Transcript Glen Harris
Joint Frequency Distributions for Future European Climate Change Glen Harris, Ben Booth, Kate Brown, Mat Collins, James Murphy, David Sexton, Mark Webb Quantifying Uncertainty in Model Predictions (QUMP) Research Theme, Hadley Centre for Climate Prediction and Research, Met Office, Exeter, UK. Jonty Rougier, Durham University. Ensembles Work Package 6.2 Meeting, Helsinki, 26-27 April 2007 1 Gulf of Finland joint frequency distribution Joint frequency distributions for annual temperature and annual precipitation anomalies, with respect to 1961-90 baseline climate. A1B forcing, 2080-2100 mean anomaly. 129 time-scaled versions of HadSM3 equilibrium response (blue points). Sample distribution of scaling error, including internal variability (black points). Medians: T=5.1K, P=12% 2 HadCM3 European Land Grid-points Exclude 4 UK points (avoid potential conflicts with UKCIP08 project). Eastward to Moscow only. Rather coarse resolution (3.752.5 deg). 102 points in this set. Finnmark North_Cape Varangerfjord Westfjord Swedish_Lapland North_Bothnia Finnish_Lapland Russian_Lapland Murmansk Kola_Peninsula Central_Norrland West_Bothnia East_Bothnia North_West_Karelia North_East_Karelia White_Sea Sognefjord Trondheim South_Norrland Western_Finland Eastern_Finland North_Ladoga Onega South_West_Archangel Telemark Oslo Svealand Gulf_of_Finland Saint_Petersburg East_Ladoga West_Vologda Gotaland Latvia Pskov Western_Tver Moscow_North Denmark West_Lithuania East_Lithuania Vitebsk Smolensk Moscow_South Holland North_Germany Berlin North_Poland Warsaw Pripet South_East_Belarus Briansk Kursk Ireland Channel Belgium_NE_France Rhine South_East_Germany Czech_Republic Slovakia_South_Poland South_East_Poland Western_Ukraine Kiev Sumi Kharkov Western_France Burgundy Switzerland Austrian_Alps Eastern_Austria Hungary North_West_Romania North_East_Romania Moldova Lower_Dniepr Donetsk South_West_France South_East_France French_Italian_Alps Po_Dolomites Slovenia_Croatia Bosnia South_West_Romania South_East_Romania Pyrenees Tuscany Albania_Montenegro Central_Balkans Eastern_Bulgaria Galicia Northern_Spain Eastern_Spain Greece West_Marmara Bosphorus Ankara Black_Sea_Turkey Northern_Portugal Central_Spain South_West_Turkey Taurus_Mountains Turkish_Euphrates Southern_Portugal Andalucia 3 Where are the uncertainties? Natural unforced variability Modelling of Earth system processes Unknown future forcing QUMP: focus on modelling uncertainties 4 QUMP approach Predictions are uncertain so… 1. Run an ensemble of simulations with a climate model in which perturbations are made to the uncertain inputs and processes. 2. Compare each model simulation with observations and assign a relative score to each. 3. Produce a weighted distribution of the forecast variable of interest. i.e.: Posterior = Prior Likelihood QUMP project pragmatically uses a Bayesian framework. 5 Parameter Perturbations – 31 quantities perturbed Large Scale Cloud • Ice fall speed. • Critical relative humidity for formation. • Cloud droplet to rain: conversion rate and threshold. • Cloud fraction calculation. Dynamics • Diffusion: order and e-folding time. • Gravity wave drag: surface and trapped lee wave constants. • Gravity wave drag start level. Convection • Entrainment rate. • Intensity of mass flux . • Shape of cloud (anvils). • Cloud water seen by radiation. Boundary layer • Turbulent mixing coefficients: stabilitydependence, neutral mixing length. • Roughness length over sea: Charnock constant, free convective value. Radiation • Ice particle size/shape. • Cloud overlap assumptions. • Water vapour continuum absorption. Sea Ice • Albedo dependence on temperature. • Ocean-ice heat transfer. Land Surface Processes • Root depths. • Forest roughness lengths. • Surface-canopy coupling. • CO2 dependence of stomatal conductance. 6 Some issues for ensemble climate prediction Limited computational resources. use HadSM3/HadCM3 models, not expensive flagship HadGEM model mainly use mixed-layer (slab) ocean models. predict pdfs for equilibrium climate response. Large number of uncertain climate model parameters. to obtain robust predictions independent of sampling, emulators are required to predict response for parts of parameter space unsampled by GCM simulation. Sample prior distributions of uncertain model parameters. use expert ranges, prior distribution shape (triangular, uniform,…) test sensitivity to sampling assumptions. Likelihood weighting. want to choose as many observational constraints as possible to down-weight unrealistic model variants. Scale equilibrium response, to create “pseudo-transient” ensemble validate scaling with GCM ensemble Physics perturbations upset radiative balance, potential for climate drift. flux-correct transient GCM simulations. 7 “Perturbed-Physics” Atmosphere-Slab Equilibrium Ensemble Simulations Typical slab member Additional simulations underway to explore interesting regions of parameter space (currently ~300 members). Distribution differences due to different sampling strategies and parameter choices. Murphy et al, 2004. Stainforth et al, 2005. Webb et al, 2006. 8 Simple example for climate sensitivity “emulated” prior predictive distribution posterior predictive distribution likelihood weighting via comparison with real world histogram of “perturbed physics” ensemble Murphy et al., 2004, Nature, 430, 768-772 9 Probabilistic Predictions - Framework 1. Perform a limited ensemble of GCM experiments with perturbed input parameters. 2. Build an emulator which can estimate the GCM output at untried parameter values. 3. Sample emulator to produce model prior predictive distributions of climate variables. 4. Use observations to produce a likelihood function and posterior (observationally-constrained) predictive distributions. 5. Sample weighted posterior distribution and time-scale with Simple Climate Model (SCM) to predict pdfs for transient regional future climate change, at GCM resolution. 6. Run ensemble of 25km Regional Climate Model (HadRM3) variants driven by equivalent GCM transient runs, and downscale responses to predict regional pdfs. 10 Emulation for any perturbed-parameter value. Emulator: statistical model designed to predict the outputs of a climate model which one could in principle run. Emulators predict not only the mean response, but also the error in the predicted response. Built from a sample of runs. Multiple linear regression; entertain many possible functional relationships for explanatory variables. Emulator error used to select interesting parameter combinations to create additional members, and improve emulator. Emulator uncertainty is propagated through to the final PDFs. Joint prior equilibrium pdf for Eng-Wales summer temperature and precipitation response, for CO2 doubling. Rougier, Sexton et al, J.Clim (submitted) 11 Compare models with observations (likelihood weighting) Each “ensemble member” gets a weight w, something like: Sum over all observables simulated variable observed variable n ( M i Oi ) 2 w exp i 1 2(var( di ) var( ei ) var( Oi )) variance of “discrepancy” variance of emulator error variance of observations (including natural variability, obs. error etc.) More precisely, model skill is likelihood of model data given some observations: n 1 log Lo (m) c log | V | (m - o)T V 1 (m - o) 2 2 Sexton et al, J.Clim (in prep) 12 Discrepancy Following Murphy et al (Nature, 2004), began collaboration with statisticians (Rougier and Goldstein, Durham Univ.) to improve robustness of predictions. Introduce “discrepancy”: Measure of uncertainty associated with model imperfection: “distance” between unknown true future climate and “best” possible choice of the uncertain model input parameters. Unknown, but we assume this distance similar to that between other climate models and our best perturbed-physics emulation of the future predictions from these same models. Discrepancy therefore also a quantification of structural modelling error. 13 Compare model prior pdf with observationally-constrained pdf observationallyconstrained posterior pdf (no discrepancy) Equilibrium warming for EnglandWales for a doubling of CO2. Observational-constraints: narrow the spread in pdf, and can also move it (e.g., less than 2C warming unlikely). model prior pdf posterior pdf, with discrepancy Discrepancy: flattens likelihood, and broadens spread in observationallyconstrained posterior. Need discrepancy to avoid overconfidence, spiky posterior distributions. D.Sexton, J.Rougier 14 Transient Ensembles Need coupled model experiments to capture timedependent climate change. Historical + A1B forcing Observations Run 17 of the perturbed atmosphere HadSM3 versions coupled instead to dynamic ocean, i.e. HadCM3 setup. Transient ensembles smaller because of spin-up, additional ocean model, and longer runtime required. Flux adjustments used to prevent model drift, and reduce SST biases. HadCRUT observed series. 15 Compare perturbed physics ensemble with multi-model ensemble Increase CO2 by 1% per annum. Spread in transient response comparable in the two ensembles. Collins et al., Clim. Dyn. 16 Scaling the equilibrium response Problem: Can only afford relatively few simulations in transient GCM ensemble (17 here). Aim: Want to predict the transient response for the 129 slab-ocean experiments (or indeed any emulated equilibrium response), if they were coupled instead to a dynamic ocean (HadCM3). Solution: Scale anomaly patterns for each slab member by global mean surface temperature anomaly ΔT(t) predicted by a Simple Climate Model (SCM) Proposed in 1990 by Santer, Wigley, Schlesinger & Mitchell as way of predicting transient regional response from slab equilibria, before fully-coupled AOGCM’s had been developed. Fj pred ( x, t ) T scm ( j , t ) s slab j ( x), s slab slab F2slab CO 2 ( x) F1CO 2 ( x ) ( x) slab T2CO 2 ( x) T1slab CO 2 ( x) F in principle any climate surface variable, e.g. mean temperature, seasonal precipitation, soil moisture, percentiles of daily Tmax 17 Time-Scaling to Produce Pseudo-Transient Ensembles 129 SCM projections for global surface temperature anomaly, using diagnosed equilibrium feedbacks (1% p.a. CO2 inc). Typical response pattern for annual surface temperature to a doubling in CO2 concentration. Frequency distributions for Northern Europe annual temperature (including scaling error). 18 Scaling Assumptions 1. 20 year mean for equilibrium response sufficient to give good signal (compared to internal variability). 2. Slab equilibrium response patterns represent transient patterns. 3. Climate anomalies linear in global temperature anomaly ΔT(t). 4. ΔT(t) can be predicted by a Simple Climate Model (SCM), driven by emulated equilibrium climate feedbacks λ. 5. Assume equilibrium climate feedbacks represent transient feedbacks. Justification and Validation Compare pattern-scaling with the 17 fully-coupled simulations to give scaling error, and include this in predicted transient distributions. Any partial failure in assumptions quantified by validation: errors in scaling bigger uncertainty. 19 Scaling – validation with 17 member GCM ensemble SCM scaled prediction GCM anomaly SCM-GCM error Global (ghg only) . Mediterranean Basin (all forcing) 20 Frequency distribution for Transient Climate Response (TCR) Assume distribution of error in scaled response to be Gaussian (no evidence to contrary). Estimate variance and bias from validation with 17 member GCM ensemble. For each region and time, sum 129 t distributions (red curve) to obtain frequency distribution (blue curve). F Fjpred ( x, t ) bias(t ) D(F , t ) t ;16 (t ) j 1 129 Parameter uncertainty more important than scaling uncertainty. (TCR: surface temperature response for years 60-80 during 1% per annum CO2 increase). Distribution shape here mainly reflects sample design, not model prior distribution. 21 Time-scaling equilibrium patterns of change Example: djf precipitation, 1% CO2 pa increase Transient regional frequency distributions, using 129 perturbed atmosphere models. Plumes of evolving uncertainty (median, 80, 90, 95% confidence ranges) Harris et al., 2006, Clim.Dyn. 27, p357. 22 Pattern scaling A1B scenario • SCM uses forcing diagnosed from GCM runs. • compare here internal variability for one GCM run (green), with parameter and scaling uncertainty (red). Improvement of scaling to reduce error Using the A1B and A1B-GHG GCM ensembles, we can calculate - additional patterns for the normalised aerosol response saero - correction patterns to represent differences between the slab and dynamic ocean response cgcm F Tghg (t ) sslab ( x) Taero (t ) saero ( x) Ttot (t )cgcm ( x) error ( x, t ) 23 Production of interim data - summary 1. Scale 129 equilibrium responses, to predict transient joint temperature-precipitation response if we were to run with dynamic ocean and A1B forcing. 2. For each equilibrium member, sample (40 times for this test) the scaling error distribution (red curve), with variance and bias obtained from validation. Still a lot more to do… 24 Gulf of Finland future annual temperature/precipitation 80%, 90% and 95% confidence ranges. 17 GCM anomalies 2080-2100 anomalies with respect to 196190 baseline. 25 European pdfs – still to do Will do - Instead of annual data, process seasonal means and produce frequency distributions, based once again on 129 member ensemble. - Data now all back so can be done. Possible (time/resource constraints) - Build emulators for selected European GCM grid-points, and at same time obtain weights to observationally-constrain model variants. - Then resample weighted equilibrium distributions and time-scale to produce observationally constrained pdfs for future European climate change (HadCM3 resolution). Unlikely at moment - Redo UKCIP08 but for other parts of European domain, down-scaling to 25km resolution. 26 Down-scaling to the UK (and Europe?): UKCIP08 Also running a 17-member 25km resolution HadRM3 (regional model) ensemble . Driven by boundary forcing from the HadCM3 A1B ensemble (1950-2100). Runs will finish in July. We will construct regression relationships between the 17 GCM and 17 RCM simulations of future climate. Then sample predicted GCM transient pdfs and use these regression models to deliver regional response pdfs at 25km scales (this will introduce further uncertainty). R.Clark, D.Sexton, K.Brown, G.Harris, many others… 27 Additional perturbed physics ensembles (PPE) Atmosphere PPE. Also done two other forcing scenarios: A1B-GHG, and B1. Will also do A1FI. RCM ensemble Probabilistic climate predictions from perturbed physics ensembles 4 additional transient ensembles Ocean PPE E Atmosphere PPE (transient response) Regional climate model PPE C M Sulphate F aerosol PPE Timescaling transient changes D Downscaling transient changes N G Terrestrial ecosystem PPE Earth System PPE H Atmosphere PPE A (equilibrium response) B Emulation of equilibrium climate in parameter space Probabilistic predictions K Structural modelling errors Observational constraints J L Murphy et al (to appear in Phil. Trans. special issue, 2007) 28 Acknowledgments QUMP Team: David Sexton, Mat Collins, Ben Booth, James Murphy, Mark Webb, Kate Brown Also: Robin Clark, Penny Boorman, Gareth Jones, B. Bhaskaran, Jonty Rougier And: Hadley Centre, Met Office, DEFRA (Department for the Environment, Food and Rural Affairs) UK Govt, ENSEMBLES, ClimatePrediction.net. Thank You. 29