Climate and health implications of future aerosol emission scenarios

Anthropogenic aerosols have a net cooling effect on climate and also cause adverse health effects by degrading air quality. In this global-scale sensitivity study, we used a combination of the aerosol-climate model ECHAM-HAMMOZ and the University of Victoria Earth System Climate Model to assess the climate and health effects of aerosols emissions from three Representative Concentration Pathways (RCP2.6, RCP4.5, and RCP8.5) and two new (LOW and HIGH) aerosol emission scenarios derived from RCP4.5, but that span a wider spectrum of possible future aerosol emissions. All simulations had CO2 emissions and greenhouse gas forcings from RCP4.5. Aerosol forcing declined similarly in the standard RCP aerosol emission scenarios: the aerosol effective radiative forcing (ERF) decreased from −1.3 W m−2 in 2005 to between −0.1 W m−2 and −0.4 W m−2 in 2100. The differences in ERF were substantially larger between LOW (−0.02 W m−2 in 2100) and HIGH (−0.8 W m−2) scenarios. The global mean temperature difference between the simulations with standard RCP aerosol emissions was less than 0.18 °C, whereas the difference between LOW and HIGH reached 0.86 °C in 2061. In LOW, the rate of warming peaked at 0.48 °C per decade in the 2030s, whereas in HIGH it was the lowest of all simulations and never exceeded 0.23 °C per decade. Using present-day population density and baseline mortality rates for all scenarios, PM2.5-induced premature mortality was 2 371 800 deaths per year in 2010 and 525 700 in 2100 with RCP4.5 aerosol emissions; in HIGH, the premature mortality reached its maximum value of 2 780 800 deaths per year in 2030, whereas in LOW the premature mortality at 2030 was below 299 900 deaths per year. Our results show potential trade-offs in aerosol mitigation with respect to climate change and public health as ambitious reduction of aerosol emissions considerably increased warming while decreasing mortality.


Introduction
Aerosol particles have a major influence on global climate (Myhre et al 2013b). They scatter and absorb radiation (Myhre et al 2013a) and are also a key component determining clouds' microphysical and consequently radiative properties (Lohmann and Feichter 2005). Anthropogenic aerosol emissions have had a net cooling effect on climate, although the exact value is highly uncertain (Myhre et al 2013b).
There is also uncertainty in future aerosol emissions; the Intergovernmental Panel on Climate Change (IPCC) used Representative Concentration Pathways (RCPs) to cover a range of possible emission scenarios. RCPs are named after the year-2100 radiative forcing in each scenario in W m −2 (RCP2.6, RCP4.5, RCP6.0, and RCP8.5). All RCPs have a declining trend of aerosol emissions in the 21st century, which would decrease the magnitude of aerosol-induced cooling and thus lead to increased global warming (Shindell et al 2013, Smith and Bond 2014, Westervelt et al 2015. In the long term, cumulative CO 2 emissions mainly determine the magnitude of climate change (Allen et al 2009, Matthews et al 2009, provided that emissions of short-lived climate forcers, including aerosols and methane, are eventually removed (Rogelj et al 2014b, Bowerman et al 2013. Moreover, stringent CO 2 emission controls would lead to significant decreases in aerosol emissions as a co-benefit (Rogelj et al 2014b). However, unmitigated CO 2 emissions could lead to increases in aerosol emissions if strict air quality policies are not enforced (Strefler et al 2014) and aerosol emission pathways can have a noticeable effect on near-term climate change. For example, Chalmers et al (2012) demonstrated that near-term global warming in a low-emission scenario (RCP2.6) was greater than in a moderate-emission scenario (RCP4.5) due to larger sulfate aerosol emission reductions despite lower CO 2 emissions. Differences in aerosol emission pathways can also affect near-term global mean warming rate, which is an important factor in climate change adaptation (Smith et al 2015).
In addition to climate impacts, aerosols affect human health by degrading air quality, resulting in premature mortality via cardiopulmonary diseases and lung cancers (Krewski et al 2009, Lepeule et al 2012, Pope III et al 2002, Burnett et al 2014. Silva et al (2013) estimated that the present-day concentration of particulate matter with a dry diameter less than 2.5 m (PM 2.5 ) due to anthropogenic aerosol emissions is responsible for about 2.1 (95% confidence interval of 1.3-3.0) million deaths annually. Lelieveld et al (2015) estimated an even higher number of 3.15 (1.52-4.60) million deaths annually for present-day outdoor PM 2.5 -caused premature mortality and projected that to increase to 6.2 million deaths annually in 2050 under a business-as-usual scenario. The adverse health effects of PM 2.5 have motivated policies to curb aerosol emissions. Likhvar et al (2015) calculated that maximum feasible reductions of aerosol emissions would lead to a 1.5 (0.4-2.4) million avoided deaths per year in 2030 compared to year 2010. However, RCPs do not span the full range of possible aerosol emission pathways as they have similar underlying assumptions (Rogelj et al 2014a, Fiore et al 2015. Consequently, different levels of future anthropogenic aerosol emissions could involve a trade-off between climate warming and human health. Most previous studies on aerosol-mitigation scenarios have concentrated either on health effects (Silva et al 2016) or relied on simple one-dimensional energy-balance models to assess the climate impacts (Rogelj et al 2015, Bowerman et al 2013, Rogelj et al 2014b, Strefler et al 2014. Stohl et al (2015) used an ensemble of earth system models to assess climate impacts of two scenarios of mitigation of short-lived climate forcers. Butt et al (2016) estimated both the mortality and radiative forcing from aerosol emissions in year 2000, but accounting for residential combustion only. In this study, we assess both the spatially resolved climate and health effects of different emission pathways for anthropogenic aerosols, spanning a wider range of potential emission pathways than the RCPs, and using a combination of global aerosol-climate model and an earth system model of intermediate complexity.

Modeling approach
We used a combination of two global climate models to calculate the effects of different aerosol emission scenarios on climate and air quality. To simulate aerosol radiative forcing and aerosol concentrations, we used the aerosol-climate model ECHAM-HAMMOZ (ECHAM6-HAM2.1) (Zhang et al 2012, Stevens et al 2013. ECHAM-HAMMOZ is a general circulation model that solves atmospheric circulation and physics at a horizontal resolution of T63 (roughly 1.9 • ×1.9 • ) and accounts for emissions, transport, removal, and microphysics of aerosol particles. The aerosol module includes aerosol species of sulfate, organic carbon (OC), black carbon (BC), sea salt, and mineral dust. Excluding nitrate, ammonium, and secondary organic aerosols means that anthropogenic aerosol forcing and PM 2.5 contribution are likely lower than they would be if these species were included (e.g. Scott et al 2014, Xu andPenner 2012). The model includes both aerosol direct and indirect effects (Lohmann et al 2007) with a semi-empiric cloud droplet activation scheme (Lin and Leaitch 1997). We conducted the ECHAM-HAMMOZ simulations with nudging of the model meteorology towards ERA-Interim reanalysis data (Dee et al 2011(Dee et al ) and prescribed monthly climatological (1979(Dee et al -2008 sea surface temperature (Hurrell et al 2008). We used aerosol emissions from the ACCMIP Emission database (http://accmip-emis.iek.fz-juelich. de/data/accmip/gridded_netcdf/accmip_interpolated/ T63/) for the historical (Lamarque et al 2010) and RCP (van Vuuren et al 2011a) periods. Natural sea salt, dust, and dimethyl sulfide emissions (DMS) were calculated online. Sea salt emissions were based on a combination of source functions by Monahan et al (1986) and Smith and Harrison (1998) (see Guelle et al (2001) for details), dust emissions were calculated according to Tegen et al (2002) with East Asia soil properties adopted from Cheng et al (2008) (see Zhang et al (2012) for details), and DMS emissions were based on Nightingale et al (2000). For a detailed evaluation of the aerosol concentrations in ECHAM-HAMMOZ, we refer the reader to Zhang et al (2012) and Partanen et al (2013).
ECHAM-HAMMOZ is computationally very expensive and cannot simulate long-term impacts on climate owing to its reliance on prescribed sea surface temperatures. Therefore, we used the University of Victoria Earth System Climate Model (UVic ESCM version 2.9) (Weaver et al 2001, Eby et al 2009 with aerosol forcing from the ECHAM-HAMMOZ runs to assess the wider climate implications of different aerosol emission pathways. The UVic ESCM is an intermediate complexity earth system model that includes a one-layer atmosphere with monthly-varying prescribed winds and simple energy-moisture balance equations, three-dimensional ocean circulation with 19 vertical levels, sea ice, terrestrial vegetation, and ocean biogeochemistry. The model is thus capable of simulating long-term effects on climate and carbon cycle, and its results are comparable to more complex earth system models (Eby et al 2013, Joos et al 2013, Arora et al 2013.

Experiment design
Our main interest in this study was the effects of aerosol emissions during the 21st century. However, we also conducted a historical simulation from year 1850-2000. First, we simulated the aerosol forcing with ECHAM-HAMMOZ at 10 year intervals. For each of these time points, we ran ECHAM-HAMMOZ continuously for three years with an additional three-month spinup period. This approach is computationally cheaper than running the model for each year with different emissions and also reduces noise from inter-annual variability in the results. Then, we calculated the mean effective radiative forcing (ERF) including both shortwave and longwave contributions for each calendar month from the three-year runs as the mean difference in net radiation at the top-of-the-atmosphere between a given emission scenario and a reference run with year-1850 anthropogenic and non-agriculture biomass burning aerosol emissions (ERF takes into account fast feedbacks such as changes in cloud cover and is thus a better predictor of temperature change (Haywood et al 2009, Lohmann et al 2010, Myhre et al 2013b). In this way, ERF for each month and grid-cell for a given year was calculated as an average of the same month over three consecutive years. Due to the interpolation, the temporal patterns showed less year-to-year variability than it would have been with every year simulated individually. Mineral dust and sea salt emissions were calculated online identically in all simulations. We used linear interpolation to calculate the aerosol forcing for the years between the 10 year intervals, and applied that forcing into the one-layer atmosphere of the UVic ESCM from 1850 until 2000; the same 10 year linear interpolation approach was also used for simulations covering the 21st century.
For simulations over the years 2001-2100 with the UVic ESCM, we chose RCP4.5 (Thomson et al 2011) as the baseline scenario. Thus, emissions of CO 2 and forcings from non-CO 2 greenhouse gases came from the RCP4.5 scenario in all of our UVic ESCM simulations and the only difference between the runs was aerosol forcing. To explore the range of different aerosol emissions across RCPs, we did simulations using aerosol forcing (calculated with ECHAM-HAMMOZ) from RCP2.6 ( van Vuuren et al 2011b), RCP4.5, or RCP8.5 (Riahi et al 2011). We designated these scenarios as 2.6AER, 4.5AER, and 8.5AER, respectively. We excluded RCP6.0 as its emissions fall between those of RCP4.5 and RCP8.5, and we only wanted to cover the range of possible aerosol emissions in the RCPs and not necessarily study each individual RCP.
Since the RCPs do not span the full range of conceivable aerosol emission scenarios (Rogelj et al 2014a, Fiore et al 2015, we built two idealized sensitivity scenarios to represent a very aggressive aerosol mitigation scenario and a scenario where aerosol emissions are allowed to grow with increasing CO 2 emissions. These new sensitivity scenarios allow us to assess climate and health effects from a wider range of possible outcomes. The sensitivity scenarios are based on RCP4.5, modifying only the anthropogenic aerosol emissions (i.e. keeping the same emissions from biomass burning, sea salt, and mineral dust as in 4.5AER). For each pair of sector and species of anthropogenic aerosol emissions (e.g. BC from agricultural waste burning or SO 2 from land transport), we calculated the ratio of the aerosol emissions and global CO 2 emissions at each year. For the high-end sensitivity scenario (designated as HIGH), we assumed that these ratios stayed at the level of the year 2005 throughout the 21st century. For the low-end sensitivity scenario (designated as LOW), we assumed that these ratios would reach a minimum level in 2030 (i.e. minimum ratio of aerosols per global CO 2 emissions found across RCPs 4.5, 6.0, and 8.5) and stay at that level afterwards. We did not consider the ratios in RCP2.6 as the global net CO 2 emissions become negative in this scenario. The ratios for the years 2010 and 2020 were linearly interpolated using values of 2005 and 2030. To calculate the actual aerosol emissions for HIGH and LOW in a given year, we multiplied the previously calculated ratios with global CO 2 emissions for that year and then distributed these emissions spatially based on RCP4.5 for that same year. Figure 1 shows the global total anthropogenic emissions of SO 2 , BC, and OC for all aerosol emission scenarios used in this study. Scaling aerosol emissions with CO 2 emissions is based on assumptions that aerosols are co-emitted with CO 2 and that sectoral CO 2 emissions follow the global trend. Even though these assumptions are less valid for some sectors, LOW and HIGH scenarios appear appropriate in the context of the current global-level sensitivity study of different aerosol emission pathways.

Calculation of aerosol health effects
We used the surface aerosol concentrations from ECHAM-HAMMOZ simulations to calculate premature mortality caused by chronic exposure to ambient PM 2.5 resulting from anthropogenic aerosol emissions. We mainly followed the methodology by Ostro (2004) with cause-specific relative risk calculated with the Integrated Exposure-Response (IER) model by Burnett at al (2014). For the reference (counterfactual) scenario, we made a simulation without even the minimal year-1850 anthropogenic aerosol emissions, but with year-1850 biomass burning, sea salt, and mineral dust aerosol emissions (this differs from the reference scenario for ERF calculations, which had year-1850 anthropogenic and non-agriculture biomass burning aerosol emissions as explained in section 2.2). For each simulation, we diagnosed the three-year mean surface PM 2.5 .
We estimated the PM 2.5 -caused mortality for ischemic heart disease (IHD), cerebrovascular disease (stroke), chronic obstructive pulmonary disease (COPD), and lung cancer (LC). To get the anthropogenic contribution, we subtracted the mortality calculated from the counterfactual scenario from the results of other scenarios. We calculated the total annual premature mortality rate E (deaths per year per unit area) using the following formula: where AF , is the attributable fraction in age group i for cause j, B , is the baseline mortality rate (deaths per year per capita), and P is the exposed population in age group i. AF , is 1−1/RR, where relative risk is based on the IER model by Burnett et al (2014). Burnett et al (2014) base their model on the following equation: (2) where z is the PM 2.5 concentration (in g m −3 ) in a given year in a given aerosol emission scenario and z cf is the counterfactual concentration that is the lower limit for any adverse health effects. Burnett et al (2014) provide an ensemble of 1000 sets of parameters , , , and z cf for each affected age group i and cause j. Following the example by Apte et al (2015), we evaluated equation (2) with all different ensemble values at PM 2.5 concentrations up to 409.9 g m −3 with an interval of 0.1 g m −3 , and created a lookup table for the ensemble mean and upper and lower bounds of the 95% confidence interval (defined by 2.5% and 97.5% quantiles) of RR. The use of ensemble mean and lower and upper bounds relies on the assumption that the 1000 parameter sets provided by Burnett et al (2014) are equally probable.
Projected changes in demographics will have a major impact on PM 2.5 -caused mortality, even if the air quality itself does not change (Silva et al 2016).
Since we wanted to focus on the health implications of changing aerosol concentrations only, we used constant baseline mortality rates in six different world regions (WHO 2008) (see figure 2 of Partanen et al (2013) and year-2010 population density data (SEDAC 2005) for all time intervals. Using the same baseline mortality rate for all countries within a single WHO region hides some spatial variation, but on a global level the differences partly cancel each other and we consider these limitations to be acceptable in the context of this global-scale sensitivity study. WHO (2008) provides baseline mortality rates and fractions of age groups of 30-44, 45-59, 60-69, 70-79, and 80+. We calculated the mortality caused by IHD and stroke separately for these age groups and by COPD and LC for population over 30 years of age. Since Burnett et al (2014) provide relative risks for more fine-grained age groups, we used the average of their relative risks to calculate the relative risk for each of our age group (e.g. relative risk for 30-44 was calculated as the average of age groups of 30-34, 35-39, and 40-44). Figure 2 shows the global mean ERF from changes in aerosol emissions through time (although the results include also biomass burning emissions, most of the effect comes from changes in anthropogenic aerosols). The maximum negative ERF of −1.41 W m −2 in the historic run occurred in 1990. In 2011, we obtained values between −1.19 and −1.30 W m −2 for the three AER simulations, which compares to the best estimate of −0.9 (−1.9 to −0.1) W m −2 from the latest IPCC report for the aerosol ERF in 2011 relative to 1750 (Myhre et al 2013b). The year-2000 global mean ERF (−1.31 W m −2 ) was also somewhat stronger than the multi-model mean of −1.17 W m −2 obtained by Shindell et al (2013). To assess the effect of interannual variability on ECHAM-HAMMOZ results, we calculated the annual ERF for each of the three individual years used to compute the mean year-2000 ERF (as explained in section 2.2, results for each 10 year interval were based on running ECHAM-HAMMOZ continuously for three years). The annual ERF ranged between −1.28 and −1.36 W m −2 . Thus, the effect of ECHAM-HAMMOZ interannual variability appears substantially smaller than inter-model uncertainty in aerosol ERF.

Aerosol radiative forcing
The transient global mean ERF was fairly similar for the three AER simulations. The magnitude of aerosol ERF decreased steadily to between −0.15 W m −2 and −0.39 W m −2 in 2100. The largest difference across the AER simulations was 0.34 W m −2 in 2040. These results are similar to results by Westervelt et al (2015). The differences were much larger for our low-and highend estimates. In LOW, the magnitude of aerosol ERF decreased quickly and was only −0.14 W m −2 in 2030. The subsequent decline (driven by decreasing global CO 2 emissions) was slower, reaching −0.02 W m −2 by 2100. In HIGH, on the other hand, the magnitude of aerosol ERF continued to increase until 2040 due to the increase of CO 2 emissions, reaching a peak value of −1.63 W m −2 . After 2040, the magnitude of aerosol ERF started to decline but it was still −0.82 W m −2 in 2100.
Aerosol forcing was strongest in the northern hemisphere and especially over Eastern Asia and over oceans nearby the shipping routes. In 2005 for 4.5AER, the aerosol forcing was strongest in the northern hemisphere and especially over Eastern Asia and over oceans, reaching values of about −12 W m −2 in these regions (figure S1(a)). In 2100 for 4.5AER, northern hemisphere oceans, East Asia, central Africa and the South Atlantic Ocean near the African coast had noticeable negative aerosol ERF (figure S1(b)). In the same year for LOW, most areas had negligible ERF as negative and positive values mostly took place next to each other (figure S1(c)). However, the east coast of North America experienced positive aerosol ERF in 2100 for both 4.5AER and LOW. This positive forcing was most likely caused by the substantial reduction in biomass burning emissions over the region. The forcing pattern in 2100 for HIGH (figure S1(d)) was generally similar to, although larger in magnitude than, 4.5AER. The spatial patterns of forcing in 2.6AER and 8.5AER were also similar to 4.5AER.

Impacts on surface temperature
As with aerosol forcing, the effects on global mean surface temperatures across AER scenarios were similar, but results differed considerably for LOW and HIGH ( figure 3(a)). The global mean surface temperature difference between the AER simulations remained less than 0.18 • C, year-2100 difference being 0.17 • C. The difference in end-of-the-century (2096-2100) global mean surface temperature between 2.6AER and 8.5AER was 0.17 • C, comparable to the corresponding aerosol-caused temperature differences between RCP2.6, RCP4.5, and RCP8.5 scenarios by Westervelt et al (2015) (0.07 • C-0.23 • C). On the other hand, the difference between LOW and HIGH was 0.86 • C at its maximum (in 2061) and 0.61 • C in 2100. It is also notable that simulated warming diverges quite rapidly over the years 2015-2040 in the LOW and HIGH scenarios, which is not the case among the AER scenarios. Previous analyses of contributions to uncertainty in future warming projections have suggested the scenario uncertainty remains quite low over the coming decades (e.g. Booth et al 2013); by contrast, our results suggest that considering a wider range of plausible aerosol scenarios would considerably increase the contribution of scenario uncertainty to near-term temperature projections.
We also calculated global mean warming rate as a 10-year running mean of global mean surface temperature change over time ( figure 3(b)). In all simulations, there was an initial peak in the warming rate (about 0.4 • C per decade) in 2002 due to rapid warming in the last years of the historical simulation (not visible in figure 3(a)). Among the AER simulations, 2.6AER had the highest warming rate with a maximum value of 0.34 • C per decade in 2037. For most of the second half of the century, 4.5AER had the highest warming rate among AER scenarios. In LOW, the warming rate increased quickly during the first decades with rapidly decreasing aerosol emissions, reaching a maximum value of 0.48 • C per decade in 2032. This was 81% higher than the warming rate with 4.5AER in the same year. On the other hand, in HIGH, the increase of aerosol negative ERF until 2040 and the gradual decrease afterwards (figure 2) resulted in the most stable warming rate of all the simulations, remaining below 0.2 • C per decade between 2008 and 2058. Figures 3(c) and (d) show the spatial differences in mean surface temperature for 2090-2100 between LOW and 4.5AER, and between HIGH and 4.5AER, respectively. These differences were very similar with respect to magnitude and spatial patterns, but had opposite signs. The largest differences were in high-latitude regions, particularly in the northern hemisphere, with surface temperature difference between LOW and HIGH reaching 1.3 • C at the North Pole. Although ERF was highly localized in specific regions, the surface temperature differences created by different aerosol emissions were fairly smooth, particularly longitudinally.

Impacts on health
We calculated the anthropogenic surface PM 2.5 contribution as the difference between a given scenario and the preindustrial year-1850 simulation without any anthropogenic emissions. The largest anthropogenic  figure S2(b)) was similar or slightly larger in North America and Europe, comparable in South America and the polluted regions of Asia, slightly higher in Northern Africa, and slightly lower in parts of Central Africa. PM 2.5 concentrations in 2005 for 4.5AER (figure S2(c)) were largely similar to those of Brauer et al (2012) (their figure 2), who combined modeling to ground and satellite measurements. The largest differences were in Northern Africa and Australia, where our model predicted higher PM 2.5 concentrations, most probably due to differences in dust emissions. Figure 4(a) shows global total premature mortality due to PM 2.5 in the different aerosol emissions scenarios for the central estimates of the RR values (see below for an uncertainty analysis). As discussed in section 2.3, these numbers represent what the Figure 4. (a) Global total annual premature mortality due to PM 2.5 using year-1850 aerosol emissions without anthropogenic contribution as the reference, and spatial differences in annual premature mortality due to PM 2.5 in 2100 between (b) LOW and 4.5AER and (c) HIGH and 4.5AER. Changes in mortality were due to changes in PM 2.5 alone, as we used present-day population and baseline mortality rates throughout. mortality would be with present-day population density and baseline mortality rates, and cannot therefore be directly interpreted as future predictions. Taking into account the combined effect of population growth and changing baseline mortality rates would give higher global total future premature mortality (Silva et al 2016). Larger mortality in all simulations would also mean that the differences between our scenarios would be larger. For example, LOW would lead to larger avoided mortality when compared to the other scenarios. Geographically, a larger fraction of global total mortality would take place in Africa due to the fastest projected population growth.
The PM 2.5 -caused mortality for 4.5AER in 2010 was about 2 371 800 (95% confidence interval of 1 329 300-2 925 200) deaths per year. As aerosol emissions decreased in all AER scenarios, premature mortality due to PM 2.5 also decreased steadily and similarly across scenarios, and was only about 525 700 (258 400-757 300) deaths per year for 4.5AER in 2100. 2.6AER had the lowest mortality across all AER scenarios. In HIGH, mortality increased until 2030, when it reached 2 780 800 (1 607 000-3 366 300) deaths per year. HIGH reached its minimum mortality in 2080 (1 150 900 (608 800-1 530 800) deaths per year), followed by a small increase as the emissions increased slightly until the end of the century. PM 2.5 -caused premature mortality in LOW decreased quickly along with aerosol emissions. It was below 300 000 deaths per year from 2030 onwards and went very close to zero (below 30 000 deaths per year) from 2080 onwards. Throughout the simulations, the uncertainty bounds of the AER simulations were overlapping with each other (table S1). On the other hand, the higher bound of LOW was below the lower bound of HIGH from year 2030 onwards. This shows that the AER scenarios are fairly similar to each other, whereas LOW and HIGH differ from each other significantly.
Even though the global mortality in all simulations was positive throughout the 21st century, the mortality was below zero in some regions near the end of the century. For example, in LOW, the mortality in 2100 was negative in India, eastern parts the United States, and in parts of South America ( figure S3(a)). The negative (i.e. avoided) mortality means that in LOW, PM 2.5 concentrations in these regions were below the reference level from year-1850 without anthropogenic aerosol emissions. This happened because biomass burning aerosol emissions in LOW, which were prescribed according to RCP4.5, were in some regions lower than in the year-1850 reference simulation. The largest difference in PM 2.5 was due to differences in organic carbon emissions ( figure S3(b)).
Looking at the spatial pattern of mortality, the differences between LOW (figure 4(b)) or HIGH (figure 4(c)) vs. 4.5AER were mostly mirror images of each other; regions with a negative difference in mortality for LOW had a positive difference of similar magnitude for HIGH. The differences were largest in heavily populated areas of Asia and Europe.
Our estimate of the present-day (2000) mortality (2 263 500 (1 267 500-2 815 300) deaths per year) was about 8% higher than the estimate of 2.1 million deaths per year by Silva et al (2013Silva et al ( ) (years 2000Silva et al ( vs. 1850 in their study) and 33% higher than the estimate of 1.7 million deaths per year by Silva et al (2016). The main reason for the higher mortality in our study was most likely that we defined the reference scenario completely without anthropogenic aerosol emissions, whereas Silva et al (2013Silva et al ( , 2016 used year-1850 emissions that include a small anthropogenic contribution. Therefore, our estimate for anthropogenic PM 2.5 concentration was higher over North America and Europe (figure S2(a)) than in Silva et al (2013). To better compare with the results by Silva et al (2013Silva et al ( , 2016, we recalculated the mortality by including anthropogenic emissions in the year-1850 reference simulation. This resulted in mortality of 1 831 700 (1 067 300-2 178 900) deaths per year for the year 2000, which is very close to the estimate by Silva et al (2016), who used the same concentration-response function as we did. On the other hand, our present-day PM 2.5 -induced mortality estimate is smaller than the 3.15 (1.52-4.6) million deaths per year estimated by Lelieveld et al (2015). Our results can also be compared to those of Likhvar et al (2015) who estimated that maximum feasible reductions of aerosol emissions would lead to 1.5 (0.4-2.4) million avoided deaths per year in 2030 compared to 2010. A comparable figure in our study is the difference in mortality between 4.5AER in 2010 and LOW in 2030 (2.1 (1.2-2.5) millions of deaths per year). Johnston et al (2012) found that results varied by a factor of 2.3 when using two different functions to quantify PM 2.5 -induced mortality resulting from landscape fires. To test the sensitivity of our results to the choice of the concentration-response function, we calculated the relative risk also using the functions given by Ostro (2004) (see supplemental material available at stacks.iop.org/ERL/13/024028/mmedia). These results did not differ considerably from the results using the IER model by Burnett et al (2014) (compare tables S1 and S2). When using Ostro (2004) to calculate relative risk, the annual global total mortality was between 12% and 27% higher than with the IER model; also, 95% confidence intervals were wider with Ostro (2004) than with the IER model.
Our premature mortality results are likely to be biased low due to the coarse model resolution. Kodros et al (2016) showed that decreasing model resolution from 2 • × 2.5 • to 4 • × 5 • decreased by 16% global total mortality caused by aerosols from domestic combustion. Lower mortality with coarser resolution is caused by potential dilution of PM 2.5 in larger grid cells as well as reduced overlap of densely populated areas and high PM 2.5 concentrations (Kodros et al 2016).

Connections between climate change and mortality
As discussed above, ambitious reductions in aerosol emissions resulted in both increased warming and greatly decreased PM 2.5 -induced premature mortality. This potential trade-off is highlighted in figure  5, where we plot the trajectories of surface temperature change and premature mortality. Although the surface temperature increase and premature mortality decrease were faster in LOW, the situation in year 2100 was not very different from that in the AER simulations. On the other hand, the trajectory in HIGH was considerably different, especially with increasing premature mortality during the first decades of the 21st century.
We assessed the premature mortality from changes in aerosol emissions only, omitting possible feedbacks that climate change could have on air quality (Fiore et al 2015, von Schneidemesser et al 2015. Greenhouse gas emission reductions can also lead to decreases in aerosol emissions (West et al 2013), which mitigates the trade-offs presented here. Aerosol emissions could also affect human mortality indirectly; in particular, heat waves are known to increase premature mortality (Mitchell et al 2016) and aerosol-induced cooling could reduce heat wave intensities (Péré et al 2011). Comparing premature mortality due to heat waves to premature mortality due to PM 2.5 exposure would prove particularly interesting and should be the topic of future studies. Unfortunately, existing exposure-response functions for mortality caused by heat waves require city-specific coefficients and daily mortality data that are not available for a global study. For example, there is evidence that the temperature threshold above which heat waves cause mortality can vary by more than 10 • C across specific cities  (Baccini et al 2008), so extrapolating the relationships available for a handful of cities to the global scale appears risky at best. Moreover, the coarse spatial and minimal internal variability of the UVic ESCM would make it difficult to diagnose heat waves from the model results. Simplistic back-of-the-envelope calculations could be made at a global level by assuming linear relationship between global mean surface temperature change and climate-change caused mortality (Löndahl et al 2010), but we deemed this approach too unreliable to calculate meaningful numbers.

Conclusions
Our results demonstrate that the three RCPs studied (2.6, 4.5 and 8.5) have similar global trends of decreasing climate effects of aerosol emissions, reflecting similar tightening air-quality policies and thus similar trends in aerosol emissions; such results are in line with Westervelt et al (2015). Similarly, the adverse health effects caused by aerosol emissions decline in the three RCPs, in agreement with the results of Silva et al (2016). To explore a larger spectrum of conceivable aerosol emission scenarios, we made idealized low-and high-end estimates for aerosol emissions. Simulations with these emission scenarios showed that if aerosol emissions increase noticeably above presentday levels, PM 2.5 -caused premature mortality might increase substantially in the future while the climate would warm much more slowly. On the other hand, very strict and rapid reductions in aerosol emissions would dramatically improve air quality, but also lead to a rapid warming as the net cooling effect of aerosols disappears. Simultaneous reductions in emissions of non-CO 2 greenhouse gases like methane would be required to minimize this rapid warming effect. The most sustainable path forward likely involves a strong mitigation of CO 2 emissions, which would reduce aerosol emissions and the associated PM 2.5 -caused mortality as a co-benefit (Rogelj et al 2014b); in such a low-warming future, differences among air quality policies would become much reduced (Strefler et al 2014), making the trade-off between climate warming and human mortality much less critical. ETH Zurich, Max Planck Institut für Meteorologie, Forschungszentrum Jülich, University of Oxford, the Finnish Meteorological Institute and the Leibniz Institute for Tropospheric Research, and managed by the Center for Climate Systems Modeling (C2SM) at ETH Zurich. We thank Anni Karhunen for rewriting the mortality scripts into Python.