Understanding the regional climate consequences of aerosols is of growing importance as emissions of aerosols and their precursors are projected to decline in most regions over the coming decades due to policies enacted to protect human health from the negative effects of air pollution (Rao et al., 2017; van Vuuren et al., 2011). Global emissions of anthropogenic aerosols and their precursors, including sulfur dioxide (SO2, precursor to sulfate aerosol), black carbon (BC), and organic carbon aerosol (OA), peaked in the 1970s and have been declining for the last few decades (Klimont et al., 2013; Smith et al., 2011; Smith and Bond, 2014). Major source regions such as the United States and Europe have also experienced decreases in anthropogenic SO2, BC, and OA emissions during this time (Leibensperger et al., 2012; Tørseth et al., 2012). Emissions in China may also be beginning to decline, whereas emissions in India continue to increase (Fontes et al., 2017; Li et al., 2017; Lu et al., 2011). As emissions of anthropogenic aerosols are phased out regionally and globally, their removal is expected to affect global and regional precipitation (Shindell et al., 2012). However, we currently lack a full understanding of the magnitude, spatial pattern, statistical significance, and underlying physical mechanisms of the precipitation response. In order to address this knowledge gap, we simulate here precipitation responses to the removal of aerosols from six world regions in three different fully coupled chemistry–climate models.

Aerosols impact precipitation primarily through two pathways: by altering the surface and top-of-atmosphere solar radiation balance (direct effect) and through microphysical effects on clouds (indirect effect) (Myhre et al., 2013). Generally, decreasing aerosol emissions results in a net enhancement of precipitation, since the reduced aerosol attenuation of incoming solar radiation results in more radiation reaching the surface, thereby resulting in more available heat for evaporation and convection (Ramanathan et al., 2001; Rosenfeld et al., 2008). Additionally, aerosol removal may enhance autoconversion and thus further increase rainfall locally via the cloud lifetime effect (Albrecht, 1989), though this effect remains uncertain (Stevens and Feingold, 2009). Aerosol composition plays a role in determining precipitation response in both the direct and indirect pathways. Pure sulfate aerosols are scattering agents, while black carbon also absorbs incoming solar radiation and therefore may impact precipitation rates in different ways (Ming et al., 2010). Sulfate aerosols and some organic aerosols are efficient cloud condensation nuclei (CCN), while black carbon aerosols do not seed liquid clouds as readily (Bond et al., 2013; Petters and Kreidenweis, 2007), at least not initially. These differences in optical and microphysical properties related to aerosol composition may lead to unique precipitation responses to the removal of individual aerosol components such as sulfate and black carbon (Andrews et al., 2010; Frieler et al., 2011).

Previous work has found that aerosols are linked to a number of regional precipitation and/or circulation responses, including the location and width of the Intertropical Convergence Zone (ITCZ), (Allen, 2015; Hwang et al., 2013; Ridley et al., 2015; Allen and Ajoku, 2016), rainfall in the Sahel (Ackerley et al., 2011; Biasutti and Giannini, 2006; Chang et al., 2011; Haywood et al., 2013; Held et al., 2005; Rotstayn et al., 2002; Westervelt et al., 2017), South Asian monsoon circulation (Bollasina et al., 2011; Menon et al., 2002), phasing of the North Atlantic Oscillation (NAO) (Fischer-Bruns et al., 2009) and North Atlantic climate variability (Booth et al., 2012), and rainfall in the US (Leibensperger et al., 2012; Shindell et al., 2012). Additional work is needed to identify robustness across multiple models and understand the physical mechanisms of these regional responses to aerosols. Westervelt et al. (2017) began this process by simulating the precipitation response to the complete removal of US anthropogenic SO2 emissions in three coupled chemistry–climate models and found statistically significant increases in Sahel rainfall in multiple models. We build here on the work of Westervelt et al. (2017) by considering, for the same three models, regional emissions removal from not only the US, but also Europe, China, India, South America, and Africa. Additionally, we expand the scope to include multiple aerosol types, including sulfate, BC, and OA in each of these regions. We identify robust (and non-robust) precipitation responses to a variety of regional aerosol perturbations and show that precipitation responses in the Sahel can be explained by a consistent physical mechanism involving a change in the interhemispheric temperature gradient and a northward shift of the ITCZ that is robust across multiple models. We choose to investigate the Sahel in more detail based on its recent climatic vulnerability to drought, which occurred over the latter half of the 20th century and was partially attributed to aerosol forcing (Ackerley et al., 2011; Biasutti and Giannini, 2006; Held et al., 2005). Similarly, precipitation in the Mediterranean has declined since the mid-20th century, although the cause of this decline is not well understood (Giorgi, 2002; Xoplaki et al., 2004). While the response of Mediterranean precipitation to climate variability has been thoroughly investigated (Dünkeloh and Jacobeit, 2003; Krichak and Alpert, 2005), the potential role of aerosol forcing has not been examined. Therefore, we use our multi-model regional aerosol perturbation framework to focus on the Sahel and Mediterranean precipitation responses in detail in addition to our more general analysis of precipitation responses around the globe.

We use an identical modeling framework as described by Westervelt et al. (2017) and Conley et al. (2017). Briefly, we employ three coupled atmosphere–ocean–land–sea–ice climate models with fully interactive chemistry of aerosols and trace gases: (1) Geophysical Fluid Dynamics Laboratory Coupled Climate Model version 3 (GFDL-CM3) (Donner et al., 2011), (2) Goddard Institute for Space Studies ModelE2 (GISS-E2-R) (Schmidt et al., 2014), and (3) Community Earth System Model version 1 (CESM1) (Neale et al., 2012). The model configuration for each is very similar to that used for the Coupled Model Intercomparison Project 5 (CMIP5). For further model description and model evaluation of relevance to precipitation response, we refer readers to Westervelt et al. (2017).

In each model, we conduct a series of long “present-day” control simulations of up to 400 years in length forced by perpetual year 2000 (2005 for NCAR-CESM1) conditions, including all emissions of aerosols and their precursors and greenhouse gas concentrations. We then conduct individual regional aerosol perturbation simulations in each model of at least 160 years and as long as 240 years, in which the anthropogenic aerosol or aerosol precursor emissions for a certain region are set equal to zero or reduced by the amount shown in Table 1. The magnitude of the emissions perturbation was chosen in order to have roughly equivalent emissions decreases across regions and models. As an example, “IN_SO2” refers to a simulation with perpetual year 2000 conditions (2005 for NCAR-CESM1) perturbed by setting all anthropogenic SO2 emissions over India to zero. Other than the regional aerosol emissions perturbation, all other model settings remain identical to the control. Long control and perturbation simulations allow us to establish statistical significance and separate forced responses from internal climate variability. We also conduct an additional set of atmosphere-only, fixed SST simulations of 40–80 years in length with control and perturbed aerosol emissions to calculate the effective radiative forcing (ERF; as defined in Myhre et al., 2013) resulting from the regional perturbations to aerosol emissions. The ERF calculation in GISS-E2 included a land surface temperature adjustment based on the global climate sensitivity parameter (Forster et al., 2016), while GFDL-CM3 and NCAR-CESM1 did not include this adjustment.

Figure 1 presents the annual mean precipitation response to a given aerosol emissions perturbation in each of the three models for six different perturbation simulations. The remaining simulations are presented in Fig. S1 in the Supplement. Hatching represents statistical significance at the 95 % level according to a simple Student's t test. Each plot is the difference between the perturbation simulation and the control (e.g., US_SO2 minus control), differenced at each exact month of the two simulations as done in Westervelt et al. (2017), and can therefore be interpreted as the precipitation response to decreasing regional aerosol emissions. The first row (Fig. 1a–c) is for zero US SO2 emissions and is discussed in detail in Westervelt et al. (2017). Generally, across all perturbations, precipitation responses are largest in NCAR-CESM1, followed by GFDL-CM3 and GISS-E2. GISS-E2 simulations were performed in a setup that does not include a cloud lifetime effect (Schmidt et al., 2014), contributing to a smaller aerosol effective radiative forcing (Table 1 and Fig. S2) and a weaker precipitation response in that model.

Global mean aerosol effective radiative forcing values for each of the models are shown in Table 1, and a scatterplot of global mean precipitation changes versus global mean aerosol effective forcing at the top of the atmosphere (TOA) is presented in Fig. S2. The aerosol ERF values are largest in NCAR-CESM1, followed by GFDL-CM3 and GISS-E2. Aerosol ERF is a factor of 2 or 3 smaller in GISS-E2 than in GFDL-CM3 and NCAR-CESM1 for some simulations. Overall, aerosol ERF is largest in NCAR-CESM1, ranging from 0 to 0.3 W m-2 depending on the regional aerosol perturbation. Across the models, we find a strong to medium linear relationship between global precipitation response and global effective radiative forcing in GFDL-CM3 (r=0.70) and GISS-E2 (r=0.5), but poor correlation in NCAR-CESM1 (r=0.23). Although global precipitation responses are known to be constrained by the atmospheric energy budget (Allen and Ingram, 2002; Ming et al., 2010; Liu et al., 2018), we find weaker correlation (e.g., r=-0.3 for GFDL-CM3) between global precipitation response and atmospheric absorption (TOA minus surface forcing) when compared to global precipitation and TOA forcing alone. Samset et al. (2016) and Liu et al. (2018) found strong correlation between global precipitation “fast” response and atmospheric absorption. Their analysis correlated precipitation responses from fixed SST simulations with aerosol ERF, whereas our analysis in Fig. S2 correlates precipitation responses from coupled model simulations with aerosol ERF, which may explain the discrepancy. Global precipitation may also be an imperfect metric for correlation if opposite-signed regional changes are largely offset in the global mean. Recently, Chung and Soden (2017) showed that aerosol indirect effects could dominate precipitation responses to aerosol perturbations, consistent with our finding that GISS-E2, lacking an aerosol cloud lifetime effect, has the smallest precipitation response.

In Fig. 2, we present precipitation responses (perturbation minus control, representing aerosol decreases, as in Fig. 1) globally averaged and averaged over two regions (which are shown in Sect. 4), the Sahel and the Mediterranean. The numbers in the upper left of each panel of Fig. 2 represent the mean precipitation for the control run for each region and each time period. Figure 2 shows that the global precipitation responses (Fig. 2a) nearly always agree in the three models. In general, NCAR-CESM1 responds the strongest to aerosol decreases, with increases in global mean precipitation up to about 0.025 mm d-1 or about 1 % of the global mean in the control simulation. Global precipitation changes in GISS-E2 and GFDL-CM3 are similar in magnitude in many of the aerosol perturbation scenarios. A total of 33 out of 34 model simulations among the various regional emissions perturbations result in a global annual mean increase in precipitation, the one exception being US_BC in GFDL-CM3 (not statistically significant). In addition to heating the surface, BC removal results in cooling aloft in the free troposphere and an increase in shortwave radiation at the surface, both of which can drive convective updrafts and result in precipitation increases. This “fast response” of precipitation to BC reductions tends to dominate the total response to BC, as shown in the Precipitation Driver Response Model Intercomparison Project (PDRMIP) results (Samset et al., 2016; Liu et al., 2018). Despite opposite-signed aerosol ERF (Table 1) between BC and sulfate perturbation simulations among the models, global precipitation responses are often in agreement in sign (e.g., EU_BC and EU_SO2 in NCAR-CESM1 and GFDL-CM3). Because of the surface heating influence of BC compared to the cooling effects of sulfate, previous research has shown that BC and sulfate perturb precipitation in opposite directions (Wang, 2007; Ramanathan and Carmichael, 2008). Our results here, while still somewhat inconclusive, suggest that in some cases, BC emissions decreases may actually increase global and regional precipitation, similar to sulfate. The black carbon aerosol global mean ERF (Table 1) varies in sign and magnitude, indicating a strong sensitivity to different model configurations for black carbon and perhaps a role for internal climate variability. In many of the black carbon simulations, the global mean aerosol ERF values reported are not statistically significant. This result highlights the fact that the influence of BC on global precipitation is still largely uncertain (Pendergrass and Hartmann, 2012; Liu et al., 2018) with major knowledge gaps still remaining (Bond et al., 2013).

Figure 1 points to an ENSO-like (El Niño–Southern Oscillation) response in the tropical Pacific. In NCAR-CESM1 and GISS-E2, there is a strong east–west dipole response in the tropical Pacific, with drying to the west and wetting to the east. These responses are some of the largest in any region and are statistically significant in NCAR-CESM1. There are also significant impacts in the tropical Pacific in GFDL-CM3, especially in CH_SO2 (Fig. 6g) and US_ALL (Fig. 6m), though the precipitation response is generally opposite in sign compared to NCAR-CESM1 and GISS-E2, with a wetting in the western tropical Pacific in GFDL-CM3 as opposed to a drying in the other two models.

We therefore extend our precipitation analysis by investigating the impact that aerosols may have on precipitation through changes in the El Niño–Southern Oscillation (ENSO). To estimate the ENSO component of the precipitation response to regional aerosol emissions decreases, we first perform a linear regression of the monthly mean precipitation fields onto a monthly ENSO index at each grid point in the control simulation of each model. We use a large-scale Indo-Pacific zonal sea level pressure (SLP) gradient index representing Walker circulation variations, which are closely linked to ENSO (Vecchi et al., 2006). The Indo-Pacific SLP gradient is defined as the difference between regional average SLP in the Indian Ocean and west Pacific (80–160∘ E, 5∘ S–5∘ N) and the central and east Pacific (160–80∘W, 5∘ S–5∘ N). The index is computed for every simulation (control and perturbation) and differences in indices are calculated between each perturbation simulation and the control simulation. The ENSO component of the precipitation response to aerosol emissions reductions, ΔPENSO, is then computed as ΔPENSO=rP:ENSOΔENSO, where rP:ENSO is the regression value (slope) between precipitation and the Indo-Pacific zonal SLP gradient index in the control simulation (one value per grid point) and ΔENSO is the difference of the index between the perturbation and the control simulation.

The ENSO component of the precipitation response to aerosol emissions reductions is shown in Fig. 6, with the same layout as Fig. 1 except with a smaller scale range (by a factor of 5). In each model, we find substantial responses mostly in the tropical Pacific and Atlantic, with changes as high as 0.1 mm d-1 or 20 % of the total precipitation response (compare with Fig. 1). With the exception of the IN_SO2 simulation in all models and the IN_BC simulation in GISS-E2, the tropical precipitation patterns in every simulation and their teleconnections in different parts of the world tend to resemble the positive phase of ENSO (El Niño). Though most of the large responses are in the tropics, there is some evidence of ENSO teleconnections, for example over the Amazon region in Brazil, for which precipitation decreases (drying) typically occur in each simulation and each model associated with the positive phase of ENSO. The ENSO component of the precipitation response is also apparent over the Indian monsoon region, manifested mostly as a drying, consistent with the positive phase of ENSO. Agreement between models is strongest for GFDL-CM3 and GISS-E2, which show similar ΔPENSO patterns for each of the different perturbation simulations in Fig. 6. All three models agree on a strong response in the US_SO2 simulation; however, there is a weaker response in NCAR-CESM1 for the rest of the perturbation simulations compared to the other two models. Since the models each show ENSO-like responses in the tropical Pacific, albeit with varying degrees of statistical significance and consistency, we conclude that no matter the emissions region or aerosol type, precipitation changes may occur via the modulation of ENSO in the tropical Pacific as a result of aerosol decreases, and these changes mostly resemble the positive phase (El Niño).

We conduct a series of 14 aerosol emissions perturbation simulations (160–240 years each) in which we perturb emissions of sulfur dioxide and/or carbonaceous aerosol within six world regions relative to a long present-day control simulation in three coupled chemistry–climate models: NOAA Geophysical Fluid Dynamics Laboratory Coupled Model 3 (GFDL-CM3), NCAR Community Earth System Model (CESM1), and NASA Goddard Institute for Space Studies ModelE2 (GISS-E2). We find local increases in precipitation near the source region for each individual aerosol perturbation (e.g., increases in Chinese precipitation for the CH_SO2 simulation), with statistical significance mostly limited to two models: NCAR-CESM1 and GFDL-CM3. We find strong tropical precipitation responses in all three models and in essentially all aerosol removal simulations. In NCAR-CESM1 and GFDL-CM3, a northward shift in the tropical North Atlantic ITCZ is associated with increased Sahel precipitation in several of the simulations in which aerosols are removed. Globally averaged, small increases in precipitation occur in nearly all (33 out of 34 simulations across the three models) aerosol emission removal simulations. Regional emissions removal of black carbon (BC) and sulfur dioxide alone both increase global mean precipitation in some cases, despite opposite-signed ERF, highlighting the uncertainties remaining in BC aerosol impacts on precipitation.

Precipitation response is weakest and largely lacks statistical significance in GISS-E2, partially attributed to the lack of a cloud lifetime effect and thus a weaker aerosol indirect effect, which was recently found to dominate tropical precipitation response to aerosols (Chung and Soden, 2017). Our results further support this conclusion, as we find the weakest radiative forcing and precipitation response in GISS-E2. Using both climate model simulations and satellite observations of a major volcanic eruption, Malavelle et al. (2017) found that aerosol-induced changes in cloud liquid water path (the cloud lifetime effect) were undetectable, suggesting that the cloud lifetime effect may be less important than the cloud albedo effect for climate models. Without sensitivity simulations that isolate the cloud lifetime component of the precipitation response to regional aerosol emissions removal, however, it is difficult to determine conclusively whether cloud microphysical or large-scale dynamical mechanisms dominate the modeled precipitation response.

We estimate the aerosol effective radiative forcing (ERF) in each perturbation simulation in each model using a series of atmosphere-only simulations with sea surface temperatures fixed to present-day modeled climatological means. The global mean ERF values are positive in all model simulations with the exception of black carbon simulations (EU_BC, US_BC), which are not statistically significant and generally fall in the range of 0 to 0.3 W m-2. ERF is largest in NCAR-CESM1, followed by GFDL-CM3 and GISS-E2. In both GISS-E2 and GFDL-CM3, global precipitation response correlates strongly to moderately (r>0.5) with global mean ERF, although NCAR-CESM1 shows a weaker correlation (r=0.3).

We further investigate the regional aerosol impact on remote precipitation and show a strong linear relationship between the change in interhemispheric temperature gradient and changes in Sahel rainfall across all of the different aerosol emission perturbation simulations. Changes in the interhemispheric temperature gradient produce an anomalous Hadley cell circulation and an accompanying northward ITCZ shift, with implications for precipitation over the Sahel. This linear relationship holds across multiple models, suggesting that regional aerosol reductions impact precipitation via the same physical mechanism, which we interpret to be a large-scale dynamical response, across different models and different aerosol perturbations. Higher-latitude regional emissions reductions (e.g., US, Europe) lead to greater change in the interhemispheric temperature gradient and thus correspondingly larger changes in Sahel rainfall than lower-latitude aerosol emissions perturbations (e.g., China, India, Africa, South America). Air pollution controls in Europe and the US may help reduce the likelihood and severity of future droughts in the Sahel and by altering the interhemispheric temperature gradient can influence precipitation in regions far removed from the emission region.

We find increases in Mediterranean wintertime precipitation in two of three models in response to most aerosol removal perturbations, implying that increases in aerosols throughout the mid-20th century could have played a role in the observed decreasing precipitation trends. In GFDL-CM3, this precipitation response can be largely explained by an aerosol-induced weakening of the prevailing westerlies and southward shift in the storm track over the Atlantic. Despite the distinction between sulfate (and organic carbon) and black carbon as scattering (cooling) versus absorbing (warming) species, respectively, we find that European precipitation, sea level pressure, and wind speed respond similarly to the removal of each of these species emissions over Europe, implying a role for cloud microphysical effects in this local climate response. This mechanism, however, is not confirmed by either GISS-E2 or NCAR-CESM1 and is therefore not robust and requires future investigation. Previous work relating aerosol forcing to North Atlantic circulation (Chiacchio et al., 2011; Fischer-Bruns et al., 2009; Allen and Sherwood, 2012) has been limited to single models and results have been inconclusive across studies. The impact of aerosols on the North Atlantic and Mediterranean climate (and the NAO) thus remains unclear and may warrant additional work with a larger model ensemble, with highly time-resolved diagnostics needed to probe the driving mechanisms more deeply and to assess robustness in a more rigorous manner.

Based on our analysis, we conclude that the El Niño–Southern Oscillation (ENSO) plays an important role in modulating the impact of regional aerosol removal on precipitation. Kim et al. (2015) found observational evidence from satellite aerosol retrievals and MERRA reanalysis data supporting a role for aerosol-driven ENSO amplification in rainfall increases over the Indian summer monsoon region by correlating elevated aerosol levels over India with Niño3.4 SST index and precipitation rates. We perform a linear regression analysis to determine the contribution of the ENSO component of the precipitation responses to the total response to regional aerosol emissions. We find the ENSO component can be as large as 20 %, especially over the tropical Pacific, with teleconnections to South Asian monsoon precipitation and Amazon wet season rainfall. Regional aerosol emissions reductions tend to cause a shift to the positive ENSO phase (El Niño as opposed to La Niña), with a few exceptions. Model agreement on the ENSO component of the precipitation response is best for the US_SO2 simulation and best between GISS-E2 and GFDL-CM3. More broadly, our findings suggest a possible anthropogenic influence on this mode of climate variability, which may complicate efforts to separate variability arising naturally from that forced by anthropogenic drivers.

Aerosol–precipitation interactions remain one of the most uncertain aspects of future climate change, especially on the regional scale (Rosenfeld et al., 2008; Michibata et al., 2016). To reduce the uncertainty of how future regional aerosol decreases will impact regional precipitation, a thorough analysis with multiple models, including several regions and aerosol types, is needed. Our results show that robust precipitation responses to regional aerosol emissions changes do occur, indicating promise for future work. One caveat of our study is that in each of the models, aerosols do not exert a microphysical effect on deep convective clouds; however, they can alter precipitation associated with deep convection through the aerosol direct effect. Overall, our findings suggest that, despite large variations between different models, there are some robust precipitation responses to aerosol emissions that warrant future investigation with additional models to pursue even more robust estimates, perhaps through model intercomparison projects such as the upcoming AerChemMIP (Aerosol Chemistry Model Intercomparison Project) (Collins et al., 2017). Other precipitation responses show little consistency across the models, raising questions as to whether the model representation is insufficient to detect a role for aerosol emissions or whether those responses are swamped by climate variability relative to any aerosol influence. A possible avenue of further study may be combining different regions into a single perturbation simulation, resulting in a larger climate response and the ability to test for additivity or linearity among the simulations. Our analysis can serve as a benchmark for future efforts with fully coupled chemistry and interactive emissions within climate models that consider emissions perturbations from a broad sampling of regions and aerosol species.