Volume 110, Issue D9
Climate and Dynamics
Free Access

Effect of climate sensitivity on the response to volcanic forcing

T. M. L. Wigley

T. M. L. Wigley

National Center for Atmospheric Research, Boulder, Colorado, USA

Search for more papers by this author
C. M. Ammann

C. M. Ammann

National Center for Atmospheric Research, Boulder, Colorado, USA

Search for more papers by this author
B. D. Santer

B. D. Santer

Program for Climate Model Diagnosis and Intercomparison, Lawrence Livermore National Laboratory, Livermore, California, USA

Search for more papers by this author
S. C. B. Raper

S. C. B. Raper

Climatic Research Unit, University of East Anglia, Norwich, UK

Alfred Wegener Institute for Polar and Marine Research, Bremerhaven, Germany

Search for more papers by this author
First published: 06 May 2005
Citations: 105

Abstract

[1] The results from 16 coupled atmosphere/ocean general circulation model (AOGCM) simulations are used to reduce internally generated noise and to obtain an improved estimate of the underlying response of 20th century global mean temperature to volcanic forcing. An upwelling diffusion energy balance model (UD EBM) with the same forcing and the same climate sensitivity as the AOGCM is then used to emulate the AOGCM results. The UD EBM and AOGCM results are in very close agreement, justifying the use of the UD EBM to determine the volcanic response for different climate sensitivities. The maximum cooling for any given eruption is shown to depend approximately on the climate sensitivity raised to power 0.37. After the maximum cooling for low-latitude eruptions the temperature relaxes back toward the initial state with an e-folding time of 29–43 months for sensitivities of 1–4°C equilibrium warming for CO2 doubling. Comparisons of observed and modeled coolings after the eruptions of Agung, El Chichón, and Pinatubo give implied climate sensitivities that are consistent with the Intergovernmental Panel on Climate Change (IPCC) range of 1.5–4.5°C. The cooling associated with Pinatubo appears to require a sensitivity above the IPCC lower bound of 1.5°C, and none of the observed eruption responses rules out a sensitivity above 4.5°C.

1. Introduction

[2] A number of studies have attempted to estimate the climate sensitivity (defined here by the equilibrium warming for a CO2 doubling (ΔT2x)) by comparing model simulations with observed climate changes over the 20th century [Andronova and Schlesinger, 2001; Forest et al., 2002; Schneider and Mass, 1977; Gregory et al., 2002; Harvey and Kaufmann, 2002]. These estimates, however, still leave wide margins of uncertainty, not least because of the noise of internally generated variability in the observations and because of substantial uncertainties in the past forcing history (arising primarily from anthropogenic aerosol forcing uncertainties). For example, if there were a substantial internally generated cooling trend in the past record or if the applied forcing underestimated the true magnitude of sulfate aerosol-induced cooling over the 20th century, the model-based warming would be too high, and the implied value of ΔT2x would be too low.

[3] An alternative that has been suggested is to use comparisons between the modeled and observed effects of volcanic eruptions for model validation and estimation of the climate sensitivity [Hansen et al., 1993; Lindzen and Giannitsis, 1998]. There are a number of difficulties with this approach, as articulated by Lindzen and Giannitsis [1998, hereinafter referred to as LG98]. First, even for the eruption of Mount Pinatubo (June 1991) where satellite data have provided us with detailed information on the properties and distribution of the volcanic aerosol, there are still substantial discrepancies between different estimates of the forcing [see, e.g., Santer et al., 2001]. Uncertainties in the forcings for earlier eruptions are necessarily larger. Model-based signals therefore have considerable intrinsic uncertainty, even for results from a single model. Second, there is a signal-to-noise ratio problem. Since the relevant response is on a monthly timescale and since the response to an individual eruption decays to a negligible amount after less than a decade, the noise of internally generated variability makes it difficult to define the response signal in the observations (although some of these noise influences, such as the effects of El Niño–Southern Oscillation (ENSO) variability, may be removed by empirical methods; see section 5). Third, short-timescale forcing events (spanning 5 years or less) are less sensitive to ΔT2x than longer-timescale processes. If the response is relatively insensitive to ΔT2x, then it becomes much more difficult to back out information about ΔT2x from any model/observed data comparison. Fourth, the sensitivity may depend on the nature of the forcing and its spatial distribution [Wigley, 1994; Joshi et al., 2003]; we assume here that any such dependence is within the uncertainties of an empirical sensitivity estimate.

[4] LG98 note that the longer-timescale response to multiple consecutive eruptions is more strongly dependent on ΔT2x (as pointed out and quantified earlier by Wigley [1991]), so this may provide an alternative way to obtain information on ΔT2x from the observational record. Unfortunately, this approach is confounded by the effects of (and uncertainties in) other forcings, both natural and anthropogenic, and by the magnitude of internally generated noise.

[5] LG98 attempted to estimate ΔT2x from the global mean temperature response to volcanic forcings using a three-box, energy balance climate model with a limited-depth (400 m) diffusive ocean. The authors show that this very simple model emulates the results of a slightly more realistic upwelling diffusion energy balance model (UD EBM) [Hoffert et al., 1980], but the LG98 model still has admitted shortcomings. In this paper we use a more detailed UD EBM to investigate the effect of ΔT2x on the response to 20th century volcanic forcings. In order to determine the credibility of the UD EBM in this context we compare its results to those obtained using a fully coupled atmosphere/ocean general circulation model (AOGCM). The climate sensitivity of the AOGCM is fixed by the model's physics and parameterizations. The simpler model, however, has a user-specified climate sensitivity. Thus, provided that the simpler model is able to match the results of the AOGCM when its sensitivity is set equal to that of the AOGCM, the simpler model may be run for a range of climate sensitivities to see how different characteristics of the response to volcanic forcing vary as the sensitivity is changed. The climate sensitivity can then be estimated by adjusting the sensitivity of the EBM to obtain a best fit match to the observed responses to individual eruptions.

2. Analytical Results

[6] We begin with a simple piece of pedagogy to provide some insights into the factors that control the response to time-dependent external forcing. This response is determined primarily by the climate sensitivity and the thermal inertia of the climate system, with the relative importance of these two factors depending on the timescale of the forcing. For very slow (multicentury) timescale forcing changes the system is able to maintain near equilibrium with the forcing, so sensitivity effects dominate. For rapid forcing changes, such as those for the seasonal insolation cycle, the response is dominated by inertia effects. Volcanic forcing lies between these two extremes, and it is easy to demonstrate that both sensitivity and inertia effects are important. We do this in an idealized way using sinusoidal forcing (which allows the forcing timescale to be uniquely defined) and employing the simplest possible climate model, a one-box model represented by
equation image
Here C is a heat capacity term, Q(t) is the applied external forcing, S is the climate sensitivity expressed as temperature change per unit radiative forcing (i.e., S = ΔT2x/ΔQ2x, where ΔQ2x is the forcing for CO2 doubling), and ΔT(t) is the change in global mean temperature.
[7] For sinusoidal forcing (Q(t) = A sin(ωt)) the solution is
equation image
where ω is in rad yr−1 (ω = 2π/T, where T is the period in years) and τ (years) is a characteristic timescale for the system, τ = SC. (The sine/cosine term can be written in the form sin(ωt + ϕ), showing that the asymptotic response follows the forcing with a lag, ϕ; however, the expanded form is more convenient here.)
[8] We now consider two end-member cases, for high-frequency and low-frequency forcing. For the latter (ω ≪ 1/τ) the asymptotic solution is simply the equilibrium response
equation image
showing no appreciable lag between forcing and response, with the response being linearly dependent on the climate sensitivity and independent of the system's heat capacity. For the high-frequency case (ω ≫ 1/τ) the solution is
equation image
showing a quarter cycle lag of response behind forcing, with the response being independent of the climate sensitivity.

[9] The critical question, then, is what is the appropriate timescale for volcanic forcing relative to the characteristic timescale (τ) for the climate system? If representative values are used for the effective heat capacity [Wigley and Raper, 1991] and S, and a realistic volcanic forcing timescale corresponding to ω = 1–3 rad yr−1 is assumed, then we can show that the response to volcanic forcing should have a relatively small but nonnegligible dependence on the climate sensitivity. This is in accord with model simulation results obtained by Wigley [1991] and by LG98. Because the effective heat capacity of the climate system depends nonlinearly on the climate sensitivity [see Wigley and Raper, 1991], the general solution for ΔT(t) also depends nonlinearly on the climate sensitivity.

3. Volcanic Response Signal

[10] We begin by showing the response to 20th century volcanic forcing for simulations with a fully coupled AOGCM, the National Center for Atmospheric Research (NCAR)/U.S. Department of Energy (USDOE) parallel climate model (PCM) [Washington et al., 2000]. (Note that PCM has a relatively low sensitivity compared with other AOGCMs [see Raper et al., 2001].) We use results from simulations carried out by Ammann et al. [2003, also C. M. Ammann et al., Coupled simulations of the 20th century including external forcing, submitted to Journal of Climate, 2005], which employ a new forcing history developed by Ammann. In total, there are 16 simulations that include volcanic forcing, comprising four-member ensembles for the following four experiments: volcanic forcing alone (V); volcanic plus solar forcing (VS); volcanic plus solar plus ozone forcing (VSO); and combined volcanic, solar, ozone, well-mixed greenhouse gases, and direct sulfate aerosol forcing (ALL). The runs begin in 1890 and end in 1999. In addition, we have four unforced control run experiments spanning the same 1890–1999 interval. There are also parallel four-member ensembles of volcano-free experiments with solar forcing alone (S); solar plus ozone forcing (SO); and solar, ozone, well-mixed greenhouse gas, and aerosol forcing (SOGA) that we make use of below.

[11] An important consideration in identifying the volcanic signal is the problem of spatial drift in the AOGCM, which can be quantified here using the four control run experiments. At the hemispheric mean scale in PCM this drift is appreciable: a warming of 0.16°C century−1 in the Southern Hemisphere and a cooling of 0.17°C century−1 in the Northern Hemisphere (NH). Comparisons of results for different forcing combinations show that these trends are common to all simulations. We minimize the problem by considering only global mean changes here for which the PCM's drift is small (−0.01°C century−1). Drift effects are removed in all the data we consider by subtracting the above trend from all simulations.

[12] Figure 1a shows the results for a single volcano-only realization. For the underlying signal the largest eruptions are those of Santa Maria (October 1902), Agung (March 1963), El Chichón (April 1982), and Pinatubo (June 1991), all of which occurred in tropical latitudes (see Figure 1a, vertical lines). We also consider results for the high–NH latitude eruption of Novarupta (June 1912). For all cases except Pinatubo and El Chichón the response is largely obscured by the “noise” of internally generated variability. Cooling around the time of El Chichón is apparent, but for this particular realization it begins well before the eruption date because of a random warming event around 1980.

Details are in the caption following the image
(a) Single realization (run B06.77) and (b and c) ensemble mean (n = 4 and n = 16, respectively) responses to volcanic forcing using the parallel climate model (PCM). Eruption dates for Santa Maria (October 1902), Novarupta (June 1912), Agung (March 1963), El Chichón (April 1982), and Pinatubo (June 1991) are shown by the vertical lines.

[13] A better estimate of the volcanic response signal can be obtained simply by averaging the four members of the V ensemble (Figure 1b). This ensemble averaging reduces the noise about the volcanic response signal by a factor of ∼2. To be more specific, for the control runs (drift corrected, see above) the interannual standard deviation over 1890–1999, averaged over four ensemble members, is 0.171°C. This provides an estimate of the AOGCM's internally generated variability that is superimposed on the volcanically induced temperature signal in any single simulation. Averaging four simulations should reduce this noise to ∼0.086°C (which it does if the four control runs are averaged). The four main eruption events now become clear, but they are still too noisy to allow confident quantification of the maximum cooling (although the situation is improved if ENSO events are statistically removed from the individual realizations).

[14] A further reduction in the noise is possible by making use of the other runs that include volcanic forcing. For these cases, results for volcano-free companion experiments to VS (i.e., S), VSO (SO), and ALL (SOGA) are available, and these may be subtracted from the “with-volcanic-forcing” cases to give residual volcano-only results (e.g., Vresid = VS − S, where S is a solar-forcing-alone ensemble).

[15] The gain here is less than might naively be expected because by virtue of their construction method the residual volcanic cases have amplified noise, which partly offsets the noise reduction that arises from the increase in sample size. For example, for the VS − S case we have
equation image
where Var[X] is the variance of X, sX is the standard deviation of X, and r(X, Y) is the correlation between X and Y. If VS and S were uncorrelated and both had the same standard deviation (which is approximately the case here), the standard deviation of VS − S would be inflated by equation image relative to that for V. Because of the common “S” signal in VS and S the two series are, in fact, weakly correlated (r(VS, S) > 0), so the noise inflation is slightly less than equation image.

[16] When all 16 volcano runs (i.e., four from V and four each from VS − S, VSO − SO, and ALL − SOGA) are averaged, the residual variability decreases to 0.059°C (Figure 1c). The reduction in noise compared with what would be obtained from a single AOGCM realization is therefore 65%, compared with the maximum possible reduction for 16 independent volcano-only runs of 75%.

[17] The improved definition of the volcanic signal as the internally generated noise is progressively reduced is clear in Figure 1. As noted above, only the Pinatubo and El Chichón signals are obvious in the single realization case. When four ensemble members are averaged, the Santa Maria, Agung, El Chichón, and Pinatubo signals become clear, but they are still not well defined quantitatively. The n = 16 ensemble mean case defines these four eruption signals much more clearly and identifies (albeit weakly) the Novarupta signal. The characteristic response signature revealed here is a rapid cooling over the first 7–18 months to a peak cooling that is (as far as can be judged given the residual noise) approximately proportional to the maximum forcing, followed by an approximately exponential relaxation back to the initial state. The relaxation time will be discussed further in section 5.

4. Validating the UD Model

[18] Even with an ensemble size of 16 the volcano signal is still partially obscured by the noise of internally generated variability. Further, these results are specific to PCM and to PCM's climate sensitivity. In order to estimate the “pure” signal and to extend the results to other sensitivities we use a simple UD EBM, namely, the model for the assessment of greenhouse gas–induced climate change (MAGICC) [Wigley and Raper, 1992, 2001; Raper et al., 1996] used in various Intergovernmental Panel on Climate Change (IPCC) reports [e.g., Cubasch et al., 2001]. First, however, it is necessary to demonstrate that MAGICC can emulate the PCM results in a like-with-like comparison.

[19] It has already been shown, as part of the IPCC Third Assessment Report (TAR) [see Raper et al., 2001; Cubasch et al., 2001, appendix], that MAGICC can emulate the responses of a range of AOGCMs for the case of forcing by a 1% compound CO2 increase (using the experiments coordinated under Coupled Model Intercomparison Project [Covey et al., 2003]). PCM was one of those models, so we use the TAR calibration results to define the model parameters in MAGICC. The parameters are the climate sensitivity, the land-ocean sensitivity ratio, the oceanic mixed layer depth, the ocean's effective vertical diffusivity, the rate of change of upwelling rate as a function of temperature, and land-ocean and interhemispheric heat exchange rates (note that MAGICC separates the globe into land and ocean “boxes” in each hemisphere). Applying these long-timescale calibration results to the much shorter timescales of a volcanic eruption is quite a severe test of the UD EBM.

[20] There is still one unspecified parameter. The primary forcing from the AOGCM simulations is produced as optical depth (OD) changes (defined at some specified frequency), while the UD EBM requires input as forcing at the top of the troposphere (in W m−2). The conversion factor between these two is uncertain. Work at the Goddard Institute for Space Studies illustrates this uncertainty. According to early work by Lacis et al. [1992] the conversion for OD at 0.55 μm (for small forcings) is 30 W m−2, according to Hansen et al. [1997] it is 27 W m−2, while according to Hansen et al. [2002] it is 21 W m−2. Results from PCM suggest a value slightly less than the Hansen et al. [2002] value. We chose a value of 20 W m−2 as a somewhat arbitrary estimate, and it is these results that are shown here. “Tuning” this value does not noticeably affect the AOGCM/MAGICC match.

[21] Another difference between the AOGCM and UD EBM experiments is in the nature of the input forcing. In the AOGCM the “forcing” is specified month by month as zonal mean loading patterns of stratospheric aerosol [Ammann et al., 2003]. In the UD EBM it is only hemispheric mean forcings that are specified, but the input is still on a monthly timescale.

[22] Figure 2 compares the ensemble mean PCM results (as shown in Figure 1c) with those for MAGICC (note the expanded vertical scale). Features in the AOGCM results that are obscured by internally generated noise become more readily identifiable when the “pure” signal is superimposed. For example, there is a short cooling episode in 1929 that coincides with the eruption of Paluweh (August 1928), and the apparently slow recovery after Agung can be seen to be an artifact of the later eruptions of Fernandina (June 1968) and Fuego (October 1974). Further, the volcano-like sharp cooling after July 1892, peaking in February 1894, can be seen to be an internal fluctuation. The most important result, however, is the very close agreement between the MAGICC and ensemble mean AOGCM results, which justifies our use of MAGICC (with different climate sensitivities) to obtain reliable estimates of how the volcanic response varies with sensitivity with some confidence.

Details are in the caption following the image
Comparison of PCM volcanic response (16-member ensemble mean) with response simulated by the model for the assessment of greenhouse gas–induced climate change (MAGICC) upwelling diffusion energy balance model. The initial (January 1890) value for MAGICC is set equal to the January 1890 value in the PCM case (0.1°C).

[23] To further quantify this agreement, we can assess the goodness of fit for the MAGICC signal (as shown in Figure 2) by subtracting this signal from each of the four AOGCM volcano-only runs and then calculating the standard deviation of the residuals. The mean of this standard deviation is 0.173°C, consistent with the mean control run variability of 0.171°C. After ensemble averaging over the four volcano-only runs the residual variability about the estimated “pure” signal is reduced to 0.090°C, very close to the expected value of 0.086°C. These results show that the fit between MAGICC and the AOGCM (with no additional tuning) is close to optimal.

5. Effect of Different Climate Sensitivities

[24] Figure 3 shows MAGICC results for climate sensitivities of 1.0, 2.0, and 4.0°C. (In MAGICC the sensitivity is specified as temperature change per unit radiative forcing, S = ΔT2x/ΔQ2x. The assumed value for ΔQ2x is 3.71W m−2, the central value used in the IPCC TAR [Cubasch et al., 2001]. The sensitivity used for PCM is 1.7°C [Raper et al., 2001].)

Details are in the caption following the image
Volcanic responses for climate sensitivities of 1.0, 2.0, and 4.0°C equilibrium warming for a CO2 doubling. Note that the initial (January 1890) value for MAGICC is set equal to zero.

[25] For maximum cooling these results show that for any given eruption the cooling depends approximately on the climate sensitivity raised to the power 0.37, with the exponent obtained by fitting a power law relationship to results spanning a wider range of sensitivity values than shown here. A similar result has been obtained by Harvey and Kaufmann [2002], who found that increasing the sensitivity by a factor of 5 increased the maximum cooling by a factor of 2 (equivalent to an exponent of 0.43). From results of LG98 (their Figure 4) the corresponding power is only 0.20. The timescale for relaxation from peak cooling back to the preeruption state is also dependent on the climate sensitivity, with slower decay for larger sensitivity (see below).

[26] A further difference from LG98 is that we find the time of peak cooling to be independent of the climate sensitivity; LG98 find that this time lags behind the time of peak forcing by 4–16 months, with greater lag for larger sensitivity. In our simulations the lag varies with eruption, ranging from 1 month (Novarupta) to 8 months (Agung) behind peak forcing (i.e., 3–13 months after the eruption date for these particular volcanoes). Differences between eruptions arise primarily because of eruption-specific forcing differences between the hemispheres.

[27] We quantified the relaxation timescale by fitting exponential decay curves for times after the maximum cooling point to the MAGICC results for the five largest eruptions, Santa Maria, Novarupta, Agung, El Chichón, and Pinatubo. In all cases the decay is slower than exponential for the first 12–16 months (only a few months for Novarupta), is well approximated by an exponential over the next 30–50 months, and is then again slower than exponential. The slow early response is a result of the initially slow removal of aerosol from the stratosphere. The later subexponential decay behavior produces a long “tail” in the response, although this is obscured in most cases by a subsequent eruption. This behavior is consistent with an increase with time in the ocean's effective heat capacity, as implied by the results of Wigley and Raper [1991]. While this behavior is clear in the EBM, it is impossible to identify it in either the observations or the AOGCM results because by the time the subexponential portion is reached, the residual cooling is invariably much less than 0.1°C and is consequently obscured by the noise of internal variability (which has a standard deviation of ∼0.17°C). For the same reason, assuming a purely exponential decay for all times provides an excellent approximation to the “true” decay curve.

[28] The best fit exponential decay times for sensitivities of 1.0, 2.0, and 4.0°C are as follows: Santa Maria, 27, 31, and 34 months; Novarupta, 17, 19, and 21 months; Agung, 29, 34, and 39 months; El Chichón, 31, 37, and 43 months; and Pinatubo, 31, 37, and 42 months. There is no statistically significant difference between these decay times except for Novarupta, where the more rapid decay is presumably because of the high-latitude Northern Hemisphere location of this volcano. (In MAGICC this geographical influence is captured by the hemispheric differential in the applied forcing and by the land-ocean and interhemispheric separations in the model.) These results are entirely consistent with the decay times assumed by Santer et al. [2001]. The decay times for temperature response are much longer than the decay times for forcing (∼12 months), a necessary consequence of the thermal inertia of the climate system. While this is rather simple physics, some recent statistical analyses of the temperature record have erroneously assumed that the decay times for volcanic forcing and response are the same [Christy and McNider, 1994; Michaels and Knappenberger, 2000; Douglass and Clader, 2002]. Conclusions drawn by such studies are suspect.

[29] The decay time's dependence on sensitivity (based on a larger sample of sensitivities than the three illustrated in Figure 3) is given approximately by
equation image
Using LG98 results (their Figure 4), the corresponding result is
equation image
LG98 therefore have much longer decay times than those implied by MAGICC, particularly for higher sensitivities. The simplicity of the LG98 model and its lack of validation against a more physically realistic model make the LG98 results suspect.

[30] Given the MAGICC results for the dependence of maximum cooling on sensitivity, it is possible, by comparing these results with observations, to estimate climate sensitivity values consistent with the observed global mean responses. (This estimation process assumes that the volcanic forcing record given by Ammann et al. [2003] is a reasonable approximation to the actual forcing that occurred over the 20th century.) We do this here using modeled and observed maximum coolings for Agung, El Chichón, and Pinatubo (it is not possible to obtain reliable maximum cooling estimates for Santa Maria or Novarupta). As noted by LG98 and others [e.g., Bradley, 1988; Robock and Mao, 1995], it is difficult to estimate the observed maximum coolings for any individual eruptions because they are obscured by other sources of natural variability on the monthly to interannual timescale. We use the estimates of Wigley [2000] here, which we reproduce to two decimals for the sake of precision (not to be confused with accuracy) as follows: Agung, 0.30°C ± 0.1°C; El Chichón, 0.24°C ± 0.15°C; Pinatubo, 0.61°C ± 0.1°C, where the plus or minus sign refers to estimated 2-sigma limits.

[31] The method used to obtain these maximum cooling estimates is an iterative procedure described by Wigley [2000] and Santer et al. [2001]. If T(t) denotes the “raw” observed temperature data, we first subtract a trial volcano series (V(t)) to give X(t) = T(t) − V(t). V(t) assumes a linear ramp to maximum cooling followed by an exponential relaxation back to the initial state (a form justified by the above model results). V(t) involves three parameters for each volcano, the time to maximum cooling, the maximum cooling, and the decay time. Second, we regress X(t) against lagged values of an ENSO index (E(t)) with the lag (usually 6 or 7 months) chosen to optimize the fit (Xest = a + bE). The ENSO sensitivity (b) is then used to remove the ENSO influence from the original data. This makes the volcano signals clearer, so improved estimates may be made for the volcano parameters. The process is repeated until stable estimates are obtained.

[32] These are the most recent maximum cooling estimates for these three volcanoes that account for the effects of ENSO, a factor that can significantly obscure the volcanic response signal [e.g., Angell, 1988; Robock and Mao, 1995; Wigley, 2000; Santer et al., 2001]. The greater uncertainty for El Chichón arises primarily from the large El Niño event that occurred at the same time in 1982/1983. The other two eruption events are less contaminated by ENSO effects and hence are less sensitive to the details of the ENSO removal procedure.

[33] From the maximum coolings obtained in this way the implied climate sensitivities are as follows: Agung, 1.28 (2.83) 6.32°C; El Chichón, 0.30 (1.54) 7.73°C; and Pinatubo, 1.79 (3.03) 5.21°C. The central numbers here correspond to the best estimate cooling values, while the other numbers correspond to the 2-sigma maximum cooling limits. The uncertainties in estimating the observed maximum cooling and the relative insensitivity of volcanic responses to the value of the climate sensitivity combine to leave large uncertainties in the implied climate sensitivity values. Forcing uncertainties imply even larger sensitivity uncertainties than shown here. Nevertheless, the results are consistent with other empirical sensitivity estimates (e.g., Andronova and Schlesinger [2001]; Forest et al. [2002]; Gregory et al. [2002]; Harvey and Kaufmann [2002], studies that are fraught with uncertainties arising from uncertainties in past forcings on decadal and longer timescales), AOGCM-based estimates [Cubasch et al., 2001; Raper et al., 2001], and the sensitivity range endorsed by IPCC. We note that the cooling associated with Pinatubo appears to require a sensitivity above the IPCC lower bound of 1.5°C and that none of the observed eruption responses rules out a sensitivity above 4.5°C.

[34] There are a number of reasons why it is not possible to follow the LG98 strategy and use the long-timescale behavior of eruptions (i.e., times greater than a few years after peak cooling) to obtain additional insight into the value of the climate sensitivity. First, volcanic forcing is only one of many forcing agents that have acted over the 20th century, and the effects of these other forcings and their uncertainties are greater than any residual long-term volcanic signal that may exist. Second, the residual cooling in the long-timescale “tails” that we obtain (Figure 3) is much less than illustrated by LG98, who have significantly longer decay times. In all cases on the basis of our results this residual cooling is sufficiently small that it must, in general, be lost within the noise of natural internally generated variability. Third, the absolute separation between simulations with different climate sensitivities, at least after the first few years, decreases with time, making the task of their separation from the noise and from each other more and more difficult as time progresses.

6. Conclusions

[35] We have defined the response to 20th century volcanic forcing on the basis of simulations with the NCAR/USDOE parallel climate model (PCM). In total, there are 16 simulations that include volcanic forcing. These multiple realizations allow us to reduce the noise due to internally generated variability by 65% and to produce a much more well defined volcano response signature than can be seen in a single realization.

[36] We then compared the AOGCM results with results obtained using a simple upwelling diffusion energy balance model. Model parameters for the UD EBM were chosen independently of the volcano simulations, using the results from 1% compound CO2 increase experiments. The agreement between the AOGCM and UD EBM volcanic forcing results was excellent (see Figure 2), justifying the use of the UD EBM to determine the volcanic response for different climate sensitivities. The UD EBM results showed the maximum cooling for any given eruption to depend approximately on the climate sensitivity raised to the power 0.37. We also found the timescale for relaxation back to the preeruption state to depend on the climate sensitivity, with slower decay for larger sensitivity, and the time of peak cooling for any given eruption to be independent of the climate sensitivity.

[37] We quantified the relaxation timescale by fitting exponential decay curves to the UD EBM results for the five largest eruptions, Santa Maria, Novarupta, Agung, El Chichón, and Pinatubo. Assuming a purely exponential decay provides an excellent approximation to the “true” decay curve. The best fit exponential decay times for climate sensitivities of 1.0–4.0°C range from 27 to 43 months (excluding Novarupta), with the decay time given approximately by τ = 30(ΔT2x)0.23.

[38] By comparing the UD EBM results with observations we estimated the climate sensitivity values consistent with the observations for Agung, El Chichón, and Pinatubo. The central sensitivity estimates are near 3°C for Agung and Pinatubo. For El Chichón the central estimate is 1.5°C, but because of the potentially larger errors in separating the volcanic signal from the near-contemporary El Niño in this case this estimate is much more uncertain. These results are consistent with other empirical sensitivity estimates and with the sensitivity range endorsed by IPCC. While it is not possible to use eruption data to narrow the sensitivity uncertainty range, the observed cooling associated with Pinatubo (the best observed volcanic eruption) appears to require a sensitivity above the IPCC lower bound of 1.5°C, while none of the observed eruption responses rules out a sensitivity above 4.5°C.

[39] Our results differ significantly from those of Lindzen and Giannitsis [1998]. These authors conclude that the observations favor a low value for the climate sensitivity on the basis of the long-timescale response to eruptions. The main reason for this difference is because the long-timescale response that we obtain, using physically more comprehensive and realistic models, is substantially less than that obtained by Lindzen and Giannitsis. The long-timescale response that we obtain is so small that it would be overwhelmed by the effects of other (uncertain) forcing factors and is certainly too small to be identified in the observational record above the noise of internally generated variability. The larger Lindzen and Giannitsis result is an artifact of their use of an oversimplified model. Our conclusion therefore is twofold: While useful information on the climate sensitivity can be obtained from the short-timescale responses to individual volcanic eruptions, it is unlikely that meaningful quantitative results can be obtained from the long-timescale responses to such forcing.

Acknowledgments

[40] This research was supported by NOAA Office of Climate Programs (“Climate Change Data and Detection”) grant NA87GP0105 and U.S. Department of Energy (DOE) grant DE-FG02-98ER62601. NCAR is supported by the National Science Foundation.