projections (for e.g. 2100) cannot
be easily verified in the way that probabilistic weather
forecasts are. Challenges in the world of palaeo
data/model comparisons too.
We did the wrong experiment
DMC
Triangle
Climate Change Projection/Retrodiction
• …however, we can still use ensemble and probabilistic
techniques in climate change projection/retrodiction
• Need a different strategy for generating the ensemble as
initial conditions are not the leading source of uncertainty
– The “multi-model” ensemble, MME
– The “perturbed-physics” ensemble, PPE
• Need something in place of the verification cycle –
assume that models which are good at reproducing
observed/palaeo climate change are also good at
simulating future climate change
Types of ensembles in palaeoclimate
• Boundary condition ensembles: understanding palaeoclimate
– Too many to mention (Valdes et al. etc)
• Multi-model ensembles (MMEs): understanding models
– PMIP, PlioMIP
• Perturbed parameter ensembles: quantifying model uncertainties
– Calibration of models for future climate prediction
• updating model parameters
– Physically-based reconstruction of palaeoclimate
• updating model state
– EBM: Hegerl et al. (2006)
– EMIC: Schneider von Deimling et al. (2006)
– GCMs slab: Annan et al. (2005); Hargreaves et al. (2006); CPDN H. Muri
PhD
– GCMs low res: CPDN Millennium; Gregoire et al. (2010)
– GCMs: Brown et al. 2007, Pope et al. (2011); Stone et al. (submitted);
Edwards et al. (in prep.); Valdes/Sagoo et al....
Multi-Model Ensemble
• A collection of the world’s climate models
• Sometimes called an “ensemble of opportunity”
• Currently coordinated by projects like CMIPCoupled Model Intercomparison Project and
housed at PCMDI, California, PMIP (LSCE, France)
• A relatively large “gene-pool” of possible
models, although it is common to share some
components
• Models are “tuned” to reproduce observed data
– although formal tuning is not performed
Perturbed Physics Ensemble
• Take one model structure and perturb uncertain
parameters and possible switch in/out different
subroutines
• Can control experimental design, systematically
explore and isolate uncertainties from different
components
• Potential for many more ensemble members
• Unable to fully explore “structural” uncertainties
• HadCM3 widely used (MOHC and
climateprediction.net) but other modelling groups
are dipping their toes in the water
For PPE’s and MME’s think cars!
Comparison of MMEs and PPEs
Global mean
temperature
change in MMEs
and PPEs under
different
scenarios
PPEs capable of
sampling global
response
uncertainties
Some Notation
y  f(x)
y = {yh,yf} historical and future climate variables (many)
f = model (complex)
x = uncertain model input parameters (many)
o = observations (many, incomplete)
• Our task is to explore f(x) in order to find y which will be
closest to what will be observed in the past and the future
(conditional on some assumptions)
• Provide probabilities which measure how strongly different
outcomes for climate change are supported by current
evidence; models, observations and understanding
Probabilistic Approach
future climate
{yh, yf}  f(x)
p( y | o)  p( y ) p(o | y )
yf
f(x1)
input space
f(x2)
x1
x
x2
f(x1)
f(x2)
o
yh
historical/observable climate
Bayesian Probabilities
• The probability expresses the
uncertainty in the prediction
(e.g. p(ΔT2100 > 6ºC)=0.05) not
the frequency of occurrence
of a particular event (ΔT2100 >
6ºC, 5% of the time)
• Fundamentally different to a
weather or seasonal forecast
prediction (which can be
verified)
• Probabilities are conditional on
assumptions; emissions
pathways for example
© Crown copyright Met Office
PPEs and the Green Sahara
PalaeoQUMP 17 HadCM3
Climateprediction.net ~60
HadSM3 (H. Muri PhD)
MMEs and PPEs and the Last Glacial
Maximum
?
MARGO
Updated from Edwards et al. (2007), Prog Phys Geog
State-dependence of uncertain
parameters
Feedback parameter (–Q/ΔT)
best
models
can constrain sea ice parameter with LGM cooling
but less relevant for warming scenario
FAMOUS PPE (Gregoire et al., 2010, Clim Dyn)
most show greater sensitivity to warming than
cooling
PalaeoQUMP 17 HadCM3 PPE
MIROC3.2 slab PPE (Annan, pers. comm.)
PMIP2 MME (Crucifix, pers. comm.)
Types of approach using PPEs
• Picking the best
– maximum likelihood, confidence sets (by any
other name...)
• Downweighting the worst
– reweighting with skill scores
• Probabilistic calibration or predictions
– reweighting within statistical framework
Picking the best model(s)
10 of 100 FAMOUS Gregoire et al. (2010) Clim Dyn
Pifalls, questions, looking forward
• Scientific
– Need reliable uncertainties of proxy-based reconstructions
– Good experimental design to avoid (much, much) pain later
– Important to learn from others about important parameters
– Different parameter sensitivity and variability in palaeoclimates
– Advantages / disadvantages of flux correction
– Maximum likelihood vs reweighting vs probability distributions
– How to estimate model uncertainty (parametric and structural)
• Technical
– Mistakes are amplified, propagated by N
– Problem of spin-up is multiplied by N
– Sufficient person power to analyse data
– Share tools to strip data and automate analysis
– Give simulations citable DOI (BADC) to crowd-source analysis