The PMIP4 simulated dryland aridity changes during the Last Interglacial

The dryland is one of the most important land ecosystems on the planet, and its changes are closely bound up with one-third of the world’s population. Investigating dryland climate change during the Last Interglacial (LIG; ∼127 ka ago) can advance our knowledge of dryland climate behaviors in an orbitally-induced warmer-than-present scenario. Based on the multiple model outputs from the Paleoclimate Modelling Intercomparison Project Phase 4, we suggest that the dryland areas during the LIG are 37% wetter than that during the preindustrial period as measured by the aridity index (AI), and 37% of the preindustrial drylands correspondingly convert to wetter subtypes. Spatially, there are hemispheric differences with drylands wetting and contracting in the Northern Hemisphere but desiccating and expanding in the Southern Hemisphere. Further diagnosis indicates that the altered precipitation is the dominant contributor to more than 72% of the AI changes, and the precipitation change is mainly attributed to the orbitally-induced redistribution of incoming insolation and heat. The secondary factor is the relative humidity change, exaggerating the AI changes in the same direction as the precipitation does. The simulation agrees reasonably with reconstructions for most regions, except Australia and southern Africa. The simulated changes in dryland aridity and the mechanism differ from that in future warming scenarios, so we claim that the LIG as a potential analogue for a warmer future does not fully hold for the dryland climate.


Introduction
Drylands, which are characterized by prolonged water deficits, cover approximately 46% of the Earth's land surface, and the productivity therein supports over a third of the world's population (Olsson et al 2019).The long-term scarcities of water make drylands prone to frequently occurring droughts, dust storms, and heat waves, constraining economic development and agricultural productivity (Robert and Thomas 2012, Gaur and Squires 2018).Whether an area is dry or not depends not only on precipitation but also on energy and aerodynamic factors (Middleton and Thomas 1997).Precipitation supplies water to the land, while energy and aerodynamic factors facilitate evapotranspiration.Aridification has been observed and simulated over the last half-century in multiple indices, including the vapor pressure deficit, the dimensionless aridity index (AI), and soil moisture, and is projected to continue in the near future (Feng and Fu 2013, Lin et al 2015, Huang et al 2016, Cook et al 2020, Deng et al 2020, Lian et al 2021).This aridification is possibly due to the increased potential evapotranspiration (PET) resulting from warming (Huang et al 2016, Deng et al 2020).Considering the vulnerability of dryland ecosystems (Moreno-Jiménez et al 2019, Berdugo et al 2020), it is vital to adequately project the aridity change.However, future climate states likely lie outside the range of the current instrumental period, and only using instrumental data is insufficient to understand the physical laws of aridity-humidity fluctuations.Therefore, paleoclimatology can be adopted to extend the research scale by providing varying climate scenarios that occurred in the past.In addition, it is still uncertain whether alterations in the AI can effectively modify environmental factors, such as vegetation, runoff, and soil moisture (Greve et al 2019, McColl et al 2022, Scheff et al 2022).Paleoclimate data present a chance to assess the impact of AI changes on shaping the landscape (Scheff et al 2017).
Early research has done much regarding the past 21 ka, because it includes abundant proxies, and the climate in this period has been extensively simulated (e.g.Scheff et al 2017, Liu et al 2021).In particular, the last glacial maximum (∼21 ka ago) and mid-Holocene (∼6 ka ago) intervals are widely investigated based on multiple models from Paleoclimate Modelling Intercomparison Project Phase 3 (PMIP3) and reconstructions (Scheff et al 2017, Liu et al 2018, 2019, Xu et al 2020).It is shown that, compared to the preindustrial period, the globally-averaged AI is nearly unchanged during the last glacial maximum and slightly wetter during the mid-Holocene, and the simulations match with the proxies in most regions; the altered precipitation and temperature are both responsible for the last glacial maximum aridity change (Liu et al 2018, Xu et al 2020), while the precipitation variation determines the mid-Holocene state (Liu et al 2019).
Beyond the aforementioned periods, the Last Interglacial (LIG, ∼129-116 ka ago), the warmest period during the past 800 ka, has drawn much attention due to its comparable warming to projected temperatures by the end of the 21st century under low emission scenarios (Gulev et al 2021).This provides an opportunity to understand the dryland climates in a warmer-than-present world.Regarding the hydrology during the LIG, earlier studies have focused on precipitation (Scussolini et al 2019, He and Zhou 2020, Williams et al 2020), river discharge, and floods (Scussolini et al 2020).However, little attention has been paid to the alteration of aridity over drylands, which is not solely dependent on precipitation but also on energy and aerodynamic factors (Middleton and Thomas 1997).Now the LIG is included in the PMIP4 for the first time (Otto-Bliesner et al 2021), enabling research to use multiple models that perform LIG simulations under a common experimental protocol (Otto-Bliesner et al 2017).In this study, aridity changes over drylands are examined using PMIP4 models based on AI, soil water, runoff, and the leaf area index (LAI).We aim to address the following issues: (1) how and to what extent does the dryland aridity change on global and regional scales?
(2) What are the main causes of the aridity changes?
(3) Are the simulations consistent with proxies?

Model and observation data
The analysis is based on 12 climate models within the PMIP4 framework (table S1), which conduct both LIG and preindustrial experiments and include all variables required in this study.The climatological differences between the LIG and preindustrial periods are analyzed.The last 50 years of simulations are taken to represent their quasi-equilibrium climate states.The LIG experiment sets forcings and boundary conditions at the early part of the LIG (∼127 ka BP) and differs from the preindustrial control experiment mainly in orbital parameters (table S2).Compared to the preindustrial period, the orbit at the LIG is characterized by larger eccentricity, higher obliquity, and perihelion closer to the boreal summer solstice, leading to higher top-of-atmosphere insolation during April to September (figure S1) (Otto-Bliesner et al 2021).The atmospheric greenhouse gas (GHG) concentrations assigned to the LIG experiment are slightly changed compared to the preindustrial control experiment, resulting in negligible climate change compared to that induced by altered orbital configurations.We use the median to measure the center tendency of multiple models, as outliers hardly influence this estimator in a small collection of models.
Observations and reanalysis data (hereafter both types are referred to as observations) during the period 1961-1990 are used to assess the model's ability to simulate the spatial distribution and aridity of modern drylands.The precipitation data are taken from the Global Precipitation Climatology Centre (GPCC) dataset that is based on gauge observations (Schneider et al 2017), and PET data are taken from the fourth version of the Climatic Research Unit gridded Time Series (CRU TS v4) (Harris et al 2020).

Definition of the AI
Dryland aridity is primarily measured by the AI, which is defined by the ratio of annual mean precipitation to PET (Middleton and Thomas 1997).A larger AI value derived from higher precipitation or lower PET corresponds to a wetter climate.PET measures evapotranspiration capacity and is computed using the Penman-Monteith equation (supplementary text S1), which summarizes the impacts of 2 m air temperatures, relative humidity, and wind speeds, as well as surface available energy (the difference between the surface net radiation and soil heat flux density) (Allen et al 1998).Temperatures and available energy provide evaporative energy, and wind and relative humidity act to sustain evaporation by moving local water vapor.Therefore, higher temperatures, available energy, or wind speeds, or lower relative humidity, corresponds to higher PET.
Drylands are identified as regions with AI less than 0.65 and can be further subdivided into hyperarid (AI < 0.05), arid (0.05 ⩽ AI < 0.20), semiarid (0.20 ⩽ AI < 0.50), and dry subhumid (0.50 ⩽ AI < 0.65) zones (Middleton and Thomas 1997).Relevant hydrological variables, including total runoff and deep-layer soil water, as well as the ecological variable LAI, are used as comparisons to show the applicability of AI in capturing aridity changes (Greve et al 2019, Cook et al 2020, Scheff et al 2021).Here, total runoff is calculated as the difference between precipitation and evapotranspiration, rather than using model runoff outputs, in order to avoid inconsistency in runoff definitions across various models (Scheff et al 2021).Deep-layer soil water is obtained by integrating the 'mrsol' variable from the surface to a depth of 2 m.Due to the limited availability of variables, soil water and LAI are derived from 5 and 8 models, respectively (table S1).
The causes of AI changes are analyzed by isolating contributions of individual variables from the AI and PET functions (Feng andFu 2013, Fu andFeng 2014), and the formulas to decompose AI and PET anomalies are provided in the supplementary material (text S1).In the following analysis, the regional average of the AI is calculated as the ratio between regionallyaveraged precipitation and PET.Percentage changes in regional-mean variables are computed by dividing regional-mean changes by preindustrial regional means.

Model performance in the modern era
The models' performance in simulating dryland characteristics is evaluated based on observations from 1961 to 1990.Observed drylands are widely distributed in the subtropics of both hemispheres and inland Asia (figure 1(a)), with hyperarid zones mainly found in northern Africa and the Arabian Peninsula, and arid and semiarid regions widely distributed over each dryland.Such a distribution agrees with the previous documentation (Prȃvȃlie 2016) and matches the surface vegetation types (Feng and Fu 2013).The multimodel median reasonably simulates this large-scale structure but overestimates the AI over Australia, inland Asia, and North America (figure 1(b)).This overestimation is partly because the preindustrial period has significant decreases in atmospheric GHG concentrations compared to the period 1961-1990, which could lead to biases in the AI by altering near-surface temperatures and precipitation (Ren et al 2013, Bonfils et al 2020).
Further, the consistency between the simulated and observed annual mean fields is quantified based on spatial correlation coefficient, normalized standard deviation, and normalized centered root-meansquare difference.If the former two statistics are closer to 1, namely the last statistic is closer to 0, the simulation better approximates the observation.According to the Taylor diagram (Taylor 2001) that graphically summarizes the three statistics (figure 1(b)), the models are able to capture the largescale feature of the dryland climatology.However, they still suffer from biases in magnitudes and spatial variances when compared to observations.These errors, especially in precipitation, are common problems in current climate models, and the underlying causes remain unresolved.In comparison, the multimodel medians outperform most models and are thus used as the best model in the following analysis.

Dryland aridity changes during the LIG
The dryland aridity change is examined with the dryland spatial extent fixed at the preindustrial level.The dryland spatial distribution in each model is presented in figure S2.Compared to the preindustrial period, global drylands are 37% wetter during the LIG as represented by the increased AI in the multimodel median (figure 2(a)).Spatially, there is an interhemispheric contrast.Northern Hemisphere drylands experience consistent wetting with an average AI increase of 81% relative to the preindustrial period, dominating the global mean change.The strongest wetting occurs over southern boundaries of northern Africa drylands.Southern Hemisphere drylands, however, are uniformly desiccated, with an average AI 17% lower than in the preindustrial period.
Toward a combined assessment of the aridity change, responses of soil water, runoff, and the LAI, are also evaluated.The results reveal a strong correspondence between the AI and hydrological metrics.Both total runoff (figure 2(b)) and deeplayer soil water (figure 2(c)) demonstrate contrasting changes between hemispheres and exhibit the strongest wetting along the southern boundaries of the northern Africa dryland, spatially matching the AI change.Quantitatively, the spatial correlation coefficients between AI and runoff changes, as well as soil water changes, are 0.91 and 0.79, respectively, statistically significant at the 99.9% confidence level.This consistency across the AI, soil water, and runoff holds for most models (figure S3).Regarding changes in the LAI (figure 2(d)), the spatial correlation coefficient versus AI changes is 0.75, significant at the 99.9% confidence level.This indicates a general agreement between the ecological metric and the AI on a large scale.Regionally, however, LAI-AI discrepancies arise in inland Asia.This is possible because the LAI largely reflects vegetation conditions in growing seasons, particularly summer in inland Asia when the AI decreases (figure S4).
In line with simulated aridity changes, there are spatial variations in simulated dryland distributions.The increased aridity leads to dryland expansions, and vice versa for the decreased aridity.As shown in the multimodel median field (figure 2(e)), from the preindustrial period to the LIG, the global dryland area reduces by 1% of the earth's land surface; 37% of preindustrial drylands convert to wetter subtypes, which is 2.5 times the magnitude of zones replaced by

Causes of AI changes
The mechanisms behind AI changes can first be elucidated by decomposing AI changes into contributions of precipitation and PET changes (supplementary text S1).As shown in figures 3(a) and (b), alterations in both precipitation and PET facilitate the AI change in the same direction.Increased precipitation and decreased PET cause wetting over Northern Hemisphere drylands, while decreased precipitation and increased PET favor drying over Southern Hemisphere drylands.Furthermore, PET changes can be decomposed into contributions of near-surface air temperatures, available energy, relative humidity, and wind speeds (figures 3(c)-(f)).Changes in relative humidity are the main reasons for PET variations for both hemispheres (figure 3(e)).Increased relative humidity causes reduced PET and subsequently increased AI over the Northern Hemisphere drylands, and vice versa for the Southern Hemisphere.Such changes in near-surface relative humidity might be linked to altered atmospheric water contents due to precipitation changes, so relative humidity alterations can be regarded as an amplifier to enlarge the influence of precipitation.
The AI changes are specifically examined over seven subregions to quantify the regional responses, and the results of individual models are presented to show the inter-model consistency (figure 4).The regions include four wetter (North America, northern Africa-Arabian Peninsula, South Asia, and inland Asia) and three dryer regions (South America, southern Africa, and Australia).Specific dryland coverage for each model is presented in figure S5.The multimodel medians show that the strongest wetting occurs in northern Africa-Arabian Peninsula and South Asia, where the AI rises by 133% and 77%, respectively.Comparatively smaller increases appear in inland Asia and North America, with values of 6% and 17%, respectively.In South America, southern Africa, and Australian drylands, the AI shows decreases of 21%, 14%, and 20%, respectively.
The precipitation change as a primary contributor accounts for 82% of the AI change for the globaldryland mean, 72%-88% for the wetter regions, and 79%-82% for the dryer regions; the relative humidity change as a secondary role is responsible for 16%, 13%-17%, and 16% of the AI changes for global, wetter, and drier drylands, respectively.This leading role of precipitation and secondary role of relative humidity are common characteristics for most models in each region.In addition, the inter-model uncertainty is quantified by the ratio of the inter-model spread,  S1).Hearts represents the overall changes in the AI.The values are normalized by dividing preindustrial regional means of the AI and thus represent percentage changes in the AI.The calculation is based solely on dryland grids in each region, and the specific dryland coverage for each model is presented in figure S5.
as measured by the interquartile range, to the absolute value of the multimodel median.The ratios are greater than 1.0 for the AI change over inland Asia and southern Africa, implying large uncertainties.For other cases, the ratios are below 0.8, indicating high agreement across models not only in the sign but also in the magnitude.
Considering the significance of precipitation changes, we examine the mechanisms for different regions, with emphasis on the primary season of occurrence.For most regions, the orbitally-induced change in atmospheric circulations plays the dominant role in altered precipitation.Over northern Africa-Arabian Peninsula and South Asia, summer as a primary contributor accounts for 72% and 79% of the annual precipitation increase, respectively (figure 5).In boreal summer, Northern Hemisphere high latitudes experience the strongest increase in incoming insolation (figure S1), resulting in intensified land-ocean and inter-hemispheric thermal contrasts (figure 6(a)).This causes the northern Africa-South Asia monsoons to strengthen and expand (figure 6(c)), leading to increased precipitation over drylands (Scussolini et al 2019, Yeung et al 2021, Chen et al 2022).The increased precipitation over northern Africa could be further strengthened by two additional processes.First, the precipitation rise heats the atmosphere, driving a low-level cyclone to its northwestern side due to the excitation of westward propagating Rossby waves (Gill 1980, Wang et al 2014).The southwesterly winds situated south of the cyclone in turn strengthen the North Africa monsoon and thus precipitation.Second, the northward surface winds in the tropical Atlantic, along with the meridional contrast in sea surface temperatures (SSTs) around 10 • N, may be reinforced by a wind-evaporation feedback (Zhao and Harrison 2012), potentially amplifying the orbitally-induced increase in precipitation over northern Africa.Over southern-hemispheric drylands, summer is also the dominant season responsible for 46%-85% of the annual precipitation deficit (figure 5).This season experiences decreased insolation, particularly in Southern Hemisphere middle to high latitudes, which causes reduced land-sea temperature contrasts (figure 6(b)), weakened continental thermal  lows (figure 6(d)), and ultimately decreased precipitation (Scussolini et al 2019, Yeung et al 2021).Over North American drylands, 48% of the precipitation increase occurs during boreal winter (figure 5).In this season, the reduced insolation amplifies the land-sea thermal contrast in the Northern Hemisphere.This causes an anomalous cyclone over the North Pacific (figure 6(d)), which favors anomalous southerly synoptic winds and accordingly could cause more precipitation across southwestern North America.Over inland Asia, 86% of the annual precipitation increase appears in autumn (figure 5) when 850 hPa westerlies are strengthened (figure S6(a)).The strengthened westerly winds cannot be attributed to attenuated low-level atmospheric baroclinicity, as indicated by weakened meridional temperature gradients (figure S6(b)).A comprehensive understanding of underlying causes may be achieved by analyzing the energy budget.Finally, the moisture budget analysis (text S2 and figure S7) provides evidence supporting the dominant role of modified atmospheric circulations in precipitation changes over drylands, except for southwestern North America and inland Asia, where the nonlinear component associated with synoptic eddy contributions on submonthly time scales is also important.

Model-data comparison
A compilation of existing terrestrial moisture records at ∼127 ka BP has been presented by Scussolini et al (2019).The proxies are mostly based on pollen, lacustrine or marine sediment composition, speleothems, and multiple proxies.In the Northern Hemisphere drylands, most records suggest a wetter-than-present climate except for a few sites (figure 2(a)).Particularly from northern Africa to the Arabian Peninsula, various types of data reflect spatially homogeneous wetting.The proxy-inferred wetting in boreal drylands is in line with the simulation.However, in Southern Hemisphere drylands, the records are insufficiently consistent (figure 2(a)).In South America, the two proxies are conflicted.In southern Africa and Australia, most records indicate a wetter climate, discordant with simulations.
Over Australia, inter-model differences in precipitation changes mainly occur during September-November (SON).The AWI-ESM-1-1-LR simulates significant increases in precipitation during this season, making it the only model that agrees with proxies (figure 4(h)).Further analysis suggests that the SON precipitation changes are linearly correlated with SST variations in the tropical eastern Indian Ocean and central-western Pacific Ocean (figure S8).This result highlights the significance of accurately simulating SST changes when modeling shifts in Australian precipitation.However, there may be inadequate simulations of SST and precipitation partly due to deficiencies in LIG experimental designs, causing discrepancies compared to proxies.However, there is a lack of significant signals in the similar correlation coefficient field for southern Africa, indicating that the model-data discrepancy might arise from other perspectives.Further discussion about the inconsistency is discussed in section 4.

Summary and discussion
This study investigates the dryland aridity change during the LIG based on PMIP4 simulations.At the global scale, drylands during the LIG are 37% wetter than that during the preindustrial period, as indicated by AI increases; 37% of the preindustrial drylands correspondingly convert to wetter subtypes, which is 2.5 times the area degraded to drier categories.Spatially, there are hemispheric differences with drylands wetting and contracting in the Northern Hemisphere but desiccating and expanding in the Southern Hemisphere.The primary contributor to AI changes lies in the precipitation change, which accounts for 82% of the AI change for the global average and 72%-88% for individual regions, and the precipitation change is mainly attributed to the orbitally-induced modifications in atmospheric circulations.The secondary factor is the near-surface relative humidity change, which exerts influence by altering the PET and contributes to 16% of AI changes for the global average and 13%-17% for individual regions.The simulations of AI show high inter-model agreement over most regions apart from some uncertainties over inland Asia and southern Africa.The spatial responses of the AI correspond well with those of total runoff and deep-layer soil water.There is also broad agreement between AI and LAI changes on a large scale, with the exception of discrepancies over inland Asia due to the seasonality implied in the ecological metric.
The simulated aridity change in our study agrees reasonably with existing proxies in Northern Hemisphere drylands but conflicts with those in southern Africa and Australia.The model-data mismatch may be partly due to deficiencies in LIG experimental designs (Otto-Bliesner et al 2021).First, the current experiment protocol does not consider the ice-sheet reduction and coherent sea-level rising during the LIG (Otto-Bliesner et al 2021).In particular, the retreated Greenland ice sheet could induce meltwater discharges that slow down the Atlantic meridional overturning circulation, which might lead to wetting over the Southern Hemisphere continents (Liu et al 2021).The absence of this process could partly explain the model-data mismatch distinguished over Australia and southern Africa.Another recent study suggests that rising sea levels during the LIG may help reduce the model-data mismatch in Southern Hemisphere temperature changes (Zhang et al 2023), although the impact on moisture requires further investigation.Second, the vegetation and dust are currently prescribed at the preindustrial level in most models, while their changes may further modify hydrological states over local and remote regions.For example, the inclusion of increased vegetation (Griffiths et al 2020, Tabor et al 2020) or reduced dust (Pausata et al 2016(Pausata et al , 2017) ) over the Sahara-Sahel region can not only amplify the local monsoon strengthening but also affect global climates by modifying atmospheric circulations.The limitations of the proxy dataset could also lead to the modeldata mismatch (Scussolini et al 2019).The age of LIG records is beyond the reach of radiocarbon dating, so there are considerable uncertainties in dating that might reach the order of hundreds to thousands of years.The LIG climate change is mainly driven by the altered precession, which has a relatively short period of 21 ka.Therefore, the dating errors of thousands of years could translate to a large fraction of the precession cycle, which might even cause a reversal change in terrestrial moisture conditions and thus an artificial mismatch between simulations and proxies.Uncertainties also come from the interpretation of proxies, including considerations of seasonality and whether site-based data accurately represent largescale features.The model's inadequate representation of reality, particularly regarding precipitation, could also be a contributing factor in the model-data mismatch.
The LIG is often considered a potential analogue to the warming future caused by elevated GHG concentrations.This may hold true for drylands over inner northern Africa, the southern Arabian Peninsula, and South Asia, where aridity reduces (Feng andFu 2013, Cook et al 2020) due to increased precipitation during both periods (Dong and Sutton 2015, Wang et al 2023), notwithstanding different mechanisms of precipitation changes (Wang et al 2023).However, for other dryland zones, including marginal areas of northern Africa drylands, the situation is different.Under future warming caused by high GHG concentrations, aridity is projected to increase primarily due to the warming-induced increase in PET, which may even reverse the wetting caused by increased precipitation in most of those regions (Huang et al 2016).By contrast, during the LIG, aridity decreases/increases in Northern/Southern Hemisphere drylands, which is primarily triggered by altered precipitation and secondly by resulting changes in relative humidity, while annual mean temperature changes are insufficiently important.Similar conditions during the LIG also occur during the mid-Holocene albeit more moderate (Liu et al 2019(Liu et al , 2021)).Overall, we argue that dryland aridity changes during orbitally-induced warm scenarios differ from those under future warm conditions caused by high GHG concentrations, and that the LIG as a potential analogue for future warming does not fully hold to the dryland climate.

Figure 1 .
Figure 1.Comparison between preindustrial control simulations and observations (GPCC and CRU) in the period 1961-1990.The dryland classification based on AI for (a) the observation and (b) multimodel median, and (c) the Taylor diagram for displaying normalized pattern statistics between each model and the observation (marked as REF) in precipitation (gray), PET (blue), and AI (red).For (c), the statistics include the spatial correlation coefficient (the azimuthal position), the normalized standard deviation (the radial distance from the origin), and the normalized centered root-mean-square difference (the distance apart from the REF); the calculation is based on data at all preindustrial dryland grids as shown in (b).

Figure 2 .
Figure 2. Changes from the preindustrial period to LIG for the median field in (a) AI, (b) total runoff, (c) 0-2 m column-integrated soil water, (d) LAI, and (e) dryland distribution.Markers in (a) represent proxies at ∼127 ka BP obtained from Scussolini et al (2019), which is reprinted with permission from AAAS.Stippling indicates regions where at least 9 out of 12 models agree on the sign of the change for (a)-(b), 4 out of 5 models for (c), and 6 out of 8 models for (d).The bottom-right corner in (b)-(d) presents spatial correlation coefficients versus AI changes.For (e), the distribution changes are plotted only where at least 9 out of 12 models agree on the sign of the AI change; preindustrial dryland subtypes that convert into adjacent or nonadjacent wetter climate categories at LIG are labeled using contracting, while the LIG subtypes that expand from preindustrial wetter categories are labeled with expanding; cross and diagonal lines denote the disappeared and newly formed drylands, respectively.

Figure 3 .
Figure 3. AI differences between the LIG and preindustrial period that are caused by changes in (a) precipitation, (b) PET, (c) near-surface air temperatures, (d) available energy, (e) relative humidity, and (f) wind speeds.The sum of (c)-(f) equals (b), and the sum of (a) and (b) equals figure 2(a).Stippling indicates regions where at least 9 out of 12 models agree on the sign of the change.

Figure 4 .
Figure 4. Changes in regional averages of the AI that are caused by individual variables.The labels on the x-axis denote the multimodel median (MM) and individual models (serial numbers can be found in tableS1).Hearts represents the overall changes in the AI.The values are normalized by dividing preindustrial regional means of the AI and thus represent percentage changes in the AI.The calculation is based solely on dryland grids in each region, and the specific dryland coverage for each model is presented in figureS5.

Figure 5 .
Figure 5.The percentage of the annual mean precipitation change (the LIG minus preindustrial period) that occurs in December-February (DJF), March-May (MAM), June-August (JJA), and September-November (SON).

Figure 6 .
Figure 6.The LIG minus preindustrial period change during JJA and DJF in the (a), (b) surface temperature ( • C) and (c), (d) geopotential height (shading; m) and wind field (arrow; m•s −1 ) at 850 hPa.Regions without cross hatching are where at least 9 out of 12 models agree on the sign of the change.