A climate change simulation starting from 1935

Climate Dynamics (1995) 11:71-84 limui¢ Uynumia © Springer-Verlag 1995 A climate change simulation starting from 1935 U Cubasch 1, GC Hegerl 2, A Heilbach 1, H H6ck ~, U Mikolajewicz 2, BD Santer 3, R Voss 1 Deutsches Klimarechenzentrum, Bundesstr. 55, D-20146 Hamburg, Germany 2 Max-Planck-Institut ft~r Meteorologie, Bundesstr. 55, D-20146 Hamburg, Germany 3 PCMDI/Lawrence Livermore National Laboratory, Livermore, CA 94550, USA Received: 28 December 1993/Accepted: 7 September 1994 Abstract. Due to restrictions in the available computing resources and a lack of suitable observational data, transient climate change experiments with global coupled ocean~atmosphere models have been started from an initial state at equilibrium with the present day forcing. The historical development of greenhouse gas forcing from the onset of industrialization until the present has therefore been neglected. Studies with simplified models have shown that this "cold start" error leads to a serious underestimation of the anthropogenic global warming. In the present study, a 150-year integration has been carried out with a global coupled ocean-atmosphere model starting from the greenhouse gas concentration observed in 1935, i.e., at an early time of industrialization. The model was forced with observed greenhouse gas concentrations up to 1985, and with the equivalent CO2 concentrations stipulated in Scenario A ("Business as Usual") of the Intergovernmental Panel on Climate Change from 1985 to 2085. The early starting date alleviates some of the cold start problems. The global mean near surface temperature change in 2085 is about 0.3 K (ca. 10%) higher in the early industrialization experiment than in an integration with the same model and identical Scenario A greenhouse gas forcing, but with a start date in 1985. Comparisons between the experiments with early and late start dates show considerable differences in the amplitude of the regional climate change patterns, particularly for sea level. The early industrialization experiment can be used to obtain a first estimate of the detection time for a greenhouse-gas-induced near-surface temperature signal. Detection time estimates are obtained using globally and zonally averaged data from the experiment and a long control run, as well as principal component time series describing the evolution of the dominant signal and noise modes. The latter approach yields the earliest detection time (in the decade 1990-2000) for the time-evolving near-surface temperature signal. For global-mean temperatures or Correspondence to: U Cubasch for temperatures averaged between 45°N and 45°S, the signal detection times are in the decades 2015-2025 and 2005-2015, respectively. The reduction of the "cold start" error in the early industrialization experiment makes it possible to separate the near-surface temperature signal from the noise about one decade earlier than in the experiment starting in 1985. We stress that these detection times are only valid in the context of the coupled model's internally-generated natural variability, which possibly underestimates low frequency fluctuations and does not incorporate the variance associated with changes in external forcing factors, such as anthropogenic sulfate aerosols, solar variability or volcanic dust. 1 Introduction Since 1989, several modelling groups have used global coupled ocean-atmosphere models to simulate the response of the climate system to a transient increase in greenhouse gas concentrations (for an overview see Houghton et al. 1992). All of these simulations follow the same basic experimental strategy. Due to a lack of comprehensive data sets describing the current state of the ocean, the oceanic part of the coupled model is spun up to an equilibrium with presen t day forcing conditions ( o r a state reasonably close to it). The oceanic and atmospheric models are then coupled together and run with present day forcing to obtain the mean state and variability of the current climate. This control experiment provides a reference state for climate change experiments with time-dependent changes in greenhouse-gas (GHG) concentrations. The control and the climate change experiments are typically integrated over time periods of 100-1000 years (Manabe et al. 1991; Manabe et al. 1992; Meehl and Washington 1993; Cubasch et al. 1992; Stouffer et al. 1994). Virtually all transient experiments performed with fully-coupled global atmosphere-ocean general circula- 72 Cubasch et al.: A climate change simulation starting from 1935 Total warming from t b / / Total f o r c i n g / CoLd start 1 . 1 , / Total ~__L__ ta tb Forcing _--1_ CoLdstart J CoLd start warming from t b Fig. 1. Schematic diagram of the "cold start" problem (after Hasselmann et al. 1993) t,, preindustrial starting point of the simulation; tb, actual starting point of the simulation tion models (AOGCMs) show a smaller increase in global mean temperature in the first few decades compared to the last decades. The magnitude of this initial temperature retardation is strongly model-dependent (Houghton et al. 1992). The slow initial temperature increase has been called the "cold start" effect (Hasselmann et al. 1993), and was first discussed by Hansen et al. (1988). The cold start is attributable to the experimental procedure: the atmospheric and oceanic components of the coupled model are at equilibrium with respect to present-day G H G concentrations. This means that the effects of G H G increases from the beginning of industrialization (about 1750) to the present are neglected (see Fig. 1). In the real world, this history of G H G forcing may have lead to the sequestering of heat in the intermediate and deep ocean. Experiments with present-day starting dates effectively ignore any previous oceanic warming. Note, however, that the cold start effect is primarily a problem of limited computer resources, which did not allow to gain enough experience to carry out the experiments in an appropriate way. Multiple A / O G C M integrations of 200-300 years in length have only become technically feasible within the last few years, so that long experiments with starting dates well before the present are only now becoming possible. Hasselmann et al. (1993) first suggested a method to correct for the cold start. This procedure makes it possible to 'correct' a transient experiment starting from 1985 G H G concentrations and to estimate how global mean temperature would have evolved if the experiment had started from pre-industrial conditions. The method involves fitting a linear response model in order to simulate the global mean behavior of the coupled ocean-atmosphere model. This response model is then used to calculate the cold start error and to correct the global mean near-surface temperature obtained by the coupled ocean-atmosphere model. This procedure has been applied in Cubasch et al. (1992, 1994) to correct various model experiments. A further evaluation by Fichefet and Tricot (1992) using a 1-dimensional model indicated that the cold start error becomes increasingly important the later an integration starts. They estimated that a Scenario A integration starting in 1990 would underestimate the warming realized in 2050 by about 15% relative to an integration starting in 1765. This error estimate was based on an assumed response time of 35 y (a measure of the oceanic thermal inertia), and the magnitude of the error was found to be highly sensitive to the choice of response time for the range which the authors considered (15, 35 and 75 y). In theory it is possible to apply a linear approach similar to that used by Hasselmann et al. (1993) to obtain a regional distribution of the cold start correction. Such a linear correction technique, however, can de-. scribe only part of the typically nonlinear interactions calculated in a comprehensive global coupled model. In the current study, we use an experiment with a global coupled ocean-atmosphere model to address the question of whether the linear assumptions of the Hasselmann et al. (1993) cold start correction are valid. We also consider the question of the extent to which the cold start influences the regional details of the greenhouse warming signal. To investigate these issues, we perform a time-dependent greenhouse warming experiment starting from the G H G concentrations appropriate to the year 1935. This is not a true pre-industrial starting point, but the results of Fichefet and Tricot (1992) indicate that the cold start error (relative to a true pre-industrial starting date of 1765) is less than 3-5% for this particular starting date (assuming oceanic response times in the range 15-75 y). Thus the benefits of a minor reduction in the cold start error due to an even earlier starting time do not at present justify the additional computing time required. The cold start is but one source of uncertainty in estimating the space-time properties of an evolving greenhouse warming signal. A further source of uncertainty is introduced by decadal-to-century time scale natural variability of the coupled atmosphere-ocean system, and the consequent sensitivity of the spacetime signal properties to initial conditions. Studies with stochastically-forced ocean general circulation models suggest the existence of oceanic circulation modes with time scales of centuries (Mikolajewicz and MaierReimer 1990). It is possible that comparable centurytime scale variability will be found in long, fully-coupled integrations. Such internally-generated fluctuations will be superimposed on any time-evolving G H G signal, and their space-time evolution may depend on initial conditions (particularly of the intermediate and deep ocean) at the start of the experiment. The possible impact of initial condition uncertainties on detection times has been analyzed by Cubasch et al. (1994) for selected atmospheric variables. In this study, a suite of four 50-y "Monte Carlo" experiments were performed, each using the same global coupled ocean-atmosphere model, but starting from years 0, 30, 73 Cubasch et al.: A climate change simulation starting from 1935 60 and 90 of a control simulation. The global mean 2 m temperature response to the G H G forcing was delayed between 11 and 37 y. This indicates that initial condition uncertainties can exert a considerable influence on the timing of the surface temperature response, an effect which is superimposed on the cold start problem. An analysis of the same suite of Monte Carlo experiments by Santer et al. (1994) showed similar initialcondition sensitivity for spatially-integrated ocean variables. The implication is that estimates of the magnitude of the cold start error are themselves sensitive to initial condition uncertainties and the model-generated natural variability superimposed on any G H G signal. Here we consider only one early industrialization experiment, but note that a more rigorous investigation of the cold start error may require a suite of different initial condition realizations with a fully coupled model. We first present a brief description of the coupled ocean-atmosphere model and the experimental design (Sect. 2), and then compare the behavior of globallyaveraged near-surface temperature in the early industrialization experiment, which starts in 1935, and the cold-start corrected Cubasch et al. (1992) Scenario A experiment (henceforth SZA), with a starting date in 1985 (Sect. 3.1.1). We then analyze the regional differences between the near surface temperature and sealevel signals in the two experiments (Sect. 3.1.2), and finally estimate detection times for the near surface temperature signal using globally and zonally averaged data and principal component time series (Sect. 3.2). A short discussion is given in Sect. 4. 2 The early industrialization experiment 2.1 The coupled model The coupled global ocean-atmosphere model used in the early industrialization experiment has been described in Cubasch et al. (1992). The atmospheric component (ECHAM-1) comprises 19 levels in the vertical and has a horizontal resolution of T21 (with a Gaussian grid of about 5.625°). It has been synchronously coupled to the large scale geostrophic (LSG) ocean general circulation model (Maier-Reimer et al. 1993), which has 11 layers in the vertical and employs a horizontal grid spacing similar to the Gaussian grid of the T21 model. The coupling procedure includes flux correction in order to minimize climate drift. The ECHAM-1/LSG coupled model has been used in a number of climate change experiments (Bakan et al. 1991; Cubasch et al. 1992, 1994). 2.2 The initial conditions The coupled model was initialized with the equivalent COs concentration appropriate to the year 1935. At this time, information about the sea surface temperature (SST) was spatially incomplete, with large data gaps in the Southern Ocean and in areas of the Pacific and Atlantic not covered by ship routes (Jones and Briffa 1992). Consequently, there is some degree of subjectivity in determining the SST initial conditions for the early industrialization (henceforth EIN) experiment. We used the following strategy: the coupled model was run for 35 years, starting from initial SST conditions of the Cubasch et al. (1992) control integration, but with radiative conditions appropriate to 1935 rather than 1985. The time scale of 35 years was chosen after calculations with a simple linear response model (for details, refer to Hasselmann et al. 1993). The 35-y integration time appears to be sufficient to adjust the SST from 1985 to 1935 conditions, although it is obviously not long enough to allow the deep ocean circulation to reach an equilibrium with the forcing. This strategy is justifiable since the differences between the observed climatological SST and the SST in the year 1935 are generally smaller than the uncercainties involved in reconstructing a spatially-complete picture of SSTs in 1935 and spinning-up the ocean-only model with these reconstructed SSTs. The flux correction was the same as that used in the experiment starting from 1985 conditions, since the differences in the climate system (i.e., mainly in the SST fields) between 1935 and 1985 are too small to justify a recalibration of the correction terms. Any errors made in using correction terms appropriate to 1985 rather than 1935 are likely to be small relative to the changes that are expected in the climate change experiments. The EIN experiment was then run for a period of 150y: from 1935-1985 with the observed record of equivalent CO2 changes (Houghton et al. 1990), and from 1985 onwards with the equivalent CO2 increase stipulated in the IPCC Scenario A. The radiative effects of greenhouse gases other than CO2 are considered implicitly (as CO2 equivalents). The radiative forcing, which is proportional to the logarithm of the 7.0 g 6.5 o L.) 0 L.) :¢ 3 6.0 o J 5.5 1925 . . . . I 1975 . . . . I 2025 . . . . I , 2075 Time (yeors) Fig. 2. Time evolution of the equivalent CO2 concentration between 1935 and 2085 according to Scenario A (IPCC 1990) 74 Cubasch et al;: A climate change simulation starting from 1935 i . . . . 289 i MC30 I ~ MC60 J ........ MC90 EIN OBS [ 288 L--- I-- 287 286 , ..~ ~'~>,~J~.~-~ ~,e 1850 j~ ,~ , i 1900 ~# I , i i , , i 1950 , i I 2000 2050 timely] COs concentration, increases slowly from 1935 to 1985 before it rises almost linearly during the subsequent century (Fig. 2). The results of the EIN experiment will be compared with those of the 100-year SZA integration, which used the same Scenario A G H G increases (but starting from 1985), and with the 100-y control simulation (henceforth CTL). These experiments have been described by Cubasch et al. (1992) and Santer et al. (1994). Where appropriate, we also compare with results from a suite of four 50-year "Monte Carlo" experiments, each with identical Scenario A G H G forcing but starting from different initial conditions of the CTL integration (Cubasch et al. 1994). 3 Results We discuss our results in time coordinates appropriate to the forcing, i.e., from 1935 to 2085 rather than from modeI years 1-150 (for EIN) or 1-100 (for SZA). This is purely for convenient comparison of the two experiments; it should not be interpreted as a prediction that a particular climate change event will occur in a certain year. 3.1 Discussion of the cold start 3.1.1 Comparison with a linear model The EIN experiment starts in the year 1935 from a global mean annually-averaged temperature which is about 0.4 K lower than the starting value for the SZA integration (Fig. 3). The global mean temperature of the E1N experiment stays at this low level and warms only marginally by the year 1985. The higher temperature of the CTL simulation at its beginning can be attributed to an insufficient spin-up period for the coupled model (5 y). A subsequent analysis has shown (Cubasch et al. 1994) that it would have been necessary to integrate the coupled model in a spin-up run for at least 70 y to allow the Arctic sea ice volume to reach equilibrium. This is , , i , 21 O0 Fig. 3. Time evolution of global mean 5-year-averaged near surface temperature for the SZA, EIN, CTL and MC experiments and the observations (IPCC 1990). The observations, which originally only exist as anomalies with respect to the average of the years 1951-1980 have been scaled to the mean value of EIN during that period caused by inconsistencies of the flux correction at the ice edge. The problems in the simulation of sea ice are probably not as serious as suggested by Meehl and Washington (1990) since, in contrast to their experiments, the SST in the Hamburg coupled model is kept close to the observational values by the flux correction. The model warming of 0.1 K from 1935 to 1985 is smaller than the observed value, which is approximately 0.25 K (Houghton et al. 1990). This discrepancy of -0.15 K is small relative to the two-standard deviation noise level estimated from a 250-y control simulation ( + 0.31 K for annual mean near surface temperature, + 0.25 for 25-y averages). The difference between observed and model warming may have a number of explanations. First, it may be attributable to the differences between the true 1935 SST and the SST field which we generate using 1935 radiative forcing conditions (see Sect. 2.2). Second, the EIN experiment has its own cold start error. Third, we are comparing an observed temperature record which represents some combination of internallygenerated natural variability and response to multiple external forcings with a simulated record which represents only natural variability and the response to timedependent G H G forcing. Fourth, we are comparing only one model realization with the observed record of temperature changes. The sensitivity to initial conditions shown by Cubasch et al. (1994) implies that the use of slightly different 1935 SST conditions (but equally plausible conditions, given the uncertainties in reconstructing a spatially-complete picture of 1935 SSTs) could have resulted in different evolutions of the global mean warming than the one simulated by the EIN experiment. And finally, the cooling in the Arctic region, because the sea ice thickness is still increasing, has a noticeable effect, even though it cannot mask the global warming due to its limited spatial extent (a more comprehensive discussion of the sea ice variability and its possible causes can be found in Sect. 3.1.2). The striking similarity between the observed and modelled global annual average temperature changes over the period 1935-1985 is fortuitous, and is somewhat Cubasch et aL: A climate change simulation starting from 1935 . . 3 . . . . . . . i . . . . . . . . . q . . . . . . . . . ~°--- SEZA ~-.~ SZA(corrected) ~ i . /." / ~ s / ~ f / ~ EIN(corrected) /~//f ~ ////J 2 [.-, < 1 ,~//~,j" .X -oJ 0 ~ "" "~" ..-,."~ . . . . . . 1935 / ~ t " ~ , , , I ~ i r i i i , p i 1985 , I i 2035 I r i i T ~ i i I i R 2085 time[y] Fig. 4. The time evolution of the global mean 5-year-averaged near surface temperature changes in the uncorrected and "cold start", corrected SZA and EIN experiments. Changes are expressed relative to the decade 1986-1995 of the respective simulation surprising given all the possible explanations (outlined already) for model-versus-observed discrepancies. In the EIN experiment the rate of temperature increase from 1985-2005 is approximately 0.1 K per decade, increasing to roughly 0.35 K per decade thereafter. This final rate of change is slightly higher than that in the IPCC "best estimate", but is virtually identical to the rate of temperature change at the end of the SZA integration. To be consistent with the definition of climate change used in the theoretical paper of Hasselmann et al. (1993), global mean temperature anomalies have been calculated relative to the average of the decade representing the years 1986 to 1995 of each experiment (Fig. 4). Under this definition the anomaly time series for both the SZA and EIN experiments start at the same level in 1985, and their behavior can be compared easily. The global-mean annually-averaged temperature changes at the end of the SZA and EIN experiments are 2.6 K and 2.9 K, respectively. This systematic difference of approximately 0.3 K in the warming realized during the last 50 years of the two simulations cannot solely be attributed to the different manifestations of internal variability, but it is at least partially a consequence of the cold start of the SZA experiment. Its behavior in terms of globally-averaged temperature lags behind the EIN integration by roughly one decade. After correction for the cold start using the linear model described by Hasselmann et al. (1993), the SZA curve closely parallels the curve for the uncorrected EIN experiment, and is generally 0.10.2 K higher than the latter. This result is due to the fact that the EIN experiment starts too late to fully avoid a cold start error, so that its (uncorrected) warming response is expected to be less than that of the cold 75 start corrected warming in SZA. The differences between global mean temperature changes in the uncorrected EIN and cold-start corrected SZA experiments are well within the noise envelope defined by the internal variability of the coupled model. These results support the findings of the theoretical study by Hasselmann et al. (1993). We also used the method of Hasselmann et al. (1993) to compute a cold start correction for the EIN experiment. The cold start error at the end of the experiment is approximately 0.25 K. Theoretically the curves for the cold-start corrected EIN and SZA experiments should be in reasonable agreement, with any differences attributable solely to internally-generated variability. The differences between the two corrected curves are within the two standard deviation noise limit established by the control run. We note, however, that the corrected temperature increase is consistently lower for the SZA experiment than for the corrected EIN integration. One possible explanation for this result is that since both cold start corrections have been calculated by fitting a linear model to relatively noisy temperature increase curves, this fit has a considerable inaccuracies (compare with Cubasch et al. 1994). The uncertainties in this fitting procedure introduce attendant uncertainties in the magnitude of the cold start corrections. While it is straightforward to correct the global mean values for the cold start, it is much harder to apply cold start corrections on a regional basis. One would have to find a response function for every grid point and every variable and apply the correction locally, which might lead to physical inconsistencies (particularly if the linear models provide poor fits to the predicted data). It is conceivable that regional corrections could be derived by some combination of the Hasselmann et al. (1993) global-mean appoach with appropriate statistical methods, such as empirical orthogonal function (EOF) analysis (e.g., by correcting principal component time series, and then using the linear combination of these components and their corresponding EOFs to derive a cold start-corrected data set). Such correction approaches must also involve considerable uncertainties. Ultimately, it would appear more sensible to estimate the spatio-temporal impact of the cold start by performing multiple integrations of a global AOGCM from the onset of the Industrial Revolution, and starting each integration from different initial conditions. In addition to the cold start problem, differences in the initial state of the coupled atmosphere-ocean system can also cause a delay in the global warming. The impact of uncertain knowledge of the initial state has been analyzed by Cubasch et al. (1994) in a suite of four 50-y "Monte Carlo" experiments (MC), each with identical SZA G H G forcing but starting from years 0 (identical to the first 50 y of SZA), 30, 60 and 90 of the Cubasch et al. (1992) CTL simulation. During the first 100 y the CTL experiment cools by 0.3 K, so the initial states of the MC experiments can show substantial differences with respect to SST. The EIN experiment can 76 Cubasch et al.: A climate change simulation starting from 1935 BOX MODEL 3.0 LSG/ECHAM Sea Level Change global start ODegree - - - s t a r t 0.25 Degree . . . . start -0.25 Degree 2.0 / ~ " . f 0.2 I ~SZA " ] / PRE I SZA(corrected) PRE(correcfed) ~6 £3. E 4* fy- 0.1 i.o ~ ¢.. -1.E 1985 , 2005 , 2025 Years , 2045 2065 2085 BOX MODEL 2m Temperature (global.) - . - . . . start 0 Degree start 0.25 Degree start -0.25 Degree ..'" 1985 20'05 ' ......... ~......... 1985 2035 time[y] ''' 2085 Fig. 6. The time evolution of the global mean 5-year-averaged sea level changes of the EIN and SZA experiments relative to the decade 1986-1995 in the respective simulations ."/ .-"..//.-'.. ¢¢~ b- -0.1 . . . . . . . . . 1935 20'25 2045 Time (years) // 20'65 2085 Fig. 5a, b. Box model simulation of the MC experiments a, absolute; b, relative to the respective initial states be considered as an additional MC experiment, particularly since its initial state in 1985 (the starting year of the M C experiments) lies within the natural variability range of the control simulation, at least in global mean terms. A comparison of the post-1985 temperature changes in the MC, E I N and S Z A experiments indicates that the rate of increase during the initial 35 y of each experiment is highly d e p e n d e n t on the initial state (Fig. 3). After the year 2020, all experiments relax to the same t e m p e r a t u r e increase independent of their initial state. The S Z A and MC30 simulations start with a global mean t e m p e r a t u r e which is warmer than the average of the control simulation, while the other two M C experiments start with conditions which are colder than average. The E I N experiment has the coldest global mean conditions of all simulations in 1985. The warming during the first few decades of the simulation is dependent on this initial state, with reduced warming corresponding to warmer than average initial states and increased warming associated with cooler than average initial conditions. This behavior can be simulated to a first approximation with a simple box model which has been tuned to reproduce the time-dependent global mean temperature changes of the global coupled model. T h r e e cases were considered, with the box model simulation starting at the mean temperature of the control simulation, at a temperature about 0.25 K warmer than the mean, and at a temperature about 0.25 K colder than the mean. As can be seen in Fig. 5a, the temperature changes in the experiments starting at warmer and colder conditions relax to the curve of the experiment starting from the mean initial value after roughly 50 years. If the temperature changes are plotted relative to the respective initial states of the three experiments (Fig. 5b), the changes appears as a slower (faster) temperature rise for warmer (colder) initial conditions. It has to be stressed that the box model is only a coarse approximation of the global model and cannot represent nonlinear feedback mechanisms. For example, it is also possible that ice-albedo feedback might contribute towards a faster temperature rise when starting from colder than average conditions. The uncorrected globally averaged sea-level change due to thermal expansion shows some lowering at the beginning of the E I N simulation, which is a result of adjustment to the reduced forcing (Fig. 6) corresponding to the greenhouse gas forcing of the year 1935. After the spin-up period of 35 years (see Sect. 2.2), only the upper layers of the ocean have reached an equilibrium. Sea level changes due to the thermal expansion of sea water, however, involve the whole water column of the oceans (Cubasch et al. 1992) with equilibrium times in excess of thousand years, i.e., outside our tech- Cubasch et al.: A climate change simulation starting from 1935 nical possibilities. The experimental compromise leads to an underestimation of the sea-level rise. In the second half of the EIN simulation sea level rises by about 2.5 cm per decade, a rate similar to that obtained in the second half of the SZA experiment. From 1985 to 2020, however, the rate of sea level increase is more rapid in the EIN experiment. The earlier starting date and longer exposure to the radiative forcing yields a sea level increase of 18 cm at the end of the EIN integration, roughly 3 cm higher than the sea level change in SZA. We also calculated cold start corrections for the sea level changes in the EIN and SZA experiments (see Cubasch et al. 1992). Since the time constant for the sea level rise is much longer than for the surface temperature changes (in this case 70 y, consistent with Mikolajewicz et al. 1990), the effect of the cold start correction is substantial. For the EIN experiment, the correction is approximately 2.5 cm in 1985 and 4.4 cm in 2085. The corrected rise for the EIN experiment is less steep than for the corrected SZA experiment, while both corrected curves are virtually identical in the final 30 years. Sea level changes in both of the cold-start corrected curves are similar to the IPCC (1990) "low estimate" of 6.8 cm in the year 2030. 3.1.2 Regional impact of the cold start. As shown by Cubasch et al. (1994) and Santer et al. (1994), the climate change signal can be sensitive to the assumptions made regarding the degree of correlation between the variability displayed in the greenhouse warming and control integrations. These studies defined climate change signals relative to either the smoothed initial state (definition 1) or the instantaneous state (definition 2) of the CTL experiment. Neither definition is directly applicable here, since the EIN experiment starts in 1935 while the CTL experiment was performed with radiative forcing corresponding to 1985 conditions. We therefore choose to define the EIN changes relative to the average of the initial decade (1936-45) of the EIN experiment (definition 3). When the last 100 years of the EIN simulation are directly compared with the SZA simulation, changes are calculated relative to the decade 1986-1995 of the respective experiment (definition 4). The zonally averaged temperature changes according to definition 3 in the EIN simulation show a gradually evolving warming, reaching a maximum of 5-7 K over the Arctic and Antarctic at the end of the experiment (Fig. 7). During the first 120 years there is substantial cooling at high latitudes in the Northern Hemisphere, with maximum values in excess of - 7 K. This extensive cooling is due to an increase in sea ice volume. This increase seems to be unrealistic and might be an indicator that the sea ice was not yet at an equilibrium with the forcing at the start of the simulation. It is, however, very difficult to get an accurate estimate of the ice volume equilibration time from the EIN experiment, because three effects are involved and are impossible to separate. Firstly, the 35-year spinup is insufficient to achieve a true equilibrium of the coupled sys- 77 tem. Although the mixed-layer may be at equilibrium with respect to the 1935 radiative forcing, the intermediate and deeper ocean is not. As has been pointed out in Cubasch et al. (1993) and Santer et al. (1994), sea ice volume in the CTL integration required at least 60 to 80 years in order to reach a quasi-stationary state. Secondly, even if the model had been spun-up for several thousand years, there would be some long-periodic fluctuations about the equilibrium of the ice volume due to natural variability of the coupled system. And finally, the forcing is increasing with time in the EIN experiment. This is associated with warming and slow melting of ice. This effect is superimposed on the first and second effects, thus making it difficult to define an ice equilibration time in EIN. Any evaluation of the ice volume trends and the connected temperature change therefore suffers from ambiguity and is almost impossible to interpret. During the first 50 years of the EIN integration, the surface temperature response (definition 3) is spatially highly noisy, particularly at mid- to high latitudes in both hemispheres (see Figs. 7, 8). In this initial phase there is little evidence of the pronounced land-sea contrast signal component which was found in the first 50 years of the SZA integration (Santer et al. 1994) and in the mean of the Monte Carlo integrations performed by Cubasch et al. (1994). This is a reflection of the lower forcing in the first 50 years of the EIN experiment (see Fig. 2). The low temperatures on the extratropical continents must be seen in connection with the sea ice increases. The initial years of the EIN simulation do not show penetration of the warming into the deep ocean in areas of convective overturning, as described by Cubasch et al. (1992). This can have two reasons. First, the radiative forcing in the initial decades is relatively small. Second, the deeper layers of the ocean have not yet reached an equilibrium with the 1935 forcing, i.e., they are still too warm for the forcing. Thus, the heat uptake is strongly reduced. The spatial distribution of near-surface temperature changes in the final decades of the EIN and SZA experiments is shown in Fig. 9a, b (definition 4). The annual mean warming patterns are qualitatively highly similar. Both patterns are dominated by a strong landsea contrast (which is enhanced in the EIN integration relative to SZA), and show a cooling to the south of Greenland and reduced warming in the northeastern Pacific. In quantitative terms, warming during the last decade of the EIN simulation is larger than in the SZA simulation almost everywhere (Fig. 9c), with the exception of isolated areas in the vicinity of the Antarctic sea ice margin, the east coast of the USA and western North Atlantic, North Africa and northern Europe, and the western Pacific. We conclude from this that despite differences in the length of the overall forcing history, the initial conditions at the start of each experiment, and the reference states used to define the changes, a common nearsurface temperature pattern emerges in both the EIN and SZA experiments (see Sect. 3.1.3). 78 Cubasch et al.: A climate change simulation starting from 1935 Fig. 7. Hovmoeller diagram of the near surface temperature anomaly of the EIN experiment relative to its first decade (1936-1945) Fig. 8. Decadal average spatial pattern of annual mean near surface temperature change for years 1986-1995 of the EIN experiment. Changes are expressed relative to the average of the first decade (1936-1945) of the EIN experiment a) ~trq (2o76 - 20S~ b) SZA (2076 - 208b') e) E I N . SZA ( 7 , O 7 6 . 2 0 ~ ~C Pig. 9a-e. The spatial pattern of the annual mean near-surface temperature change in the final decade (2076-85) of the a EIN and h SZA experiments. Changes are expressed relative to the decade 1986-95 of the respective experiment. e The difference field (EIN minus SZA) for the two temperature change patterns 79 Cubasch et al.: A climate change simulation starting from 1935 a) industrialization experiment is also characterized by larger areas where sea level actually decreases, principally in the vicinity of the Antarctic coast and in regions of strong convective overturning. 3.1.3 EOF analysis of surface temperature changes. As Fig. 10a, b. Surface elevation. The spatial pattern of annual mean sea level changes in the final decade (2076-85) of the a EIN 2076-2085 minus EIN 1986-1995 and b SZA experiments (SZA 2076-2085 minus SZA 1986-1995) in Cubasch et al. (1992) and Santer et al. (1994), we performed a spatial EOF analysis in order to compare the dominant near-surface temperature signal patterns in the EIN and SZA greenhouse warming experiments. Temperature changes were defined relative to the decade 1986-95 in each experiment (i.e., years 1~-10 of •.SZA and 51-60 of EIN), and the first 50 years of the EIN integration were excluded. The EOF 1 patterns are dominant in both experiments, explaining 90.3% (EIN) and 84.3% (SZA) of the total spacetime variance (see Fig. 11 and Table 1). The EOF 1 patterns of the two greenhouse warming experiments are highly similar, regardless of whether the spatial means are subtracted (r) or included (r*) in the computation of the pattern correlation (r=0.73, r*=0.93; Table 1). As noted in the previous section, the common signal pattern emerges despite differences in the length of the overall forcing history, the initial conditions at the start of each experiment, and the reference states used to define the changes. A similar pattern correspondence was noted by Cubasch et al. a) The meridional overturning in the North Atlantic in the decade 2076 to 2085 is diminished by more than 30% relative to the CTL simulation in the case of the EIN experiment, which is a slightly higher reduction than in the SZA simulation. As discussed in Mikolajewicz et al. (1990) and Cubasch et al. (1992) the longer exposure of the surface water masses to positive P - E (precipitation minus evaPoration ) anomalies due to the reduced meridional exchange leads to a decrease in surface salinity at high latitudes and an increase in salinity in the subtropics (not shown). Mikolajewicz et al. (1990) first showed that GHGinduced ocean circulation changes can lead to regional changes in sea level which may be several times larger or smaller than the global mean increase in sea level due to thermal expansion. Figure 10 illustrates that the amplitude of the sea level response pattern response is larger in the EIN experiment than in the SZA integration. Increases >__24 cm occur over wide regions in the EIN integration; e.g., between Australia and South America and over much of the Arctic Ocean. The early Fig. l l a , b. The first EOF for annually-averaged near-surface temperature change in the a EIN and b SZA experiments b) Weight 8O Cubasch et al.: A climate change simulation starting from 1935 Table 1. Between-integrationpattern correlations and explained variance for EOF 1 of the EIN, SZA and CTL experiments EINso EINa5o Explained 55.8 Variance (%) EINso 1.0 EIN~so SZA CTL 90.3 SZA CTL 84.3 52.5 100 III (D Z < .< > ,.-., l.lJ z 0.15 (0.25) 1.0 0.06(0.21) 0.73 (0.93) 1.0 0.15 (0.18) 0.05 (0.15) 0.30 (0.03) 1.0 All correlations are computed with (r) and without removal of the spatial mean component (r*; in brackets). EINso: EOFs computed using data from years 1-50. EIN~5o:EOFs computed using data from years 51-150 (1992) and Santer et al. (1994) in comparing the nearsurface temperature EOF 1 patterns in the SZA experiment and an integration with instantaneous doubling of equivalent CO2. The suggestion is that these pattern similarities are at least partly related to the increasing disequilibrium between the rate of warming over land and ocean, i.e., the physics governing heat uptake and transport in the ocean and atmosphere eventually imprints itself on the surface temperature response. The dominant EIN signal pattern is essentially uncorrelated with the dominant mode of the control simulation (r=0.05; r*=0.15). A similar result is obtained for the correlation between the EOF 1 patterns of SZA and CTL (r=0.30; r* =0.03). When the EIN and SZA anomaly data are projected onto EOF 1 of SZA, the loading of the EIN data in 2085 is approximately 14% larger than the corresponding SZA loading, a further indication of the larger signal amplitude in the EIN simulation (see Fig. 15). We also computed near-surface temperature EOFs for the first 50 years of the EIN experiment, with anomalies defined relative to the initial decade (1935-44). The resulting EOF 1 pattern (not shown) is characterized by strong Arctic cooling, and slight warming at low latitudes and in northern Asia. It explains less than 56% of the variance, and is only weakly correlated with EOF 1 of SZA (r=0.06; r* =0.21) and the EOF 1 pattern computed for the final two-thirds of the EIN experiment (r= 0.15; r* =0.25). This indicates that the climate change signal which dominates the latter part of the EIN simulation is not clearly discernible during the first 50 years. An analysis of cumulative explained spatial variance as a function of time and number of EOFs confirms this result (Fig. 12). During the first 40 years of the EIN experiment, the EOF 1 explains only a small fraction of the total spatial variance of annually-averaged 2 m temperature changes, i.e., the signal is completely obscured in this first phase by noise. A large number of EOF patterns (50) are required to explain _>80% of the spatial variance. Thereafter the variance explained by the EOF 1 pattern grows linearly, explaining ___50% ot the spatial variance after 70 ~/ : ; , , 80 ;, J •t Ill 60 _ 40 ~, . ~ :2 &, I '~ \: ~ , ,','h"~['" t " ; ::: ". ..: 2O _ . ",: : : ~:,'d • : :.- ~ i .-:,."1 " ~ , A U IIJ / ~" [ ............ ..... EOFS 1- 2 EOFSi - ~o . EOFS 1 - 80 ....... EOFS1-50 ~ ] - 0 0 30 60 90 120 150 TIME (YEARS) Fig. 12. Cumulative spatial variance as a function of time and number of EOFs for the EIN experiment. EOFs were computed using annually-averagednear-surface temperature anomaly data from the full 150 years of the EIN experiment, with anomalies defined relative to the smoothed initial decade (1936-45) of the EIN integration years, and reaching an asymptotic level of ___90% after about 100 years. This behavior should be contrasted with that of the SZA experiment, where the dominant EOF 1 pattern explained _>50% of the spatial variance after only 40 years (Santer et al. 1994). It is possible that averaging over multiple realizations, each with identical forcing but different initial conditions (see Cubasch et al. 1993) might achieve better results in extracting a near-surface temperature signal from the natural variability noise, even for time scales of only 50 years and for comparatively low G H G forcing (i.e., from 1936-1985). We also computed between-experiment pattern correlations for the (decadal average) temperature change patterns in the final decade of various simulations. Annually-averaged near-surface temperature changes in the years 141-150 of the EIN experiment (definition 4) are highly correlated with changes in years 91-100 of SZA (r=0.79; r*=0.96), but are less correlated with the change pattern in the final decade of the Cubasch et al. (1992) instantaneous COa doubling experiment (r=0.40; r* =0.69). This smaller correlation is caused by the different response pattern in the instantaneous CO2 doubling experiment, which is a direct consequence of the non-linear response behavior of the coupled model to different CO2 increase rates (Hasselmann et al. 1993). In all cases the r* values are higher, since they include the global mean warming component, while r subtracts this and focuses on smaller spatial scales. Additionally, the temperature change pattern of the final decade of the EIN simulation has been correlated with the final decade (years 41-50) of the four MC experiments described in 3.1.1. Correlations range from 0.29 to 0.71 for r and from 0.50 to 0.91 for r*. As mentioned, we may regard the years 51-100 of the EIN simulation as yet another Monte Carlo experiment start- 81 Cubasch et al.: A climate change simulation starting from 1935 ing at still a different initial state (see Fig. 3). If we calculate again the correlation of its final decade (i.e., years 91-100) with the final decade of the EIN simulation (years 141-150), the correlation is higher (r*=0.91; r=0.81) than that achieved when the final decade of any other Monte Carlo simulation is correlated with that of the EIN simulation. The correlation is even higher than the value obtained when the mean of the Monte Carlo realizations is correlated with the final decade of the EIN simulation (which has been shown in Cubasch et al. 1994, to have a stronger correlation with the climate change pattern of SZA than the single realizations). This result is not due to the fact that both patterns have been extracted from the same simulation, since a similar comparison of the change pattern for years 41-50 and 91-100 of the SZA simulation yielded lower correlations. It illustrates rather that the annually-averaged near-surface temperature signal in the EIN experiment is larger and more rapidly discernible than in any of the Monte Carlo simulations. This may be partly attributed to its relatively cool initial state (see Fig. 3, Fig. 5), and partly to the reduced cold start error. 2 1935 1985 2035 2085 time[y] 3.2 Estimates o f detection time Fig. 13. Global mean 5-year-averaged near-surface temperature changes for the EIN experimentrelative to the interannual variability of the control integration. Temperature changes are defined relative to the initial decade of the EIN experiment. The shaded areas denote the 95% confidencelimits for the EIN and CTL changes under the assumptionthat the interannual variability is identicalin the two integrations.The EIN global-meantemperature signal would be detectable in the decade 2015-25 relative to the CTL natural variabilitynoise Figure 13 shows the evolution of the globally averaged temperature changes in the EIN experiment. In order to estimate the latest time it would take to detect the temperature change caused by the increased G H G forcing, one requires some estimate of the natural variability noise, particularly of the low-frequency noise, since the G H G signal evolves on time scales of decades or longer. In the model world, the control run represents the natural variability noise of the coupled atmosphere-ocean system in the absence of external forcing. Unfortunately, reliable estimates of low-frequency noise are difficult to obtain, both from such control integrations and from observed data (see Santer et al. 1995; Hegerl et al. 1994). This is due to the limited duration of the CTL experiment and the observation period (Jones and Briffa 1992). Here, as a first approximation, we use the modelgenerated interannual variability in order to characterize the noise of the CTL integration. It must be stressed that this approach to detection is the simplest possible. The application of a more sophisticated approach which attempts to estimate the low-frequency noise properties of both observed and modelled data is the subject of another paper (Hegerl et al. 1994). The shading enclosing the zero-line in Fig. 13 represents twice the standard deviation of global mean annually-averaged temperature in the control experiment (+ 0.31 K). If we assume that the interannual variability is unchanged in the EIN experiment, we can draw a shaded range of the same size around the EIN globalmean temperature curve. The EIN experiment leaves the shading around the current control climate state in the decade 2015-2025. In the model world this is the time one would be able to separate the global mean climate change signal from the natural variability (at a confidence level of 95%). We have noted previously (Sect. 3.1.2) that there is substantial Arctic cooling during the first decades of the EIN experiments, at least some of which may be attributable to the slow equilibration of Arctic ice volume (see Cubasch et al. 1993). The high latitudes of both hemispheres also display the largest variability in the control simulation (Cubasch et al. 1992; Santer et al. 1994). It is obviously desirable to filter out this highlatitude noise in order to detect a greenhouse warming signal. We therefore excluded these regions by averaging temperatures between 45°N-45°S. In this case the variability of the control simulation is reduced to +0.19K (two standard deviations). The EIN signal now emerges from the shading defining the CTL experiment natural variability noise in the decade 20052015 (Fig. 14). In the context of model signal and noise, therefore, the exclusion of high latitudes from the analysis yields a global warming signal which is detectable one decade earlier than in the case where the mean is computed using full global data. There is potential to achieve still earlier detection times if we use information about the full spatial patterns of near-surface temperature. We first assume that the EOF 1 pattern of the SZA experiment is the dominant climate change signal (as in Cubasch et al. 1992), and then project temperature data from the SZA, EIN and CTL experiments onto this pattern. The resultant principal component (PC) time series are shown in Fig. 15. The amplitude of the SZA EOF 1 pattern is larger in the EIN temperature data than in the SZA data. As 82 Cubasch et al.: A climate change simulation starting from 1935 f , r . . . . . . . . i i i i i i i i i 1935 r . . . . . . . . . i i t i i i I i i i 1985 i . . . . . . . i i r ~ i ' i i i ' r 2035 I i ' i ' i 2085 time[y] Fig. 14. As for Fig. 13, but for near-surface temperature changes averaged between 45°N and 45°S. The exclusion of high-latitude noise leads to an earlier detection time for the EIN signal (in the decade 2005-15) than in the case of global mean temperature 120 120 . . . . . . I . . . . . . ~ I . . . . . . . . ' I/ i 60 E~. ,," We can then use the CTL principal component time series to estimate the natural variability noise envelope, as in the case of global mean temperature and temperature averaged over 45°N-45°S. If the S Z A data are used for the projection, their loading on the first E O F exceeds the climate model noise by about the year 2000. The EIN signal emerges from the noise envelope of the control experiment by approximately the year 1990. Hence the EIN experiment produces a signal which is significant at the 95% level about one decade earlier than the S Z A signal and more than two decades earlier than for the global mean of the E I N experiment. It must be stressed that these detection times are only valid in the context of the model-generated natural variability, which does not incorporate the variance associated with changes in external forcing factors, such as sulfate aerosols, solar variability or volcanoes (see e.g., Hansen et al. 1988; Charlson et al. 1991; FriisChristensen and Lassen 1991; Taylor and Penner 1994). It is also likely that some of the temperature fluctuations in the control run represents residual climate drift rather than bona fide natural variability of the coupled system. The detection procedure which we use here is the simplest possible. It uses interannual variability rather than focusing on longer time scales, which was not feasible due to the relatively short duration of the control run. The study does not represent the time behavior of the G H G signal, e.g., by considering trends, and it does not make any attempt to optimize the signal-to-noise ratio by spatial rotation (Hasselmann 1979; Santer et al. 1995; Hegerl et al. 1994) or by a combined space-time rotation (Hasselmann 1993). A more rigorous approach towards the estimation of signal detection times involves some form of optimization strategy. 0 ~NilNi::iiNiT:iii:: 4 Discussion - 3 0 / , , ~ , , ~ , , , , , i , , ~ , , r . , , , , ~ , , , , , 1935 1950 1965 1980 1995 2010 2025 201,0 2055 2070 2085 Year Fig. 15. Time evolution of the (slightly smoothed) projection of the SZA, EIN and CTL near-surface temperature anomaly data on EOF 1 of SZA. Since the dominant EOF 1 patterns of SZA and the CTL integration are nearly orthogonal, this projection is a simple means of filtering out natural variability noise. S h a d e d areas denote the 95% confidence limits for the EIN and CTL changes, computing using the CTL projection time series. The EIN signal is detectable earlier (in the decade 1990-2000) than if globally-averaged temperature or temperature averaged between 45°N-45°S are used noted by Santer et al. (1994), the E O F 1 modes of the S Z A and CTL experiments are only weakly correlated, and the loading of the CTL data on the dominant S Z A mode is very small. Projection of signal and noise data onto S Z A E O F 1 therefore provides a rather simple way of filtering out much of the natural variability noise of the control run. We have shown that the 1935 starting date of the early industrialization (EIN) experiment alleviates some of the cold start problems which affected the Cubasch et al. (1992) S Z A integration, which commenced in 1985. Both integrations were performed with the same global coupled atmosphere-ocean model and used identical G H G forcing from 1985-2085. The global mean near surface temperature change in 2085 is about 0.3 K (about 10%) higher in the early industrialization experiment than in the S Z A integration. Comparisons between the experiments with early and late start dates show considerable differences in the amplitude of the regional climate change patterns, particularly for sea level. The cold start correction for the S Z A experiment obtained by Hasselmann et al. (1993) using a linear response model is a good approximation for the correction yielded by the EIN experiment for the global mean value. The dominant near-surface temperature signal patterns are highly similar in the two integrations. The common signal pattern emerges despite differences in 83 Cubasch et al.: A climate change simulation starting from 1935 the length of the overall forcing history, the initial conditions at the start of each experiment, and the reference states used to define the t e m p e r a t u r e changes. Regionally, the warming in the final decade of the E I N experiment is almost everywhere larger than the warming in the last decade of the S Z A experiment. T h e regional differences b e t w e e n the E I N and the S Z A simulation in terms of the pattern of sea level change are large, particularly in the Arctic Ocean and between Australia and South America. T h e E I N e x p e r i m e n t predicts a time evolution of the climate change signal which is not inconsistent with observed changes in global m e a n near-surface t e m p e r ature (see H o u g h t o n et al. 1990) i.e., the climate change signal is not clearly discernible during the first 50 years of the integration, which coincides with the observations for the years 1935 to 1990. However, it dominates the latter part of the simulation. W e applied the simplest possible a p p r o a c h in order to estimate the detection time for the E I N near-surface t e m p e r a t u r e signal, using the m o d e l - g e n e r a t e d interannual variability in a 250-year control run in order to define the b a c k g r o u n d noise. It is m o r e appropriate to use the m o d e l low-frequency variability in order to establish the natural variability noise. O u r preliminary signal-to-noise analysis with the interannual noise indicates that the global average change in annual m e a n near-surface t e m p e r a t u r e could be detected during the decade 2015 to 2025. T h e use of a spatial average for the latitude belt 45°N-45°S successfully filters out the large high latitude noise c o m p o n e n t and excludes areas where projected changes are likely to be influenced by p r o b l e m s associated with residual drift of ice volume and the sea-ice p a r a m e t e r i z a t i o n in general (Cubasch et al. 1994). It enables a detection of the near-surface t e m p e r a t u r e signal in the m o d e l world in the decade 2005 to 2015. It has to be stressed that all detection times are valid in the m o d e l world only, which does not include E N S O variability and p r o b a b l y underestimates the long periodic fluctuations (Hegerl et al. 1994). Additionally, the m o d e l does not consider changes in external forcing factors, such as anthropogenic sulfate aerosols, solar variability or volcanic dust and their influence on the systems variability. A further i m p r o v e m e n t in detection time can be obtained by using information about the dominant signal and noise spatial patterns. The projection of data f r o m the control run and the E I N e x p e r i m e n t onto E O F 1 of S Z A is a simple way of filtering out much of the natural variability noise, and yields a detection time around the year 1990. Using this approach, the near-surface t e m p e r a t u r e signal can be detected a b o u t one decade earlier in the E I N experiment than in the S Z A integration, thus illustrating the i m p o r t a n c e of avoiding the cold start. We stress, however, that the E I N experim e n t is only one realization of the evolution of climate change f r o m an early industrialization state, and does not permit analysis of uncertainties caused by imperfect knowledge of the initial conditions. Acknowledgements. We wish to thank K. Hasselmann, L. Bengtsson, and H. yon Storch for their advice and helpful discussions, J. Mitchell and G. Meehl for their helpful review. M. Zorita extended the control simulation up to 250 years, while P. Lenzen and R. Gaedtke provided data handling support. All diagrams were expertly drawn by N. Noreiks and M. Grunert. This research has been supported by the German Ministry for Research and Technology (BMFT), the Max-Planck-Gesellschaft, the Freie und Hansestadt Hamburg and the European Community Environmental program. One of the authors (B. D. Santer) was supported by the US Department of Energy, Environmental Sciences Division, under contract W-7405-ENG-48 with Lawrence Livermore National Laboratory. The authors are grateful to the staff of the DKRZ for their excellent technical assistance. References Bakan S, Chlond A, Cubasch U, Feichter J, Graf HF, Grassl H, Hasselmann K, Kirchner I, Latif M, Roeckner E, Sausen R, Schlese U, Schriever D, Schult I, Schumann U, Sielmann F, Welke W (1991) Climate response to smoke from the burning oil wells in Kuwait. Nature 351 : 367-371 Charlson RJ, Langer J, Rohde H, Leovy CB, Warren SG (1991) Perturbation of the Northern Hemisphere radiative balance by backscattering from anthropogenic sulfate aerosols. Tellus 43ab: 152-i63 Cubasch U, Hasselmann K, HOck H, Maier-Reimer E, Mikolajewicz U, Santer BD, Sausen R (1992) Time-dependent greenhouse warming computations with a coupled ocean-atmosphere model. Clim Dyn 8:55-69 Cubasch U, Santer BD, Hellbach A, Hegerl GC, HOck H, MaierReimer E, Mikolajewicz U, St6ssel A, Voss R (1994) Monte Carlo climate change forecasts with a global coupled oceanatmosphere model. Clim Dyn 10:1-19 Fichefet T, Tricot C (1992) Influence of the starting date of model integration on projections of greenhouse-gas-induced climatic change. Geophys Res Lctt 19:1771-1774 Friis-Christensen E, Lassen K (1991) Length of the solar cycle: an indicator of soalar activity closely associated with climate. Science 254: 698-700 Hansen J, Fung I, Lacis A, Rind D, Lebedeff S, Ruedy R, Russel G (1988) Global climate changes as forecast by GISS 3-d model. J Geophys Res 93:9341-9364 Hasselmann K, Sausen R, Maier-Reimer E, Voss R (1993) On the cold start problem with coupled atmosphere-ocean models. Clim Dyn 9: 53-61 Hegerl GC, von Storch H, Hasselmann K, Santer BD, Cubasch U, Jones PD (1994) Detecting anthropogenic climate change with an optimal fingerprint method. Rep. No. 142, MPI for Meteorologie, Hamburg, Germany Houghton JT, Jenkins GJ, Ephraums JJ (eds) (1990) Climate change. The IPCC scientific assessment. Cambridge University Press, Cambridge, UK Houghton JT, Callander BA, Varney SK (eds) (1992) Climate change 1992. The supplementary report to the IPCC scientific assessment. Cambridge University Press, Cambridge, UK Jones PD, Briffa KR (1992) Global surface air temperature variations during the twentieth century: Part 1, spatial, temporal and seasonal details. The Holocene 2, 2:165-179 Maier-Reimer E, Mikolajewicz U, Hasselmann K (1993) Mean circulation of the Hamburg LSG OGCM and its sensitivity of the thermohaline surface forcing. J Phys Oceanogr 23:731757 Manabe S, Stouffer RJ, Spelman MJ, Bryan K (1991) Transient responses of a coupled ocean-atmosphere model for gradual changes of atmospheric CO2. Part I: annual mean response. J Clim 4:785-818 Manabe S, Spelman MJ, Stouffer RJ (1992) Transient responses of a coupled ocean-atmosphere model for gradual changes of 84 atmospheric CO2. Part II: seasonal response. J Clim 5:105126 Meehl GA, Washington WM (1990) CO2 climate sensitivity and snow-sea-ice albedo parameterization in an atmospheric GCM coupled to a mixed layer ocean model. Clim Change 16: 283-306 Meehl GA, Washington WM, Karl TR (1993) Low-frequency variability and CO2 transient climate change. Part I: time-averaged differences. Clim Dyn 8:117-133 Mikolajewicz U, Maier-Reimer E (1990) Internal secular variability in an ocean circulation model. Clim Dyn 4:145-156 Mikolajewicz U, Santer BD, Maier-Reimer E (1990) Ocean response to greenhouse warming. Nature 354:589-593 Santer BD, Brtiggemann W, Cubasch U, Hasselmann K, HOck H, Maier-Reimer E, Mikolajewicz U (1994) Signal-to-noise anal- View publication stats Cubasch et al.: A climate change simulation starting from 1935 ysis of time-dependent greenhouse warming experiments. Part 1: pattern analysis. Clim Dyn 9:267-285 Santer BD, Mikolajewicz U, Brtlggemann W, Cubasch U, Hasselmann K, HOck H, Maier-Reimer E, Wigley TML (1995) Ocean variability and its influence on the detectability of greenhouse warming signals accepted by JGR (ocean) Santer BD, Cubasch U, Hegerl GC, Mikolajewicz U (1993b) The use of general circulation models in detecting climate change induced by greenhouse gases. PCMDI Report 10, Lawrence Livermore National Laboratory, Livermore, CA 94550 Stouffer RJ, Manabe S, Vinnikov K Ya (1994) Model assessment of the role of natural variability in recent global warming. Nature 367: 634-636 Taylor KE, Penner JE (1994) Anthropogenic aerosols and climate change. Nature 369:734-737