Assessing Outcomes in Stratospheric Aerosol Injection Scenarios Shortly After Deployment

Stratospheric aerosol injection (SAI) is a proposed form of climate intervention that would release reflective particles into the stratosphere, thereby reducing solar insolation and cooling the planet. The climate response to SAI is not well understood, particularly on short‐term time horizons frequently used by decision‐makers and planning practitioners to assess climate information. We demonstrate two framings to explore the climate response in the decade after SAI deployment in modeling experiments with parallel SAI and no‐SAI simulations. The first framing, which we call a snapshot around deployment, displays change over time within the SAI scenarios and applies to the question “What happens before and after SAI is deployed in the model?” The second framing, the intervention impact, displays the difference between the SAI and no‐SAI simulations, corresponding to the question “What is the impact of a given intervention relative to climate change with no intervention?” We apply these framings to annual mean 2 m temperature, precipitation, and a precipitation extreme during the 10 yr after deployment in two large ensembles of Earth system model simulations that comprehensively represent both the SAI injection process and climate response, and connect these results to implications for other climate variables. We show that SAI deployment robustly reduces changes in many high‐impact climate variables even on these short timescales where the forced response is relatively small, but that details of the climate response depend on the model version, greenhouse gas emissions scenario, and other aspects of the experimental design.

Global and regional climate responses to SAI are not well understood on the short time horizons of 10 yr or fewer often used by decision-makers and planning practitioners to assess climate information (e.g., Bolson et al., 2013;DePolt, 2021;Pearman & Cravens, 2022). Previous SAI modeling experiments have provided useful insights into general implications of the intervention, such as the potential for SAI to reduce global mean temperature, the inability of SAI to counteract impacts linked directly to CO 2 concentration, and the risk of rapid climate change if SAI is stopped ("termination shock"; e.g., Bony et al., 2013;Jones et al., 2013;Kwiatkowski et al., 2015;Rasch et al., 2008;Tilmes et al., 2009;Trisos et al., 2018). Many of these experiments (e.g., the Geoengineering Model Intercomparison Project; Kravitz et al., 2011) have relied on models with limited representations of relevant Earth system processes including atmospheric chemistry, stratospheric dynamics, and aerosol microphysics (e.g., McCusker et al., 2015;Quaglia et al., 2023). Many of the SAI scenarios in these experiments are implemented in highly idealized ways, such as by prescribing the aerosol optical depth fields or reducing the model solar constant (Kravitz et al., 2011), which can produce a very distinct climate response from when SAI is more realistically represented with interactive aerosols Ferraro et al., 2015;Visioni et al., 2021). These limitations leave large gaps in scientific knowledge of the climate response to SAI scenarios on spatiotemporal scales relevant to policymakers, planning practitioners, and questions of how SAI would affect climate risk inequality (Buck et al., 2014;NASEM, 2021;Pearman & Cravens, 2022).
The Geoengineering Large Ensemble (GLENS, Tilmes et al., 2018a) and Assessing Responses and Impacts of Solar climate intervention on the Earth system with stratospheric aerosols (ARISE-SAI-1.5, Richter et al., 2022) projects are the first large ensembles of Earth system model simulations that comprehensively represent processes most important to realistically portray SAI and employ strategically placed SAI to meet specific intervention goals. GLENS and ARISE-SAI-1.5 each contain parallel ensemble simulations: one following a climate change trajectory with no SAI, and one where SAI is deployed. This design helps separate the forced response to SAI from the influence of climate change and internal variability.
The parallel simulations in GLENS and ARISE-SAI-1.5 provide extensive insight into the climate response to SAI. We propose two framings that utilize these parallel simulations to efficiently display multiple perspectives on the climate response to SAI. The first, which we call a snapshot around deployment, displays the change over time within the SAI simulations. This corresponds to the question, "What happens before and after SAI is deployed in the model?" The second, the intervention impact, describes the difference between the SAI and no-SAI simulations. This addresses the question, "What is the impact of a given intervention relative to climate change with no intervention?" We apply these two framings to explore the climate response to SAI in the decade after SAI is deployed in GLENS and ARISE-SAI-1.5. Using our two framings together allows us to explore both the tangible climate response through the snapshot around deployment and place these changes in context to climate change with no SAI with the intervention impact. We apply our framings to annual mean 2 m temperature, annual mean precipitation, and the simple intensity index (a measure of extreme precipitation), and connect these results to global and regional impacts on an assortment of other climate variables selected for their familiarity in Earth science and importance for human and ecological impacts. We discuss commonalities and differences between the climate responses in the GLENS and ARISE-SAI-1.5 scenarios of SAI deployment.
Our work addresses the goals of NASEM (2021) to "advance knowledge relevant to decision making" and "develop policy-relevant knowledge." Consistent with this NASEM report and the broader social science literature, we explicitly distinguish our goals from research on the practical deployment of SAI about which critical ethical and governance concerns exist (Burns et al., 2016;NASEM, 2021). We intend our study simply to support the informed discussion of the potential risks and benefits of SAI.

Description of Simulations
We use model output from the GLENS (Tilmes et al., 2018a) and ARISE-SAI-1.5 (hereafter ARISE; Richter et al., 2022) experiments to explore the climate response to SAI. These are large ensembles of SAI modeling experiments performed in fully interactive Earth system models, with strategically placed SAI to meet specific temperature targets. We summarize key aspects of the design of GLENS and ARISE, and refer readers to Tilmes et al. (2018a) and Richter et al. (2022) for more comprehensive descriptions of the experiments. GLENS and ARISE both employ the Community Earth System Model (CESM) with the Whole Atmosphere Community Climate Model (WACCM) as its atmospheric component, albeit different versions of the model that will be described shortly. WACCM includes 70 vertical layers (model top 140 km) to explicitly simulate the stratosphere and lower mesosphere, and uses a 1.25° longitude × 0.9° latitude horizontal resolution. The representation of processes thought to be most important for SAI and its climate response, including stratospheric dynamics, heterogeneous chemistry, and aerosol production, show good agreement with observations of the mean state and anomalous conditions under volcanic aerosol loading (Gettelman et al., 2019;Richter et al., 2017).
Each of the two experiments contains two parallel ensemble simulations: one following a future greenhouse gas forcing scenario with no SAI, and one where SAI is also deployed. A proportional-integral feedback-control algorithm (known as the "controller") annually adjusts the amount of sulfur dioxide continuously released at four latitudes (30° and 15°N/S, all at 180°E) intended to maintain global mean temperature, the pole-to-pole temperature gradient, and the pole-to-equator temperature gradient at some specified target value MacMartin et al., 2014MacMartin et al., , 2017. These targets aim to ensure planetary circulations under SAI change less than if only global mean temperature were to be targeted (Cheng et al., 2022;Tilmes et al., 2018a;Visioni et al., 2021).
GLENS and ARISE each portray a unique intervention scenario where SAI is deployed to maintain specific goals against a particular greenhouse gas forcing. GLENS uses a no-mitigation emissions trajectory (Representative Concentration Pathway [RCP] 8.5;van Vuuren et al., 2011) with SAI deployed to maintain a global mean temperature target near 2020 values (Tilmes et al., 2018a). This yields a large signal-to-noise ratio useful to isolate the forced response to the SAI intervention over time. ARISE is run with a moderate-mitigation scenario (Shared Socioeconomic Pathway [SSP] 2-4.5; Riahi et al., 2017) and a temperature target of approximately 1.5°C above the IPCC AR6 pre-industrial definition (IPCC, 2021;Richter et al., 2022). ARISE illustrates one plausible future where the use of SAI complements current mitigation strategies to achieve Paris Agreement goals .
There are several differences between the experimental design of GLENS and ARISE. We summarize these in Table 1, and provide details on their implications here.
1. The two experiments use different model versions: GLENS uses CESM1(WACCM5), while ARISE uses CESM2(WACCM6). Thus, GLENS and ARISE exhibit different spatial patterns of the forced response due to model dependencies, particularly the depiction of subtropical and Southern Ocean low clouds (Fasullo & Richter, 2023;Gettelman et al., 2019). CESM1(WACCM5) is described by Hurrell et al. (2013) and , and CESM2(WACCM6) by Danabasoglu et al. (2020) and Gettelman et al. (2019). 2. The two experiments have different forcing scenarios: GLENS uses RCP8.5 while ARISE uses SSP2-4.5, which yields distinct spatial patterns of the forced response. RCP8.5 and SSP2-4.5 differ in many ways that affect these spatial patterns, including the depiction of land use and aerosol emissions, but the primary influence is the different CO 2 concentration which operates through a direct effect on clouds and precipitation (e.g., Fasullo & Richter, 2023;Rugenstein et al., 2016;Sherwood et al., 2015). These scenario dependencies are largest in the mid-latitudes and subtropics (Fasullo & Richter, 2023).  , 2022). Due to differences in model physics and definitions of the preindustrial baseline, it is more precise to discuss temperature targets in terms of the time averages implemented in each experiment. For additional context, we note these global mean temperature targets correspond to approximately 1.1°C above preindustrial in GLENS and 1.5°C in ARISE when the IPCC AR6 definition is used (IPCC, 2021). 4. The method of generating ensemble spread is different between GLENS and ARISE. The ocean state in every member of GLENS is branched off of the first member of the CESM Large Ensemble (Kay et al., 2015), in which the Atlantic Meridional Overturning Circulation (AMOC) is strengthening. Hence, each realization of GLENS begins with a strengthening state of the AMOC and ensemble spread is generated only through differing atmospheric initial conditions. On the decadal time horizons we emphasize in this study, the memory of these ocean initial conditions has not fully dispersed (Fasullo et al., 2018;Tilmes et al., 2018a). This increases oceanic heat transport into the North Atlantic and influences European climate (Fasullo et al., 2018). Longer-term trends in AMOC strength may be caused by changes in precipitation that impact salinity and temperature gradients in the ocean (Fasullo & Richter, 2023). In contrast, the ocean states from five separate SSP2-4.5 simulations dispersed over multiple decades are used in addition to differing atmospheric initial conditions to generate ensemble spread in ARISE . Thus, ARISE samples ocean internal variability more widely than GLENS. 5. Aerosol is injected at lower altitudes in GLENS (≈ 25 km) than in ARISE (≈ 21 km). Injection height affects stratospheric chemistry, but has few other effects on the climate as long as the altitude is above the tropopause (Tilmes et al., 2018b;Y. Zhang et al., 2022). Injection heights near 20 km are consistent with near-future aerospace technology (D'Oliveira et al., 2016;Moriyama et al., 2017;W. Smith et al., 2022).
Because the controller maintains meridional temperature gradient targets under different spatial patterns of the forced response between GLENS and ARISE, it injects aerosol in disparate latitudinal amounts throughout each experiment. This leads to distinct global distributions of aerosol optical depth ( Figure 1 in Fasullo & Richter, 2023) with corresponding differences in the regional climate response. These model and scenario dependencies imply that the results of GLENS and ARISE are specific to these scenarios; consistency of a result between the simulations does not imply increased confidence that the result is true of any general SAI deployment. Thus, GLENS and ARISE are best compared to illuminate the differences in the climate response produced by two SAI scenarios simulated in physically comprehensive models.

Analysis Metrics
We use the 5 yr prior to SAI deployment ( in GLENS, 2030( -2034 in ARISE) as pre-intervention reference periods, and the 5 yr period beginning 5 yr after deployment (2025( -2029( in GLENS, 2040( -2044 in ARISE) as a post-intervention reference period while remaining close to the deployment year. The ensemble sizes of the GLENS and ARISE experiments (Table 1) increase the number of years available for analysis, allowing us to average over many realizations of relatively short spans of time (Deser et al., 2012;Maher et al., 2021;Tebaldi et al., 2021). We develop two framings to investigate the climate response to SAI. Our first framing, which we call a snapshot around deployment, depicts change over time within the SAI experiments: the difference between 2025-2029in GLENS, and 2040-2044and 2030-2034 in ARISE. This can be phrased as answering the question: "What happens before and after SAI is deployed in the model?" Our second framing, the intervention impact, is the SAI and no-SAI difference for the 2025-2029 period in GLENS and the 2040-2044 period in ARISE. This can be expressed as answering the question: "What is the impact of a given intervention relative to climate change with no intervention?" This was inspired by the "world avoided" perspective used to study the Montreal Protocol (e.g., Morgenstern et al., 2008).
We structure our framings to focus on the short-term climate responses occurring in the first 10 yr after SAI deployment. This near-term timeframe has been analyzed with respect to the atmospheric dynamical response to SAI (e.g., Richter et al., 2018;Tilmes et al., 2017), but has seen little exploration with respect to climate impacts. Policymakers and planning practitioners often assess climate information on time horizons of 10 yr or fewer (e.g., Bolson et al., 2013;DePolt, 2021;Keys et al., 2022;Pearman & Cravens, 2022). Thus, we portray our results consistently with how information could hypothetically be used for decisions about SAI deployment, governance, and evaluation. The signal-to-noise ratio of the forced response to SAI is smaller on this time horizon than when trends are calculated over a longer span of time. We use timeseries ( Figure 2) to complement ensemble mean global maps of our two framings (Figures 3-8). These allow us to display the longer-term evolution of a variable and emphasize the contribution of internal variability for a specific region. The longer-term evolution may be different from the short-term response, and helps place our work in context to previous results in the literature. We show timeseries for 2010-2069 to span the period where the output from both GLENS and ARISE are available. We use the CESM2 (WACCM6) Historical simulations (Danabasoglu et al., 2020) to supplement the period 2010-2014 before the ARISE no-SAI simulations begin.
As an Earth system model, CESM provides a breadth of model output including variables that represent the atmosphere, ocean, land surface, and ecology. This allows many aspects of the Earth system response to SAI to be assessed holistically. We examined a wide variety of variables in developing this paper. Here, we present a subset that is familiar in climate science, has links to human impacts, and whose representation in CESM has been evaluated against observations (Danabasoglu et al., 2020;Fasullo, 2020;Hurrell et al., 2013). We describe our selected variables below. Surface temperature: We calculate the annual mean 2 m temperature from monthly output. For temperature and all other variables, we define regions following the IPCC Working Group 1 Fifth Assessment Report Annex (van Oldenborgh et al., 2013), except when specified otherwise. We illustrate the regions we discuss in this work in Figure S1 in Supporting Information S1. In addition to the regions highlighted in Figure 2, we provide timeseries of 2 m annual mean temperature for all IPCC-defined regions in the archive linked in Text S2 in Supporting Information S1. Tropical nights: We use tropical nights from the World Climate Research Program's Expert Team on Climate Change Detection and Indices (ETCCDI) set of extreme indices as an example of a temperature extreme. Tropical nights are the annual count of days where the minimum temperature exceeds 20°C (68°F) (X. Zhang et al., 2011). High nighttime temperatures increase mortality, particularly in urban areas without widespread air conditioning (Buechley et al., 1972;Laaidi et al., 2012;Rathi et al., 2021;Sillmann & Roeckner, 2008). We calculate tropical nights from daily minimum temperature using Pyclimdex (Groenke, 2022). Tye et al. (2022) comprehensively explore ETCCDI extremes in GLENS; no such assessment has been completed for ARISE. Sea surface temperature (SST): We calculate the annual mean SST from monthly output at the surface level of the ocean component in CESM. Marine heatwaves: We identify marine heatwaves as events where daily mean SST exceeds the daily local 90th percentile (computed over 2010-2020) for longer than 5 days (Hobday et al., 2016;Oliver, 2022). This definition is standard in public communication and the scientific literature (e.g., Benthuysen et al., 2018;Holbrook et al., 2020;MHIWG, 2022). Marine heatwaves occur at many locations around the world (K. E. Smith et al., 2021), and we select a point in the Leeuwin Current (30.63°S, 112.5°E) where they have been frequently observed to harm local ecology (Chandrapavan et al., 2019;Holbrook et al., 2020). We apply a left-aligned 5 yr rolling sum of days to smooth interannual variability in Figure 2l. Sea ice extent: We show sea ice extent in its minimum month for both hemispheres-September for the Arctic and February for the Antarctic (Parkinson, 2019;Stroeve et al., 2012). Sea ice extent is the sum of grid cell areas with an ice fraction greater than 0.15 in the atmospheric component of CESM (NSIDC, 2020). Precipitation: We derive annual mean precipitation from monthly total precipitation. To describe South Asian Monsoon rainfall, we use the conventional dynamical definition of June through September mean precipitation between 10°N-40°N and 80°E-100°E (Geen et al., 2020). We provide timeseries of annual mean precipitation for all IPCC-defined regions at the archive provided in Text S2 in Supporting Information S1. Simple intensity index: We use the ETCCDI simple intensity index to illustrate changes in a precipitation extreme. The simple intensity index measures the precipitation amount divided by the number of days with precipitation (X. Zhang et al., 2011). This is a standard metric to analyze trends in precipitation intensity (e.g., Alexander et al., 2006;Ayugi et al., 2021). Following Ayugi et al. (2021), we define the East African region as spanning 12°S-5°N and 28°E-42°E to capture relevant regional climate features. We calculate the simple intensity index using Pyclimdex (Groenke, 2022).

Robustness
Regional trends in the model output are due to combinations of the forced response to the SAI intervention, the direct effect of CO 2 concentration, and internal variability (Fasullo & Richter, 2023). We define a metric called robustness (ρ) to quantify where the signal from the forced response to SAI is large relative to noise from internal variability and the response to climate change. Because the parallel ensemble simulations in GLENS and ARISE are identical aside from the presence of the SAI intervention, consistent differences between the SAI and no-SAI members are likely due to the response to the SAI intervention. Robustness quantifies this consistency as the count ρ of each SAI realization whose temporal mean over a given time period falls outside (exceeds or subceeds) a user-defined quantity of no-SAI realizations (denoted as B; B = 11 for GLENS, B = 6 for ARISE given differing ensemble sizes [ Table 1]).
Sufficiently large values of robustness indicate a consistent ("robust") forced response to the SAI intervention. We refer to results as "robust" if they fall outside the 90% confidence bounds of the robustness distribution expected by chance (i.e., robustness distribution computed via 10,000 pairs of vectors randomly sampled from a uniform distribution). "Robust" results have ρ ≥ 15 members in GLENS and ρ ≥ 7 members in ARISE. These thresholds are statistically significant at the p < 0.1 level (p = 0.02 for GLENS, p = 0.05 for ARISE) under a binomial test-however, we emphasize that the exact choice of threshold is subjective. We apply image muting to de-emphasize (gray out) points that are not robust (ρ < 15 members for GLENS, ρ < 7 members for ARISE) without removing data from our maps (Tomkins et al., 2022).
To help build intuition, we provide an example of results considered robust and not robust for a case where the SAI ensemble members subceed the no-SAI members ( Figure S6 in Supporting Information S1). Robustness is a non-parametric test that leverages the parallel large ensemble design of GLENS and ARISE to rigorously convey the consistency of a result, without requiring assumptions about the statistical distribution of the variable of interest.
We formalize robustness mathematically in Equation 1. For each longitude θ and latitude ϕ (grid point), we compute the robustness ρ θ,ϕ which is the maximum of the number of SAI realizations that exceed or subceed B number of no-SAI realizations. is the time mean over a given period for a variable for each SAI realization r z , ̃{ } denotes time means of a variable for B number of no-SAI realizations r {B} , and Z is the size of the SAI ensemble. The robustness calculation is repeated for every latitude and longitude to generate a map of robustness for a given variable ( Figure S2 in Supporting Information S1).
The unit of robustness is "number of ensemble members" inherited from the cardinality operator n(). Robustness is non-negative and bounded by the size of the SAI ensemble: ρ θ,ϕ = Z is the upper bound (21 in GLENS, 10 in ARISE). We detail the algorithm to calculate robustness and the binomial test for statistical significance in Text S1 in Supporting Information S1.

Results
GLENS and ARISE both maintain global mean temperature close to their respective target values, while the no-SAI RCP8.5 and no-SAI SSP2-4.5 scenarios continue warming globally throughout the period (Figure 2a).
Thus, global mean temperature shows a clear forced response to the SAI intervention. For each timeseries shown in Figure 2, the envelope around the ensemble mean illustrates a range of internal climate variability by spanning the maximum to minimum value across the ensemble at each year. The ensemble sizes for each scenario differ and are given in Table 1. While forced trends are visible in the ensemble mean for many of the timeseries (Figure 2), internal climate variability is substantial especially for regional scales and noisier variables such as precipitation (e.g., Figure 2h). The ensemble spreads of the SAI and no-SAI scenarios overlap for all quantities in the time periods shortly after deployment when the forced response is small. Thus, internal variability can mask the forced response to the SAI intervention for any individual realization. Our results suggest climate variability may lead to the "perceived failure" of SAI on short time horizons across many variables (Figure 2) regardless of the true forced response from SAI, as previously shown for temperature alone (Keys et al., 2022).
We use global maps corresponding to the snapshot around deployment and intervention impact framings (Figures 3-8, see Methods for details) to explore the ensemble mean response of temperature, precipitation, and the simple intensity index within the decade after SAI deployment. We refer to timeseries in Figure 2 to connect results from our framings to the evolution of a variable over a longer period of time and to display the spread due to internal variability.

What Happens Before and After SAI Is Deployed in the Model?
We begin our discussion with 2 m temperature (Figure 3), as it is the variable directly targeted by the SAI intervention. Some global warming is visible in the GLENS snapshot (Figure 3a) due to the rapid warming rate in the underlying RCP8.5 emissions trajectory. The GLENS experimental design maintains global mean temperature at 2020 levels, which leaves some warming relative to the 2015-2019 mean which defines our pre-intervention snapshot baseline. The SSP2-4.5 forcing scenario used in ARISE yields a much more moderate rate of warming as compared to RCP8.5 and a smaller relative change between the 2030-2034 baseline and the deployment of SAI in 2035. Hence, the snapshot around deployment for ARISE does not display substantial planetary-scale warming.
The subpolar North Atlantic Ocean stands out as the region experiencing the largest temperature trends (Figure 3). The sign of the trend is opposite in each experiment: warming in GLENS (Figure 3a), but cooling in ARISE ( Figure 3b). This difference is driven by the opposite-signed AMOC evolution in GLENS and ARISE. The strengthening AMOC throughout the simulation period in GLENS increases oceanic heat transport into the North Atlantic (Fasullo et al., 2018); in contrast, the AMOC weakens in ARISE, although it remains stronger than in the no-SAI SSP2-4.5 scenario (Fasullo & Richter, 2023;Richter et al., 2022). These trends in the AMOC are likely due to memory of the ocean initial conditions on the short timescales shown in the snapshot around deployment (Fasullo et al., 2018;Tilmes et al., 2018a). On a longer time horizon, the direct effect of CO 2 concentration on precipitation may drive a forced response in the AMOC by altering oceanic salinity and temperature gradients (Fasullo & Richter, 2023); however, this long-term effect would not be visible in our snapshot around deployment.
In general, regional changes in annual mean 2m temperature after SAI deployment are much smaller in GLENS and ARISE (Figure 3) than in no-SAI RCP8.5 and no-SAI SSP2-4.5 ( Figure 1). All regions (save Northern Europe in GLENS, discussed shortly) defined by the IPCC WG1-AR5 Atlas (van Oldenborgh et al., 2013) have a similar temperature response over time to the global mean. We provide 2 m annual mean temperature timeseries for each IPCC region at the archive linked in Text S2 in Supporting Information S1 to illustrate the universality of this response. We highlight the Amazon region ( Figure 2e) as an example of the typical evolution of annual mean temperatures on a regional scale. The response in the Amazon is very similar to the global response: GLENS and ARISE are maintained near their pre-deployment values, while the no-SAI scenarios continue to warm through the period. Temperature trends are robust for all land area in GLENS outside Antarctica, and almost all land area in ARISE. We reiterate that "robust" trends fall outside the 90% confidence bounds of the distribution expected by chance, and a full description of the robustness metric can be found in Section 2.3.
Northern Europe (Figure 2i) experiences moderate warming throughout the period in GLENS. While this warming has been previously shown to occur on late-century timescales (Banerjee et al., 2021;Fasullo et al., 2018;Tilmes et al., 2018a), we show this warming is already robust within the decade after deployment (Figure 3a). One cause of this warming may be a forced positive trend in the North Atlantic Oscillation driven by stratospheric heating from the absorption of radiation by the sulfate aerosols injected by the SAI intervention (Banerjee et al., 2021). The strengthening AMOC could also contribute to this regional warming by importing heat from lower latitudes (Fasullo et al., 2018). In contrast, we find Northern European warming does not occur in ARISE on any timescale (Figures 2i and 3b). The differing responses in Northern Europe emphasize that climate responses in an individual scenario may be particular to that strategy, and cannot be assumed to be general features of all SAI interventions.
We now turn to the snapshot around deployment for precipitation (Figure 4). Due to its large internal variability (Deser et al., 2012), fewer regional trends are robust for precipitation rather than temperature on our short timescale of 10 yr after SAI deployment. Precipitation robustly decreases over portions of the tropical Pacific in GLENS (Figure 4a). On longer timescales, similar trends emerge over much of the basin and are responsible for a decrease in globally averaged precipitation (Figure 2b). Since these changes primarily affect precipitation over the ocean, land-only precipitation trends in GLENS are small even later into the century ( Figure S3 in Supporting Information S1). The decrease in tropical oceanic precipitation may be related to the direct effect of CO 2 concentration in the RCP8.5 emissions pathway or the circulation response to stratospheric heating, but the precise underlying dynamics are not well understood (Bony et al., 2013;Simpson et al., 2019). Trends in global precipitation in ARISE are difficult to identify (Figure 2b), which indicates a more moderate injection strategy may minimize impacts on the global hydrologic cycle.
The location of equatorial precipitation associated with the intertropical convergence zone (ITCZ) shifts southward in GLENS and ARISE (Figure 4), but we show these changes are not robust on the short timescale of 10 yr after deployment. The controller in GLENS and ARISE minimizes the impacts on the ITCZ by maintaining the pole-to-pole and pole-to-equator temperature gradients which are primarily responsible for ITCZ location (Cheng et al., 2022;Kang et al., 2018;Undorf et al., 2018). Other processes, such as differences in the relative aerosol burden between the Northern and Southern Hemispheres, changes in heat transport by the AMOC, or stratospheric heating can still influence ITCZ location (Cheng et al., 2022;Ciemer et al., 2021;Haywood et al., 2013;Iles & Hegerl, 2014;Moreno-Chamarro et al., 2019). Thus, on longer timescales, the controller reduces but cannot fully eliminate shifts in the ITCZ location (Cheng et al., 2022). Targeted modeling experiments and observations after volcanic eruptions indicate that much larger ITCZ migrations are possible under SAI strategies that do not consider planetary temperature gradients (Cheng et al., 2022;Haywood et al., 2013). ITCZ location can be targeted more successfully in simulations that use different controller targets specifically tuned to this feature (Lee et al., 2020).
Early modeling results indicated certain SAI strategies could cause large decreases in South Asian Monsoon precipitation . Any changes to the monsoon directly affect water availability and agricultural productivity in a densely populated region, with further impacts on global food supply (Gadgil & Rupa Kumar, 2006;Kulkarni et al., 2016). We find that South Asian Monsoon precipitation robustly decreases in GLENS (Figure 4a) even on the short timescales of the snapshot around deployment, although the magnitude of the change is smaller than the increase throughout the period in no-SAI RCP8.5 (Figure 2j). Late in the century in GLENS, monsoon failures double in frequency due to circulation changes induced by stratospheric heating from the extremely large aerosol burden (Simpson et al., 2019). In contrast, South Asian Monsoon precipitation remains largely unchanged in both SSP2-4.5 and ARISE across short ( Figure 4b) and long timescales (Figure 2j). Thus, we conclude that impacts on monsoon precipitation are dependent on the SAI strategy rather than a general feature of this type of intervention. Visioni et al. (2020) previously showed monsoon impacts varied in modeling experiments where SAI was limited to certain seasons. The difference in base state between the CESM2 SSP2-4.5 and CESM1 RCP8.5 simulations indicate that model dependencies and the greenhouse forcing scenario may be especially important to the monsoonal climate response.
On our short time horizon of the decade after deployment, global changes in the simple intensity index are very noisy without clear forced responses ( Figure 5). This illustrates how internal climate variability can remain the dominant driver of certain high-impact climate variables after SAI deployment. Precipitation extremes exhibit the largest internal variability of any quantity examined here. The 10-member ensemble of ARISE, in particular, is not sufficient to isolate the forced response to SAI on precipitation extremes for regional spatial scales and short timescales after deployment. An ensemble size of 40 members or more may be necessary to reliably isolate forced trends (Kirchmeier-Young & Zhang, 2020).

What Is the Impact of a Given Intervention Relative to Climate Change With No Intervention?
GLENS and ARISE both avert warming around the globe ( Figure 6); that is, they robustly remain cooler than their respective no-SAI scenarios. This impact is evident even in the decade immediately following deployment and is nearly single-signed worldwide ( Figure 6). The Arctic experiences the greatest averted warming (Figure 6), because the controller greatly reduces Arctic amplification by maintaining the pole-to-equator temperature gradient in addition to global mean temperature. In either GLENS or ARISE, no regions experience robust warming relative to climate change in the ensemble mean. Regions that warm relative to a pre-intervention baseline, namely Northern Europe in GLENS, still experience averted warming relative to climate change (Figure 6a). This illustrates the value of using our two framings together: the snapshot around deployment shows the tangible climate response, while the intervention impact places these changes in context to climate change with no SAI.
In highly localized areas where trends are weak in both the SAI and no-SAI scenarios, the intervention impact is small and noise from internal variability can make it appear as if the intervention has exacerbated warming.
This effect can be seen in very small portions of the Southern Ocean in GLENS ( Figure 6a) and the northeastern Pacific Ocean in ARISE (Figure 6b) that display a positively signed intervention impact. Note these regional features are not robust and thus are grayed out. Internal variability may mask the impact of SAI for any individual realization, which complicates how the effectiveness of an intervention could be perceived on short timescales after deployment (Keys et al., 2022).
We connect the averted warming in global mean temperature to implications for the evolution over time of other Earth system variables. Global sea surface temperature (Figure 2c) responds very similarly to global 2 m temperature ( Figure 2a). Sea ice loss is halted in the Arctic and Antarctic (Figures 2g and 2k) in both GLENS and ARISE. The impact is most dramatic in GLENS; in no-SAI RCP8.5, the Arctic experiences ice-free minima by mid-century while SAI keeps sea ice near present-day values. The SAI scenarios have the potential to slow or avert feedbacks involving sea and land ice. Arctic sea ice thickness is maintained alongside sea ice extent ( Figure  S4 in Supporting Information S1), indicating ice-insulation feedbacks that can cause rapid sea ice loss (e.g., Burt et al., 2016) could be averted. In Antarctica, preventing sea ice loss prevents the exposure of coastal ice shelves to ocean waves which may make land ice less likely to collapse (Massom et al., 2018). Exploring the impacts from SAI on the cryosphere in more depth is a clear avenue for future research.
Mid-latitude tropical nights increase drastically in the no-SAI scenarios ( Figure 2d) and are associated with the planetary-scale expansion of the tropics (Rajaud & de Noblet-Ducoudré, 2017). SAI interventions in GLENS and ARISE both limit this process, maintaining tropical nights near pre-intervention values. Averting increases in tropical nights could mitigate impacts from heat waves, as high overnight temperatures worsen mortality during these events (e.g., Buechley et al., 1972;Laaidi et al., 2012). While heat extremes are mitigated under GLENS or ARISE, extreme cold may be worsened relative to no-SAI climate change scenarios (Tye et al., 2022). More detailed risk analysis is necessary to quantify tradeoffs in exposure to extreme cold and heat.
Temperature extremes in the ocean are also impacted by the averted warming under SAI. In GLENS and ARISE, increases in marine heatwave frequency are prevented for a point off the coast of Western Australia (Figure 2l). In the no-SAI scenarios, this location reaches a near-permanent marine heatwave state by mid-century. Marine ecosystems are increasingly affected by compound hazards: combinations of stressors including direct anthropogenic impacts, ocean acidification, and temperature extremes (e.g., Chandrapavan et al., 2019;Gruber et al., 2021). While SAI only mitigates temperature extremes, lessening one component of compound hazards may allow ecosystems to stay within their capacity for resilience (Bernhardt & Leslie, 2013).
Due to the large internal variability of precipitation (Deser et al., 2012), regional impacts are not robust over much of the globe in the decade after deployment (Figure 7). Robust regional precipitation responses are particularly difficult to identify in ARISE (Figure 7b), as both the ensemble size and SAI forcing are smaller than in GLENS. Still, certain impacts of the SAI intervention on precipitation oppose notable no-SAI climate change trends. For example, precipitation in the Southern Hemisphere subtropics decreases in response to climate change when meridional SST gradients in the South Pacific Ocean are impacted by the rate of change in global mean temperature (Sniderman et al., 2019). As GLENS and ARISE both maintain global mean temperatures, they avert this transient climate response (Figure 7). GLENS and ARISE both robustly oppose increases in precipitation in portions of the Arctic that occur in the no-SAI scenarios ( Figure 7) associated with rapid warming from Arctic amplification (Figure 2f). Warmer air temperatures support exponentially larger saturation vapor pressures, a trend that is reinforced by increased evaporation from the open ocean due to sea ice loss (Bogerd et al., 2020). Alaska Native communities are highly vulnerable to climate change impacts, particularly those from increased precipitation (Melvin et al., 2017;Shearer, 2012). The potential to avert these impacts indicates SAI could mitigate regional climate risk inequality in certain cases, although far more analysis is needed to draw any broader conclusions. Increased temperature and precipitation together yield more vegetation growth in the Arctic (Dial et al., 2022;Elmendorf et al., 2012). Arctic vegetation can exacerbate warming by decreasing surface albedo and increasing local water vapor mixing ratios, which accelerates ice loss and encourages further plant growth (Swann et al., 2010). This positive feedback is considered a possible tipping point in the Earth system (Crump et al., 2021;Heijmans et al., 2022). SAI may prevent this process by preventing increases in temperature and precipitation, although further research would be necessary to examine this in detail.
As discussed previously, changes in the simple intensity index are largely not robust (Figure 8) due to the influence of internal variability. Still, certain robust regional impacts are more apparent relative to climate change in the intervention impact than relative to the pre-intervention baseline in the snapshot around deployment. Globally, GLENS and ARISE both reduce the simple intensity index over land relative to the no-SAI scenarios (Figure 8, Figure S5 in Supporting Information S1). We highlight the East African region, specifically, due to its high exposure to extreme precipitation events (Adhikari et al., 2015;Nicholson, 2017;Wainwright et al., 2021). The simple intensity index decreases in East Africa relative to no-SAI climate change, although this trend is robust for only a portion of the area on the timescale of 10 yr after SAI deployment. Over the course of the simulation period, regional simple intensity index decreases in GLENS and is maintained in ARISE (Figure 2h) in contrast to increasing trends in the no-SAI scenarios. Elsewhere, regional trends in the simple intensity index are generally  not robust. We provide timeseries of the simple intensity index for each IPCC-defined region in the archive linked in Text S2 in Supporting Information S1.

Conclusions
We present two ways to frame output from Earth system modeling experiments with parallel intervention and no-intervention ensemble simulations, which we call the snapshot around deployment and the intervention impact. Our framings directly address the research questions: "What happens before and after the intervention is deployed in the model?" and "What is the impact of a given intervention relative to climate change with no intervention?" We apply our framings to GLENS and ARISE, the first SAI modeling experiments performed by large ensembles of fully interactive Earth system models. We explore these questions in the decade after SAI deployment, a policy-relevant time horizon that has not been widely explored in the literature with respect to SAI impacts. We use our framings to efficiently describe many aspects of the climate response to SAI, including 2 m temperature, annual mean precipitation, and the simple intensity index (a measure of precipitation extremes). We observe certain commonalities between the SAI scenarios relative to their respective no-SAI scenarios: annual mean temperature is maintained at target values after deployment both globally and for nearly all IPCC-defined regions, sea ice loss is halted at both poles, and increases in certain marine and terrestrial heat extremes are prevented. However, GLENS and ARISE each portray a distinct scenario of SAI deployment with its own design and climate response. Results that are consistent between the simulations still cannot be taken to be true of any general SAI deployment.
Our study is the first to synthesize results from GLENS and ARISE together. We focus on a short-term time horizon of the 10 yr after deployment, which is consistent with timescales frequently used by policymakers and planning practitioners to assess climate information (e.g., Bolson et al., 2013;DePolt, 2021;Keys et al., 2022;  Pearman & Cravens, 2022). This differentiates our work from existing literature on the climate response in GLENS or ARISE, which usually examines time horizons later in the century in order to obtain a larger forced signal from the SAI intervention (e.g., Camilloni et al., 2022;Pinto et al., 2020;Richter et al., 2022;Simpson et al., 2019;Tilmes et al., 2018a;Tye et al., 2022). We intend our data analysis to provide a point of entry for researchers or educators unfamiliar with SAI, and include an archive of timeseries depicting each of the variables used with our framings for all IPCC regions (linked in Text S2 in Supporting Information S1).
Our framings can be used with any modeling experiment that has parallel intervention and no-intervention simulations. In particular, we see an opportunity to apply these framings to planned ARISE-SAI experiments that explore a wider variety of temperature targets, deployment dates, and Earth system models . As our framings directly address concrete questions of the climate response to SAI, they could also motivate a more comprehensive regional risk analysis constructed in collaboration with planning practitioners and members of affected communities (e.g., Adelekan & Asiyanbi, 2016;DePolt, 2021).
We show that while large forced responses to SAI are visible in the ensemble mean within the decade after deployment in GLENS and ARISE, internal variability can mask impacts in individual realizations. The noise from internal variability has important implications for three key open problems highlighted by NASEM (2021): detection, monitoring, and social perception of any climate intervention. Machine learning methods have shown promise for rapid detection of the surface climate response to SAI despite the influence of internal variability . Improved understanding of the data most useful to detect SAI could help constrain potential observational platforms for long-term monitoring. Regardless of the true forced climate response, the noise from internal variability may influence the perceived success or failure of any climate intervention-or climate action more broadly (Diffenbaugh et al., 2023;Keys et al., 2022).
GLENS and ARISE provide high-fidelity depictions of two useful scientific knowledge-building scenarios (Talberg et al., 2018). However, these scenarios are geopolitically idealized: they depict SAI as an uninterrupted worldwide project ("global action" scenarios) with a controller limiting disruptions to global mean climate. Thus, the results from these specific scenarios do not generalize to any given SAI intervention. The differences between GLENS and ARISE demonstrate that even global action scenarios with many commonalities can produce distinct climate responses, due to factors such as model dependency, the ocean initial conditions, and the direct effects of the underlying greenhouse gas emissions forcing scenario. To explore a scenario of interest, it will be necessary to explicitly model that scenario; the results cannot be assumed to track those of GLENS or ARISE. Future modeling should widely explore the scenario design space, with possible examples of candidates including unilateral ("rogue actor") deployment and environmental peacebuilding (Buck, 2022;Fitzgerald, 2016).

Data Availability Statement
The processed model output used throughout this work, code for reproducibility, and additional timeseries described in Text S2 in Supporting Information S1 are archived at the Open Science Foundation (Hueholt, 2022, https://doi.org/10.17605/OSF.IO/5A2ZF). This repository additionally includes a datasheet describing the data adapted from best practices from software engineering (Gebru et al., 2021). The original GLENS model data set from which the data in this work was derived can be obtained from NCAR , https://doi. org/10.5065/D6JH3JXX). The original ARISE data set from which the data in this work was derived (all SAI members and 5 no-SAI members) are located on the NCAR Climate Data Gateway (Richter, 2022, https://doi. org/10.5065/9kcn-9y79). The remaining 5 no-SAI members are available from the NCAR Climate Data Gateway at (Mills et al., 2022, https://doi.org/10.26024/0cs0-ev98). All ARISE data may also be accessed from Amazon Web Services (NCAR, 2022, registry.opendata.aws/ncar-cesm2-arise/). The complete CESM2(WACCM6) Historical runs from which the data in this work was derived are available at Earth System Grid (Danabasoglu, 2019, https:// doi.org/10.22033/ESGF/CMIP6.11298).