The response of terrestrial ecosystem carbon cycling under different aerosol-based radiation management geoengineering

. Geoengineering has been discussed as a potential option to offset the global impacts of anthropogenic climate change and at the same time reach the global temperature targets of the Paris Agreement. Before any implementation of geoengineering, however, the complex natural responses and consequences of such methods should be fully understood to avoid any unexpected and potentially degrading impacts. Here we assess the changes in ecosystem carbon exchange and storage among different terrestrial biomes under three aerosol-based radiation management methods with the baseline of RCP8.5 using an Earth system model (NorESM1-ME). All three methods used in this study (stratospheric aerosol injection, marine sky brightening, cirrus cloud thinning) target the global mean radiation balance at the top of the atmosphere to reach that of the RCP4.5 scenario. The three radiation management (RM) methods investigated in this study show vastly different precipitation patterns, especially in the tropical forest biome. Precipitation differences from the three RM methods result in large variability in global vegetation carbon uptake and storage. Our ﬁndings show that there are unforeseen regional consequences under geoengineering, and these consequences should be taken into account in future climate policies as they have a substantial impact on terrestrial ecosystems. Although changes in temperature and precipitation play a large role in vegetation carbon uptake and storage, our results show that CO 2 fertilization also plays a considerable role. We ﬁnd that the effects of geoengineering on vegetation carbon storage are much smaller than the effects of mitigation under the RCP4.5 scenario (e.g., afforestation in the tropics). Our results emphasize the importance of considering multiple combined effects and responses of land biomes while achieving the global temperature targets of the Paris Agreement.


Introduction
The Paris Agreement, adopted under the Convention of the Parties of the United Nations Framework Convention on Climate Change (UNFCCC) in 2015, aims to limit the temperature increase to 2 • C and strive for 1.5 • C above pre-industrial levels (UNFCCC, 2015). This temperature target is very ambitious considering the rate of current warming, as such goals require not only strong mitigation efforts (e.g., Ro-gelj et al., 2016Ro-gelj et al., , 2018; , but also application of negative emission technologies or carbon dioxide removal (CDR) (IPCC, 2018). Geoengineering has been discussed as a potential option to offset the global impacts of anthropogenic climate change and at the same time help reach the global temperature targets. The complex natural responses and consequences of such methods, however, should be fully understood before implementation of geoengineer-ing to avoid any unexpected and potentially degrading impacts.
By definition, geoengineering is a deliberate attempt to modify the climate system on a sufficiently large scale to alleviate the impacts of climate change (Crutzen, 2006). Two broad categories of geoengineering, which are persistently discussed in the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC, 2013), are CDR and solar radiation management (SRM). CDR methods aim to capture CO 2 from the atmosphere and store it in reservoirs, where it stays isolated from the atmosphere for a significant period of time. This could be done in a number of different ways, from afforestation to direct air capture of CO 2 with long-term geological storage technology (Lawrence et al., 2018). SRM methods, on the other hand, aim to modify the atmospheric radiative budgets by reducing the amount of solar radiation reaching the Earth's surface to alleviate anthropogenic global warming. We hence refer to these methods as radiation management (RM) in this study following Schäfer et al. (2015).
Due to the long thermal inertia in the climate system and limitations on the maximum removal rate of CO 2 , CDR would likely require more time to lower global temperatures (Zickfeld et al., 2017) compared to RM methods. On the other hand, several proposed RM methods could stabilize or even reduce global temperature within a few years (Lawrence et al., 2018). The benefits of RM methods may not only be in reducing the current rate of increase in atmospheric temperatures, but also in mitigating climate extremes likely caused by warming (Irvine et al., 2019). Despite this encouraging potential, studies have shown numerous undesirable climatic and biophysical side effects of RM, particularly related to sudden termination of RM (e.g., Keller et al., 2014;Lauvset et al., 2017;Lee et al., 2019;Robock et al., 2009;Tjiputra et al., 2016). These studies point out that upon sudden termination of RM, the climate system will return to its "unmitigated" state within a few decades. This may lead to very large rates of change in the climatic state unless there is a solution to reduce atmospheric CO 2 concentrations. Nevertheless, our understanding on how RM influences vegetation carbon (C) dynamics at regional scales remains limited, with only a few studies published focusing on single or simplistic RM methods (Dagon and Schrag, 2019;Muri et al., 2014Muri et al., , 2018Naik et al., 2003;Tjiputra et al., 2016;Xia et al., 2016;Yang et al., 2020;Zhang et al., 2019).
In this study, we assess the response of different terrestrial biomes in their ecosystem C exchange and storage under three different RM methods using an Earth system model. There are a number of different methods studied within RM, including the aerosol-injection-based ones used in this study. The three RM methods considered in this study are stratospheric aerosol injection (SAI), marine sky brightening (MSB), and cirrus cloud thinning (CCT). The mechanisms through which different methods stabilize the climate are quite different; SAI and MSB regulate shortwave radia-tion and CCT modifies terrestrial radiation. Among the three, the most studied is SAI (e.g., Robock, 2016;Tilmes et al., 2015;Tjiputra et al., 2016), which involves increasing the backscatter of solar radiation to space by introducing a reflective aerosol layer in the stratosphere. Bright and reflective aerosols also form the foundation of another method, namely MCB (Ahlm et al., 2017;Latham et al., 2012). The principle here is to inject aerosols such as sea salt into low cloud layers over the tropical oceans to make these more effective at reflecting incoming radiation, hence reducing surface warming. If such spraying is done outside the typical cloud deck areas, the brightness of the aerosols themselves may also cool the climate (Ahlm et al., 2017). Hence, the term "marine sky brightening (MSB)" has been used in the literature (Schäfer et al., 2015;Ahlm et al., 2017), since the sky, and not just the clouds, is brightened. Furthermore, there is a less studied method referred to as CCT (Gasparini et al., 2020;Mitchell and Finnegan, 2009;Kristjánsson et al., 2015), which aims to cool by letting more longwave radiation escape to space by removing or thinning out highlevel ice clouds (cirrus clouds). This could also be done by seeding with aerosols. Since thinning of cirrus clouds would primarily act on the longwave range of the spectrum, as opposed to the other two aforementioned methods, we refer to the methods used in this study collectively as RM rather than the commonly used term SRM to be inclusive of CCT.
The modeling study by Muri et al. (2018) demonstrates that all three of these methods could potentially stabilize atmospheric temperature and reduce net radiative forcing on climate. Side effects, however, may persist as these methods alter atmospheric circulation and precipitation patterns. Studies from the Geoengineering Model Intercomparison Project (GeoMIP; Kravitz et al., 2015) demonstrate that there is substantial regional climate variation in response to different methods, scenarios, and models (e.g., Stjern et al., 2018;Wei et al., 2018;Yu et al., 2015). As a result, different terrestrial ecosystems exhibit varying patterns in vegetation production (net primary productivity, NPP). Analyses of vegetation responses show that global mean and high-latitude NPP have different patterns Lee et al., 2019). This is likely due to different RM methods resulting in different patterns of precipitation in particular. In addition to temperature and precipitation, different biomes are limited by different environmental factors, such as growing season length, dry season length, availability of sunlight for photosynthesis, and soil fertility.
This led us to investigate the following questions.
(1) What are the key factors affecting future vegetation under different RM applications? (2) If there are regional differences in environmental change under RM applications, which terrestrial biomes are affected the most in ecosystem C uptake and storage? (3) What is the impact of geoengineering termination on vegetation and terrestrial C storage? (4) What are the effects of RM applications on global vegetation compared to lower emissions and mitigation scenarios (i.e., RCP4.5)?

Model description (NorESM)
We conducted three different aerosol-based geoengineering experiments using the fully coupled NorESM1-ME, with which we investigated the impacts of idealized scenarios of aerosol-based geoengineering under the high-CO 2 RCP8.5 and the target temperature scenario RCP4.5. NorESM1-ME is based on the Community Earth System Model (CESM; Gent et al., 2011). Some of the key differences in NorESM1-ME from CESM are (1) a more sophisticated tropospheric chemistry-aerosol-cloud scheme (Kirkevag et al., 2013), (2) a different ocean circulation model based on the Miami Isopycnic Coordinate Ocean Model (MICOM) with extensive modifications , and (3) the ocean biogeochemical model, which originated from the Hamburg Oceanic Carbon Cycle (HAMOCC) model (Tjiputra et al., 2013). Both the land and atmospheric components have a horizontal resolution of 1.9 • latitude × 2.5 • longitude with 26 vertical levels in the atmosphere, whereas the ocean model employs a displaced pole grid with a nominal ∼ 1 • resolution and 53 isopycnal layers.
The land component of NorESM1-ME is CLM4 . The land C cycle module in CLM4 includes carbon-nitrogen (CN) coupling that is prognostic in CN and vegetation phenology. As a result, plant photosynthesis is also limited by the nitrogen (N) availability (Thornton et al., 2009). The CLM4 has separate state variables for C and N, which are followed through separately in leaf, live stem, dead stem, live coarse root, dead coarse root, and fine root pools. There are two corresponding storage pools representing short-term and long-term storage of non-structural carbohydrates and labile N. Sources and sinks of mineral N are implemented in the form of atmospheric deposition, biological N fixation, denitrification, leaching, and losses due to fire events. The CLM4 photosynthesis uses both direct and diffuse radiation for sunlit leaves and only diffuse radiation for shaded leaves . The plant functional types (PFTs) and land cover change distribution in CLM4 are prescribed and updated annually according to the Coupled Model Intercomparison Project phase 5 (CMIP5) global land use and land cover change dataset (Lawrence et al., , 2012. The transient PFT and land cover fields take into account historical and future climate change under the RCP8.5 scenario (1850-2100); these were implemented using the harmonized land use change scenarios and Integrated Assessment Model, respectively. Details on PFT, terrestrial C and N cycling, and land cover implementation in the CLM4 model are described in Lawrence et al. (2011). For this study, NorESM is run with a fully interactive prognostic C cycle (i.e., in emission-driven mode).

Aerosol-based geoengineering experiments
Two of the RM methods used in this study aim to reduce the amount of solar radiation reaching the surface to alle-viate global warming through spraying of aerosols into the atmosphere: SAI (e.g., Crutzen, 2006;Robock, 2016) and MSB (Ahlm et al., 2017;Latham, 1990). Another technique aims to increase the amount of outgoing thermal radiation to space by reducing the cover of high-level ice clouds: CCT (Storelvmo et al., 2013). Increasing application of RM was used to lower the total radiative forcing in the RCP8.5 baseline simulation down to a temperature level corresponding to RCP4.5, as described in Muri et al. (2018) and similar to the G6sulfur experiment of GeoMIP . The RM is started in the year 2020 on the background of the RCP8.5 scenario and continued until the end of the century. The mean of three ensemble members was used for each case. In the year 2101 the RM was ended. One ensemble member was extended for another 50 years for each case such that the effects of sudden termination of large-scale RM may be assessed.
The aerosol-based RM experiments were implemented as follows.

Stratospheric aerosol injections (RCP8.5 + SAI)
Since there is no interactive stratospheric aerosol scheme in NorESM1-ME, stratospheric aerosol properties were prescribed based on the approach of Tilmes et al. (2015), although different reference cases are used. In simulations with the ECHAM5 model, sulfur dioxide was released at ∼ 2 km of altitude (60 hPa) in a grid box at the Equator. The interactive aerosol microphysics module within the general circulation model of ECHAM5 (Niemeier et al., 2011) calculated the resulting distribution of sulfate aerosols in the stratosphere. The aerosol optical depth and distribution represented by the zonal aerosol extinction, singlescattering albedo, and asymmetry factors were implemented in NorESM1-ME and are described in more detail in Niemeier and Timmreck (2015). A number of test runs were performed to establish how much aerosol was needed to offset the anthropogenic radiative forcing between RCP8.5 and RCP4.5. The resulting aerosol layer corresponds to equivalent emissions of 5 Tg(S) yr −1 in 2050, 10 Tg(S) yr −1 in 2075, and as much as 20 Tg(S) yr −1 in 2100.

Marine sky brightening (RCP8.5 + MSB)
The sea salt emissions parameterization in NorESM1-ME is coupled to the cloud droplet number concentrations. In this way, the emissions of sea salt may interact with cloud processes, including brightening effects. Emissions of sea salt aerosols were uniformly increased at latitudes of ±45 • . This follows the approach of , and the emissions are increased over a wider latitude band to achieve an effective radiative forcing of −4 W m −2 more readily. The medium-sized aerosol bin has been found to be the most efficient at brightening clouds in NorESM . Aerosol emissions were hence increased for the accumulation-mode size, with a dry number modal radius of 0.13 µm and geometric standard deviation of 1.59, corresponding to a dry effective radius of 0.22 µm. Sea salt emission increases were of the order of 460 Tg yr −1 at the end of the century.

Cirrus cloud thinning (RCP8.5 + CCT)
With regards to cirrus cloud thinning, the Muri et al. (2014) method was used. The fall speed of all ice crystals at temperatures below −38 • C was increased. This is a typical temperature for homogeneous freezing to start occurring. The coverage of ice clouds in the CMIP5 ensemble was assessed by Li et al. (2012); NorESM was found to perform reasonably compared to satellite observations and is indeed one of the better-performing models. The terminal velocity of ice was increased by a factor of 10 by 2100, i.e., within the observational range (Gasparini et al., 2017;Mitchell, 1996).

Analysis of biomes
We follow the definition of different land biomes as in Tjiputra et al. (2016), with plant functional types (PFTs) in the CLM4 that represent certain biomes merged together (e.g., the boreal forest biome includes boreal needleleaf evergreen trees, boreal needleleaf deciduous trees, boreal broadleaf deciduous trees, and boreal broadleaf deciduous shrub PFTs in the CLM4). The biomes are static by taking a 20-year mean (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000) of the PFT distribution from the surface dataset. See Fig. S1 in the Supplement for the overall distribution of the biomes used in this study. We note that projected land use change characteristics are very different in RCP8.5 and RCP4.5 . While there is an increase in cropland and grassland (driven by food demand of an increasing population) at the expense of forested land in RCP8.5, there is an increasing area of forest due to assumed reforestation programs in the mitigation scenario RCP4.5 .

Global-scale responses under RM applications
The three RM methods alter the direct visible radiation (DVR) and diffuse visible radiation (FVR) in different directions, with little impact on the level of atmospheric CO 2 concentrations (Fig. 1). The differences in direct and diffuse radiation are attributed to how the radiation management methods are implemented; each differs in affecting longwave and shortwave radiation . Regardless of the methodological differences, all RM methods are able to reduce the net radiation at the top of the atmosphere and the global mean air temperatures close to the RCP4.5 level as expected. Global land surface air temperature (TSA) increases at a slower rate until the end of the 21st century under all three RM scenarios compared to the baseline RCP8.5 scenario, in which there is approximately 2.3 • C of difference between the RM and non-RM world at the year 2100. Under CCT application, the increase in global precipitation is somewhat higher than RCP8.5 as explained in Muri et al. (2018). CCT keeps the level of precipitation close to RCP8.5 until the year 2100 due to an amplified hydrological cycle from increased latent heat flux (Kristjánsson et al., 2015). CCT has been shown to lead to an increase in precipitation in previous studies (Jackson et al., 2016;Kristjánsson et al., 2015;Muri et al., 2018), whereby the radiative cooling of the troposphere increases the latent heat flux at the surface and hence alters the precipitation rates. SAI shows a reduced rate of increase in global precipitation similar to RCP4.5. Under MSB application, the rate of global precipitation increase falls between the SAI and CCT.
There is a large overall increase in global mean NPP until the end of the 21st century in the RCP8.5 scenario and under the three RM scenarios (Fig. 1), whereas only a small increase in NPP is simulated under the RCP4.5 scenario. At the same time, there is a large increase in the rate of soil organic matter decomposition (heterotrophic respiration: HR) in the RCP8.5-based experiments. Relatively small NPP differences are observed between RCP8.5 and the RM simulations compared to the RCP4.5 scenario. This illustrates that the CO 2 fertilization effect is much larger in regulating NPP than the effects of temperature and precipitation, as the levels of temperature and (in the case of SAI and MSB) precipitation are similar between the RCP4.5 scenario and the three RM methods on a global scale.

Regional differences in temperature and precipitation
There are no discernable spatial patterns shown in the changes in direct and diffuse radiation except that changes in direct radiation under the CCT application are more concentrated in the tropics. The SAI method shows considerably increased diffuse radiation throughout the global land areas compared to the baseline RCP8.5 scenario (Fig. S2). While TSA exhibits similar patterns across different RM applications, the precipitation patterns are more variable over space across different RM methods. The global spatial patterns of precipitation towards the end of the century (mean of 2070-2100; Fig. 2) show that CCT generally increases precipitation in the tropics and Mediterranean region relative to RCP8.5. In particular, MSB tends to increase precipitation over extratropical land more than SAI due to the regional application of the forcing (Alterskjaer et al., 2012). The spatial patterns of precipitation change in MSB mostly follow those of CCT, but the magnitude of change is smaller. On the other hand, SAI shows overall decreases in precipitation, particularly in the tropics, relative to RCP8.5. All three methods show a decrease in precipitation in the East Asia region. The differences in temperature and precipitation across the tree RM methods in different land biomes of the world show that there is no noticeable difference in mean annual temperature across the three different RM methods (Fig. 3). There is a cooling imbalance across the three RM forcings; the tropics tend to cool more than high latitudes and cooling is more pronounced in the ocean than on land, with a stronger southern hemispheric cooling for CCT . We show that precipitation patterns vary across the three methods in different biomes. In all biomes, SAI application results in the largest decrease in precipitation, followed by MSB, relative to the RCP8.5 scenario. Under CCT application, precipitation even increases beyond the RCP8.5 level. The precipitation differences across the three methods are large, particularly in the tropics and the midlatitudes, where CCT application results in higher precipitation rates than the other two methods. The differences in precipitation are amplified over time until the end of the 21st century. According to Muri et al. (2018), shortwave-radiation-based geoengineering methods exhibit a strong reduction in global precipitation levels relative to RCP8.5 but also relative to RCP4.5. CCT leads to a slight increase in global precipitation, even over the RCP8.5 levels; however, land precipitation patterns in different biomes vary. Aggregated over all biomes, precipitation changes are much smaller than over the total (ocean + land) area. Particularly, precipitation is not reduced much below the RCP4.5 levels for SAI and MSB, as in Muri et al. (2018) (compare their Fig. 2 with Fig. 1 in this study).

Biome-specific C uptake and release rate
The spatial patterns and the magnitude of NPP change under the three methods show distinct differences. There are common spatial patterns in the NPP decrease in the northwestern part of Amazonia, equatorial Africa, and eastern Asia in the three RM experiments (Fig. 2). But overall, the large increase in NPP in Europe and equatorial South America, particularly for the CCT experiment, compensates for the decreases elsewhere, hence creating a general lack of deviation as a whole from the RCP8.5 scenario (see Fig. 1). It is clear from the comparison shown for precipitation (Fig. 2) that the NPP changes are most correlated with the spatial changes in precipitation.
Under the CCT application, there is a strong increase in NPP in the tropics and the Mediterranean region but a decrease in East Asia. MSB does not show a noticeable change except increased NPP in eastern Amazonia. The spatial pattern of NPP in MSB is similar to CCT, but the magnitude is smaller in MSB. There is a strong decrease in NPP under SAI application, particularly in the tropics. These overall patterns follow similar spatial patterns as the precipitation and are highly correlated as expected (Fig. S3). The differ-ences in NPP are largely dominated by three biomes: tropical forest, grass-shrubland, and temperate forest (Fig. 4). NPP and HR in MSB and SAI simulations negatively deviate from the RCP8.5 simulation, whereas in CCT both remain at a similar level as RCP8.5 in tropical forest, grassshrubland, and temperate forest, likely due to an increased precipitation level in these biomes. But since temperature is a stronger regulator of NPP and HR in high-latitude biomes, CCT simulations also exhibit decreased NPP and HR compared to the RCP8.5 scenario. Additionally, we do not observe any noticeable changes in seasonality for NPP and leaf area index (LAI) between RM methods and the RCP8.5 sce- nario (Figs. S4 and S5) as seen in Dagon and Schrag (2019). Although there is spatial variability in precipitation patterns, there is no change in seasonality between the three RM methods and the baseline RCP8.5 scenario.
Overall, the varying precipitation patterns may be the strongest driver of the responses of global-scale C uptake and release. Changes in diffuse radiation are found to affect photosynthesis (Keppel-Aleks and Washenfelder, 2016) as diffuse radiation can be more efficient in photosynthesis (e.g., Gu et al., 2002). Under these assumptions, increases in diffuse radiation and decreases in direct radiation under SAI are expected to increase plant photosynthesis (Mercado et al., 2009;Xia et al., 2016). Increases in diffuse radiation are known to positively affect photosynthesis up to a threshold of the ratio between diffuse and total radiation at around 0.4-0.45 (Knohl and Baldocchi, 2008;Mercado et al., 2009). Across the three RM methods, this ratio ranges from 0.29 (CCT) to 0.4 (SAI) at the end of the RM application in the year 2100. The responses of NPP under changes in diffuse radiation to different RM applications exhibited in our study suggest that changes in diffuse radiation may not be as large a driver of NPP change at the global level as temperature and precipitation. Under the coupled framework of an ESM, it is very difficult to decompose the direct single effects of climatic factors due to interactions (Zhang et al., 2019), and separate simulations are necessary to directly quantify this. The N limitation implemented in the CLM4 has been shown to limit C uptake by 74 % relative to the C-only model (Thornton et al., 2007), but CLM4 still exhibits NPP biases in the tropics . It is important to note that despite the strong CO 2 fertilization and the increase in diffuse radiation, NPP in some parts of the tropics decreases under the SAI application, likely due to the strong decrease in precipitation (Figs. 2, S3 and S4).

Biome-specific C storage
Vegetation C storage in different biomes illustrates that global vegetation C storage changes are dominated by the responses of the tropical forest biome (Fig. 5). Under the baseline RCP8.5 scenario, global vegetation C storage decreases due to reduced tropical forest, temperate forest, and grass-shrubland area as part of the land use change scenario used in RCP8.5 . Compared to the baseline RCP8.5 scenario, vegetation C in Arctic tundra, boreal forest, and tropical forest biomes are affected the most under RM applications. In Arctic tundra and boreal forest biomes, all three RM scenarios result in a slightly reduced accumulation of vegetation C compared to the RCP8.5 scenario, likely due to decreased temperature, exhibiting the temperature limitation in high-latitude biomes. In tropical forest, SAI application reduces vegetation C storage relative to the RCP8.5 scenario, but CCT application slightly increases C storage due to increased precipitation (Fig. 2). The magnitude of change in global vegetation C at the end of the century due to ap-plication of different RM methods is up to 10 PgC. On the other hand, the magnitude of vegetation C reduction due to the different underlying land use change scenarios in RCP4.5 and RCP8.5 is up to 100 PgC (Figs. 5, S7). These differences are attributed to increased forest and grassland area as part of the RCP4.5 scenario . This highlights the fact that large-scale changes in vegetation C storage depend much more on anthropogenic land use change than on additional perturbations caused by application of RM in our simulations.
In tropical forest, the differences in vegetation C storage appeared to be correlated with precipitation patterns, whereby decreases and increases in precipitation in the three different methods regulate vegetation C storage. Differences in vegetation and soil C storage in the temperate zone (temperate forest and grass-shrubland), however, did not always correspond directly to varying precipitation patterns. For instance, an approximately 100-120 mm difference in mean annual precipitation shown in temperate forest and grassshrubland biomes between the SAI and CCT methods does not portray differences in vegetation C storage (Figs. 3, S6).

Effects of RM termination
Upon sudden termination of RM applications, the levels of radiation, temperature, and precipitation quickly converge to the baseline RCP8.5 scenario (Figs. 1, 3, 4). Note that the temperature does not increase to exactly the same level as the RCP8.5 scenario, which has been observed in previous studies and is due to the thermal inertia of ocean heat uptake (Tjiputra et al., 2016). As the temperature and precipitation patterns converge towards the RCP8.5 scenario, NPP also becomes similar to the RCP8.5 scenario (Fig. 1). The soil C storage decreases as RM is terminated, and towards the end of the simulation in the year 2150, soil C storage in all three RM methods is at a similar level (Fig. S6), but the magnitude is still higher than under the RCP8.5 scenario by 10 PgC globally. The likely accumulation of soil C under RM applications may be viewed as one of the positive effects of geoengineering, which was supported by a recent multimodel comparison study (Yang et al., 2020). Globally, land C accumulation associated with RM would remain on land for at least 50 years following termination . Although the termination effects seem catastrophic due to their rapidity in particular, some studies suggest that realistically the most extreme cases would be unlikely as termination could be avoided by geopolitical agreement once deployed (Parker and Irvine, 2018).

Implications and limitations
Reduced atmospheric temperature and precipitation under RM have large effects on vegetation C storage compared to the baseline scenario, RCP8.5. Under the RCP4.5 scenario, the rate of C uptake denoted as NPP is slower due  to reduced temperature, precipitation, and atmospheric CO 2 levels (Fig. 1). However, global vegetation C storage is far greater than RCP8.5 and the three RM simulations, which are based on underlying RCP8.5 scenario assumptions (Fig. 5), due to the larger forest and grassland areas in the RCP4.5 scenario . As a result, the difference in global vegetation C between the RCP4.5 scenario and the rest of the RCP8.5-based scenarios is nearly 170 PgC. This strongly suggests that on a global scale, areal changes in vegetation and land surface management play very important roles when accounting for global-scale vegetation C storage. We suggest taking this point into account when comparing the different pros and cons of technological applications such as geoengineering and mitigation options such as afforestation.
Our results suggest that even with reduced temperature stress created by RM applications, the productivity of vegetation in the three most productive biomes on Earth may be reduced due to changing precipitation patterns (particularly SAI). Therefore, considering the changes (i.e., reduction) in precipitation alone, RM may have negative effects on nonirrigated crops or food production globally. Nevertheless, the effects of CO 2 fertilization in the future are suggested to compensate for the deleterious impacts of both RM-induced temperature and precipitation changes (Pongratz et al., 2012;Xia et al., 2016). Although not directly investigated in this study, different RM methods have been shown to induce various climate extremes in addition to mitigating them, which will have profound effects on the physiology of vegetation (Aswathy et al., 2015). Indeed, some studies show seasonal variation in temperature under geoengineering (Dagon and Schrag, 2019), although we did not observe this in our study. This is not within the scope of our study, but it could be an interesting point to consider in future studies.
We acknowledge that CLM4 has numerous limitations that prevent it from accurately estimating global-scale soil C storage, and therefore we do not make an estimation of global soil C storage. But here, we compare soil C storage under different methods to understand the factors controlling the difference across the three RM methods. Soil C storage increases under RM applications compared to the baseline RCP8.5 scenario (Figs. S6 and S7) because the decrease in temperature slows the rate of soil organic matter decomposition by microorganisms. An increase in total soil C is also simulated under the RCP4.5 scenario (Fig. S7) likely due to the combination between increased vegetation C accumulation and slight reduction in soil respiration. There is an increase in soil C storage under the RCP8.5 scenario in the early 21st century due to increased NPP, but ultimately soil C decreases quickly due to accelerated soil respiration (Fig. 1). In different biomes, temperate forest exhibits the largest difference across the three RM methods; soil C storage under the SAI method is nearly 1.0 PgC higher than CCT at the end of the 21st century. This is likely due to lower precipitation in SAI, which reduced the rate of decomposition.

Conclusions
We show that the three different RM applications mainly differ in the precipitation patterns, which in turn affect differences in global-scale NPP. The precipitation differences across the three RM applications are the most pronounced in the tropics and midlatitudes, where SAI application results in the largest decrease in precipitation, followed by MSB and CCT relative to the RCP8.5 scenario. Tropical forest shows the largest variability in NPP and vegetation C storage, as the precipitation patterns vary the most across the three methods in the tropics compared to other biomes. Ultimately, all three RM applications investigated in this study reduced the surface temperature to the level of the RCP4.5 scenario, with vegetation C uptake and storage being affected due to different temperature and precipitation patterns created by the different RM methods. Our results illustrate that there are regional differences in the biogeochemical cycles under the application of large-scale RM and suggest that such effects should be taken into consideration in future shaping of climate policies. Although changes in temperature and precipitation play a large role in vegetation C storage capacity, CO 2 fertilization plays a considerable role in terrestrial C dynamics that can overshadow the effects of temperature and precipitation. Furthermore, changes in vegetation C storage under large-scale RM applications are much smaller than exhibited under the RCP4.5 scenario, which uses climate mitigation efforts through afforestation in the tropics. Hence, it is important to consider the multiple combined effects and responses of land biomes when applying different strategies to reach the global temperature targets of the Paris Agreement.
Data availability. The model simulations used in this study are archived and available on the Norwegian Research Data Archive server (https://doi.org/10.11582/2019.00007; Tjiputra, 2019).
Author contributions. HM and JT received funding; HM and JT designed and conducted simulations; HL and AE analyzed the data; HL, HM, and JS wrote the paper; all authors contributed to editing the paper.
Competing interests. The authors declare that they have no conflict of interest.
Acknowledgements. The simulations were performed on resources provided by UNINETT Sigma2 -the National Infrastructure for High-Performance Computing and Data Storage in Norway, accounts nn9182k, nn9448k, NS2345K, and NS9033K. We thank the two anonymous reviewers for providing constructive comments, which greatly improved the paper.
Financial support. This research was supported by the Research Council of Norway projects EXPECT (grant no. 229760/E10), EVA (grant no. 229771), and HiddenCosts (grant no. 268243) as well as the Bjerknes Centre for Climate Research strategic project SKD-LOES.
Review statement. This paper was edited by Ben Kravitz and reviewed by two anonymous referees.