Uncertainties, Limits, and Benefits of Climate Change Mitigation for Soil Moisture Drought in Southwestern North America

Over the last two decades, southwestern North America (SWNA) has been in the grip of one of the most severe droughts of the last 1,200 years, with one third to nearly one half of its severity attributable to climate change. We analyze how the risk of extreme soil moisture droughts in SWNA, analogous to the most severe 21‐year (≥ in magnitude to 2000–2020) and single‐year (≥ in magnitude to 2002) events of the last several decades, changes in projections from Phase 6 of the Coupled Model Intercomparison Project. By the end of the 21st century, SWNA experiences robust (R ≥ 0.80) soil moisture drying and substantial increases in extreme single‐year drought risk that scale strongly with warming, spanning an 8%–26% probability of occurrence across +2–4 K. Notably, our results show that 21‐year droughts analogous to 2000–2020 are up to 5 times more likely than extreme single‐year droughts under all levels of warming (≈50%). These high levels of 21‐year drought risk are largely invariant across scenarios because of large spring precipitation declines in half the models, shifting SWNA into a drier mean state. Despite projections of this sweeping and ostensibly inevitable increase in 21‐year drought risk, climate mitigation reduces their severity by reducing the magnitude of extreme single‐year droughts during these events. Our results emphasize both the importance of preparing SWNA for imminent increases in persistent drought events and constraining projected precipitation uncertainty to better resolve future long‐term drought risk.

; widespread ecosystem disruptions (Anderegg et al., 2013;Fettig et al., 2019;Ganey & Vojta, 2011;Schwalm et al., 2012;Williams et al., 2013); and rapid increases in annual forest fire area (Abatzoglou & Williams, 2016). While such persistent multi-decadal droughts, often termed megadroughts, are common in the paleoclimate record , this ongoing event is unprecedented in the instrumental period and is the second most severe 21-year drought period of the last 1,200 years ( Figure 1; updated soil moisture reconstruction from Williams et al. (2020)).
The 2000-2020 drought was likely triggered by precipitation deficits connected to natural variability and ocean dynamics in the tropical Pacific (Delworth et al., 2015;Lehner et al., 2018;Seager, 2007). However, recent work has attributed one third to nearly one half the severity of the associated soil moisture deficits to climate change through anthropogenically driven increases in evaporative demand . Such a large contribution from warming is consistent with conclusions from other studies demonstrating that climate change has intensified soil moisture, streamflow, and snow droughts in the western US (Berg & Hall, 2017;Marlier et al., 2017;McCabe et al., 2017;Mote et al., 2016Mote et al., , 2018Williams et al., 2015;Woodhouse et al., 2016;Xiao et al., 2018). It is also consistent with climate change projections indicating that future warming will continue to exacerbate drought in the region (Ault, 2020;Cook et al., 2015Cook et al., , 2018Cook et al., , 2019Cook et al., , 2020Overpeck & Udall, 2020;Spinoni et al., 2020;Zhao et al., 2020). Beyond these fundamental expectations of a warmer and drier future, however, there remain some significant uncertainties regarding how climate change will affect drought dynamics. These include how future drying will compare to droughts in the instrumental and paleoclimate record; how various processes, including precipitation and evaporative demand, will ultimately contribute to droughts in the region; how responses of short-(≤1 year) and long-(≥ COOK ET AL.  10 years) term drought events may differ; and the effectiveness of climate change mitigation for ameliorating future drought risk and severity.
We use the latest state-of-the-art climate model simulations from Phase 6 of the Coupled Model Intercomparison Project (CMIP6)  and an update of the Williams et al. (2020) soil moisture reconstruction to investigate how the risk and severity of both short-and long-term soil moisture droughts in SWNA will change with warming at the end of the 21 st century. This updated reconstruction targets the same soil moisture data set as in Williams et al. (2020), with the following exceptions. First, observational values are now available through summer 2020. Second, humidity data in the US were replaced with the latest version from the Parameter-elevation Regressions on Independent Slopes Model (PRISM) climate datasets (https://prism.oregonstate.edu/notices/notice_20191029.php). Third, precipitation and temperature data for locations outside the US were replaced with the latest version of the Climate Research Unit climate grids (https://crudata.uea.ac.uk/cru/data/hrg/cru_ts_4.04/).
We focus on analogs of two recent events to analyze in these simulations: the 21-year drought that began in 2000 and is ongoing as of spring 2021, and the single most extreme drought year during this period, 2002. We evaluate how droughts equal or greater in magnitude to these recent events are likely to change under a range of warming scenarios, focusing on three primary research questions: (a) How does soil moisture drought change with warming in the CMIP6 projections and how do the projections compare to drought variability over the last 1,200 years?; (b) How do the risks for events that are equal or greater in magnitude to 2002 and 2000-2020 change with warming, and what are the driving processes?; and (c) To what extent can climate mitigation reduce the risk of the single-and 21-year drought analogs that we investigate?

Soil Moisture Reconstruction
We assessed drought variability in the instrumental and paleoclimate record using an updated version of the summer season (June-July-August; JJA) 0-200 cm soil moisture reconstruction from Williams et al. (2020). This tree-ring based reconstruction spans 800-2020 C.E. and targets 0-200 cm soil moisture estimates from an empirically constrained model that accounts for inputs from precipitation, losses via evapotranspiration, and carryover storage between seasons. The reconstruction has high skill over SWNA, with average cross validation R 2 > 0.80 after 1400 C.E. and R 2 > 0.70 back to 800 C.E. We use the SWNA regional average (125°-105°W, 30°-45°N) time series from the reconstruction for comparison with the CMIP6 simulations, using the tree-ring reconstructed values for 800-1900 C.E. and direct instrumental values for 1901-2020. Soil moisture units are in standardized z-scores with unit standard deviation (σ), and the regionally averaged reconstruction is calibrated to the observed mean and variance during 1901-1983.
We analyzed drought analogs based on the most recent drought period in this reconstruction through summer of 2020, an extended version of the event that was the focus of the Williams et al. (2020) study (Figure 1). For extreme single-year droughts, we used 2002 as our analog, the single driest year of the last several decades with a soil moisture anomaly of −2.6σ, ranking as the 9 th driest year in the full reconstruction back to 800 C.E. We also analyzed changes in 21-year drought events using the the 21-year mean soil moisture from 2000-2020, which had a magnitude of −0.77σ and was tied with the late 16 th -century megadrought as the second driest 21-year soil moisture anomaly since 800 C.E. While the degree to which such rankings are statistically distinguishable is sensitive to various uncertainties, the reconstruction nevertheless highlights the exceptional nature of recent drought in SWNA.

CMIP6 Multi-Model Ensemble
We used climate model simulations from the CMIP6 database, a set of standardized experiments contributed by climate modeling groups from around the world . We analyzed the "historical" (1850-2014) simulations conducted as part of the core "DECK" experiments  and three forcing scenarios ("Shared Socioeconomic Pathways"; SSPs) for the 21 st century (2015-2100) from Scenar-ioMIP (O'Neill et al., 2016). The historical simulations used estimates of natural (e.g., volcanic eruptions, solar, and orbital variability) and anthropogenic (e.g., greenhouse gas emissions, aerosols, land use change) COOK ET AL.
10.1029/2021EF002014 climate forcings to simulate climate change and variability over the instrumental period, while the SSP projections are based on scenarios of greenhouse gases and other forcings consistent with various assumptions regarding possible trajectories of economic growth, energy use, climate mitigation efforts, and global governance. The three SSPs we employed are taken from Tier 1 of ScenarioMIP and span a range of forcing and warming levels: SSP1-2.6 (+2.6 W m −2 radiative imbalance at the end of the 21 st century; low forcing sustainability pathway), SSP2-4.5 (+4.5 W m −2 ; medium forcing middle-of-the-road pathway), and SSP3-7.0 (+7.0 W m −2 ; medium-to high-end forcing pathway).
We constructed our 18-member multi-model ensemble (MME; summarized in Table 1) by selecting models that provided the mrsol variable (in each soil layer, the mass of water in all phases, including ice; kg m −2 ) for continuous ensemble members from the historical simulations through all three SSP scenarios (1850-2100). We used a single realization from each model, and all models are equally represented within and across the three SSP scenarios. Where available, we also analyzed the following diagnostics within the MME: tas (2-m near surface air temperature; K), pr (precipitation rate, all phases; mm day −1 ), hurs (near-surface relative humidy; percent), and mrro (total runoff, including drainage through the base of the soil model; mm day −1 ). Global warming levels in our ensembles relative to the 1851-1880 near pre-industrial baseline are presented in Figure 2. The three SSPs cover a range of warming levels in the latter part of the 21 st century (2071-2100), with MME median warming of +2.1 K (SSP1-2.6), +2.9 K (SSP2-4.5), and +3.9 K (SSP3-7.0).

CMIP6 Postprocessing and Analyses
To facilitate comparisons with the 0-200 cm JJA tree-ring based soil moisture reconstruction, we post-processed the individual-layer soil moisture in mrsol by interpolating to 200 cm and standardizing the resulting time series to match the mean and variability in the reconstruction. For every month, we integrated each model's soil moisture from the surface through all layers where the bottom layer depth was above 200 cm.
COOK ET AL.  We then added a fractional estimate of soil moisture from the next layer that includes 200 cm based on its relative depth in that layer. Second, we calculated summer season average soil moisture by averaging the 200-cm interpolated soil moisture across JJA, and then calculated a land-area weighted average of the JJA mean over the same spatial domain as the SWNA reconstruction time series (125°W-105°W, 30°N-45°N). Finally, we rescaled the 200 cm JJA soil moisture from each model to match the mean and variability of the SWNA reconstruction over 1851-1950. We also used tas and hurs from CMIP6 to calculate the vapor pressure deficit (VPD) for each grid cell and month individually. We approximated the saturation vapor pressure using: where e s is saturation vapor pressure (kPa), e 0 = 0.6113 kPa, R v = 461 J × K −1 × kg −1 , T 0 = 273.15 K, and T is the temperature (K; using tas). We then calculated actual vapor pressure (e a ) as the product of hurs and e s (after converting % hurs values to fractional values), and the vapor pressure deficit as e s minus e a .

Analyses
The bulk of our analyses are focused on regional average soil moisture time series for SWNA from the CMIP6 models. We acknowledge that this broad region encompasses a diversity of environments and ecosystems, and that such large-scale averaging may mask over important sub-regional heterogeneity in drought responses to climate change. This averaging is necessary, however, to provide consistent time series for comparison with the reconstruction of Williams et al. (2020), which is used to (a) define the drought analogs we are investigating and (b) establish the baseline of historical drought variability against which the CMIP6 projections are evaluated. We nevertheless also show spatial maps of the MME mean response in the CMIP6 models to communicate spatially heterogenous drought responses within SWNA and to put the changes in the region in the context of changes across North America.
To analyze spatially explicit changes in the MME mean, we equally weighted each model and averaged all models and years for two thirty-year periods: 2071-2100 and 1971-2000. We then calculated the difference at each grid cell (2071-2100 minus 1971-2000) and assessed robustness in the ensemble response using the robustness metric R from Knutti and Sedlacek (2013), which incorporates information on the magnitude and sign of the MME change, variability within each simulation, and the spread across models. A higher model spread or smaller signal will decrease R. Values of R will increase if the mean, shape of the COOK ET AL.
10.1029/2021EF002014 5 of 19 distribution, or variability changes between the two time periods. A value of R = 1.0 indicates perfect agreement across models in the ensemble response. For this study, we used a robustness threshold of R ≥ 0.80, which Knutti and Sedlacek (2013) referred to as "good agreement." Values below this threshold, indicating non-robust responses, are shaded by diagonal hatching.
To investigate changes in drought risk, we analyzed the annually resolved SWNA time series and a smoothed version with a 21-year running mean applied in the reconstruction and models. We used the annual time series to investigate changes in the risk of single-year drought events equal to or drier than 2002 in the observed record (−2.6σ) in the CMIP6 projections. Similarly, we used the 21-year smoothed time series to investigate how 21-year drought events with trailing mean values equal to or drier than 2000-2020 (−0.77σ) change with warming. The trailing mean in the models is calculated so that 2071 (the first year of the late 21 st century period we analyze) is the mean of 2051-2071, the value for 2072 is the mean of 2052-2072, and so forth to the final year 2100, which is the mean of 2080-2100. We calculated "risk" as the fraction of years within the reconstruction or MME with values matching or exceeding the magnitude of the 2002 and 2000-2020 droughts from the reconstruction. In the MME, we do this calculation for the historical (1851-2020) and late 21 st century (2071-2100) time periods, pooling together all years and models together during these intervals.

Drought Mechanisms in Climate Model Projections
Projected changes in soil moisture and runoff drought in response to increased greenhouse gas warming depend largely on changes in evapotranspiration and precipitation. Increased evapotranspiration is a major driver of projected drying trends in many regions, occurring in response to strong warming (Knutti & Sedlacek, 2013) and increased evaporative demand (Ault et al., 2016;Cook et al., 2014Cook et al., , 2015Scheff & Frierson, 2013) in all seasons and regions. Uncertainties in actual evapotranspiration changes, and their contribution to surface drying, are largely attributable to how plant water use changes in response to warming and increased carbon dioxide concentrations, with plants acting to either ameliorate (Swann et al., 2016 or amplify (Mankin et al., , 2019 surface water losses. There are also large uncertainties in precipitation within climate model projections, with often large seasonal and spatial differences in the sign, magnitude, and robustness of responses (Knutti & Sedlacek, 2013) and a strong influence of internal climate variability . Total precipitation responses to warming may therefore increase or decrease drought severity and risk, depending on the region and season analyzed, though in many cases even large increases in precipitation are unlikely to completely negate warming-driven increases in evapotranspiration . Beyond changes in total precipitation, warming also increases the fraction of precipitation falling as rain rather than snow (Krasting et al., 2013;, causes earlier melting of the snowpack (Musselman et al., 2021), and reduces total snowpack water storage (Mote et al., 2018;Shi & Wang, 2015). Such shifts can lead to significant seasonal changes in both runoff (Li et al., 2017; and soil moisture (Gergel et al., 2017;Harpold et al., 2015;Marvel et al., in review), even if changes in total precipitation are negligible. More detailed analyses of seasonal changes in important hydrologic variables in the CMIP6 projections across our study region are provided in Marvel et al (in review).

Precipitation and Vapor Pressure Deficit
As in the previous generations of climate models (Knutti & Sedlacek, 2013), the sign, magnitude, and robustness of precipitation responses to warming in our MME is highly variable across regions and seasons (Figure 3), and largely consistent with other CMIP6 analyses Ukkola et al., 2020). The strongest, most robust increases in precipitation occur in northern and eastern North America during winter (December-January-February; DJF) and spring (March-April-May; MAM) and across Mexico in the fall (September-October-November; SON). Robust precipitation declines are not nearly as extensive, limited to the southwestern US in spring; in Mexico in winter, spring, and fall; and in a band across the Pacific Northwest, Northern Plains, and Midwest in the summer. Over SWNA (dashed black box), robust precipitation increases are confined primarily to the northern half of the domain during the winter, along with more modest increases during summer under the strongest warming scenario. Robust declines in precipitation Increasing VPD is the dominant driver of increases in evaporative demand in climate change projections, especially during the warm season. VPD changes are uniform in sign, with robust increases across North America in spring and summer ( Figure 4). These robust increases occur as a consequence of the strong dependency of VPD on temperature, which increases everywhere and is one of the most consistent and robust responses of any climate variable in climate model projections (Knutti & Sedlacek, 2013). Increases in VPD are especially large over SWNA, the Pacific Northwest, and the Central Plains, especially during the summer, which likely reflects the exponential response of saturation vapor pressure to temperature.

Runoff and Soil Moisture
Total runoff decreases across most of western North America in all three scenarios during MAM and JJA, affecting Mexico, SWNA, the Pacific Northwest, and the northern Plains ( Figure 5). During MAM in SWNA, runoff declines are strongest and most robust in the southern part of the domain (southern California, Arizona, New Mexico), collocated with the most robust precipitation declines. MAM runoff also declines in more northern parts of SWNA in both seasons, a region with large increases in precipitation during previous months (DJF) and non-robust changes or slight increases in precipitation during MAM. For this part of the SWNA domain, this decline likely reflects increased partitioning of precipitation to rain rather than snow and earlier melting of the snowpack, leading to earlier and lower total runoff during the warm season months. One location in southern Idaho shows strong and robust increases in total runoff during JJA, though it is not clear what, specifically, is driving this hyper-local response. Robust soil moisture drying is much less extensive during spring ( Figure 6, top panels), confined primarily to the Pacific coast, Arizona, New Mexico, Texas, and Mexico. Reduced soil drying in spring relative to the warm season is likely a consequence of less snow but increased total precipitation during the cold season, which allows this moisture to infiltrate into the soil and carry over into the spring. Declines in soil moisture are much more intense and widespread in JJA ( Figure 6, bottom panels), especially over SWNA. Precipitation changes during this COOK ET AL.  . This MME includes all models in Table 1 except BCC-CSM2-MR, which did not provide hurs (surface relative humidity) from the required ensemble members. Hatching indicates regions with non-robust responses (R < 0.80).  season are non-robust, though negative precipitation anomalies in MAM likely contribute along with the strong increases in evaporative demand in both of these seasons.
As outlined in the introduction, the expression of the 2002 and 2000-2020 drought in our observational soil moisture index is clearly reflected in other drought metrics, including water resources, vegetation health, and wildfire. Recent studies, however, have highlighted the complexity of vegetation-drought interactions, which can influence how drought is expressed across the hydrologic cycle, especially in climate model projections (Lemordant et al., 2018;Mankin et al., 2017Mankin et al., , 2018Mankin et al., , 2019Swann et al., 2016;Swann, 2018). Plant water-use responses to climate and atmospheric carbon dioxide concentrations could theoretically ameliorate surface drying projected by aridity measures (Swann et al., 2016), or even repartition moisture flows to reduce drought risk (Betts et al., 2007). At the same time, however, there is evidence that plant responses to CO 2 do not account for the full gap between aridity measures and prognostic model quantities like soil moisture (Scheff et al., 2021), even as SWNA appears to be one region where the aridity measures and prognostic variables agree (Mankin et al., 2019;Scheff et al., 2021). It is therefore conceivable that, unlike extreme droughts during the 20 th century, soil moisture responses in the future may not be indicative of overall declines in surface water availability. To assess whether our soil moisture index adequately represents declines in overall surface water availability in the projections, we compare changes in runoff ( Figure 5) and soil moisture ( Figure 6) across SWNA. During JJA, robust declines in both quantities occur across most of SWNA in the projections from our CMIP6 MME. This confirms that changes in the soil moisture index that we use to define extreme drought are consistent with overall declines in the surface water resources available to supply societal and ecological demands.

Soil Moisture Trends and Extreme Drought
Soil moisture time series from the MME show that the drying evident at the end of the 21 st century (Figure 6) is part of a long-term trend that begins in the ensemble in the latter part of the twentieth century (Figure 7a). Trend comparisons between the soil moisture reconstruction and individual models in our MME for three different intervals are shown in Figure 7b. The observed drying (negative) trend from 1901-2020 is well captured by the MME, matching closely to the ensemble median trend. Median ensemble drying is larger than observed for the latter half of the twentieth century , though the observed trend still falls within the range simulated by the models. The observed drying from 1950-2020 is relatively weak because SWNA experienced a major drought in the mid-twentieth century at the beginning of this trend estimate (Heim, 2017;Seager et al., 2005). By contrast, no model in the ensemble has an equivalent or larger drying trend than observed for the most recent period . The decades preceding the 2000-2020 drought were some of the wettest in the reconstruction , and only two other 41-year periods had larger observed drying trends, both during this interval (1978-2018 and 1979-2019).
To investigate if the models can even produce such large magnitude short-term drying trends, we calculated all possible 41-year trends in the ensemble for the available instrumental period (1851-2020). Across the ensemble, 41-year trends of equal or greater magnitude to the observed trend from 1980-2020 occur five times. More generally, the models produce large magnitude trends with a similar frequency to the reconstruction; for example, the fifth percentile trends from the observations and models are −0.03 and −0.04 σ per year, respectively. The models in the ensemble are therefore capable of reproducing extreme short-term drying trends similar to what was observed in recent decades. The mismatch between models and observations for 1980-2020 is thus likely a consequence of how rare such severe drying trends are in both observations and models, and the limited sampling in the model ensemble during this period.
We now evaluate the single-and 21-year drought analogs for the latter part of the 21 st century (2071-2100) in our ensemble. In all three SSPs, the distributions of single-and 21-year mean soil moisture anomalies shift towards drier conditions (Figures 8a and 8b) with substantially increased risks of 2002 and 2000-2020 drought analogs (Figures 8c and 8d). For single-year soil moisture, model distributions are unimodal, and the historical simulated distribution matches closely with the reconstruction: conditions equal or drier than 2002 occur in 0.74% of years in our reconstruction and in 0.88% of all years pooled from the instrumental period (1851-2020) from our model ensemble (n = 3,060). The risk of these events increases substantially in the SSPs and scales strongly with the forcing. Across all possible years from 2071-2100 pooled from our ensemble (n = 540), the frequency of these events increases to 8.3% of all years in SSP1-2.6, 14.4% in SSP2-4.5, and 25.9% in SSP3-7.0.
Over the instrumental period, the ensemble simulated 21-year mean soil moisture distributions have an extended dry tail compared to the reconstruction and a higher prevalence of extreme 21-year drought periods. Indeed, 2000-2020 drought analogs, while still exceedingly rare, are much more common in the model historical simulations (4.9% of all years) compared to the reconstruction (0.25% of all years). The 21-year mean soil moisture distributions from the SSPs have a pronounced bimodality that is absent in the single-year soil moisture and a much higher risk for 2000-2020 drought analogs: 47.0% of all years in SSP1-2.6 have a 21year mean that matches or exceeds 2000-2020, with only slightly higher frequencies in SSP2-4.5 (53.3%) and SSP3-7.0 (53.5%). Only six of the 18 models do not simulate any 2000-2020 drought analogs in any of the SSPs: ACCESS-CM2, GISS-E2-1-G, IPSL-CM6A-LR, KACE-1-0-G, MPI-ESM1-2-HR, and UKESM1-0-LL.

Drought Responses Across the MME
To investigate the bimodality in the 21-year mean soil moisture response, and the apparent lack of sensitivity of 21-year drought risk to the magnitude of warming, we examined single-year ( Figure 9a . Solid colored lines are the MME median in each shared socioeconomic pathway (SSP), while the shading indicates the interquartile range across models in the MME. (b) SWNA soil moisture trends for the reconstruction (red dots) and individual models from the MME (black dots) for three historical intervals. Historical simulations in the models end in 2014; for these trend estimates, we therefore extended the time series through 2020 using output from the SSP2-4.5 scenario.
(BCC-CSM2-MR, CanESM5, CESM2, CESM2-WACCM, CNRM-CM6-1, CNRM-ESM2-1, EC-Earth3-Veg, MRI-ESM2-0, NorESM2-MM), the median in all three SSPs exceeds the 2000-2020 threshold (Figure 9b, vertical dashed line). We categorize this half of the ensemble as "high-risk," and it is these models that create the drier peak in the full MME 21-year mean soil moisture distributions (Figure 8b). The remaining nine "low-risk" models (ACCESS-ESM1-5, ACCESS-CM2, GISS-E2-1-G, IPSL-CM6A-LR, KACE-1-0-G, MI-ROC6, MPI-ESM1-2-LR, MPI-ESM1-2-HR, UKESM1-0-LL) have more modest drying or even increased soil moisture, creating the secondary, less dry peak in Figure 8b. This low-risk ensemble includes all six models that do not simulate any 2000-2020 drought analogs in the projections. Because median values from the low-risk models do not cross the 2000-2020 threshold in most SSPs, it is the nine high-risk models mainly driving the increase in 21-year drought risk. Furthermore, because the differences between the high and low-risk models are larger than the differences across SSPs within models, the forcing scenario has little influence on 21-year drought risk in the full MME.
To understand why the response is so different between the high and low-risk models, we compared the climate sensitivities between the model groups and the regional precipitation and temperature responses. Median climate sensitivity in the high-risk models is 4.8 K/2 × CO 2 , moderately higher than for the low-risk models, which have a median climate sensitivity of 3.9 K/2 × CO 2 . The range of climate sensitivities within the two model groups is comparable across the two, however: 2.5-5.6 K/2 × CO 2 in the high-risk and 2.6-5.3 K/2 × CO 2 for the low-risk. Additionally, we observe that even models with nearly the same sensitivities across the two groups can have dramatically different soil moisture responses over the SWNA. Contrast, for example, the strong drying in BCC-CSM2-MR with the much more muted response in GISS-E2-1-G COOK ET AL. ( Figure 9), despite the fact that both models have approximate climate sensitivities of 3.0 K/2 × CO 2 . It therefore seems unlikely that climate sensitivity specifically can explain the differences across these two sets of models.
Declines in spring precipitation are larger and more robust in the ensemble mean of the high-risk models compared to the low-risk models, especially across the southern half of SWNA (Figure 10). At the same time, the low-risk models show a modest increase in precipitation during the summer, a signal that is absent in the high-risk models. Together, these differential precipitation responses would amplify soil moisture drying in the high-risk models and reduce it in the low-risk models, likely accounting for the majority of the difference in soil moisture drought between these two model groups. Warming during spring and summer is also slightly higher in SWNA over Arizona and New Mexico in the high-risk models (not shown), the same area with the amplified declines in spring precipitation. While this slightly enhanced warming could amplify soil moisture losses by increasing evaporative demand, this model response could instead be a consequence of the enhanced soil moisture drying in the high-risk models (e.g., an increased Bowen ratio).
COOK ET AL.

10.1029/2021EF002014
13 of 19 Even if warming mitigation offers little capacity for reducing the occurrence of 2000-2020 drought analogs, there may still be mitigation benefits from reducing the intensity or severity of these events. To test this, we sampled single-year soil moisture from all 2000-2020 drought analogs across our full 18-member MME. Consistent with our previous analyses (Figure 8a), distributions of single-year soil moisture during these events are unimodal and shift progressively drier with increased warming (Figure 11a). Additionally, the risk of 2002 single-year drought analogs during these 21-year events scales strongly with warming (Figure 11b): 18.1% (SSP1-2.6), 24.3% (SSP2-4.5), and 37.2% (SSP3-7.0). This demonstrates that, while mitigation may have little effect on the occurrence of droughts at least as severe as the 2000-2020 analog, it would still have large benefits for reducing the overall severity of these multi-decadal droughts and the risk of the most extreme single-year droughts during these events.

Discussion & Conclusions
As in the previous generation of climate models Cook et al., 2018), the latest projections from CMIP6 demonstrate that future warming will likely increase soil moisture drought risk across SWNA to levels far outside the envelope of natural variability from the last millennium. Here, we analyzed model-projected responses to anthropogenic climate change of both single-and 21-year soil moisture droughts, focusing on analogs of recent events (2002 and 2000-2020)   events scales strongly with the magnitude of warming, even the most conservative warming scenario (SSP1-2.6, +2.1 K) results in an approximate 50% likelihood of a 21-year drought at least as severe as 2000-2020 by the end of the 21 st century. This strong response of long-term droughts in the models is a consequence of large spring season precipitation declines in half the models in our ensemble, where the median soil moisture response exceeds the dryness of 2000-2020 in all warming scenarios. However, though warming mitigation may not be effective at reducing the likelihood of occurrence of droughts similar to 2000-2020, there are clearly strong benefits for alleviating the overall magnitude of drying, the intensity of these droughts, and the risk of extreme single-year droughts similar to 2002.
Our results from the latest generation of climate models in CMIP6 are broadly similar to earlier studies (Ault et al., 2016;Cook et al., 2015;Overpeck & Udall, 2020;Spinoni et al., 2020;Zhao et al., 2020), which also found large projected declines in warm season soil moisture and runoff over this region of North America in the twenty-first century. In the one study that specifically investigated climate mitigation effects on multi-decadal droughts, Ault et al. (2016) found that the risk of these events was strongly reduced with climate mitigation, an apparently contrary result to our own findings. However, Ault et al. (2016) focused on events that were more persistent (35-year) and modest in severity (−0.5σ) than the shorter and more severe 2000-2020 event we used as an analog for our analyses. Ault et al. (2016) also found that the mitigation benefits were primarily a consequence of differences in the magnitude of warming across the SSPs, while in our CMIP6 ensemble 21-year drought risk was most strongly related to precipitation differences across models and not temperature or precipitation differences across scenarios. Divergent results across these two studies are therefore likely due, at least in part, to differences in both how drought was defined and climate in the models and scenarios within the different ensembles.
Our analyses clearly highlight the importance of precipitation uncertainties in climate change projections of drought. As in the last generation of climate models (Knutti & Sedlacek, 2013), precipitation in the CMIP6 projections is one of the most uncertain and least robust responses in the models Ukkola et al., 2020), especially in mid-latitude regions. This is clear in our ensemble, where much of the drying and especially the increase in 21-year drought risk is forced by larger precipitation declines in the models comprising the high-risk half of our ensemble. At the same time, drought risk is ameliorated in the low-risk models because these show robust precipitation increases during the summer. Although large and robust spring precipitation declines for SWNA are expected (Ting et al., 2018), improving confidence in drought projections in this, and other regions, would still strongly benefit from better constraining the magnitude of these precipitation responses. While not detailed here, there are also major uncertainties for surface COOK ET AL. drought connected to treatments of the land surface and vegetation processes within the models, which can act to ameliorate (Swann, 2018) or amplify (Mankin et al., 2019) soil moisture and runoff droughts. Within our ensemble, projected soil moisture declines during JJA are also associated with reductions in total runoff over SWNA, indicating that, in aggregate, vegetation processes are not sufficient to counteract surface drying forced by changes in precipitation and evaporative demand.
As of spring 2021, the ongoing persistent drought that began in 2000 is continuing to have large impacts across SWNA (Freedman & Dormido, 2021;National Drought Mitigation Center, 2021), a herald for what may be a much drier future in the region with continued warming. As a consequence of its severity and persistence, this drought has motivated substantial action among managers, regional governments, stakeholders, and private investors across the western US to more aggressively and proactively address water resource issues in the face of declining availability and increased demand (Howe, 2021). Results from our analyses demonstrate the clear value of warming mitigation for reducing drought severity and risk in SWNA. Under even the lowest warming scenario, however, increases in drought severity and risk are likely by the end of the 21 st century. Building a resilient future in the region will therefore require additional and ongoing adaptation measures and policies to manage water resources in a world that will be warmer and drier compared to the past.
Acknowledgments BIC, APW, and KDM supported by NOAA MAPP NA19OAR4310278. BIC and APW also supported by NASA's Modeling, Analysis, and Prediction program (MAP-16-0081). APW also supported by NSF P2C2 AGS-1703029. JSM and JES supported by NOAA MAPP NA20OAR4310414, and JSM is supported by Burke Award. JES was additionally supported in part by the US National Science Foundation under grant AGS-1805490.