Abstract
The 'Greening' and subsequent desertification of the Sahara during the early to mid-Holocene is a dramatic example of natural climate change. We analyse a suite of simulations with a newly palaeo-conditioned configuration of the HadCM3 coupled model that is able to capture an abrupt desertification of North Africa during this time. We find that this model crosses a threshold of moisture availability for vegetation at around 6000 years before present. The resultant rapid reduction in vegetation cover acts to reduce precipitation through moisture recycling and surface albedo feedbacks. Precursor drying events which are not directly forced also indicate that the model is close to a critical moisture level. Similar precursor-like events appear in a Holocene record from the East of the continent, hinting that the natural system may resemble some of the properties of this model simulation. The overall response is not fundamentally altered by the inclusion of solar irradiance variations or volcanic eruptions. The simulated timing of the abrupt transition is mostly controlled by orbital forcing and local positive feedbacks, but it is also modulated to some extent by the state of the atmosphere and ocean. Comparisons with proxy records across North Africa show good agreement with the model simulations, although the simulations remain overly dry in the East. Overall, a threshold response may present a useful model of the real transition, but more high-resolution palaeoclimate records would help to discriminate among the predictions of climate models.

Original content from this work may be used under the terms of the Creative Commons Attribution 4.0 license. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
1. Introduction
Climate change already poses a great risk to human society and ecosystems (Hoegh-Guldberg et al 2018, Lee et al 2021) and abrupt climate change represents a significant but uncertain additional risk (Lenton et al 2019). Palaeoclimate examples provide a rich archive of the natural examples of abrupt climate transitions (Brovkin et al 2021). An increasing focus on potential tipping points over recent decades is largely motivated by information about past events from palaeoclimate records (e.g. Broecker et al 1985). Apart from early warning diagnostics (e.g. Boers 2021), climate or Earth system models (ESMs: coupled climate models that comprise general circulation models (GCMs) of the atmosphere and ocean coupled to Earth system components such as dynamic vegetation or atmospheric chemistry) are a main tool for predicting the future occurrence of abrupt change. Despite this, climate models and ESMs fail to satisfactorily reproduce the majority of palaeo-instances of abrupt change (Valdes 2011), probably because these models are not tested against past events during their development, meaning that the emergent properties of the Earth system in these models are not calibrated with actual abrupt events (Hopcroft and Valdes 2021).
North Africa is one key region for considering abrupt climate change (Petit-Maire and Guo 1996, Brovkin et al 1998, Claussen et al 1998, 1999, 2017, de Menocal et al 2000, Hopcroft and Valdes 2021). During periodic 'African Humid Periods' (AHPs) much of today's Sahara desert was transformed by wetter conditions to a region of shrub and savannah vegetation (Hoelzmann et al 1998, Hély et al 2014), wetlands (Brostrom et al 1998, Krinner et al 2012, Chen et al 2021) and human settlements (Kuper and Kropelin 2006, Manning and Timpson 2014). The strong environmental response in North Africa is due to the profound influence of the seasonality of insolation on the tropical hydrological cycle which is controlled by long-term variations in Earth's orbit around the Sun (e.g. Kutzbach and Street-Perrott 1985). During phases of low precession when eccentricity is not too low (i.e. Earth's orbit is elliptical), northern hemisphere insolation is stronger during the summer than today and this intensifies the past monsoon rainfall (Kutzbach 1981). The periodic greening of the Sahara during AHPs is among the most striking manifestations of this orbital forcing of monsoons.
The most recent AHP terminated around 4000–6000 years ago and was, in places, a rapid return to desert conditions (Petit-Maire and Guo 1996, de Menocal et al 2000, Tierney et al 2017). The abrupt climate changes are indicative of thresholds in the system. The primary candidate being the positive feedback between vegetation cover and rainfall (Charney 1975, Charney et al 1975). The vegetation feedback has been shown in three-dimensional model simulations (Kutzbach et al 1996, Claussen and Gayler 1997, Texier et al 2000, Levis et al 2004) and moisture recycling by vegetation probably acts as an additional positive feedback (Kutzbach et al 1996, Levis et al 2004, Hopcroft et al 2017). The oceanic thermal response to orbital changes acts to amplify the monsoon response (Braconnot et al 1999). Mineral dust has also been considered (Pausata et al 2016), but its role is small when up-to-date physical dust properties are implemented in models (Hopcroft and Valdes 2019, Braconnot et al 2021).
Positive feedbacks between land and the atmosphere can imply the presence of thresholds and hence abrupt transitions (Brovkin et al 1998, Claussen et al 1998, 2017). Climate and ESMs largely fail to reproduce the greening of the Sahara during the Holocene (Joussaume et al 1999, Harrison et al 2015, Brierley et al 2020), although with exceptions (Claussen and Gayler 1997, Dallmeyer et al 2020). Two recent studies show the potential of abrupt changes in this region (Dallmeyer et al 2020, Hopcroft and Valdes 2021). Open questions remain concerning the realism of threshold behaviour. For example, abrupt change seems to dominate palaeoclimate records in some regions of Africa (Tierney and deMenocal 2013, Trauth et al 2015, Tierney et al 2017) but not others (Kropelin et al 2008, Tjallingii et al 2008). Model simulations also disagree on the spatial expression (Liu et al 2006, Brovkin and Claussen 2008, Claussen et al 2017). Other uncertainties include the sensitivity of the transition out of the AHP to high-frequency climate forcing such as due to solar irradiance variations or volcanic eruptions.
In this work we analyse a suite of transient climate simulations of the Holocene. We use an updated version of the coupled climate model HadCM3 which is able to reproduce the Holocene greening including a threshold return to desert conditions (Hopcroft and Valdes 2021). In addition to a simulation forced by varying orbit, ice-sheets and greenhouse gas concentrations, we incrementally include forcings due to solar irradiance variations, volcanic eruptions and two reconstructions of early anthropogenic land-use. We evaluate the properties of abrupt changes in this suite of simulations with respect to the different forcings applied and make comparisons with a suite of palaeoclimate reconstructions.
2. Methods
2.1. Climate model
We use a suite of transient simulations covering the Holocene from 10 ka BP to 0 ka BP (BP: before CE 1950) (Hopcroft and Valdes 2021). These use the coupled GCM HadCM3 (Gordon et al 2000) which has been widely used in IPCC reports and other studies (e.g. Schaller et al 2016, Armstrong et al 2019, Sime et al 2019, Hopcroft et al 2020, 2021, Buizert et al 2021). HadCM3 has a horizontal resolution of with 19 levels in the atmosphere and with 20 levels in the ocean. The land surface is represented with a tiling of surface types within MOSES2.1d and includes interactive vegetation (labelled with 'd') as simulated by TRIFFID (Cox 2001). Some updates to the original HadCM3 are described by Valdes et al (2017). This configuration is HadCM3B-M2.1d (B—University of Bristol).
We made two sets of changes based on comparisons with palaeoclimate reconstructions for the mid-Holocene (6000 yr BP). We altered several parameters in the atmosphere (Hopcroft et al 2021) and the vegetation moisture stress function (Hopcroft and Valdes 2021). The modified atmosphere shows increased sensitivity of convection to orbital forcing (Hopcroft et al 2021). The new formulation for vegetation moisture stress prevents too much bare soil in semi-arid regions (Hopcroft and Valdes 2021). All simulations in this work use the new configuration which we label HadCM3BB-M2.1d-v1.0 (BB: Bristol/Birmingham version 1.0).
2.2. Transient Holocene scenarios
We forced HadCM3BB-M2.1d-v1.0 (HadCM3 hereafter) with Holocene variations in orbital parameters (Berger 1978), trace gas mixing ratios of CO2, CH4 and N2O from ice-cores adjusted to the AICC2012 chronology (Veres et al 2013) and land-ice and sea-level derived from ICE-6G (Argus et al 2014, Peltier et al 2015). The former are updated every timestep whereas the sea-level and ice-sheet extent are updated every 500 years and the ice-sheet elevation and land orography are linearly interpolated at each model month between the 500 year steps of the ICE-6G reconstruction. All other model boundary conditions remain constant through time at their pre-industrial settings. This setup is a continuation of the PMIP4 deglaciation transient experiments (Ivanovic et al 2016). The initial conditions are identical in each case and are taken from the 10 ka BP state of a PMIP4 deglaciation simulations with HadCM3B-M2.1 in which freshwater is routed from Eurasian ice-sheets to the Arctic ocean (Snoll et al 2022). This basic configuration is labelled Baseline.
Four additional simulations were then branched from Baseline incrementally adding solar irradiance, volcanic eruptions and anthropogenic land-use. Solar variations are reconstructed at a decadal timestep for the Holocene by Vieira et al (2012). A second simulation (+Solar) was branched from Baseline at 10 ka BP. Volcanic aerosol optical depth variations are from the HolVol reconstruction by Sigl et al (2021a, 2021b). The volcanic simulation (+Solar+Volc) was branched from +SOLAR at 9.0 ka BP. Reconstructions of past land-use are derived by combining estimates of regional human population evolution with estimated land-use per capita to produce the land-used through time (e.g. Kaplan et al 2011, Klein Goldewijk et al 2011). The resultant uncertainty increases further back in time partly because the per-capita land-use is subject to debate. Here we adopt two contrasting estimates of past land-use in order to span the current uncertainty on Holocene land-use. One simulation is performed with the HYDE v3.2 and is labelled +Solar+Volc+ALU(HYDE) (Klein Goldewijk et al 2017) and a second labelled +Solar+Volc+ALU(KK10) uses the KK10 Holocene reconstruction (Kaplan et al 2011), where the latter assumes much higher per-capita land-use back in time. Both were branched from +Solar+Volc at 8.5 ka BP. The simulations are summarised in table 1.
Table 1. Summary of the five transient Holocene model simulations with HadCM3BB-M2.1d-v1.0.
Run name | Orbit | GHG | Ice and sea-level | Solar | Volcanic | Land-use | Start date (yr BP) |
---|---|---|---|---|---|---|---|
Baseline | B78 | Ice-core | ICE-6G | — | — | — | 10 000 |
+Solar | B78 | Ice-core | ICE-6G | Vieira et al | — | — | 10 000 |
+Solar+Volc | B78 | Ice-core | ICE-6G | Vieira et al | HolVol | — | 9000 |
+Solar+Volc+ALU(KK10) | B78 | Ice-core | ICE-6G | Vieira et al | HolVol | KK10 | 8500 |
+Solar+Volc+ALU(HYDE) | B78 | Ice-core | ICE-6G | Vieira et al | HolVol | HYDE | 8500 |
a B78: Berger (1978). b Ice-core: Loulergue et al (2008), Schilt et al (2010), Veres et al (2013), Bereiter et al (2015), Köhler et al (2017). c ICE-6G: Argus et al (2014), Peltier et al (2015). d Vieira et al: Vieira et al (2012). e HolVol v1.0: Sigl et al (2021a, 2021b). f KK10: Kaplan et al (2011). g HYDE: Klein Goldewijk et al (2017).
2.3. Testing stability of the land-atmosphere simulation
We configured a suite of equilibrium snapshot simulations to explicitly diagnose bistability over North Africa. We ran paired snapshot simulations at discrete intervals over the time from 9 kyr BP to 1 kyr BP. These are initialised from respective points in the Baseline transient simulation but are configured with constant boundary conditions for ice, greenhouse gases and orbit. For each time period two simulations were initialised with fixed vegetation cover. The first ('green') is prescribed with 95% vegetation coverage (80% C4 grass, 10% C3 grass and 5% shrub) over the Sahara (12.5–35∘ N and 20∘ W–30∘ E), and the second ('brown') is prescribed with 5% vegetation coverage split evenly between C3 and C4 grasses and shrubs. Outside of the Sahara the vegetation distribution is equal in the two simulations. The vegetation cover is fixed for the first 300 model years and only the leaf area index and canopy height are allowed to evolve. Subsequently, the simulations are branched into a continuation of the fixed vegetation simulation and a dynamic vegetation branch where the distribution of plant functional types (PFTs) is allowed to interactively evolve. There are a total of four simulations for each time point: fixed 'green' and 'brown', and dynamic 'green'-initialised and 'brown'-initialised. These four branches are continued for a further 200 years. The final 30 years are used to calculate climatologies.
3. Results
3.1. Simulated rainfall response in North Africa
The simulated vegetation and rainfall averaged over North West Africa (from 20–30∘ N by 20∘ W–5∘ E) in the five transient simulations are shown in figures 1 and S1, respectively. The annual June-July-August-September (JJAS) -mean precipitation is used to identify periods of abrupt change. Here an abrupt transition is identified where the 30 year mean precipitation difference between two periods spaced by no more than 50 years exceeds 180 mm yr−1. In figure 1 the abrupt changes in precipitation are shaded in green for shifts to wetter conditions and in brown for drying phases.
Figure 1. Simulated vegetation coverage over North West Africa (20∘ W–5∘ E by 20–30∘ N) in five transient Holocene simulations. Left column: simulated 100 year mean vegetation cover. Middle column: as for the left column but showing the abrupt change around 6000 yr BP. The vertical green and brown lines indicate which years are calculated as having abrupt positive or negative changes in precipitation respectively in this region and are identical to those in figure S1. In the left and middle columns the Baseline simulation is shown in all panels (black in the top row and grey lines for the remaining rows) for comparison. Right column: the running-variance of vegetation cover using a moving window of 200 years (thin lines) and 2000 years (thicker lines). Model versions are indicated in the left column. Major volcanic eruptions are indicated by red triangles.
Download figure:
Standard image High-resolution imageIn the Baseline simulation the vegetation cover starts at around 80% and then rapidly drops to around 35% at 6100 yr BP. Subsequently both vegetation cover and precipitation gradually decline towards the present day. The abrupt change in Baseline is associated with an increase in variance in the centuries up to the cross-over point as shown in figure 1 (Hopcroft and Valdes 2021). The increase in variance is a potential indicator of threshold in the underlying system (e.g. Scheffer et al 2009). Another feature is the rapid and sustained variations in vegetation (and rainfall) in earlier years at around 7250, 6600 yr BP and the approximately 500 years preceding the final collapse. These precursor-like events suggest that the system is nearing a critical threshold.
The introduction of additional forcings in the model does not appreciably alter these basic characteristics. The remaining four simulations that introduce solar, volcanic and land-use forcing also show abrupt reductions in vegetation coverage and rainfall (figure 1). The variance of vegetation coverage increases towards 6000 yr BP in all five simulations (figure 1). The simulations also show numerous centennial-millennial phases of rapid reductions in vegetation. These, however, are not synchronous across the suite of runs. As indicated by the green and brown vertical bands in figure 1, there is diversity in frequency, duration and onset times. The introduction of solar forcing alone does increase the frequency of abrupt rainfall variations. With the inclusion of volcanic forcing (as in figures 1(g)–(i)), the frequency of these events is comparable to the simulation (Baseline) with no high-frequency forcing. Land-use has less of an obvious impact.
The final rapid collapse of the vegetation in north West Africa occurs at some point between 6500–6000 yr BP in all simulations except the KK10 run. In the Baseline simulation the collapse is perhaps most dramatic and occurs at 6100 yr BP. The collapse occurs around 100 years later with the inclusion of solar and solar and volcanic forcing. With the less intense (HYDE) land-use the collapse occurs around 100 years earlier at around 6250 yr BP. In the KK10 run there are several oscillations with wetter phases occurring until around 5250 yr BP. The more intense reconstruction of human land-use has therefore potentially helped to sustain wetter conditions in this region. Increased rainfall in western north Africa is potentially a result of the cooling over Europe arising from the increase in surface albedo due to deforestation and its effect on the stationary waves and jet structure. A more in-depth analysis of this result is planned in future work.
3.2. Vegetation threshold over North Africa
The results from the set of fixed and dynamic vegetation snapshots are shown in figure 2. These simulations are labelled 'brown' and 'green'—start, i.e. spun-up with 95% or 5% bare soil, respectively. The fixed vegetation simulations show that there is a substantial impact on the hydrological cycle caused by prescribing vegetation across the Sahara. There are two branches in figure 2(b), the upper branch of 'green' runs is substantially wetter than the 'brown' simulations. The results in figure 2 can help to explain the results in the fully transient simulation. In fully transient mode, the model starts on the 'green' branch and subsequently the precipitation gradually declines with waning summer insolation. Eventually, the system crosses a threshold which is shown by the increase in variance shown in figure 1, i.e. the system is critically slowing-down and approaching a transition (e.g. Scheffer et al 2009). Below this threshold vegetation is less likely to survive and the system 'jumps' down onto the lower 'brown' branch. The jump is equivalent to the moment in the transient simulation when the model shows an abrupt reduction in rainfall and vegetation coverage at around 6200 yr BP.
Figure 2. Absence of bistability or hysteresis for rainfall or vegetation over North Africa. Fractional vegetation cover (a) and precipitation (b) in fixed vegetation snapshot simulations initialised with 95% ('green') or 5% ('brown') vegetation coverage. (c) and (d) equivalent simulations with dynamic vegetation enabled. The averages are calculated over North Africa: 20–30∘ N by 20∘ W–5∘ E.
Download figure:
Standard image High-resolution imageThe snapshot simulations with dynamic vegetation show that there is little evidence for bistability or hysteresis over North Africa. These simulations all converge on similar vegetation and precipitation levels, irrespective of the initial conditions (brown or green). This result is robustly reproduced across the time interval studied and follows a 300 year spin-up meaning that other aspects of the climate system are close to quasi-equilibrium with the two states before vegetation dynamics are enabled.
3.3. Tipping point and system 'flickering'
The spatial precipitation anomalies associated with phases of rapid vegetation contraction and rainfall reduction in North West Africa are shown in figure 3. The anomalies are calculated using only years that pass the criteria for an abrupt reduction in rainfall as described above for figure 1. That is where the 30 year mean precipitation difference between two periods spaced by no more than 50 years exceeds 180 mm yr−1. These events are indicated with brown lines in figure 1. We took the JJAS seasonal-mean for the wettest 25% of years shaded brown and subtracted the driest 25% to give the spatial precipitation anomaly associated with the flickering events. We only used events that occur between 8 and 6 kyr BP.
Figure 3. Spatial distribution of JJAS-mean precipitation change (mm day−1) during abrupt drying events. The anomalies are calculated as the difference between the wettest and driest quartiles averaged across the transitions shaded in brown in figure 1 that occur between 8 and 6 kyr BP. Stippling indicates where the anomalies are statistically significant (at the 95% level) according to a non-parametric Kolmogorov-Smirnov test.
Download figure:
Standard image High-resolution imageThe anomalies are concentrated over North West Africa but extend to the eastern coast of Africa at around 10∘ N. A long palaeoclimate record of water availability from the Chew Bahir basin in southern Ethiopia shows very similar apparent 'flickering' of the hydrological cycle in the run up to the end of the Holocene Africa Humid Period (Trauth et al 2015). The model-data agreement in terms of 'precursor' events provides a potentially important validation of one of the key predictions of these simulations. If it can be verified with other records, it would mean that the threshold properties of the model (e.g. the increasing variance and the results of the fixed and dynamic vegetation simulations) may be representative of the natural system.
Other evidence for wet conditions in East Africa during the early to mid-Holocene are not captured by the model. Savannah grassland biomes probably extended up to around 26∘ N (Jolly et al 1998, Harrison 2017), consistent with other palaeo-hydrological evidence (Tzedakis 2007). In the model bare soil instead dominates the area from 18∘ N. The simulated footprint of the humid phase is therefore biased towards West of the continent. If in reality there were 'precursors' drying events, their spatial footprint would probably also have extended further East than simulated with this model.
3.4. Terminations: role of initial conditions
Figure 4 shows how initial conditions in the ocean and atmosphere affect the evolution of the termination of the humid phase in the model. Here the Baseline simulation was branched at 6500 yr BP with six different sets of initial conditions in both the atmosphere and ocean (figure 4(a)) and with five different sets of atmosphere conditions (figure 4(b)). The differing starting conditions were taken from the remaining simulations shown in figure 1 with different forcing, along with a sixth in-progress simulation that include solar-ozone forcing that is not otherwise discussed. These provide a diverse set of initial conditions. The timeseries are aligned by the final abrupt rainfall drop using the same criteria as applied in figure 1. The non-aligned timeseries are shown in the supporting information figure S2. One ensemble member does not meet the criteria for alignment and so this simulation is only shown in the supporting information.
Figure 4. Simulated JJAS-mean precipitation over North West Africa (20–30∘ N by 20∘ W–5∘ E) from 6.5 to 5.5 kyr BP in an ensemble of perturbed initial condition simulations branched from Baseline at 6.5 kyr BP. (a) Initialised with different atmosphere and ocean states; (b) initialised with different atmosphere states. The simulations are aligned in time with respect to the main abrupt event. The label identifies the initial conditions used (ini-), where for example, ini-0 indicates the original initial conditions as shown in the top panel of figure 1, and ini1-5 are perturbed initial conditions.
Download figure:
Standard image High-resolution imageThe initial conditions do not show a particularly strong influence on the precipitation response. The difference between initialising all simulations with identical ocean conditions and those with different ocean starting points is limited. There is some evidence for stronger multi-decadal variability in the ensemble with varying ocean conditions. However, all simulations show similar abrupt behaviour and oscillations between wetter and much drier regime before the collapses. The timing of the abrupt collapses also varies considerably across the ensemble. These results suggest that simulated positive feedbacks in the region are dominant and that external conditions, whilst important are not the controlling factors on the timing of the abrupt changes.
3.5. Comparison with Holocene climate reconstructions
We compare simulated precipitation with palaeoclimate records from four sites around North Africa. We use leaf-wax hydrogen isotope records from a marine core in the Gulf of Aden located at 12∘ N, 44∘ E (Tierney and deMenocal 2013), a core in Lake Bosutwami (6∘ N, 1∘ W) (Shanahan et al 2015), a hydrological index from the West African coast (20∘ N, 18∘ W) (Tjallingii et al 2008) and reconstructed precipitation averaged over four marine cores located between 18 and 31∘ N also near the West African coast (Tierney et al 2017). We compare these timeseries with equivalent regional averages of JJAS simulated precipitation in the five model simulations in figure 5.
Figure 5. Simulated Holocene precipitation and reconstructed precipitation or hydrological change in North Africa from four locations: (a) East Africa (Gulf of Aden: 12∘ N, 44∘ E) (Tierney and deMenocal 2013), (b) North West Africa (18–31∘ N, coastal) (Tierney et al 2017), (c) West Africa (20∘ N, 18∘ W) (Tjallingii et al 2008), and (d) central West Africa (6∘ N, 1∘ W) (Shanahan et al 2015). The core locations and the masked regions in the model used for comparison are shown in the map in the top right. The masks applied to the model in panels (a) and (c) overlap.
Download figure:
Standard image High-resolution imageThe results show good agreement in the timing of rainfall decline in North West Africa. As discussed by Hopcroft and Valdes (2021), three of the four reconstructions by Tierney et al (2017) show a nearly synchronous, abrupt decline in precipitation which is synchronous with the simulated precipitation decline in North West Africa. The inclusion of additional forcings from solar, volcanic and land-use do not alter this model-data agreement substantially.
Further east, the model reproduces the trends in the leaf-wax hydrogen isotopes, showing a mostly precessional signal with a gradual reduction in precipitation. The reconstruction shows an abrupt drying event at around 4000 yr BP, which is not reproduced by the model. Similar centennial to millennial variability is seen in the Lake Bosumtwi record at around 6000 yr BP as in the East Africa core, and this is also not seen in any of the model simulations or in the palaeorecords from West Africa. The model-data discrepancy here hints at a regional signature that may originate from other feedbacks or forcings. The precipitation decline is well-matched in terms of timing in the Western and Eastern sites (Tierney and deMenocal 2013, Tierney et al 2017), although the record by Tjallingii et al (2008) shows a later decline than in these records or in the simulations.
Further south the simulated precipitation declines later figure 5(d) but it appears too early when considering the records by Tjallingii et al (2008) and especially the Shanahan et al (2015) as shown in figures 5(c) and (d). It is unclear how the regional differences in timing should be interpreted. The differences across the four selected records probably exceeds what could reasonably be expected given that the monsoon is generally relatively spatially coherent. Moreover, when compared to a larger database of hydrological reconstructions, the model resolves the timing of the end of the AHP very well (Hopcroft and Valdes 2021). The differences across the region require further investigation. If the spatial difference in the timing is real, then this may require more detailed representation of the land-surface in the model—e.g. with many more PFTs to represent a more spatially incoherent response, as discussed further below.
4. Discussion
The transient response of the model (figure 1) is consistent with a threshold in the system and is not associated with inherent bistability or hysteresis. The threshold most probably derives from the dependence of vegetation on moisture availability as illustrated in figure 2(b) by the differences in precipitation between the 'green' and 'brown' states. The switching between the upper and lower branches provides an effective explanation for the transient model results. As a region crosses the threshold it 'jumps' from the upper green branch down to the lower brown one.
The realism of the simulated threshold is difficult to evaluate. The good agreement with some palaeoclimate records (e.g. figure 5) could be taken as evidence in favour of these properties displayed in the model. However, there are simplifications in the land-surface scheme that need to be addressed in future work. The limited number of PFTs in HadCM3-M2.1 may bias the model towards abrupt transitions (Claussen et al 2013, Groner et al 2015, 2018). A fuller representation of PFTs could smooth out transitions for example. Dynamic disturbance due to fire is not represented. Fire probably has an important role in the structure of vegetation in this region (Lu et al 2018) and so its response to climate changes. In one ESM, the addition of dynamic fire disturbance was shown to induce hysteresis behaviour of vegetation cover (Lasslop et al 2016). In our simulations the absence of fire may artificially dampen abrupt responses. Future work should evaluate how an expanded set of PFTs with dynamic fire could shape the response of the end of the Africa Humid period.
The spatial extent of the simulated humid phase is largely controlled by the response of vegetation to the North and a teleconnection that forces subsiding air over the Eastern Mediterranean (Rodwell and Hoskins 1996). The representation of this monsoon-desert teleconnection varies between GCMs (Cherchi et al 2014) but this probably explains why some of this region remained dry throughout the Holocene (Claussen and Gayler 1997, Hopcroft and Valdes 2021). However, the model may not achieve a realistic balance of competing influences in this region, meaning that the area that remains dry is too large and the simulated hydroclimatology of this region is overly stable. Potential reasons for this could include the structural aspects of the model (e.g. resolution or large-scale circulation biases) or the choice of parameter values. If the latter is the case it may be possible to further refine the simulation of the Africa Humid Period. The model parameter tuning of HadCM3 used here only targeted the regional average rainfall over the Sahara (Hopcroft et al 2021). A more refined approach could account for the spatial distribution of what is known about the Holocene expression of the AHP (e.g. Harrison 2017, Tierney et al 2017, Cheddadi et al 2021).
5. Conclusions
The greening and subsequent aridification of the Sahara during the Holocene has proven an especially challenging target for climate and ESMs over several decades (Braconnot et al 2021). Model simulations that cannot reproduce this past regime may also fail to reliably predict future change in monsoon regions. It has been clear for several years that climate and ESM responses can be sensitive to the structure and parameter values within certain parameterisations (Murphy et al 2004, Stensrud 2007) and these can therefore be optimised to better capture past events like the Sahara 'greening'. Our palaeo-conditioned ESM is found to satisfy many aspects of the Holocene AHP including wetter and greener conditions in North west Africa with relatively rapid shifts to a drier regime punctuating this. The modelled threshold is mostly confined to the North and West of the continent in agreement with some palaeorecords. Finally the model produces rather gradual changes elsewhere.
Overall it appears that this threshold-determined model provides a relatively powerful representation of the natural system. It highlights that coupled global models are able to produce abrupt climate changes comparable to real palaeoclimate events. Future work is needed to refine the understanding of the response in the east of the continent and to evaluate the vegetation feedback in more comprehensive models. High-resolution records that can constrain either the seasonality (e.g. Cheddadi et al 2021) or the decadal-to-centennial variability over several centuries towards the end of the AHP will be particularly useful for this.
Acknowledgments
P O H is supported by a University of Birmingham fellowship. P O H is grateful for fruitful discussions with Paul Wilson, Anya Crocker, Chuang Xuan, Gilles Ramstein and Martin Claussen. This paper is TiPES contribution 167: P J V has received funding from the European Union's Horizon 2020 research and innovation programme under Grant Agreement No. 820970.
Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: https://www.paleo.bristol.ac.uk/ummodel/scripts/papers/. Data will be available from 01 May 2022.
Data and code availability
Model simulations were run on the University of Birmingham's BEAR HPC facility www.bear.bham.ac.uk. All model output is available for further analysis from www.paleo.bris.ac.uk/simulations. Analyses were performed with the NCAR Command Language http://dx.doi.org/10.5065/D6WD3XH5.
The Met Office released the HadCM3 source code via the Ported Unified Model release (www.metoffice.gov.uk/research/collaboration/unified-model/partnership, and um_collaboration@metoffice.gov.uk). The main repository for the Met Office Unified Model (UM) version presented here can be viewed at http://cms.ncas.ac.uk/code_browsers/UM4.5/UMbrowser/index.html. Code modifications for the HadCM3B configuration as described by Valdes et al (2017) are available from https://doi.org/10.5194/gmd-10-3715-2017. The version of the model used here (HadCM3BB) requires updates to the atmosphere dx.doi.org/10.6084/m9.figshare.12311360 (labelled REVopt) and vegetation schemes https://doi.org/10.6084/m9.figshare.13650062.v1.
The forcings used in HadCM3 are available as follows: the solar irradiance data: https://doi.org/10.17617/1.5U, the Holocene volcanic eruptions optical depth: https://doi.pangaea.de/10.1594/PANGAEA.928646, the KK10 Kaplan et al (2011) land-use: https://doi.org/10.1594/PANGAEA.871369. and the HYDE v3.2 land-use: https://doi.org/10.17026/dans-25g-gez3.
The palaeoclimate records used are available as follows: Tjallingii et al (2008): https://doi.org/10.1594/PANGAEA.705111, Tierney and deMenocal (2013): www.ncdc.noaa.gov/paleo/study/15537, Tierney et al (2017): www1.ncdc.noaa.gov/pub/data/paleo/contributions_by_author/tierney2017/, and Shanahan et al (2015): http://ncdc.noaa.gov/paleo/study/18355.