Occurrence of pristine aerosol environments on a polluted planet

Significance

Uncertainty in aerosol forcing of climate since the preindustrial era hampers efforts to quantify the sensitivity of global temperature to radiative perturbations caused by human activity. Because forcings are referenced to preindustrial conditions, a large part of the uncertainty will be reduced only by accurately defining pristine aerosol conditions before air pollution. We show that pristine conditions should still be observable on a few days per month in many regions of the Earth. However, pristine cloudy regions, which are of most importance for forcing uncertainty, occur almost entirely in the Southern Hemisphere. Reduction in uncertainty of predominantly Northern Hemisphere forcing may therefore have to rely on measurements from a different hemisphere, which will limit the extent to which uncertainties can be reduced.

Abstract

Natural aerosols define a preindustrial baseline state from which the magnitude of anthropogenic aerosol effects on climate are calculated and are a major component of the large uncertainty in anthropogenic aerosol−cloud radiative forcing. This uncertainty would be reduced if aerosol environments unperturbed by air pollution could be studied in the present-day atmosphere, but the pervasiveness of air pollution makes identification of unperturbed regions difficult. Here, we use global model simulations to define unperturbed aerosol regions in terms of two measures that compare 1750 and 2000 conditions—the number of days with similar aerosol concentrations and the similarity of the aerosol response to perturbations in model processes and emissions. The analysis shows that the aerosol system in many present-day environments looks and behaves like it did in the preindustrial era. On a global annual mean, unperturbed aerosol regions cover 12% of the Earth (16% of the ocean surface and 2% of the land surface). There is a strong seasonal variation in unperturbed regions of between 4% in August and 27% in January, with the most persistent conditions occurring over the equatorial Pacific. About 90% of unperturbed regions occur in the Southern Hemisphere, but in the Northern Hemisphere, unperturbed conditions are transient and spatially patchy. In cloudy regions with a radiative forcing relative to 1750, model results suggest that unperturbed aerosol conditions could still occur on a small number of days per month. However, these environments are mostly in the Southern Hemisphere, potentially limiting the usefulness in reducing Northern Hemisphere forcing uncertainty.

Natural aerosol and precursor gas emissions are an important part of the climate system (1, 2). Improved understanding of how natural emissions determine aerosol concentrations in different environments is important for reducing the uncertainty in model estimates of cloud radiative forcing over the industrial period (3, 4). Even under the assumption that natural emissions do not change with time, the magnitude of preindustrial (PI) to present-day (PD) aerosol−cloud forcing is very sensitive to natural emissions and processes in the PI because of the nonlinear relationship between aerosol concentrations, cloud drop concentrations, and cloud albedo (3–7). Other more complex cloud adjustments are also likely to respond sensitively to small changes in aerosol under clean conditions (6, 8). Improved understanding of the PI “baseline aerosol” is necessary to reduce the uncertainty in aerosol−cloud forcing estimates, hence the total anthropogenic radiative forcing and thereby also the climate sensitivity which depends on it (4, 9). Natural emissions may also change in the future due to climate change (1), causing a natural aerosol radiative feedback on climate (10). To reduce these large climate model uncertainties, it is important to know where on Earth we can study aerosols in environments that most closely resemble a natural, unperturbed state (4, 11).

It has been argued that regions in which aerosols are unperturbed by air pollution no longer exist in today’s atmosphere (12). If this were so, then much of the uncertainty in indirect forcing may be irreducible (4). Even in regions dominated by natural emissions of sea spray, volcanic sulfates, marine dimethyl sulphide, or terrestrial biogenic volatile organic compounds, the aerosol state can still be strongly perturbed by long-range transport from anthropogenic sources (e.g., refs. 10–13). Studies of natural emissions and processes have typically focused on remote regions such as the high northern latitude boreal forest (14), the Brazilian rainforest (15, 16), and the Southern Ocean (17). However, the choice of location tends to be based on the physical remoteness of the site and the strength of local natural emissions, but with little consideration of how closely the aerosol state truly resembles unperturbed conditions.

Several definitions of pristine, natural or “clean background” aerosol environments have been used when analyzing observations, including particle number concentrations (18), the concentration of a particular species such as carbon monoxide or particulate black carbon (19), the location (14, 16), or a combination of factors (6). However, operational definitions suffer from not knowing how much the environment is influenced by a pervasive background of anthropogenic aerosol, which is unlikely to be detectable in observations. It is also not always possible to define pristine environments in terms of the lowest observed aerosol concentration at a particular site because often such conditions are associated with strong scavenging by precipitation and will not represent and behave like the true climatological state in the PI. Remote oceans can provide an insight into how clouds respond to changes in aerosol starting from a very low aerosol baseline (6), but are unlikely to be good analogs for aerosol in all PI regions, which will often have been strongly affected by emissions from natural forest fires (20), volcanic activity (3), or terrestrial biogenic emissions (21). There is therefore no single globally applicable aerosol state that defines the preindustrial atmosphere. Rather, the state changes spatially and seasonally. Global models provide an alternative, and perhaps the only, way of estimating the properties and behavior of aerosols in a preindustrial reference state, and can at least point to where a pristine aerosol state is likely to be observable today.

The purpose of this paper is to define regions of the Earth where PD aerosol (year 2000) looks and behaves most like it did in a PI reference state (year 1750). We use the term pristine to refer to such regions, meaning aerosol in an original (i.e., preindustrial) state, although the term can often be used misleadingly to imply extremely low aerosol concentrations. Anthropogenic emissions in 1750 were not zero (22), and our reference year is therefore not truly “prehuman” (12), but is appropriate for defining the properties and behavior of aerosols in a reference state that is used for radiative forcing calculations (23). We focus our analysis on cloud condensation nuclei (CCN) because of the high sensitivity of cloud radiative forcing to the aerosol load (24).

Based on the results of a global aerosol microphysics model (3, 4, 25), we use two measures of similarity of aerosol between the PI and PD. First, we compare the aerosol states: On how many days are PD CCN concentrations similar to those in the PI? Air pollution tends to be episodic in remote regions, so it is not appropriate to compare longer time averages because a few polluted days will make the whole period appear mildly polluted. Secondly, we compare the aerosol response: How similar is the CCN response (or sensitivity) to perturbations in emissions and processes in the PI and PD? The similarity of aerosol responses is an essential additional factor in defining pristine regions, which cannot be addressed using observations alone. It is conceivable that two periods in the aerosol historical record could have similar CCN concentrations (the aerosol state) but respond differently to perturbations because those two states were generated through a different series of emissions and/or processes. By comparing the responses in the PI and PD, we can identify regions where the PD aerosol is “behaving” like it did in the PI.

To test the similarity of aerosol states, we quantify changes in CCN concentration between 1750 and 2000. To test the similarity of aerosol responses in 1750 and 2000, we compare the sensitivity of CCN to 28 model parameters representing natural and anthropogenic aerosol emissions, microphysical processes, and model structures, with the parameter ranges based on expert elicitation (4, 26). To ensure that the statistics of PI−PD similarity reflect only the change in emissions and associated aerosol processes, we used the same (year 2008) global 3D meteorological reanalyses in 1750 and 2000, which eliminates meteorology as a source of variability between the two years.

Results and Discussion

Properties of the Preindustrial Aerosol.

Fig. 1 shows CCN concentrations simulated using the median setting of the model parameters (Table S1) for PI and PD conditions in January and July (a full year of PI CCN concentrations is shown in Fig. S1). We define CCN concentrations as the number concentration of soluble particles with a dry diameter equal to or greater than 50 nm at 915 hPa, approximately cloud base for stratiform low-level clouds. Contrary to previous suggestions (12), our model results suggest that a significant land−ocean contrast in CCN concentrations could also have existed in the PI. Averaged over a year, the mean PI CCN concentrations over Northern Hemisphere (NH) land regions are 36% higher than the Southern Hemisphere (SH) land regions (NH = 186 cm⁻³ and SH = 137 cm⁻³), while, over ocean regions, they are almost identical (NH = 61 cm⁻³ and SH = 64 cm⁻³). Fires are an important source of high PI CCN concentrations, particularly in Africa, South America, and northern boreal regions in summer. The PI land−ocean contrast in CCN concentrations is particularly large in July, lying between 0 cm⁻³ and 100 cm⁻³ over most extratropical ocean regions and 100 cm⁻³ and 500 cm⁻³ over the majority of the land, with peak CCN concentrations greater than 2,500 cm⁻³ in African savannah burning regions and greater than 5,000 cm⁻³ in Siberian boreal regions. Another factor affecting the PI land−ocean contrast is the higher abundance of biogenic organic compounds over land, which causes the growth of small nucleated particles to CCN sizes (27). The model we use here does not account for the role of organic compounds in the first stages of nucleation itself, which would further increase the land−ocean contrast (21).

Fig. 1.

PI (1750) and PD (2000) monthly mean CCN concentrations in January and July. CCN concentrations are defined as the aerosol concentration with a dry diameter above 50 nm and calculated at cloud base (∼915 hPa).

Histograms (Fig. 2) of PI and PD CCN concentrations (based on Fig. 1 and Fig. S1) show that although the occurrence of high CCN concentrations is similar in the PI and PD, the most frequent oceanic CCN concentrations are centered around 80 cm⁻³ compared with about 180 cm⁻³ over land. The PD histograms show a much greater influence of pollution over land compared with oceanic regions, with the most frequent CCN concentration remaining around 80 cm⁻³ over the ocean, but increasing to more than 500 cm⁻³ over land. Despite the differences between PI and PD aerosol concentrations, there is considerable overlap of the CCN histograms, particularly over oceans at low CCN concentrations, suggesting that regions exist today that could be analogs for PI environments.

Fig. 2.

PI and PD global modeled annual mean CCN concentrations over land and ocean grid cells, calculated from daily mean CCN concentrations at cloud base (∼915 hPa). The third color indicates overlap of the two distributions. The maximum CCN bin concentration is set to 1,000 cm⁻³, although a small fraction exists at higher concentrations.

The estimated uncertainty (SD) in PI CCN concentrations varies spatially and temporally (Fig. S2). The parametric uncertainty was calculated by performing a Monte Carlo sampling of validated Bayesian emulators conditioned on an ensemble of 168 model simulations covering the joint parameter space of the 28 model parameters (see SI Methods). This approach generates a probability density distribution of CCN concentrations caused by the uncertainty in the model input parameters, including their interactions (26). By sampling the model uncertainties, we are able to estimate a plausible range of CCN concentrations in the PI under the assumption that the sources of uncertainty are the same as in the PD (Table S1). Over marine regions, modeled uncertainty in CCN concentrations is typically 20–40% (SD divided by the mean, shown in Fig. S3) between ±60° latitude, up to about 70–90% at higher latitudes and over 100% near the Antarctic continent and very high latitude Arctic regions. In some continental regions, uncertainties exceed 100% of the mean in regions dominated by fires. To further assess how structural uncertainties potentially alter the pristine regions presented in this study, future studies using a range of models would need to be undertaken.

The Occurrence of Pristine Days.

Fig. 3 shows the similarity of CCN concentrations in the PI and PD in terms of the number days that concentrations are within ±20% (a full year is shown in Fig. S4). The threshold of ±20% in concentration is based on the estimated CCN measurement uncertainty across the majority of datasets compiled by Spracklen et al. (28), and Fig. S5 shows the daily fraction of the Earth defined as pristine when this threshold is set to values ranging between ±0% and ±100%.

Fig. 3.

The occurrence of pristine days in January, April, July, and October, based on two definitions. Colors show the number of days per month on which PI and PD CCN concentrations differ by no more than ±20% in that grid cell at cloud base (∼915 hPa). Stippling shows regions where the sensitivities of PI and PD CCN to 28 model parameters are similar (r² ≥ 0.9) in that grid cell at cloud base. Pristine regions are those that exhibit both a similar PI and PD CCN concentration and a similar PI and PD response to the 28 parameters.

The occurrence of pristine CCN regions in the PD atmosphere is highly variable in space and time, with the frequency of pristine days lying between 0% and 100% in a given month. Averaged over a full year, approximately one third (Table 1) of the SH has CCN concentrations similar to the PI, with a maximum spatial coverage over SH ocean regions in the SH summer when every day of the month approaches pristine conditions. In contrast, less than 9% of the NH is pristine, with a maximum coverage of about 15%, also in the summer. In many major shipping regions [e.g., Capaldo et al. (29)], anthropogenic perturbations to aerosol concentrations are large enough that the region is classed as nonpristine. The equatorial Pacific Ocean is the most persistently pristine environment, most likely due to the dominant local marine emission source and effective barrier to NH interhemispheric transport of anthropogenic pollution provided by the intertropical convergence zone. In agreement with observational (19) and modeling studies (11), we identify the southern Pacific Ocean (approximately 20°S–60°S, 90°W –180°W) as a large region close to pristine, especially during SH summer when monthly mean PD CCN concentrations in this region are in the range of 53–285 cm⁻³ (median 104 cm⁻³), when the main natural source of CCN is from DMS-derived sulfate aerosol (30). A Southern Ocean summertime band of pristine CCN exists between 50°S and 65°S, with generally low monthly mean PD CCN concentrations of 20–153 cm⁻³ (median 58 cm⁻³), when natural emissions of sea spray are the dominant aerosol source (17). The midlatitude Pacific and Atlantic Oceans deviate from a pristine state for more of the year than at higher and lower latitudes, mainly due to assumed increases in emissions from South American and African tropical fire regions (31) which is assumed to be due to increased anthropogenic activity. Generally, SH continental land masses have sparse regions of pristine CCN concentrations.

Table 1.

Fraction of the Earth defined as pristine

Month	Pristine Fraction
	Global	Ocean	Land	NH	SH
Jan	0.27 (0.32)	0.34 (0.40)	0.09 (0.13)	0.02 (0.06)	0.53 (0.59)
Apr	0.18 (0.25)	0.21 (0.30)	0.08 (0.12)	0.00 (0.03)	0.35 (0.46)
Jul	0.05 (0.14)	0.06 (0.16)	0.02 (0.08)	0.06 (0.15)	0.04 (0.13)
Oct	0.05 (0.11)	0.07 (0.14)	0.01 (0.05)	0.03 (0.09)	0.07 (0.13)
Annual	0.12 (0.21)	0.16 (0.26)	0.02 (0.10)	0.02 (0.08)	0.22 (0.34)

Pristine defined as PI to PD CCN concentration ±20% and similar PI to PD CCN response to 28 parameters covering natural and anthropogenic emissions, processes and model structures. Values in parentheses show the fraction when the concentration change only is considered.

In the NH, prolonged pristine periods generally occur only over continental regions above 60°N, such as in boreal Canada (32) and Russia (33), where aerosol is affected strongly by natural forest fire emissions. Here, CCN concentrations are highly variable, but generally range from 100 to 1,000 cm⁻³. The high Arctic (75°N and above) is frequently pristine during the NH summer, with low monthly mean CCN concentrations in July of 39–142 cm⁻³ (median 55 cm⁻³), but strongly and persistently polluted during winter and spring, consistent with the seasonal cycle of Arctic haze controlled by scavenging processes (34, 35). There are almost no marine pristine days during NH winter and spring, and very few regions are persistently pristine over a month. In NH midlatitude regions, there are no pristine days at any time of the year. In particular, the North Pacific Ocean is impacted by transport of pollution from East Asia to North America (36) and is a region where we find no pristine days in the main transport periods.

Our confidence in the extent and location of pristine regions depends on the modeled CCN uncertainty as well as the assumed tolerance used to compare PI and PD CCN (set at ±20%, as above). Fig. S6 shows the effect of changing the tolerance to 10%, 30%, and 50%. This range of tolerances is comparable to the modeled relative CCN uncertainty, which is ∼20–50% in the main pristine regions (Fig. S3). We have most confidence in regions with a low relative modeled CCN uncertainty (Fig. S3) and a high number of pristine days. The optimum pristine region by this definition is the central Pacific. Although relative uncertainties can be fairly high in other pristine regions, we expect the model uncertainties to be correlated in the PI and PD, giving us more confidence in the model results than indicated by Fig. S3. While meteorological variability is likely to cause interannual variability in the precise location of pristine regions, the principal pristine regions (the remote Pacific, Southern Ocean, Arctic, etc.) will be more climatologically persistent features.

The distribution of pristine days is similar at ∼2.5 km above sea level (a.s.l.) (Fig. S7), while at ∼5 km a.s.l., long-range transport of anthropogenic emissions cause changes in the spatial distributions of SH pristine regions (Fig. S8). However, these free tropospheric aerosols do not affect cloud-base CCN concentrations in the boundary layer, and if they are mixed down to lower altitudes, then they will be included already in our analyzed fields at ∼850 m a.s.l.

Changes in CCN Sensitivity Between 1750 and 2000.

Stippling in Fig. 3 shows regions where the aerosol sensitivity to perturbations of 28 model parameters covering emissions, microphysical processes and model structures (26) is similar in 1750 and 2000. The parameter sensitivities in each grid cell at cloud base (915 hPa) were calculated by variance decomposition of the CCN distributions that were used to generate the uncertainty in CCN concentrations (Fig. S2). The fractional contribution to variance of each parameter is termed the main effect index (26, 37). We correlated the main effect indices for the 28 parameter perturbations in each grid cell (see the SI Methods for a full description) and define regions to be pristine, somewhat arbitrarily, if the coefficient of correlation is greater than 0.9. Fig. 4 shows these correlations at four representative sites.

Fig. 4.

The similarity of CCN sensitivities in the PI and PD to 28 parameters covering natural emissions (green), anthropogenic emissions (red), and processes (blue) for every month in the year at four different sides: (A) Melanesia (1°S, 151°E), (B) eastern Atlantic (38°N, 21°W), (C) Brazilian rainforest (1°S, 66°W), and (D) northeast China (38°N, 111°E). For more information on individual marker descriptions, please see Table S1.

In regions where CCN concentrations exhibit about 20 or more pristine days in a month, the CCN response to model parameters is generally also similar in both periods. For regions with about 10–19 d of pristine CCN concentration, there is partial overlap with regions with similar CCN sensitivity, while for regions of with less than 10 d per month of pristine CCN concentrations, CCN sensitivities are now different than the PI.

The equatorial Pacific is the largest region that is closest to pristine all year round in terms of the similarity of the aerosol state and the aerosol sensitivity, while other regions vary seasonally between being pristine and not. Other regions with a high number (20 or more) of pristine days and similar CCN sensitivity in one or more months include parts of Alaska and Yukon, the Southern Ocean, Melanesia, southwest Greenland, and the southern Indian Ocean.

Fig. 4 shows the change in PI to PD CCN sensitivity at four grid cells representing four different sites. Melanesia (Fig. 4A) has a very similar CCN sensitivity and state in 1750 and 2000, where 99% of January/April/July/October modeled days have PD CCN concentrations within ±20% of PI concentrations. Prevailing winds in this region originate over the Pacific Ocean, bringing clean background air masses and natural marine aerosol (38). Fig. 4A shows that both 1750 and 2000 CCN sensitivities are dominated by the same parameters relating to natural emissions and aerosol microphysical processes [volcanic SO₂ and biogenic secondary organic aerosol (SOA) as well as model uncertainties in the Aitken mode width, boundary layer nucleation rates, CCN activation diameter, and the pH of cloud droplets]. In all months, the fraction of variance attributable to anthropogenic emissions is less than about 1%. CCN concentrations and behavior in this location are therefore clearly driven by natural processes.

The eastern Atlantic (Fig. 4B) has a different CCN response and state in the two periods. It is located in a region of high aerosol−cloud radiative forcing (4). Many of the main effect indices have changed over the industrial period, indicating an anthropogenic influence on the behavior of the aerosol. In particular, anthropogenic SO₂ emissions and production of SOA from anthropogenic compounds contribute more to CCN variance in the PD than in the PI, and natural emissions of DMS and biomass burning contribute less. This reduction in sensitivity to natural emissions will suppress natural aerosol−climate feedbacks (39).

The Brazilian rainforest (Fig. 4C) has a similar CCN response but dissimilar state in the two periods. Although the aerosol state is strongly anthropogenically perturbed by increased biomass burning emissions, these emissions dominate the aerosol sensitivity in both the PI and PD (Fig. 4C). While our model results suggest that fires are the largest contributor to PI CCN concentrations, major uncertainties exist as to the magnitudes of historic biomass burning emissions (20). However, in many fire-dominated regions, the response of CCN to changes in emissions is similar in both the PI and PD (e.g., Fig. 4C), and these regions are still likely to be useful analogs of PI CCN behavior.

Northeast China (Fig. 4D) has a very different CCN response and state in the two periods. Annual mean CCN concentrations at this site are the furthest from PI CCN baseline concentrations of any other location. While PI CCN concentrations are sensitive to the flux and size of biofuel emissions (from early human activity) and biogenic SOA, PD CCN concentrations are sensitive to the flux and size of fossil fuel emissions and the fraction of sulfate formed on subgrid scales from anthropogenic SO₂ emissions.

Model results spanning the grid cells 14°S to 22°S and 121°W to 130°W, which overlaps the most with the pristine region studied by Koren et al. (6) (13°S to 22°S and 121°W to 130°W), suggest that even in these remote regions PD CCN concentrations in July are up to 40% higher than in the PI. The cause of the enhanced CCN is increased biomass burning emissions. Also, the CCN response to volcanic SO₂ emissions in the region has approximately halved since the PI, and an additional PD contribution from anthropogenic SOA concentrations is seen.

The Overlap of Pristine Regions and Aerosol−Cloud Radiative Forcing.

Aerosol measurements under pristine conditions would be most useful if they were made in regions where there is an aerosol−cloud radiative forcing, so that both the clean and perturbed aerosol−cloud processes could be observed. However, on average, such regions are of course not pristine today. To assess the overlap of forcing and pristine conditions, Fig. 5 shows the relationship between 1750-to-2000 monthly mean aerosol indirect radiative forcing (see SI Methods) and the number of pristine days per month. As expected, the general relationship shows that regions with the highest monthly mean forcing have the lowest number of pristine days (gray markers in Fig. 5). For example, grid boxes with greater than −5 Wm⁻² forcing have generally less than five pristine days per month. When the additional constraint of similarity of CCN sensitivities is applied (tan markers in Fig. 5), it is possible to observe pristine days only in regions with a monthly mean forcing less than about −2 to −3 Wm⁻², which is in the lowest quartile of our forcing estimates.

Fig. 5.

The relationship between 1750-to-2000 monthly mean aerosol indirect radiative forcing and the occurrence of pristine aerosol conditions. Maps show number of days in which pristine conditions (PI:PD CCN concentrations within ±20% and similar response to the 28 parameters in both time periods) and average low cloud fraction (≥0.3; stippling) overlap.

If we do not restrict ourselves to observing pristine aerosol in regions of forcing, but just in regions of low cloud, then extensive regions can be found with pristine days of 0–31 per month. The black markers and maps in Fig. 5 show pristine regions with higher than average low cloud cover (fraction ≥0.3). These pristine low-cloud conditions generally occur over the major marine stratocumulus decks in the SH. It is in these regions where aerosol and related process measurements would be most useful in constraining the PI baseline using PD observations. However, in making aerosol−cloud measurements, there is a compromise to be reached between a useful number of observable pristine days and the magnitude of the forcing in that region. Regions with a small number of pristine days will enable strong cloud perturbations to be observed, while regions with a high number of pristine days will experience only weak or brief aerosol perturbations.

These results show that, although cloud radiative forcing and pristine regions are in general spatially anticorrelated, meteorological variability means that in regions with nonzero radiative forcing, there are days that are likely to approach PI aerosol conditions.

Conclusions

If the large uncertainty in aerosol−cloud radiative forcing between PI and PD periods is to be constrained by measurements, it is important to characterize aerosol in regions of today’s atmosphere that most closely resemble PI conditions. While there is likely to always be uncertainty associated with predicting the PI atmosphere, simply because natural emissions may have changed with time, we have defined regions in today’s atmosphere that are similar enough to PI conditions to be explored further. Our joint analysis of changes in CCN concentration alongside the changes in the sensitivity of CCN to emission and process perturbations provides a complete model picture of where on Earth we can observe pristine aerosol concentrations and behavior.

We have identified regions and seasons with the highest likelihood of observing pristine aerosol in terms of the number of model days per month pristine aerosol could be observed. However, ultimately, it is necessary to define pristine aerosol based on measurements because measurements are needed for model evaluation. There is no universal operational definition of pristine aerosol because natural aerosols vary substantially. The mass concentration of a species associated with anthropogenic emissions is often used as an anthropogenic tracer [e.g., carbon monoxide (19) or black carbon (40)]. However, such a definition is appropriate only if the species is not part of the local natural aerosol—clearly not the case if natural forest fire aerosols are being studied.

Our identified pristine regions may not be the only places to observe PI-like aerosol behavior. Although our model suggests that even the clean region of the Pacific studied by Koren et al. (6) has a different aerosol state and response to parameter perturbations today compared with the PI, it may still behave in a way that is sufficiently similar to the PI to be informative. However, clouds appear to be highly sensitive to aerosols at very low concentrations (3, 4, 6) and strong regime shifts can occur in some aerosol-cloud systems (8, 41), so even small perturbations of the aerosol state away from PI conditions may be important.

Our analysis has focused on daily mean regional CCN concentrations at cloud base. Precipitating shallow clouds can strongly scavenge aerosol on timescales shorter than 1 d (42, 43), leading to transient and localized conditions of very low aerosol concentration, even when the region is generally perturbed relative to the PI state by anthropogenic aerosol. The question arises whether such PD locally scavenged clean environments can be used to experimentally evaluate modeled aerosol−cloud interaction as an analog of the regional-scale response to changes in aerosols that occurred between the PI and PD. Given the different behaviors of single clouds and regional-scale cloud systems (44), it is unlikely that local processes will be informative about regional PI to PD changes that we present here. It is also important to recognize that pristine CCN environments could still be perturbed by light-absorbing aerosols either within or above the clouds (45, 46). Such effects, and associated fast adjustments of the cloud system, may alter the extent of PI-like aerosol−cloud environments shown in this study. Furthermore, pristine aerosol days may be associated with different meteorological conditions than polluted days, which would make it difficult to separate meteorological and aerosol influences on cloud behavior in PD observations.

Pristine low-cloud regions are almost entirely in the SH. To reduce the uncertainty in regions of NH forcing, we need to characterize the natural aerosol state either directly in these perturbed NH regions (which seems challenging) or in regions of the SH that are appropriate analogs for the NH. From a model uncertainty reduction perspective, an appropriate analog implies that the parameters controlling CCN sensitivity in the PI SH are the same as those controlling sensitivity in the NH. Our PI simulations of CCN suggest that CCN concentrations may have been higher in the NH than the SH, because of a larger influence of terrestrial emissions. Further research is needed to determine whether these differences limit what we can learn about NH aerosol from SH measurements.

The PI NH/SH contrast, combined with the rarity of pristine days in the NH, may mean that we have to accept that some of the PI to PD aerosol−cloud forcing uncertainty will be irreducible. Regardless of how well we can observe and simulate aerosol−cloud interaction in today’s atmosphere, a large part of the forcing uncertainty—that part associated with the unknown baseline aerosol state—will remain.

Methods

We used a global aerosol microphysics model (25), within the Toulouse Off-line Model of Chemistry and Transport (TOMCAT) (47), to simulate daily mean CCN concentrations in both the PI and the PD (see SI Methods for further details). Temperature fields are provided from European Centre for Medium-range Weather Forecasts, while low cloud cover is from the International Satellite Cloud Climatology Project climatology. CCN concentrations were modeled based on the median setting of 28 parameters used in Lee et al. (26), which are listed in Table S1. Emissions for both the PI and PD are listed in Table S2 and mostly follow Dentener et al. (22). Anthropogenic fossil fuel emissions are assumed to be zero in the PI, although a small anthropogenic biofuel component to the atmosphere exists (22). Natural emissions of sea spray, biogenic organic volatile compounds (which forms secondary organic aerosol), and volcanic sulfur dioxide are the same in the PI and PD. PI emissions of biofuel and biomass burning, which follow Dentener et al. (22), were derived by scaling by population, land cover, and crop production and are assigned as natural aerosol, although, in reality, these emissions could be a result of both anthropogenic (land/agricultural clearance) and natural (wildfire) activity. The model has previously been extensively evaluated against observations (see SI Methods), and model−observation bias is low in both clean and polluted regions (28). A variance-based sensitivity analysis of PI CCN concentrations to perturbations of the 28 model parameters covering emissions, microphysics, and model structures was performed in the same way as the PD analysis of Lee et al. (26) (see SI Methods). The 28 model parameters and their ranges are defined in Table S1, with a full description of what each parameter does in the model in Lee et al. (26). Parameters relating to biomass burning emissions are assigned as natural aerosol in both PI and PD. Accurate source type identification of biomass burning aerosol with similar fuel type is currently impossible to disentangle in the atmosphere once the aerosol becomes well mixed. In combination with the similar assumption of PI BB emissions, we therefore expect that our identification of pristine aerosol in fire-dominated environments will be an upper limit. The modeled CCN concentrations and the sensitivity analysis used 2008 meteorology in each modeled time period. We define CCN concentrations as the aerosol concentration with a dry diameter above 50 nm that are reported at 915 hPa, typical of cloud base for stratiform low-level clouds. An analysis of the vertical profile of pristine regions (see SI Methods) shows little variation within the boundary layer. Radiative forcing values reported in Carslaw et al. (4), which were derived from the same set of PI and PD experiments, were also used in this study.