© UKCIP 2006 © Crown copyright Met Office Probabilistic climate projections from the decadal to centennial time scale WCRP Workshop on Regional Climate,

© U

KC

IP 2

006

© Crown copyright Met Office

Probabilistic climate projections from the decadal to centennial time scale

WCRP Workshop on Regional Climate, Lille, June 2010

James MurphyMet Office Hadley Centre

© U

KC

IP 2

006

Contents

• Sources of uncertainty

• Survey of alternative approaches

• More detail on a methodology developed at MOHC for UK climate projections and ENSEMBLES

• Interpretation and limitations

Emissions uncertainty

Internal variability

Uncertainty in climate projections

Te

mp

era

ture

Ch

an

ge

(d

eg

C) High Emissions

Low Emissions

E&W Precipitation

Pre

cip

ita

tio

n A

no

ma

ly (

mm

)

Source: UKCIP02

our incomplete understanding of climate processes and inability to model them perfectly

Modelling uncertainty

Change in summer precipitation (%), 2080-99 relative to 1980-99, SRES A2, IPCC AR4 models

Multi-model ensembles (MMEs)

Key Strengths

• Each member extensively tested – credibility derived from tuning and validation against a wide range of observables

• Constructed from a large pool of alternative components – samples different structural assumptions

• The source of much of our knowledge of projected future changes

Some Limitations

• Not designed to sample modelling uncertainties in a systematic fashion (“ensemble of opportunity”). No obvious “best” way of determining the distribution of possible changes of which the MME is a sample.

• Rather small. Difficult to get robust estimates of most likely changes, or associated uncertainties, in noisy quantities like regional changes in extreme events

Probability distributions of regional temperature changes from multi-model ensembles

• Substantial assumptions needed to convert the ensemble results into probabilities

• Different methods make different assumptions and get different results.

• e.g. Can errors in each model realisation of future climate be assumed independent, and randomly distributed about the true, unknown future ?

2080-99 relative to 1980-99, SRES A1B, Mediterranean Basin, derived from AR4 models. From Tebaldi and Knutti, 2007

DJF JJA

© U

KC

IP 2

006

Errors common to all models

Systematic (black) and random (white) contributions to errors in climate mean spatial fields of different climate variables in a multimodel ensembles of atmosphere-mixed layer ocean models.Collins et al (2010, in press). See also Knutti et al (2010).

Caveat: Part of the apparent systematic component may actually arise from observational biases.

Probabilistic projections derived from GCM-RCM matrix experiments

JJADJF

Déqué (2009), Déqué and Somot (2010)

ASK - An alternative approach

• Aim to produce probabilities which are as model-independent as possible, and determined by uncertainties in observations of historical climate.

• Idea is to develop robust, well-understood transfer functions which link something we want to predict with some physically related observable.

• Often termed the “ASK” approach – see, e.g., Allen et al (2000), Stott and Kettleborough (2004), Allen and Ingram (2005), Piani et al (2005), Stott et al (2006), etc..

• The transfer function (“emergent constraint”) needs to be robust across different models.

• May be harder to find robust emergent constraints for regional variables

Observationally-constrained pdfs of the transient climate response

Obtained from optimal fingerprint analysis: calculate a distribution of factors by which the simulated historical warming attributable to greenhouse gases can be scaled up and down while remaining consistent with observations, and assume that fractional errors in the historical response remain unchanged in future.

Applying the observational constraint scales up the best-estimate response of the low sensitivity model (green star), bringing it closer to the other models. The model dependence is not totally removed, however.

Stott et al., 2006

© U

KC

IP 2

006

Another alternative approach based on perturbed physics ensembles

• Relatively large ensembles designed to sample modelling uncertainties systematically within a single model framework

• Executed by perturbing model parameters controlling key model processes, within expert-specified ranges

• Key strength: Allows greater control over experimental design cf multi-model “ensembles of opportunity”

• Key limitation: does not sample “structural modelling uncertainties”, e.g. changes in resolution, or in the fundamental assumptions used in the model’s parameterisation schemes – need to include results from other models to account for these.

• Describe an implementation based on the HadCM3 model

© Crown copyright Met Office

Structural modelling errors

Large perturbed physics ensemble sampling uncertainties in time-evolving 21st century climate change at high spatial resolution

Probabilistic projections

Observational constraints

Ideal system for probabilistic projections based on perturbed physics ensembles

Computational resources can’t support this yet, so the method involves a larger set of affordable steps

© U

KC

IP 2

006

Inputs to probabilistic projections for UKCP09

ObservationalConstraints

StructuralModel Errors

RegionalClimate Model Carbon Cycle

Atmosphere

Sulphate Aerosol

Ocean

ProbabilisticClimate

Projections

© U

KC

IP 2

006

Three stages

• Probabilistic projections of the equilibrium climate change in response to doubled CO2 at 300km resolution

• Further steps to obtain probabilistic projections of time-dependent climate change at 300km resolution

• Downscaling to obtain projections at 25km resolution

© U

KC

IP 2

006

Simulations of equilibrium climate change

• Used the atmosphere-mixed layer (“slab”) ocean configuration of the model, HadSM3

• Obtained expert-specified prior distributions for multiple (31) uncertain model parameters controlling surface and atmospheric physical processes

• Ran an ensemble of 280 simulations (@300km horizontal resolution) of both present day climate and the equilibrium response to doubled CO2

• Allowed us to sample uncertainties in processes contributing the largest uncertainties to large-scale-regional climate changes at reasonable expense.

..gives a large sample of possible changes (e.g. summer UK rainfall)

© U

KC

IP 2

006

Converting ensemble simulations into probabilistic projections of equilibrium climate change

• Used a general Bayesian framework designed for making future projections of real world systems using simulations from complex but imperfect models (Goldstein and Rougier, 2004; Rougier, 2007)

• Key ingredients included:

• An emulator, trained on the available ensemble runs and used to estimate values for historical climate variables and the equilibrium response to doubled CO2 at points in parameter space not sampled by a GCM simulation

• Discrepancy, an estimate of the additional uncertainties due to structural model errors which cannot be resolved by varying poorly-constrained model parameters

• A set of observations to use in estimating the relative likelihood that different model variants (i.e. different points in parameter space) give a true representation of the real climate system.

• Could then integrate over the model parameter space, weighting projections according to relative likelihood and accounting for effects of structural errors, to obtain probabilistic projections.

Estimating discrepancy Discrepancy represents model errors (arising from missing or structurally deficient

representations of processes) which cannot be resolved by varying uncertain parameters

Estimated by using an international ensemble of 12 alternative slab models (AR4, CFMIP) as set of proxies for the real system.

For each multimodel ensemble member, find a few points in the HadSM3 parameter space which give the closest historical and climate change simulations that we can find.

The outstanding mismatches are then estimates of the effects of missing or structurally deficient representations of processes in HadSM3.

Pool these distances over all 12 multimodel ensemble members to give an estimated distribution for discrepancy

© U

KC

IP 2

006

Global climate sensitivity

Mean impact of discrepancy

Discrepancy estimates do not account for errors common to all models

Simulations of time-dependent climate change using HadCM3 coupled atmosphere-ocean ensembles

• Smaller 17 member ensembles due to resource limitations

• Uses a subset of the multiple perturbation parameter sets used in the cheaper equilibrium simulations

• Can then build relationships between the equilibrium and transient responses…

• .. and hence produce large pseudo-ensembles of 21st century climate realisations by applying the scaling to estimates of equilibrium changes for which we have no corresponding transient simulation.

Observations

Historical + A1Bforcing

“Timescaling” approach to emulate large ensembles of

transient climate change projections

Equilibrium feedbacks (emulated)

Normalized equilibrium response pattern (emulated)for a doubling in CO2 conc.

Simple Climate Model projections for global surface temp. anomaly

Correction pattern representing differences between slab and dynamic ocean response

+

PDFs

Sampling uncertainties in other Earth system processes

• Further 17 member perturbed physics ensembles sampling uncertainties due to:

• Ocean transport processes, sulphur cycle processes and terrestrial ecosystem processes in HadCM3

© U

KC

IP 2

006

Dynamical downscaling to 25km scale

• Ran an 11-member ensemble of perturbed physics regional model variants at 25km resolution.

• Driven by boundary forcing from the HadCM3 A1B transient simulations (1950-2100).

• Used regression relationships between the changes simulated by the global and regional models to convert estimates of climate change at 300km global model grid boxes into estimates for 25km grid boxes, admin and river-based regions.

© U

KC

IP 2

006

Effects of downscaling on future projections

Winter precipitation changes for the 2080s relative to 1961-90, with (right) and without (left) the downscaling contribution

UKCP09 probabilistic projections

Three different emission scenarios

Seven different timeframes

25km grid, 16 admin regions, 23 river-basins and 9 marine regions

10% probability level Very likely to be greater than

90% probability level Very likely to be less than

50% probability “Central estimate”

UKCP09 provides probabilities which measure how strongly different outcomes for climate change are supported by current evidence (models, observations, understanding of known uncertainties)

Varies, but typically no single source dominates.

How important are different sources of uncertainty?

Uncertainties in winter precipitation changes for the 2080s relative to 1961-90, at a 25km box in SE England

Testing the robustness of the results

• Projections inevitably depend on expert assumptions and choices

• However, sensitivities to some key choices can be tested

Changes for Wales, 2080s relative to 1961-90

© U

KC

IP 2

006

Comparison of UCKP09 and ASK approaches

Coloured lines show 2.5th, 10th, 50th (thick), 90th and 97.5th percentiles of projected past and future changes

Temperature changes for Northern Europe

© U

KC

IP 2

006

Changes in 20 year mean temperature and precip, A1B forcing.

• UKCP09 methodology (minus downscaling) applied to European regions as part of ENSEMBLES

Probabilistic projections for Europe at 300km scale

Summary

• A number of methodologies for probabilistic projections have been developed

• The scope (types of uncertainty considered), inputs (model projections, observations), methodologies and outputs (global, regional, univariate, multivariate, emissions scenarios, etc) vary substantially

• All results are conditional on the input information and the assumptions made.

• The sensitivity to key assumptions should be clearly stated, and tested as far as possible.

• Different techniques should be compared.

• Some methods are more comprehensive than others, but they are all expressions of the spread of future projections conditioned on current models and understanding.

• So, results will change as the models and understanding improves

• Important to communicate this to users