The archeological site EH is located on the Moroccan Atlantic coast in the Rabat–Témara region (33∘57′08.9′′ N, 6∘55′32.5′′ W). The climate of the area is marked by a strong seasonal variability (Fig. 1). Summer is relatively warm and dry, precipitation is almost absent, and evaporation is particularly high, while winter is relatively cool and short and is the rainy season (Sobrino and Raissouni, 2000; Lionello et al., 2006). Inferences based on species presence and abundances suggest that, in the past, the site underwent significant climatic fluctuations, which resulted in a succession of relatively arid (humid) and open (closed) environments at EH (Stoetzel, 2009; Stoetzel et al., 2011, 2012a, b). As a result, paleolandscapes of the Late Pleistocene are described as open steppe or savanna-like lands with patches of shrubs, woodlands and water bodies, the latter expanding during wet periods, especially during the mid-Holocene (Stoetzel, 2009; Stoetzel et al., 2012a, 2014).

EH displays a high-resolution stratigraphy, and its layers have revealed an impressive taxonomic richness and delivered an significant amount of large and small vertebrate remains (Michel et al., 2009; Stoetzel et al., 2011, 2012b). Its stratigraphy is currently divided into 11 layers (Fig. 2a) (each layer is abbreviated as “L” followed by the number of the layer; for consistency the present is referred to as “L0”). Among these layers eight are well studied and considered in this study. All of these eight layers have been dated (Ben Arous et al., 2020b). Three different methods were used: AMS-14C (Nespoulet and El Hajraoui, 2012; Marquer et al., 2022), OSL (Jacobs et al., 2012) and combined US–ESR (Janati-Idrissi et al., 2012; Ben Arous et al., 2020a). AMS-14C is the most reliable dating method of the three and is used for recent layers (L1 and L2). However, for older layers beyond the reach of AMS-14C (L3, L4a, L5, L6, L7 and L8), OSL and combined US–ESR methods are used. However, when applied to the same layer, these two methods present discrepancies. This is the case, for example, of L5 dated at ∼ 60 ka by combined US–ESR and at ∼ 100 ka by OSL (Fig. 2f). In addition, when two consecutive layers are dated with these different methods, dates can be inconsistent with the relative position of the layers in the stratigraphy, as is the case for L8 dated at ∼ 90 ka using combined US–ESR and L7 at ∼ 110 ka using OSL (Fig. 2f). To overcome this issue, we choose to set two separate dating hypotheses (D's) of the stratigraphic sequence. In the following, D1 refers to the chronological sequence based on AMS-14C and combined US–ESR dates, and D2 refers to the sequence based on AMS-14C and OSL dates. The D's are presented in Fig. 2f.

Paleoenvironments at EH have been inferred based on two different kinds of proxies: species presence and stable-light-isotope composition. Regarding species presence, paleoenvironments are reconstructed using the taxonomic habitat index (THI). The THI is a paleoecological index which allows one to reconstruct the paleolandscape based on the species composition of the community (Stoetzel, 2009; Stoetzel et al., 2011, 2014). Each species is associated with its preferred habitat based on the assumption that species' ecological preferences were the same in the past as they are today. From that a qualitative composition of the paleolandscape is deduced, expressed as a percentage. Data are from Stoetzel et al. (2014) and are presented in Fig. 2b. Note that the oasis habitat was not considered because its percentage of presence does not vary over the EH sequence.

Isotope-based inferences are from Meriones teeth. They provide information about the diet and the environment of studied individuals (Longinelli and Selmo, 2003; Navarro et al., 2004; Royer et al., 2013). Two isotopic fractions are considered: δ18O and δ13C. Plants consumed by small mammals are sensitive to environmental conditions, which shows up in their δ13C and δ18O values. Because of that, they are good indirect indicators of aridity, seasonal variation and vegetal cover (Longinelli and Selmo, 2003; Blumenthal et al., 2017; Blumenthal, 2019). We used δ18O and δ13C means, minimums and maximums and reconstructed mean annual precipitation (MAP) computed from δ18O from Jeffrey (2016). Data are presented in Fig. 2c, d and e.

We ran a new set of paleoclimate simulations covering the Late Pleistocene to mid-Holocene period using the model LMDZOR6A (Hourdin et al., 2020). This model is the atmosphere–land surface component of the IPSL–CM6A–LR coupled model (Boucher et al., 2020) that has been used to run the CMIP6 ensemble of past, present and future coupled climate simulations (Eyring et al., 2016), including the mid-Holocene (Braconnot et al., 2021), last interglacial (Sicard et al., 2022; Otto-Bliesner et al., 2021) and Pliocene (Haywood et al., 2016) periods. It has an atmospheric resolution of 144 points in longitude, 143 points in latitude and 79 vertical levels (144 × 143 × L79). Compared to previous IPSL model versions it has a finer spatial and vertical resolution and an improved representation of atmospheric and land surface processes. In addition, when the model is run for present-day conditions (with sea surface temperature prescribed to the monthly climatology of the observed Atmospheric Model Intercomparison Project (AMIP) sea surface temperature (SST) fields; Boucher et al., 2020, 2018), the simulated climatology of temperature and precipitation over the region is consistent with observations (Figs. 1 and 3). The simulated annual mean cycle of surface air temperature and precipitation matches quite well with the observed one, despite an overestimation of summer temperature and slight 0–1-month shift in the seasonal cycle of precipitation (Fig. 3).

We produced a total of six paleoclimate simulations with the LMDZOR6A model for the key periods of the EH sequence: midH (for the mid-Holocene period), earlyH (for the early Holocene period), midMIS3 (for the mid-MIS3 period), lateMIS4 (for the late MIS4 period), midMIS4 (for the mid-MIS4 period) and MIS5d (for the MIS5d period). The different simulations differ in terms of prescribed Earth orbital parameters, atmospheric trace gas composition, ice-sheet configuration and SST in order to represent the climate conditions of the different periods (see Table 2 for details). For this we make use of pre-existing paleoclimate and control simulations that have been run in the last 10 years with different versions of the IPSL model (Marti et al., 2010; Dufresne et al., 2013; Boucher et al., 2020) (details are available in Table 1). For all these simulations the vernal equinox is prescribed to occur on 21 March at noon (GMT), following the Paleoclimate Modeling Intercomparison Project (PMIP; Kageyama et al., 2018) recommendations.

We also ran a control simulation Ctrl (see Table 2 for details), representative of the present-day climate. The SST boundary conditions used in Ctrl correspond to the mean annual cycle of the SST estimated from current observations used for the AMIP (Boucher et al., 2018) and repeated in time. This Ctrl simulation will be considered to be the reference for the current climate in our ensemble of new atmospheric simulations.

Unfortunately, simulated paleoclimate SSTs are not directly comparable because they were performed using different versions of the IPSL model (Table 1). This is due to the fact that the various versions of the model are characterized by different physical representations, resolutions and tuning. In particular, these different model versions show different present-day SST biases when compared to observations (Fig. 4). When these paleoclimate SSTs simulated using different model versions are used as boundary conditions in simulations run with LMDZOR6A (instead of the AMIP SST field), they translate to different representations of the seasonal cycle of surface air temperature and precipitation at EH (Fig. 3). The magnitude of the seasonal cycle is overestimated, and precipitation is underestimated, with almost no seasonality when IPSL–CM5A or IPSL–CM5A2 control SST is used as boundary condition. The differences that can be attributed exclusively to the differences in the SST biases are significant and may deteriorate the simulation of other variables of interest, thus complicating the intercomparison between our new LMDZOR6A simulations. Indeed, the differences between these simulations for the periods of interest at EH could then result from the difference in bias between the various versions of the model rather than representing significant climate differences between periods. In order to produce a homogeneous set of simulations we thus apply a correction to the SST fields (described below). The objective is to correct differences due to model versions while preserving differences related to climate variation between past periods. These corrected SST fields are used as the prescribed paleoclimatic SST boundary conditions in our new ensemble of LMDZOR6A simulations.

We correct the simulated SST of each coupled simulation for the systematic bias of the model. This is done by removing the SST bias corresponding to the model version used. In other words, corrected SSTs were obtained according to the formula 1SSTcor=SSTsim+(SSTamip-SSTmod), where SSTsim is the SST from the coupled simulation; SSTamip is the AMIP SST; SSTmod is the SST of the model version for current days; and SSTcor is the corrected SST, which is used as a boundary condition in our new LMDZOR6A simulations. The underlying assumption in this correction scheme is that the mean annual SST bias cycle, i.e., SSTmod - SSTamip for each, is stationary in time. Note also that boundary condition files of pre-existing (coupled) simulations were interpolated on a 143 × 144 × L79 grid to be compatible with the grid of LMDZOR6A. The corrections are applied to daily SST values of the modern climate so that the changes in season from the orbital parameters are properly accounted for in the new daily SST imposed as a boundary condition given the way the vernal equinox is prescribed in the model.

Configuration details for the new set of simulations are summarized in Table 2. The length of all these new simulations is 50 years, which is long enough given the fact that LMDZOR6A is an atmospheric model. In the following annual mean cycles are estimated from the last 30 years of each LMDZOR6A simulation.

To characterize the large-scale climate over the area, we worked on the mean annual cycle of the four grid cells containing EH. From simulations Ctrl, midH, earlyH, midMIS3, lateMIS4, midMIS4 and MIS5d we extracted nine output variables. We choose to focus on variables that are likely to directly or indirectly influence the landscape and/or the biotic environment. Based on a review of the ecological literature, we selected tsol (temperature at surface, C∘) (Gillooly et al., 2001; Yom-Tov and Geffen, 2006; Ebrahimi-Khusfi et al., 2020), precip (precipitation, millimeters per month) (Yom-Tov and Geffen, 2006; Alhajeri and Steppan, 2016; Ebrahimi-Khusfi et al., 2020), qsurf (specific humidity, kg kg-1) (Hovenden et al., 2012; Alhajeri and Steppan, 2016), w10m (wind speed at 10 m, m s-1) (McNeil, 1991; Tanner et al., 1991; Chapman et al., 2011; Pellegrino et al., 2013), sols (solar radiation at surface, W m-2) (Monteith, 1972; Fyllas et al., 2017), drysoil_frac (fraction of visibly dry soil, %) (Paz et al., 2015) and humtot (total soil moisture, kg m-2) (Paz et al., 2015). Two additional variables are computed: the diurnal temperature range tsol_ampl_day (C∘; Alhajeri and Steppan, 2016) computed from tsol_max (daily maximum temperature) and tsol_min (daily minimum temperature) as tsol_max - tsol_min and the hydric stress hyd_stress (mm d-1; Martínez-Blancas and Martorell, 2020) computed from evapot (potential evaporation) and evap (evaporation) as evapot - evap.

We explore variation in annual means (mean values of the variables) and annual standard deviations (amplitude of seasonal variation). The means and standard deviations are computed on the annual mean cycle estimated from the last 30 years of each LMDZOR6A simulation; thus they represent, respectively, the annual mean and the seasonal amplitude of the variable. In all analyses, climate variables are standardized between periods.

Potential biases related to calendar effect (the change in length of seasons and months between periods; Bartlein and Shafer, 2019; Joussaume and Braconnot, 1997) were considered and discounted in the statistical analyses by characterizing the annual mean cycles by their annual means and standard deviations. Indeed, these metrics are only slightly, if at all, affected by the calendar effect, in particular when they are properly weighted by the number of days in the months (estimated using the modern or a celestial calendar). This choice is made to keep the physical consistency between the different variables that nonlinearly depend on moisture and temperature.

The association of paleoclimate simulations with the stratigraphic layers of EH is based on age proximity. This step is performed for both D1 and D2 (the D's presented in Sect. 2.1.2). As a result, we obtain two hypothetical paleoclimate sequences corresponding to EH sequence.

In order to easily visualize the climate proximity and differences between EH layers we used a principal component analysis (Jolliffe and Cadima, 2016). This analysis allows one to find new uncorrelated variables that successively maximize variance (the principal components). They are computed from the eigenvectors and eigenvalues of the covariance–correlation matrix (correlation matrix in our case, as all the variables are standardized). Then, the data can be visualized along the leading principal components, which maximizes the amount of information displayed. It allows us to visualize EH layers along the two main axes of variance in climate data, resulting in the proximity between stratigraphic layers (points) in the plot representing the proximity between the climate of the layers. The principal component analyses are performed on the annual mean and standard deviations (seasonal amplitude) of climate variables.

The two hypothetical paleoclimate sequences corresponding, respectively, to D1 and D2 are presented in Fig. 8. The two principal components analyses are shown in Fig. 9. To visualize how climate variables structure the climate space of these two principal component analyses, biplots are available in Supplement Fig. S2.

Based on D1, our results indicate four major climate transitions (Fig. 8). The first occurs between L8 and L5. In L5 the climate is wetter and colder, with more precipitation, more soil humidity, increased wind speed, less hydric stress and a smaller portion of dry soil. Humidity, precipitation and wind speed also show a significant seasonal variability. The second transition, less marked, is between L5 and L4a. Climate in L4a is rather similar to the climate in L5, but precipitation and humidity have increased. Seasonal variation are globally more pronounced. The third transition is between L3 and L2. The climate changes drastically between these two periods, with hotter and drier conditions in L2 corresponding to the end of the last deglaciation. Temperature, solar radiation, water stress and soil dryness all increase, coupled with a decrease in precipitation, soil moisture and wind speed. All these changes are consistent with aridification or desertification from L2. Surprisingly, however, specific humidity is enhanced, which is partly consistent with the decrease in precipitation and wind speed at this coastal location and is concomitant with large changes in the atmospheric circulation patterns over this region from L2 (Fig. 5). The last climatic transition is more subtle and occurs between L1 and L0. The environment in L0 seems closer to the one in L8, with more seasonal variability in temperature, solar radiation and water stress (Fig. 8).

Overall, there is an alternation of two main climate types. This partition is confirmed by the principal component analysis shown in Fig. 9. The first principal component explains 65.85 % of the observed variance (of the standardized variables) and splits the layers of EH into two climate regimes. The first regroups L3, L4a and L5 and is defined by humid and windy conditions, with a high seasonal variability. The second includes L0, L1, L2 and L8 and is characterized by hot and dry conditions. The second axis, explaining 30.28 % of the variability, divides the latter group into two subgroups: L1 and L2, with high seasonal variability, and L0 and L8, with a lower seasonal variability.

Regarding D2, we observe three significant and abrupt climatic transitions (Fig. 8). The first happened between L5 and L4a. In L4a, the climate is much windier, and temperatures are colder, with a substantial increase in diurnal temperature range. Precipitation and soil moisture increase significantly, while hydric stress decreases. The climate also presents an overall higher seasonal variability. These tendencies persist in L3. The second transition is between L3 and L2. The soil is drier, and the hydric stress increases greatly as well as solar radiation and temperature. Precipitation and soil moisture are less important. Conditions in L1 are close to those in L2, but a bit colder with a less marked water stress. The last climate transition, smoother than the previous ones, occurs between L1 and L0 and is mainly marked by a global decrease in seasonal variation.

Results associated with D2 suggest that three types of climate succeeded one another at EH. The first group is composed of L8, L7, L6 and L5; the second of L3 and L4a; and the third of L2, L1 and L0. As for D1, this partition is supported by the principal component analysis presented in Fig. 9. The first principal component, which explains 52.10 % of the observed variance, separates the group containing L8, L7, L6 and L5 from the one composed of L3 and L4a. The second principal component, explaining 42.35 % of the observed variance, separates the group of L2, L1 and L0 from others. L8, L7, L6 and L5 are characterized by a hot and dry climate and a low seasonal variation. L3 and L4a are defined by a wet and windy climate, with significant precipitation and high seasonal variability. Finally, L2, L1 and L0 present a hot environment associated with a significant water stress.

Results of 2B-pls and pairwise correlations between climate and paleoenvironmental variables are presented in Fig. 10. Regarding D1, no significant covariation is found between THI values and climate variables. However, 2B-pls shows that there is a statistically significant covariation between isotope data and all climate variables and between isotope data and the second principal component (climate variables contributing to the second principal component are detailed in Fig. 9). Specifically, this second principal component is positively correlated with δ18O mean, maximum and minimum values and δ13C maximum values. Conversely, D2 presents neither a statistically significant result for 2B-pls nor correlations, meaning that for D2 paleoclimate inferences and paleoenvironmental proxies are not consistent.