3.1. Budget of Injected Sea Salt Particles
 The mass emissions and the approximate burden of sea salt from both natural and artificial sources in the different simulations are presented in Table 2. Natural global sea salt emissions were approximately 5500 Tg yr−1 in all simulations with a variation of about 1% due to different wind conditions. Although the contribution of the artificial emissions to the total emitted sea spray mass was very small (0.4–7.4% depending on the simulation), they were responsible for 4–88% of the total number emissions of sea spray. As expected, multiplication of the baseline source function in simulations 3 × GEO and 5 × GEO resulted in a nearly linear increase in global mean emissions. For example in GEO the annual injected sea salt mass was 20.6 Tg yr−1 and in 3 × GEO 62.1 Tg yr−1. When geoengineering emissions were applied over all oceans (ALL), the sea salt burden was 27% higher than in the control simulation (CTRL); in all other geoengineering simulations the burden increased by less than 10%.
Table 2. The Simulated Budget of Sea Salt Particles From Both Natural and Artificial Sourcesa
|Simulation||Natural Emissions (Tg yr−1)||Artificial Emissions (Tg yr−1)||Burden (Tg)||Burden Anomaly (Tg)|
|3 × GEO||5510.6||62.1||15.0||0.9|
|5 × GEO||5467.3||103.0||15.5||1.4|
 Figure 3apresents the difference in total sea salt burden between the simulations GEO and CTRL, which is used as an estimate of the sea salt mass from artificial emissions. Note that since the ECHAM5.5-HAM2 model does not differentiate between sea spray particles from natural and artificial sources, the time-averaged global burden of the injected particles cannot be calculated exactly. In the baseline simulation GEO, the total sea salt burden was 0.2 Tg higher than the total sea salt burden of 14.1 Tg in the CTRL simulation (Table 2). This is an upper estimate for the sea salt burden from geoengineering injections since the life-time of the background particles was probably higher in the simulation GEO than in the simulation CTRL due to decreased scavenging by wet deposition. Within the geoengineered regions, the sea salt column burden from artificial sources was far from uniform and varied by roughly an order of magnitude (Figure 3a). Highest column burdens were found near the coast of South America and further away from the coasts of North America and Africa. The inhomogeneous distribution was caused primarily by variation in the wind speed dependent injection flux (Figure 3b) and transport of sea salt, but also by microphysical processes such as scavenging by wet deposition.
Figure 3. (a) The 10-year mean difference in total sea salt burden between the simulations GEO and CTRL and (b) the 10-year mean of sea salt number injections in the simulation GEO.
Download figure to PowerPoint
 As can be assumed based on the small size of injected particles, they had a longer life-time than the natural sea spray particles on average. While the exact life-time of injected aerosol cannot be calculated as the model does not differentiate between particles from different sources, the life-time can be estimated by dividing the difference in sea salt burden between CTRL and geoengineering simulations with the total mass flux from artificial emissions. The mean life-time estimate of the injected sea salt particles was between 3.1 (ALL) and 5.2 days (3 × GEO) in the geoengineering simulations while the mean life-time of sea salt in simulation CTRL was only 0.9 days. The long life-time of the injected particles caused the relative increase in total sea salt burden to be higher than the relative increase in total sea salt emissions.
 It is also noteworthy that in our baseline case GEO, the mean number flux of artificial sea salt emissions was 57% of the flux assumed in the cloud-system-resolving model study ofWang et al. . In our simulations 3 × GEO and 5 × GEO the corresponding number flux was 73% and 187% higher than that in the study by Wang et al. .
3.2. Effect of Injection Rate on Clouds
 Figure 4 presents the mean cloud top CDNC in the CTRL simulation without artificial sea spray emissions together with the relative CDNC increase in the geoengineering simulations GEO, 3 × GEO and 5 × GEO. Here the mean cloud droplet number concentration for each simulation was calculated by sampling only over cloudy time steps. In the marine clouds in CTRL simulation (Figure 4a), the mean CDNC at cloud top varied between 20 and 424 cm−3 with a mean of 100 cm−3.
Figure 4. The 10-year mean cloud droplet number concentration at cloud top: (a) absolute values in CTRL, and relative increase (with respect to CTRL) in (b) GEO (c) 3 × GEO and (d) 5 × GEO. Note that graphs have different scales.
Download figure to PowerPoint
 The mean cloud top CDNC in the three geoengineered regions are given in Table 3. The simulated background CDNCs in the optimal geoengineering regions in the simulation CTRL (regional means 112–163 cm−3) are consistent with in situ measurements made in marine stratocumulus clouds off the Chilean coast during the VOCALS-Rex campaign which ranged between 80 and 400 cm−3 [Zheng et al., 2011]. On the other hand, the model-predicted CDNCs in the CTRL simulation tend to be somewhat higher than those from MODIS satellite retrievals [e.g.,Quaas et al., 2006]. The MODIS 5-year-mean CDNC (1 March 2000–28 February 2005) from retrieval byQuaas et al.  for our North Pacific region was 84 cm−3 (versus 112 cm−3 in the model), for the South Pacific 93 cm−3 (versus 132 cm−3) and for the South Atlantic 97 cm−3 (versus 163 cm−3).
Table 3. Mean Values of Cloud Droplet Number Concentration (CDNC) at Cloud Top, Its Relative Change Compared to CTRL (ΔCDNC), Effective Radius of Cloud Droplets (reff) at Cloud Top and Liquid Water Path (LWP)a
| ||CTRL||GEO||3 × GEO||5 × GEO||SMALL GEO||LARGE GEO|
|LWP (g m−2)||99||151||219||261||296||103|
|LWP (g m−2)||107||154||225||253||287||111|
|LWP (g m−2)||75||120||187||229||247||81|
 There are several possible explanations for the difference in CDNC between model simulations and satellite measurements. First, the simulated aerosol fields from which the CDNC are calculated online may be too high. Unfortunately, aerosol data from satellites and in situ measurements are insufficient for a conclusive comparison of the modeled aerosol concentrations against observations. In the regions where the modeled aerosol concentrations were too high in the CTRL simulation, we would be likely to underestimate the efficiency of sea spray cloud seeding. Second, the simulated mean updraft velocities in the optimized regions were very high (1.0–1.4 m s−1). In field measurements, typical updraft velocities range between 0.2 and 0.4 m s−1, although much higher as well as lower values have also been measured [Lu et al., 2009; Meskhidze at al., 2005]. In the light of these measurements, the updraft velocities in ECHAM5.5-HAM2 may lead to overestimation of activated cloud droplets. This model feature can causeunderestimation of cloud seeding efficiency since it is likely to overestimate the CDNC from background particles more than that from the injected particles (which are mostly large enough to activate at any reasonable updraft). Third, CDNCs retrieved from remote sensing observations suffer from many uncertainties. CDNC retrievals are based on the cloud optical thickness (COT) and cloud droplet effective radius reff which are lower order cloud properties retrieved from remote sensing observation. CDNC retrieval is especially sensitive to changes in reff, for which the values differ significantly between different retrieval algorithms and remote sensing instruments [e.g., Bennartz, 2007; Breon and Doutriaux-Boucher, 2005; Maddux et al., 2010] translating to large uncertainty in CDNC.
 Figures 4b–4d show that in the simulation GEO, the artificial sea salt injections changed the cloud droplet concentrations substantially but with a high spatial variation. The highest relative increases in CDNC were situated close to the coastline. For example, close to the California coast in the North Pacific region, the relative increase was in the range of 90–150% while further to the ocean it was only in the range of 30–70%. The likely reason for the strong response near the coasts is discussed in later in this section. Although the relative increases of cloud top CDNC were higher in 3 × GEO and 5 × GEO, the spatial pattern of increase was similar in all three geoengineering simulations. It is noteworthy that the effect of artificial sea salt emissions extended also outside the emission regions, and in some areas as far as 1500 km away from emission regions the CDNC at cloud top increased by 10% in the simulation GEO.
 Table 3 summarizes the mean relative changes in cloud top CDNC in the three optimized regions. In simulation GEO, the changes in CDNC for North Pacific, South Pacific and South Atlantic regions were 74, 80 and 75%, respectively. It could be expected that the increase in CDNC would be sublinear with respect to the magnitude of the emission flux. However, in our simulations the increase of CDNC followed the multiple of the baseline mass flux superlinearly: averaged over all optimized regions, CDNC increased in GEO by 75%, in 3 × GEO by 253% and in 5 × GEO by 408%. There are several reasons for this superlinearity. First, the particle concentration in the soluble accumulation mode increased superlinearly. This was probably due to weakened scavenging by wet deposition as the mean precipitation decreased in the emission regions by 2–5% in most simulations (second indirect effect). Second, vertical velocities were higher with higher artificial sea salt fluxes. Third, due to the modal aerosol description, the mean diameter of the accumulation mode increased more (compared to CTRL) in 3 × GEO and in 5 × GEO than in GEO, which makes activation to cloud droplets more probable.
 When assuming the baseline flux GEO, the mean absolute values of CDNC in the emission regions remained below 375 cm−3 (Table 3), which is a value assumed in several previous climate model studies of sea spray geoengineering [Latham et al., 2008; Jones et al., 2009]. However, in simulation 3 × GEO the mean CDNC at cloud top over all emission regions was 458 cm−3 and in 5 × GEO it was 658 cm−3. Even these values are significantly lower than the 1000 cm−3 assumed by Rasch et al. .
 The strong response in CDNC in the three optimized regions can be explained by their relatively low background CDNC (regional mean values in CTRL 112–163 cm−3, Table 3) and the stratocumulus clouds that reside at heights of 100–500 m (Figures 5d–5f). Since the strongest effect of sea spray injections was limited to the lowest ∼2 km (Figures 5a–5c), these low-lying clouds are an ideal target for cloud whitening. These low clouds also occur frequently, which explains the strong radiative flux perturbation in these regions when all oceanic regions are geoengineered (simulation ALL,Figure 2). Note, however, that frequent cloud occurrence was not the cause of the high mean absolute CDNC values in these regions as only cloudy time steps were considered.
Figure 5. (a–c) 10-year mean regional mean cloud droplet number concentration at different altitudes in three emission regions. Mean values are sampled only over cloudy time steps. (d–f) Regional mean cloud cover profiles for three emission regions. Note that profiles of the simulations CTRL and LARGE GEO almost overlap.
Download figure to PowerPoint
 Our study shows a much stronger cloud response to sea spray injections than the previously published study by Korhonen et al.  who assumed similar number injection fluxes (increase in CDNC is mainly dependent on number instead of mass flux, provided that injected particles are large enough). Korhonen et al.  simulated a regional mean increase of only 20% or less with a flux comparable to GEO (mean in this study: 75%) and of 163% or less with a flux comparable to 5 × GEO (mean in this study: 408%). Several factors may contribute to this difference.
 First, the regional mean background CDNCs in the control simulation without geoengineering were somewhat higher in the previous study (143–177 cm−3) compared to this study (112–163 cm−3).
 Second, Korhonen et al. used a chemical transport model with prescribed meteorology and could therefore not include cloud feedbacks on the aerosol concentration. In our study these feedbacks are simulated, and they increase the life-time of the injected particles and thus the change in the CDNC. The increase in particle life-time is evident e.g., from the superlinear increase in the particle number concentration in the soluble accumulation mode when the magnitude of the artificial sea salt emissions is increased (not shown).
 Third, the reference altitudes for CDNC in these two model studies were not exactly comparable. While Korhonen et al. calculated their CDNC fields offline at an approximate cloud base altitude of 1 km, ECHAM5.5-HAM2 indicates that stratocumulus clouds in the seeding regions are frequently found at 100–500 m altitude (Figure 5). We recalculated the CDNC fields using the model results of Korhonen et al.  at 360 m altitude and found that the relative CDNC change increased by only up to 5 percentage points in the simulation corresponding to GEO and by up to 40 percentage points in the simulation corresponding to 5 × GEO. Thus, the differences in cloud altitude alone cannot explain the differences in CDNC enhancements between the two studies.
 Fourth, Korhonen et al. simulated a significant suppression of supersaturation when the artificial emissions were present. This meant that some of the smaller background particles forming cloud drops in their control simulation did not activate to droplets in the geoengineering simulations, thus reducing the relative CDNC increase. This effect is lacking in our study, since ECHAM5.5-HAM2 simulates much higher effective updrafts (1.0–1.4 m s−1) than assumed by Korhonen et al.  (Gaussian distribution with mean of 0 m s−1 and standard deviation of 0.25 m s−1) and thus the small background particles that activated in our simulation CTRL, activated also in the geoengineering simulations. Since the updraft velocities assumed by Korhonen et al.  are lower than in typical measurements [Lu et al., 2009; Meskhidze at al., 2005], it is possible that Korhonen et al.  overestimated the supersaturation suppression effect. However, recalculating their CDNC fields using a significantly higher updraft of 0.8 m s−1 increased the predicted relative CDNC changes by only up to 13 percentage points in the simulation corresponding to GEO and decreased the CDNC change in the simulation corresponding to 5 × GEO (because at high updraft smaller background particles activated to droplets). Therefore, the differences in updraft velocities alone cannot explain the differences in CDNC enhancements between the two studies.
 Another difference between this study and that of Korhonen et al.  is that they found the maximum increase in relative CDNC further off the coast over the open ocean. Korhonen et al.  explained their result with the high background concentration of anthropogenic aerosols buffering the relative CDNC change close to the continents. The anthropogenic effect on aerosol fields is also visible in our simulations and thus the difference between the two studies cannot be explained by different emissions.
 The most likely reason for this discrepancy has to do with the structure of a modal aerosol description such as HAM2. Due to anthropogenic pollution, the dry geometric mean diameter of the soluble accumulation mode in the control simulation (CTRL) was smaller in the nearest 500 km from the coasts than further out over the ocean but within the geoengineered regions (approximately 140–160 nm and 180–210 nm, respectively) and the organic fraction is higher close to the continents. Because the injected particles have a dry diameter of 250 nm and a higher solubility than the background particles, the mean diameter of the accumulation mode as well as the particle solubility increased close to the continents in the geoengineering simulations relative to the CTRL simulation. Since the standard deviations of the modes are fixed in HAM2, this means that the size and solubility of the background accumulation mode particles is somewhat overestimated. Because of this, some of the background particles that did not activate in the CTRL simulation form cloud droplets in the geoengineering simulations. As a result, modal aerosol models such as ECHAM5.5-HAM2 tend to overestimate the CDNC in geoengineering simulations in anthropogenically influenced regions and thus our results on cloud seeding efficiency in these regions are likely to be an upper estimate. It should be noted thatKorhonen et al.  used a sectional aerosol model in their study and thus did not suffer from this effect.
 Figure 6a shows that artificial sea salt emissions caused also an increase in cloudiness (second indirect effect). For example, in the simulation GEO the mean total cloud cover over the North Pacific region increased from 59.4% to 63.3%, over the South Pacific from 59.3% to 61.7% and over the South Atlantic from 46.4% to 50.8%. The areas with the strongest increase in total cloud cover were roughly the same ones which had the highest increase in CDNC (Figures 4b–4d) and liquid water path (LWP) (not shown). The mean LWP in the optimized regions increased by 44–61% in GEO compared to CTRL (Table 3). It should be noted, however, that modeling the second indirect effect with global climate models has still many uncertainties. For example, compared to satellite observations, climate models tend to overestimate the effect of aerosol number concentration increase on the LWP, and underestimate the correlation between aerosol optical depth and cloud fraction [Quaas et al. 2009].
Figure 6. (a) Change of total cloud cover (percentage points) and (b) relative change (%) of cloud droplet effective radius at cloud top in GEO compared to the simulation CTRL. The 10-year mean values for total cloud cover and cloud droplet effective radius were used.
Download figure to PowerPoint
 The first indirect effect, i.e., the decrease in effective radii of the droplets when their number concentration increases, was also evident in all simulations (Figure 6b and Table 3). In GEO, the decrease in cloud top effective radius was 9% when all seeding regions were considered. As an example, in the North Pacific region the mean effective radius was 13.5 μm in CTRL and 12.0 μm in GEO. Higher injection rates further decreased the effective radii of the cloud droplets: in 3 × GEO the mean decrease over the three emission regions was 15% and in 5 × GEO 16%. However, there was significant spatial variation within and between the emission regions (Figure 6b). The strongest effect was found in the North Pacific region where in simulation 5 × GEO the mean decrease in cloud top effective radius was over 4 percentage points higher compared to the other two regions.
3.3. Effect of Particle Injection Size on Clouds
 If the mass flux of seawater sprayed from the vessels is limited by technological constraints, the size of the injected aerosol particles plays an important role. In order to investigate the effect of injection size on the cloud seeding efficiency, we conducted two additional simulations with different geometric mean diameters for the injected particles: in simulation SMALL GEO, the diameter was set to 100 nm, and in simulation LARGE GEO to 500 nm. Note that the total mass flux in both of these simulations was the same as in GEO, and thus the number flux increased by 1460% in SMALL GEO and decreased by 88% in LARGE GEO compared to the simulation GEO.
 Despite the small size of the injected particles in the simulation SMALL GEO, most of them activated to cloud droplets leading to a mean cloud top CDNC of 1046 cm−3 and a mean relative increase of 707% in the three emission regions. This relative increase was notably higher than the CDNC increase in GEO (75%) or even in 5 × GEO (408%), although it did not follow the magnitude of the number flux linearly (number flux in SMALL GEO was 1460% higher than in GEO) as it did in simulations GEO, 3 × GEO and 5 × GEO. The high efficiency of SMALL GEO can be seen also in Figures 5a–5c, which show that CDNC was about 1500 cm−3 at altitudes typical for stratocumulus clouds.
 These results indicate that decreasing the particle injection size can in many situations be a much more effective way to improve the cloud seeding efficiency than increasing the seawater mass flux from the spraying vessels. It must be remembered, however, that the injection size cannot be decreased much below 100 nm if one wants to be sure that the particles activate to cloud droplets in typical stratocumulus updrafts. It should be also noted, that the large updraft velocities in ECHAM5.5-HAM2 may bias the number of activated droplets high in the SMALL GEO run in which the injected particles are small.
 Table 3 shows that the changes in cloud droplet effective radius and LWP were similar in SMALL GEO and 5 × GEO. The regional mean effective radii at cloud top in SMALL GEO were 10.7–11.2 μm, which is 1.6–2.5 μm smaller than in the CTRL simulation. On the other hand, the regional mean LWP in SMALL GEO varied between 247 g m−2 and 296 g m−2 while the corresponding values for CTRL were between 75 g m−2 and 107 g m−2.
 Increasing the geometric mean diameter to 500 nm in the simulation LARGE GEO decreased the artificial sea salt number flux by 88% compared with that in GEO. As expected, this decrease was seen in the much lower CDNC increase compared to other geoengineering simulations. Cloud top CDNC increased by only 18% in the South Atlantic region and even less in the other two emission regions (Table 3). Correspondingly, the mean vertical profiles of CDNC in LARGE GEO (Figures 5a–5c) were almost identical to profiles in CTRL. Changes in effective radius and LWP were also almost negligible in LARGE GEO compared to the CTRL simulation (Table 3).
3.4. Radiative Effects
 The global and regional mean values of radiative flux perturbation (RFP), aerosol direct effect (difference in aerosol direct forcing of all atmospheric particles between a geoengineering and the control simulation) and aerosol indirect effect (the difference of RFP and the direct effect) for all the simulations are summarized in Table 4. The calculation of direct and indirect effects is explained in more detail in section 2.4. While the direct and indirect effects cannot be unambiguously separated, the chosen method can be used to investigate the approximate relative contributions of aerosol direct and indirect effects to the total radiative effects.
Table 4. Global and Regional Mean Values of the Radiative Effectsa
| ||ALL||GEO||3 × GEO||5 × GEO||SMALL GEO||LARGE GEO|
 When the artificial sea salt emissions were limited only to the three optimized regions, the strongest RFP was achieved in the 5 × GEO simulation (global mean RFP −2.2 Wm−2) followed by SMALL GEO (−2.1 Wm−2). This is slightly surprising because the relative increase of the mean CDNC over the optimal regions was clearly higher in the simulation SMALL GEO (707% versus 408% in 5 × GEO). This apparent discrepancy is due to the much larger direct aerosol effect in the 5 × GEO simulation (−0.5 Wm−2 versus −0.1 Wm−2 in SMALL GEO) and was mainly caused by the larger mass flux in the simulation 5 × GEO. Furthermore, the indirect effects started to saturate at high CDNC values.
 As expected, the regional radiative effects were much stronger than the global mean effects in all our simulations. For example, in simulation SMALL GEO (global mean RFP −2.1 Wm−2) the regional means were −35.9 Wm−2, −40.3 Wm−2 and −36.0 Wm−2 for the optimized North Pacific, South Pacific and South Atlantic regions, respectively. It is noteworthy that such high local forcings could have significant impacts on the atmospheric dynamics as well as on local marine ecosystems.
 When the sea spray injections were limited to the three optimized regions, none of our simulations could produce high enough RFP to counteract the doubling of carbon dioxide concentrations from the pre-industrial era (estimated forcing +3.7 Wm−2 [Forster et al., 2007]). However, when the injections were extended over all oceans (ALL), the global mean RFP was −5.1 Wm−2 which would be more than enough to compensate for the CO2 doubling.
 Our results can be compared against those of Jones et al. , who modified clouds over an area equivalent to our optimized areas (3.3% of Earth's surface) and obtained a global mean RFP of −0.97 Wm−2. This is comparable to the RFP in our GEO simulation (−0.8 W m−2). Our GEO simulation predicts clearly lower CDNC (regional means varied between 194 and 286 cm−3) than the 375 cm−3 assumed by e.g., Jones et al. but an increase in cloud cover by 2–5 percentage points. Furthermore, we predict also a non-negligible direct effect (−0.1 W m−2) which was omitted by Jones et al. . Thus, in our study the aerosol direct effect compensated to some extent for the lower indirect effects.
 In all the simulations with injections restricted to the three optimal regions (apart from the run LARGE GEO), the aerosol indirect effects dominated over the aerosol direct effects. The extreme example was SMALL GEO, for which the global mean direct effect was only −0.1 Wm−2 compared to the indirect effect of −2.1 Wm−2. However, in most simulations the absolute values of the regional mean direct effect were quite significant. For example, in simulation GEO they were −1.2, −1.1 and −0.8 Wm−2 for the North Pacific, South Pacific and South Atlantic regions, respectively. However, these values are still much lower than the corresponding regional mean values of the indirect effect (−14.5, −15.2 and −14.0 Wm−2, respectively). It is also worth noting that the global mean direct effect depended nearly linearly on the mass flux of artificial sea salt emissions, as in simulation 3 × GEO it was 3.1 times and in simulation 5 × GEO 4.6 times that of GEO. On the other hand, the indirect effect was clearly sublinear, being −0.7 Wm−2 in GEO and −1.7 Wm−2, i.e., only 1,4 times larger, in 5 × GEO.
 Note that for the South Atlantic region the calculated mean direct effect in the simulation SMALL GEO was +0.7 Wm−2 (Table 4). This positive value is an example of how changes in cloud cover can affect the direct effect calculated by the difference of total aerosol forcing between two different simulations (see section 2.4). In this region, the aerosol optical depth (AOD) was 87% higher in the simulation SMALL GEO compared to CTRL. Furthermore, the direct effect using the clear-sky values was −4.2 Wm−2, which shows that the artificial sea salt emissions had a cooling effect, although the calculated total-sky direct effect was positive due to highly increased cloud cover.
 Figure 7 shows the geographical distributions of the direct and indirect effects for the simulations ALL and GEO. Areas with the strongest direct and indirect effect do not overlap since in regions with persistent cloud cover the aerosol direct effects are of minor importance. For example, on the coast of Africa the strongest direct effect was found on the western edge of the emission region, but the strongest indirect effect next to the coast where low clouds occur more frequently. Generally, the geographical distribution of the direct effect in the simulation GEO (Figure 7b) was similar to the estimate for the burden of sea salt originating from artificial emissions (Figure 3b).
Figure 7. The 10-year mean radiative effects in the simulations ALL and GEO. (a) The aerosol direct effect in the simulation ALL and (b) the aerosol direct effect in the simulation GEO. (c) The aerosol indirect effect calculated as the difference between radiative flux perturbation (RFP) and the aerosol direct effect in the simulation ALL and (d) the corresponding aerosol indirect effect for the simulation GEO. Note that direct and indirect effects have different scales.
Download figure to PowerPoint
 Unlike in the other simulations, in simulation ALL the aerosol direct effect was a significant part of the total radiative effect. The global mean direct effect was −1.5 Wm−2 while the indirect effect was −3.6 Wm−2. Between latitudes 16°N and 5°S, the direct effect was even stronger than the indirect effect: the zonally averaged direct effect over ocean was about −3 Wm−2 compared to the indirect effect of about −2 Wm−2. There are two main reasons for this high relative importance of direct effect in ALL compared to the other simulations. First, large parts of the ocean area have either a low total cloud cover or only few low-altitude clouds (e.g., close to the equator) that can be substantially affected by the sea salt injections. In these regions, the simulated aerosol indirect effects were fairly unimportant. Second, in the other simulations all the emission areas were highly clouded and thereby the aerosol direct effect was relatively unimportant compared to aerosol indirect effects. In the simulation ALL, these stratocumulus regions covered only a small fraction of the total emission area.
3.6. Sensitivity of the Results to Ultrafine Sea Salt Emissions
 As discussed in section 2.3, one limitation with the ECHAM5.5-HAM2 model is that it does not include natural ultrafine sea spray emissions (dp < 100 nm) which can contribute significantly to cloud condensation nuclei especially in remote marine areas [Pierce and Adams, 2006]. It is therefore possible that our model underestimates the background aerosol concentration and thus overestimates the relative CDNC increase and radiative forcing in the geoengineering simulations. In order to estimate the sensitivity of our results to the lack of ultrafine sea spray emissions, we made two additional simulations (UF-SS CTRL and UF-SS ALL), which included a simplified implementation ofMårtensson et al.  parameterization for ultrafine sea salt emissions (described in section 2.3.2) but were in other respects identical to simulations CTRL and ALL, respectively.
 With the ultrafine sea salt emissions, the regional mean cloud top CDNCs in simulation UF-SS CTRL were 128, 153 and 198 cm−3for the optimized regions in the North Pacific, South Pacific and South Atlantic, respectively. These values are on the average about 20% higher than in simulation CTRL. The geographical pattern of the relative difference between UF-SS CTRL and CTRL is similar to increase of CDNC due to geoengineering. The strongest enhancement in background CDNC is seen in the polar regions and in the three stratocumulus regions where low-altitude clouds are abundant.
 Figures 9a and 9bshow the relative increase of cloud top CDNC in simulations ALL (with respect to CTRL) and UF-SS ALL (with respect to UF-SS CTRL).Figure 9c shows the difference in the percentage change of CDNC due to geoengineering between the simulation with ultrafine sea spray included and the corresponding simulation with ultrafine sea spray not taken into account as described by the following formula:
where UF-SS denotes that natural ultrafine sea salt emissions were included in the simulation. The measure of this quantity is percentage points. Thus,Figure 9c shows how much the ultrafine sea salt emissions affect the efficiency of cloud modification. In Figure 9c, negative values indicate that introducing ultra-fine sea spray decreases the efficiency of geoengineering and positive values that ultra-fine sea spray enhances the effects of geoengineering.
Figure 9. The effect of ultrafine sea salt emissions on cloud droplet number concentration (CDNC) at cloud top, when artificial sea salt emissions are placed over all ocean area. (a) Relative change (%) between geoengineering simulation without natural ultrafine sea salt emissions (ALL) and control simulation (CTRL). (b) Relative change (%) between geoengineering simulation with natural ultrafine sea salt emissions (UF-SS ALL) and control simulation with natural ultrafine sea salt emissions (UF-SS CTRL). (c) The sensitivity of CDNC change to natural ultrafine sea salt emissions (equation (3)). The unit in Figure 9c is percentage points. CDNC was sampled over cloudy time steps over the 10-year simulation time.
Download figure to PowerPoint
 By and large the CDNC enhancements in the simulations with and without ultrafine sea salt emissions look very similar (Figures 9a and 9b). For the majority of the globe, the sensitivity of CDNC change to ultrafine emissions was less than 15 percentage points (Figure 9c). However, some exceptions exist. For example, close to the South Pole (55–62 °S) the difference between the two set-ups was quite large with zonally averaged CDNC enhancement about 20–40 percentage points lower when ultrafine sea salt was included. This is probably due to the very low natural CDNC in the Southern Ocean: as the natural background concentration was smaller in CTRL than in UF-SS CTRL, a high relative increase was also easier to achieve in the former simulation.
 In the three optimized geoengineering regions, the CDNC change was not highly sensitive to ultrafine sea salt. The largest effect was found in the South Atlantic region where the increase in CDNC was 20 percentage points smaller in the simulation with ultrafine sea salt emissions. In the North and South Pacific, the changes in CDNC enhancement were only −3 and −5 percentage points, respectively. Compared to the total CDNC enhancement in geoengineering simulations (regional means 84–105% in simulation ALL), these changes are not very significant.
 However, even these relatively small changes in CDNC had a notable effect on the radiative flux perturbation. Global mean RFP in UF-SS ALL (compared to UF-SS CTRL) was −4.5 Wm−2 while it was −5.1 Wm−2in ALL (compared to CTRL). The regional difference in RFP between UF-SS and standard simulations was largest in the South Atlantic region (5.7 Wm−2), where the CDNC increase due to geoengineering was strongest. On the other hand, the ultrafine sea salt emissions had almost negligible effect on aerosol direct forcing due to their small size. In the Arctic (60°N–90°N) the addition of ultrafine sea salt emissions had little effect on summertime RFP. The summertime (JJA) mean RFP was −2.1 Wm−2in both ALL and UF-SS ALL.
 All in all, these sensitivity simulations indicate that our results presented in previous sections are not highly sensitive to excluding natural ultrafine sea salt emissions. One possible exception is the polar regions which have low background aerosol concentrations in the accumulation mode.