The Community Earth System Model (CESM, version 1.2) with the Community Atmosphere Model version 4 (CAM4; Neale et al., 2013) and Community Land Model version 4 (CLM4; Oleson et al., 2010; Lawrence et al., 2011) is employed in this study. The CESM incorporates prognostic carbon–nitrogen (CN) equations and the dynamic vegetation model into the land surface model. The model simulates carbon dynamics such as productivity, phenology, decomposition, and allocation with carbon and nitrogen fluxes at the ecosystem–atmosphere interface by the CN prognostic equations. This new CN parameterization makes a substantial difference in vegetation simulation by limiting gross primary production (GPP) with soil mineral nitrogen in the climate simulations particularly (Thornton et al., 2007; Castillo et al., 2012; Arora et al., 2020). 1GPP=GPPp⋅1-CFexcessGPPp GPP becomes the potential GPP (GPPp) if there is no nitrogen limitation, and nitrogen limitation is linked to soil organic matter in the model. Nitrogen limitation on GPP is expressed by CFexcess, which is decided by soil mineral nitrogen (Nsminn). Nsminn is a function of the conversion rate of soil organic nitrogen to mineral nitrogen (NFsoil4n→sminn) and is affected by the amount of soil carbon (soil4c) eventually (Eqs. 2–9). It was reported that GPP in CESM version 1 was smaller than other CMIP6 models because of the larger nitrogen limitation effect. More details on the CN parameterization in the CLM can be found in Thornton et al. (2002, 2009) and Kluzek (2012). The description of the variables used in this study including those in Eq. (1) are listed in Appendix A. 2CFexcess=CFavail_alloc-CFalloc3CFavail_alloc=GPPp-MR4CFalloc=Nplant_demand⋅fpg+Nretrans⋅kc_allockn_alloc5fpg=NuptakeNplant_demand6Nplant_demand=CFavail_alloc⋅kn_allockc_alloc-Nretrans7Nuptake=NsminnΔt-NFimmob8Nsminn=Nsminn+NFsoil4n→sminn⋅dt9NFsoil4n→sminn=soil4c⋅kdecompdt⋅10

The albedo–precipitation feedback has also been considered to be one of the important aspects in the land–atmosphere interaction and should be clearly examined in the GS simulation (e.g., Charney, 1975; Cess, 1978; Bonfils et al., 2001; Houldcroft et al., 2009; Levine and Boos, 2017). Precipitation increases if surface albedo decreases because of more vegetation through the stratification of the atmosphere, particularly in drylands. However, in the CMIP5 (Coupled Model Intercomparison Project 5) EMH simulations, soil moisture is irrelevant to surface albedo in most land surface models (LSMs) (e.g., Takata et al., 2003; Zeng, 2005; Houldcroft et al., 2009; Oleson et al., 2010; Watanabe et al., 2010; Vamborg et al., 2011). This study uses CLM4 for simulation of land surface processes. In the CLM4, surface albedo is parameterized as a function of soil moisture in the calculation of soil background albedo. In this parameterization, soil background albedo decreases with increases in soil moisture of the first soil layer as follows. 10αsoil=αsat+Δ11Δ=0.11-0.40θ1θ1≤0.2750θ1>0.275 In addition, the soil texture distribution is classified by sand and clay percentage, and soil has 20 color classes from the low to high albedo. In the lowest albedo (darkest soil), dry (saturated) albedo is 0.08 (0.04) and 0.16 (0.08) in the visible and near-infrared (NIR) range, respectively. In the highest albedo (lightest soil), the dry (saturated) albedo is 0.36 (0.25) and 0.61 (0.50) in the visible and NIR range, respectively. αsat is equal to the albedo of the saturated soil.

For our CESM simulations of the GS, the horizontal resolution is 4∘×5∘ in global scale with 26 vertical levels. The initial land surface data are made using CLM4 offline simulations for 600 years in the CN-accelerated decomposition mode (Thornton and Rosenbloom, 2005), and then using 1200-year CLM4 simulations with CNDV (carbon–nitrogen–dynamic–vegetation) and atmospheric forcing data for the years 1948–1972 based on Qian et al. (2006). The present (i.e., 0K) and EMH (i.e., 8K) simulations are 500-year climate simulations using the pre-industrial and EMH conditions, respectively (Table 1). The EMH period is set to 8000 years ago (8K) in this study because the 8K period was relatively greener than other periods (Jolly et al., 1998; Groner et al., 2018).

Additional sensitivity experiments using the EMH conditions are conducted with different land states to examine the impacts of soil nitrogen, soil texture, and the mega-lake on the GS simulation (Table 1). Soil carbon and nitrogen are negligible in the current Sahara. To the best of our knowledge, there is no information on soil organic carbon and nitrogen in North Africa during the EMH. It is reasonable to consider more soil carbon and nitrogen in the GS than in the present-day Sahara because of their close coupling with biomass (Brooks, 2003; Wang et al., 2010; Yang et al., 2014). To consider the possibility of relatively larger soil carbon and nitrogen in the GS, more soil carbon and nitrogen are added with less sandy soil (i.e., loamier soil with more soil organic matter) in the 8KCN and 8KCNS experiments based on Jolly et al. (1998) to prescribe soil organic matter in the GS (i.e., 10–22∘ N, 345∘ W–40∘ E) using nearby grassland grid values (Fig. 1). In these sensitivity experiments, the feedbacks between vegetation and increased soil carbon and nitrogen evolve dynamically in the model through litterfall and decomposition processes.

In all EMH simulations, the Earth's orbital parameters and atmospheric CO2 concentration are prescribed based on Berger (1978) with pre-industrial SST (Table 2), and atmospheric dust is reduced following Pausata et al. (2016) (Fig. 2). In addition, Lake Mega-Chad is considered in our EMH simulations by assigning 73 % of the two grids near the Lake Chad region (14–18∘ N, 15∘ E) as lake area to consider the extensive Lake Chad in the EMH (Leblanc et al., 2006; Quade et al., 2018). Our analysis focuses on three critical regions in North Africa around the border of the vegetation in our simulation: North Africa at 20∘ N (solid boxed area in Fig. 3, NA hereafter) and the Sahara–Sahel (SS hereafter) region. These regions are selected as the analysis region because the NA region is closely related to the onset and progress of the African monsoon (Levis et al., 2004) and the SS region was sensitive to precipitation–land interaction (Kutzbach et al., 1996). The Sahara–Sahel region is divided into the northern Sahara–Sahel region (dotted box area in Fig. 3, NSS hereafter) and southern Sahara–Sahel region (dashed box area in Fig. 3, SSS hereafter) with a common latitudinal band around 18∘ N in this study. The NSS is the boundary region of vegetation in the EMH, and the SSS shows substantial increases in precipitation with changes in the orbital forcings (more information in the next section). These two regions share the 18∘ N latitude band for considering the common properties between two regions and the model resolution.

The 8K simulation shows typical climate patterns in the Sahara–Sahel region during the EMH, and there are substantial differences of surface climates between the 0K and 8K experiments in North Africa (Fig. 3). The summer downward radiation at the top of the atmosphere increases by 6 % in the SS region compared to the present because of the change in the orbital forcing and lower amount of dust in the EMH. This increase in solar radiation leads to an increase in surface shortwave radiation and net radiation by 30 and 21 W m-2, respectively, as well as subsequent temperature warming by 1.8 ∘C in the NA region in the 8K experiment (Fig. 3). The intensified land–sea thermal contrast during the EMH makes spatial changes in precipitation and wind patterns over Holocene North Africa compared to the present climate. Notably, the increases in air temperature and shortwave radiation are not substantial in the SS region (less than 10 % compared to the NA region) because of the increase in precipitation and clouds with the northward migration of the monsoon by the intensified land–sea thermal contrast (Fig. 3d). Indeed, the meridional wind in the 18∘ N region changed from northerly (-0.7 m s-1) to southerly (0.3 m s-1) in the 8K experiment (Fig. 3g, h), indicating a northward shift of the Intertropical Convergence Zone (ITCZ) (i.e., the convergence zone of the trade winds) to above 18∘ N. This ITCZ shift caused a favorable condition for a moister Sahara by transporting more moisture and subsequently more precipitation to the SS region (Chikira et al., 2006; Larrasoaña et al., 2013). Compared to the pre-industrial simulation (i.e., 0K), summer precipitation in the EMH simulation (i.e., 8K) increases by 18 % (from 4.0 to 4.7 mm d-1) and 15 % (from 2.5 to 2.9 mm d-1) in the SSS and NSS region, respectively (Fig. 3e, f).

The pre-industrial simulation captures the observed spatial extent of the current Sahara based on MODIS (Moderate Resolution Imaging Spectroradiometer) land surface data (Fig. 4a) (Broxton et al., 2014). Importantly, the model does not simulate an increased vegetation fraction, despite the increased precipitation over the SS region during the EMH (Fig. 4b). Reconstruction of vegetation based on the pollen records reveals that grass and trees expanded up to 26 and 20∘ N over North Africa during the GS period, respectively (Hély et al., 2014; Hopcroft et al., 2017). Also, leaf wax data indicate that the GS extended as far north as 31∘ N (Tierney et al., 2017). In contrast to the proxy data, vegetation fraction and GPP show negligible changes in North Africa and even a decrease in western Africa and the southern border of the current Sahara in the 8K simulation (Fig. 4b, e). It is notable that Lake Mega-Chad did not make substantial changes in simulated summer precipitation northward movement of the 200 and 412 mm yr-1 precipitation isohyet for 20 % grassland cover, which is similar to Contoux et al. (2013) (Figs. S1, S2 in the Supplement).

Hopcroft et al. (2017) attributed such underestimation of the GS to an excessive precipitation requirement for vegetation growth in the model (Table 3). Our investigation also shows that the minimum amount of precipitation for vegetation growth is 577 mm yr-1 in the 8K experiment, and this precipitation threshold is larger than the observation value of 281 mm yr-1 (Hopcroft et al., 2017). This larger threshold indicates that our model has relatively little vegetation cover with respect to the same amount of precipitation and is not fully understood yet in the ESM (Hopcroft et al., 2017). Our investigation reveals that this excess water requirement by vegetation growth becomes smaller with improved water use efficiency (WUE) and proper assignment of soil organic matter and soil texture, which is discussed in the next section.

The 8KCN experiment is designed to study the effect of an increase in soil nitrogen by vegetation on the GS and simulates typical surface climate patterns during the EMH (not shown here). Importantly, the 8KCN simulation reproduces larger precipitation and vegetative areas with an increase in GPP in the SS region compared to the 8K experiment (Fig. 4). Vegetation fraction increases by about double in the SS region (21 % and 8 % in the SSS and NSS regions, respectively) in the 8KCN experiment (Table 4). The 8KCN experiment also shows a 2 times larger vegetation fraction than the 8K experiment around the boundary of precipitation changes between the present and EMH (about 18∘ N). Particularly, the 8KCN experiment shows 3.3 and 5.3 times larger GPP in the NSS and SSS regions, respectively, compared to the 8K experiment (Table 5). Consequently, the precipitation threshold for vegetation decreases to 412 mm yr-1 in the 8KCN from 577 mm yr-1 in the 8K experiment together with increased WUE values (Figs. 5 and 6). With this increase in vegetation fraction, the desert–grassland boundary moves one grid box northward in the 8KCN simulation compared to the 8K experiment (Fig. 4i).

Our analysis shows that this increased vegetation in the 8KCN results from moderated nitrogen limitation on photosynthetic carbon uptake (Eq. 1 and Fig. 7). The nitrogen downregulation fraction defined by GPP/GPPp is about 0.5 and 0.3 in the NSS and SSS regions, respectively, in the 8K experiment but decreases by about 0.1 in the 8KCN experiment. Our further simulation indicates that vegetation fraction and vegetation types depend on the amount of soil nitrogen (Fig. S3). The increases in vegetation fraction and tree cover fraction are approximately proportional to the amount of soil nitrogen. The vegetation fraction increase is substantially smaller in the 8KCNh (i.e., half nitrogen simulation of the 8KCN) compared to the 8KCN (Fig. S3). Trees in the 8KCN are also substantially replaced by grass and shrub in the simulation with less soil nitrogen (8KCNh).

We speculate that albedo–precipitation feedback is reinforced with more vegetation in the 8KCN experiment. In the 8K experiment, the surface albedo decreases compared to the 0K experiment, mainly because of the increase in soil moisture by the enhanced precipitation. In the 8KCN simulation, the surface albedo diminishes more because the vegetation increases more, and the surface albedo decreases with increasing vegetation cover. Our further analysis suggests that such feedback reinforcement makes an additional 0.1 mm d-1 precipitation and expansion of vegetation, especially in the SSS region (Fig. 8a and b).

Soil is classified as sandy soil for the present-day Sahara in the 8K and 8KCN simulations like previous studies on the GS (e.g., Hopcroft et al., 2017; Groner et al., 2018). It is, however, plausible that soil in the GS includes more nutrients and humus than sandy soil (i.e., loamier) with the extensive vegetation in this humid region because soil organic matter is formed by decomposition of litter and woody debris (Levis et al., 2004; Koven et al., 2013). Such soil type changes make impacts on soil color as well as soil hydraulic and thermal properties. Compared to sandy soil, loamy soil has larger soil porosity and heat capacity but smaller saturated hydraulic conductivity because it contains more slit and clay than sandy soils. Consequently, loamy soils have larger water-holding (i.e., soil porosity) and heat capacities (Kutzbach et al., 1996; Levis et al., 2004). A previous study based on the offline dynamic vegetation simulation of the GS shows that the simulated greening of the Sahara is sensitive to changes in soil thermal and hydraulic properties without bare-soil albedo changes (Lu et al., 2018).

Our coupled ESM simulation of 8KCNS also shows that soil texture leads to significant changes in vegetation and climate in North Africa. Vegetation cover fraction shows a 78 % increase over the SS region when the soil nitrogen and soil texture are modified all together (i.e., 8KCNS experiment) compared to the 8KCN experiment (Fig. 9e; Table 4). Notably, those feedback processes are critical in the western SSS, which is coincident with the region classified as sandy soil in the ESMs (Fig. 9f–l). A vegetation fraction increase is clearly observed, especially in the western SSS (<10∘ E) in the 8KCNS experiment, with a 2∘ N northward movement (15 to 17∘ N) of the 20 % vegetation coverage border and its related GPP increases. This result suggests that the western SSS is a hot spot from the perspective of soil texture in the EMH simulations.

Our analysis indicates that such substantial sensitivity is mainly related to changes in soil hydraulic and thermal properties, but the darker soil color of loamy soils also regulates climate and vegetation in the study region. The relatively darker soil color of loamy soils in the 8KCNS experiment leads to smaller surface albedo, and its increase in reinforced by increases in summertime precipitation (see Eq. 9; Fig. 9a–b). That is, the darker soils enhance absorption of solar radiation (i.e., increase in surface net radiation) and more available energy is partitioned into latent heat fluxes (i.e., smaller Bowen ratio), leading to more precipitation in North Africa (Fig. 9c–e), possibly through the albedo–precipitation feedback and its related moisture recycling (Hopcroft et al., 2017; Levine and Boos, 2017; Groner et al., 2018). Also, previous studies indicate that enhanced evapotranspiration results in more rainfall by a weakened African easterly jet (AEJ) (Cook, 1999; Wu et al., 2009; Rachmayani et al., 2015). Our simulation also reveals that the AEJ is weakened with more evaporation and summertime rainfall, which is not found in the 8KCN experiment.