英文摘要: | The question of how climate model projections have tracked the actual evolution of global mean surface air temperature is important in establishing the credibility of their projections. Some studies and the IPCC Fifth Assessment Report suggest that the recent 15-year period (1998–2012) provides evidence that models are overestimating current temperature evolution. Such comparisons are not evidence against model trends because they represent only one realization where the decadal natural variability component of the model climate is generally not in phase with observations. We present a more appropriate test of models where only those models with natural variability (represented by El Niño/Southern Oscillation) largely in phase with observations are selected from multi-model ensembles for comparison with observations. These tests show that climate models have provided good estimates of 15-year trends, including for recent periods and for Pacific spatial trend patterns.
The differences between climate model forecasts and projections1 have come to prominence over interpretation of model simulations of recent temperature trends. A key difference between a climate forecast and a climate projection is that the former attempts to account for the correct phase of natural internal climate variations whereas the latter does not and can not. A weather forecast attempts to account for the growth of particular synoptic eddies and is said to have lost skill when model eddies no longer correspond one to one with those in the real world. Similarly, a climate forecast of seasonal or decadal climate attempts to account for the growth of disturbances on the timescale of those forecasts. This means that the model must be initialized to the current state of the coupled ocean–atmosphere system and the perturbations in the model ensemble must track the growth of El Niño/Southern Oscillation2, 3 (ENSO) and other subsurface disturbances4 driving decadal variation. Once the coupled climate model no longer keeps track of the current phase of modes such as ENSO, it has lost forecast skill for seasonal to decadal timescales. The model can still simulate the statistical properties of climate features from this point, but that then becomes a projection, not a forecast. The Coupled Model Intercomparison Project 5 (CMIP5) series of coupled climate models have been run in climate projections mode5 and, to a limited extent, in decadal forecasting mode6. The models run as climate projections apply best estimates of the historical sequence of radiative forcing of climate for the past (until 2005), followed by specified future forcing scenarios. That means that until 2005 the models attempt to stay in sequence only with the year to year and decade to decade fluctuations in climate caused by the historical variation in radiative forcing (and not with internal variations). Decadal variations in surface climate are due to a range of factors7, 8. These factors include radiative forcing variations, but they are not the only, or even the most important, factors. Natural internal variations in the climate system also drive decadal variations9, 10, 11. These can occur for example through variations in the rate at which the ocean circulation takes up the additional heat added to the atmosphere from greenhouse forcing12, 13. Long-term variations in the preference of the coupled system for La Niña and El Niño states (the Pacific Decadal Oscillation14 (PDO)) change the rate of ocean heat uptake and are a key driver of decadal variability13, 15. In the CMIP5 models run using historical forcing there is no way to ensure that the model has the same sequence of ENSO events as the real world. This will occur only by chance and only for limited periods, because natural variability in the models is not constrained to occur in the same sequence as the real world. For any 15-year period the rate of warming in the real world may accelerate or decelerate depending on the phase of ENSO predominant over the period. That means that for a set of model projections well calibrated to the range of natural variability, there will be some 15-year periods where the observed rate of warming is in the low tail of the distribution of model trends for that 15-year period, and some 15-year periods where the observed rate of warming is in the high tail of the model distribution. These cases are illustrated by the 15-year observed trend and CMIP5 model trend distribution for 1998–2012 (Fig. 1a) and 1984–1998 (Fig. 1b). These two periods are no more meaningful in evaluating projections than any other 15-year periods, and we focus on the former here only to evaluate claims made about model projections in the most recent 15-year period7, 16.
Here we develop and apply an alternative simple, but natural (unforced), method to analyse model projections in phase with the real world. The models are not given any information about the observed state in this approach. The method takes advantage of the fact that, by chance, some of the CMIP5 model runs will be at least partially in phase at any given time with internal variability in the real world. One of the major processes driving variability in the rate of ocean heat uptake is ENSO (ref. 13). The long-term cycle of ENSO in switching between El Niño-preferred periods (slower ocean heat uptake decades) and La Niña-preferred periods (faster ocean heat uptake decades) will by chance line up in some of the CMIP5 models. To select this subset of models for any 15-year period, we calculate the 15-year trend in Niño3.4 index24 in observations and in CMIP5 models and select only those models with a Niño3.4 trend within a tolerance window of ± 0.01K y−1 of the observed Niño3.4 trend. This approach ensures that we select only models with a phasing of ENSO regime and ocean heat uptake largely in line with observations. In this case we select the subset of models in phase with observations from a reduced set of 18 CMIP5 models where Niño3.4 data were available25 and for the period since 1950 when Niño3.4 indices are more reliable in observations. Figure 4 shows the running 15-year trends for observations in red (Goddard Institute for Space Studies18 (GISS) top row; Cowtan and Way17 bottom row) and for the subset of CMIP5 models that fell within the Niño3.4 trend tolerance window in blue in the left column. The right column shows model 15-year trends for the subset of models in grey that were furthest from the observed Niño3.4 slope (least in phase with observations), where the subset of models is constrained so that it contains the same number of models that fell within the best-fit tolerance window. The size of plotting symbol for model trends is proportional to the number of models that fell within the Niño3.4 tolerance window. The solid lines are loess-smoothed fits to the trend points. The loess smoothing is weighted by the number of models that contributed to each observation. The shaded areas surrounding each loess line represent approximate 95% confidence intervals.
|