On the Importance of Paleoclimate Modelling for Improving Predictions of Future Climate Change

Home , Climate model

Clim. Past, 5, 803–814, 2009 www.clim-past.net/5/803/2009/ Climate © Author(s) 2009. This work is distributed under of the Past the Creative Commons Attribution 3.0 License.

On the importance of paleoclimate modelling for improving predictions of future climate change

J. C. Hargreaves and J. D. Annan Global Change Projection Research Program, Research Institute for Global Change, JAMSTEC, 3173-25 Showa-machi, Kanazawa-ku, Yokohama City, Kanagawa, 236-0001, Japan Received: 21 July 2009 – Published in Clim. Past Discuss.: 29 July 2009 Revised: 23 October 2009 – Accepted: 1 December 2009 – Published: 21 December 2009

Abstract. We use an ensemble of runs from the MIROC3.2 in the past when the climate was rather different to today, AGCM with slab-ocean to explore the extent to which mid- because a model that does a good job of modelling paleocli- Holocene simulations are relevant to predictions of future climates may reasonably be considered preferable to one that mate change. The results are compared with similar analy- does not. However, this hypothesis is so far largely untested. ses for the Last Glacial Maximum (LGM) and pre-industrial Similar processes do not necessarily govern all past and fu- control climate. We suggest that the paleoclimate epochs can ture climate changes, so it is arguable that what we really provide some independent validation of the models that is should seek are those climate changes in the past that are in also relevant for future predictions. Considering the paleo- some way analogous to or informative regarding the future climate epochs, we find that the stronger global forcing and changes we expect to see. hence larger climate change at the LGM makes this likely As observational evidence tends to become more sparse to be the more powerful one for estimating the large-scale and uncertain at more distant times, much effort has been changes that are anticipated due to anthropogenic forcing. focussed on more recent paleo-climates such as the mid- The phenomena in the mid-Holocene simulations which are Holocene (6 ka BP) and Last Glacial Maximum (LGM, most strongly correlated with future changes (i.e., the mid 21 ka BP). Model simulations of these epochs have formed to high northern latitude land temperature and monsoon pre- the centrepiece of Paleoclimate Modelling Inter-comparison cipitation) do, however, coincide with areas where the LGM Projects PMIP and PMIP2 (Joussaume et al., 1999; Bracon- results are not correlated with future changes, and these are not et al., 2007a). However, there is relatively little research also areas where the paleodata indicate significant climate directly addressing the extent to which these paleoclimate changes have occurred. Thus, these regions and phenomena epochs are informative or the processes governing the cli- for the mid-Holocene may be useful for model improvement mate changes analogous to those effecting future change. and validation. For example, one might question how essential it is to ac- curately simulate the response and climate feedbacks due to the huge ice sheets present over large parts of the Northern Hemisphere at the LGM, since for the modern climate and 1 Introduction in coming decades, we expect at most small changes in the Model predictions of long term climate change cannot be val- amount of terrestrial ice. However, at the LGM, there was idated through repeated forecast/analysis cycles in the same also a significant decrease in the the forcing from greenhouse manner as weather forecasts, as the necessary time scale is gases (primarily CO2) compared to present day levels, and too long. Models are routinely evaluated in terms of how we may expect that the response to this forcing involves sim- well they represent current climate, although this only ap- ilar processes to those relevant to future change. pears to provide rather limited evidence for their future per- Annan et al. (2005) and Hargreaves et al. (2007) inves- formance (Whetton et al., 2007; Abe et al., 2009). Re- tigated an ensemble of the MIROC3.2 AGCM (coupled to searchers have, therefore, been motivated to look for times a slab ocean) in which parameters were constrained by ob- servations of the current climate state. While such an approach can investigate the range of responses arising from Correspondence to: J. C. Hargreaves parametric uncertainties, an acknowledged limitation is that (jules@jamstec.go.jp) the use of a single model precludes exploration of structural

Published by Copernicus Publications on behalf of the European Geosciences Union. 804 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling limitations of the model. They found a reasonably strong cor- 760 ppm to 650 ppm) and a change in the orbital forcing cal- relation between the climate changes for the LGM and dou- culated from Berger (1978). The change to the orbital forc- bled CO2, especially away from the large Northern Hemi- ing affects the seasonal cycle and also changes the lengths sphere ice sheets. Using a simpler atmospheric model (but of the seasons by up to 5 days (Joussaume and Braconnot, a three-dimensional ocean), Schneider von Deimling et al. 1997). The use of the classical calendar for both experi- (2006) obtained broadly comparable results. However, Cru- ments results in a mismatch between monthly results when cifix (2006) examined results from 4 distinct AOGCMs and compared to a true solar calendar. However, the analysis dis- found no evidence of a relationship between the climate cusssed in this paper pertains to assessing correlations be- changes seen at the LGM and 2×CO2 states, although with tween variables over the ensemble, and we expect the effect such a small ensemble, statistical significance would be hard of this calendar inaccuracy on our results to be small since it to achieve. Thus, the results obtained here must be consid- affects all ensemble members equally. The PMIP2 protocol ered in the context of the single model that underlies the for the LGM consists of a significant decrease in the levels experiments. Since multi-model experiments are obviously of greenhouse gases compared to the modern climate, the in- outside our capabilities, we hope that our results will encour- clusion of fixed massive ice sheets over the Northern Hemi- age other groups to attempt similar investigations. sphere and a small change in the orbital forcing (see Bracon- In this paper we explore the climate changes for the mid- not et al., 2007a for details). We also performed the PMIP2 Holocene to see to what extent modelling that epoch may pre-industrial control (CTRL) and a future climate experi- help improve forecasts of future climate, and compare our ment which has an increased carbon dioxide concentration results to similar analysis of LGM simulations. Although but is otherwise identical to the pre-industrial. paleoclimate studies are gaining in popularity, most comparison with data for GCMs is made against the present day climatology. Therefore, to further set this work in context 3 Method we also compare the results to similar analyses for the pre- industrial control simulation. The mid-Holocene might ap- For our experiments we use a 36 member ensemble of runs pear at first to be a rather weak analogue for future climate of the MIROC3.2 AGCM with slab ocean, at T21 resolu- change since at this epoch there are no large changes in the tion. The ensemble generation method closely follows An- net forcing or greenhouse gas levels. Moreover, the over- nan et al. (2005), in which an ensemble Kalman filter was all climate changes at the mid-Holocene are rather different used for multivariate parameter estimation by tuning the in character compared to those expected to occur in the fu- model to various fields of modern climatological data, pro- ture (Mitchell, 1990). Thus, we do not expect the climate ducing a 40 member ensemble. However, in this experi- changes to be a close analogue in the sense of being di- ment we only allowed 13 model parameters to vary (selecting rectly applicable to future projections. Rather, we are look- those which had been found to most strongly influence model ing for climate phenomena that respond to the same uncer- results). We use a slab ocean model to reduce the equilibra- tain inputs, so that, as Mitchell (1990) proposes, analysis of tion time of the model, as is usual for such perturbed param- paleoclimate changes might be useful for the estimation of eter ensemble experiments (e.g. Murphy et al., 2004). This model parameters that also control future changes. Paleo- requires the calculation of implied heat fluxes at each grid data from the mid-Holocene do indicate some significant box (the so-called Q-flux field) to mimic the lateral trans- changes which may be relevant to future climate change. The port of heat (Russell et al., 1985). Although in principle this most prominent example is the evidence that the monsoon re- flux field should integrate to zero for an equilibrium state, gions extended further northward than today, a result shown if no particular care is taken, the Q-flux field may have a most dramatically by evidence for greening of parts of the large nonzero integral, indicating that the model is far from present day Sahara desert (Jolly et al., 1998). Current pre- radiative balance under modern conditions. We found that dictions for the future changes of the monsoon under global in our previous work, there was a substantial radiative im- change are highly uncertain, with disagreement in the mod- balance of around 10Wm−2 both for the ensemble members, els over the sign of the precipitation change (e.g. Fig. 10.9, and also for the unperturbed model when run at T21 resolu- Solomon et al., 2007). Therefore in this paper we analyse tion. Therefore, in this experiment, we imposed an additional monsoon changes as well as zonal and global changes. constraint on the global and annual average of the Q-flux to reduce the radiative imbalance to around 2Wm−2, similar to or lower than the results of Collins et al. (2006). Further de- 2 Boundary conditions for mid-Holocene and LGM tails of the ensemble are described in Yokohata et al. (2009). climate simulations All the runs are at least 40 years long, with the last 20 years averaged to provide the results. The imposition of the Q-flux The forcing for the mid-Holocene in the PMIP2 proto- constraint resulted in an increase in the climate sensitivity of col (Braconnot et al., 2007a) consists of a small change to the the ensemble, to 7.0±3.7 ◦C (2 standard deviations). For the CH4 concentration relative to the pre-industrial level (from doubled CO2 experiment a number of the ensemble members

Clim. Past, 5, 803–814, 2009 www.clim-past.net/5/803/2009/ J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 805 with higher control temperature and high sensitivity became Another important issue is that of the model inadequacy, unstable once the global mean surface temperature exceeded that is the fact that the model cannot simulate reality per- ∼296 K. According to the IPCC AR4, climate sensitivity is fectly for any set of parameter values. Moreover, different likely to lie in the range 2–4.5 ◦C (IPCC 2007: summary for model ensembles may well exhibit somewhat different rela- Policymakers, Solomon et al., 2007). In this experiment we tionships between past and future climate, depending on their are, therefore, exploring the climate behaviour for the un- structure and parameterisations. The relationship is a prop- likely higher values of climate sensitivity. Given the interest erty of the model and experimental details, rather than the in the possible consequences of such high sensitivities, in or- climate system itself (which has one past and future trajec- der to obtain a stable ensemble which could reasonably be tory). Therefore, while the existence of relationship between √used for analysis, we performed another experiment using past and future climate (within a model ensemble) is a pre- 2 times pre-industrial CO2 levels which reduces the equi- requisite in order for paleoclimate tuning to affect the future librium warming by 50% (for those models which do equi- predictions, this may not be sufficient by itself to assure that librate). Apart from one ensemble member which crashed the tuning will actually lead to improvements. We cannot early on, and is also pathological for the LGM climate, all address this question within this paper, but hope that the is- the other 39 ensemble members remained stable for a long sues relating to this question can be further explored using run (>150 years)√ with this forcing. Here we use the results ensembles of different models (e.g. Yokohata et al., 2009). from the larger 2×CO2 ensemble but double the climate changes to ease comprehension of the values in the context of other literature, and refer hereafter to the result as 2×CO2. 4 Results For the LGM, one other ensemble member produced a state about 6 ◦C colder than the ensemble mean so this is consid- 4.1 Analysis of annual averages on the global scale ered to not represent a reasonable LGM and is excluded from the ensemble. Such abnormal cooling can be produced by a The difference between the control and mid-Holocene an- physically unrealistic cloud or sea-ice feedback and is a well- nually and globally averaged global 2 m temperatures (here- known limitation of slab ocean models. The final 36 member after, T2) is rather small (0.3±0.2 ◦C at 1 standard deviation), ensemble is produced by further excluding two more ensem- reflecting the small change in the mean climate forcing. The ble members which produced similarly very cold states for a correlation coefficient between the temperature change and LGMGHG simulation (pre-industrial conditions with LGM climate sensivity is less than 0.1. These results are not unex- greenhouse gases), that is not discussed in this paper, but is pected given the net forcing change is a small proportion of the subject of other analyses currently underway. the seasonal and regional forcing. For the LGM simulations, The underlying basis for our work is the expectation MIROC shows a fairly strong correlation coefficient of 0.74 that our ensemble of models with varying parameters rep- between the change in T2 from control to LGM, and climate resents at least a large subset of our uncertainties concern- sensitivity, consistent with but slightly higher than in previ- ing the physical processes controlling past and future cli- ous work (Annan et al., 2005). These results are illustrated mate change. We note that in support of this claim, the fu- in Fig. 1. For reference, the level of statistical significance at ture projections for monsoon changes in this model ensem- 1% for a sample size of 36, assuming independent samples, ble (discussed in more detail later) include both increases is 0.42 according to Student’s t-test. While in our results we and decreases in precipitation, similar to that exhibited by have used the 1% significance level as a guide to the signif- the IPCC ensemble of opportunity mentioned above. The icance of our results, it should be noted that in this work we first-order relationship between past and future data can be have considered over 1000 different correlation coefficients, expressed as their covariances, which describes how the un- meaning that a number of false positives are to be expected in certainties in the future changes are related to uncertainties the results. A larger model ensemble, which is not at present in past changes. If the covariance is low, then information computationally feasible, would improve the statistical con- about the past will not influence future predictions, but if the fidence of our results. covariance is high, then this implies that information about For clarity and brevity, the three experiment-control differ- the past will propagate into predictions. Therefore, we now ences for the 2×CO2, LGM, and mid-Holocene experiments explore our ensemble results to investigate where such rela- are abbreviated to D2×CO2, DLGM and D6ka, respectively. tionships may exist. This approach is similar to the general Table 1 shows, for T2 and precipitation, the correlation co- principles underlying Observing System Simulation Exper- efficients between D2×CO2 and the other experiments. For iments (Arnold Jr. and Dey, 1986), by which the value of T2, the control is significantly correlated with D2×CO2, but observational data for prediction systems can be considered. the DLGM shows a much stronger correlation. Note also However in this work here do not quantify the likely bene- that the control itself is also correlated with DLGM, indi- fits, which also depends on the precision of the observational cating that the control and LGM may provide only partially evidence that might be available. independent constraints on modelled climate sensitivity. We illustrate this through regression models which predict the www.clim-past.net/5/803/2009/ Clim. Past, 5, 803–814, 2009 4 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling

806 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling

red: (CTRL-LGM) low QFLUX ensemble vs. Climate sensitivity Table 1. Correlation between the global temperature and precipita- blue: (CTRL-LGM) 120 member ensemble vs. Climate sensitivity tion for the various experiments. magenta: (6ka-CTRL) low QFLUX ensemble vs. Climate sensitivity

correlation coe s red: 0.74 D2×CO2 CTRL D6ka DLGM blue: 0.60 T2 ppt T2 ppt T2 ppt T2 ppt magenta: 0.09 D2×CO2 T2 X 0.98 0.45 −0.71 0.09 0.12 −0.74 −0.54 D2×CO2 ppt – X 0.51 −0.63 0.09 0.14 −0.75 −0.62 CTRL T2 – – X – – – −0.30 –

C) CTRL ppt – – – X – – – 0.04 o T2 (

4.2 Analysis of annual averages on the zonal scale

One of the potential difficulties of using paleo-climate data to validate climate models is the mismatch in spatial scales between the GCMs, which are most reliable at the largest o Climate Sensitivity ( C) scales, and less so at smaller scales, versus paleoclimate data, which is generally highly local in nature. So, while the global Fig. 1. Correlation between the temperature changes for 2×CO2 calculations presented above indicate that there are links be- and the two experiments, LGM and mid-Holocene. The blue dots tween the processes effecting climate changes at the LGM show the results from previous work on the LGM, indicating that the and 2×CO2 climates, they do not shed any light on which Figure 1. Correlation between the temperatureresults for changes the new for low 2× Q-fluxCO2 and ensemble the two are experiments, consistent, LGM although andmid-Holocene. of The blue dots show the results from previous work on theslightly LGM, indicating higher sensitivity. that the results The contrast for the new between low Q-flux the strong ensemble signal are consistent,regions although may be of effectively slightly used to validate and improve the higher sensitivity. The contrast betweenfor the the strong LGM signal and very for smallthe LGM global and change very small for the global mid-Holocene change for the is mid-Holocenemodelled is predictions clear. of future climate change. In addition, clear. the changes in climate at the mid-Holocene, being caused by changes in the patterns of seasonal forcing, are expected to Table 1. Correlation between the global temperature and precipitation for the various experiments. be far greater on zonal to regional scales than at the global global D2×CO2 temperature change using the DLGM and average. The albedo forcing due to ice sheets at the LGM control temperaturesD2×CO2 (singlyCTRL or jointly) D6ka as predictors. DLGM The is also strongly linked to latitude. Thus we consider the T2 ppt T2 ppt T2 ppt T2 ppt control value alone explains 21% of the total variance, with relationship in the ensemble between the zonally averaged D2×COthe2 T2 LGM X explaining 0.98 0.45 55%− and0.71 both 0.09 control 0.12 and−0 LGM.74 − to-0.54 variables for D2×CO2, DLGM, D6ka, CTRL and the global × − − − D2 COgether2 ppt combining X X to 0.51explain0 61%..63 0.09 0.14 0.75 0.62 scale changes for 2×CO2. The mid-Holocene results pro- − CTRL T2For D2 X×CO X2, there X is an– extremely – high – correlation0.30 be- – vide a very small contribution, explaining less than 1% of CTRLtween ppt temperature X X and X precipitation, X – so it – is not – surpris- 0.04 the variance. The zonally averaged annual average temper- ing to also see significant correlation coefficients for the atures for CTRL, D2×CO2, D6ka and DLGM are shown in correlation of control and DLGM temperatures with pre- Fig. 2. The heavier lines are the ensemble mean, whereas the cipitation for D2×CO2. Although the correlations indi- lighter weight lines show the one standard deviation widths ing to also see significant correlation coefficients for the control and DLGM results each explain almost 40% of the cate that a higher control temperature leads to increased cli- of the ensembles. The results from the correlation analysis correlation of control and DLGM temperatures with pre- total variance of the D2×CO precipitation change, and in mate sensitivity, and increased global precipitation change2 are shown in Fig. 3. The red lines show the correlation of cipitation for D2×CO . Although the correlations indicate combination they explain 75% of the total variance. Again, 2 for D2×CO , higher control precipitation leads to a smaller that a higher control temperature leads to2 increased cli- the mid-Holocene makes a negligiblezonal contribution. and global temperature, and the blue show the correla- precipitation change for increased CO . In contrast, for the mate sensitivity, and increased global precipitation change 2 tion of zonal and global precipitation. As paleoclimate data LGM, the global precipitation changes are negative, and the for land and ocean is often analysed and synthesised indepen- for D2×CO2, higher control precipitation leads to a smaller larger the change in precipitation4.2 the Analysis greater of the annual precipita- averages ondently, the zonal the correlation scale analysis is done for the land and ocean precipitation change for increased CO2. In contrast, for the LGM, the global precipitation changestion change are negative, for increased and the CO2One. The of the negligible potential correlation difficulties of usingseparately paleo-climate (solid and data dashed lines, respectively). As shown larger the change in precipitationbetween the greater the the control precipita- and DLGMto validate precipitation climate suggests models that is the mismatchin Fig. 3 ina, spatial for D2 scales×CO2 the temperature is strongly corre- these aspects of the LGM and control climates may be de- tion change for increased CO2. The negligible correlation between the GCMs, which are mostlated reliable with climate at the largest sensitivity over all latitudes. For precipita- between the control and DLGMtermined precipitation by (and suggests therefore that usefulscales, for constraining)and less so at smaller different, scales, versustion, although paleoclimate the correlation data, is significant for most latitudes, these aspects of the LGM and controlindependent climates processes may be de- in the model,which is and generally thus the highly informa- local in nature.it is mostlySo, while not the as global strongly correlated with global precipita- termined by (and therefore usefultion for constraining) from both epochs different, may combinecalculations effectively. presented A above similar indicatetion that change, there are particularly links be- over the land. The CTRL (Fig. 3b) independent processes in the model,regression and thus analysis the informa- to the previoustween paragraph the processes reveals eff thatecting the climatetemperatures changes at are the most LGM strongly correlated with climate sen- tion from both epochs may combinecontrol effectively. and DLGM A similar results eachand explain 2×CO2 almostclimates, 40% they of the do not shedsitivity any over light the on land which in the Northern Hemisphere, while the regression analysis to the previoustotal paragraph variance reveals of the that D2 the×CO2regionsprecipitation may be change, effectively and used in to validatenegative and correlation improve for the precipitation, aluded to above, arises combination they explain 75%modelled of the total predictions variance. of Again, future climatefrom change. the low In to addition, mid-latitude ocean and the land around 40◦ the mid-Holocene makes a negligible contribution. north and south. Journalname www.jn.net

Clim. Past, 5, 803–814, 2009 www.clim-past.net/5/803/2009/ J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 5 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 807 the changes in climate at the mid-Holocene, being caused by changes in the patterns of seasonal forcing, are expected to be tion in the southern ocean at the LGM, the mid-to-high lat- far greater on zonal to regional scales than at the global aver- itude temperatures for the Northern Hemisphere at the mid- Holocene and the whole of the Northern Hemisphere tem- age. The albedo forcing due to ice sheets at the LGM is also peratures and tropical precipitation for the pre-industrial cli- strongly linked to latitude. Thus we consider the relationship mate. Given the very high correlation between the global in the ensemble between the zonally averaged variables for temperature and precipitation for D2×CO2, the correlations C D2×CO2, DLGM, D6ka, CTRL and the global scale changes o for zonal temperature with global precipitation (not shown) look almost identical to the red lines in Fig. 3, and indi- for 2×CO2. The zonally averaged annual average temperatures for cates that improving the temperature should also result in Control Tempertaure Tempertaure improved precipitation. CTRL, D2×CO2, D6ka and DLGM are shown in Fig. 2. D2xCO2 D6ka Correctly predicting globally averaged climate change, The heavier lines are the ensemble mean, whereas the lighter DLGM while still an important achievement scientifically, may not weight lines show the one standard deviation widths of the be as useful as knowing what the climate change will be ensembles. The results from the correlation analysis are in a particular region. In order to consider future changes shown in Fig. 3. The red lines show the correlation of zonal on a somewhat smaller scale, Fig. 4 shows the correlation and global temperature, and the blue show the correlation of of the zonally averaged D2×CO2 with CTRL, D6ka and zonal and global precipitation. As paleoclimate data for land DLGM. For temperature, the curves look rather similar to those in Fig. 3. This comes as no surprise, given the strong and ocean is often analysed and synthesised independently, zonal-global correlation of temperature at D2×CO2 shown the correlation analysis is done for the land and ocean sep- in Fig. 3a. For precipitation, the results are, however, some- arately (solid and dashed lines, respectively). As shown in what different to those in Fig. 3. The clearest correlation is Fig. 3a, for D2×CO2 the temperature is strongly correlated the southern ocean region for DLGM. Apart from this, the with climate sensitivity over all latitudes. For precipitation, correlation coefficient rises above the nominal noise level at although the correlation is significant for most latitudes, it Precipitation mm/dy various latitudes for all three experiments, mostly in the tropics. The correlation between zonal temperature for the exper- is mostly not as strongly correlated with global precipitation iments and zonal precipitation for D2×CO2 is not shown, but change, particularly over the land. The CTRL (Fig. 3b) tem- is similar in shape to what would result from the multiplica- peratures are most strongly correlated with climate sensitiv- tion of the blue lines in Fig. 3a with the red lines in Fig. 4. In ity over the land in the Northern Hemisphere, while the nega- latitude(o) other words, the correlation tends to decrease in significance tive correlation for precipitation, aluded to above, arises from in those regions where the blue lines in Fig. 3a lie within the low to mid-latitude ocean and the land around 40◦ north Fig. 2. Top: annually averaged temperature for the control climate, the magenta band. So, for improving climate change pre- and the differences between the control and the simulated climates. diction on the zonal scale the temperature may be similarly and south. FigureThe thinner 2. Top: lines annually show the 1 averaged standard deviationtemperature ranges for of the the control en- cli-constrained on the global and regional scales, with the LGM For D6ka (Fig. 3c) the correlation coefficients are mostly mate,semble. and Bottom: the diff theerences same as between the top plot, the butcontrol for precipitation. and the simulated cli-being useful for the tropics and the high latitudes, and the below the level of the noise, except for part of the North- mates. The thinner lines show the 1 standard deviation ranges of thecontrol and mid-Holocene contributing in the northern hemi- ern Hemisphere where the correlation is reasonably strong ensemble. Bottom: The same as the top plot, but for precipitation.sphere, while for precipitation all three experiments may pro- over the land in particular. Interestingly, this is the same lat- For D6ka (Fig. 3c) the correlation coefficients are mostly vide some useful information, but none is particularly domi- itude range for which the correlation coefficient for DLGM below the level of the noise, except for part of the North- nant. lookern Hemisphere almost identical where the to correlation the red lines is reasonably in Fig. 3,strong and indi-Since there is considerable overlap with the results for the (Fig. 3d) temperature is rather weak, due to the existance of over the land in particular. Interestingly, this is the same lat- cates that improving the temperature should also resultcontrol in with those for DLGM and D6ka, one may wonder the large ice sheets (which was previously discussed in Har- itude range for which the correlation coefficient for DLGM if all we really need in order to build better models is to greaves et al., 2007). In addition, for DLGM, as we noted in improved(Fig. 3d) temperature precipitation. is rather weak, due to the existance of better fit them better to the modern climate. Figure 5 in- previous work (Hargreaves et al., 2007), although the corre- theCorrectly large ice sheets predicting (which globally was previously averaged discussed climate in Har- change,vestigates this question for MIROC. It shows the correla- lation for temperature is generally very strong, small biases whilegreaves still et al. an, 2007 important). In addition, achievement for DLGM, scientifically, as we noted in may nottion of DLGM and D6ka with CTRL for temperature. For in the control sea-ice extent may strongly influence the tem- beprevious as useful work as (Hargreaves knowing et what al., 2007 the), climate although changethe corre- willD6ka, be there is no significant correlation in the land over the lation for temperature is generally very strong, small biases Northern Hemisphere region where the significant correla- perature in the sea-ice region for increasing CO2, leading to in a particular region. In order to consider future changes in the control sea-ice extent may strongly influence the tem- tion is found in Fig. 4b. Similarly, for DLGM the correla- no significant correlation for temperature in the sea-ice re- on a somewhat smaller scale, Fig. 4 shows the correlation perature in the sea-ice region for increasing CO2, leading to tion is weak for the ocean and only marginally significant gions for DLGM. The situation for precipitation is, however, ofno the significant zonally correlation averaged for D2 temperature×CO2 with in the CTRL, sea-ice D6ka re- andfor small regions in the tropics. These results indicate that reversed with the correlation coefficient being rather strong DLGM.gions for For DLGM. temperature, The situation the for curves precipitation look is, rather however, similarthe to paleo-climate experiments may be exercising different for DLGM over the ocean from 40–70◦ S. thosereversed in Fig. with 3. the This correlation comes coefficient as no surprise, being rather given strong the strongaspects relevant for climate change, than the control exper- for DLGM over the ocean from 40–70◦ S. From these results it would appear that the best data to use zonal-global correlation of temperature at D2×CO2 showniment. This suggests that improving the modelling of the to improve the model predictions would be the temperatures in Fig.From 3a. these For results precipitation, it would appear the results that the are,best data however, to use some-paleoclimates, over and above the present day climate, can in the tropics and very high latitudes along with precipita- whatto improve different the model to those predictions in Fig. would 3. The be clearest the temperatures correlationbe is expected to further improve future predictions. in the tropics and very high latitudes along with precipitation in the southern ocean at the LGM, the mid-to-high lat- the southern ocean region for DLGM. Apart from this, the itude temperatures for the Northern Hemisphere at the mid- correlation coefficient rises above the nominal noise level at Holocene and the whole of the Northern Hemisphere tem- variouswww.clim-past.net/5/803/2009/ latitudes for all three experiments, mostly in the trop- Clim. Past, 5, 803–814, 2009 peratures and tropical precipitation for the pre-industrial cli- ics. The correlation between zonal temperature for the exper- mate. Given the very high correlation between the global iments and zonal precipitation for D2×CO2 is not shown, but temperature and precipitation for D2×CO2, the correlations is similar in shape to what would result from the multiplica- for zonal temperature with global precipitation (not shown) tion of the blue lines in Fig. 3a with the red lines in Fig. 4. In www.jn.net Journalname 6 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling

808 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling

T2 land T2 ocean PPT land PPT ocean (a) D2xCO2(zonal) correlated with D2xCO2(global) Correlation coefficient

(b) CTRL(zonal) correlated with D2xCO2(global) (a) CTRL(zonal) correlated with D2xCO2(zonal) Correlation coefficient Correlation coefficient

(c) D6ka(zonal) correlated with D2xCO2(global) (b) D6ka(zonal) correlated with D2xCO2(zonal) Correlation coefficient Correlation coefficient

T2 land (d) DLGM(zonal) correlated with D2xCO2(global) (c) DLGM(zonal) correlated with D2xCO2(zonal) T2 ocean PPT land PPT ocean Correlation coefficient Correlation coefficient

o latitude( ) latitude(o)

Fig. 3. Correlation of the annually averaged global changes for the 2×CO2 experiment with the annually averaged zonal changes for all the experiments, for both precipitation and 2 m temperature. The results are split into separate results for land and ocean. The magenta band shows the region which does not achieve signiﬁcance at the 99% level.

4.3 Analysis ofFigure seasonal averages 3. Correlation of the annuallyAnalysis averaged on the zonal global scale presents changes a more for interestingFigure 4. Correlation of the annually averaged zonal changes for picture. The dashed line in the bottom plot of Fig. 6 indi- the 2×CO2 experiment with the annually averaged zonal changes the 2×CO2 experiment with the annually averaged zonal changes Tsushima and Manabe (2001) suggested that the global re- cates the zonal variation of the correlation between the con- sponse of the climatefor all system the to experiments, seasonal variation for may both be precipitationtrol seasonal signal and (JJA-DJF 2 m temperature temperature. over land), and cli-for all the experiments, for both precipitation and 2 m temperature. mate sensitivity for our model ensemble. The correlation is analagous to theThe global results changes are that splitoccur under into global separate results for land◦ and◦ ocean.◦ The ◦ The results are split into separate results for land and ocean. The warming. Using this assumption, they argued that since the stronger around 25–30 N, 70 N, 70 S and 25–30 S. In ad- cloud feedback effectmagenta was small band for the shows annual the global region tem- whichdition these does same not parts achieve of the globe significance also show the strongestmagenta band shows the region which does not achieve significance × perature variation,at it the could 99% also be level, small for according the case of an- to Student’scorrelation t-test. for D2 CO2 with climate sensitivity. It would,at the 99% level, according to Student’s t-test. thropogenic global warming. In our ensemble, however, we therefore, appear that the MIROC results are similar in char- do not find a significant correlation between the present-day acter to those of HadSM3 in Knutti et al. (2006), who also global seasonal temperature signal and climate sensitivity. found that many Northern Hemisphere extratropical land sphere region where the significant correlation is found in Clim. Past, 5, 803–other814, 2009 words, the correlation tends to decrease in www.clim-past.net/5/803/2009/ significance Fig. 4b. Similarly, for DLGM the correlation is weak for in those regions where the blue lines in Fig. 3a lie within the ocean and only marginally significant for small regions the magenta band. So, for improving climate change pre- in the tropics. These results indicate that the paleo-climate diction on the zonal scale the temperature may be similarly experiments may be exercising different aspects relevant for constrained on the global and regional scales, with the LGM climate change, than the control experiment. This suggests being useful for the tropics and the high latitudes, and the that improving the modelling of the paleoclimates, over and control and mid-Holocene contributing in the northern hemi- above the present day climate, can be expected to further im- sphere, while for precipitation all three experiments may pro- prove future predictions. vide some useful information, but none is particularly domi- nant. 4.3 Analysis of seasonal averages Since there is considerable overlap with the results for the control with those for DLGM and D6ka, one may wonder if Tsushima and Manabe (2001) suggested that the global re- all we really need in order to build better models is to better sponse of the climate system to seasonal variation may be fit them better to the modern climate. Figure 5 investigates analagous to the global changes that occur under global this question for MIROC. It shows the correlation of DLGM warming. Using this assumption, they argued that since the and D6ka with CTRL for temperature. For D6ka, there is no cloud feedback effect was small for the annual global tem- significant correlation in the land over the Northern Hemi- perature variation, it could also be small for the case of an-

Journalname www.jn.net 6 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling

T2 land T2 ocean PPT land PPT ocean J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 809 (a) D2xCO2(zonal) correlated with D2xCO2(global) Correlation coefficient

(b) CTRL(zonal) correlated with D2xCO2(global) (a) CTRL(zonal) correlated with D2xCO2(zonal) Correlation coefficient Correlation coefficient

(c) D6ka(zonal) correlated with D2xCO2(global) (b) D6ka(zonal) correlated with D2xCO2(zonal) Correlation coefficient Correlation coefficient

T2 land (d) DLGM(zonal) correlated with D2xCO2(global) (c) DLGM(zonal) correlated with D2xCO2(zonal) T2 ocean PPT land PPT ocean Correlation coefficient Correlation coefficient o latitude( ) latitude(o)

Fig. 4. Correlation of the annually averaged zonal changes for the 2×CO2 experiment with the annually averaged zonal changes for all the experiments, for both precipitation and 2 m temperature. The results are split into separate results for land and ocean. The magenta band shows the region which does not achieve signiﬁcance at the 99% level.

Figure 3. Figure 4. Correlation of the annually averaged global changesregions for showed a positive correlationCorrelation between climate of the sen- annuallyimproving averaged predictions aszonal that at changes the present-day, for and more the 2×CO2 experiment with the annually averaged zonal changessitivity and the magnitudethe 2×CO of the2 seasonalexperiment signal. Based with on theuseful annually than the averaged mid-Holocene. zonal Of course changes in order to take ad- for all the experiments, for both precipitation and 2 m temperature.these results it wouldfor seemall the likely experiments, that although the forseasonal both precipitationvantage of this relationship, and 2 m we temperature. would need paleodata that signal alone could not be used to precisely estimate climate is informative of the seasonal cycle at those times, and de- The results are split into separate results for land and ocean.sensitivity, The improvingThe the results present-day are seasonal split into signal separatein the composing results the for seasonal land climate and signal ocean. from Thepaleo-data is not magenta band shows the region which does not achieve significancemodels may helpmagenta improve the bandprediction shows of future the changes. region whichgenerally does currently not feasible, achieve although significance this result may become Since the forcings for the mid-Holocene climate largely of more practical importance as proxies are better understood at the 99% level, according to Student’s t-test. consist of changesat in the the seasonal 99% level,forcing we according might also hope to Student’sand modelled. t-test. that getting the correct seasonal response in the model for the mid-Holocene will help improve our climate model, and its 4.4 Monsoon regions ability to predict future change. However, we find little evidence to supportsphere this. Figure region6 shows that where changes the in sea- significantAs mentioned correlation in the Introduction, is found one of the in reasons for other words, the correlation tends to decrease in significancesonal cycle for DLGM of a similar magnitude (and opposite studying the mid-Holocene epoch is the evidence for in- sign) to the changesFig. for the 4b. increased Similarly, CO2 climate, for whereas DLGMcreased the monsoon correlation precipitation, is and weak for vegetation for further in those regions where the blue lines in Fig. 3a lie withinthose for D6ka are much smaller. The correlation between north in the African monsoon region, compared to the present the ocean and only marginally significant for small regions the magenta band. So, for improving climate changeDLGM pre- and climate sensitivity is quite high over the high lat- day. Given this regional-scale evidence, we examine the cor- itude bands and moderatein the in tropics. the mid-latitudes. These For D6ka, results in indicaterelations of the that model the responses paleo-climate to changes in forcing in diction on the zonal scale the temperature may be similarlycontrast, the correlation coefficients mostly do not rise above these regions. We take our definition of the monsoon re- the level of the samplingexperiments noise. These may results be suggest exercising that gions di fromfferentBraconnot aspects et al. (2007a relevant) and Ohgaito for and Abe- constrained on the global and regional scales, with the LGM the seasonal cycleclimate at the LGM change, may actually than be as useful the for controlOuchi experiment.(2007). For the Asian This monsoon suggests region, Braconnot being useful for the tropics and the high latitudes, and the that improving the modelling of the paleoclimates, over and control and mid-Holocene contributing in the northern hemi-www.clim-past.net/5/803/2009/ Clim. Past, 5, 803–814, 2009 above the present day climate, can be expected to further im- sphere, while for precipitation all three experiments may pro- prove future predictions. vide some useful information, but none is particularly domi- nant. 4.3 Analysis of seasonal averages Since there is considerable overlap with the results for the control with those for DLGM and D6ka, one may wonder if Tsushima and Manabe (2001) suggested that the global re- all we really need in order to build better models is to better sponse of the climate system to seasonal variation may be fit them better to the modern climate. Figure 5 investigates analagous to the global changes that occur under global this question for MIROC. It shows the correlation of DLGM warming. Using this assumption, they argued that since the and D6ka with CTRL for temperature. For D6ka, there is no cloud feedback effect was small for the annual global tem- significant correlation in the land over the Northern Hemi- perature variation, it could also be small for the case of an-

Journalname www.jn.net J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 7

(a) CTRL JJA-DJF temperature ( D2xCO2 D6ka DLGM o C)

810 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling

(b)

(a) D6ka(zonal) correlated with CTRL(zonal) T2 land T2 ocean Correlation coefficient

(b) DLGM(zonal) correlated with CTRL(zonal) Correlation with climate sensitivity

Latitude (o) Correlation coefficient

latitude(o) J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 7 Fig. 5. Correlation of the annually averaged zonal temperatures for the CTRL experiment with the annually averaged zonal temperaureFigure 6. (a) Seasonal (JJA-DJF) zonally averaged temperature changes for the LGM and mid-Holocene. The results are split into separate results for land and ocean. The magenta band shows the region which does not achieve significance at the 99% level. difference over land for the control climate (dot-dashed line) and the differences between the control and the simulated climates (solid Figure 5. Correlation of the annually averaged zonal temperatures et al. use a region focussed around northern India (70–lines). The thinner lines show the 1 standard deviation ranges of the for the CTRL experiment with the100 annually◦ E, 20–40◦ averagedN), while Ohgaito zonal and temper- Abe-Ouchi use a region which extends further east as far as Japan (70–140◦ ensemble.E, (b) Ensemble correlation of the seasonal change with (a) CTRL

aure changes for the LGM andJJA-DJF temperature ( mid-Holocene.◦ The results are split D2xCO2 22–40 N). We compare results from both these regions. Weclimate sensitivity. D6ka ◦ into separateDLGM results for land andalso ocean. analysed The the magenta West African band monsoon shows region (20 W– ◦ ◦ ◦ the region which does not achieve30 E, signiﬁcance 10 N–25 N). at the 99% level, according to Student’s t-test. Data on land for the mid-Holocene and LGM are typically

o taken from remnants of biological material, so may contain C) signal alone could not be used to precisely estimate climate information on a mixture of seasonal temperature, precipitation and other environmental factors. In the next few yearssensitivity, improving the present-day seasonal signal in the (b) we expect researchers to move towards utilising more of thismodels may help improve the prediction of future changes. (a) D6ka(zonal) correlated with CTRL(zonal) T2 land information, for example, by directly modelling the climate T2 ocean thropogenic global warming.proxies. In our Therefore ensemble, we present however, results for the we monsoon cli- Since the forcings for the mid-Holocene climate largely do not find a significant correlationmate on a between monthly basis. the The monthly present-day mean temperature results for the different regions are shown in Fig. 7. Figure 7consistc, of changes in the seasonal forcing we might also hope Correlation coefficient global seasonal temperature signald show the and correlations climate of monthly sensitivity. temperatures for CTRL,that getting the correct seasonal response in the model for the DLGM and D6ka with those for D2×CO2. The results are (b) DLGM(zonal) correlated with CTRL(zonal) Correlation with climate sensitivity Analysis on the zonal scalesimilar presents for both regions, a more with the interesting correlation being strongestmid-Holocene will help improve our climate model, and its for DLGM and moderate for CTRL, while there is no sig- o picture.Latitude ( The) dashed line in thenificant bottom correlation plot for D6ka. of Fig. In this 6 case indi- we do also findability to predict future change. However, we find little ev- significant correlation between the CTRL and DLGM tem-

Correlation coefficient Fig. 6. (a) Seasonalcates (JJA-DJF) the zonally zonal averaged variation temperature dif-of the correlation between the con- idence to support this. Figure 6 shows that changes in sea- ference over land for the control climate (dot-dashed line) and the peratures, suggesting that these data do not test independent latitude(o) differences betweentrol the controlseasonal and the signalsimulated climates (JJA-DJF (solid temperatureaspects of the model. over Figure land),7e, f show and the correlation cli- of thesonal cycle for DLGM of a similar magnitude (and opposite CTRL, DLGM and D6ka temperatures with precipitation for Figurelines). 6. The(a) thinnerSeasonal lines (JJA-DJF) show the zonally1 standard averaged deviation temperature ranges of the ensemble. (b) mateEnsemble sensitivitycorrelation of the seasonal for ourchange model with D2 ensemble.×CO2. Here the results The are correlation less significant although is theresign) to the changes for the increased CO2 climate, whereas difference over land for the control climate (dot-dashed line)◦ and ◦ ◦ ◦ theclimate differences sensitivity. betweenstronger the control andaround the simulated 25–30 climates (solidN, 70 areN, a small 70 numberS and of 25–30 months whereS. there In ad-is some signifi-those for D6ka are much smaller. The correlation between Figure 5. Correlation of the annually averaged zonal temperatures lines). The thinner lines show the 1 standard deviation ranges of the cant correlation for DLGM in the larger Asian monsoon re- for the CTRL experiment with the annually averaged zonal temper- ensemble. (b) Ensembledition correlation these of the same seasonal parts change of with thegion globe and also also for DLGM show and the CTRL strongest in West Africa. FigureDLGM8 and climate sensitivity is quite high over the high lat- aure changes for the LGM and mid-Holocene. The results are split climate sensitivity. into separate results for land and ocean. The magenta band shows correlation for D2×CO2 with climate sensitivity. It would, itude bands and moderate in the mid-latitudes. For D6ka, in the region which does not achieve significance at the 99% level, Clim. Past, 5, 803–814, 2009 www.clim-past.net/5/803/2009/ according to Student’s t-test. therefore, appear that the MIROC results are similar in char- contrast, the correlation coefficients mostly do not rise above signal alone could not be used to precisely estimate climate sensitivity, improvingacter the present-dayto those seasonal of HadSM3 signal in the in Knutti et al. (2006), who also the level of the sampling noise. These results suggest that models may helpfound improve the that prediction many of future Northern changes. Hemisphere extratropical land re- the seasonal cycle at the LGM may actually be as useful for thropogenic global warming. In our ensemble, however, we Since the forcings for the mid-Holocene climate largely do not find a significant correlation between the present-day consist of changesgions in the seasonal showed forcing awe positive might also hope correlation between climate sensi- improving predictions as that at the present-day, and more global seasonal temperature signal and climate sensitivity. that getting the correct seasonal response in the model for the Analysis on the zonal scale presents a more interesting mid-Holocene willtivity help improve and our the climate magnitude model, and its of the seasonal signal. Based on useful than the mid-Holocene. Of course in order to take ad- picture. The dashed line in the bottom plot of Fig. 6 indi- ability to predictthese future change. results However, it would we find little seem ev- likely that although the seasonal vantage of this relationship, we would need paleodata that cates the zonal variation of the correlation between the con- idence to support this. Figure 6 shows that changes in sea- trol seasonal signal (JJA-DJF temperature over land), and cli- sonal cycle for DLGM of a similar magnitude (and opposite mate sensitivity for our model ensemble. The correlation is sign) to the changes for the increased CO2 climate, whereas stronger around 25–30◦ N, 70◦ N, 70◦ S and 25–30◦ S. In ad- those for D6ka arewww.jn.net much smaller. The correlation between Journalname dition these same parts of the globe also show the strongest DLGM and climate sensitivity is quite high over the high lat- correlation for D2×CO2 with climate sensitivity. It would, itude bands and moderate in the mid-latitudes. For D6ka, in therefore, appear that the MIROC results are similar in char- contrast, the correlation coefficients mostly do not rise above acter to those of HadSM3 in Knutti et al. (2006), who also the level of the sampling noise. These results suggest that found that many Northern Hemisphere extratropical land re- the seasonal cycle at the LGM may actually be as useful for gions showed a positive correlation between climate sensi- improving predictions as that at the present-day, and more tivity and the magnitude of the seasonal signal. Based on useful than the mid-Holocene. Of course in order to take ad- these results it would seem likely that although the seasonal vantage of this relationship, we would need paleodata that www.jn.net Journalname J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 9

J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 811

(a) Asia (b) West Africa CTRL DLGM D6ka

C C D2xCO2 o o Tempertaure Tempertaure Tempertaure Tempertaure

(e) correlation with D2xCO2 precip (f) correlation with D2xCO2 precip Correlation coefficient Correlation coefficient

Month Month

Fig. 7. Correlation of monthly 2 m temperature results for the monsoon region. Left plots, Asia; solid lines, Braconnot et al. (2007a) region; dashed lines: Ohgaito and Abe-Ouchi (2007) region. Right plots, West Africa, Braconnot et al. (2007a). (a, b) Temperatures for CTRL, Figure 7. Correlation of monthly 2 m temperature results for the monsoon region. Left plots, Asia; solid lines, Braconnot et al. (2007a) DLGM, D6ka and D2×CO2. (c, d) Correlation of monthly temperature for CTRL, DLGM, D6ka with monthly temperature for D2×CO2. region; dashed(e, f) Correlation lines: Ohgaito of monthly and temperature Abe-Ouchi for CTRL, (2007) DLGM, region.D6ka with monthly Rightprecipitation plots, West for D2 Africa,×CO2. Braconnot et al. (2007a). (a, b) Temperatures for CTRL, DLGM, D6ka and D2×CO2. (c, d) Correlation of monthly temperature for CTRL, DLGM, D6ka with monthly temperature for D2×CO2. (e,shows f) Correlation the results for the of precipitation monthly temperature for CTRL, DLGM for CTRL, and 5 DLGM, Discussion D6ka with monthly precipitation for D2×CO2. D6ka. Figure 8a, b show the precipitation for CTRL, DLGM, × D6ka and D2 CO2 for the two regions. It is notable here that For MIROC, there is evidence that improving both the pre- × the changes are rather large for D2 CO2, DLGM and D6ka industrial and LGM temperatures and precipitation should when compared to the base annual cycle of the control run. influence both the estimates of climate sensitivity and global × All experiments, but particularly D2 CO2 have a large en- precipitation change. For estimating future changes on the semble spread from May to October, indicating a high degree zonal scale, evidence from the pre-industrial, LGM and of uncertainty in the ensemble predictions. Figure 8b,c show mid-Holocene should all prove useful. Moving to the re- × the results for the correlations with D2 CO2 precipitation. gional scale of the monsoon, the LGM and pre-industrial For Asia, only D6ka shows any significant correlation with climates are clearly the most useful for improving fu- × (a) Asia (b) West Africa precipitation at D2 CO2, while for Africa D6ka has signfi- ture temperatures. For the all-important prediction of the cant correlation coefficientsCTRL for 7 months of the year, and the DLGM monsoon rainfall change, the evidence is less strong, but in CTRL for 4 months. The correlationD6ka of the CTRL with D6ka our results, the mid-Holocene precipitation, and to a lesser D2xCO2 (not shown) is only significant for the month of November. extent the pre-industrial temperature and precipitation would To summarise the results for the monsoon regions, im- seem to be most useful. proving temperatures in the monsoon regions for LGM and CTRL simulations would be expected to improve future pre- Adding to the potential value of the mid-Holocene climate Precipitation mm/dy dictions of temperature.Precipitation mm/dy For precipitation the evidence is is the fact that in the regions for which we found the most sig- much weaker, but the most notable feature is that, for the nificant results in our model-model correlations, there is also monsoon regions, the results for the mid-Holocene are at clear evidence of changes in climate from the mid-Holocene least as strong as those for the other epochs. data. For the mid to high latitude land in the Northern Hemi- sphere (Sects. 4.1 and 4.2), there is evidence from the proxy

(c) correlation with D2xCO2 precip data which may be(d) correlation used for with model D2xCO validation.2 precip In particular,

www.clim-past.net/5/803/2009/ Clim. Past, 5, 803–814, 2009 Correlation coefficient Correlation coefficient

Month Month

Figure 8. Correlation of monthly precipitation results for the monsoon region. Left plots, Asia; solid lines, Braconnot et al. (2007a) region; dashed lines: Ohgaito and Abe-Ouchi (2007) region. Right plots, West Africa, Braconnot et al. (2007a). (a, b) Precipitation for CTRL, DLGM, D6ka and D2×CO2. (c, d) Correlation of monthly precipitation for CTRL, DLGM, D6ka with monthly precipitation for D2×CO2.

www.jn.net Journalname J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 9

(a) Asia (b) West Africa CTRL DLGM D6ka

C C D2xCO2 o o Tempertaure Tempertaure Tempertaure Tempertaure

(e) correlation with D2xCO2 precip (f) correlation with D2xCO2 precip Correlation coefficient Correlation coefficient

Month Month

Figure 7. Correlation of monthly 2 m temperature results for the monsoon region. Left plots, Asia; solid lines, Braconnot et al. (2007a) region; dashed lines: Ohgaito and Abe-Ouchi (2007) region. Right plots, West Africa, Braconnot et al. (2007a). (a, b) Temperatures for CTRL, DLGM, D6ka and D2×CO2. (c, d) Correlation of monthly temperature for CTRL, DLGM, D6ka with monthly temperature for D2×CO2. (e, f) Correlation of monthly temperature for CTRL, DLGM, D6ka with monthly precipitation for D2×CO2.

812 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling

(a) Asia (b) West Africa CTRL DLGM D6ka D2xCO2 Precipitation mm/dy Precipitation mm/dy

Month Month

Fig. 8. Correlation of monthly precipitation results for the monsoon region. Left plots, Asia; solid lines, Braconnot et al. (2007a) region; dashedFigure lines: 8. CorrelationOhgaito andof monthly Abe-Ouchi precipitation(2007) region. results for Right the plots,monsoon West region. Africa, LeftBraconnot plots, Asia; et solidal. (2007a lines,). Braconnot(a, b) Precipitation et al. (2007a) for region; CTRL, DLGM,dashed D6kalines: and Ohgaito D2×CO and2 Abe-Ouchi. (c, d) Correlation (2007) region. of monthly Right precipitation plots, West for Africa, CTRL, Braconnot DLGM, etD6ka al. (2007a). with monthly(a, b) precipitationPrecipitation for for D2 CTRL,×CO2. DLGM, D6ka and D2×CO2. (c, d) Correlation of monthly precipitation for CTRL, DLGM, D6ka with monthly precipitation for D2×CO2. there is evidence for warmer climates compared to present In this context is worth noting that modern GCMs do not in the northern latitudes of Eurasia (Tarasov et al., 1999; agree on the sign of the change for future precipitation in Bigelowwww.jn.net et al., 2003). the monsoon regions (Meehl et al., 2007), whereasJournalname those that One of the clearest features of the mid-Holocene climate have been run for PMIP2 do all agree on an increase for the found in proxy records is the change in vegetation type in the mid-Holocene compared to present (Braconnot et al., 2007a). Sahara north of the present-day monsoon region compared Despite only using a single model in our experiments, our re- to the present day. To date most climate models have not sults are consistent with this. As Fig. 8a, b imply, the precip- managed to reproduce this feature (Braconnot et al., 2007a itation changes in the monson season are all positive for the and references therein). For the mid-Holocene, while there is D6ka experiment, while for the D2×CO2 experiment we see a general increase in the monsoon precipitation and a north- both increases and decreases among the ensemble members. ward shift in the ITCZ over Africa (Braconnot et al., 2007b), For the annual mean rainfall in the three monsoon regions, very few models (and none of our ensemble members) have the precipitation change is negative for 40% of the D2×CO2 produced enough of a climate change to induce significant members but negative for only 1% of the D6ka members. vegetation changes in the Sahara (Braconnot et al., 2007a). For practical reasons, an in depth analysis of the processes It is of some concern that there should be such strong ev- and feedbacks governing the different climate changes stud- idence for bias across so many models, indicating that the ied here is outside the scope of this paper, and is left as a sub- models do not properly represent the processes governing ject for future studies. We are also aware that our study fo- the mid-Holocene. If that is the case, then our finding corre- cusses on only one model, and that the results are not nec- lations between the mid-Holocene and the future climate in essarily directly tranferrable to other models, or even fu- the model-space may not translate directly for the real world, ture development of MIROC. For example, the (unperturbed) since the missing processes may overwhelm our results. On T42 version of MIROC3.2 does not suffer from the large the other-hand, the fact that there is such a clear signal in radiative imbalance that we found with the lower resolution the paleo-data gives us some good evidence in the past with T21 version (and discussed in Sect. 3), but due to compu- which to validate and improve the model. Given the short- tational constraints we were limited to using the latter. As age of data for validating climate changes, and the fact that has previously been noted, correlations across multiple mod- we do get a significant correlation between mid-Holocene els tend to be weaker than in individual model ensembles, and increased CO2 climates actually suggests that further im- as there are many more differences between different mod- proving the simulation of the mid-Holocene climate would els than can be represented by changing a limited number be a suitable focus for those wishing to improve predictions of parameters in one model (Crucifix, 2006). In addition, of the future. the MIROC3.2 slab-ocean model omits some potentially important feedbacks, such as the effect of ocean dynamics and

Clim. Past, 5, 803–814, 2009 www.clim-past.net/5/803/2009/ J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling 813 vegetation response. While previous work with MIROC3.2 Education, Culture, Sports, Science and Technology (MEXT), and has indicated that the influence of the ocean dynamics on by the Global Environment Research Fund (S-5-1) of the Ministry the monsoon is not of first order importance (Ohgaito and of the Environment, Japan. Abe-Ouchi, 2007), recent work integrating a dynamic vegetation model (LPJ) into the standard version of MIROC3.2 Edited by: V. Masson-Delmotte has shown that for increased CO2, the dynamic vegetation amplifies the increase in precipitation in the monsoon regions (O’ishi and Abe-Ouchi, 2009). It is, therefore, possible that the inclusion of dynamic vegetation could bring the models References closer to modelling the green Sahara at the mid-Holocene. Therefore, it would be helpful if other similar analyses could Abe, M., Shiogama, H., Hargreaves, J., Annan, J., Nozawa, T., be undertaken with different models incorporating different and Emori, S.: Correlation between Inter-Model Similarities in feedbacks. Spatial Pattern for Present and Projected Future Mean Climate, SOLA, 5, 133–136, 2009. Annan, J. D., Hargreaves, J. C., Ohgaito, R., Abe-Ouchi, A., and Emori, S.: Efficiently constraining climate sensitivity with pale- 6 Conclusions oclimate simulations, SOLA, 1, 181–184, 2005. Arnold Jr., C. and Dey, C.: Observing-systems simulation experi- On the whole we find that the LGM, with its relatively large ments: Past, present, and future, B. Am. Meteorol. Soc, 67, 687– negative change in CO2 forcing, is more likely to be useful 695, 1986. for constraining future climate change, despite the noise in Berger, A.: A simple algorithm to compute long term variations of the correlations introduced by the responses to large ice sheet daily or monthly insolation, Tech. Rep. 18, Universite` Catholique changes which do not seem relevant to future climate change. de Louvain, Belgium, 1978. Bigelow, N., Brubaker, L., Edwards, M., et al.: Climate change and For large-scale variables such as climate sensitivity, and the Arctic ecosystems: 1. Vegetation changes north of 55 N between seasonal signal on zonal scales, analysing the behaviour of the last glacial maximum, mid-Holocene, and present, J. Geo- the mid-Holocene would seem to be less helpful. The one phys. Res, 108, 8170, doi:10.1029/2002JD002558, 2003. exception to this is the northern high latitude land, where Braconnot, P., Otto-Bliesner, B., Harrison, S., Joussaume, S., Pe- at the LGM the large ice sheets causes the relation to the in- terchmitt, J.-Y., Abe-Ouchi, A., Crucifix, M., Driesschaert, E., ˆ creased CO2 climate change to be weak, whereas the changes Fichefet, Th., Hewitt, C. D., Kageyama, M., Kitoh, A., Laine,´ in mid-Holocene temperatures over land are, in fact, mod- A., Loutre, M.-F., Marti, O., Merkel, U., Ramstein, G., Valdes, erately well correlated with both climate sensitivity and the P., Weber, S. L., Yu, Y., and Zhao, Y.: Results of PMIP2 coupled simulations of the Mid-Holocene and Last Glacial Maximum - zonal changes for increased CO2. On the more regional scale of the African and Asian monsoons, the mid-Holocene shows Part 1: experiments and large-scale features, Clim. Past, 3, 261– interesting results for the precipitation changes, although the 277, 2007, http://www.clim-past.net/3/261/2007/. LGM is again more relevant for temperature. Braconnot, P., Otto-Bliesner, B., Harrison, S., Joussaume, S., Pe- The case for continuing modelling of the mid-Holocene terchmitt, J.-Y., Abe-Ouchi, A., Crucifix, M., Driesschaert, E., is supported by the availability of observational evidence Fichefet, Th., Hewitt, C. D., Kageyama, M., Kitoh, A., Loutre, which indicates climatic conditions significantly different M.-F., Marti, O., Merkel, U., Ramstein, G., Valdes, P., Weber, from present in both these regions, which may be used to L., Yu, Y., and Zhao, Y.: Results of PMIP2 coupled simulations validate the models. There is evidence for a warmer cli- of the Mid-Holocene and Last Glacial Maximum - Part 2: feed- mate compared to present in the northern latitudes of Eurasia backs with emphasis on the location of the ITCZ and mid- and from proxy data. It is also well known that the modelled high latitudes heat budget, Clim. Past, 3, 279–296, 2007, monsoon changes at the mid-Holocene are insufficiently dra- http://www.clim-past.net/3/279/2007/. matic: there should be more rain further north, particularly in Collins, M., Booth, B., Harris, G., Murphy, J., Sexton, D., and Africa (Braconnot et al., 2007a). While the caveats relating Webb, M.: Towards quantifying uncertainty in transient climate to model inadequacy mentioned in Sect. 3 should be kept in change, Clim. Dynam., 27, 127–147, 2006. mind, because the effect of increased CO on monsoon pre- 2 Crucifix, M.: Does the Last Glacial Maximum constrain climate cipitation is highly uncertain, the mid-Holocene should not sensitivity?, Geophys. Res. Lett., 33, L18 701, doi:10.1029/ be ignored. 2006GL027137, 2006. Hargreaves, J. C., Abe-Ouchi, A., and Annan, J. D.: Linking glacial and future climates through an ensemble of GCM simulations, Acknowledgements. The authors would like to thank A. Abe-Ouchi Clim. Past, 3, 77–87, 2007, for encouragement and discussions which led to the development http://www.clim-past.net/3/77/2007/. of this work, R. Ohgaito for crucial assistance with MIRCO3.2, and Jolly, D., Harrison, S., Damnati, B., and Bonnefille, R.: Simulated the two referees, M. Crucifix and A. Ganopolski for their helpful climate and biomes of Africa during the late quaternary compar- suggestions. This work was supported by Innovative Program of ison with pollen and lake status data, Quaternary Sci. Rev., 17, Climate Change Projection for the 21st Century of the Ministry of 629–657, 1998. www.clim-past.net/5/803/2009/ Clim. Past, 5, 803–814, 2009 814 J. C. Hargreaves and J. D. Annan: Importance of paleo-modelling

Joussaume, S. and Braconnot, P.: Sensitivity of paleoclimate simu- Russell, G., Miller, J., and Tsang, L.: Seasonal oceanic heat trans- lation results to season definitions, J. Geophys. Res., 102, 1943– ports computed from an atmospheric model, Dynam. Atmos. 1956, 1997. Oceans., 9, 253–271, 1985. Joussaume, S., Taylor, K., Bracconot, P., et al.: Monsoon changes Schneider von Deimling, T., Held, H., Ganopolski, A., and Rahm- for 6000 years ago: results of 18 simulations from the Paleocli- storf, S.: Climate sensitivity estimated from ensemble simula- mate Modeling Intercomparison Project (PMIP), Geophys. Res. tions of glacial climate, Clim. Dynam., 27, 149–163, 2006. Lett., 26, 859–862, 1999. Solomon, S., Qin, D., Manning, M., Chen, Z., et al.: Climate change Knutti, R., Meehl, G., Allen, M., and Stainforth, D.: Constraining 2007: The physical science basis. Contribution of the Working climate sensitivity from the seasonal cycle in surface tempera- Group I to the Fourth Assessment Report of the Intergovernmen- ture, J. Climate, 19, 4224–4233, 2006. tal Panel on Climate Change, 2007. Meehl, G. A., Stocker, T. F., Collins, W. D., et al.: Global Climate Tarasov, P., Guiot, J., Cheddadi, R., et al.: Climate in northern Eura- Projections, in: Climate Change 2007: The physical science ba- sia 6000 years ago reconstructed from pollen data, Earth Planet sis. Contribution of the Working Group I to the Fourth Assess- Sc. Lett., 171, 635–645, 1999. ment Report of the Intergovernmental Panel on Climate Change, Tsushima, Y. and Manabe, S.: Influence of cloud-feedback on an- chap. 10, Cambridge University Press, Cambridge, United King- nual variation of global mean surface temperature, J. Geophys. dom and New York, NY, USA, 2007. Res., 106, 22635–22646, 2001. Mitchell, J.: Greenhouse warming: Is the Mid-Holocene a good Whetton, P., Macadam, I., Bathols, J., and O’Grady, J.: As- analogue?, J. Climate, 3, 1177–1192, 1990. sessment of the use of current climate patterns to evaluate re- Murphy, J. M., Sexton, D. M. H., Barnett, D. N., Jones, G. S., Webb, gional enhanced greenhouse response patterns of climate models, M. J., Collins, M., and Stainforth, D. A.: Quantification of mod- Geophys. Res. Lett., 34, L14 701, doi:10.1029/2007GL030025, elling uncertainties in a large ensemble of climate change simu- 2007. lations, Nature, 430, 768–772, 2004. Yokohata, T., Webb, M., Collins, M., Williams, K. D., Yoshimori, Ohgaito, R. and Abe-Ouchi, A.: The role of ocean thermodynam- M., Hargreaves, J. C., and Annan, J. D.: Structural similarities ics and dynamics in Asian summer monsoon changes during the and differences in climate responses to CO2 increase between mid-Holocene, Clim. Dynam., 29, 39–50, 2007. two perturbed physics ensembles by general circulation models, O’ishi, R. and Abe-Ouchi, A.: Influence of dynamic vegetation J. Climate, in press, 2009. on climate change arising from increasing CO2, Clim. Dynam., 33(5), 645–663, doi:10.1007/s00382-009-0611-y, 2009.

Clim. Past, 5, 803–814, 2009 www.clim-past.net/5/803/2009/