Differing precipitation response between Solar Radiation Management and Carbon Dioxide Removal due to fast and slow components

Solar Radiation Management (SRM) and Carbon Dioxide Removal (CDR) are geoengineering methods that have 10 been proposed to prevent climate warming in the event of insufficient greenhouse gas emission reductions. Here, we have studied temperature and precipitation responses to CDR and SRM with the RCP4.5 scenario using the MPI-ESM and CESM Earth System Models (ESMs). The SRM scenarios were designed to meet one of the two different climate targets: to keep either global mean 1) surface temperature or 2) precipitation at the 2010-2020 level via stratospheric sulfur injections. Stratospheric sulfur fields were simulated beforehand with an aerosol-climate model, with the same aerosol radiative properties 15 used in both ESMs. In the CDR scenario, atmospheric CO2 concentrations were reduced to keep the global mean temperature at approximately the 2010-2020 level. Results show that applying SRM to offset 21st century climate warming in the RCP4.5 scenario leads to a 1.42% (MPI-ESM) or 0.73% (CESM) reduction in global mean precipitation, whereas CDR increases global precipitation by 0.5% in both ESMs for 2080-2100 relative to 2010-2020. In all cases, the simulated global mean precipitation change can be represented as the sum of a slow temperature-dependent component and a fast temperature20 independent component, which are quantified by regression method. Based on this component analysis, the fast temperatureindependent component of CO2 explains the global mean precipitation change in both SRM and CDR scenarios. Based on the SRM simulations, a total of 163-199 Tg(S) (CESM) or 292-318 Tg(S) (MPI-ESM) of injected sulfur from 2020 to 2100 was required to offset global mean warming based on the RCP4.5 scenario. To prevent a global mean precipitation increase, only 95-114 Tg(S) was needed and this was also enough to prevent global mean climate warming from exceeding 2 degrees above 25 preindustrial temperatures. The distinct effects of SRM in the two ESM simulations mainly reflected differing shortwave absorption responses to water vapor. Results also showed relatively large differences in the individual (fast versus slow) precipitation components between ESMs.


3
Problems and side effects associated with SRM have been discussed extensively (Robock et al. 2009;Royal Society 2009).
One fundamental problem is that compensating GHG induced warming with SRM would decrease global mean precipitation through the direct radiative effect described above. This can be understood as follows. Given a GHG concentration increase, less outgoing longwave (LW) radiation escapes to space, causing surface temperatures to increase until a new equilibrium is 70 achieved. SRM methods aim to offset this temperature increase by reducing incoming shortwave (SW) solar radiation. Thus, even though the total radiative flux may be the same between an increased GHG + SRM scenario and the unperturbed climate, the atmospheric SW and LW radiative fluxes differ. This has been shown in general to lead to a decrease in global mean precipitation (Bala et al., 2008). In general, the suite of climate responses arising from a LW radiation change cannot be fully compensated by modifying SW radiation. Use of SRM thus involves a tradeoff between temperature and precipitation on the 75 global scale.
CDR methods are considered less risky than SRM as these methods remove CO2 from the atmosphere and thus reduce the atmospheric GHG concentration (Royal Society 2009). However, climate change is not necessarily a reversible process due to factors such as sea and glacier ice melt, sea level rise, and carbon cycle changes (Frölicher and Joos 2010;Wu et al., 2015). In 80 addition, climate does not adapt immediately to a change in radiative forcing. For example, due to ocean thermal inertia global temperatures will continue to change for decades or even centuries after a given radiative forcing perturbation. It is therefore important that CDR scenarios be studied to assess climate responses beyond changes to global mean temperature.
In section 3 of this study the temperature and precipitation responses to CDR and SRM are simulated with two Earth System 85 Models (ESMs). The mechanisms driving global mean precipitation changes are assessed by separately examining the temperature-dependent slow response and radiatively-induced fast response for differing magnitudes of SRM and CDR. This methodology can be used to better understand impacts of CDR and SRM. Unlike in several previous studies, here, fast and slow responses are quantified by a regression method instead of a fixed sea surface temperature (SST) method (Duan et al., 2018;Myhre et al., 2017;Samset et al., 2016). An advantage of the regression method is that it separates total temperature-90 dependent and independent responses while in the fixed-SST method land temperature adjustments are included in the temperature-independent fast response. We also study regional disparities in temperature and precipitation responses for both geoengineering techniques, and estimate the SO2 emission amounts required to keep either temperature or precipitation at present-day levels.

95
In section 4 we simulate three geoengineering scenarios against the Representative Concentration Pathway 4.5 (RCP4.5) scenario (Thomson et al., 2011). We examine two SRM scenarios designed to address two different climate targets: keeping either global mean 1) surface temperature or 2) precipitation at 2010-2020 level via stratospheric sulfur injections. A CDR scenario is designed to keep the global mean temperature at approximately the 2010 level. We used an aerosol-climate model to simulate stratospheric aerosol fields, and two separate ESMs (MPI-ESM and CESM) to simulate the climate response to 100 SRM and CDR.

Models
This study was conducted using three climate models: one aerosol-climate model with fixed sea surface temperature, and two ESMs. We first simulated stratospheric aerosol fields with the aerosol climate model ECHAM-HAMMOZ. We then 105 implemented the radiative properties of these fields in two ESMs-the Max Planck Institute Earth System Model (MPI-ESM) and the National Center for Atmospheric Research's Community Earth System Model (CESM)-for simulation of the various scenarios. For each scenario, we run a three-member ensemble with both ESMs.

ECHAM -HAMMOZ
We defined the radiative properties of the aerosol fields resulting from stratospheric injections of sulfur dioxide (SO2) with the 110 MAECHAM6.1-HAM2.2-SALSA global aerosol-climate model (Bergman et al., 2012;Kokkola et al., 2008;Laakso et al., 2016;Stevens et al., 2013, Zhang et al., 2012. In this model, the ECHAM atmospheric module  is coupled interactively with the HAM aerosol module (Zhang et al., 2012). The HAM module calculates the emissions, removal, and radiative properties of aerosols along with the associated gas and liquid phase chemistry. The model includes the SALSA explicit sectional aerosol scheme (Kokkola et al., 2018), which describes aerosols based on number and volume size 115 distributions with 10 and 7 size sections for soluble and insoluble particles, respectively. The model simulates the microphysical processes of nucleation, condensation, coagulation, and hydration. The model was configured as described by Laakso et al. (2017). Simulations were performed at T63L47 resolution, which corresponds approximately to a 1.9° x 1.9° horizontal grid with 47 vertical levels reaching up to ~80 km. The model accurately simulates stratospheric aerosol loads and radiative properties based on observations of the 1991 Mt. Pinatubo eruption (Laakso et al., 2016;Kokkola et al., 2018). It 120 should be noted that this model configuration does not simulate the quasi biennial oscillation at L47 resolution. Hydroxyl radical (OH), which impacts the oxidation of SO2 to sulfate as well as ozone concentration, is accounted for through prescribed monthly mean fields.

MPI-ESM and CESM
MPI-ESM (Giorgetta et al., 2013) consists of the same atmospheric model (ECHAM6.1) as ECHAM-HAMMOZ, and the 125 MPI-ESM simulations here also employed the same T63L47 resolution as the ECHAM-HAMMOZ simulations described above. MPI-ESM includes the JSBACH active land model (Reich et al., 2013) (Ilyina et al., 2013). The tropospheric aerosol climatology of Kinne et al. (2013) is used in all scenarios. CESM version 1.2.2 (Hurrell et al., 2013) consists of the Community Atmospheric Model (CAM4) which is used with horizontal resolution of 0.9° x 1.25° and 26 vertical levels up to 40 km (finite volume grid). It is coupled to the Parallel Ocean Program (POP2) ocean model, the Community Land Model (CLM4) and Community Ice CodE (CICE4) sea ice model.

Implementing prescribed aerosol fields in ESMs
To examine the effects of solar radiation management by stratospheric sulfur injections we implemented prescribed sulfate 135 fields in ESMs as described by Laakso et al. (2017). First, we used ECHAM-HAMMOZ to simulate aerosol fields resulting from gaseous SO2 injections. These simulations include a 2-year spin-up period followed by a 5-year steady-state period. From this 5-year period, mean stratospheric values for aerosol optical depth (AOD), single scattering albedo (SSA), and the asymmetry factor (ASYM) were archived as monthly output in 14 SW bands plus 16 LW bands for (absorption) AOD. We then implemented these fields in the two ESMs as prescribed zonal and monthly mean fields. ECHAM-HAMMOZ and MPI-140 ESM share the ECHAM atmosphere model, which itself uses the Rapid Radiative Transfer Model. Because the same resolution (T63L47) was employed for both ECHAM-HAMMOZ and MPI-ESM, the only differences in aerosol radiative properties between the models was caused by zonal and monthly averaging of the radiative properties (however, the total AOD did not vary between the two). In the case of CESM, aerosol fields from the ECHAM-HAMMOZ simulations had to be interpolated horizontally to 0.9° x 1.25° and to 26 vertical levels. Because CAM4 uses different wavelength bands than does ECHAM (7 145 LW bands and 19 SW bands), we interpolated the aerosol optical properties accordingly.
The above implementation ensures that SRM radiative effects are consistent in both ESMs, while also enabling longer-term analyses since computationally expensive aerosol microphysics are prescribed rather than simulated online. The aerosol radiative effects are nevertheless based on explicit simulations of aerosol microphysics and of the resulting aerosol size 150 distribution and spatial-temporal variability. Our methodology is therefore more physically realistic compared to approaches that simply reduce the solar constant or apply idealized zonally homogenous aerosol fields. Realistic simulation of aerosol microphysics is necessary for robust prediction of the associated radiative effects, which depend on the size, properties, and location of the particle. In the stratosphere, particle life times are roughly one year, so that microphysical processes such as coagulation and condensation play a greater role than in the troposphere. As a result of these microphysical processes, radiative 155 forcing from stratospheric sulfur injections does not increase linearly with the amount of injected sulfur and thus radiative impacts cannot be scaled linearly based on level of the injection level (Niemeier et al., 2015).

Simulations
To simulate SRM stratospheric aerosol fields, we performed six SRM and one control simulation with ECHAM-HAMMOZ.
Here, SO2 was injected continuously throughout the simulation at 20 km altitude between 10 N and 10 S latitudes. Each of the 160 ECHAM-HAMMOZ simulations included injection of 1, 2, 3, 4, 5 or 6 Tg(S)/yr. We divide the ESM simulations into two groups: 1) component analysis simulations, and 2) scenarios. Component analysis simulations are performed to enable subsequent separation of the slow (temperature-dependent) and fast (temperatureindependent) responses to the specific forcing agent based on a regression method (Gregory et al., 2004). In this method, an 165 individual forcing agent (CO2 or SRM) is added to the steady state climate conditions, and different climate variables are regressed against the global mean surface temperature change. The fast and slow responses for a specific forcing agent is then obtained from the fitted regression line. Specifically, the fast temperature-independent response is derived as the intercept (zero temperature change), while the slow temperature-dependent response is derived as the slope. This analysis is done for three purposes: 1) to evaluate the implementation of the stratospheric aerosol fields across the two ESMs, 2) to quantify 170 differences in radiative forcing and climate sensitivity between models under a specific forcing agent, and 3) to separate the fast and slow precipitation responses of the forcing agents. A total of 9 scenarios are simulated with both ESMs: preindustrial, 6 SRM scenarios with 1, 2, 3, 4, 5, 6 Tg(S) injections, and 2xCO2 and 4xCO2 conditions. All component analysis simulations start from a radiatively balanced climate for pre-industrial conditions. A forcing agent 175 (CO2 or SRM) is introduced at the outset of the simulation while other conditions are kept pre-industrial levels. We simulated three 20-year ensemble members for each component analysis scenario in Table 1.
Scenario simulations were based on RCP 4.5 (Moss et al., 2010;van Vuuren et al., 2011), and included: i) one baseline scenario with no geoengineering (RCP45), ii) two SRM scenarios designed to keep global mean surface temperature (SRM-TEMP) or 180 precipitation (SRM-PRECI) at 2010-2020 mean values, and iii) one CDR scenario designed to keep global mean surface temperature at the 2010-2020 mean value (CDR). In each case, three ensemble members were simulated for years 2010-2100.
In the RCP 4.5 scenario, radiative forcing stabilizes several decades before the end of the simulations (year 2100), leading to warming clearly below that seen high emission scenario (RCP 8.5), but above the targets defined in the Paris Agreement. For 185 the SRM-TEMP and SRM-PRECI scenarios, the global mean temperature or precipitation was kept close to the 2010-2020 mean via changing level of stratospheric sulfur injections.
Xyear-1 + Xyear-2 are the corresponding global mean value in the two preceding years. A running window of two preceding years is used to avoid undue influence from natural variability in global mean temperature or precipitation. Use of a longer window is suboptimal because the temperature or precipitation change the year following an SRM adjustment then does not carry sufficient weight for the subsequent evaluation. This can lead to overly large temperature or precipitation changes before the 200 need to act is recognized.
The A parameter is a threshold value and set to 0.2 K in SRM-TEMP, which based on our test simulations is generally larger than natural variability. For SRM-PRECI A is defined to correspond to a 0.5% change in the global mean precipitation in the model. If both of the above conditions are false, the stratospheric sulfur injections are maintained at the previous year's level. 205 SRM simulations are initialized with 1 Tg(S)/yr injections at year 2020.
An approximation inherent in this approach is that transitory ramp-up and ramp-down periods in the stratospheric aerosol burden with 1 Tg(S)/yr changes in SRM are not taken into account. Thus the simulated SRM changes take place faster than would occur in the real world. For example, the ECHAM-HAMMOZ simulation with 5 Tg(S)/yr injections requires 6 months 210 to achieve 70% of the ultimate steady state aerosol optical depth (AOD) (533nm) after starting from background conditions..
When sulfur injections are suspended in the ECHAM-HAMMOZ simulation, the AOD decreases by roughly by 40% over the course of the first year. However, since the sulfur changes in our ESM simulations are only ±1 (Tg(S)/yr)) and do not usually occur in consecutive years we can assume that neglecting this time lag does not significantly alter our overall results."

215
In the CDR scenarios, CO2 removal was likewise initialized at year 2020. Here, the annual CO2 increase based on RCP4.5 was counteracted by a 1% annual removal of the atmospheric CO2 concentration. This process was continued until year-2070, when radiative forcing is stabilized in the RCP4.5 scenario. Accounting for both RCP4.5 emissions and CDR, the total atmospheric CO2 concentration is then reduced yearly by 0.3 -0.6 % between 2020-2070 ( Fig. 1). Removing 1% of atmospheric CO2 in 2020 corresponds to negative emissions of approximately 8.7 GtC/yr. As carbon cycle feedbacks (i.e., outgassing from natural 220 carbon sinks) lower the efficiency of CDR (Tokarska and Zickfeld, 2015), the actual amount of sequestered carbon would in reality need to be even higher than this. Achieving such high negative emissions in 2020 would be virtually impossible. The rate required is higher than the maximum estimated sustainable potential of the highest-potential negative emission technologies (Fuss et al., 2018), without even considering competition between the methods. Among SR15 scenarios pursuing the most aggressive CDR, the median carbon sequestration rate for the primary employed method (bioenergy with carbon 225 capture and storage) reaches ~4 GtC/yr in 2100 . Thus, the CDR scenario employed here should be considered as an idealized high-end carbon removal scenario, and we do not speculate how CDR could be achieved and do not study impacts of any specific CDR technology. All non-CO2 GHG concentrations and other forcings in the CDR scenario are the same as in RCP45.
Fast and slow components of radiation and precipitation was quantified from the component analysis simulations by regressing variables of interest against temperature. These simulations were 20 year long. In each case, three ensemble members were simulated. Simulations were initiated in stable preindustrial conditions. In addition, forcing agent (CO2 or SRM) were included, which causes radiative imbalance and results in warming or cooling. Then, annual global mean values were regressed against temperature to separate temperature-dependent and independent responses. 235

Evaluating the implementation of stratospheric sulfur aerosol fields in MPI-ESM and CESM
We evaluated the stratospheric aerosol implementation by comparing clear-sky aerosol radiative forcing in the two ESMs with that in ECHAM-HAMMOZ. The ECHAM-HAMMOZ simulations were performed with fixed sea surface temperatures, with aerosol radiative forcing calculated based on the change between a scenario with stratospheric sulfur injection and the control simulation. To calculate the corresponding radiative forcing in the ESMs, a regression (Gregory) method was used (Gregory 240 et al., 2004) (Fig. 2), which also provides the climate feedback parameter which can be used to analyse different responses in two ESMs. First, we calculated the clear-sky shortwave flux and temperature anomaly compared to the stable preindustrial conditions (Preind simulation) for each year individually, and performed a linear regression between the two variables. Then, we obtained a radiative forcing as the clear-sky shortwave flux anomaly of the linear regression line at zero temperature anomaly (i.e., when the climate system has not yet adjusted to the forcing). 245 The SW radiative forcing in both ESMs was in good agreement with that in ECHAM-HAMMOZ (dashed lines). Radiative forcings were slightly smaller (i.e. less negative) in MPI-ESM than ECHAM-HAMMOZ, likely due to differing background conditions (preindustrial in MPI-ESM versus year-2000 in ECHAM-HAMMOZ, and thus more extensive ice cover in the MPI-ESM simulations). The zonal distribution of radiative forcing also agrees well between the models (not shown). 250 Stratospheric aerosols absorb some LW radiation, and the LW radiative forcing in MPI-ESM agrees well with that in ECHAM-HAMMOZ. However, CESM exhibits 37% (on average) weaker LW radiative forcing than ECHAM-HAMMOZ. This is probably due to the different radiative transfer models in CESM-CAM4 (9 LW radiation bands) and ECHAM-HAMMOZ (16 LW radiation bands). However, LW radiative forcing was small compared to the SW forcing, and this underestimation does not significantly affect the results or conclusions of this study. Since LW radiative forcing (warming effect) is weaker while 255 SW radiative forcing (cooling) is stronger in CESM than in MPI-ESM, SRM resulted in slightly more clear-sky cooling in CESM.
We see in Fig. 2 that SW radiative forcing does not increase linearly with the amount of injected sulfur. This is because more sulfur condenses onto existing particles, and small particles coagulate more efficiently with larger particles when the sulfate 260 burden is increased. This leads to lower particle numbers and larger particle sizes per unit sulfur injected (Heckendorn et al., quite linearly with the amount of injected sulfur as shown, as also demonstrated by Niemeier and Timmreck (2015).
Earth's outgoing radiation is following linearly changes in temperature (Koll and Cronin 2018), an effect apparent in Fig. 2 c  265 and d. However, SW radiation also changes as a function of temperature and we found that this change is fairly linear. The resulting feedback was positive, amplifying cooling in the SRM scenarios and amplifying warming in the case of a CO2 increase. The radiative fluxes in Fig. 2 are clear-sky, and this SW feedback is thus caused mainly by ice cover and albedo changes along with changes in atmospheric absorption. The SW feedback was much larger in CESM (all-scenario average of 0.96 W/m2K) than in MPI-ESM (0.50 W/m2K). There was no large difference in surface albedo change between models 270 (supplementary Fig. S1). However, clear-sky SW absorption (net clear-sky SW flux at top of the atmosphere (TOA) -net clearsky SW flux at surface) was linearly dependent on surface temperature by 0.98 W/m 2 K in MPI-ESM and 0.85 W/m 2 K in CESM (supplementary Fig. S2). We attribute this to atmospheric shortwave response of the change in atmospheric water vapor due to the temperature change The differing model response likely originates from the distinct radiation schemes and spectral resolutions in MPI-ESM and CESM. This argument is supported by Fildier and Collins (2015), who likewise derived a larger 275 SW absorption response to temperature in MPI-ESM compared to models that include CAM4.
Overall, we find that the clear-sky aerosol radiative forcings in the two ESMs are in good agreement with ECHAM-HAMMOZ.
However, the same stratospheric sulfur fields yielded 8% weaker (on average) total (SW+LW) clear-sky radiative forcing in MPI-ESM than in CESM. 280 Figure 3 shows Gregory plots for the total TOA all-sky (clouds also taken into account) radiative forcing. In this case, the total SRM radiative forcing was 22% weaker in MPI-ESM than in CESM. On the other hand, the radiative forcing due to increased CO2 concentrations was larger in MPI (orange and red markers in Fig. 3), but the difference was relatively small and is explained by different cloud radiative forcings between models. The impact on SW radiation was larger than it was on LW 285 radiation. The overall result is that the same stratospheric sulfur injection led to larger and faster cooling in CESM than in MPI-ESM (Fig. 3). During the 20-year simulation period, stratospheric sulfur injections of 6 Tg(S)/yr (SRM6) led to slightly over -1 K global mean cooling (left-most green hexagon markers in Fig. 3a) in MPI-ESM but closer to -2 K in CESM (Fig   3b). Global mean warming after 20-year 2xCO2 and 4xCO2 simulations was consistent between the models. However, there was a nearly 2 x larger radiative imbalance in MPI-ESM compared to CESM by the end of the simulations. If these simulations 290 reached radiative equilibrium, the climate would presumably therefore be warmer in MPI-ESM than in CESM.

Temperature-independent fast and temperature-dependent slow precipitation responses
Precipitation responses can be divided into a temperature-independent fast response, which takes place immediately when some forcing agent is introduced, and a slow response caused by the temperature change and subsequent feedbacks (Myhre et al., 2017). Because of climate (e.g., ocean) inertia, precipitation will change slowly along with temperature even in the case of 295 abrupt radiative forcing changes. Here, we separately quantified these fast and slow responses based on the regression method described earlier. Results are shown in Fig. 4. Fast response was obtained by intersection of fitted line and the y-axes (T=0), and slope of the linear fit shows the slow response due to the temperature change. Fast responses are mainly driven by changes in atmospheric absorption (Samset et al., 2016). A change in absorbed radiation modifies the amount of energy transferred between the TOA and surface. This energy transfer is then largely compensated by a change in latent heat flux (evaporation), 300 in turn changing precipitation. Changes in CO2 concentration affect LW atmospheric absorption while SRM primarily modifies SW reflection and is affecting relatively less to how much radiation is absorbed Figure 4 shows that an atmospheric CO2 increase led immediately to a decrease in global mean precipitation. However, this CO2 increase simultaneously warm the climate, which eventually led to a precipitation increase. After 2-5 years, this 305 temperature-dependent slow component exceeded the immediate radiative component, and global mean precipitation was then larger than in the absence of a CO2 increase. On the one hand, stratospheric sulfur aerosols (SRM1-6) also absorb some radiation (Fig. 2b), but on the other hand, relatively more solar radiation is reflected and thus less is absorbed by the background atmosphere. We therefore saw only a small total temperature-independent increase in global mean precipitation for most SRM cases. Overall increasing CO2 decreases precipitation via the fast component and increases precipitation via the slow 310 temperature component (Fig. 5). Fast precipitation impacts were significantly larger for CO2 changes than for SRM (shown in legends in Fig 4.), and therefore the fast precipitation component of SRM was omitted in Figure 5 for clarity.
As Fig. 4a shows, the fast precipitation responses in MPI-ESM differed from those in ECHAM-HAMMOZ, despite the fact that the same atmosphere model was used in both cases. This may result from differing background conditions between the 315 models, land temperature change in ECHAM-HAMMOZ with fixed SST, or noise of in the yearly mean values of MPI-ESM simulations.
Based on the scenarios examined here, the average global precipitation change scales with global mean temperature with a proportionality coefficient of 2.54 (SD 0.27) %/K in MPI-ESM and 2.26 (SD 0.13) %/K in CESM. These values are robust for 320 temperature changes caused by CO2 and SRM forcings. Our results thus support prior findings that the slow precipitation response is not dependent on the forcing agent (Kvalevåg et al., 2013).

Results from simulated scenarios
In the scenario runs (Table 1), the years 2010-2100 were simulated for RCP45 and for geoengineering the RCP4.5 climate via SRM or CDR. Results are discussed below. 325

Change in global mean temperature
Global mean temperature and precipitation anomalies relative to 2010-2020 are shown in Fig. 6. Under RCP45, the global mean temperature increased by 1.30 K and 1.20 K over the 2010-2020 average in MPI-ESM and CESM, respectively. These changes were slightly below the CMIP5 multi-model mean of 1.35 K (Knutti and Sedláček, 2012). During the same period, global mean precipitation increased by 1.76-1.78 % under RCP4.5, also below the CMIP5 multi-model mean (2.66 %). 330 In the SRM-TEMP scenario, the global mean surface temperature was kept close to the present-day value via stratospheric sulfur injections. This reduced global mean precipitation in both ESM simulations (Fig. 6). The reduction was significantly larger in MPI-ESM (-1.42 %) than in CESM (-0.73 %). These differences are explored in Sect. 5. Given the SRM-TEMP results, it is not surprising that when global mean precipitation is maintained at the 2010-level in the SRM-PRECI scenario, 335 the climate warms. SRM-PRECI warming in MPI-ESM is 0.64 K over 2010-2020-level, substantially larger than was seen in CESM (0.27 K). This is consistent with the disparate model results for SRM-TEMP.
Overall, in both models, the majority of the global mean climate warming seen in RCP45 was compensated in SRM-PRECI. The CDR scenario led to a 0.10 (MPI-ESM) and -0.11 (CESM) K change in global mean temperature by the end of the century 345 (2080-2100) compared to present-day (2010-2020). There was thus no significant difference in global mean temperature between the CDR and SRM-TEMP scenarios at the end of the century. The largest difference in global mean temperature between these scenarios was seen immediately after the onset of geoengineering, when the CDR temperature was larger than in SRM-TEMP. Under CDR, the global mean temperature only starts to decrease post-2040. This is because CDR acts slowly atmospheric CO2 to reduce global temperatures whereas similar cooling can be gained with stratospheric sulfur injection much 350 faster (Royal Society 2008). In the CDR scenario, CO2 removal was suspended in year-2070, when atmospheric CO2 concentrations have returned to their 1976 levels. The global mean temperature at that time was close to the present-day value, and did not change significantly through the end of the century (when the rate of change in atmospheric CO2 matches that seen in RCP 4.5). Thus, even this very optimistic CDR scenario is insufficient for cooling the climate to pre-21st century levels.

Change in global mean precipitation
Although the global mean surface temperature in the CDR scenario was the same at the end of the century (2080-2100) as at the beginning of the century (2010-2020), the global mean precipitation was over 0.5 % larger in both ESMs. In Sect. 3.3, we showed that the precipitation impacts of SRM and CO2 can be separated into a temperature-independent fast component and 360 a temperature-dependent slow component. Here we use that framework to examine precipitation responses across the different geoengineering scenarios. Precipitation is also affected by non-CO2 GHGs, tropospheric aerosols, and land-use changes, all of which can induce their own temperature-independent fast components. For our purposes this can be assumed to be the same across all scenarios. Radiative forcings are generally assumed to be additive (Marvel et al., 2015). If we assume based on supplementary Fig. S3 that the overall fast response depends only on absorbed radiation, it follows that the fast responses of individual forcing agents are also additive. In Sect. 3.3 we also showed that the slow temperature-dependent component does not depend on the applied 375 forcing. We can thus describe the global mean precipitation change as the sum of the temperature dependent slow component (a×ΔT) and all fast components (Fläschner et al., 2016): where a, and c are model-specific coefficients, b is a function of the SRM level, T is the simulated global mean surface temperature, CO2 preind is the preindustrial CO2 concentration, ∆CO2 is the atmospheric CO2 change relative to the preindustrial 380 value, and BG is the background fast component, assumed to be the same for all scenarios. Coefficient a is obtained from the scenario-ensemble mean slope in  Fig 4). To calculate coefficient c, we again assume that fast precipitation response is linearly dependent on absorbed radiation. Radiative forcing due to CO2 varies logarithmically with concentration (Etminan et al., 2016) and thus the fast precipitation response for CO2 is also assumed to 385 be logarithmically dependent on CO2 concentrations (see supplementary material Fig. S4) . We calculated the fast precipitation response for three different CO2 concentrations: preindustrial, 2×CO2, and 4×CO2. The coefficient c can then be calculated from a logarithmical fit of the fast response versus CO2 concentration across these three scenarios. This approach yields c values of 4.5 (%) for MPI-ESM and 4.0 (%) for CESM. Finally, we calculated the BG component as the 5-year running mean residual between the first three terms of the right hand side of Eq. (2) and the modelled precipitation (ΔP in Eq.(2)) based on 390 the RCP45 scenario. Note that if Eq. (2) is used only to study precipitation difference between modelled scenarios, the BG component is not needed (see supplementary material Fig. S5). However here we also wish to examine precipitation changes relative to 2010-2020, and the BG term is thus included here. year-2100 the mean differences between the Eq. (2) results and the actual model quantities were ranged from -0.01% to 0.04% 400 for MPI-ESM and from -0.16% to 0.05% for CESM. Fig. S5 in supplementary material shows the precipitation responses under the geoengineering scenarios as anomalies relative to the RCP4.5 case. The plotted precipitation differences in Fig. S5 are thus independent of the BG component. We see from this figure that the individual components can be reliably used to understand the drivers of precipitation change for each scenario. This fast component related to increasing atmospheric CO2 (plus other GHG and absorbing aerosol) probably also explains why the increase of observed global mean precipitation has not increased significantly, despite the fact that climate has warmed (Allan et al., 2014).

410
Under the SRM-TEMP scenario (Fig. 7 c and d), the global temperature change (and thus the slow precipitation component) was small, as is the fast precipitation component due to sulfate aerosols. However, the fast component due to CO2 was as large as in RCP45. This fast (radiative) component from CO2 is the main reason that SRM generally leads to a decrease in global mean precipitation when used to fully offset GHG-induced warming. On the other hand, in the SRM-PRECI scenario (Fig. 7 e and f) the climate was cooled to the point that the temperature-dependent slow component balances the fast radiative 415 components (CO2, SRM and background), so that the net precipitation change was close to zero.
The CDR scenario led to a slight increase in global mean precipitation, despite no significant net change in global mean temperature. Figures 7 g and f show that this was also explained by the fast radiative component of CO2. As in SRM-TEMP end of the century, reducing atmospheric absorption and thus increasing global precipitation compared to 2010-2020.
Although the global mean precipitation response was approximately the same in both ESMs in RCP45 and CDR, a closer look at the underlying drivers shows that only the radiative component of CO2 was consistent across models. The temperaturedependent response differs between ESMs, driving divergent precipitation impacts. This was resulted from slightly different 425 temperature response and hydrological sensitivity between ESMs. In RCP4.5 temperature-dependent slow component was 32% larger in the end of the simulation (2080-2100) with MPI-ESM than in CESM simulation. In CDR magnitude of the slow component was the same between models (0.28% in MPI-ESM and -0.24% in CESM at the end of the simulation), but the sign was different. However, this effect was compensated by differing non-CO2 background responses, which also changed over the course of the simulated century. Figure 7 shows that this BG response is very different between the models and even 430 have a different sign. In MPI-ESM non-CO2 fast components were causing 0.48% decrease on precipitation at the end of the simulation (2080-2100) compared to the beginning (2010-2020) while in CESM non-CO2 forcers were increasing precipitation by 0.23% Thus, it is merely fortuitous that the net precipitation response was similar between models in the CDR and RCP45 scenarios.

435
The BG radiative components impacting precipitation include a range of factors including non-CO2 GHG (methane, nitrous oxide, ozone, CFCs etc), tropospheric and background stratospheric aerosols, and land-use change-with differing treatments between models. Radiative transfer modeling also differs between the ESMs. As shown in Sect. 3.1-3, radiative forcing and (particularly) atmospheric absorption-and thus latent heat flux and precipitation-in the ESMs responded differently to the various forcing agents. Thus, it is not surprising that the BG precipitation component, which is affected by several different 440 forcing agents, also differs between models. Figure 8 shows the amount of sulfur required to keep global temperature or precipitation at current levels through the end of the 21st century. All scenarios started with injections of 1 Tg(S)/yr in year-2020, and the amount of required sulfur then increased along with the RCP4.5-driven warming. In all cases, more sulfur was needed to compensate for RCP4.5 warming 445 than for the associated precipitation increase (see cumulative injection amount on right-hand axes). As shown in Sect. 4.2, the fast, CO2-driven radiative component partly offsets the temperature-driven precipitation component caused by global warming.

Sulfur injections
Thus, in the SRM-PRECI scenario, the sulfur aerosol only needs to compensate for the (already partly offset) precipitation effect of changing temperatures, rather than for the total temperature change (as is the case in SRM-TEMP).
Based on these simulations, in a total of 107-113 Tg(S) and 95-114 Tg(S) was required to prevent a simulated precipitation increase between years 2020 and 2100 in MPI-ESM and CESM, respectively (scenario SRM-PRECI). These 80-year totals are slightly larger than the amount of SO2 emitted each year in the mid 1970s, when annual emissions were roughly 75 Tg(S)/yr . Global sulfur emissions have since decreased; however, China alone emitted over 100 Tg(S) SO2 between 2006(Li et al., 2017. However, the lifetime of aerosols derived from surface emissions is on the order of days, and 455 the cooling impact is therefore much smaller than in the case of stratospheric injection. In the SRM-PRECI scenario, yearly injections are 3 Tg(S)/yr or less, with the exception of occasionally higher injections for one MPI-ESM ensemble member. 3: the all-sky forcing for a given amount of sulfur was significantly (22-33 %) larger in CESM than in MPI-ESM. 465 Figure 8 also shows some limitations of the climate-control algorithm used here. At times the change in SRM injection amount was too large, leading to an overly-strong climate response. In some cases the ensuing compensatory change in injection amount then overshoots the desired climate response in the opposite direction. This led to rapid fluctuations between SRM levels, as seen for example between years 2070-2080 in MPI-ESM ensemble member 2 for the SRM-PRECI scenario. Such 470 effects could be avoided by smaller injection increments, by using a more sophisticated algorithm that could better separate large natural variations in temperature or precipitation from long-term changes or defining geoengineering strategy in advance by using e.g linear response theory (Bódai et al., 2020). We also noted that the introduction of 1 Tg(S)/yr in 2020 led to an overly-large precipitation response for all simulations under scenario SRM-PRECI. However, the above effects do not affect the overall results and conclusions shown here. 475

Regional climate responses
While the SRM-TEMP and CDR scenario simulations led to similar global mean temperatures by the end of 21st century, the regional responses were quite different. Figures 9 a and b map the temperature difference between these two scenarios in both ESMs for the last 20 years of the 21st century. We see that the SRM-TEMP scenario led to cooler tropics and warmer high latitudes than the CDR scenario in both ESMs. These regional discrepancies have been demonstrated in prior studies (Kravitz 480 et al., 2013;Laakso et al., 2017) and point to a fundamental problem with the SRM approach. Aerosols primarily affect incoming SW radiation, while GHGs affect LW thermal radiation, and the meridional gradient is steeper for SW than for LW radiation. Consequently, compensating for a global mean LW change by modifying SW radiation leads to zonally-dependent differences. This issue can be reduced by concentrating the SRM injections in mid and high latitudes, or via seasonal adjustment of the sulfur injection area (Laakso et al., 2017). Overall, however, the temperature differences over land between 485 scenarios were rarely statistically significant (indicated by hatching in Fig. 9, 10% and 20% of land area in MPI-ESM and CESM respectively).
Figures 9 c and d compare the SRM-TEMP scenario to present-day climate (2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019)(2020). In MPI-ESM, the regional SRM-TEMP versus present-day temperature differences were significantly larger than those between SRM-TEMP and CDR at the 490 end of the century. However, this was not the case in the CESM simulations. It should be kept in mind that comparing years 2080-2100 from the SRM-TEMP scenario with 2010-2020 (as present-day) is a somewhat arbitrary choice, and that the comparison reflects not only geoengineering impacts but also climate change under RCP4.5. In addition, even though the global mean temperature was similar between these two periods, the climate was relatively stable in 2080-2100 but was warming in 2010-2020 climate. The regional patterns seen in Fig. 9 c) and d) thus depend to a degree on the choice of reference 495 years, and not only the impacts of geoengineering.
Regional temperature anomalies for other scenarios are provided in the supplement (Fig. S6). Overall, RCP45 led to larger warming at high latitudes than at low latitudes when compared to CDR for years 2080-2100. The corresponding regional patterns in SRM-PRECI were similar to those in RCP45 but with reduced magnitude. Nevertheless, warming in SRM-PRECI 500 relative to the CDR scenario was statistically significant almost everywhere in both models. Figure 10 shows the relative precipitation differences between the SRM-PRECI and CDR scenarios in boreal winter (DJF) and summer (JJA) in 2080-2100. Globally, CDR led to 0.5% more precipitation than SRM-PRECI in both models. However, this precipitation change was not regionally or seasonally homogeneous. A key conclusion is that these changes were rarely 505 statistically significant (hatching in Fig. 10) and that there was often not good agreement between models.
Both models did show broadly similar responses over tropical oceans, especially over the Eastern Pacific and Atlantic. This was probably caused by an Intertropical Convergence Zone (ITCZ) shift due to the zonal temperature difference between SRM-PRECI and CDR (SRM-PRECI led to more warming in high versus low latitudes compared to CDR). Generally, the 510 responses seen in Fig. 10 were larger in MPI-ESM than in CESM, likely due to the significantly warmer climate in MPI-ESM under SRM-PRECI. Supplemental Figures S7 and S8 show that when comparing temperature in SRM-PRECI and CDR, simulations with MPI-ESM led to much greater warming in DJF and (especially) JJA over Europe, Australia and South America when compared to CESM. Figure 10 shows that the corresponding precipitation responses were also significantly different over these areas. Precipitation responses for the other studied scenarios are shown in the supplement (Figs. S9 and 515 S10). As with the results in Fig. 10, spatial features of these precipitation responses were rarely statistically significant. To increase confidence in how SRM and CDR would affect regional precipitation distributions, longer simulations or larger ensemble sizes are necessary.

Discussion and conclusions
Here, we have studied different scenarios in which global mean warming and precipitation changes are compensated by solar 520 radiation management (SRM) or carbon dioxide removal (CDR) during the 21st century. We carried out simulations using two Earth System Models, MPI-ESM and CESM, with SRM based on stratospheric aerosols first simulated with the aerosolclimate model ECHAM-HAMMOZ. SRM was used for two scenarios in which the magnitude of sulfur injections was controlled to maintain global mean temperature or precipitation at year 2010-2020 levels in the RCP4.5 scenario. Additionally, an idealized CDR scenario (also based on RCP4.5) was performed that included 1%/yr removal of atmospheric CO2. We 525 examined the resulting global mean precipitation changes by dividing the response into temperature-dependent and temperature-independent components. These model-specific components were defined based on a regression method using simulations with fixed climate conditions, and that included a constant SRM treatment or an abrupt change in atmospheric CO2 concentrations.

530
Our work supports prior studies in showing that the ratio of the global precipitation change to the global temperature change for SRM is larger than for an atmospheric CO2 perturbation (e.g. Bala et al., 2008). Thus, less sulfur was needed to compensate for the global mean precipitation change under RCP45 than to compensate for the corresponding temperature. Our results  Smith et al., 2011Smith et al., ) http://sedac.ciesin.columbia.edu/data/set/haso2-anthro-sulfur-dioxide-emissions-1850Smith et al., -2005. This simultaneously reduced global mean warming by 50% and 78% based on the MPI-ESM and CESM simulations, respectively (compared to the 2010-2100 RCP4.5 temperature increase in the absence of SRM). 540 While completely preventing global mean warming in this century (in RCP4.5) would require much more sulfur than preventing a change in global precipitation, the total sulfur required was comparable to that emitted globally at the surface from anthropogenic sources during the first five years of the 21st century (274 Tg(S), Smith et al., 2011). However, maintaining a constant global mean temperature in this way led to a significant reduction in global mean precipitation (-1.42% with MPI-545 ESM and -0.73% with CESM) compared to present-day climate. Our component analysis showed that this precipitation decrease was caused by the temperature-independent radiation component resulting from the CO2 increase in the RCP 4.5 scenario. Under RCP45 without SRM, this component was overridden by the temperature-dependent effect on precipitation from global warming. When this temperature component was compensated by SRM, the CO2 component remains and global mean precipitation decreases. It should be noted that this is the case for all SRM methods and not only for stratospheric 550 aerosols. SRM itself had only a small temperature-independent fast effect on precipitation.
In the CDR scenario, the annual CO2 increase based on RCP4.5 was counteracted by a 1% annual removal of the atmospheric CO2 concentration until year 2070. This was found to slow down warming significantly and to return the global mean temperature to its present-day (2010-2020) value. The atmospheric CO2 budget is currently increasing at roughly 4 GtC/yr. In 555 our CDR scenario, 8.7 GtC/yr of CO2 was removed at year-2020. Our scenario should be considered as an idealized high-end CDR scenario as achieving this high CO2 removal rates in a few years would not be feasible due to technological, economic, social, and political constraints. The results highlight the challenge in substantially slowing global warming, and suggest that entirely preventing global-mean warming during this century solely via CDR without significant cuts in CO2 emissions is probably not achievable. 560 Even though global mean temperature at end of the CDR simulation was the same as at the beginning, global mean precipitation increased (~0.5 %) in both ESMs. To date, we have not seen as large increase in global mean precipitation as would be expected only based on the temperature increase (Allan et al., 2014). This is because the fast radiation-driven precipitation effect is largely compensating the slower temperature-dependent component from warming. However, over time, the temperature 565 component will dominate, and a significant increase in global mean precipitation is expected. If atmospheric CO2 is removed as in the CDR scenario, the temperature component is prevented from increasing, but simultaneously a positive fast CO2 precipitation component is induced by the reduction of CO2, increasing global mean precipitation. It is thus difficult to prevent an increase in global mean precipitation via GHG reduction. However, global precipitation changes are also driven by the fast radiative components of aerosols and non-CO2 GHGs, and future precipitation will depend on how these emissions evolve 570 over time.
RCP45 and CDR scenarios led to a similar global mean precipitation response between the two ESMs. However, regression analysis revealed that this was fortuitous. The precipitation response to changing temperature and CO2 concentrations differed between the ESMs, but these differences were masked by offsetting background (BG) effects related to other GHGs and 575 tropospheric aerosols. Large differences in the primary drivers of precipitation change can therefore exist between ESMs even when the ESMs predict similar net changes. A more detailed component analysis, with BG effects separated into relevant subcomponents, is therefore needed. The Precipitation Driver Response Model Intercomparison Project (PDRMIP) may help address this issue (Myhre et al., 2017).

580
Similar component analyses as done here on the global scale (Sect. 4.2) can in principle be performed regionally. However, for regional analyses (e.g., applying Eq. (2) for a single model grid box), the dry static energy flux divergence of the atmosphere needs to be taken into account . This term depends on the neighbouring grid boxes and is not linear or independent from other components. Because of this and natural variability, regression analyses to quantify the fast and slow precipitation components either regionally or for individual grid boxes will be subject to noisier data than in the global 585 case. However, preliminary analyses reveal regions where the approach appears promising, and we therefore recommend further evaluation of this potential in subsequent work.
Overall, this study shows that global mean temperature-independent fast and temperature-dependent slow precipitation responses caused by CDR and SRM can be quantified by the regression method. When these components are known, the 590 global mean precipitation change can be presented as the sum of the temperature-dependent slow component and all fast components. Our results show that the fast responses of CO2 have a major role in the resulting precipitation impacts, when CO2 induced global warming is slowed down by geoengineering. If global warming is prevented by stratospheric sulfur injections while atmospheric CO2 concentration still increases, the global mean precipitation is decreased due to the fast response of increasing atmospheric CO2. On the other hand, less sulfur is required to keep the global mean precipitation stable, 595 because the fast precipitation response to increased CO2 is the opposite of the slow precipitation response resulting from warmer climate. Without SRM, temperature response overruns CO2 fast response (as in RCP45). Also in our CDR scenario, the global mean precipitation increase was explained by the positive fast precipitation response to reduced CO2. As we showed here, separating precipitation to fast and slow response is a useful method to analyse differing precipitation responses between different geoengineering techniques. This framework can thus help to understand and anticipate temperature and precipitation 600 responses in different time scales and geoengineering scenarios, where SRM and CDR are used potentially simultaneously. In principle, this method can also be used to study precipitation response in any scenario, if the temperature change and forcing agents are known.
Code and data availability. The data from the model simulations and implemented model codes are available from the authors 605 upon request.
Author contributions. AL designed the research, performed the experiments, carried out the analysis and prepared manuscript.
All author contributed ideas, interpretation and discussion of results and to writing the paper.