Figure 8.1: Decadal mean SST errors relative to the GISST climatology for (a) the non-flux adjusted model HadCM3 (Gordon et al., 2000), (b) the previous generation, flux adjusted model HadCM2 (Johns et al. 1997), (c) the HadCM2 model when run without flux adjustments (Gregory and Mitchell, 1997). The figures are from representative periods after at least 100 years of each control run. Multi-century drifts in each run are much smaller than the differences between the runs. The errors are smallest in (b), because the flux adjustments were chosen specifically to minimise the errors. The errors in (c) result from a number of complex feedbacks. The model was designed to work in flux adjusted mode and it is possible that the non-flux adjusted SST errors could have been reduced by relatively minor "tuning" of the model.
In this chapter, we assess climate models on the basis of their ability to simulate present and past climates. What it means to simulate a particular climate state is linked to the question of model initialisation. Ideally, given a "perfect model", and "perfect knowledge" of the present climate state, one could simply initialise a climate model with the present state. Then, given perpetual present day forcing (from trace substances and solar radiation), one might expect the model to remain close to the present state, perhaps with some level of variability. However, in practice, this ideal is not achieved and a model initialised in this way adjusts from the initial state. This adjustment has been characterised by two time-scales (Bryan, 1998): in the initial "fast" adjustment, the atmosphere, land surface, ocean mixed layer and sea ice reach a state of near equilibrium, typically taking 5 to 50 years. Surface temperature error patterns (e.g., Figure 8.1a) are generally established on this time-scale, and persist over many centuries of integration, but there typically remain slight imbalances in the surface heat and fresh water fluxes. These imbalances drive a second, slower, adjustment phase (often called "climate drift") which takes place over centuries to millennia and involves adjustment of the deep ocean to the surface imbalances. This "climate drift" can make interpretation of transient climate change simulations difficult, so models are generally allowed to adjust to a state where such drifts have become acceptably slow, before starting climate change simulations. A number of techniques have been used to achieve this (see Stouffer and Dixon, 1998), but it is not possible, in general, to say which of these procedures gives "better" initial conditions for a climate change projection run.
A number of model runs have now been made which are forced by the historical record of natural and anthropogenic forcing from the mid-19th century to the present. As well as avoiding the "cold start" problem in climate projections (SAR Chapter 6, Section 6.2.4 and Chapter 9), these runs can be compared against historical observations and form a valuable model evaluation tool (Section 8.6.1).
Given present day greenhouse gas concentrations, most coupled models at the time of the SAR had difficulty in obtaining a stable climate near to the present day state. Therefore "flux adjustment" terms were often added to the surface fluxes of heat, water and (sometimes) momentum which were passed from the atmosphere to the ocean model. Flux adjustments are non-physical in that they cannot be related to any physical process in the climate system and do not a priori conserve heat and water across the atmosphere-ocean interface. The flux adjustments were specifically chosen to give a stable and realistic simulation of present surface climate (especially the sea surface temperature and sea-ice cover), and were often as large as the annual mean model fluxes themselves. The need to use such adjustments was clearly a source of uncertainty: the approach inherently disguises sources of systematic error in the models, and may distort their sensitivity to changed radiative forcing. Models which did not use flux adjustment produced unrealistic simulations of fundamental aspects of the climate system such as the strength of the North Atlantic thermohaline circulation (SAR Chapter 5, Table 5.5).
Recently a number of coupled models have emerged with greatly improved surface climatologies without using flux adjustments. Figure 8.1a shows SST errors from one such model, about 100 years after initialisation. Errors are generally less than 2°C, and the error pattern shown is stable over several centuries of integration. However some larger errors are seen, in particular a cooling in the North Pacific, a warming in the eastern tropical ocean basins (probably due to a lack of stratocumulus cloud there), and a warming in the southern ocean. These errors appear to be common to a number of more recent non-flux adjusted models (Action de Recherche Petite Echelle Grand Echelle/Océan Parallélisé (ARPEGE/OPA1), Guilyardi and Madec 1997; Climate System Model (CSM 1.0), Boville and Gent 1998; Hadley Centre Coupled Model (HadCM3), Gordon et al., 2000), but other models show different error patterns (ARPEGE/OPA2, Institut Pierre Simon Laplace/Coupled Atmosphere-Ocean-Vegetation Model (IPSL-CM2), Barthelet et al., 1998a,b). The surface climatologies in several non-flux adjusted models are now considered good enough and stable enough to use those models for climate change projections. Typical flux adjusted models do show smaller SST errors, because the flux adjustments are chosen specifically to minimise those errors (Figure 8.1b). For comparison, Figure 8.1c shows the SST errors when the older, flux adjusted model was run in non-flux adjusted mode.
It appears that the success of the recent models which do not require heat flux adjustments is related to an improved ability to simulate the large-scale heat balances described in Chapter 7 (Weaver and Hughes, 1996; Guilyardi and Madec, 1997; Johns et al., 1997; Bryan, 1998; Gordon et al., 2000). Improvements to both atmospheric (Section 184.108.40.206.2) and oceanic (Section 220.127.116.11) components of the models have played a part in this advance. Such models have the advantage over flux-adjusted models that, provided the large-scale balances are obtained using a physically justifiable choice of model parameters, these models are physically self-consistent representations of the climate system. However, in some cases only very loose physical constraints can be placed on the model parameters.
The fresh water budget is more complex than the heat budget because of the effects of land surface processes, rivers and sea ice. Water budget errors are potentially far reaching because there is no direct feedback between surface salinity errors and the surface fresh water flux, for example, persistent freshening at high latitudes could lead to a collapse of the ocean thermohaline circulation (Manabe and Stouffer, 1997; see also Chapter 7, Section 7.6.2). Some aspects of the large-scale hydrological cycle are subject to large observational uncertainty (Wijffels et al., 1992), and this has inhibited evaluation and improvements in the water budget. Nonetheless, some models are now able to produce stable multi-century runs without water flux adjustments.
Marotzke and Stone (1995) show that using flux adjustment to correct surface errors in the control climate does not necessarily correct errors in processes which control the climate change response. Flux adjustments can also result in spurious multiple equilibrium states of the tropical (Neelin and Dijkstra, 1995) and thermohaline (Dijkstra and Neelin, 1999) ocean circulation. On the other hand, a good representation of, say, sea-ice extent may be important to produce the correct magnitude of ice-albedo feedback under climate change, and it may be preferable to use flux adjustments to give a good sea-ice distribution than to omit the flux adjustments but to have a poorer sea-ice extent. Overall, differences have been seen in the climate change response of flux adjusted and equivalent non-flux adjusted models (Fanning and Weaver, 1997b; Gregory and Mitchell, 1997), but it is not clear whether the differences are due to the flux adjustment itself, or to the systematic errors in the non-flux adjusted model. The only practical way to resolve this issue may be to continue the progress which has been made towards models which achieve good surface climatology without flux adjustment, whereupon the effect of flux adjustments will cease to be of concern.
Other reports in this collection