1 Introduction

The West African Monsoon (WAM) system is the main conveyance of the region's annual rainfall, with the maximum contribution occurring during the summer months of June–September (Diallo et al. 2014, 2016; Akinsanola et al. 2015, 2016; Sylla et al. 2015; Klutse et al. 2016; Akinsanola and Zhou 2018). Its seasonality largely influences many of the socio-economic activities over the West Africa (WA) region, such as rain-fed agriculture, hydroelectric power generation, and water resource management (Lebel et al. 2003; Odoulami and Akinsanola 2017). Although this region is identified as a climate change hotspot in Africa (Martin and Thorncroft 2013), there is no consensus amongst various ensembles of global and regional climate models on the future changes in rainfall, in terms of both magnitude and, sometimes, even sign (e.g. Cook and Vizy 2006; Dosio et al. 2020, 2021a). Therefore, assessing the plausibility of modeled rainfall is crucial for establishing the reliability of projections (James et al. 2018; Doblas-Reyes et al. 2021). In addition, understanding the reasons behind models’ biases can be beneficial for the improvement of models’ physical and dynamical formulations.

Broadly, the WAM consists of a seasonal change in the atmospheric circulation dynamics, inducing strong temporal and spatial modifications in the precipitation’s pattern. This reversal in atmospheric circulation begins when the African land (Sahelian region) warms while the moist regions along the coasts of the Guinea Gulf cool down. This thermal contrast (and the consequent pressure gradient) raises moisture-laden winds from the Gulf of Guinea. Large thunderstorm systems develop when these moist winds move up onto the overheated African continent. These storms are the only annual recharge of water for the countries bordering the Gulf of Guinea and the Sahel. Moisture in the Gulf of Guinea, known as the monsoon flux, originates from southeasterlies pumped from the South Atlantic high-pressure system, and which recurve into southwesterlies while crossing the equator. The eastern Atlantic also feeds the WA via the West African Westerly Jet (WAWJ) around 10°N (Pu and Cook 2012).

The WAM is the result of a multitude of complex physical processes and mechanisms acting at diverse scales. These include remote or large-scale forcing factors such as the anomalous sea surface temperature (SST), the Pacific Decadal Oscillation (PDO), the El Niño-Southern Oscillation (ENSO) (e.g. Zhang and Zhou 2011; Diatta and Fink 2014); drivers at regional scale such as the WAWJ, African Easterly Jet (AEJ), Tropical Easterly Jet (TEJ) and African Easterly Waves (AEWs) (e.g. Cook 1999; Nicholson and Grist 2003; Akinsanola et al. 2017; Bercos-Hickey and Patricola 2021); finally, local forcings such as the surface vegetation and continental land surface conditions (e.g. Wang and Eltahir 2000; Clark et al. 2001; Sylla et al. 2016). The role of aerosols emanating from the Saharan and Sahelian regions on the WAM variability has been also established (Marcella and Eltahir 2014; N’Datchoh et al. 2018; Doblas-Reyes et al. 2021).

Numerous studies have attempted, with varying degrees of success, to model the WAM system in order to understand future changes in monsoon rainfall. Some of them are based on global climate models (GCMs) participating in the Coupled Model Intercomparison Project (CMIP-3 to 6; Meehl et al. 2007; Taylor et al. 2012; Stouffer et al. 2017), and on their multi-model ensembles (e.g. Kamga et al. 2005; Diallo et al. 2012; Monerie et al. 2012, 2021; Biasutti, 2013; Seth et al. 2013; Mariotti et al. 2014; Watterson et al. 2014; James et al. 2015). However, the GCMs’ coarse grid resolution poses serious challenges to the simulation of mesoscales and local processes that modulate the monsoon system including, for instance, organized convection, land–atmosphere-ocean interactions, surface temperature gradients, AEWs, and soil moisture (Koster et al. 2004; Cook and Vizy 2006; Watterson et al. 2014; Birch et al. 2014; James et al. 2018). Several other works (e.g. Sylla et al. 2015; Akinsanola et al. 2015; Akinsanola and Zhou 2018; Diallo et al. 2014, 2015, 2016; Klutse et al. 2016; Dosio and Panitz 2016; Dieng et al. 2017; Dosio et al. 2020) used regional climate models (RCMs) to dynamically downscale the results of CMIP5 GCMs onto a finer (~ 50 km) horizontal grid, within the framework of internationally coordinated activities such as the COordinated Regional climate Downscaling EXperiment (CORDEX; Giorgi et al. 2009) project, and the West African Monsoon Modeling and Evaluation (WAMMEBoone et al. 2010; Druyan et al. 2010; Xue et al. 2016) project. The downscaling is supposed to enhance the representation of smaller-scale physiographic processes while also adding local topographical information that is missing in GCMs (Giorgi and Gutowski 2015; Dosio et al 2019). In particular, the ability of RCMs to add value to the driving GCM in simulating precipitation characteristics (especially higher-order statistics) has been shown in several studies (e.g. Pinto et al. 2016; Fotso-Nguemo et al. 2017; Gibba et al. 2019) although the added value is not always clear (see e.g. Dosio and Panitz 2016). In addition, differences between model results may stem from the different configuration and parameterizations (see e.g. Sørland et al. 2021) and the “improved” precipitation field may be the result of wrongly simulated mechanisms (Tamoffo et al. 2020). Recently, a new generation of higher-horizontal resolution (~ 25 km) simulations has been developed within the CORDEX-Coordinated Output for Regional Evaluations (CORDEX-CORE) project (Gutowski et al. 2016). CORDEX-CORE results are designated to be utilized jointly across regions of the world for the study of regional climate processes and for the analysis of the impact of climate change on different sectors, including vulnerability assessment and adaptation options (Giorgi et al. 2021).

Some studies have already analyzed the CORDEX-CORE results over Africa. For instance, Gnitou et al. (2021) and Ilori and Balogun (2021) investigated the performances of CORDEX-CORE RCMs in simulating seasonal and annual precipitation over WA; however, these studies were merely descriptive and did not investigate the causes of RCM biases. Ashfaq et al. (2020) provided an assessment of future characteristics of monsoon systems, including the West and East African monsoon, but results were based on a single CORE RCM. Teichmann et al. (2020) analyzed the mean climate change signal at a global scale from CORE and the CMIP5 GCMs, and Coppola et al. (2021) provided an analysis of projected impact-relevant indices based on CORE, CORDEX, CMIP5 and CMIP6 runs. Finally, Dosio et al. (2021a) compared present and future precipitation daily characteristics from the available CMIP, CORDEX and CORDEX-CORE simulations over Africa. However, these studies did not investigate the mechanisms behind differences across models’ ensembles. In addition, none of these studies specifically investigated the ability of CORDEX-CORE RCMs to reproduce the WAM system, and, in particular, the mechanisms responsible for the models’ biases (as done for Central Africa, by e.g. Tamoffo et al. 2021). Although the methodology is similar, given the large uncertainty in models' performances and projections over WA and where models do not agree even on the sign of future precipitation change, (see e.g. Dosio et al. 2020) a process-based analysis of RCMs focusing on WA is a very timely and critically needed.

To this end, this study aims to thoroughly analyze the representation of the WAM system in two CORDEX-CORE RCMs in order to highlight the causes of models biases by means of a process-oriented evaluation approach (James et al. 2018). The goal of the work is to highlight misrepresented physical processes associated with models’ rainfall biases so as to provide guidance for their improvement. The plausibility of modeled rainfall climatology is therefore questioned in order to stimulate discussion on the fitness for purpose of RCMs, especially when their results are used to develop society's responses to the disastrous consequences of climate change. The paper is structured as follows. Section 2 describes the data and methods. The rainfall climatology of RCMs is characterized in Sect. 3. Section 4 examines how models simulate the land–atmosphere-ocean feedback. Section 5 looks at how models of the atmospheric circulation react to the land processes. The final section provides a summary and conclusions.

2 Data and methods

2.1 Modeled and observed data

In this work, two CORDEX-CORE RCMs were used: the REgional MOdel version 2015 (REMO2015; Jacob and Podzun 1997; Jacob 2001; Remedio et al. 2019) and the fourth generation of the Regional Climate Model (RegCM4-v7; Giorgi et al. 2012). Although currently, four RCMs participated in the CORDEX-CORE project, only the two models used here provided results for the entire atmospheric column necessary for the process-based evaluation. The models are run over the CORDEX-Africa domain (AFR-22) at 0.22° × 0.22° horizontal resolution, using 27 vertical layers for REMO2015 and 23 for RegCM4-v7. Other details of the models’ configurations are listed in Table 1. Two sets of simulations are performed: the ‘evaluation’ runs (1981–2010), i.e. when RCMs are driven by ERA-Interim reanalysis (Dee et al. 2011), and the ‘historical’ runs (1970–2005), i.e. using the results of CMIP5 GCMs as lateral boundary conditions. It should be noted that, following the CORDEX-CORE protocol, SSTs are prescribed by the driving GCMs (or reanalysis), and are not modified by RCMs, as they do not include an ocean model.

Table 1 Detailed configuration of CORDEX-CORE RCMs used in this study, as described in Teichmann et al. (2020)
Full size table

The purpose of the evaluation runs is to distinguish systematic biases (RCM internal errors) from those inherited from large-scale boundary forcing, in a “quasi-perfect forcing mode” (see e.g., Doblas-Reyes et al. 2021). The historical runs are driven by three CMIP5 GCMs, namely HadGEM2-ES, MPI-ESM-LR/MR, and NCC-NORESM1 (see more details about driving GCMs in Table 2).

Table 2 Details of the driving GCMs and names of RCM experiments used in this study
Full size table

Numerous observational datasets, including gauge-based, satellite-derived, and reanalysis products have been included in the analysis (Table 3). In fact, observed data feature large discrepancies over many regions in Africa. Recently, Dosio et al. (2021b) demonstrated the impossibility of using a single ‘best’ performing observational dataset to realistically represent rainfall features across Africa. Furthermore, they argued that assessing model results based on a very limited observational dataset may not be only pointless but can be even misleading.

Table 3 Description of the reanalysis and the satellite/gauge datasets employed for the evaluation in this study
Full size table

For all datasets, variables utilized include precipitation, sea surface temperature (SST), mean sea level pressure (MSLP), evaporation, zonal and meridional wind components (u,v), vertical wind velocity (ω), specific humidity (q), surface pressure (SP), air temperature (ta), and geopotential height (z). All variables are available at the monthly time scale from 1980 to 2005, except ARC2 and CHIRPS2 precipitation, which is available on a daily basis for 1983–2005 and 1981–2005 respectively. The meridional wind component (v-wind) at a daily time scale is also used to estimate the models’ representation of the AEWs.

2.2 Methods

The climatology of WA (5°-20°N; 18 W°-16°E, Fig. 1) consists of a unimodal rainfall regime during the monsoon season with the precipitation peak occurring in August. Although precipitation over the coasts of the Gulf of Guinea follows a bimodal regime (Dosio et al. 2021b), here we focus on the season July–August-September (JAS), which provides the largest precipitation amount over the region, especially the Sahel, and which is the prevailing period of the WAM (Nicholson 2013).

Fig. 1
figure 1

Long-term mean (1983–2005) rainfall biases in the reanalysis datasets and RCM simulations (units: mm.day−1). The CHIRPS2 dataset is used as reference. Stippling highlights the grid points where the rainfall bias is statistically significant at the 95% confidence level using the Student’s t-test. The red and black boxes denote the Guinea Coast (18°W-16°E, 5°N-10°N) and the Sahel region (18°W-16°E, 10°N-20°N), respectively. The West Africa (WA) region (5°-20°N; 18 W°-16°E) is that resulting from the combination of the black and red boxes

Full size image

First, the ability of RCMs to represent the observed spatial pattern and seasonality of rainfall is evaluated. Analyses are performed separately for the region along the coasts of the Gulf of Guinea (hereafter, Guinea Coast 18°W-16°E, 5°N-10°N; red box in Fig. 1) and the Sahel (18°W-16°E, 10°N-20°N; black box in Fig. 1). To understand the reasons behind RCMs' rainfall biases, we also evaluate the ability of the models to simulate a number of well-known drivers of the WAM system. These include land-sea interactions through SST, MSLP, SP variabilities and moisture transport, the dynamics of upper- and mid-tropospheric easterly jets (TEJ and AEJ), low-level westerly flows (monsoon flux), AEWs, and convection (Cook 1999; Nicholson and Grist 2003; Patricola and Cook 2007; Pu and Cook 2012; Vizy and Cook 2017; Akinsanola et al. 2017; Bercos-Hickey and Patricola 2021).

To associate rainfall biases with the atmospheric circulation, the JAS vertically integrated mass-weighted moisture transport is examined using the following equation:

$$QV=\frac{1}{g}{\int }_{sp}^{Ptop}qVdp$$
(1)

The moisture transport QV can be separated into zonal (Qu in Kg.m−1.s−1) and meridional components (Qv in Kg.m−1.s−1). Similarly, the total wind speed V (in m.s−1) is separated into its zonal u and meridional v components. g is the gravitational acceleration (in m.s−2), q is the specific humidity (in g.Kg−1), SP is the surface pressure and Ptop is the pressure of the top-level (in Pa).

To highlight sources of moisture surplus or deficit, transient regional moisture convergence or divergence across each WA border and at each pressure level was estimated. In a given region, the total moisture convergence (divergence), i.e. the sum of inwards (outwards) moisture across the borders, can be split into the zonal (West–East; Qλ) and the meridional (South-North; Qφ) directions as follows (see e.g. Zheng and Eltahir, 1998):

$${Q}_{\lambda }=\int qud\lambda $$
(2)

and

$${Q}_{\phi }=\int qvd\phi $$
(3)

where λ are longitudes and φ latitudes. The region can be considered as a rectangle of length X and width Y, X and Y being a set of segments dl through which the atmospheric moisture flows inwards and outwards. Thereby, moisture fluxes flowing across western (Qλwest), eastern (Qλeast), southern (Qφsouth), and northern (Qφnorth) boundaries can be expressed respectively as follows:

$$ Q_{\lambda west} = \int_{lwest} Q_{\lambda } dl{\text{ and Q}}_{\lambda east} = \int_{least} Q_{\lambda } dl $$
(4)
$$ Q_{\phi south} = \int_{lsouth} Q_{\phi } dl{\text{ and Q}}_{\phi north} = \int_{lnorth} Q_{\phi } dl $$
(5)

where

dl is the distance between two grid points computed as follows:

$$dl=\theta \times \frac{\pi }{180}\times R$$
(6)

with θ being the dataset’s horizontal resolution in the zonal or meridional direction, and R (in m) the Earth’s radius.

The integration limits are defined as.

$$ l_{west} = 18^{\circ} W, l_{east} = 16^{\circ}, l_{south} = 5^{\circ}N \; {\rm and} \; l_{north} = 20^{\circ} N. $$

Within the integration limits, the net zonal (QZ), meridional (QM), and total (QT) column of the regional moisture convergence or divergence are estimated as:

$$ Q_{Z} = Q_{\lambda west} - Q_{\lambda east} {\text{ and Q}}_{M} = Q_{\phi south} - Q_{\phi north} $$
(7)
$${Q}_{T}={Q}_{Z}+{Q}_{M}$$
(8)

As expressed in Eq. 7, positive (negative) values of \({Q}_{Z}\) mean that inward (outward) moisture flux through the western (eastern) border 18°W (16°E) is greater than outward (inward) moisture flux through the eastern (western) border 16°E (18°W), hence the zonal moisture flux balance \({Q}_{Z}\) is convergent (divergent). Likewise, positive (negative) values of \({Q}_{M}\) indicate that the amount of moisture flux entering (exiting) the region via the southern (northern) boundary 5°N (20°N) is greater than that exiting (entering) across the northern (southern) frontier, hence the meridional moisture flux balance \({Q}_{M}\) is convergent (divergent).

For identifying the AEJ and TEJ, we filtered the zonal component of the wind field and retained only the u-wind greater or equal to 6 m.s−1 from the mid- to upper-troposphere following Nicholson and Grist (2003). The AEJ lies within the layers 700–600 hPa whereas the TEJ interacts around 200 hPa. For tracking the AEWs, the daily meridional wind component at 700 hPa is employed (e.g. Diedhiou et al. 1999). The signals of the atmospheric disturbances associated with the AEWs’ activity are highlighted using a 2–10 day passband filter applied over the time series during the JAS season. The variance of the filtered result is then computed at each grid point to approximate the mean seasonal AEWs' activity. The large 2–10 day temporal band is used to account for the two kinds of AEWs featuring the WA and namely 3–5 and 6–9 day period waves (Diedhiou et al. 1998; Wu et al. 2012) as also done in previous studies by Patricola et al. (2018) and White and Aiyyer (2021).

Finally, biases in convection are examined by means of two thermodynamic metrics, namely, the moist static energy (h) and the equivalent potential temperature (θe) defined as:

$$h={c}_{p}T+gz+Lq$$
(9)
$${\theta }_{e}={T}_{e}{\left(\frac{{P}_{O}}{P}\right)}^{\frac{{R}_{d}}{{c}_{p}}}\approx \left(T+\frac{{L}_{v}}{{c}_{p}}r\right){\left(\frac{{P}_{O}}{P}\right)}^{\frac{{R}_{d}}{{c}_{p}}}$$
(10)

In Eq. 9, the first two terms on the right-hand side represent the dry static energy (DSE) input; CpT is the sensible heat, gz is the potential energy and Lq is the latent static energy (LSE); Cp is the specific heat at constant pressure, T is the air temperature, g is the gravitational constant, z is the geopotential height, L is the latent heat of condensation and q is the specific humidity. In Eq. 10, Te is the equivalent temperature, P ∈ [1000 – 100 hPa], Po is the standard reference pressure (1000 hPa), Rd is the specific gas constant for air at constant pressure, Lv is the latent heat of evaporation and r is the mixing ratio of water vapor in air.

3 Characterizing RCMs’ rainfall

Figure 1 shows the biases, compared to the CHIRPS2 dataset used here as reference, in long-term (1983–2005) mean precipitation for three reanalysis products as well as the RCMs simulations (both evaluation and historical runs). It is noteworthy to point out that reanalysis and RCM biases with regards to other observational products such as CRU-TS4.04, GPCC-v8, GPCP-v3.2 and the ensemble mean of all observational products, look very similar to that of CHIRPS2 (not shown). Therefore, using any of these observational products does not significantly alter conclusions. Here, CHIRPS2 is chosen as reference because of its highest horizontal grid spacing (0.05°) which allows for minimizing errors from the interpolation process. Reanalysis products are used to evaluate the RCMs’ performances to reproduce the physical processes and mechanisms associated with rainfall biases. It is important to note that reanalyses are obtained by data assimilation from in-situ observational products using a reanalysis model. However, in situ stations are sparse and the density of the observation network varies from one region to another over Equatorial Africa. In regions where the surface measurement stations are widely scattered, the reanalysis data are more the result of the assimilation model than the observation station. Therefore, as for climate models, it is important to assess how well these reanalyses reproduce the observed climatology. This is also the reason why reanalysis products are used here for qualitative purposes (i.e. to assess if the physical processes or mechanisms identified in the reanalysis data are also present in the RCM runs) rather than quantitative (i.e. to compare the intensity or magnitude of the physical processes or mechanisms between reanalyses and RCM runs) like observational products.

All reanalysis datasets and most of the RCM runs (REMO-NorESM1, RegCM4-HadGEM2 and RegCM4-NorESM1) show a north (negative) - south (positive) dipole-like rainfall bias. MERRA and NCEP show a very similar distribution of precipitation, with overestimation over the Guinea Coast and underestimation over the Sahel. The ERA5 pattern is qualitatively similar, although the magnitude of the bias is usually smaller. On the contrary, the evaluation run REMO-ERA exhibits a north (positive) - south (negative) dipole-type rainfall bias. Other simulations (REMO-HadGEM2, REMO-MPI-LR and RegCM4-MPI-MR) simulate positive rainfall biases almost over the whole of WA. Finally, the evaluation simulation RegCM4-ERA features dry biases over most parts of WA, apart from the wet biases confined within latitudes 7°–12°N.

By comparing the spatial distribution of rainfall biases of different experiments, it emerges that both the large-scale boundary conditions from GCMs and the internal physics of RCMs co-influence the results of downscaling. The influence of the boundary conditions on the RCM results is clear, as the two RCMs simulations driven by the same forcing show similar patterns of biases (although local differences exist). In particular, both RCMs simulations driven by NorESM1 show biases very similar to those of the reanalyses, whereas those driven by MPI-LR/MR show a general overestimation of precipitation. Evaluation runs show a somehow different behavior, with a general underestimation over the Guinea Coast that is not visible in MERRA-2 and NCEP2, and only visible in ERA5 over the Western coasts and the Nigeria/Cameroon’s mountains. This topographic effect can indeed be missing in the lower resolution reanalyses (MERRA and NCEP). In fact, all RCM simulations show negative rainfall biases over the region. Notably, satellite products (such as CHIRPS2) have difficulties in retrieving precipitation over complex topography (Dosio et al. 2021b).

The effect of the RCM internal physical parameterization is also important. In fact, some similarities exist across the individual RCM simulations independently of the driving GCM/reanalysis. For instance, the REMO-ERA bias’ pattern is different from that of RegCM4-ERA although the two experiments share a common boundary condition. Also, in all experiments (apart from those driven by NorESM1) REMO shows a larger positive bias over the Sahel compared to RegCM4, which, on the contrary, tends to underestimate precipitation, especially over the central and western regions.

JAS mean precipitation over the Guinea Coast shows large uncertainties: Fig. 2 highlights that there is little consensus across both observational products and model simulations over the region (Dosio et al. 2021b). In general, over the Guinea Coast, GCM driven RCM simulations overestimate observed precipitation (compared to CRU and GPCC, with rainfall surplus within the range 0–4.5 mm.day−1. In contrast, evaluation (i.e. reanalysis driven) runs underestimate observations, especially in the region between 5° and 9°N, with rainfall deficit up to 2 mm.day−1. Over the Sahel, uncertainty in both observations and model results is smaller. However, some of the model simulations (e.g. REMO-ERA, REMO-NorESM1, RegCM4-NorESM1) are, for certain latitudes, outside the range of observations. Over the ocean (latitudes < 5°N) most models tend to greatly overestimate the precipitation, especially around the Equator.

Fig. 2
figure 2

Latitudinal distribution of JAS precipitation (mm.day−1), spatially averaged between 18°W-16°E. Observational datasets used include rain-gauge, satellite, and combined products. The data cover the period 1980–2005, apart from the ARC2 dataset, which covers the period 1983–2005, and the CHIRPS2 dataset, which covers the period 1981–2005. Model simulations from REMO2015 (solid lines) and RegCM4-v7 (dashed lines) include both the evaluation and historical runs. The shaded gray band represents the standard deviation of the observations (CRU, GPCC, NIC131-gridded, and CHIRPS2) multi-annual means, from 1983 to 2005

Full size image

Overall, the basic pattern of the WAM rainfall climatology is well captured in the current generation of CORDEX-CORE RCMs, despite the large uncertainties in both model simulations and observational products. Observational uncertainty, associated with the low density of the observation network in Africa (Nicholson et al. 2018), makes model evaluation of variables such as precipitation unreliable (e.g., Maidment et al. 2015; Tapiador et al. 2017, 2019; Dosio et al. 2021b). Instead, a more valuable method for model evaluation and formulation of physical parameterization is the examination of how models reproduce the physical mechanisms related to rainfall climatology (James et al. 2018). This aspect is investigated in the following sections.

4 Land–atmosphere-ocean interactions: SST and MSLP

Figure 3 shows the mean JAS climatology of SST over the eastern Atlantic Ocean. All reanalyses show very consistent SST patterns, with the portion of the ocean bordering the most western countries featuring the highest values of SST (0°-10°N), in correspondence with the monsoon rainband (Fig. 1).

Fig. 3
figure 3

Mean (1980–2005) seasonal SST climatology (1st row) and models’ biases (2nd and 3rd rows). The climatology is obtained from reanalysis data (HadISST, ERA5, and ERSST). The biases are computed relative to ERA5 reanalysis, from the REMO2015 and RegCM4-v7 experiments. The stippling denotes areas where the difference between the dataset under consideration and the ERA5 reanalysis dataset is statistically significant at the 95% confidence level by means of the Student’s t-test

Full size image

The evaluation simulations REMO-ERA and RegCM4-ERA feature very weakly warm SST biases over most of the ocean, although cold SST biases along the coastlines between 0° and 10°S can be noted. All the historical simulations show similar SST bias patterns, consisting of strong temperature overestimation over the whole Gulf of Guinea (more extended to the west of the ocean in the RCMs-MPI-LR/MR simulations), and negative SST biases towards the interior of the ocean. Balas et al. (2007) demonstrated the strong complexity of the relationship between SST and rainfall over western equatorial Africa, which is highly regionally and seasonally dependent. Factors such as general warming or cooling of the tropical oceans and Atlantic SST anomalies in the Gulf of Guinea are associated with precipitation anomalies over WA (Newell and Kidson 1984; Cadet and Nnoli 2007; Vizy and Cook 2001; Vizy, 2002). These authors demonstrated that positive SST anomalies enhance the evaporation over the Gulf of Guinea, which in turn, strengthens northward moisture advection into WA, despite the weakening of the meridional temperature gradient. While this process increases precipitation over the Guinea Coast, it rather reduces precipitation over the Sahel because the subsidence strengthens over the equator in order to compensate for the potential vorticity. Therefore, we next verified whether warm SST biases prescribed by driving GCMs are accompanied by increased evaporation.

Figure 4 displays the mean climatology of evaporation from ERA5 and MERRA2, together with the RCMs’ evaporation biases relative to ERA5. Apart from the evaluation run REMO-ERA, which shows negative biases over nearly the whole Gulf of Guinea, all the other simulations feature strong positive biases over the most part of the eastern Atlantic Ocean. Over land, experiments from REMO2015 show evaporation bias patterns very different from those of RegCM4. REMO-ERA and REMO-HadGEM2 simulate a north (Sahel, positive) - south (Guinea Coast, negative) dipole-like evaporation bias that is reversed in REMO-NorESM1. REMO-MPI-LR features a prevalence of slight positive biases in most parts of WA. On the other hand, most of the RegCM4 experiments show strong positive biases in the entire WA except RegCM4-ERA and RegCM4-NorESM1, which show negative biases in a small portion of the northern Sahel. Differences in the patterns of evaporation biases of the two RCMs over the Sahel might induce differences in atmospheric circulation. In fact, the underestimating of evaporation in REMO experiments may indicate amplified warming over the Sahel, owing to the limited latent heating necessary to cool land surface and to distribute heat to the atmosphere, thereby contributing to modify meridional temperature gradients despite warm biases of SST over the Guinea Gulf (Vizy and Cook 2017). On the contrary, the overestimating of evaporation in RegCM4 runs might induce opposite effects. These assumptions will be verified in Sect. 5. Notably, rainfall bias patterns are somehow similar to those of the evaporation in overall RCM experiments (see Fig. 1): in fact, the increase in rainfall causes an increase in soil moisture, which, in turn, increases evaporation because of unsaturated soils.

Fig. 4
figure 4

Mean (1980–2005) seasonal evaporation (E in mm.day−1) climatology (1st row) and E biases (2nd and 3rd rows). The climatology is obtained from the reanalysis data of ERA5 and MERRA2. The biases are computed relative to ERA5 reanalysis, from the REMO2015 and RegCM4-v7 experiments. The stippling occurs where the difference between the dataset under consideration and the ERA5 reanalysis dataset is statistically significant at the 95% confidence level by means of the Student’s t-test

Full size image

Another important process modulating the land-atmosphere-ocean feedback is the variability in MSLP and SP. In fact, the land-sea thermal contrast (∇T) induces a land-sea pressure gradient (∇P) that enhances low-level westerly and southwesterly moisture transport inland, and, eventually, precipitation (Patricola and Cook, 2007; Vizy et al. 2013). Figure 5 highlights that, except RegCM4-ERA which slightly overestimates MSLP, all other simulations underestimate it, including the South Atlantic High-pressure system (20°–30°S), in part or in full. Figure 6 reveals that although all simulations adequately reproduce the seasonality of the land-ocean temperature and pressure contrasts (∇T, ∇P), most RCMs simulate a weaker ∇T relative to ERA5 and MERRA2. Consequently, most models underestimate ∇P, especially at the monsoon onset (June–July) compared to ERA5, and during all the monsoon time (JJAS) compared to MERRA2. A direct consequence of differences in ∇T and ∇P between model runs is the difference in the amount of moisture imported into land regions, associated with the strength of the monsoon flow. In fact, weak temperature and pressure gradients weaken the monsoon flow, reducing the amount of inwards moisture, thereby suppressing convection, especially over the Sahel where moisture availability is limited. Conversely, strong land-ocean thermal and pressure gradients displace advected moisture from the ocean further into the continent, therefore weakening (strengthening) the moisture availability over the Guinea Coast (Sahel), and hence convection. Modeled biases in ∇T and ∇P might also affect the simulated West African westerly jet (WAWJ; Pu and Cook 2012), which is responsible for moving moisture from the eastern Atlantic onto the WA. This is discussed in the following section.

Fig. 5
figure 5

Mean (1980–2005) seasonal SLP climatology (in the top row) and SLP biases (in the middle and bottom rows). The climatology is obtained from reanalysis data (ERA5, MERRA-2, and NCEP2). The biases are computed relative to ERA5 reanalysis, from REMO2015 and RegCM4-v7 experiments. The stippling occurs where the difference between the dataset under consideration and the ERA5 reanalysis dataset is statistically significant at the 95% confidence level by means of the Student’s t-test

Full size image
Fig. 6
figure 6

Seasonality of the near-surface a land–ocean temperature difference (thermal contrast; ∇T; K) and b land-surface pressure and ocean sea-level pressure difference (∇P; hPa) between the interior of the continent (15°W-16°E, 10°-30°N) and the southeastern Atlantic Ocean (15°W-16°E, 0°–20°S), for the ERA5 (red), MERRA-2 (black), and NCEP2 (blue) reanalysis data, and for the REMO2015 (solid lines) and RegCM4-v7 (dashed lines) experiments, over the period 1980–2005

Full size image

5 The response of the atmospheric circulation

We first examine how RCMs simulate the regional atmospheric column moisture availability (Fig. 7). The total column moisture convergence/divergence (QT) is separated into the zonal (QZ) and meridional (QM) components to comprehensively identify sources of biases. Reanalyses and RCM runs agree on the shape of the vertical profile of moisture content in all components, but differences in intensity remain. Uncertainties are larger for QZ than QM in mid-layers. In the bottom layers, reanalysis products and RCMs show large variations in all components of the convergence/divergence. For instance, all models overestimate moisture convergence for QZ between 950 and 700 hPa compared to ERA5 (and partly NCEP2). In particular, some RCMs do not show any divergence in the entire column. REMO-ERA shows a profile of QZ very close to that of ERA5 between 925 and 700 hPa, whereas RegCM-ERA shows very small values of convergence/divergence for the entire column, similar to MERRA2. In the QM component, REMO runs and the NCEP2 reanalysis are closer along the tropospheric column compared to other reanalysis datasets. They feature moisture convergence from the bottom layers up to 700 hPa and slight moisture divergence between 700 and 600 hPa. RegCM4 simulations show stronger moisture convergence at 850 hPa relative to all reanalyses and REMO simulations, and become similar to REMO and reanalyses in other layers, apart from RegCM4-ERA which models moisture divergence around 850 hPa. Therefore, strong moisture convergence in QT in lower layers (1000–850 hPa) is a result of both QM (strong) and QZ (weak). QT is divergent in mid-layers (800–500 hPa), as a result of divergence in QZ (strong) and QM (weak).

Fig. 7
figure 7

Vertical profile of the regional moisture convergence/divergence (kg.m−1 s−1) across the WA, divided into the (left) zonal (QZ) and (centre) meridional (QM) directions. (right) The total column moisture convergence/divergence (QT); MD = moisture divergence (negative values) and MC = moisture convergence (positive values). Data used are from reanalyses ERA5 (black), MERRA-2 (blue), and NCEP2 (red), and from experiments REMO2015 (solid lines) and RegCM4-v7 (dashed lines) over the period 1980–2005

Full size image

To understand the reasons behind these differences, the ability of models to properly distribute moisture across WA is assessed in Fig. 8. Results show that stronger low-level QZ is associated with stronger modeled WAWJ and low-level westerlies over the coast of the Gulf of Guinea. Also, a stronger low-level QM is related to a large amount of southeasterlies turning into southwesterlies while crossing the Guinea Gulf. However, patterns of the zonal integrated moisture transport (Qu) reveal divergences between simulations. In fact, REMO-ERA is close to reanalysis data with a widespread area of positive Qu covering the entire Guinea Coast and the Sahel. The RegCM4-ERA Qu pattern is instead similar to those of the historical simulations, consisting of higher positive but less widespread Qu values in the Sahelian region. It is worth noting that some simulations (e.g. REMO-HadGEM2, RegCM4-MPI-MR) that feature precipitation wet biases over the Sahel (Fig. 1), show total moisture transport and Qu similar to those featuring dry biases (e.g. RegCM4-ERA). Also, reanalysis products that feature strong precipitation dry biases in the Sahel, feature further inward moisture into the region, which is not the case for some experiments (e.g. REMO-NorESM1, RegCM4-NorESM1) featuring dry biases over the Sahel, but accompanied with little moisture convergence onto the region. This suggests that the low-level circulation is not the only process causing differential wetting/drying between datasets.

Fig. 8
figure 8

Mean (1980–2005) seasonal climatology of lower-layer (1000–850 hPa) zonal vertically integrated moisture flux (Qu; kg.m−1 s−1; shaded) superimposed with vertically integrated (1000–850 hPa) total moisture transport (QT; kg.m−1 s−1; vectors). Data used are from reanalyses (ERA5, MERRA-2, and NCEP2) and from the REMO2015 and RegCM4-v7 experiments. Negative values indicate easterly flows and positive values are westerly flows

Full size image

We, therefore, investigated other processes driving the WAM system. The African Easterly Jet (AEJ) and the Tropical Easterly Jet (TEJ) are of great importance in modulating moisture transport in mid- (700–600 hPa) and upper (around 200 hPa) tropospheric layers respectively. The AEJ modulates the development of the rainy season by increasing/decreasing the number of mesoscale convective systems (MCSs). This is also associated with a strong/weak baroclinic instability over the Sahel (Nicholson and Webster 2007). In turn, rainfall feedback on the jet through strengthening or weakening in the meridional surface temperature gradient; the strong TEJ can accentuate the convective activity by increasing upper-level divergence, thereby favoring ascent motions in the bottom layers that leads to high rainfall over the Sahel (Grist and Nicholson 2001; Nicholson and Grist 2003; Nicholson and Klotter 2020; Bercos-Hickey and Patricola 2021). The vertical profile of zonal wind is thereby shown in Fig. 9, where the AEJ and TEJ are also highlighted. REMO simulations are generally closer to reanalyses compared to RegCM4 ones. In fact, according to all reanalysis datasets, the mid-tropospheric AEJ shows a strong (> 14 m.s−1) core located over the Sahel, around 15°N. At the same time, the upper-tropospheric TEJ core (> 16 m.s−1) is located over the Guinea Coast (around 5°N). REMO runs simulate weaker AEJ and TEJ compared to all reanalyses, with differences ranging within 2–4 and 4–6 m.s−1 respectively. All RegCM4 runs show only a very weak AEJ, and a southward shifting and underestimated TEJ core. Two distinct processes are associated with models’ wetness and dryness over the Sahel and seem to be driven in major part by the AEJ: the poleward or equatorward shifting of AEJ and the jet strength. In fact, experiments that feature rainfall dry biases simulate an anomalous equatorward AEJ (e.g. REMO-NorESM1, RegCM4-ERA, RegCM4-NorESM1); those simulating precipitation wet biases experience a weaker and northward AEJ (e.g. REMO-ERA; REMO-HadGEM2; RegCM4-MPI-MR). This is consistent with previous findings by Nicholson and Webster (2007) who showed that an equatorward (poleward) shifting of AEJ may promote moisture divergence (convergence) from (into) the region, therefore suppressing (enhancing) convective activity due to the reduced (increased) moisture availability. The rainfall dry biases over the coasts of the Gulf of Guinea in the two evaluation experiments (REMO-ERA and RegCM4-ERA) are related to stronger (compared to all reanalyses) mid-level tropospheric easterly divergence. This process dries mid-layers and probably reduces MCSs thereby weakening the convection.

Fig. 9
figure 9

Latitude-height cross-sections of zonal wind (u-wind; m.s−1; shaded), averaged over longitudes 10°W-10°E (following Nicholson and Grist, 2003). Overlaid dashed contours indicate the mean seasonal location of AEJs (wind speed > 6 m.s−1). Data used are from the reanalyses ERA5, MERRA-2, and NCEP2 and the experiments REMO2015 and RegCM4-v7, from 1980 to 2005. The stippling occurs where the difference between the dataset under consideration and the ensemble mean of the three reanalysis products is statistically significant at the 95% confidence level by means of the Student’s t-test. The red bars define the Guinea Coast and the black bars denote the Sahel

Full size image

Figure 9 also shows that in the lower layers (> 850 hPa), the extent to which the models simulate the penetration of the monsoon flow inland differentiates the experiments that represent wet biases over the Sahel from those that represent dry biases. In the first case, the monsoon flux deeply enters the continent up to 13°N (REMO-ERA, REMO-HadGEM2, REMO-MPI-LR). Although the absence of the model outputs at 1000 hPa pressure level makes it difficult to analyze the RegCM4 runs, this would be the case for RegCM4-MPI-MR. In the second case, the strength of the monsoon flow is weak, resulting in a less inland penetration of moisture and therefore an unevenly distributed moisture across the region (RegCM4-ERA, RegCM4-HadGEM2, REMO-NorESM1 and RegCM4-NorESM1).

The coupling between the AEJ and AEWs is well known, as well as the role AEWs play in the Sahel rainfall system (Fink and Reiner 2003; Nicholson 2013). Indeed, the existence of AEWs was first attributed to a joint baroclinic-barotropic instability associated with the vertical and horizontal shear of the AEJ during the monsoon time from June to October (Nicholson and Grist 2003). Newer studies showed that AEWs may require a genesis mechanism such as convection coupled with the AEJ’s instability (Thorncroft et al. 2008; Leroux and Hall 2009). AEWs are the main precursor of the Atlantic tropical cyclone genesis and strongly modulate mesoscale deep convection (Dieng et al. 2014). We discussed previously the biases in the modeled AEJ, which may indicate biases in simulated AEWs. The way the models represent the mean seasonal climatology of the AEWs’ activity is illustrated in Fig. 10 through the variance of the 2–10 day passband filtered meridional component of the wind (note that the daily meridional wind component is not available for REMO2015 RCM, therefore, only results based on RegCM4-v7 experiments are shown). Biases in the modeled AEWs’ activity are consistent with simulated rainfall biases, which corroborate the aforementioned findings suggesting biases in the simulated AEJ as the main drivers of models’ wetness or dryness over the Sahel. In fact, when a model simulates stronger AEWs’ activity, it also overestimates rainfall (e.g. RegCM4-MPI-MR). Conversely, the simulation of weaker AEWs’ activity results in less precipitation, especially over the Sahel (e.g. RegCM4-ERA, RegCM4-HadGEM2 and RegCM4-NorESM1). The AEWs’ troughs are linked to a cyclonic convective circulation in the lower and middle layers of the troposphere. Thus, an underestimated (overestimated) AEWs' activity leads to a weak (strong) southwesterly flow and thereby a weak (strong) moisture transport to feed MCSs embedded within the AEJ (Dieng et al. 2014). We previously showed that REMO2015 experiments better represent the AEJ than RegCM4-v7 experiments (Fig. 9). This would likely suggest that REMO2015 experiments also better represent AEWs than those of RegCM4-v7, although the two RCMs have simulated a weaker AEJ-AEWs (compared to reanalyses) complex system, especially for RegCM4-v7. AEWs being the main mechanism controlling the convection over WA (Diedhiou et al. 1999), the presence of biases in AEWs may indicate biases in simulated convection.

Fig. 10
figure 10

Mean (1980–2005) seasonal climatology of the variance (in m2.s−2) of the 2–10 day passband filtered 700 hPa daily meridional component of the wind (v-wind) (1st and 2nd rows) and variance bias (3rd row). The climatology is obtained from reanalysis data (ERA5 and NCEP2) and from RegCM4-v7 experiments. The biases are computed relative to ERA5 reanalysis, from RegCM4-v7 experiments. Daily v-wind data are not available for REMO2015. The stippling occurs where the difference between the dataset under consideration and the ERA5 reanalysis dataset is statistically significant at the 95% confidence level by means of the Student’s t-test. The white contour in each plot denotes the 6 m·s−1 zonal wind contour, averaged between 700–600 hPa and is indicative of the mean location of the AEJ core

Full size image

Finally, Fig. 11 shows how convection responds to different regional atmospheric circulation. It should be noted that the term convection here refers to the wind ascent motions because, in CORDEX-CORE RCMs, the convection is parameterized and not explicitly resolved. As vertical velocity is not available from RegCM4, the analysis is presented for REMO2015 only. Results show that precipitation dry biases over the Sahel in reanalyses are linked to shallow ascents; reversely over the Guinea Coast, rainfall wet biases are results of deep convection. Also in RCM experiments, rainfall dry and wet biases are associated with shallow and deep-ascent motions, respectively. This corroborates the aforementioned findings on the way simulated atmospheric circulations modulate the convection through the variability in the moisture availability.

Fig. 11
figure 11

Height-latitude cross-section of the vertical velocity (ω in Pa.s−1; shaded), the vertical motion of meridional wind and ω (vectors), and the monthly precipitation (mm/day; red lines). Data used are from ERA5, MERRA-2 and NCEP2 reanalysis data, and from the REMO2015 and RegCM4-v7 experiments over the period 1980–2005. The vertical velocity data are not available for RegCM4-v7 experiments. As the units are Pa.s−1, negative values of omega means upward motions. The stippling indicates the grid points where the difference between the dataset under consideration and the ensemble mean of the three reanalysis products is statistically significant at the 95% confidence level using the Student’s t-test

Full size image

To understand the reasons for the convection bias in more detail we explored the modeled vertical structure of moist static energy (h, Fig. 12) and equivalent potential temperature (θe, Fig. 13). The moist static energy helps understanding how the large-scale tropical circulation interacts with the regional atmospheric moisture column, thus providing a comprehensive evaluation of how well RCMs simulate these tropical disturbances (stability/instability) that initiate convection. Results show that the way biases in modeled atmospheric stability/instability are associated with convection biases (Fig. 11) differs from one RCM run to another. This process would be also different from what is observed in reanalyses. For instance, all reanalysis products feature precipitation dry biases over the Sahel (Fig. 1): however, they feature high values of both h (Fig. 12) and θe (especially ERA5 and NCEP2; see Fig. 13) in the lower layers of the atmosphere, a sign of strong disturbances, which should lead to deep convection. However, all reanalysis datasets show shallow convection over the Sahel (Fig. 11). On the other hand, some RCMs (REMO-ERA, REMO-HadGEM2, REMO-MPI-LR) do simulate deep convection over the Sahel (Fig. 11) and high values of h and θe in the bottom layers (Figs. 12 and 13), a sign of strong atmospheric instability, but they overestimate precipitation over the region (Fig. 1). Finally, as suggested by Grist and Nicholson (2001) and Nicholson and Grist (2003), modeled convection over the Sahel may be influenced by dynamic factors (AEJ in mid-tropospheric layers) rather than thermodynamic ones. Our results indicate a need for further investigation of the relationship between atmospheric instability, convection and precipitation in both reanalysis and climate models.

Fig. 12
figure 12

Latitude-height cross-sections of the moist static energy (h; kJ.kg−1). Data used are from reanalysis data (ERA5, MERRA-2, and NCEP2) and from the REMO2015 and RegCM4-v7 experiments over the period 1980–2005. The 331-K contour of the equivalent potential temperature (θe; black line) highlights the convection band. The red lines are the latitudinal migration of the rainband (mm.day−1). The stippling occurs where the difference between the dataset under consideration and the ensemble mean of the three reanalysis products is statistically significant at the 95% confidence level by means of the Student’s t-test

Full size image
Fig. 13
figure 13

Latitude-height cross-sections of the equivalent potential temperature (θe in K). Data used are from reanalysis data (ERA5, MERRA-2, and NCEP2) and from the REMO2015 and RegCM4-v7 experiments over the period 1980–2005. The 331-K contour shows the convection band. The red lines are the latitudinal migration of the rainband (mm.day−1). The stippling occurs where the difference between the dataset under consideration and the ensemble mean of the three reanalysis products is statistically significant at the 95% confidence level by means of the Student’s t-test

Full size image

6 Conclusions and discussion

In this study, we investigated the performances of two new generation high resolution, (~ 25 km) CORDEX-CORE RCMs in simulating observed precipitation over West Africa during the monsoon season (JAS). The rainfall biases are analyzed and the physical underlying processes associated with these biases are examined to understand how well the models represent the WAM system. Drivers of the WAM system used for the models’ diagnostic include land-ocean feedback processes via SST, land surface temperature, MSLP, land surface pressure variability and evaporation; the regional atmospheric circulation such as the AEJ, AEWs, TEJ, WAWJ and finally convection and associated thermodynamic (h and θe) factors. Some RCM experiments show a north (negative/positive)-south (positive/negative) dipole-like rainfall bias, while others simulate positive or negative rainfall biases over the majority of WA. In this study, we found that the following physical mechanisms are potential drivers of precipitation biases across simulations:

  1. 1.

    Most GCM driven RCMs overestimate SST (which is inherited from the driving GCM) over the whole of the Guinea Gulf. Enhanced SST increases moisture availability through enhanced evaporation. However, RCMs show weaker MSLP over the eastern Atlantic Ocean, including the South Atlantic High-pressure system in part or in full, which can result in too little moisture advection from the ocean towards the Guinea Gulf. Inland, surface pressures are also slightly underestimated including African thermal lows, but this does not alter models' precipitation biases owing to the lack of involvement of lows’ pressure systems in the atmospheric circulation anomalies.

  2. 2.

    Biases in the modeled land-ocean temperature and pressure contrasts (∇T and ∇P) also induce biases in the simulated monsoon flux strength, leading to modifications in the latitudinal positioning of the monsoon convective system. Indeed, simulations that show rainfall wet biases over the Sahel also feature moisture being advected further inland, compared to those representing dry biases over the same area. Specifically, strong (weak) ∇T and ∇P induce more moisture to be advected over the Sahel (Guinea Coast) through southeasterlies recurving into southwesterlies and via the West African westerly jet (WAWJ).

  3. 3.

    The AEJ plays an important role in models’ rainfall biases, especially over the Sahel: indeed, RCM simulations that feature precipitation dry biases over the Sahel, simulate an equatorward shifting of the AEJ, promoting strong moisture divergence from the region, thereby drying mid-tropospheric layers and suppressing convection. Contrastingly, simulations with precipitation wet biases over the same domain exhibit a poleward (northward) shifting and weaker AEJ, conditions that favor important amounts of moisture convergence into the region, therefore moistening mid-tropospheric layers and increasing convection. These results adequately corroborate the known rainfall-AEJ coupling process over the Sahel (Nicholson and Grist 2003; Bercos-Hickey and Patricola 2021). Although rainfall biases are interpreted here as being induced by that of related AEJ, the reverse may be also true, owing to feedback in processes. Bercos-Hickey et al. (2020) showed that the Saharan mineral dust and the West African precipitation feedback on the AEJ by modifying the meridional temperature gradient over North Africa.

  4. 4.

    Biases in the modeled AEJ induce biases in modeled AEWs with a strong implication on simulated rainfall climatology, especially over the Sahel. RCM runs that feature weaker (stronger) AEJ also model weaker (stronger) AEWs’ activity leading to dry (wet) rainfall biases. This result reveals a lack of coupling between AEWs and rainfall, and is consistent with previous findings by Skinner and Diffenbaugh (2013) using CMIP3 GCMs. They showed that the force of modeled AEWs’ activity is strongly dependent on the simulated intensity and positioning of the AEJ and low-level westerlies. The cyclonic convective circulation has likely weakened because of the underestimating strength of the AEJ and the subsequent AEWs’ activity.

  5. 5.

    Over the Sahel, rainfall dry biases in the RCMs are associated with shallow ascent motions, whereas wet biases to deep-ascent motions. However, the way the moist static energy (h) and equivalent potential temperature (θe) are associated with convection differs between reanalyses and RCM experiments, and from one RCM run to another. We found that over the Sahel, strong h and θe values are related to moist convection in some RCM experiments (e.g. REMO-ERA, REMO-HadGEM2, REMO-MPI-LR, RegCM4-MPI-MR), whereas the same conditions drive, instead, dry convection in reanalyses. Also, REMO-NorESM1 which features precipitation dry biases over the Sahel shows rather high values of h and θe in the lower atmospheric layers, a condition of atmospheric instability that should induce convection and therefore moistening the region.

Numerous studies have demonstrated the strong influence of large-scale processes on the WAM system, for instance, the Pacific Decadal Oscillation, the El Niño-Southern Oscillation or the North Atlantic Oscillation (Zhang and Zhou 2011; Diatta and Fink 2014). These factors are not investigated in the present study although their representation in the driving GCMs might have a strong impact on the RCMs' rainfall biases. This suggests that the current work is far from exhaustive in terms of the sources of precipitation biases in CORDEX-CORE RCMs. Also, here (and in addition to other studies, e.g. Dosio et al. 2021a; Tamoffo et al. 2021), it is shown that there are still important rainfall biases in the results of the dynamical downscaling conducted with CORDEX-CORE RCMs over Africa, even if the horizontal resolution is twice that of the original CORDEX simulations. Although increasing models' resolution is proven to be beneficial where regional- and local-scale processes play an important role in a region’s climate, it is not sufficient to overcome many model deficiencies (Doblas-Reyes et al. 2021). Wu et al. (2020) demonstrated the dominance effect of the model’s physics parameterization over the resolution on precipitation climatology across Africa. They showed that the model physics controls the spatial pattern of seasonal rainfall biases while the resolution mainly influences the precipitation intensity, thereby suggesting the necessity of models’ formulation for a better representation of the climate. The manner models are configured could also be an important source of uncertainties, for instance, when the convection is parameterized instead to be explicitly resolved; or when RCMs are not coupled with an ocean model, in order to physically modify SST prescribed by driving GCMs. In the first case, promising results are becoming available from convection-permitting models (CPM), i.e. RCMs run at very high resolution (around 5 km or less) so that convection can be explicitly simulated. These simulations showed significant improvements in small-scale precipitation characteristics (such as intensity and diurnal cycle) and notable effects on projected changes in extremes. However, given the significant amount of computational resources involved in the simulations, results are available only for a very limited period of time (10 years or less) (Kendon et al. 2019; Kouadio et al. 2020; Senior et al. 2021). In the second case, the RCM community has started to develop models that incorporate more components of the climate system, like regional oceans, or biogeochemistry (Doblas-Reyes et al. 2021). However, the majority of dynamical downscaling using RCMs is still done with SST prescribed by the driving GCM. This was indeed part of the CORDEX protocol. Nevertheless, the results of the downscaling employed in this paper capture well the basic features of the WAM system, as most of the analyzed physical underlying processes confirm the reasonableness of the simulated monsoon rainfall climatology.

Finally, as mentioned, the direct comparison of present and future climate as simulated by RCMs and GCMs has been performed in several previous studies (see e.g. Dosio et al. 2021a and reference within), including the analysis of the added value of downscaling (e.g. Dosio et al. 2015; Gibba et al. 2019). However, the quantification of added value by means of a specific, thorough process-based analysis of the performances of the RCMs compared to those of the driving GCMs is still missing (see e.g. discussion in Dosio et al. 2019) but outside the scope of this particular work (which focuses on the performances of RCMs only), and is left for future research.