Sensitivity experiments using a coupled model initialized from analysed atmospheric and oceanic observations are used to investigate the potential for interannual-to-decadal predictability. The potential for extending seasonal predictions to longer time scales is explored using the same coupled model configuration and initialization procedure as used for seasonal prediction. It is found that, despite model drift, climatic signals on interannual-to-decadal time scales appear to be detectable. Two climatic states have been chosen: one starting in 1965, i.e. ahead of a period of global cooling, and the other in 1994, ahead of a period of global warming. The impact of initial conditions and of the different levels of greenhouse gases are isolated in order to gain insights into the source of predictability.
Ensemble forecasting is a practical tool for estimating flow-dependent predictability due to uncertainties in both initial data and model equations (Palmer & Hagerdon 2006). Ensemble forecast techniques have been used in operational meteorology for approximately 20 years. Following research by Miyakoda et al. (1986), the first operational ensemble forecast was made for the monthly time scale (Murphy & Palmer 1986). The technique was then introduced into operational medium-range weather forecasting in the early 1990s (Palmer et al. 1993; Toth & Kalnay 1993; Molteni et al. 1996), and into operational seasonal forecasting in the late 1990s (e.g. Stockdale et al. 1998). In the early 2000s, ensemble forecasting became a standard technique in both short-range weather forecasting (Tibaldi et al. 2006) and centennial climate-change prediction time scales (Murphy et al. 2004; Stainforth et al. 2005).
Following studies by Griffies & Bryan (1997), research is now aimed at assessing the viability of introducing ensemble forecast techniques into decadal time-scale prediction (Boer 2000; Collins 2002; Collins & Allen 2002; Smith et al. submitted). From a practical point of view, this time scale is of particular relevance to decisions on investment needed to mitigate and adapt to climate change. Here, we give a first attempt to make decadal ensemble re-forecasts from directly analysed atmosphere ocean initial conditions (ICs), i.e. using the same techniques as used in operational seasonal prediction.
On decadal forecast time scales, potential predictability arises from both radiative forcing (from greenhouse gases (GHGs) and other anthropogenic and natural agents) and atmosphere/ocean ICs. Thus far, however, work has mostly concentrated on decadal predictions either in the perfect model context (e.g. Boer 2000; Collins & Allen 2002) or with the coupled model started close to its own equilibrium, i.e. near the model attractor (Smith et al. submitted). One important reason for this is that most coupled models are affected by sizeable drifts. Thus, by adopting the above-mentioned approaches the problem of the model drift is reduced. In this paper, we investigate the prospects for the predictability of a coupled system which is integrated forward from directly analysed ICs, which are therefore as close as possible to the real attractor (as opposed to the model attractor).
More specifically, we discuss here decadal ensemble integrations of a global coupled ocean–atmosphere model, initialized on two dates: one in 1965 (preceding a decade of global cooling) and the other in 1994 (preceding a decade of global warming). Differences are found between the two forecast ensembles, indicating potential for predictability. Based on supplementary integrations, we address whether these differences arise primarily owing to the different GHG forcing (out of all radiative forcings, here we limit our attention to GHGs as they are deemed to be the principal cause of anthropogenic climate change) or the different ICs.
In §2, the coupled model used in this study is described. An estimate of the drift of this coupled model is analysed in §3. After removing the drift a posteriori, we investigate the interannual signals in §4 and the impact of GHG forcings is evaluated in §5. Section 6 presents a summary of this work.
2. Description of the system
The coupled general circulation model used in this study is composed mainly of an atmospheric model, a land model and an ocean model. The atmospheric/land model is the ECMWF integrated forecasting system (IFS),1 in its cycle 28r1 version. It is run at TL95 horizontal resolution (approx. 200 km) with 40 vertical levels. The ICs for the atmosphere are provided by the ECMWF ERA-40 reanalysis (Uppala et al. 2005). The ocean model is the OPA model (Madec et al. 1998) in its global configuration, ORCA2, with a 2° resolution and a meridional refinement near the equator. There are 31 levels in the vertical, with the highest resolution (10 levels) in the upper 150 m. A restoring term to climatology (Levitus NODC) is added within the potential temperature and salinity equations, poleward of 60° of latitude.
Three different types of ICs are used for the ocean model. The first type is generated by forcing the ocean model with heat, momentum and fresh water fluxes from ERA-40 reanalysis but ‘correcting’ the excessive ERA-40 precipitation following Troccoli & Kållberg (2004). The other two types are produced by adopting, in addition to what done for the first type, two variants of the same three-dimensional variational assimilation developed by Weaver et al. (2003).
The oceanic and atmospheric models exchange fluxes every day through the Oasis software (Terray et al. 1995). The atmosphere is driven by SST and sea ice, and the ocean is driven by heat, momentum and moisture fluxes. Solar radiation, needed also to calculate the effect of penetrative radiation within the upper layers of the ocean, is passed separately from the other heat fluxes. No flux correction is applied to the exchanged fluxes.
The configuration of this coupled model is the same as that used for seasonal forecast experiments carried out under the EU project ENhAnced ocean data assimilation and ClimaTe prediction2 (ENACT), and is similar to that in the EU project DEMETER (Palmer et al. 2004). This is the first attempt to extend the coupled model integrations with IFS beyond the 6–12 months integration typical of seasonal prediction. The coupled model is run in fact for 20 years starting from two ICs: 1 May 1965 and 1 May 1994. These two dates have been especially selected (i) because they are ahead of a decade of global cooling of −0.05°C and of global warming of 0.28°C, respectively, and (ii) to avoid periods during which major volcanic eruptions occurred (e.g. in 1963 and 1991). Six members for each start date have been run. Each member differs from each other by their oceanic ICs and/or forcing conditions. Features of all the runs are shown in table 1, and the values and the rate of change of the GHGs are given in table 2.
3. Drift in the coupled model
In the absence of corrective measures such as ‘flux correction’, climate models are likely to be affected by drifts (also known as bias). Since only some relaxation of the ocean model at high latitudes is applied in our case, the coupled integrations discussed here are no exception. Figure 1 shows an estimate of the evolution of the coupled model drift for global maps of 2 m temperature. This estimate of drift is computed as the difference between the average of the model integrations and the ERA-40 reanalysis.
Three main features can be seen in figure 1. First and most obvious is the size of the drift especially over land. During the first year, with the exception of the high latitudes, the size of the drift is within ±2°C (figure 1a). Then, in the next two periods (2–4- and 5–8-year averages; figure 1b,c) the drift becomes more accentuated. Apart from the Kuroshio extension and the Circumpolar current, the drift over the ocean is smaller and within ±2–3°C. But over land, such as over the Southeast Asia, West Australia and the mountainous regions of the Americas, the distance between model integration and observations is distinct.
Second, the drift pattern is established relatively early in the integration. Most of the major drift apparent in the latter two periods is in fact already present in the first year of integration. Indeed, the large warm drift over the Eurasian continent and North America is discernable well within the first year. Equally, the cold drift over the tropical Pacific Ocean and over the US West Coast, India and Australia is also noticeable in the first year. The spatial correlation between these maps is indeed high: 0.81 for year 1 versus years 2–4 (figure 1a,b), 0.93 for years 2–4 versus years 5–8 (figure 1b,c) and 0.80 for year 1 versus years 5–8 (figure 1a,c). Further integrations are needed to confirm this result, but if generalized this would imply that relatively short integrations could be used in order to study the long-term systematic errors, such as via the tendency budget analysis of Klinker & Sardeshmukh (1992) and Rodwell & Palmer (2007).
Third, from the sequence of the three periods, the magnitude (and pattern) of the drift in global 2 m temperature appears to reach an equilibrium (see also the inset of figure 4) in the 5–8-year period. Other fields such as mean sea-level pressure and the zonal wind component at 10 m follow a similar behaviour (not shown).
Finally, it is worth pointing out that the drift pattern present in these integrations is also seen in similar integrations in which a different ocean model (HOPE instead of OPA) and a different version of the atmospheric model were used. Other coupled models may have completely different biases, and in fact the Met Office GloSea–HADAM3 coupled model does have a very different drift pattern (e.g. over the Eurasian continent the bias is mostly negative and over India it is positive). However, the magnitude of the drift is comparable to our coupled model.
Despite the large size of the drift we will show in §4 that, by using a simple post-processing methodology, it is still possible to analyse the interannual and decadal signal of these integrations, even on time scales relevant to climate change.
4. Interannual and decadal climate signal
The main purpose of this work is to investigate whether the interannual and decadal variability indicates the potential for useful predictions.
The approach used here to remove the model drift a posteriori is based on the same post-processing technique adopted by long-range weather forecasting. This involves computing a reference time-dependent model climate with which the individual members are compared, i.e. the same model climate is subtracted from each ensemble member as a function of the lead time. The model climate is normally computed by taking several realizations of integrations starting from ICs with similar characteristics (e.g. same month for several years). In our case, given the limitations of running very long experiments, we will use eight members for the model climate, i.e. members 1–4 for 1965 and the corresponding ones for 1994. Clearly, more realizations of the model integration would define better the climatology. However, it is not unusual to consider a limited set of start dates in studies whose aim is to carry out initial investigations.3 Besides, we looked at the errors in the individual start dates and they are very similar to each other (not shown).
In what follows, the climate is computed by taking the eight integrations (four for each start date) which have realistic GHGs. We will consider only the differences (anomalies) of each member from the climate.4 Analogous anomalies are computed for the ‘observations’ which, in the case of temperature, are again taken from the ERA-40 reanalyses. When comparing the integrations to observations, we restrict the integrations to their first 10 years in order to have observations for the period considered.
We look first at the interannual variability of the global 2 m temperature. Figure 2 compares the temporal variability of the mean model anomaly (black thick line) along with its spread, as measured by the extremes of the four curves (grey shading), with the anomaly of the observations (black thin line). As noted earlier, the two chosen periods are characterized by global temperature trends of −0.05°C per decade and 0.29°C per decade, respectively, for 1965 and 1994. The trend in the mean of the integrations is negative for the 1965 start date and equal to approximately −0.06°C per decade, hence in agreement with the observed one. Given the linearity of the post-processing adopted, the trend in the 1994 start date is necessarily equal and opposite to that of the 1965 start, i.e. equal to 0.06°C per decade. Despite the smaller magnitude of the warming trend for the 1994 start integrations compared with the observed one, the signs of the trends are captured by the model. By considering the 95% CI associated to the linear trend, the observed and experiment trends are in good agreement (table 3).
Similar results hold even at regional level, as for instance for the North Atlantic (50° W–10° W/40° N–60° N) area, as shown in figure 3 and table 3. For the 1965 start, the trend is −0.14°C per decade for the mean of the integrations and −0.81°C per decade for the observations. For the 1994 start, the mean trend of the integrations is again equal and opposite, 0.14°C per decade, and for the observations it is 0.41°C per decade.
Overall, these results show that, at least in terms of near-surface temperature, there appears to be a predictable signal in the ICs chosen. The next step is to understand to what extent this signal is due to the different ICs and, instead, what is the contribution of the GHGs forcing. This is addressed in §5.
5. Impact of initial conditions and greenhouse gas forcing
The integrations presented in §4 are characterized by two main differences: the ICs and the values of the GHGs. The latter difference arises not only in the absolute values of the GHGs in 1965 and 1994—with 1994 with higher values in all the five GHGs considered—but also in the rate of change (or trend), again with 1994 having a larger rate of change than 1964 (table 2).
In order to understand the sensitivity of the coupled model to the GHGs forcing conditions, i.e. to different levels of GHGs, two additional integrations for each start date were carried out: one in which the GHGs were kept fixed at 1965 levels and the other at 1994 levels (ensemble members 5 and 6 in table 1).
The impact of GHGs on global 2 m temperatures can already be observed in the first few years of integration, as shown in figure 4. The grey lines represent the ‘best estimate’ (the integration with time-varying GHGs, member 1, table 1) for each of the two start dates and are plotted here for reference. The black thick lines—the integrations with fixed 1994 GHGs levels otherwise identical to member 1—separate within the first few years of integration from their reference run. The black thick line is particularly responsive from the 1965 start for which there is a large increase in temperature of approximately 0.7°C in only 1 or 2 years, approximately 2 years after the start of the run. Arguably, the variation in GHGs is quite dramatic, especially that for CO2 which effectively suddenly jumps from 320 to 359 ppmv. However, for the analogous integration, i.e. when the GHGs change from 1994 to 1965 level (black thin line on the right-hand side of figure 4), the transition to a cooler state is less abrupt. This might be an indication of an asymmetric response of the climate system to changes in GHGs, even if it is not possible to draw any firm conclusion with only one realization per start date. However, over the whole 20-year integration, the temperature variation for the 1965 GHGs starting in 1994 is larger than that for the 1994 level starting in 1965 (black thick line on the left-hand side of figure 4), approximately 0.6°C versus approximately 0.3°C, even though the magnitude of the change in GHGs is the same.
By dividing the 20-year time series in subsets and taking a sliding window of 1-year increments, it is possible to estimate whether and how rapidly the distribution of pairs of experiments will diverge/converge. To do this, the pair of distributions of each subset is tested for significance. Under the null hypothesis that pairs of distributions are identical, a simple two-sample two-sided t-test, which also takes into consideration the temporal dependencies of the time series via their autocorrelation, has been applied.
When we consider the pairs of integrations in which different GHG values are applied to the same IC (the black thick/black thin pairs in figure 4), the separation of time series in both pairs, at 0.05 significance level, happens within the first few years of integration. The information contained in the ICs is therefore rapidly obliterated by the sudden change in GHGs concentration. Analogous conclusions are drawn from the comparison between the other two pairs of integrations (the black thick/black thick and black thin/black thin pairs in figure 4). Differences between these integrations test the impact of ICs in the presence of constant GHGs. Despite starting from ICs, which differ by more than 0.5°C, the global 2 m temperatures of these two pairs become indistinguishable at 0.05 level between year 5 and year 10 of integration.
To summarize the results of these sensitivity experiments, GHGs have a rapid impact on the global 2 m temperature when a sudden GHGs variation is applied to the same ICs. That GHGs have a noticeable impact within the first months of integration was to be expected (Doblas-Reyes et al. 2006). We have found that by years 5–6 the global mean 2 m temperature appears to be determined entirely by the values of the GHGs. A similar impact is observed when the same GHGs variation is applied to different ICs, although in this case the impact of GHGs is slower.
A parallel can also be drawn with the work of Collins & Allen (2002): the first pairs of experiments (black thick/black thin pairs in figure 4) can be viewed as an indication of predictability to changing GHG forcing (predictability of the second kind), while the latter pairs (black thick/black thick and black thin/black thin pairs in figure 4) give an indication of predictability to a change in ICs (predictability of the first kind). Although it is difficult to directly compare Collins & Allen's results with ours, as in their case the change in GHGs was more gradual, there seems to be an agreement in the level of predictability of the first kind, for which ICs are predictable for global 2 m temperatures of order of a few years, and also for the predictability of the second type, which dominates for time scales longer than approximately 10 years.
Not only do the GHGs affect the globally averaged 2 m temperature, but their signature can also be seen at regional level, as for instance for the African continent. Figure 5 shows the impact of GHGs in the predictability of the second kind in terms of statistical significance for the extended boreal summer (June–July–August–September, JJAS). Many parts of the African continents (and of the Mediterranean basin) are statistically warmer at 90% or higher in the first 10 years of integration (figure 5a) with 1994 GHG levels compared to 1965 levels. The following 10 years (years 10–20; figure 5b) reinforce the separation between the runs with different GHGs, showing their impact over a large portion of the region considered.
Even for the predictability of the first kind, the GHGs signal is apparent (figure 6). Parts of the African continent as well as large areas in the Atlantic and Indian Oceans are statistically warmer in the integration with 1994 ICs (figure 6a). However, all these differences disappear in the following 10 years (years 10–20; figure 6b) when the IC signal has been completely dominated by the GHGs forcing.
A similar global and regional analysis has also been carried out for precipitation. Some signal due to the GHGs forcing can be seen in this case too, but the results are not as conclusive as for the 2 m temperature and need further investigation in order to ascertain whether some useful predictable signal can be extracted.
6. Summary and outlook
With this work, we explored the potential for predictability of a coupled model on the decadal time scale. The coupled model was initialized from directly analysed ICs, i.e. as close as possible to the observed climate state as normally done in numerical weather and seasonal climate forecasting.
An ensemble of 20-year forecast integrations were run starting from two dates: one in 1965 (preceding a decade of global cooling) and the other in 1994 (preceding a decade of global warming). Considerable differences were found between the two forecast ensembles, indicating the potential for predictability. Supplementary integrations were carried out to ascertain the extent to which these differences arose from the different GHG concentrations as opposed to different ICs. The results here suggest that ICs are a source of predictability for global 2 m temperature only for the first few years of the coupled model integration. Not surprisingly, beyond the first few years GHGs seem to be the major source of predictability, in agreement with, for example, Collins & Allen (2002). Similar results were obtained for a regional analysis of the African continent. However, the precipitation signals remain elusive at this stage and further investigation is needed to ascertain the impact of ICs and GHGs on this variable.
The interannual-to-decadal time scale is very important both in terms of science and decision making. Regarding the latter, investment/policy decisions on adaptation to climate change would greatly benefit from improved forecast information on this time scale. Given the relatively large impact of radiative forcing changes we found on the decadal time scale, even climate-change mitigation policy might benefit from such information.
As for the scientific aspect of this study, the debate is still open on whether, for this time scale, the initialization of the coupled model should be done keeping the model close to the model attractor, as traditionally done in climate-change type of simulations, or as close as possible to the observed climatic attractor, as traditionally done in weather and seasonal forecasts. As articulated in the World Climate Research Programme's recent strategic initiative, an ultimate goal is to achieve a ‘seamless’ system which combines weather and climate prediction. It is apparent that current limitations lie with the magnitude of model drift. In this paper, we have attempted to show that the weather/seasonal forecast approach is viable on longer time scales, provided the drift is handled a posteriori, e.g. in a similar way to what is done for seasonal forecasts.
Even if predictions in the presence of drift are possible, it is nonetheless desirable to try to reduce the drift. There are several paths to reducing climate drift. One is to improve the model formulation/parametrization. Indeed, more recent versions of the IFS atmospheric model than that used in this study show a reduced surface temperature drift in the first few months of integration. Another path is to improve, or even re-think, the initialization methodology. Yet another is to make simulations at significantly higher resolutions than are currently possible. However, any of these paths is expensive either in terms of human resources and/or computing power. Given the importance of the problem of quantifying accurately the threat of climate change, reducing drift by one or more of these paths is paramount.
A.T. has been supported by the European Union projects ENACT (EVK2-2001-00077) and MERSEA (AIP3-CT-2003-502885). The authors would like to thank David Anderson, Paco Doblas-Reyes, Malcolm MacVean, Jean-Jacques Morcrette, Tim Stockdale and Antje Weisheimer for their fruitful discussions.
One contribution of 13 to a Theme Issue ‘Ensembles and probabilities: a new era in the prediction of climate change’.
↵See http://www.ecmwf.int/research/EU_projects/ENACT/index.html for the ENACT final report.
↵This was the case for instance in most papers which appeared in the Report of the World Climate Programme Workshop ‘Simulation of Interannual and Intraseasonal Monsoon Variability’ (March 1992) and where only the summers of 1987 and 1988 were considered.
↵In the case of two start dates, it is equivalent to refer to anomalies with respect to their climate or to their direct difference. The advantage with the former approach is that it can be easily generalized to a larger number of start dates.
- © 2007 The Royal Society