Decadal Prediction and Predictability and Mechanisms of Climate Variability

1. Decadal Prediction and Predictability andMechanisms of Climate Variability Edwin K. Schneider COLA SAC Meeting April 2012

2. A Brief History of COLA’s Decadal Pillar • Decadal prediction component was added to COLA Omnibus research in 2009. • Prior to that COLA has been solely focused on shorter seasonal-to-interannual time scales. • Is there evidence for decadal predictability? Wide open field.

3. Outline of Research Areas • Analysis of observational and model data • Scientific basis of decadal predictability • Detection and attribution • Empirical prediction • Isolate modes of decadal variability • Re-examination of the assumption of a static annual cycle • Statistical tools: development and application • Relative entropy • Ensemble Empirical Mode Decomposition (EEMD) • Modeling • Decadal predictability • CFSv2 CMIP5+ runs • Multimodel evaluation • Mechanisms and predictability of decadal climate variability • Interactive Ensemble CGCM • Role of weather noise • Climate sensitivity • Isolate the causes of differences between models

4. Selected Results: Analysis of Observational and Model Data

5. Decadal SST Influences on Indian Monsoon Krishnamurthy and Krishnamurthy

6. Regression Analysis: Effect of ENSO and PDO on Great Plains Precipitation

7. In-phase of ENSO and PDO favors Great Plains drought/floodOut-of-phase of ENSO and PDO favors Great Plains neutral PDO NINO3.4 Hu and Huang

8. Consequences of Definition of Annual Cycle: Traditional vs. EEMD ENSO Phase Locking to Annual Cycle Wu, Schneider, Kirtman, Sarachik, Huang, and Tucker

9. QBO in a 2D EEMD Analysis QBO Hu and Huang

10. Selected Results:Decadal Modeling Research

11. CFS-based Decadal Prediction • Decadal prediction and predictability using the NCEP CFS (version 2) CGCM. • COLA Team: Cash, DelSole, Huang, Kinter, Klinger, Krishnamurthy, Lu, Marx, Schneider, Stan, Zhu • Collaborations: NCEP EMC and CPC; IRI; GSFC; IITM (Thanks to NCEP for providing the model, data sets and technical assistance and NASA for computing resources)

12. CFS-based Decadal Prediction • COLA is participating in the CMIP5 decadal prediction enterprise, including running the prescribed protocol and providing the output data to the CMIP5 archives for public consumption.  • The work is being done collaboratively with NCEP, including experimental design, sharing results and possible joint papers.  • The preliminary results were similar to and consistent with what other groups are finding, esp. the problem of insufficient sample size, which motivated a more exhaustive set of hindcasts.  • Our goal is to use the results to establish the scientific basis for decadal prediction.

13. Seamless Prediction: Feedback of Decadal Predictions on Shorter Time Scale Predictions By using the same model that is used for operational seasonal prediction, our results can have an impact on the way operational climate prediction is done, including identifying and quantifying erroneous and/or pathological behavior of the prediction model and dependency on the ocean initialization method. 

14. CFS-based Decadal Prediction • Complete and analyze the CMIP5 “core” hindcast/forecast cases as a baseline. • Produce additional runs to address problems with the experimental design. • Conduct additional experiments to address issues of model bias, improve the hindcasts.

15. COLA-NCEP Collaboration: CMIP5 Decadal Predictions • Identical model being used by both groups – enables direct comparison of results • COLA uses ECMWF (NEMOVAR) ocean initial conditions 1960-present in CFSv2. • NCEP uses CFSR ocean initial conditions, available 1980-present in CFSv2.

16. Technical Description • Model • CFS version 2 provided by NCEP EMC • Identical to model used by NCEP for operational S-I prediction and CMIP5 (to be documented in Saha et al. 2012) • Initial data • Atmosphere, land, sea ice: CFSR reanalysis (1980-present) • Ocean: NEMOVAR (ECMWF) interpolated to CFS (1960-present) • 4-member ensembles • 10 year predictions from Nov. 1960, 1965, 1970, 1975, 1980, 1985, 1990, 1995, 2000, 2005 • Extend to 30 years for 1960, 1980, 2005 cases • Computer resources • NASA Pleiades (Thanks to NASA)

17. Results: Predictability of Atlantic SST Variability • Evaluate predictability of three Atlantic SST indices in the CMIP5 decadal prediction ensemble means. • Atlantic Multidecadal Variability (TAV) • North Atlantic “Tripole” (associated with NAO) • Tropical meridional mode (TAV for “tropical Atlantic variability”). • Compare CFSv2 predictions with other CMIP5 models.

18. COLA CMIP5 Decadal Prediction Database

19. Atlantic Multidecadal Variability SST Index 1960-2010 ECMWF ICs Index region 80°W-0°E,0°N-59°N Color key to line plots Smoothed model forecast Smoothed persistence forecast Model bias Smoothed bias Observed

20. Atlantic Multidecadal Variability SST Index 1980-2010 ECMWF ICs NCEP ICs

21. Tropical Atlantic Meridional Mode SST Index 1960-2010 Index region 80°W-30°E,5°S-20°N minus 60°W-30°E, 20°S-5°S ECMWF ICs Color key to line plots Smoothed model forecast Smoothed persistence forecast Model bias Smoothed bias Observed

22. North Atlantic Tripole SST Index 1960-2010 ECMWF ICs Index region 60°W-40°W,40°N-55°N minus 80°W-60°W, 25°N-35°N Color key to line plots Smoothed model forecast Smoothed persistence forecast Model bias Smoothed bias Observed

23. Extended NINO3.4 Predictability • Sample of 10 decadal predictions is too small to make robust inferences about interannual or longer time scale predictability. So … • Fill out the cases to include at least 2 member ensembles out to 3 years lead time for all years 1960-2008 (no volcanoes).

24. Multiyear NINO3.4 Index 1960-2012 Index region 170°W-120°W,5°S-5°N Color key to line plots Model forecast Persistence forecast Model bias Smoothed bias Observed

25. Multiannual Predictability Using All Years 1960-2009 AMV TAV Tripole

26. Sea Ice/AMOC/Salinity Biases in CFS

27. How Serious a Problem is CFSv2 AMOC bias? Consider AMOC in CFSv1 Huang, Hu, Schneider, Wu, Xue, and Klinger 2012

28. CFSv2 AMOC in 30-year runs CFSR Ocean Initial Conditions NEMOVAR Ocean Initial Conditions Huang and Zhu

29. Are AMOC Biases Crucially Important for Decadal Prediction? • The scientific basis for decadal prediction of internal AMV variability has been supposed to be due to AMOC heat flux variability. • On the other hand, the AMV index seems to have decadal predictability for CFSv2 despite AMOC bias. Why? • AMOC heat flux potentially provides a substantial positive feedback to externally forced climate change (prediction of externally forced variability). • Several CMIP5 models have AMOC biases similar to CFSv2 (private communications).

30. Mechanisms of Climate Variability

31. Attribution and Predictability of Global Mean Temperature • Traditional view: acceleration of global warming in recent decades is externally forced. • Another view: recent acceleration is due to an internal mode. • Internal variability: is it the response to weather noise?

32. Global Mean Surface Temperature5-yr Running Mean CCSM3 20C3M Externally Forced Component (5 year running mean) .4 0 -.4 Observed (GISS analysis)

33. Influence of the Most Predictable Mode DelSole and Shukla

34. CCSM3 Response to Observed External Forcing 1870-2000 (20C3M) CCSM3 CMIP3 Colors = ensemble members Black= ensemble mean Red = CCSM3 ensemble mean Blue = CCSM3 IE Ensemble Members vs. Ensemble Mean Interactive Ensemble vs. Ensemble Mean

35. Attribution Results Using the Interactive Ensemble (Model World) • The interactive ensemble simulation captures the externally forced global mean temperature variability as given by the CCSM3 20C3M ensemble mean. • Since the interactive ensemble filters out the weather noise forcing of the ocean, land, and sea ice, the variability of the CCSM3 20C3M ensemble members relative to the ensemble mean can be attributed to weather noise forcing.

36. Understanding the Differences Between Coupled and Uncoupled Simulations • New results since the last SAC meeting: • Question: In a perfect model world, is the weather noise statistically the same in 1) a coupled model and 2) an atmospheric model forced by the SST from the coupled model? • Assuming that the SST forced response is the same in coupled and uncoupled simulations.

37. Diagnosis of the Weather Noise A long CGCM simulation provides “observations” AMIP ensemble is forced by CGCM SST For any field: Weather Noise = CGCM – (AMIP ensemble mean)

38. Hurrell et al. 2009 BAMS

39. Ratio of Surface Heat Flux Anomaly Standard Deviations CGCM:AGCM Total Noise <1 =1

40. Explanation Variance (total) = variance (signal) + variance (noise) + covariance (signal, noise) • Signal is identical in coupled and uncoupled by construction. • Noise variance is the same in coupled and uncoupled. • Covariance is different • Uncoupled: covariance(signal, noise) = 0 • Coupled: covariance(signal, noise) ≠ 0 because the noise forces the signal.

41. So What? • Who cares about surface heat flux variability? You can’t even measure it.

42. Ratio of Precipitation Anomaly Standard Deviations CGCM:AGCM Total Noise <1 =1

43. A Real World Consequences: Tier 2 Prediction • “Samescaling” CFSv2 seasonal forecasts using GFS • use the AGCM component (GFS) of the CGCM, forced by the SST predicted by the CGCM (CFSv2). Shukla and Zhu

44. Precipitation Standard DeviationSeasonal Forecasts CMAP AGCM CGCM

45. Summary • Observations: applications of new techniques • Information theory: new techniques • Relative Entropy • EEMD • Modeling • Decadal prediction: industrial strength effort • Mechanisms for internal variability: novel methodology, beautiful results

46. Questions for Omnibus Proposal • How can we build on the CMIP5 prediction experiments to better define the scientific basis for decadal predictability? • Initialization? • Model bias abatement? • Idealized experiments? • Auxiliary models (like interactive ensemble)? • Multimodel strategy? • Seamless strategy?