Stratospheric aerosol injection (SAI) is being studied as a possible supplement to emission reduction to temporarily mitigate some of the risks associated with climate change. This approach would reduce global mean temperature by adding aerosols, or their precursors, to the lower stratosphere, to increase solar reflection before it reaches the surface. Existing research has shown that SAI does not bring the climate back to a previous state but creates a novel climate . The choice of how much to inject at different latitude(s) and seasons(s) we refer to as the strategy, and it is clear that different strategies will lead to different climate outcomes . As such, one cannot evaluate the benefits and risks associated with SAI without first considering the choice of strategy. To do this, it is important to better understand the design space: for any given set of climate goals, does there exist a strategy that can simultaneously meet them, or what is the “best” strategy to balance those goals? How much does the strategy itself then depend on the chosen goals? How much do the outcomes then depend on what the strategy is designed for – that is, are there underlying trade-offs between achieving different goals, and how much does this matter for assessing the outcomes?

Most existing studies evaluate the potential climate impact of SAI by simulating some particular defined strategies in global climate models. Few studies attempt to explore the broader range of possible climate responses . first treated SAI as an optimization problem, with 3 degrees of freedom of specified aerosols optimized for several climate goals. explore the trade-offs between different climate goals but based on idealized patterns of solar reduction, via reducing the incoming solar radiation directly in the climate model in multiple different spatial and seasonal patterns. This study takes a similar approach in principle to these pioneering studies but for the first time evaluates trade-offs between multiple different climate goals based on simulations of stratospheric aerosol injection, albeit in a single climate model.

A key enabler for this research is understanding how “big” the design space for SAI is, that is, how many different latitudes or seasons of injection can be chosen to optimize over before the differences from adding additional degrees of freedom are too small to matter: injection at two different latitudes very close to each other might technically have a different influence on the climate system, but the differences would not be detectable over a reasonable time frame. The design space here is chosen based on the conclusion from , that there are of order 6–8 degrees of freedom available for a cooling of up to 1–1.5 °C (though it is important to recognize that this is quite a considerable amount of cooling and not a level that would likely be reached for several decades after the start of any deployment, if ever). We consider annually constant injection at the Equator, 15° N and S, 30° N and S, and spring-time injection at 60° N and S, described more in the next sections; these reasonably span the space.

Herein we use existing climate model simulations for injection at these latitudes or combinations of latitudes and use optimization tools to determine the “best” SAI strategy – that is, the best combination of injection rates across the seven latitudes above – to minimize different objective functions. This allows us to address the questions above. If a climate model simulation were conducted with the injection rates determined from this optimization process, it of course would not exactly match the predicted outcomes, and thus it is critical to contextualize the differences between outcomes when optimizing for different metrics relative to both the error introduced by the assumption of linearity and the uncertainty caused by natural variability. The latter influences both how confident we are in how the climate model might respond to a particular choice of injection rates and also how well an observer could distinguish between different strategies – are the differences large enough to matter? The answer to that clearly depends on how much warming is being offset by SAI as the forced response to SAI will scale with the cooling, while the variability will not.

The next section describes the climate model and the existing simulations that we use in our analyses. Section 3 then describes extracting a linear model from simulation output, the validation that linearity is a reasonable approximation, and the formulation as a constrained optimization problem. Section 4 explores different trade-offs. First, if the number of climate goals is equal to or fewer than the number of degrees of freedom, then it may be possible (depending on specific goals) to find strategies that simultaneously satisfy all of the goals; a 3-degree-of-freedom example of this is the multi-objective strategy described in and subsequently used in various studies such as GLENS and ARISE-SAI . Second is the more general case where there are many distinct climate goals, such as minimizing the spatial root mean square (rms) across the planet of the difference in annual-mean temperature or precipitation relative to some earlier reference climate of the same global mean temperature. In this case, no strategy will simultaneously satisfy all of the objectives. Changing the relative weighting between goals, for example between rms temperature and rms precipitation anomalies, yields a Pareto front against which existing strategies can be compared.

All simulations of SAI strategies considered in this study are conducted in version 2 of the Community Earth System Model with the Middle Atmosphere (MA) chemistry configuration of the Whole Atmosphere Community Climate model, version 6, as the atmospheric component, CESM2(WACCM6-MA) . CESM2(WACCM6) is a fully coupled Earth system model consisting of atmosphere, land, ocean, sea ice, land ice, river runoff, and surface wave components. The atmospheric component has 70 vertical layers that extend from the Earth's surface to about 140 km in altitude, with a horizontal resolution of 0.95° in latitude and 1.25° in longitude. Compared to the comprehensive troposphere–stratosphere–mesosphere–lower thermosphere (TSMLT) chemistry configuration, the MA version of WACCM6 used here reduces computing time by about 35 % by neglecting non-methane hydrocarbon species and reactions in the troposphere . While it significantly reduces the computing cost, this simplified chemistry configuration yields climate, variability, and climate sensitivity that are comparable to the TSMLT version .

As described earlier, estimate that there are about 6–8 degrees of freedom in the SAI design space for offsetting global warming by 1–1.5 °C and suggest that a complete set of injection choices that yield detectably different surface climate responses includes injections at the following latitudes: 60, 30, 15° N, the Equator, 15, 30, and 60° S, with all but the high-latitude cases injecting the same amount throughout the year. At high latitudes, spring injection is more efficient as the aerosol lifetime is shorter, and there is minimal benefit from having aerosols in the winter; injection at 60° N is thus only during March, April, and May (MAM) and injection at 60° S only in September, October, and November (SON).

A number of different climate model simulations have been conducted in CESM2(WACCM-MA) using different combinations of these injection latitudes, summarized in Table ; all of these branch from the SSP2-4.5 simulation in 2035, with continued evolution of other non-SAI concentrations from the Shared Socioeconomic Pathway (SSP) 2-4.5, . The SSP2-4.5 simulation is also conducted in CESM2(WACCM6-MA) from the year 2015 to 2099, with three ensemble members.

At each latitude, single-latitude fixed-injection-rate simulations have been conducted with 12 Tg yr⁻¹ injection of SO₂. The setup and some results from the first 10 years of simulation are described in and , where CESM(WACCM) is compared with other climate models; these cases were extended to 35 years (from 2035 through 2069), with two ensemble members each in and a third ensemble member added in ; additional simulations (one ensemble member each) have also since been conducted for the 60° N and 60° S cases. These sensitivity experiments are intended to understand how the climate responds to a given injection location and not to represent plausible scenarios for deployment. We also use the “Arctic high” simulation from , where injection is only at 60° N but with a variable injection rate to maintain Arctic September sea ice at a target level; we average the sensitivities between this case and the fixed-injection-rate simulation at the same latitude and season.

simulate four hemispherically symmetric strategies, injecting either only at the Equator (“EQ”), equal amounts at 15° N and 15° S (“15N + 15S”), equal amounts at 30° N and 30° S (“30N + 30S”), or equal amounts in the spring at 60° N and 60° S (“60N + 60S”). Each simulation uses a feedback algorithm to adjust the injection rate to maintain the global mean temperature (denoted here as T0) at 1.0 °C above the pre-industrial level following the definition in ; this corresponds to maintaining temperature at the averaged value over 2008 to 2027; this time period will also be used as the reference climate below. In addition, the multi-objective strategy from has been simulated in this model with the same target for global mean temperature; this strategy injects SO₂ at 15° N, 30° N, 15° S, and 30° S to maintain not just T0 but also a measure of interhemispheric temperature gradient (T1) and Equator-to-pole temperature gradient (T2) using a feedback algorithm. Maintaining T1 in this climate model under SSP2-4.5 requires more injection in the Southern Hemisphere than the Northern Hemisphere ; this multi-objective strategy injects only 4 % SO₂ at 30° N, while 15° S, 15° N, and 30° S receive about 52 %, 30 %, and 14 % of injected SO₂ respectively.

Only seven cases are actually necessary here to span the space, but the availability of additional cases allows us to assess how well a linear quasi-static approximation captures the simulated response (discussed in detail in Sect.  below). In addition to any nonlinear dependence on injection rate, the pattern of climate response will differ slightly between a case that gradually ramps up injection rates to maintain T0 and one that has a step change in injection rate, even if the final global mean temperature is similar, due in part to the two cases yielding a different state for rate dependent responses, such as the Atlantic Meridional Overturning Circulation (AMOC) . For spanning the space here, we choose a set that spans the design space while minimizing the errors introduced from the linear and quasi-static assumption; we thus pick to the extent possible cases that maintain constant global mean temperature rather than the fixed-injection-rate cases and cases where the change in interhemispheric temperature gradient is not too large, as we expect these situations to be those where the linear quasi-static approximation will be better. For use in the analysis in Sects. 4–5 we thus use the four hemispherically symmetric strategies, the multi-objective strategy, and complete that set with two single-latitude cases, 60 and 30° N. While it would be conceptually simpler to just use the single-latitude injection cases, the ability to usefully predict the response of a linear combination of strategies is crucial here, and thus it is worth the slight extra complication to minimize the errors from the linear quasi-static approximation as much as possible. Note that the multi-objective strategy has been used in several single-model simulation datasets for impact analyses (e.g., GLENS and ARISE-SAI, as noted earlier), and more recently the simpler hemispherically symmetric 30N + 30S strategy has been chosen for the next set of model intercomparisons under GeoMIP , and the 60N + 60S strategy has also received some attention . Indeed, a possible way to think about the results of optimization in Sect. 5 is to ask whether these existing strategies are good enough for evaluating the policy-relevant impacts of a deployment of SAI or whether there are other strategies that should be considered more carefully.

Existing studies have only explored strategies that can manage up to three climate goals at the same time, by adjusting injection rates at up to four latitudes simultaneously (but only adjusting 3 degrees of freedom independently; see ). In principle, with 7 independent degrees of freedom available, it might be possible to simultaneously satisfy seven distinct goals. However, this is not in general possible. Using the constrained quadratic optimization formulation introduced above, we can evaluate whether there exists a strategy that can simultaneously achieve m climate goals (m≥1) by looking for solutions J(x)=0 (or close to zero). We consider the following five climate goals, in addition to maintaining the global mean temperature T0: P0, T1, T2, ITCZ, and SSI. This was explored by , but only 30° S, 15° S, 15° N, and 30° N were used in that study.

We first note that while T0 and P0 can indeed be simultaneously managed (the 60N + 60S strategy nearly does this, by “undercooling” the tropics where most of the precipitation is), these two cannot be managed at the same time as any of the additional variables above. We therefore restrict our attention to the remaining large-scale metrics and assess whether T0, T1, T2, ITCZ, and SSI can be managed simultaneously. This is not quite achievable in this climate model but nearly so, with the residual differences in these metrics relative to the reference period smaller than the uncertainty due to natural variability. Figure  compares this injection strategy and the seven existing strategies in terms of changes in T1, T2, ITCZ, and SSI relative to the reference period (2008–2027). The injection distribution across latitudes in this optimized strategy is similar to that of the multi-objective strategy but with most of the 15N + 15S injection being replaced by equatorial injection.

Past strategy design has typically considered targeting several large-scale climate metrics (e.g., ). However, focusing solely on large-scale climate metrics can potentially be misleading; e.g., the 60N + 60S strategy can simultaneously balance both global mean temperature and global mean precipitation, but there is considerable spatial variation in temperature in particular relative to the reference climate. In this section, we therefore turn our focus to minimizing regional (grid-scale) changes from the reference period, and in particular on trade-offs, through two examples: first, changing the relative weighting between area-weighted mean-square temperature changes and precipitation changes over the entire Earth surface and then between temperature and precipitation minus evaporation (P-E), over land only. There are of course an infinite number of cases one could explore, but two is sufficient to illustrate broader conclusions. The objective function representing the area-weighted mean-square temperature and precipitation differences from the reference period is given by 10J(x)=γ∑j∑iwij∑k=1k=7xkβk,Tij2+(1-γ)∑j∑iwij∑k=1k=7xkβk,Pij2, where i is the latitude index, i∈[1,192], and j is the longitude index, j∈[1,288]. The change in temperature in the region indexed as ij is ∑k=1k=nxkβk,Tij, and the change in precipitation in the corresponding region is ∑k=1k=nxkβk,Pij, and the weighting is chosen here to weight by area. The values of γ and 1-γ are the relative weights for temperature and precipitation, respectively, where γ∈[0,1].

The optimal solutions as a function of γ form a Pareto front, which shows the trade-off between prioritizing temperature vs precipitation; this is shown in Fig.  along with the existing strategies. The top-left-most point on the Pareto front is temperature-optimal (γ=1), and the bottom-right-most point on the Pareto front is precipitation-optimal (γ=0). The hemispherically balanced strategies are much closer to the Pareto front than the single-hemisphere strategies and non-SAI case. In this model, the rms temperature change is greatly reduced regardless of the strategy, except for the cases that are not at all hemispherically balanced.

While the predicted behavior of the optimized cases must by definition be “better” (for the metrics being optimized) than any of the existing strategies, the most critical question is how much better and in particular two slightly related questions: (i) whether the differences are meaningful given the information we have and (ii), if they are meaningfully different, whether the differences are large enough to matter in a deployment. The extent to which general conclusions from a single model study would hold in the real world also matters; of course specific quantitative results will differ, but even with a single model it is certainly valuable to gain insight into the extent to which multi-latitude optimization is likely to help achieve a desired climate, the extent to which the solution depends on the chosen metric, etc.

There are two sources of error in predicting the response to a linear combination of injection choices in this particular climate model. One is the extent to which the results will inevitably differ from the linear and quasi-static approximation, and the other is that the estimate of the response to any simulated strategy includes the effects of natural variability. Figure  (right panel) illustrates the magnitude of both of these effects relative to the differences across strategies.

First, the purple square represents the value of the multi-objective strategy predicted using the linear combination method from Sect. . The predicted value is about half-way between the simulated multi-objective strategy and the temperature-optimal point on the Pareto front. This suggests that the difference between the multi-objective strategy and the temperature-optimal strategy may not be as significant as the plot suggests.

To evaluate the effect of natural variability, Fig.  (right panel) also shows the rms temperature and precipitation change of each individual ensemble member for each simulated strategy and also the Pareto front calculated using only a single ensemble member (small circles, empty triangles) and the ensemble means (large circles, solid black squares). This shows how distinct each strategy is in the context of internal variability. The difference between the 15N + 15S strategy and the multi-objective, EQ, and 30N + 30S strategies is of similar size to the variation due to internal variability, making these strategies only detectably different from each other for these metrics because we have multiple ensemble members available. The strategy predicted to optimize rms temperature differs from the multi-objective case by an amount comparable to our ability to distinguish them given the finite simulations that we have. The magnitude of this second effect could be reduced by running more and more ensemble members, but of course in the real world there is only ever a single ensemble member; even if one simulated many more ensemble members and used those to more precisely define an optimum strategy (and even if it were the optimum strategy in the real world, not just the model), an observer might not be able to tell the difference between the optimum and the multi-objective strategy unless the level of cooling were much larger than the level used (roughly 1 °C cooling) in the simulations used herein. Arguably, this assumes that climate models adequately represent the signal-to-noise ratio of responses to perturbations. However, some studies have suggested that models tend to underestimate the predictable component of forced response, in particular on the regional scale .

Figure  shows the results of the optimization for each ensemble member individually. This demonstrates that for any given ensemble member, and thus likely for any given real-world scenario, there exists strategies that do better than any existing strategy in these criteria. However, it is impossible to know in advance what the specific realization of natural variability will look like, so it is impossible to predict what the better strategies would be.

Nonetheless, the predicted optima deviate further from the existing strategies as the relative weighting on precipitation is increased. Figure  shows the spatial temperature and precipitation changes from the reference period at different points along the Pareto front. The optimal temperature and balanced strategies have similar patterns, whereas the optimal precipitation strategy undercools the tropics and overcools the poles, which is consistent with results of prior studies that show that SAI at higher latitudes has a lesser effect on precipitation .

There are two broad observations from exploring the trade-off between minimizing the spatial mean-square temperature deviations from a reference climate and the spatial mean-square precipitation deviations. First is that the existing multi-objective strategy is pretty good at minimizing the overall mean-square temperature changes. The 30N + 30S and 15N + 15S hemispherically symmetric strategies might be similarly good in other models that do not have a strong cloud fast adjustment to CO₂, which is a key factor responsible for CESM2(WACCM)'s need to inject more in the Southern Hemisphere to balance T1 . Second is that it seems likely that the mean-square precipitation response could be improved somewhat relative to the existing multi-objective strategy without substantially affecting the mean-square temperature response, but these differences will likely be too small to matter in the early decades of deployment when the cooling levels are small. The simulations on which the results herein are based cooled by of order 1 °C, yet the scatter from natural variability alone is substantial relative to the gains from optimization. Any realistic deployment scenario is likely to take at least decades, if ever, to reach such a substantial cooling level.

Figure  shows the SO₂ injection rates across the seven latitudes for the strategies along the Pareto front discussed above. Temperature-oriented strategies have most or all of the injection at the Equator, 60° N, and 60° S, whereas precipitation-oriented strategies distribute injection across more latitudes. Notably, equatorial, 60° N, and 60° S are favored by this optimization algorithm despite being the latitudes that are not part of the multi-objective strategy, which is the best-performing pre-existing strategy based on the rms temperature metric.

There are of course an infinite set of metrics that could be considered, and temperature and precipitation are just two; we therefore consider a second example to see if similar conclusions hold. While temperature and precipitation are commonly explored, precipitation minus evaporation (P-E) may be more relevant to impacts, and further, it is changes over land that are most relevant for most impacts. Figure  shows the results of optimizing for temperature over land vs P-E over land. The error induced by the linear approximation method is larger than the difference between the multi-objective strategy and much of the Pareto front. Thus not only is the difference too small to be meaningful between the existing multi-objective strategy and the strategy predicted to minimize land-average temperature, but this also holds for a considerable portion of the Pareto front.

Figure  shows the injection rates for these strategies. These land-oriented strategies utilize more latitudes than globally oriented ones.

The analyses for the rms temperature vs precipitation trade-off and the rms land temperature vs rms land precipitation minus evaporation trade-off were repeated, with the values in each grid cell normalized by the standard deviation in that grid cell during the reference period. The results were very similar and can be found in the Supplement.

The optimization framework also allows straightforward exploration of additional questions, such as how important it is to be able to inject at specific latitudes. For example, the optimized strategies rely heavily on equatorial injection, yet the multi-objective strategy does not include equatorial injection at all. In order to investigate how much the injection at the Equator contributes to the optimality of the found optimal strategies shown in Fig. , we ran the same optimization with equatorial injection removed as an option. Figure  (left panel) shows the injection rates at different latitudes for this case compared to the full seven-latitude case. The equatorial injection is mostly replaced with 15° N and 15° S. This also saves about 3 Tg yr⁻¹ of SO₂ injection. Figure  shows the Pareto front for the no-EQ case compared to the full seven-latitude case. The no-EQ Pareto front has slightly larger residual climate changes, but the difference is far less than the error potentially induced by the linear quasi-static approximation used. This shows that avoiding equatorial injection can reduce the needed injection rate without having a significant effect on the ability of SAI to counteract the effect of GHG, and more generally, injection strategies utilizing fewer number of injection latitudes can produce indistinguishably different responses, suggesting that there may be effectively fewer than the 6–8 degrees of freedom found by .

Previous work has considered designing SAI strategies by selecting at most three distinct climate goals, selecting at least as many latitudes of injection, and then designing strategies to balance those climate goals by adjusting injection rates across the different latitudes. However, research suggests that there should be of order 6–8 meaningfully independent degrees of freedom available by selecting additional injection latitudes and/or seasons. This raises the obvious question of the extent to which that would allow “better” SAI strategies to be designed, where better is of course only meaningful in the context of specifically chosen climate goals, and thus it is also critical to know how much the optimal strategy depends on the choice of goals (e.g., on who gets to decide what matters and what does not).

This study introduces a method of creating SAI strategies that are optimized to meet certain climate goals. With 6–8 degrees of freedom, it should be possible to meet more than three climate goals simultaneously, and indeed we have demonstrated that this is true; e.g., it is possible to simultaneously maintain global mean temperature, interhemispheric and Equator-to-pole temperature gradients, as well as the latitude of ITCZ, and Arctic September sea ice extent.

However, there are many more variables that matter than available degrees of freedom, and it is not possible to fully restore the entire climate state back to an earlier reference period. (This is not entirely due to inherent characteristics of SAI but also partially a result of the fact that the earlier reference climate is in a transient, and not equilibrated, state; .) We explore this through several examples. First, we consider how well regional temperature and precipitation goals can be met simultaneously. There is an inherent trade-off between keeping the spatial rms of temperature and the spatial rms of precipitation as close as possible to reference period levels – better performance for one will result in worse performance for the other, leading to a Pareto-optimal front as the relative weighting is changed. For any such relative weighting of temperature and precipitation, there is a strategy that meets these goals better than any strategy that has been simulated thus far. However, the extent to which these strategies are meaningfully better or detectably different from existing strategies is not so clear. The figures in Sect.  show that the difference between the simulated and predicted multi-objective strategy is similar to the difference between the simulated multi-objective strategy and the temperature-focused optimal strategies on the Pareto front. This means that it is not possible to determine, at least in this study, if there is a strategy that can actually do better than the multi-objective at maintaining temperatures around the world, at least using the rms metric. Figure  shows that natural variability, at least in the model, is similarly large, adding another layer of uncertainty as to whether such a difference could be detected in models or in the real world. There is, though, a strategy in the center of the Pareto front that maintains temperature just as well as the multi-objective strategy but is likely detectably better at maintaining precipitation for of order 1 °C cooling or more – at lower cooling, these differences would be more difficult to detect.

Changing the goals to rms temperature change over land and rms precipitation minus evaporation change over land yields similar results. The difference between the multi-objective strategy and the strategy that minimizes land-averaged temperature changes is perhaps even less significant in this case. The spatial patterns of the responses are very similar to those from the global temperature vs precipitation optimization, suggesting that the exact choice of metric used may not be particularly important.

One interesting detail of the results is the injection latitudes chosen by the optimization algorithm. Most of the strategies focus the injection at the Equator, 60° N, and 60° S. When designing the multi-objective strategy , the Equator was not chosen as an injection latitude because it results in a high aerosol concentration in a narrow latitude band near the Equator, while dividing that injection between 15° N and 15° S produces a more uniform aerosol optical depth (AOD). The high latitudes 60° N and 60° S were also not originally chosen as the aerosol lifetime in the stratosphere is much shorter compared to injection at lower latitudes . Despite this, their combined effect does the best job at counteracting the effects of GHG on temperature and precipitation. However, removing equatorial injection as an option for the optimization results in that injection being mostly shifted to 15° N and 15° S, and the ensuing climate responses are not detectably worse. This suggests that there are fewer meaningfully distinct degrees of freedom than the 6–8 suggested by . We note that the underlying assumption here is that climate models adequately represent signal-to-noise ratio of responses to perturbations. However, some studies suggest that even state-of-the art climate models can underestimate the predictable component of forced response, in particular on the regional scale , which if true could influence the results here.

A key question is whether there are additional strategies that ought to be considered in understanding the range of climate responses from SAI beyond those already identified – for example, are there ways that it could be optimized to reduce side effects beyond those strategies already considered in the literature? A critical observation in answering this question is that the amount of cooling that is likely in the first few decades of deployment is probably much less than 1 °C, and by that time there will be substantially more information available about the effects of SAI and its dependence on strategy. For this range of cooling (much of the analysis herein is scaled per degree Celsius of cooling), our conclusion is that the differences between existing strategies and optimized strategies are often not very large, particularly for temperature-based metrics. There may be strategies that do a little bit better on precipitation-based metrics, but even there the differences may not be large enough to change the answer of whether or not to begin a deployment of SAI.

There are several ways in which this study could be built upon in the future. The first is to use more climate models; this study only uses one, CESM2-WACCM(MA), and performing this analysis in other models may yield different results . However, it is unlikely that the qualitative conclusions would change . Even with just one climate model, more ensemble members of each basis strategy could be run. This would result in greater precision of the results as well as increased power in detecting differences between strategies. It would not, however, change the conclusion that many of the differences found are comparable or smaller than internal variability.

Another way to improve this study would be with a better estimation method. This study uses a linear quasi-static approximation, the limitations of which are discussed in Sect. . Development of dynamic climate emulators for SAI has been gaining attention recently . If a more advanced emulator that is nonlinear or accounts for the different timescales that exist in the climate system were developed, it could be used to perform this analysis with more accurate results.

This study only uses a small set of objective functions for optimization, and the analysis could be extended to consider other metrics, including other climate variables and non-annual-mean quantities. Herein we weight departures from a reference climate quadratically, but other dependencies could be explored, including non-symmetric ones.

Lastly, exploring the space of what can be achieved with SAI is valuable, but actually designing a strategy to achieve these goals would require additional effort due to the inevitable uncertainties and nonlinearities; even achieving something close to a predicted optimum in the same climate model would likely require use of a feedback controller that adjusts the injection at each latitude over time.