The historical near‐surface O3 concentrations over Africa (36° S–30° N, 17.5° W–50° E) for the period 2000–2019 are simulated in the GEOS‐Chem model version 13.4.1, which is driven by meteorological fields from the Modern-Era Retrospective analysis for Research and Applications version 2 (MERRA-2; Gelaro et al., 2017). The model simulation has a horizontal resolution of 2° latitude × 2.5° longitude, and includes 47 vertical layers from the surface up to 0.01 hPa. It incorporates fully coupled O3–NOx-hydrocarbon-aerosol chemical mechanisms (Park et al., 2004; Pye et al., 2009; Mao et al., 2013), with about 300 species participating in over 400 kinetic and photochemical reactions (Bey et al., 2001). Vertical mixing process within the planetary boundary layer is characterized using a non-local scheme (Lin and McElroy, 2010), and stratospheric O3 chemistry utilizes the linearized O3 parameterization scheme (LINOZ; McLinden et al., 2000).

Anthropogenic emissions of O3 precursors, including non-methane VOCs (NMVOCs), NOx and CO from 2000 to 2019 are derived from the Community Emissions Data System (CEDS) version 2021_04_21 (O'Rourke et al., 2021). Biogenic emissions are calculated online based on the Model of Emissions of Gases and Aerosols from Nature (MEGAN) version 2.1 (Guenther et al. 2012). Biomass burning emissions are obtained from the Global Fire Emissions Database (GFED) version 4 (van der Werf et al., 2017). NOx emissions from soil sources are estimated online according to an updated version of the Berkeley–Dalhousie Soil NOx Parameterization scheme proposed by Hudman et al. (2012). Lightning-induced NOx emissions are calculated online following the algorithm developed by Ott et al. (2010) and Murray et al. (2012). Methane (CH4) concentrations are prescribed using spatially interpolated monthly average observations from the National Oceanic and Atmospheric Administration (NOAA) Global Monitoring Division (GMD; Murray, 2016). Previous studies have shown that GEOS-Chem model reproduces the magnitude and spatial distributions of observed O3 concentrations across Africa (Han et al., 2018; Yan et al., 2019; Wang et al., 2025a).

The CMIP6 repository contains multi-model climate projections for various SSPs based on alternative scenarios of future emissions and land use changes (O'Neill et al., 2016). In this study, we collect meteorological variables from CMIP6 models under four different scenarios. SSP1-2.6 represents a low-forcing scenario that considers comprehensive climate mitigation strategies to effectively reduce greenhouse gas emissions and progress toward sustainable development goals. SSP2-4.5 is an intermediate-forcing scenario that follows the business-as-usual pathway. SSP3-7.0 corresponds to a medium-to-high forcing scenario. SSP5-8.5 denotes a high-forcing scenario aiming to achieve climate adaptation alongside rapid development through intensive utilization of fossil-fueled resources.

To perform a robust prediction of future near-surface O3 concentrations in Africa, climate projections from 18 global climate models participating in CMIP6 are applied, including ACCESS_CM2, ACCESS-ESM1-5, CanESM5, CESM2-WACCM, CMCC-CM2-SR5, EC-Earth3, EC-Earth3-Veg, FGOALS-f3-L, FGOALS-g3, GFDL-ESM4, INM-CM5-0, IPSL-CM6A-LR, MIROC6, MPI-ESM1-2-HR, MPI-ESM1-2-LR, MRI-ESM2-0, NorESM2-LM, and NorESM2-MM. These models provide the major meteorological variables essential for predicting near-surface O3, such as air temperatures (at 2 m, 850 hPa, and 500 hPa), wind fields (at 850 and 500 hPa), incoming shortwave radiation at the surface, near-surface relative humidity, precipitation rate, total cloud cover, and sea level pressure under the four different scenarios. In order to minimize the inconsistencies of initial conditions between CMIP6 models and MERRA-2 reanalysis data, the future meteorological fields under different scenarios from CMIP6 models are adjusted based on the differences between CMIP6 historical meteorological variables and MERRA-2 reanalysis data during 2000–2019 following Li et al. (2022, 2023, 2024). Additionally, to discuss the role of climate change in the projected variation of O3 concentrations driven mainly by anthropogenic emissions, the monthly O3 simulation outputs under four SSPs scenarios from 8 CMIP6 models are adopted, including BCC-CSM2-MR, FGOALS-g3, IPSL-CM6A-LR, MPI-ESM1-2-HR, MPI-ESM1-2-LR, MRI-ESM2-0, NorESM2-LM, and NorESM2-MM.

As one of the traditional ML algorithms, Random Forest (RF) model can process high-dimensional data with lower computational costs compared to the CTMs (Breiman, 2001). This ensemble approach excels at dealing with nonlinear and complex relationships between input and target variables, and it provides stable and reliable predictions of O3 in many previous studies (Wei et al., 2022; Xu et al., 2023). In this study, to predict future near-surface O3 over Africa, we integrate a set of relevant features as predictors, including O3 concentrations from GEOS-Chem model simulations, O3 precursor emissions, meteorological variables, topography (TOPO), land cover (LC), normalized difference vegetation index (NDVI), and population density (POP). Considering the autocorrelation between O3 and its covariates varies across space and time, the RF model also incorporates spatiotemporal information, including month of the year (MOY), longitude (LON) and latitude (LAT) across the Africa domain. The details of input data applied in this study are summarized in Table 1.

We first predict future biogenic isoprene emissions based on RF model, which are then used for predicting future near-surface O3 concentrations across Africa. The RF model for isoprene is trained using emission output from MEGAN2.1 in GEOS-Chem model, MERRA-2 meteorological variables, TOPO, LC, NDVI, POP, MOY, LAT, and LON over 2000–2009 and 2011–2019. The 2010 records are used to validate the performance of RF model, because land surface air temperature over Africa reached its peak during 2000–2019, which facilitates future projections in a warming climate. This data splitting strategy ensures the critical information related to rising temperature trends is captured in RF model projections. Based on trained RF model, the future monthly isoprene emissions from 2020 to 2054 under different climate scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) in Africa are then predicted using future meteorological fields from CMIP6.

Secondly, to train the RF model for predicting future monthly near-surface O3 concentrations in Africa, we integrate O3 simulations from GEOS-Chem model, emissions and concentrations of O3 precursors, the ratio of VOCs emission to NOx emission, MERRA-2 meteorological variables, and auxiliary data. The samples over 2000–2009 and 2011–2019 are selected as training dataset and the remaining data over 2010 as testing data. We conduct two experiments to quantitatively estimate the direct (via altering meteorological conditions) and indirect (via altering biogenic isoprene emissions) effects of climate change on future O3 in Africa. By feeding the varying meteorological parameters into the trained model while fixing isoprene emissions in 2020, the climate-driven O3 variations through changing meteorological conditions are explored (O3_MET). Both the isoprene projections and CMIP6 multi-model projections from 2020 to 2054 under four scenarios are fed to the trained RF model to predict O3 (O3_ALL), and the difference of O3_ALL – O₃_MET represents the impacts of changing isoprene emissions under climate change (O₃_NAT).

The RF model's predictive performance was optimized through hyperparameter tuning. The best hyperparameters (n_estimators = 600, min_samples_split = 2, max_features = “sqrt”, bootstrap = “True”) of RF model for predicting isoprene emissions and O3 concentrations are tuned separately by 10-fold cross-validation (Rodriguez et al., 2010). To assess the accuracy of RF model, statistical metrics such as coefficient of determination (R2), mean absolute error (MAE), root mean square error (RMSE), and mean relative error (MRE) are calculated by comparing the outputs of RF model and GEOS-Chem model. Moreover, we employ the Shapley Additive explanation (SHAP) approach (Lundberg and Lee, 2017) to quantify the importance of individual factors determining the RF model's predictions. The SHAP provides a value that reflects the contribution of each input variable to specific predictions, and has been widely adopted in atmospheric environmental studies to enhance the interpretability of the ML model (Hou et al., 2022; Stirnberg et al., 2021).

Exposure to elevated temperatures and O3 pollution, exacerbated by global warming, significantly increases human health risks. We assess the future mortality ratio (MR) attributable to ambient O3 and temperature changes, respectively, under different scenarios in Africa from 2020 to 2054, using the following equations: 1MRozone=∑iRRozone,im∑jRRozone,jn2MRtemperature=∑iRRtemperature,im∑jRRtemperature,jn3RRozone=expβ1C-C04RRtemperature=expβ2T-T0 where MRozone (MRtemperature) stands for the MR due to O3 pollution (temperature) changes. RRozone,i and RRtemperature,i represent relative risks caused by O3 and temperature exceedance in a future (2050–2054) month i, respectively. RRozone,j and RRtemperature,j denote relative risk caused by O3 and temperature exceedance on a baseline (2020–2024) month j, respectively. m is the total number of months during future period, and n is the total number of months of baseline period. β1 indicates the O3 concentration response factor, with a value of 1.39×10-3 (95 % confidence interval: 8.9597×10-4, 1.8822×10-3) per ppb (Wang et al., 2025b). C0 is the theoretical minimum risk exposure level (TMREL) for O3 exposure, which ranges from 29.1 to 35.7 ppb (GBD 2019 Risk Factors Collaborators, 2020). Following Malashock et al. (2022), we use the medium value of 32.4 ppb in this study. C reflects the O3 concentrations predicted by the RF model. β2 signifies the temperature response factor, and each 1 °C increase in temperature over Africa is associated with a 1.3×10-2 (95 % confidence interval: 0.4×10-2, 2.2×10-2) rise in mortality risk (Cromar et al., 2022). T0 is the threshold value indicating minimum mortality temperature (MMT), an important indicator for characterizing the health impacts of global heating. The 50th percentile of temperature distribution corresponding to the MMT over Southern Africa was calculated to be 25 °C, which is used as the reference value for this study (Tobías et al., 2021). T reflects air temperature from the CMIP6 dataset. Similar calculation methodology has been applied in previous studies (Lee and Kim, 2016; Wang et al., 2022).

We evaluate the ML model using a test dataset from 2010. Figure 1a and b illustrate the consistency between RF model projections and GEOS-Chem model simulations for isoprene emissions and near-surface O3 concentrations in Africa, respectively. The RF model exhibits a strong predictive capability in reproducing both isoprene emissions (R2=0.97, MRE = 17 %) and near-surface O3 concentrations (R2=0.94, MRE = 6 %). These robust statistical metrics confirm the trained RF model's suitability for predicting future O3 concentrations over Africa under a warmer climate.

The isoprene emissions simulated by GEOS-Chem model during 2000–2019 reveal apparent regional discrepancies across Africa (Fig. 2a). The maximum isoprene emissions originate from evergreen broadleaf trees over Central Africa, which has the second largest tropical rainforest in the world. Southern Africa exhibits relatively lower isoprene emissions due to its sparse coverage of deciduous broadleaf trees. Northern Africa, dominated by deserts and semi-arid grasslands, contributes little to isoprene emissions. The isoprene projections estimated by the RF model perfectly capture the spatial distributions of isoprene emissions across Africa, with a correlation coefficient (R) between the RF projections and GEOS-Chem simulations of 0.99 and a normalized mean bias (NMB) of -0.2 %. In addition, the O3 concentrations from 2000 to 2019 simulated by GEOS-Chem model indicate that O3 pollution is predominantly concentrated in Southern Africa, as well as parts of Northern Africa and Central Africa, with concentrations over 40 ppb (Fig. 2b). The RF model accurately reproduces this spatial characteristic, which aligns well with the O3 simulation results (R=0.99, NMB = -0.02 %).

We further evaluated the RF-predicted O3 concentrations for 2020–2025 against ground-based observations from the South African Air Quality Information System (SAAQIS), which provides routine air quality measurements from monitoring stations across South Africa (Fig. S1 in the Supplement). We selected 94 monitoring sites with valid observations for at least three years during this period. To match the 2° × 2.5° model resolution, site-level mean O3 concentrations were aggregated within each model grid cell, yielding 11 grid-cell mean observations. These values were compared with RF-predicted O3 averaged over 2020–2025 and across the four SSP scenarios. The RF predictions reproduce a moderate degree of spatial variability across South Africa with a correlation coefficient of 0.54, but systematically overestimate observed O3 concentrations, with MAE of 11.92 ppb, RMSE of 12.93 ppb and NMB of 46.7 %. Although the grid-cell averaging reduces the mismatch between point observations and model grid cells, the monitoring sites are concentrated around Pretoria and coastal South Africa. Thus, unresolved urban emission gradients, local topography, and coastal meteorology, as well as the limited spatial coverage of the observations, may contribute to the disagreement.

To gain deeper insights into the driving factors behind RF model outputs, we utilize SHAP, an explainable artificial intelligence technique, to quantify the contribution of each feature to the projections of biogenic isoprene emissions (Fig. 3a) and O3 concentrations (Fig. 3b) in Africa. The LC and NDVI are identified as the strongest positive drivers of isoprene emissions. This is consistent with the physics that biogenic isoprene is emitted mainly from terrestrial vegetation, and its emissions are directly influenced by LC types and the associated plant species (Guenther et al., 2006; Rosenkranz et al., 2015). Among all meteorological variables, air temperature plays a dominant role in regulating isoprene emissions, exhibiting a positive correlation, in line with prior work (e.g., Singsaas and Sharkey, 1998). The isoprene emission rates show opposite responses to wet weather conditions, as increases in the frequency and intensity of precipitation can influence plant functions (Strada et al., 2023). For O3 concentrations, meteorological factors – relative humidity, air temperature, precipitation, cloud cover, and incoming solar radiation – exert substantial influence, with temperature and solar radiation positively associated with O3 levels. Additionally, precursor emissions constitute a vital component in O3 formation. The VOCs-to-NOx emission ratio, used here as an indicator of O3 production regime in this study, is also a critical determinant of projected O3.

The projected changes in RF-predicted isoprene emissions over Africa in 2050 (averaged over 2050–2054) relative to 2020 (averaged over 2020–2024) are shown in Fig. 4. The isoprene emissions show increasing trends under all four future scenarios. Shaped by the land cover characteristics in Africa, there is a distinct north-south spatial distribution pattern in the changes in isoprene emissions. Central Africa has the vast tropical forests and woodlands, serving as the major source of isoprene emissions and exhibiting the largest increases in biogenic isoprene emissions across Africa, with a maximum growth rate exceeding 1.0 gm-2yr-1 under four scenarios. In contrast, the land use type in Northern Africa is dominated by barren land, resulting in consistently low isoprene emission levels. Compared to the period of 2020–2024, the projected changes of isoprene in Northern Africa for 2050–2054 are less than 0.1 gm-2yr-1, while the isoprene emissions increase by 0.1–0.5 gm-2yr-1 over Southern Africa under all SSPs.

According to the feature contributions derived by SHAP analysis, biogenic isoprene emissions largely depend on meteorological fields, with air temperature (at 2 m) identified as the most critical factor (Fig. 3a). A comparison of future meteorological variables between 2050–2054 and 2020–2024 under different scenarios demonstrates that near-surface air temperatures elevate throughout the entire African continent in 2050–2054, with SSP5-8.5 presenting the strongest warming (Fig. 5). Therefore, isoprene emission rates are anticipated to rise, especially in the strong warming scenarios (i.e. SSP3-7.0 and SSP5-8.5). Besides, rising air temperature and dry conditions can further enhance isoprene release. The enhancement of projected isoprene emissions induced by drought stress is particularly pronounced over Central and Southern Africa (Fig. 6).

Figure S2 shows the spatial distributions of biogenic isoprene emission changes during 2050–2054 compared to 2020–2024 in March–April–May (MAM), June–July–August (JJA), September–October–November (SON), and December–January–February (DJF) under the four scenarios. The changes in isoprene emissions do not show obvious seasonality, but there are relatively larger increases in MAM and DJF than other seasons.

Figure 7a shows the changes in annual mean near-surface O3 concentrations in response to climate change under different scenarios, as projected by the RF model using future meteorological fields from 18 CMIP6 models (O3_MET). The projected O3 exhibits an overall increasing trend during 2050–2054 relative to 2020–2024, indicating a climate penalty on air quality over most African regions. Future increases in air temperature (Fig. 5), along with reductions in relative humidity (Fig. 6) and cloud cover (Fig. 8) will enhance photochemical O3 production, leading to substantial increases in O3 levels, with maximum increases over 1.0 ppb, even reaching 2.0 ppb, under strong warming scenarios. Under the weak warming scenario SSP1-2.6, the climate-driven O3 increases are relatively small, with increases of less than 1.0 ppb across Africa.

The VOCs-to-NOx emission ratio was included as a predictor to represent, in a simplified way, the nonlinear response of O3 to precursor emissions. It is not used here as a diagnostic of a uniform O3 production regime. Precursor sensitivity can vary spatially and seasonally. NOx-rich urban environments and dry season biomass burning plumes can favor VOC-sensitive conditions and are closely linked to enhanced O3 over southern Africa (Swap et al., 2003; Wang et al., 2025a). In the O3_NAT sensitivity experiment, only biogenic isoprene emissions are allowed to vary, whereas biomass burning and other non-isoprene precursor emissions are held fixed at their baseline values. Accordingly, Fig. 7b indicates an annual mean and model-derived O3 decrease of less than 0.5 ppb over Central and Western Africa in response to the isolated isoprene perturbation. This result should not be generalized to urban or biomass burning-influenced regions or seasons. The total effects of changing meteorological factors and biogenic isoprene emissions on O3 variations are similar to those due to changes in meteorological factors alone, as depicted in Fig. 7c (O3_ALL). In summary, the climate-driven changes in meteorological parameters exert a more dominant role in regulating future O3 concentrations over Africa through changing physical and chemical processes, while the changing biogenic isoprene emissions exerts a minor role.

Climatology of meteorological fields over Africa has obvious seasonal characteristics, and the formation of near-surface O3 is tightly coupled with climate-driven processes. Figures S3–S6 show the spatial distribution changes in seasonal O3 concentrations influenced by meteorological fields and biogenic isoprene emissions under climate change. O3 concentration increases the most during MAM, with a maximum rise reaching up to 3.8 ppb over Central and Southern Africa under the SSP5-8.5 scenario (Fig. S3). In contrast, parts of Southern Africa show an O3 decrease of 0.2–1.0 ppb during SON under all scenarios (Fig. S5). Figure S5a shows that this decrease occurs in the O3_MET experiment, in which meteorological fields vary while biogenic isoprene emissions are fixed at their 2020 values. The similar spatial pattern in Fig. S5a and c indicates that, within the present sensitivity approach, the projected SON O3 decrease over Southern Africa is mainly driven by meteorological changes. In the RF model, this response is associated with higher sea level pressure (Fig. S7) and its negative relationship with O3 (Fig. 3b). Nevertheless, SON overlaps with the late dry season biomass burning period over much of southern Africa, and biomass burning is known to strongly affect regional O3 chemistry through emissions of NOx and VOCs (Archibald et al., 2010; Swap et al., 2003). Additionally, a decrease exceeding 0.5 ppb is expected across Central Africa in DJF under the strong warming scenarios (Fig. S6), largely driven by the rapid increases in precipitation frequency (Fig. S8). Moreover, increasing isoprene emissions also contribute to the decrease in O3 concentrations by 0.2–0.5 ppb over Eastern and Central Africa during JJA (Fig. S4).

Figure 9 shows the relative changes in future projected mortality burden in 2050 compared to that in 2020 corresponding to changes in O3 concentrations and air temperatures across Africa under the four future scenarios. The mortality due to the changes in O3 pollution driven by changes in meteorological factors and biogenic isoprene emissions are separately investigated. It is worth noting that the potential amplification or suppression effects of interactions between O3 and temperature on mortality are not considered in this study.

Mortality ratios (MRs) over Africa are presented in 2050 relative to 2020, with temperature-related MRs (MR_temperature) increases much higher than those due to the climate-driven increases in O3 exposure levels (MR_ozone), particularly under the SSP3-7.0 and SSP5-8.5 scenarios. This demonstrates that increases in extreme heat in a warmer future are projected to cause substantially more deaths than the climate-driven increases in extreme O3 concentrations across Africa. The MR_temperature ranges from 1.007 to 1.015, with an average of 1.0115. The MRs due to climate-driven O3 changes related to the meteorological parameters (MR_ozone−met) show a slight increase from 1.0002 to 1.0006, as the climate warms. In contrast, MRs attributed to climate-driven O3 changes related to biogenic isoprene emissions (MR_ozone−nat) do not show any noticeable change under different scenarios, which exceed 1.0 only under the SSP3-7.0 scenario. Nevertheless, these suggest that environmental conditions for human health in Africa will deteriorate in a warming climate with more frequent extreme heat together with intensified O3 pollution.