An overview of decadal climate predictability in a multi-model ensemble by climate model MIROC
- First Online:
- Cite this article as:
- Chikamoto, Y., Kimoto, M., Ishii, M. et al. Clim Dyn (2013) 40: 1201. doi:10.1007/s00382-012-1351-y
- 1.3k Downloads
Decadal climate predictability is examined in hindcast experiments by a multi-model ensemble using three versions of the coupled atmosphere-ocean model MIROC. In these hindcast experiments, initial conditions are obtained from an anomaly assimilation procedure using the observed oceanic temperature and salinity with prescribed natural and anthropogenic forcings on the basis of the historical data and future emission scenarios in the Intergovernmental Panel of Climate Change. Results of the multi-model ensemble in our hindcast experiments show that predictability of surface air temperature (SAT) anomalies on decadal timescales mostly originates from externally forced variability. Although the predictable component of internally generated variability has considerably smaller SAT variance than that of externally forced variability, ocean subsurface temperature variability has predictive skills over almost a decade, particularly in the North Pacific and the North Atlantic where dominant signals associated with Pacific decadal oscillation (PDO) and the Atlantic multidecadal oscillation (AMO) are observed. Initialization enhances the predictive skills of AMO and PDO indices and slightly improves those of global mean temperature anomalies. Improvement of these predictive skills in the multi-model ensemble is higher than that in a single-model ensemble.
KeywordsDecadal climate prediction Multi-model ensemble AMO PDO
Regional climate predictions for the coming decade have gained considerable attention since they can help to solve socioeconomic problems owing to climate change (Hibbard et al. 2007; Cox and Stephenson 2007). These predictions are one of the most challenging issues faced by the climate community (Murphy et al. 2010). To predict regional climate variability on decadal timescales, we need to consider the internally generated atmospheric and oceanic variabilities as well as externally forced variability related to global warming (Hawkins and Sutton 2009). Using a coupled atmosphere-ocean general circulation model (AOGCM) with initialization technique for assimilating observation in the ocean and prescribed natural and anthropogenic forcings (e.g., greenhouse gases and aerosol concentrations, solar cycle variations, and major volcanic eruptions), a recent series of studies has successfully predicted regional climate variability on decadal timescales (Smith et al. 2007; Keenlyside et al. 2008; Pohlmann et al. 2009; Mochizuki et al. 2010; Chikamoto et al. 2012). In order to provide useful future climate information for the coming decades, further improvement of predictive skills on decadal timescales is an important topic in the climate community.
Experiments of decadal climate prediction are restricted to computational resources because such studies need to perform ensemble model integrations over long-term periods utilizing several types of experiments designed by the Coupled Model Intercomparison Project-5 (CMIP5) (Taylor et al. 2009; Murphy et al. 2010). When we increase the resolution of AOGCM, model performance may improve because of a more realistic representation of small-scale phenomena (Sakamoto et al. 2012; Delworth et al. 2011), which may have a positive impact on the improvement of predictive skills. Nevertheless, increase in the resolution using limited computational resources leads us to reduce the number of ensemble members which may underestimate the ensemble spread during prediction. In addition, for decadal climate prediction, definition of a model climate drift during prediction, which may disrupt the prediction, requires a large number of hindcast samples coupled with a large number of ensemble members. Decadal climate predictability is also affected by uncertainty due to model deficiencies such as incorrect representation of physics, climatological field, internal variability, externally forced variability, and climate sensitivity.
A multi-model ensemble performed by many research centers is a useful approach to reduce model uncertainty under limited computational resources. For example, Jin et al. (2008) showed that the El Niño-Southern Oscillation (ENSO) predictive skill of a multi-model ensemble with ten different AOGCMs is higher than those of any individual model and hindcast experiments based on persistence and a dynamic statistical model. van Oldenborgh et al. (2012) investigated the decadal predictive skill in hindcast experiments using a multi-model ensemble of four European AOGCMs and indicated that sea surface temperature (SST) in the North Atlantic and Pacific Oceans regionally has the predictive skills beyond the persistence. Among these hindcast experiments using four AOGCMs, one group adopted a strategy of anomaly initialization (i.e., the observed anomalies are incorporated into the model anomalies) in order to reduce the model climate drift during the experiments, whereas the other three groups initialized the predictions by using full-field observations in order to provide initial conditions most similar to the real climate system (Doblas-Reyes et al. 2010). These studies suggest that the multi-model ensemble approach is an effective tool for decadal climate prediction.
In this study, we examine predictable components on decadal timescales by using the multi-model ensemble approach and discuss the number of ensemble members suitable for decadal climate prediction. In addition, we address the impact of anomaly initialization for decadal climate prediction by comparing the skills of initialized and uninitialized predictions. To reduce the model climate drift during prediction, we conducted decadal climate prediction experiments on the basis of anomaly initialization using AOGCMs with low, medium, and high resolutions (Tatebe et al. 2012; Mochizuki et al. 2012; Chikamoto et al. 2012).
The remainder of this paper is organized as follows. Section 2 describes the three versions of our model and experiments to assess the predictability of decadal climate variability. In Sect. 3, we evaluate the geographical distribution of the predictable area related to surface air and oceanic subsurface temperature anomalies. The predictability of the major climate variability on decadal timescales in terms of the global mean temperature, Atlantic multidecadal oscillation (AMO), and Pacific decadal oscillation (PDO) is examined in Sect. 4. We conclude and discuss the study in Sect. 5.
2 Model and experiments
In this study, we use the AOGCM MIROC that was cooperatively developed by the Atmosphere and Ocean Research Institute of The University of Tokyo, National Institute for Environmental Studies (NIES), and Japan Agency for Marine-Earth Science and Technology. MIROC’s atmospheric component is based on a global spectral dynamical core and a standard physics package (K-1 Model Developers 2004) developed at the Center for Climate System Research (CCSR)/NIES/Frontier Research Center for Global Change (FRCGC). The ocean general circulation model is the CCSR ocean component model (COCO; Hasumi 2006) which includes a sea ice model (K-1 Model Developers 2004; Komuro et al. 2012). No correction is applied in exchanging heat, water, and momentum fluxes between the atmosphere and the ocean. In addition, a land model that incorporates a river module is used (Takata et al. 2003; K-1 Model Developers 2004).
Summary of our experimental design for the three MIROC versions
20C3M & A1B
Every 5-year from 1960
20C3M & RCP4.5
Every 5-year from 1961
20C3M & RCP4.5
Every 5-year from 1961
Number of ensemble members in each experiment
In MIROC3m, the NoAS, ASSM, and HCST experiments are based on 10-member ensembles (Table 3). Nozawa et al. (2005) and Shiogama et al. (2007) performed 10-member ensemble simulations for the 1850–2000 period using the historical data of natural and anthropogenic forcings and the 2001–2030 period using the A1B-type emissions scenario of the Special Report on Emissions Scenarios (Nakicenovic et al. 2000). These ensemble simulations are started from different initial states obtained from the pre-industrial control run and correspond to the NoAS experiment in MIROC3m. Therefore, each member of the NoAS experiment has a different phase of internally generated variability such as ENSO, PDO, and AMO. From each member of the NoAS experiment, we obtained 10-member initial states on January 1, 1945, and conducted 10-member ensemble assimilations over the 1945–2009 period. During data assimilation, the observed temperature and salinity anomalies in the upper 700-m ocean depth (Ishii et al. 2006) are incorporated into MIROC. Although each ensemble member in the ASSM experiment tends to have a same phase of oceanic internally generated variability, our simple assimilation approach would have different trajectories among ensemble members, particularly in the atmospheric high-frequency states. On the basis of the 10-member assimilation experiment, we also obtained 10-member initial states and conducted 10 sets of 11-year-long, 10-member ensemble hindcast experiments initialized on January 1, 1960, 1965, 1970, 1975, 1980, 1985, 1990, 1995, 2000, and 2005. These experimental designs in MIROC3m are the same as those in Mochizuki et al. (2010).
In MIROC5, 3-member ensemble simulations in the NoAS experiment were conducted for the 1850–2005 period using the historical forcing dataset updated for CMIP5 and for the 2006–2100 period using the emissions scenario of the Representative Concentration Pathway (RCP4.5; Moss et al. 2008, 2010). Using two initial conditions on January 1, 1945, taken from 3-member simulations of the NoAS experiment, we conducted 2-member assimilation experiment for the 1945–2007 period. In the ASSM experiment, the observed temperature and salinity anomalies from the surface to 3,000-m depth (Ishii and Kimoto 2009) were assimilated into MIROC5 in a similar manner to MIROC3m. In the HCST experiment, we conducted 10 sets of 10-year, six-member ensemble hindcast experiments initialized in 1961, 1966, 1971, 1976, 1981, 1986, 1991, 1996, 2001, and 2006. Six ensemble members were obtained from the lagged average forecasting (LAF) method with the previous 3- and 6-month lags on the basis of the two-member ASSM experiment. For example, for the 1961 initialization, predictions commenced on July 1, 1960, October 1, 1960, and January 1, 1961.
The experimental design of MIROC4h is almost the same as that of MIROC5. The main differences between these experimental designs are the periods in the NoAS and ASSM experiments, the number of ensemble members, and the assimilation procedure. In MIROC4h, the NoAS experiment has three ensemble members for 1950–2007 but one member for 2008–2035. To reduce computational cost, the ASSM experiment was conducted with one member for 1950–2007. In terms of horizontal resolution, a small-scale eddy in the ocean is resolved in MIROC4h but not in the observation (Ishii and Kimoto 2009). When we assimilate the observation on a 1° × 1° horizontal grid into the eddy resolving model such as MIROC4h, analysis increment estimated by our system adopted in MIROC3m and MIROC5 acts to reduce the small-scale eddy resolved by MIROC4h. To retain the small-scale eddy, large-scale components of analysis increment estimated from a spatially smoothed field is incorporated into MIROC4h (Tatebe et al. 2012).
To distinguish climate signals from unrealistic signals arising from the model climate drift during prediction, drifts estimated in each model were excluded in the HCST experiment on the basis of INTERNATIONAL CLIVAR PROJECT OFFICE (2011) and Chikamoto et al. (2012). In each model, differences between the model predicted anomaly at a forecast lead time and the corresponding observational anomaly at a valid time are averaged over 9 sets initialized from 1960 or 1961 to 2000 or 2001 every 5 years. As shown in Chikamoto et al. (2012), the HCST experiment in MIROC5 displays a temperature drift during prediction, such as El Niño, for the initial 3-year lead time while a small temperature drift appears in the North Atlantic in MIROC3m. Although we used anomaly assimilation techniques to reduce the model climate drift, an inconsistency in the thermocline depth between the observation and the model may cause a model drift. Dynamical processes that induce the model climate drift during prediction will be reported elsewhere. The multi-model ensemble approach may reduce these systematic biases, as explained in Sect. 3.
After removing model drift during prediction in each model, multi-model ensemble in the HCST experiment is obtained from an average over three ensemble means in each model. Although initial dates starting from initial conditions depend on ensemble members and model versions as shown above, forecast lead time is defined as departure from initial dates on January 1, 1961, 1966, 1971, 1976, 1981, 1986, 1991, 1996, 2001, and 2006. In other words, the HCST experiment in MIROC3m are regarded as the LAF method at 1-year lag. The anomalies are defined as departures from the climatological mean value for 1971–2000 period in each model and are linearly interpolated into MIROC3m grid.
3 Overview of decadal climate predictability
The second SVD mode shows a mixture of PDO-like pattern in the Pacific and the AMO-like pattern in the North Atlantic (right panels in Fig. 2). In the Pacific region, observed SAT anomalies become positive in mid-latitude regions of the North and South Pacific and negative in the tropical Pacific and the high-latitude region of southern Pacific. This SAT pattern in observation is similar to the regression map on the PDO index (definition of PDO index is described in Sect. 4.3): the uncentered pattern correlation coefficient is −0.74 over the Pacific region (50°S–70°N, 100°E–70°W). In the Atlantic, positive SAT anomalies in observation are limited around Greenland, which appears at 5 years later after the mature phase of AMO. In fact, uncentered pattern correlation coefficients between the second SVD mode and the regression map of SAT on the AMO index are 0.47 simultaneously but increases up to 0.63 at the 5-year lag over the Atlantic region (50°S–70°N, 60°W–0°). These observed signals in the Pacific and Atlantic regions also appear in the second SVD mode of our multi-model ensemble (Fig. 2d). As pointed out by Solomon et al. (2010), distinguishing between internally generated and externally forced variations is a difficult problem due to some assumption intrinsic to an analysis technique. Our SVD analysis is one of the useful tools to roughly estimate predictability associated with natural and externally forced variations in observational and model represented fields under the small sampling such as our HCST experiment (10 samples in this paper).
The first and second SVD modes of SAT anomalies explain 97 and 1.4 % of the total covariance between the observation and the multi-model ensemble. In other words, the most predictable component of SAT anomalies in our current hindcast system is the global warming signal. When we focus on the ocean subsurface temperature variability instead of SAT, the covariance ratio explained by the second SVD mode considerably increases (to 20 %) as shown by Mochizuki et al. (2012). This high covariance ratio suggests that ocean subsurface temperature variability and the global warming signal are also important for the long-term predictability of SAT anomalies.
In summary, predictability on decadal timescales is prominent in the subsurface ocean rather than at the sea surface. Near the sea surface, externally forced components such as the global-warming signal are the main factors that explain decadal predictability, whereas the response patterns of SAT to external forcing are different between the observation and the model. On the other hand, initialization enhances the predictive skills of temperature variability, particularly in the subsurface ocean. Time series of temperature variability in the subsurface ocean show dominant natural variability such as PDO and AMO. In the next section, we focus on the predictability of the major climate variability: the global mean temperature, AMO, and PDO.
4 Internally generated and externally forced variabilities
As we have seen in the previous section, our HCST experiment, compared to the NoAS experiment, shows the improved predictive skill of temperature variability for the initial several years. When we focus on subsurface temperature variability, long-term predictability for a decade appears in the regions near the centers of action associated with PDO and AMO. In this section, we will show that the major decadal climate variability such as PDO and AMO is predictable for several years.
4.1 Global mean temperature
The predictive skill of global mean temperature variability is slightly improved in the HCST experiment compared to the NoAS experiment although internally generated variability has a small impact on the globally averaged temperature variation. As a result of the small improvement by initialization, predictive skills in the multi-model ensemble, measured by ACC and RMSE, are slightly higher in the HCST experiment than in the NoAS experiment (Fig. 7). In the recent decade, for example, the weakened warming trend in the observation is well simulated in the HCST experiment but not the NoAS experiment, particularly for the 2–4-year lead time (Fig. 7c, d). Mochizuki et al. (2010) suggested that a cool phase of PDO contributed to the weakened warming trend for the 2006–2008 period.
AMO is known as the multi-decadal SST variability in the North Atlantic (Delworth and Mann 2000; Murphy et al. 2010). Observational and modeling studies have indicated that AMO is related to rainfall variability in north east Brazil (Folland et al. 2001) and African Sahel regions (Folland et al. 1986; Rowell et al. 1995; Rowell 2003), summertime climate in North America and western Europe (Enfield et al. 2001; McCabe et al. 2004; Sutton and Hodson 2005), and Atlantic hurricanes (Goldenberg et al. 2001; Knight et al. 2006). For example, McCabe et al. (2004) suggested that recent droughts over the United States during 1999–2002 can be partly explained by a positive phase of AMO. This positive phase may have also affected the recent European summer heat wave during 2003 (Sutton and Hodson 2005). Although the mechanism of AMO is still controversial due to the insufficient historical observation in the deep ocean, study on AMO predictability is an important topic for climate prediction in terms of decadal timescales.
Although the predictive skill in MIROC3m is low for 2–4- and 5–9-year lead times, the multi-model ensemble of the three MIROC versions has a remarkable predictive skill. In particular, for 2–4-year lead time, the multi-model ensemble shows the best predictive skill as compared to each MIROC ensemble version (Fig. 12b, e). As shown in Sect. 2, our HCST experiments have a different number of ensemble members among three MIROC versions. Due to the small number of ensemble members, MIROC4h shows unstable predictive skills of the AMO index (circle in Fig. 12): the lowest for the 1-year lead time, highest for the 2–4-year lead time, and medium for the 5–9-year lead time. Our higher predictability in the multi-model ensemble suggests a possibility that useful predictive skills will also be obtained by the CMIP5 multi-model ensemble in many research centers.
In our multi-model ensemble, the HCST experiment shows higher predictive skill of the AMO index than the NoAS experiment for the initial 5-year lead time but lower at 7-year and 5–9-year lead times (Figs. 9e, f, 10). The lower predictive skills of the HCST experiment for longer lead time may be related to large forecast uncertainty. From 4-year to 5-year lead times, predictive skills of the AMO index suddenly decrease in both of the HCST and NoAS experiments (Fig. 10). Associated with this decrease, magnitude of ACC error bar also increases in the HCST experiment from 0.14 to 0.2 during this period. After 5-year lead time, the error bar in the HCST experiment becomes more than 2.0 while that in the NoAS experiment retains about 0.15. In addition to this large uncertainty in the HCST experiment, predictive skill of the AMO index suffers from a small sampling of hindcast because our analysis period covers only one phase of AMO. When we focus on the longer lead time such as 7-year or 5–9-year, the recent upward trend of the AMO index becomes prominent due to a limited sampling of the downward trend during 1960s (Fig. 9e, f). In our multi-model ensemble, however, the reason for the high predictive skill of AMO in the NoAS experiment is still unknown. To reveal predictability of AMO, a perfect model approach and a reproduction of climate for the past century are an important topic for future studies.
The PDO is the dominant climate variability in the North Pacific on decadal timescales (Mantua et al. 1997). During the positive phase of PDO, SST anomalies became cool in the central North Pacific and warm along the western coast of North America, and vice versa. Meehl et al. (2009) pointed out that the abrupt climate change observed in 1970s (known as the climate shift) contributed to decadal variability in PDO from negative to positive phases as well as externally forced variability. Mochizuki et al. (2010) indicated that the initialized ensemble hindcast in MIROC3m demonstrated a predictive skill of PDO for several years. In their follow-up paper (Mochizuki et al. 2012), reasonable predictability of PDO was also obtained by the initialized hindcasts in MIROC5 but not in MIROC4h due to the small number of ensemble members. On the basis of these hindcast experiments with different skills of PDO, we focus on the impact of the multi-model ensemble on the predictability of PDO in this subsection.
To obtain a time series for the PDO index in the observation and the HCST and NoAS experiments, we used a common definition of PDO. In a manner similar to the definition in Mantua et al. (1997), we defined the spatial pattern of PDO as the leading empirical orthogonal function (EOF) of the observed and linearly detrended VAT300 anomaly (Ishii and Kimoto 2009) over the North Pacific (north of 20°N) during the 1955–2010 period. For this spatial pattern of PDO, time coefficients were obtained by the linearly detrended projection of VAT300 anomalies in the observation and the HCST and NoAS experiments. These linearly detrended time series are defined as the PDO index in this paper. As shown in Fig. 2a, c, the spatial patterns of temperature trend are not identical for the observation and the HCST experiment: SAT anomalies in the North Pacific have almost no trend in the observation but a warming trend in the HCST experiment. This discrepancy implies model deficiencies related to the response to external forcing.
Previous studies have suggested that PDO can be regarded as the North Pacific expression of the basin-wide SST variability called the Interdecadal Pacific Oscillation (IPO) (Power et al. 1999; Folland et al. 2002; Meehl et al. 2009). In fact, subsurface temperature anomaly in the North Pacific is correlated to that in the South Pacific with a 4-year lag (Fig. 6a, b); ACC is 0.32 and exceeds the statistical significance at the 95 % level by using a two-side Student t-test with 50 degrees of freedom. However, the mechanism of PDO or IPO is still controversial, e.g., in terms of the origin of PDO or IPO, the tropics or extratropics (Liu 2012). In addition, the three MIROC versions have different characteristics of tropical SST variability; the amplitude of ENSO is high in MIROC5 but low in MIROC3m (Watanabe et al. 2010). To enhance the predictive skill on decadal timescales, physical processes contributing to interannual-to-decadal ENSO phenomena and tropical Pacific variability are an important topic for future studies.
5 Discussion and conclusions
We have investigated the predictability of decadal climate variability by a multi-model ensemble using three versions of MIROC. In our HCST experiments, initial conditions were obtained in a similar manner on the basis of anomaly assimilation in each model. We used the same external forcing to drive the initialized and uninitialized runs for each model. Therefore, uncertainties due to both initial errors and model imperfection are contained in the HCST ensemble, which can be used to assess the initialization impacts by comparing the initialized and uninitialized runs with sufficient ensemble size.
Our multi-model ensemble shows that the most predictable component in SAT variability originates in external condition (Figs. 1, 2, 7) and that internally generated variability is of secondary importance in SAT variability on decadal timescales. However, this does not imply that internal variability targeted by initialization is negligible in near-term climate prediction. For example, with respect to the global mean temperature variability, the impact of initialization has not been considerably clear in the case of the twentieth century but has been significant in the case of the recent decade (Fig. 7c, d); RMSEs of global mean temperature anomalies at 2–4-year lead time in the HCST and NoAS experiments are 0.05 and 0.06 by excluding hindcasts initialized on 2001 and 2006 (not shown), but 0.06 and 0.10 by including these initialized hindcasts, respectively. Furthermore, when we focus on the subsurface ocean, the observed interannual-to-decadal variability is large compared to the linear trend or the externally forced variability estimated by the NoAS experiment, which is consistent with long-term predictability related to AMO and PDO signals (Figs. 4, 5, 6). Our results suggest that the long-term predictable memory related to AMO and PDO resides in the subsurface ocean. Near the sea surface, ocean temperature variability is strongly affected by unpredictable atmospheric high-frequency noise and externally forced variability. These variations may reduce the predictive skills of SAT variability associated with AMO and PDO.
In our current prediction system, AMO and PDO indices have the predictive skills for several years (Figs. 10, 12, 14, 16) although the mechanisms of these phenomena are still unknown. The SVD analysis indicates that a predictable component of surface temperature anomalies shows a mixture of PDO-like and AMO-like patterns (Fig. 2). Some recent modeling and observational studies suggest that a low-frequency component of PDO on multi-decadal timescales is responsible for AMO with a lag of about a decade (Zhang and Delworth 2007; d’Orgeville and Peltier 2007). Possible physical processes involved in the relationship between PDO and AMO may be clarified by further analyzing the initialized HCST experiments, which will be a topic of future work.
To contribute to the next report of IPCC AR5, hindcast and forecast experiments will be conducted using many CMIP models with different ensemble sizes. Our multi-model ensemble shows higher predictive skills in the dominant ocean temperature variability such as AMO and PDO compared to any of the single-model ensemble. This suggests that the multi-model ensemble approach in CMIP5 also has the potential to increase the predictability by reducing errors in single models. Further studies that unravel physical processes of the predictable natural variability are desirable in order to enhance our understanding of these processes and increase the predictive skills of decadal climate variability.
The manuscript benefited from the constructive comments of two anonymous reviewers. This study was supported by the Japanese Ministry of Education, Culture, Sports, Science and Technology through the Innovative Program of Climate Change Projection for the twenty-first Century. The simulations were performed with the Earth Simulator at the Japan Agency for Marine-Earth Science and Technology and the NEC SX-8R at the National Institute for Environmental Studies.
This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.