Simulations were conducted using the CESM2(WACCM6) , a fully coupled ocean-atmosphere model with interactive tropospheric and stratospheric chemistry and aerosols. The model simulates aerosol formation and growth through an interactive, two-moment modal aerosol microphysics scheme (MAM4; ), allowing sulfate aerosols to evolve over time based on the simulation of nucleation, coagulation, condensation and removal processes. However, MAM4 uses assumptions of internal mixing for the size distribution of different species, whereas mass is tracked separately . While stratospheric and tropospheric chemistry are fully interactive, photolysis rates are calculated using lookup tables, taking into account the overhead ozone column and clouds but excluding the effects of aerosols , thus excluding the direct effect of the aerosols on actinic fluxes .

Simulated PM_2.5 components include sulfate (SO₄), secondary organic aerosols (SOA), primary organic matter (POM), salt, dust and black carbon (BC). However, the model does not include explicit ammonium or nitrate aerosol chemistry, which can be important contributors to PM_2.5, particularly in ammonia-rich regions . This omission may lead to an underestimate of absolute PM_2.5 concentrations and associated health impacts in certain areas, and future work should aim to understand if such an omission could impact heavily our conclusions: interactions between the formation of nitrate and sulfate aerosols are complex , and while recent observations have shown that it is the absence of sulfate aerosols that favors fine particulate nitrate formation in some environment , more work is needed to understand what impact this would have under SAI scenarios. Furthermore, has indicated that estimates of PM_2.5 may overestimate the contribution of dust due to the inclusion of larger dust particles.

These limitations notwithstanding, the interactive chemistry–climate framework of WACCM allows us to capture coupled meteorological, chemical, and radiative feedbacks that are central to evaluating the air quality response to stratospheric aerosol injection . CESM2(WACCM6) has been evaluated against earlier model versions and observations – including NASA ATom aircraft profiles , Tropospheric Ozone Assessment Report (TOAR) surface ozone data , and Measurements of Pollution in The Troposphere (MOPITT) carbon monoxide observations – showing good agreement with ozonesonde data and seasonality of surface ozone, though with some regional spatial biases. Previous evaluations have also shown that WACCM reproduces the large-scale distributions of tropospheric ozone and key pollutants, as well as climatological patterns of aerosols, with skill comparable to other climate models . These assessments further support the suitability of this model for investigating the relative changes in air quality under SAI.

The baseline ensemble (i.e. without SAI) follows the Shared Socioeconomic Pathway 2 with middle-of-the-road increases in greenhouse gas emissions, leading to a radiative forcing of 4.5 W m⁻² by 2100 , and is hereafter referred to as SSP2-4.5. In the ARISE-SAI-1.5 (see Supplement Fig. S2) ensemble, under the same emission scenario, SO₂ is injected annually at four fixed latitudes (15° N, 15° S, 30° N, 30° S) at approximately 21.5 km of altitude starting in year 2035, and run until 2070, with injection rates adjusted at the beginning of each year to offset continuing warming under the SSP2-4.5 emissions pathway, with the aim of maintaining global mean surface temperatures and their large-scale gradients at the 1.5 °C above preindustrial level (defined as the mean over 2020–2039 to ensure better consistency with other climate models, ). The ARISE-SAI-1.0 simulations follow the same protocol, but SAI is used to cool by a further 0.5 °C compared to the targets in ARISE-SAI-1.5.

A 10-member ensemble is produced for all three cases to account for internal climatic variability . For regional assessments of mortality and mortality-related factors, we will focus our analyses on the ARISE-SAI-1.5 case, whereas results from the ARISE-SAI-1.0 will be provided for global, temporal and injection-related analyses in order to highlight the linearity (or lack thereof) of the SAI response with the injection rates.

Here we describe how we calculated the impact on mortality rates attributed to changes in the simulated changes in ambient surface PM_2.5 and ozone. All mortality estimates in future scenarios are calculated using the fixed 2020 population distribution: this approach isolates the effects of air quality changes by removing confounding influences from projected population growth or redistribution.

Mortality is estimated using the health impact function : 1Mi,d,a,t=BMRd,a,t×Pi,a,2020×AFi,d,a,t where M is the mortality for CESM grid i from disease d for age group a and year t; P is the number of population in 2020 with each age group a in grid i; BMR is the national base mortality rate for disease d, age group a and year t; AF is the attributable fraction which estimates the proportion of deaths in a population that can be attributed to a specific exposure to disease d or risk factor from epidemiological studies. For PM_2.5, we use the AF associated with noncommunicable diseases and lower respiratory infections (NCD + LRI). For ozone, we use the AF associated with cardiovascular and respiratory diseases. Whereas previous studies attributed PM_2.5 exposure to cardiovascular and respiratory diseases and ozone exposure solely to respiratory diseases, here we attribute cardiovascular disease to ozone exposure, which aligns with more recent epidemiological findings and improves the completeness of ozone-related health impact assessments.

For PM_2.5, AF is calculated using the exposure–response function from the Global Exposure Mortality Model (GEMM; ), which provides improved estimates across a wide range of ambient PM_2.5 concentrations. GEMM is particularly effective in low-income and high-pollution regions where the older Integrated Exposure–Response (IER) functions tend to underperform due to limited observational data and less robust extrapolation at high exposure levels : 2AFi,d,a,t=1-1RRi,d,a,t;whereRRi,d,a,t=expθ×log⁡(Ci,tα+1)1+exp⁡-Ci,t-μvandRRi,d,a,t=1whenCi,t<2.4µgm-3 where C is the ambient PM_2.5 concentration (µg m⁻³); RR is the relative risk of morality at any concentration; θ, α, μ and v are empirical coefficients from the GEMM which are specific for each age group.

For ozone-attributable mortality, we convert surface ozone to the ozone season maximum daily 8 h average (OSMDA8; ppb) using hourly surface ozone data for each experiment and each ensemble member. OSMDA8 calculates the highest 6-month rolling average daily 8 h average ozone concentration, which reflects the highest average ozone concentration over a 6-month period. OSMDA8 is the metric used by the Global Burden of Disease for quantifying the health effect from long-term ozone exposure and is used in the World Heath Organization's air quality guidelines . To calculate the ozone-attributable risk fraction, we calculate the AF for cardiovascular and respiratory disease separately and then combine the associated mortality. 3AFi,d,a,t=1-exp⁡-β(Xi,t-Xmin);whereAF=0whenXi,t<Xmin where X represents the spatially and temporally resolved grid-cell level OSMDA8; Xmin represents the theoretical minimum risk exposure concentration and β represents a model-parameterized slope of the log-linear relationship between concentration and health from epidemiological studies. For chronic respiratory disease mortality, we apply a β of ln⁡(1.06) per 10 ppb ozone (95 % confidence interval (CI) 1.03–1.10) derived by GBD 2019 . For cardiovascular disease mortality, we apply a β of ln⁡(1.028) per 10 ppb ozone (95 % CI 1.010–1.047) . A summary of the RR and disease d used to calculate mortality associated with PM_2.5 and ozone is provided in Table . For more detail, Fig. S1 in the Supplement shows how AF changes with increasing PM_2.5 and ozone concentrations.

Our BMRs are drawn from the International Futures (IFs) health model, providing dynamic, age and disease-specific mortality projections consistent with policy interventions following the SSP2-4.5 scenario . The IFs health model is a comprehensive, integrated modeling platform used to explore long-term global health dynamics. This represents a more realistic approach compared to the use of static BMRs in previous studies (e.g. ).

Population (P) for each age group was calculated by using the global population density dataset based on Shared Socioeconomic Pathways (SSP) and the ratio of the population for each age group to the total population retrieved from the SSP database developed by the International Institute for Applied Systems Analysis and the National Center for Atmospheric Research for each country. The raster of nation-states was retrieved from the Gridded Population of the World, Version 4 (GPWv4): National Identifier Grid and is used to aggregate the calculated mortality to country-level mortality estimates. We further categorize the world into 21 regions following the GBD Study based on epidemiological similarities and geographic proximity.

Other studies estimating air pollution-related mortality have typically calculated mortality uncertainty based on the central intervals of the parameters used in the AF calculations . However, less attention has been given to the uncertainty arising from internal model variability: this is important as internal variability can drive regional air quality differences . Thus, our analysis account for uncertainty arising from climate ensemble spread, while applying confidence intervals for β (for ozone) and RR (for PM_2.5).

Consistent with previous studies , ARISE-SAI-1.5 exhibits an overall reduction in precipitation relative to the increase observed in SSP2-4.5 (Fig. f): these changes are due to both the avoidance of the temperature-related Clausius-Clapeyron increase expected under climate change, as well as to changes in the strength and position of the Intertropical Convergence Zone and the Hadley circulation in the different scenarios. While some regions do not exhibit statistically significant changes in surface PM_2.5 relative to SSP2-4.5 (2030–2039), other areas, such as Central America and central Sub-Saharan Africa, show statistically significant reductions. In these regions, PM_2.5 reductions coincide with increases in precipitation (Fig. d–f), suggesting that enhanced wet scavenging may play a role. However, the overall broader spatial pattern of PM_2.5 changes does not consistently track precipitation trends (Fig. g, h and j), and are not statistically significant, indicating that internal variability, circulation changes, vertical mixing, and/or aerosol-cloud interactions likely contribute to changes in PM_2.5. Thus, while precipitation influences PM_2.5 in some regions, it does not fully explain the spatial distribution or statistical significance of PM_2.5 changes, and many features of the PM_2.5 response should be interpreted carefully due to limited ensemble robustness.

Figure  indicates that dust and SOA, rather than sulfate, are the dominant contributors to total PM_2.5 concentrations across most regions in ARISE-SAI-1.5. Considering, however, that CESM2(WACCM6) is known to overestimate overall dust concentrations e.g.,, it is possible that this may contribute to dust appearing as the dominant PM_2.5 species in many regions, and this should therefore taken into account when interpreting Fig. . This potential upper bias does not affect the qualitative conclusion that sulfate is not the primary driver of PM_2.5 changes in our simulations, but it does mean that the relative prominence of dust should be interpreted with caution.

While some of the edges of the geographical ranges where each species dominates are stippled, indicating that fewer than 90 % of ensemble members agree at the grid level, this ensemble uncertainty does not alter the overall conclusion that non-sulfate species dominate global PM_2.5.

Under SSP2-4.5 (not shown), the spatial distribution of the dominant PM_2.5 species is broadly similar to Fig. 2.; SO₂ does not generally emerge as the dominant species except in particularly pristine environments (e.g., high latitudes; ) or in regions that are already extremely polluted. While it is true that sulfate can still drive relative changes in PM_2.5 even when not dominant in absolute terms, our subsequent analysis of mortality (Sect. 3.2) shows that the changes in PM_2.5 concentrations and PM_2.5-related mortality are not driven by sulfate. Specifically, the spatial and temporal patterns of PM_2.5-related mortality changes align more closely with changes in non-sulfate species and are shaped by precipitation and circulation-driven effects such as wet scavenging and regional aerosol transport.

Figure j–l shows % changes in surface ozone exposure. Interpreting these changes requires accounting for multiple mechanisms, including SAI-induced impacts on stratospheric ozone and its transport to the surface, and changes in ozone in-situ photochemical processing driven by changes in temperature and photolysis. SAI influences stratospheric ozone through multiple pathways, including alterations in heterogeneous chemical reactions on aerosol surfaces, modifications in photolysis rates due to changes in actinic flux from changes in the overhead ozone column and aerosol absorption and scattering, and dynamical changes in stratospheric circulation and temperature patterns that can impact ozone transport and distribution . Injection strategy also plays a key role: in ARISE-SAI-1.5, SO₂ is injected primarily in the Southern Hemisphere (SH) during 2060–2069 to restore hemispheric temperature gradients affected by the asymmetric warming in the underlying SSP2-4.5 simulations, resulting in an asymmetric stratospheric aerosol burden and consequently an asymmetric ozone response . However, as discussed before, our study does not include tropospheric chemistry changes caused by direct aerosol-driven changes in photolysis. As a result, our analysis does not capture potential tropospheric ozone responses caused by aerosol scattering . A study by using a previous version of WACCM (WACCM4), but modified to include online Tropospheric Ultraviolet and Visible (TUV) model calculations, showed that the exclusion of aerosol optical depth from the TUV calculations only resulted in a small difference in the overall ozone column changes due to minimal differences in the overall ozone loss rates, leading us to conclude that this shortcoming in our simulations is not likely to significantly impact our conclusions.

This stratospheric asymmetry propagates to the troposphere. Specifically, ozone concentrations exhibit a robust hemispheric asymmetry: decreases occur throughout much of the SH troposphere, while increases appear across the Northern Hemisphere (NH) (Figs. l and f). These hemispheric differences arise from distinct underlying mechanisms. In the SH, the reduction in surface ozone is primarily driven by aerosol-driven catalytic ozone loss in the Antarctic stratosphere alongside any changes in polar vortex strength and large-scale stratospheric transport , and the resulting reduction in STE.

In contrast, the NH surface ozone increases are likely not driven by changes in STE. Although stratospheric ozone increases occur in the NH lower-to-mid stratosphere, this signal does not extend to the surface. Hence, the NH surface ozone changes likely reflect the SAI-induced changes in in-situ tropospheric chemical processing. In particular, water vapor concentrations decrease in the troposphere in ARISE-SAI-1.5 compared to SSP2-4.5 (Fig. 3l) as the result of large scale near-surface cooling (Fig. 1c). This reduces chemical ozone loss in the free-troposphere, as indicated by an increased net (i.e. production minus loss) photochemical ozone production (Fig. 3c). Due to rapid tropospheric mixing timescales, the resulting NH ozone increases extend to the surface, even despite negative (particularly between 0 to 50° N) NH surface net production changes under SAI (Fig. 3i). This behavior is consistent with previous work demonstrating that reductions in temperature and humidity can suppress photochemical ozone formation in NO_x-rich environments . The zonal-mean percent changes in NO_x and OH between ARISE-SAI-1.5 and SSP2-4.5 are provided in the Supplement (Fig. S14).

To further test this interpretation, we repeated our analyses in simulations with simulated SAI injections but no changes in tropospheric anthropogenic emissions (i.e. in a preindustrial climate) and observed qualitatively similar ozone responses (not shown), reinforcing our finding that these changes arise from stratospheric chemistry, transport, and in-situ oxidant perturbations, consistent with previous findings on SAI-driven ozone redistribution e.g.,.

This section presents the estimated mortality impacts of SAI under the ARISE-SAI-1.5 protocol, relative to SSP2-4.5. We first examine changes in PM_2.5-related mortality resulting from SAI, followed by an assessment of ozone-related mortality. Together with showing ensemble-averaged results, we also highlight in the following maps the large inter-ensemble and inter-ensemble spread when calculating mortality based on yearly model output. Local air quality is strongly dependent on meteorological conditions such as precipitation rates, heatwaves and atmospheric inversions. Global warming itself has been postulated to strengthen many of these conditions as well . Therefore, it is important to interpret our estimates within this broader context.

Figures  and a show the annual global deaths resulting from changes in PM_2.5 concentration and the average PM_2.5-related deaths (per 100 000 people) by country, respectively. Globally, we estimate that SAI leads to a reduction of ∼ 151 000 premature deaths from PM_2.5 under ARISE-SAI-1.5 (2060–2069), relative to SSP2-4.5 (2030–2039), with ensemble member estimates ranging from -140 000 to -164 000. In comparison, SSP2-4.5 (2060–2069) results in a reduction of ∼ 165 000 premature deaths relative to 2030–2039 levels, with a range of -148 000 to -177 000. This yields a net increase of ∼ 14 000 premature deaths in ARISE-SAI-1.5 compared to SSP2-4.5 during 2060–2069, with an ensemble range of -7000 to +21 000. These estimates, along with the standard deviation shown in Fig. , illustrate the substantial variability in projected PM_2.5-related deaths.

The changes in PM_2.5-related mortality for each country in Fig. e are roughly consistent with the geographical changes in PM_2.5 shown in Fig. i. In particular, the PM_2.5-related changes in certain countries (e.g., Russia and several in Sub-Saharan Africa) primarily reflect regions where internal variability, rather than an SAI-driven signal, dominates, consistent with the broader spatial pattern of PM_2.5 that are not statistically significant (Fig. i). In Fig. a, we compute the ensemble-averaged global deaths resulting from SSP2-4.5 with added changes in individual PM_2.5 components between ARISE-SAI-1.5 and SSP2-4.5 to isolate the influence of each component on global mortality. Among the components, incorporating changes in the dust PM_2.5 produce the largest deviation from the unmodified SSP2-4.5 baseline. The scenario with dust-only modifications results in fewer global deaths than the SSP2-4.5 baseline, which is likely due to the nonlinearity in the ozone-attributable risk function. However, when changes in all PM_2.5 components are combined, the resulting mortality aligns with the increased PM_2.5-related mortality observed in ARISE-SAI-1.5. For other components such as salt, BC, POM, SOA and SO₄, the resulting mortality estimates largely overlap the unmodified SSP2-4.5 baseline. In particular, the changes in global deaths attributable to SO₄ are small relative to other components, implying that sulfate-driven PM_2.5 mortality changes are modest compared to the total. Therefore, we conclude that SAI's contribution to PM_2.5-related mortality is small compared to the overall changes projected due to future air quality policies (on the order of ∼ 1 %, versus ∼ 10 % from policy-driven improvements), with internal variability among ensemble members and changes from other PM_2.5-related species playing a dominant role in driving uncertainty in our mortality estimates.

For ozone-related mortality, Figs.  and b show the annual global total deaths resulting from changes in ozone concentration and the average ozone-related deaths by country, respectively. We estimate that SAI leads to a reduction of ∼ 102 000 premature deaths from ozone exposure under ARISE-SAI-1.5 (2060–2069), relative to SSP2-4.5 (2030–2039), with an ensemble range of -91 000 to -108 000. By comparison, SSP2-4.5 (2060–2069) results in an estimated reduction of ∼ 89 000 premature deaths from ozone exposure relative to SSP2-4.5 for 2030–2039, with a range of -77 000 to -97 000. The net difference between ARISE-SAI-1.5 and SSP2-4.5 during 2060–2069 is approximately −14 000, with a range of -7000 to -25 000.

The geographic distribution of ozone-related mortality changes (Fig. e) shows that the largest reductions occur primarily in the Southern Hemisphere, most notably over Southern Sub-Saharan Africa, Southeast Asia and South America. This spatial pattern aligns with the hemispheric asymmetry in the tropospheric ozone response observed in Fig. , where greater reductions in ozone concentrations occur in the SH and parts of Asia. In Fig. b, the evolution of global ozone-related deaths over time is consistent with the time series of global OSMDA8 (Fig. b). Overall, no clear long-term trend is evident in PM_2.5 and ozone-related mortality, as any underlying signal may be masked by the large ensemble variability in projected deaths (Fig. ). Geographically, both ozone-and PM_2.5-related mortality changes exhibit substantial spatial variability, driven by regional differences in how ozone and PM_2.5 concentrations respond to shifts in atmospheric chemistry, circulation, and precipitation patterns under SAI.

Figure  shows how global changes in mortality due to ozone and PM_2.5 evolve over time in our simulations. When aggregated globally, it is evident that the largest change in air-pollution related mortality is due to decreases in precursors and pollutants under the SSP2-4.5 scenario. Differences between the futures with and without SAI, and those between different amount of SAI cooling, are much smaller on a per-year basis, and in most cases within the range of variability for the ensemble estimates. This demonstrates that the direct impact of deposited sulfate is limited, and climatic factors minimally impact PM_2.5 changes under SAI. In contrast, global ozone-related mortality is slightly lower in ARISE-SAI-1.0 than in ARISE-SAI-1.5, likely due to larger SAI-induced SH extra-tropical lower stratospheric ozone loss and the resulting reduction in ozone STE.

Figure  provides an alternative way of examining this relationship by plotting ensemble means against injection rates in the SAI scenarios. For both simulations, PM_2.5-related mortality shows no clear linear scaling with increasing injection rates, as substantial ensemble variability and factors other than SAI affecting the evolution of mortality rates with time dominate the relationship. Ozone-attributable mortality remains consistently negative across the entire injection range, indicating a reduction in ozone-related deaths under both ARISE-SAI-1.5 and ARISE-SAI-1.0.

For PM_2.5-related mortality in particular, our component attribution analysis suggests that the primary driver of changes is not sulfate itself, but rather arises from changes in dust and secondary SOA concentrations (Figs.  and a). Regional reductions in PM_2.5, particularly over Central America and central Sub-Saharan Africa, align with areas of increased precipitation, highlighting the role of wet deposition and circulation-driven suppression of natural aerosol sources (Fig. ). However, the widespread lack of statistically significant precipitation or PM_2.5 changes across ensemble members suggests that internal variability and regional circulation shifts, rather than sulfate burden alone, govern the spatial and temporal patterns of PM_2.5-related health outcomes under SAI.

For ozone, the mortality reductions appear more discernible. SO₂ is primarily injected in the SH, leading to decreased SH extra-tropical lower stratospheric ozone concentrations and the resulting reduction in SH surface ozone from reduced STE overwhelming any in-situ changes in tropospheric ozone chemistry there. In the NH, on the other hand, surface ozone increases due to the suppressed photochemical destruction under drier and colder troposphere (Fig. ). These changes reflect the role of not only hemispheric asymmetries in sulfate burden alone but also of STE and chemical processing arising from circulation changes and altered chemical regimes in shaping global ozone responses and associated health outcomes under SAI.

Taken together, these findings emphasize that air pollution-related health impacts under SAI are not governed mainly by the magnitude of SO₂ injected, but rather by the complex suite of dynamical, chemical, and aerosol responses in the Earth system, many of which are nonlinear and strongly influenced by internal variability. While our two large ensemble SAI simulations show no evidence for linear scaling with respect to injection rate, we acknowledge that longer simulations and additional scenarios would be needed to more fully characterize how air quality related mortality is dependent on the SAI scenario under consideration.

Globally, ARISE-SAI-1.5 reduces total pollution-attributable mortality relative to a future without intervention (SSP2-4.5) by 0.4 %, driven by a 0.9 % increase in PM_2.5 and 1.3 % reduction in ozone-related deaths (Fig. a–b). However, the direction and magnitude of health outcomes vary substantially across GBD super-regions. For instance, large % increases in PM_2.5-related mortality occur in regions such as Central, Western and Eastern Europe. In contrast, regions like the Caribbean and Central Latin America exhibit reductions in PM_2.5-attributable mortality, highlighting the heterogeneous and sometimes adverse regional impacts of SAI.

For ozone-related mortality, the ensemble spread is also large, both in magnitude and spatial extent, especially in regions such as the Western and Eastern Sub-Saharan Africa and the Caribbean. Furthermore, while national base mortality (cardiovascular, respiratory, and NCD + LRI baseline mortality rate) declines from 2030–2039 to 2060–2069 across all regions, the magnitude of these changes is relatively small compared to the much larger shifts seen in air quality-related mortality.

In many regions, the large ensemble spread reflects uncertainties not only in the magnitude but also in the sign of the projected impact on air quality related mortality. This spread arises from internal climate variability, which influences key drivers of air quality – such as atmospheric circulation, precipitation patterns, and chemical processing – and leads to diverging pollutant concentrations across ensemble members, even under identical forcing scenarios. These findings also highlight the spatial heterogeneity in health responses to SAI. While global or hemispheric trends may point to a net decline in ozone-related mortality and an increase in PM_2.5-related mortality, such aggregates can mask substantial regional disparities. As a result, careful evaluation of region-specific trade-offs is critical when assessing the overall public health implications of SAI deployment.