Case studies for initialized decadal hindcasts and predictions for the Pacific region

Transcription

1 GEOPHYSICAL RESEARCH LETTERS, VOL. 39,, doi: /2012gl053423, 2012 Case studies for initialized decadal hindcasts and predictions for the Pacific region Gerald A. Meehl 1 and Haiyan Teng 1 Received 1 August 2012; revised 22 October 2012; accepted 25 October 2012; published 30 November [1] Case studies involving notable past decadal climate variability are analyzed for the mid-1970s climate shift, when the tropical Pacific warmed over a decade and globally averaged temperature rapidly increased, and the early 2000s hiatus when the tropical Pacific cooled over a decade and global temperatures warmed little. Ten year hindcasts following the CMIP5 decadal climate prediction experiment design are analyzed for those two periods using two different initialization techniques in a global coupled climate model, the CCSM4. There is additional skill in the initialized hindcasts for surface temperature patterns over the Pacific region for those two case studies over and above that in free-running historical simulations with the same model. A 30 year hindcast also shows added skill over the Pacific compared to the historical simulations. A 30 year prediction from the initialized model simulations shows less global warming for the period than the free-running model projection for that same time period. Citation: Meehl, G. A., and H. Teng (2012), Case studies for initialized decadal hindcasts and predictions for the Pacific region, Geophys. Res. Lett., 39,, doi: /2012gl Introduction [2] Decadal climate prediction focuses on near-term climate change (defined as 10 to 30 years in the future) using global coupled climate models initialized with observations [Meehl et al., 2009a, 2012; Hurrell et al., 2009; G. A. Meehl et al., Decadal climate prediction: An update from the trenches, submitted to Bulletin of the American Meteorological Society, 2012]. Attempts at decadal climate prediction with initialized climate models have showed some increases in skill in hindcasts (defined as an initialized prediction for a time period in the past) compared to uninitialized simulations from the same models [e.g., Smith et al., 2007; Keenlyside et al., 2008; Pohlmann et al., 2009; Mochizuki et al., 2010; Doblas-Reyes et al., 2011; Fyfe et al., 2011]. [3] In this paper we first analyze initialized model simulations for ten year hindcasts for two case studies. One, the so-called mid-1970s climate shift [Trenberth and Hurrell, 1994; Meehl et al., 2009b] involved relatively rapid warming over a decade or so in the tropical Pacific that was associated with both the response to increasing greenhouse gases (GHGs) and an internally generated transition of the Interdecadal Pacific Oscillation, (IPO) [Power et al., 1999; Meehl and Hu, 2006] from negative to positive [Meehl et al., 2009b; Meehl and Arblaster, 2011] (the Pacific Decadal 1 National Center for Atmospheric Research, Boulder, Colorado, USA. Corresponding author: G. A. Meehl, National Center for Atmospheric Research, PO Box 3000, Boulder, CO 80307, USA. (meehl@ucar.edu) American Geophysical Union. All Rights Reserved /12/2012GL Oscillation (PDO) is the North Pacific component of the basin-wide IPO, and both have similar patterns). The other case study is the so-called early 2000s hiatus when, for the first decade of the 21st century, globally averaged temperature had little warming trend and the tropical Pacific was characterized by a transition to the negative phase of the IPO [Chen et al., 2008; Burgman et al., 2008; Meehl et al., 2011; Meehl and Arblaster, 2012]. Such notable decadal variability involves interactions between the climate system response to external forcing, mainly increasing GHGs, and internallygenerated variability that could be predictable assuming there are distinct physical mechanisms producing that variability, such as for the IPO [Meehl and Hu, 2006]. This interplay between externally forced and internally generated decadal climate variability has been shown to be additive for rapid warming in periods like the mid-1970s shift with a positive phase of the IPO, with internally generated cooling of the tropical Pacific in the negative phase of the IPO overcoming the externally forced warming to contribute to periods like the early 2000s hiatus (G. A. Meehl et al., Externally forced and internally generated decadal climate variability associated with the Interdecadal Pacific Oscillation, submitted to Journal of Climate, 2012). [4] Here we analyze two sets of ten year hindcasts, each with ten member ensembles (following the CMIP5 experiment design [Taylor et al., 2012]) from two different initialization techniques for case studies of the mid-1970s shift and the early 2000s hiatus. Then we show results from a 30 year hindcast for initial states near 1980, and predictions for near term climate for from 2005 initial states. Following previous decadal climate prediction studies for the Pacific region [Meehl et al., 2010; Mochizuki et al., 2010], we focus here on that region in addition to globally averaged surface air temperature time evolution. 2. Model Description and Experiments [5] The global coupled climate model analyzed here is the Community Climate System Model version 4 (CCSM4) that includes a finite volume nominal 1 degree ( ) 26 level version of the atmospheric model CAM4, with components of ocean, land and sea ice [Gent et al., 2011]. The ocean is a version of the Parallel Ocean Program (POP) with a nominal latitude-longitude resolution of 1 (down to 1/4 in latitude in the equatorial tropics) and 60 levels in the vertical, with ocean grid points having a uniform 1.11 spacing in the zonal direction, and 0.27 near the equator, extending to 0.54 poleward of 35 N and S. No flux adjustments are used in CCSM4. [6] The two initialization techniques used in the CCSM4 decadal climate prediction experiments are 1) the hindcast method, and 2) the Data Assimilation Research Testbed 1of6

2 Figure 1. Time series of annual mean globally averaged surface air temperature ( C) from NCEP/NCAR reanalysis (black) and ensemble mean of CCSM4 free-running 20th century simulation during and RCP4.5 during (green), and the decadal prediction experiments (red for the hindcast and blue for DART initialization method). Shading indicates range of ensemble members of the free-running simulations. The small red/blue circles represent the year-1 predictions. The 30-year predictions from the 2006 initial states during the period of are represented by the dotted lines. Ensemble size of the CCSM4 simulations is shown in the legend. (DART) method. In the former, the ocean model is run for several iterations over the 20th century with observed atmospheric forcing which imprints the observations into at least the upper ocean [Yeager et al., 2012]. Though the model initial states are not exactly those of the observed climate system and returning repeatedly to the early 20th century tends to bias the earlier initial states warm, they include some element of the model systematic errors and therefore experience less drift when run for hindcasts or predictions. In the latter, DART is a weakly coupled data assimilation method that takes observations from ocean and atmosphere and assimilates them into the atmosphere and ocean model (G. A. Meehl et al., Decadal climate prediction: An update from the trenches, submitted to Bulletin of the American Meteorological Society, 2012; A. Karspeck et al., An ensemble adjustment Kalman filter for the CCSM4 ocean component, submitted to Journal of Climate, 2012). The advantage of this method is that the model is initialized relatively close to the observed climate system state, but then has more rapid drifts to its own systematic error state. The two sets of experiments are referred to as the hindcast and DART hereafter. Both methods were run following the CMIP5 protocol with initial states for the hindcasts approximately every five years. The start dates are January 1st of 1961, 1966, 1971, 1976, 1981, 1986, 1991, 1996, 2000, 2001, 2002, 2003, 2004, 2005, 2006 for the hindcast, and 1975, 1980, 1985, 1990, 1995, 2000, 2001, 2002, 2003, 2004, 2005, and 2006 for DART. For each start year we performed 10-member ensemble simulations and the spread in the 10-member initial states represents the effects of weather noise. For the hindast method, ensemble members were generated from different atmospheric initial conditions from daily restart files from the un-initialized simulations, and for DART the ensemble members were started from different ocean assimilated states with the same atmospheric conditions. The hindcast runs starting from 1961, 1981 and 2006, and DART runs starting from 1980 and 2006 were extended to 30 years. [7] Both sets of hindcast experiments are bias-adjusted following CLIVAR [2011] and as described in Yeager et al. [2012]. Essentially, climatological monthly mean differences for the ten-year hindcasts are computed and composited by month following the initial dates of all the hindcast simulations. Then these average time-evolving monthly mean differences are subtracted from each member of each ten-year hindcast to remove the model systematic error, leaving the model signal from external forcing and internally generated decadal climate variability. Most of the model drift occurs in the first ten years [Goddard et al., 2012], so that for the 30 year hindcasts and predictions, the year ten monthly mean bias adjustments are subtracted from each succeeding year from year 11 to year 30. As discussed in Goddard et al. [2012], there are a number of complications and ambiguities that arise from bias adjustment methods, some involving sampling issues and others related to the nature of model systematic errors that may not be the same over the entire duration of the second half of the 20th century (the time period for the hindcasts). Therefore, a significant caveat that accompanies these results is that the bias adjustment calculations may introduce further errors when comparing model results to observations. Additionally, there are indications that bias adjustment can also increase the skill of the non-initialized simulations, which suggests that removing model systematic errors from any predictions could improve the skill. This is the topic of further research currently in progress. [8] This study focuses on surface air temperature. Bias adjustment was carried out using NCEP/NCAR reanalysis [Kalnay et al., 1996] as a reference as it has better spatial coverage in the polar regions and Southern Hemisphere. Use of other observational data sets produced qualitatively similar results. For example, the main features are consistent with the UK Met Office Hadley Centre and the University of East Anglia Climatic Research Unit (HadCRUT3) [Brohan et al., 2006] combined land surface temperature and sea surface temperature (SST) analysis. 3. Results [9] Figure 1 shows the time series of the free-running 20th century historical simulations with CCSM4 compared to observations and the ensemble average bias-adjusted ten year hindcasts from the two initialization methods. In general for the latter part of the 20th century, the free-running CCSM4 simulations are somewhat warmer than the observations [Meehl et al., 2012]. The warmer initial states for the hindcast method in the 1960s were noted above as a consequence of that method. Since volcanoes are included in the hindcasts, they all reflect the two major volcanic eruptions (El Chichon in the early 1980s and Pinatubo in the early 1990s) as a drop of temperatures shortly after those eruptions. The mid- 1970s shift is seen as a jump in the observed globally averaged temperatures from the mid-1970s to the early 1980s, while the 2000s hiatus is evidenced by a nearly zero trend of global temperatures over the early 2000s. Therefore, for the mid- 2of6

3 1970s shift we choose available initial states near the start of the shift (January 1976 for the hindcast method, January 1975 for the DART method, so that the first year of the prediction is 1976 and 1975, respectively). [10] The observations for the prediction period (5 year average of years 3 7 of the prediction to compare to the DART method) minus the reference period, as well as the prediction period (to compare to the hindcast method) minus the reference period, show a warming of the tropical Pacific and Indian Oceans with cooling in the northwest and southwest Pacific, and a globally averaged temperature increase of 0.25 C and 0.24 C, respectively (Figures 2a and 2b; 0.17 C and 0.16 C for HadCRUT3). The free-running model results for those same periods (Figures 2e and 2f) show a general warming over most of the Pacific, with a pattern correlation with the observations of and +0.17, respectively, for the domain shown in Figure 2 (all pattern correlations in this paper are centered, meaning the area-mean is removed; a Monte Carlo test was performed that involved taking the 1000 year CCSM4 control run and calculating the pattern correlations of 100,000 random patterns, and the 95th percentile is a pattern correlation of 0.59; skillful predictions have higher values of pattern correlation and lower values of root mean square error (RMSE)). RMSE values are 0.27 and 0.31, respectively. Meanwhile, both the initialized model simulations in Figures 2c and 2d show a closer resemblance to the observed pattern, with pattern correlations of and +0.79, and RMSE values of 0.28 and Thus, there is an improvement in pattern correlation with DART initialization but nearly the same RMSE. As can be seen in Figure 1, the DART simulation does not have as large a globally averaged temperature signal (+0.06 C compared to the observations of C). The hindcast method is closer to the observed value (+0.20 C compared to the observations of C) with a larger amplitude pattern correlation and lower RMSE. The uniform warming throughout the period in the free-running simulations is C, somewhat larger than the observations. Thus, the initialization produces mostly improved regional patterns of sea surface temperature change compared to the free-running simulations, with the warm phase of the IPO in the tropical Pacific as in the observations and cooling in the northwest and southwest Pacific. [11] For the early 2000s hiatus in Figures 2g 2j, we show results from the initial state of January 2004, with the first year of the prediction being 2004 (results from initial states in January 2003 and January 2005 have similar results and are not shown). As with the mid-1970s shift that showed a positive phase of the IPO in the tropical Pacific, the (years 3 7 average of the prediction) minus difference in the observations shows a negative phase of the IPO with negative SST anomalies in the eastern equatorial Pacific and positive anomalies in the northwest and southwest Pacific (Figure 2g). The more uniform warming of the free-running simulations shows positive SST anomalies in the tropical eastern Pacific, and a pattern correlation with the observations of only 0.07 and RMSE of 0.28 (Figure 2i). The globally averaged temperature increase is C (0.15 C for HadCRUT3) for the observations compared to the greater warming in the free-running simulation of C. Meanwhile, both initialization methods capture elements of the observed negative phase of the IPO, with pattern correlations with the observations of for DART (Figure 2h) and for the hindcast method (Figure 2j), with reduced RMSE values (compared to the free-running simulation) of 0.15 and 0.18, respectively, indicating less model error in the initialized simulations. Each of the initialized simulations also shows less global warming (closer to the observations) than the free-running simulation, with values of C and C respectively. It is notable that, as described by Meehl et al. [2009b, Externally forced and internally generated decadal climate variability associated with the Interdecadal Pacific Oscillation, submitted to Journal of Climate, 2012], the early-2000s hiatus initialized simulations must overcome the general tendency for warming from the external forcing by producing negative SST anomalies in the eastern Pacific from the negative phase of the internally generated IPO. [12] A 30 year hindcast made with both initialization techniques with an initial state of 1980 for DART and 1981 for the hindcast method, such that the first years of the predictions are 1980 and 1981, respectively, is shown in the top part of Figure 3. This hindcast, verifying for the 20 year period , is intended to be a reference for the 30 year prediction for the 20 year average of , the near-term period used in the IPCC AR5 (Figures 3e 3g). The observations (Figure 3a; minus ) show a warming of the tropical Pacific, with cooling in the northwest and southwest Pacific and warming of the Indian Ocean, and a globally averaged temperature increase of C (0.41 C for HadCRUT3). The free-running simulation in Figure 3c shows almost uniform warming and a low pattern correlation with observations in Figure 3a of 0.04, a RMSE of 0.27, and a larger value for global temperature increase of C. However, both initialized simulations (Figures 3b and 3d) show elements of the observed pattern, with greater relative warming in the equatorial Pacific compared to cooling in the north or northwest and southwest Pacific regions. Pattern correlations with observations are higher than the free-running simulations, with values of for each, with reduced error as indicated by lower RMSE values of 0.22 and Values of globally averaged temperature increase are C for DART and C for the hindcast method, somewhat lower than the observed value and considerably lower than the free-running simulation. Thus, even for a 30 year hindcast, there does seem to be an advantage in the initialization methods over the free-running simulations, with greater pattern correlations and reduced RMSE, even after most of the effects of the initialization have presumably become less of a factor [e.g., Branstator and Teng, 2012]. As noted above, the improvement may partly be related to the effects of the bias adjustments and is under further investigation. [13] For the 30 year prediction from the 2006 initial state, the ensemble average globally averaged time series of surface air temperature in Figure 1 shows both initialization methods produce lower temperature increases than the freerunning simulation, and in fact maintain the slow rate of warming seen in the early 2000s until about The freerunning simulation has a warming for the period minus the reference period of of C, while the value from the DART simulation is C, and C from the hindcast method. With regards to the pattern of SST change, there is almost uniform warming over the entire Pacific with greater warming in the tropics in 3of6

4 Figure 2. Spatial distribution of five-year mean surface air temperature changes ( C) (a f) during the mid-1970s shift ( minus for DART initialized in 1975 and minus for the hindcast initialized in 1976) and (g j) during the early 2000s hiatus period ( minus for both methods initialized in 2004). Area-weighted globally averaged surface air temperature anomalies are shown at the top right of each panel. 4of6

5 Figure 3. Spatial distribution of twenty-year mean surface air temperature change ( C) (a d) during the mid-1970s shift ( minus for DART initialized in 1980 and hindcast initialized in 1981) and (e g) during the early 2000s hiatus period ( minus for both methods initialized in 2006). Area-weighted globally averaged surface air temperature anomalies are shown at the top right of each panel. the free-running simulation (Figure 3g). Elements of a negative IPO pattern linger in both the DART (Figure 3e) and hindcast method (Figure 3f), with some negative SST anomalies in the off-equatorial eastern Pacific and relatively greater warming in the northwest and southwest Pacific, as well as in the tropical western Pacific and eastern Indian Oceans. Thus the initialized predictions are showing less global warming than the free-running simulation with the same model, with a different pattern of predicted SST anomalies in the Pacific and Indian Oceans involving elements of a weak negative IPO pattern. [14] Acknowledgments. The authors thank Robert Burgman and one anonymous reviewer for constructive comments on the paper. Portions of this study were supported by the Office of Science (BER), U.S. Department of Energy, Cooperative Agreement DE-FC02-97ER62402, and the National Science Foundation. The National Center for Atmospheric Research is sponsored by the National Science Foundation. [15] The Editor thanks the two anonymous reviewers for their assistance in evaluating this paper. References Branstator, G., and H. Teng (2012), Potential impact of initialization on decadal predictions as assessed from CMIP5 models, Geophys. Res. Lett., 39, L12703, doi: /2012gl of6

6 Brohan, P., J. J. Kennedy, I. Harris, S. F. B. Tett, and P. D. Jones (2006), Uncertainty estimates in regional and global observed temperature changes: A new data set from 1850, J. Geophys. Res., 111, D12106, doi: /2005jd Burgman, R. J., A. C. Clement, C. M. Mitas, J. Chen, and K. Esslinger (2008), Evidence for atmospheric variability over the Pacific on decadal timescales, Geophys. Res. Lett., 35, L01704, doi: /2007gl Chen, J., A. D. Del Genio, B. E. Carlson, and M. G. Bosilovich (2008), The spatiotemporal structure of Twentieth-Century climate variations in observations and reanalyses. Part II: Pacific pan-decadal variability. J. Clim., 21, , doi: /2007jcli CLIVAR (2011), Data and bias correction for decadal climate prediction, CLIVAR Publ. Ser. 150, 4 pp., Int. CLIVAR Proj. Off., Southampton, U. K. Doblas-Reyes, F. J., M. A. Balmaseda, A. Weisheimer, and T. N. Palmer (2011), Decadal climate prediction with the European Centre for Medium-Range Weather Forecasts coupled forecast system: Impact of ocean observations, J. Geophys. Res., 116, D19111, doi: / 2010JD Fyfe, J. C., W. J. Merryfield, V. Kharin, G. J. Boer, W.-S. Lee, and K. von Salzen (2011), Skillful predictions of decadal trends in global mean surface temperature, Geophys. Res. Lett., 38, L22801, doi: /2011gl Gent, P., et al. (2011), The Community Climate System Model version 4, J. Clim., 24, , doi: /2011jcli Goddard, L., et al. (2012), A verification framework for interannual-todecadal predictions experiments, Clim. Dyn., doi: /s , in press. Hurrell, J., et al. (2009), A unified modeling approach to climate system prediction, Bull. Am. Meteorol. Soc., 90, , doi: / 2009BAMS Kalnay, E., et al. (1996), The NCEP/NCAR 40-year reanalysis project, Bull. Am. Meteorol. Soc., 77, , doi: / (1996) 077<0437:TNYRP>2.0.CO;2. Keenlyside, N., et al. (2008), Advancing decadal-scale climate prediction in the North Atlantic Sector, Nature, 453,84 88, doi: /nature Meehl, G. A., and J. M. Arblaster (2011), Decadal variability of Asian-Australian monsoon-enso-tbo relationship, J. Clim., 24, , doi: /2011jcli Meehl, G. A., and J. M. Arblaster (2012), Relating the strength of the tropospheric biennial oscillation (TBO) to the phase of the Interdecadal Pacific Oscillation (IPO), Geophys. Res. Lett., 39, L20716, doi: /2012gl Meehl, G. A., and A. Hu (2006), Megadroughts in the Indian monsoon region and southwest North America and a mechanism for associated multi-decadal Pacific sea surface temperature anomalies, J. Clim., 19, , doi: /jcli Meehl, G. A., et al. (2009a), Decadal prediction: Can it be skillful?, Bull. Am. Meteorol. Soc., 90, , doi: /2009bams Meehl, G. A., A. Hu, and B. D. Santer (2009b), The mid-1970s climate shift in the Pacific and the relative roles of forced versus inherent decadal variability, J. Clim., 22, , doi: /2008jcli Meehl, G. A., A. Hu, and C. Tebaldi (2010), Decadal prediction in the Pacific region, J. Clim., 23, , doi: /2010jcli Meehl, G. A., et al. (2011), Model-based evidence of deep-ocean heat uptake during surface-temperature hiatus periods, Nat. Clim. Change, 1, , doi: /nclimate1229. Meehl, G. A., et al. (2012), Climate system response to external forcings and climate change projections in CCSM4, J. Clim., 25, , doi: /jcli-d Mochizuki, T., et al. (2010), Pacific decadal oscillation hindcasts relevant to near-term climate prediction, Proc. Natl. Acad. Sci. U. S. A., 107, , doi: /pnas Pohlmann, H., et al. (2009), Initializing decadal climate predictions with the GECCO oceanic synthesis: Effects on the North Atlantic, J. Clim., 22, , doi: /2009jcli Power, S., et al. (1999), Interdecadal modulation of the impact of ENSO on Australia, Clim. Dyn., 15, , doi: /s Smith, D., et al. (2007), Improved surface temperature prediction for the coming decade from a global circulation model, Science, 317, , doi: /science Taylor, K. E., R. J. Stouffer, and G. A. Meehl (2012), An overview of CMIP5 and the experiment design, Bull. Am. Meteorol. Soc., 90, Trenberth, K. E., and J. Hurrell (1994), Decadal atmosphere-ocean variations in the Pacific, Clim. Dyn., 9, , doi: /bf Yeager, S., et al. (2012), A decadal prediction case study: Late twentiethcentury North Atlantic ocean heat content, J. Clim., 25, , doi: /jcli-d of6