Climate-driven deterioration of future ozone pollution in Asia predicted by machine learning with multi-source data

Ozone (O₃) is a secondary pollutant in the atmosphere formed by photochemical reactions that endangers human health and ecosystems. O₃ has aggravated in Asia in recent decades and will vary in the future. In this study, to quantify the impacts of future climate change on O₃ pollution, near-surface O₃ concentrations over Asia in 2020–2100 are projected using a machine learning (ML) method along with multi-source data. The ML model is trained with combined O₃ data from a global atmospheric chemical transport model and real-time observations. The ML model is then used to estimate future O₃ with meteorological fields from multi-model simulations under various climate scenarios. The near-surface O₃ concentrations are projected to increase by 5 %–20 % over South China, Southeast Asia, and South India and less than 10 % over North China and the Gangetic Plains under the high-forcing scenarios in the last decade of 21st century, compared to the first decade of 2020–2100. The O₃ increases are primarily owing to the favorable meteorological conditions for O₃ photochemical formation in most Asian regions. We also find that the summertime O₃ pollution over eastern China will expand from North China to South China and extend into the cold season in a warmer future. Our results demonstrate the important role of a climate change penalty on Asian O₃ in the future, which provides implications for environmental and climate strategies of adaptation and mitigation.

1 Introduction

Tropospheric ozone (O₃) is a secondary air pollutant, formed by photochemical oxidation of non-methane volatile organic compounds (NMVOCs) and carbon monoxide (CO) in the presence of nitrogen oxides (NO_x= NO + NO₂) and sunlight. It has adverse effects on human health (Malley et al., 2017; Cakmak et al., 2018), vegetation growth (Yue et al., 2017; Mills et al., 2018), and climate change (Checa-Garcia et al., 2018; Gaudel et al., 2018). A better understanding of the causes of changes in O₃ concentrations is useful for developing effective environment and climate strategies. Since the mid-1990s, Asian regions, including South Asia, East Asia, and Southeast Asia, have experienced the fastest O₃ increase rate of 2–8 ppb per decade at remote surface sites and in the lower free troposphere across the world (IPCC, 2021). A number of air quality monitoring stations have been established in China since 2013 to measure real-time near-surface particulate matter, O₃, and other air pollutants. The measurements showed an increasing trend of urban warm-season daily maximum 8 h average (MDA8) O₃ concentrations of 2.4 ppb (5 %) yr⁻¹ that is faster than any other regions worldwide during 2013–2019 (Lu et al., 2020). However, many regions in Asia lack O₃ observations with sufficient spatial and temporal coverage. Also, most of the present regional observations are collected only near population clusters, which are not representative of the entire region (Zhou et al., 2022).

To supplement the limited near-surface O₃ measurements, many studies utilized global and regional models with comprehensive physical and chemical processes to simulate O₃ concentrations (Zhu et al., 2017; Gao et al., 2020; Yang et al., 2022). Moreover, statistical models have also been used to estimate O₃ concentrations (Chen et al., 2020; Zhang et al., 2020). In recent years, machine learning (ML) approaches, such as random forest (Xue et al., 2020; Wei et al., 2022), neural network (Di et al., 2017), support vector machine (Su et al., 2020), extreme gradient boosting (Liu et al., 2020), and ensemble learning (Liu et al., 2022), have been widely applied to estimate O₃ levels based on potential influential factors (e.g., precursor emissions, meteorological conditions, land use, surface elevation, gross domestic product, population density, and geographical variables). The abovementioned previous studies utilizing the ML methods showed high computational efficiency and accuracy, with an overall R² between the observed and predicted O₃ concentrations in the range of 0.7–0.9.

Meteorological factors and synoptic conditions play important roles in affecting O₃ pollution (Fu and Tai, 2015; Gong and Liao, 2019; Yin et al., 2019; Liu and Wang, 2020; Dang et al., 2021). Gong and Liao (2019) illustrated that hot, dry, and stagnant weather conditions are favorable for the formation and persistence of severe O₃ pollution over northern China. High air temperature along with intense incoming shortwave radiation accelerates both photochemical reaction rates and natural precursor emissions for O₃ production (Jacob and Winner, 2009). Under high relative humidity conditions, O₃ concentrations decrease due to many complex physical and chemical mechanisms (Jeong and Park, 2013; Kavassalis and Murphy, 2017; Lu et al., 2019; Li et al., 2021a). Cloud and precipitation impact O₃ levels through reducing the downwelling solar radiation and washout of pollutants (Toh et al., 2013). Anomalous sea level pressure patterns can affect the long-range transport of O₃ by influencing atmospheric circulation (Santurtún et al., 2015). By changing the air stagnation condition and transport of pollutants, wind fields can also affect O₃ concentrations in local and downwind areas of emission sources (Doherty et al., 2013).

Future climate change corresponding to the different climate scenarios can impact O₃ through altering meteorological conditions (Wang et al., 2013; Fu and Tian et al., 2019). Using regional climate fields downscaled from general circulation models to investigate potential O₃ variations in the US due to changing climate, Fann et al. (2015) projected the MDA8 O₃ to increase by 1–5 ppb as daily maximum average temperature increases by 1–4 ^∘C in 2030 relative to 2000. Colette et al. (2015) estimated that the climate penalty for future summertime near-surface O₃ reaches 0.99–1.5 ppb by the end of the 21st century (2071–2100) in Europe compared to present-day levels using an ensemble of eight global coupled climate–chemistry models under the RCP8.5 (Representative Concentration Pathway) scenario. Through fixing sea surface temperature at present-day and future conditions in five atmospheric-only models as part of AerChemMIP (Aerosol Chemistry Model Intercomparison Project), Zanis et al. (2022) projected the climate change penalties and benefits on global near-surface O₃ concentrations from 2015 to 2100 under scenarios of SSP3-7.0 (Shared Socioeconomic Pathway). They found O₃ reductions in most regions of the globe, except a robust O₃ climate penalty of 1–2 ppb ^∘C⁻¹ in South Asia and East Asia under global warming following the SSP3-7.0 pathway. However, SSP3-7.0 is not a good representative scenario for both air quality and climate in Asia. The emissions of greenhouse gases (GHGs) and air pollutants over East Asia in SSP3-7.0 are assumed to significantly increase in the near future and remain at high levels in the middle of the 21st century, compared to SSPs (Li et al., 2022), while the emissions of air pollutants have been cut by a lot since the 2010s in the real world (Wang et al., 2021). The GHGs and pollutant emissions are very likely to continually decline in the future, related to the carbon neutrality commitment (Cheng et al., 2021).

In this study, we aim to better characterize the impact from future climate change on Asian O₃ pollution using multiple state-of-the-art modeling tools and data. It is important for policymakers that mitigating global climate change potentially has positive benefits to surface air quality through meteorological factors, not only the reduction in fossil fuel co-emissions. The near-surface O₃ concentrations covering 2020–2100 in Asia are projected using an ML method integrated with multi-source data, including assimilated O₃ data that combine ground observations across China and simulations from a global 3-D chemical transport model (GEOS-Chem), meteorological fields under various climate scenarios from the latest Coupled Model Intercomparison Project Phase 6 (CMIP6) multi-model simulations, and other auxiliary data (e.g., emissions, land use, topography, population density, and spatiotemporal information). ML approach gives the capacity to explore many scenarios more rapidly and for longer time periods than the chemical transport model process-based modeling. Details of the data and methodology used in this study are described in Sect. 2. Section 3 analyzes the results of climate-driven O₃ variations over different key regions of Asia. Section 4 summarizes the main conclusions and discusses potential uncertainties in this study.

Figure 1 The structure and specific schematics for predicting future O₃ concentrations under four scenarios based on the machine learning (ML) method. MET: meteorology; EMIS: emissions; LAT: latitude; LONG: longitude; MOY: month of the year (see other definitions in Table 1).

2 Materials and methods

2.1 GEOS-Chem model description

Figure 1 illustrates the procedures for predicting future near-surface O₃ over Asia under four scenarios. To assimilate O₃ data for the ML model training, the near-surface O₃ concentrations over Asia from 2014 to 2019 are firstly simulated using the nested-grid version of the 3-D GEOS-Chem model (version 12.9.3), driven by the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2), reanalysis meteorological data (Gelaro et al., 2017). The nested GEOS-Chem has 47 vertical layers from the surface up to 0.01 hPa, with a horizontal resolution of 0.5^∘ latitude × 0.625^∘ longitude over the Asia domain (11^∘ S–55^∘ N, 60–150^∘ E). The lateral boundaries of chemical tracer concentrations are provided by global simulations at 2^∘ latitude × 2.5^∘ longitude horizontal resolution. The model includes fully coupled aerosol–O₃–NO_x–hydrocarbon 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). The stratospheric O₃ chemistry is simulated through a linearized O₃ parameterization scheme (LINOZ; Mclinden et al., 2000), and the planetary boundary layer mixing is calculated by a non-local scheme (Lin and McElroy, 2010). GEOS-Chem has shown a good performance in reproducing spatiotemporal distributions of O₃ concentrations (e.g., Ni et al., 2018; Li et al., 2019).

Figure 2 Spatial distributions of (a) observed near-surface O₃ concentrations across China and (b) assimilated O₃ concentrations over Asia in 2014–2019. Correlation coefficient (R) between observed and assimilated O₃ and the normalized mean bias (NMB $=\sum$ (Observed − Assimilated) / ∑ Assimilated ×100 %) are given at the bottom left of panel (a).

The historical (2014–2019) anthropogenic emissions of O₃ precursor gases, including NO_x, NMVOCs, and CO, utilized in the nested domain are obtained from the Community Emissions Data System (CEDS; Hoesly et al., 2018) version 2021_04_21, which fully considered the recent emission reductions in China related to clean air measures. The biomass burning emissions are acquired from the Global Fire Emissions Database, Version 4 (GFED4; van der Werf et al., 2017). Biogenic emissions of NMVOCs from the Model of Emissions of Gases and Aerosols from Nature (MEGAN) version 2.1 are employed, with updates from Guenther et al. (2012). Soil NO_x sources are calculated with an updated version of the Berkeley–Dalhousie Soil NO_x Parameterization scheme (Hudman et al., 2012). NO_x emissions from lightning are as described by Murray et al. (2012), and the vertical distribution of emissions follows Ott et al. (2010).

2.2 Ground O₃ observations

To improve the performance of the ML model in predicting O₃ concentrations, the nationwide hourly near-surface O₃ concentrations in China during 2014–2019 are obtained from the Chinese Ministry of Ecology and Environment (MEE) and used for O₃ data assimilation, which has been widely used to examine pollution over China in previous studies (K. Li et al., 2020, 2021; Qian et al., 2022). The observational network had about 500 monitoring sites in 2013 and expanded to more than 1500 sites after 2019, covering 360 cities in mainland China. In this study, the quality controlled hourly O₃ observations in 360 cities are averaged within each 0.5^∘ latitude × 0.625^∘ longitude grid of the GEOS-Chem model.

2.3 Data assimilation

The assimilation system, which is used to combine the O₃ observations across China with results from GEOS-Chem simulations, is based on a three-dimensional variational (3DVar) data assimilation (Kalnay, 2003; Evensen et al., 2022). The goal of the 3DVar is to find the maximum likelihood estimation of a state vector x, which represents the O₃ concentrations here in this study, given the available observations y through minimizing the cost function:

$$\begin{array}{}\text{(1)}& \begin{array}{rl}J\left(x\right)& ={\displaystyle \frac{\mathrm{1}}{\mathrm{2}}}{\left(\mathit{x}-{\mathit{x}}^{\mathrm{b}}\right)}^{\mathrm{T}}{\mathbf{B}}^{-\mathrm{1}}\left(\mathit{x}-{\mathit{x}}^{\mathrm{b}}\right)\\ & +{\displaystyle \frac{\mathrm{1}}{\mathrm{2}}}{\left(\mathit{y}-\mathrm{H}\left(\mathit{x}\right)\right)}^{\mathrm{T}}{\mathbf{O}}^{-\mathrm{1}}\left(\mathit{y}-\mathrm{H}\left(\mathit{x}\right)\right),\end{array}\end{array}$$

where x^b represents the priori simulation; B is the empirical background covariance matrix formulated as a product of the uncertainty in the simulated value and a distance-based correlation matrix C; and the individual element is calculated as

$$\begin{array}{}\text{(2)}& {\mathbf{B}}_{i,j}=\mathrm{0.2}\times {\mathit{x}}_i^{\mathrm{b}}\times \mathrm{0.2}\times {\mathit{x}}_j^{\mathrm{b}}\times {\mathbf{C}}_{i,j},\end{array}$$

where we have used 20 % choice to characterize uncertainty of the O₃ simulation, with the correlation matrix empirically set as

$$\begin{array}{}\text{(3)}& {\mathbf{C}}_{i,j}=e^{-{\left(\frac{d_{i,j}}{\mathrm{200}\mathrm{km}}\right)}^{\mathrm{2}}/\mathrm{2}},\end{array}$$

where d_i,j represents the spatial distance between grid cell i and j.

H denotes the linear observation operator that converts the simulation results into the observational space. Here all observations are assumed to be independent, and therefore O is a diagonal covariance matrix storing the square of the observation uncertainty, which is also set as 20 % similarly.

Comparisons between observed and assimilated O₃ concentrations over 2014–2019 are shown in Fig. 2. The overall correlation coefficient (R) is 0.94, and the normalized mean bias (NMB) is −0.1 %, suggesting that the assimilated data have an excellent representation of O₃ observations and minimize the uncertainties of GEOS-Chem simulations in China.

Table 1 Summary of detailed datasets used in this study. AVHRR: Advanced Very High Resolution Radiometer; ESA CCI: ESA Climate Change Initiative; SRTM: Shuttle Radar Topography Mission.

2.4 Predicting O₃ using a machine learning method

In this study, a random forest (RF) model is used to predict O₃ concentrations, similar to our previous studies (H. Li et al., 2021, 2022), with input data of assimilated O₃ concentrations in China that combine observations and results from GEOS-Chem model simulations, GEOS-Chem simulated O₃ concentrations outside of China, MERRA-2 meteorological variables, O₃ precursor emissions, land cover (LC), the normalized difference vegetation index (NDVI), topography (TOPO), population density (POP), the month of the year (MOY), and the geographic location of each model grid as spatiotemporal information. Details of the datasets are summarized in Table 1.

For predicting future climate-driven near-surface O₃ concentrations, the ML model is trained with samples over 2014–2018 and the remaining 2019 data are used for model validation. To obtain an optimal ML model, hyperparameters are firstly tuned using the 10-fold cross-validation (Rodriguez et al., 2010). The best hyperparameters (n_estimators =200, min_samples_split =2, max_features = “sqrt”, bootstrap = “True”) of the ML model are utilized. Several statistical metrics, including the coefficient of determination (R²), mean absolute error (MAE), root mean square error (RMSE), and mean relative error (MRE), are used to evaluate the performance of ML model. Then the climate-driven near-surface O₃ concentrations during 2020–2100 under four SSPs (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) in Asia can be estimated using the trained ML model with varying meteorological factors under the climate change scenarios. Both anthropogenic and natural emissions of O₃ precursors are fixed at the present-day levels for the prediction.

2.5 Meteorological fields from CMIP6 multi-model simulations

Monthly meteorological parameters under four different future climate scenarios, including SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5 (a representation of low-, intermediate-, medium–high-, and high-forcing levels, respectively), are fed to an ML model to predict near-surface O₃ concentrations. The Scenario Model Intercomparison Project (ScenarioMIP) as part of CMIP6 provides multi-model projections of climate variables driven by future emission and land use changes under different SSPs (O'Neill et al., 2016). In our study, meteorological fields, such as air temperature (at 2 m, 850 hPa, and 500 hPa), wind fields (at 850 and 500 hPa), surface relative humidity, incoming shortwave radiation at the surface, total cloud cover, precipitation rate, and sea level pressure, are chosen as the key meteorological predictors for near-surface O₃ concentrations, which are obtained from 18 global climate models, i.e., ACCESS-CM2, ACCESS-ESM1-5, CanESM5, CESM2-WACCM, CMCC-CM2-SR5, EC-Earth3-Veg, EC-Earth3, 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. Before being applied to the ML model, future meteorological fields from ScenarioMIP are adjusted by their potential bias, characterized as the difference in their historical climatological mean (2014–2019) and MERRA-2 following Li et al. (2022). It minimizes the inconsistencies in the initial conditions in models and reanalysis data.

3 Results

3.1 Predictive capability of the machine learning model

The predicted monthly O₃ concentrations over Asia in 2019 by the ML model are in good agreement with the assimilated O₃ data constructed with observations and GEOS-Chem model results (Fig. 3). The overall R² between the predicted and assimilated O₃ concentrations is as high as 0.92, and the ML model has a low MRE of 9 % in predicting O₃ concentrations over the Asia domain. Overall, these statistical indices indicate that the RF model is promising for predicting the spatial distributions and temporal variations in near-surface O₃ concentrations over Asia, which can provide a practical means for studying long-term variations in O₃ under the future climate change.

Figure 3 Density scatterplots of predicted vs. assimilated monthly near-surface O₃ concentrations (ppb) in 2019 over Asia. The gray and red lines are the 1:1 line and linear regression line, respectively. Statistical metrics including the number of samples (N), correlation of determination (R², unitless), root mean square error (RMSE, ppb), mean absolute error (MAE, ppb), and mean relative error (MRE, %) are shown at the top left.

Meanwhile, the ML model predictive capability for each grid cell over the entire domain during 2014–2019 is further evaluated and demonstrated in Fig. 4. Regarding the spatial performance, the estimated O₃ concentrations are highly correlated to the assimilated data in most regions of Asia with small biases, indicating a strong spatial predictive ability of the RF model. More than 80 % of land areas have an R² greater than 0.9. In terms of model uncertainties, about 95 % of land areas have an RMSE (MAE) of less than 3 ppb (2 ppb) (parts per billion). Furthermore, approximately 86 % of land areas show small modeling bias with MRE below 5 %. Note that several grid cells show MRE over 5 % but still below 15 %, which is related to the data assimilation using monitored and simulated O₃ concentrations in China and the coarse resolution for coastal areas and islands over Southeast Asia.

Figure 5 shows the importance score of independent variables that contribute to the prediction of the trained ML model, which called Gini importance and implies the influence of input features on the target variable in the ML model. The results suggest that among all the input predictors, relative humidity, incoming solar radiation at the surface, and topography are the top-three most influential variables for the model construction of near-surface O₃ in Asia, with importance scores of 15 %, 12 %, and 10 %, respectively. The primary importance of relative humidity has also been reported in previous studies (e.g., Han et al., 2020; Qian et al., 2022). Other meteorological parameters, such as cloud cover, sea level pressure, air temperature, and precipitation, also have a substantial impact on the O₃ estimates, with importance scores ranging from 4 % to 8 %. In the ML model, the emissions of three primary O₃ precursors, including NMVOCs, NO_x, and CO, have a relatively low importance score of 4 %–5 % individually due to the spatiotemporal diversity of O₃ production regimes. However, it is noted that the O₃ variations in different regions are dominated by different meteorological factors (Weng et al., 2022). The importance score of each independent feature quantified in this study can only reflect the overall importance across Asia, which is less representative of any specific regions.

Figure 4 Spatial distributions of the performance statistics of the ML model with regard to (a) R² (unitless), (b) RMSE (ppb), (c) MAE (ppb), and (d) MRE (%) during 2014–2019 over Asia.

Figure 5 Importance scores of independent variables (meteorological parameters, emissions, land use, topography, and population density) used in the ML model for predicting future near-surface O₃ concentrations over Asia.

Figure 6 The spatial distributions of absolute (ppb) and percentage difference (%) of surface O₃ level between 2025 (2020–2029 mean) and 2095 (2091–2100 mean) driven by climate change under four scenarios (a, e) SSP1-2.6 (SSP126), (b, f) SSP2-4.5 (SSP245), (c, g) SSP3-7.0 (SSP370), and (d, h) SSP5-8.5 (SSP585). The box-outlined areas in (d) are North China (NC; 35–41^∘ N, 105–120^∘ E), South China (SC; 22–33.5^∘ N, 105–120^∘ E), Southeast Asia (SEA; 9.5^∘ S–19^∘ N, 93.75–140^∘ E), South India (SI; 8–18^∘ N, 73.125–80.625^∘ E), the Gangetic Plains (GP; 21.5–23.5^∘ N, 85.625–92.5^∘ E; 23.5–27^∘ N, 76.25–92.5^∘ E; and 27–30^∘ N, 76.25—81.25^∘ E), and the Tibetan Plateau (TP; 28–31^∘ N, 81.875–102.5^∘ E and 31–38^∘ N, 73.125–102.5^∘ E). No overlaying hatch pattern indicates statistical significance with 95 % confidence from a two-tailed t test.

3.2 Predicted future climate-driven O₃ variations

Figure 6 shows the predicted absolute and percentage changes in annual mean near-surface O₃ concentrations in response to climate change between the first and last decades of 2020–2100 based on the future meteorological fields from the 18 CMIP6 models. Figure 7 shows the time series of the regional averaged values over six sub-regions of Asia during 2020–2100. Under the global warming trends of all future scenarios, the climate-driven near-surface O₃ concentrations increase constantly from 2020 to 2100 over many key regions in Asia, such as North China (NC), South China (SC), Southeast Asia (SEA), South India (SI), and the Gangetic Plains (GP), except the Tibetan Plateau (TP). The O₃ concentrations over SC, SEA, and SI are projected to increase considerably with the maximum increase up to 5 ppb (20 %) in 2095 (2091–2100 mean) compared to 2025 (2020–2029 mean) under the SSP5-8.5 scenario, revealing a strong O₃–climate penalty in most Asian regions. The climate-driven changes in O₃ concentrations are smaller under the scenarios of less warming, especially in SSP1-2.6 that has O₃ changes of less than 5 % across Asia. These suggest that future climate following low emissions and sustainable pathways is more favorable for the mitigation of O₃ pollution in Asia than high-forcing scenarios.

The strong O₃–climate penalty over eastern China can be attributed to the particularly high O₃ precursor emissions (Fig. S1 in the Supplement), relative to western China, which lead to a positive local net O₃ production close to sources in a warming climate (Fig. S2) (Zanis et al., 2022). The absolute and percentage changes in regional averaged near-surface O₃ concentrations between 2025 and 2095 under the four scenarios are shown in Fig. 8. The climate-driven changes in O₃ concentrations are gradually stronger from the north (2 %–3 %) to the south (3 %–8 %) of China, which demonstrates that the changes in meteorology exert a greater impact on near-surface O₃ concentrations over SC than NC under future climate change. By the end of the 21st century, the relative humidity will decrease (Fig. S3) and downward solar radiation will increase (Fig. S4) over SC compared to those values in 2025, which are conducive to the O₃ productions, while NC has the opposite changes. Moreover, cloud cover will decrease more over SC than NC (Fig. S5), contributing to the larger increase in O₃ productions and concentrations over SC than NC in a warming climate.

Figure 7 Time series (2020–2100) of annual mean near-surface O₃ concentrations (ppb) driven by climate change under the four scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) over North China (NC; 35–41^∘ N, 105–120^∘ E), South China (SC; 22–33.5^∘ N, 105–120^∘ E), Southeast Asia (SEA; 9.5^∘ S–19^∘ N, 93.75–140^∘ E), South India (SI; 8–18^∘ N, 73.125–80.625^∘ E), the Gangetic Plains (GP; 21.5–23.5^∘ N, 85.625–92.5^∘ E; 23.5–27^∘ N, 76.25–92.5^∘ E; and 27–30^∘ N, 76.25–81.25^∘ E), and the Tibetan Plateau (TP; 28–31^∘ N, 81.875–102.5^∘ E and 31–38^∘ N, 73.125–102.5^∘ E). The black lines are the averages of the predicted O₃ based on meteorological fields from 18 CMIP6 models.

Figure 8 Absolute (a, ppb) and percentage (b, %) changes in projected near-surface climate-driven O₃ concentrations in 2095 (2091–2100 mean) relative to 2025 (2020–2029 mean) over the six selected key regions of Asia, including NC, SC, SEA, SI, GP, and TP under four future climate scenarios. The error bars indicate 1 standard deviation.

In South Asia, climate change also enhances O₃ concentrations by <5 % over GP and SI (Fig. 8), due to the massive precursor emissions (Fig. S1) and O₃ productions. Over SI, the decreases in relative humidity (Fig. S3) and cloud amount (Fig. S5) and increases in downward solar radiation at the surface (Fig. S4) favor photochemical production of O₃ and induce the large increases in O₃ concentrations in this region. Averaged over SEA, O₃ concentrations driven by higher temperature (Fig. S2), more downward solar radiation (Fig. S4), and lower relative humidity (Fig. S3) and cloud cover (Fig. S5) in 2095 are projected to increase O₃ concentrations by 5 %–7 % in SSP3-7.0 and SSP5-8.5 and 0 %–3 % in SSP1-2.6 and SSP2-4.5 scenarios, relative to 2025 (Fig. 8).

The Tibetan Plateau (TP), known as the highest topography in China with more solar radiation at the surface, has strong stratosphere–troposphere exchanges of O₃ compared with other regions, leading to high O₃ concentrations over this region (Fig. S6). Climate-driven O₃ concentrations are projected to decline by less than 2 % over TP from 2025 to 2095 (Fig. 8). It is likely because less solar radiation (Fig. S4) and more frequent occurrence of rainy weather (Fig. S7) in the future would reduce the local chemical production of O₃.

Figure 9 The spatial distributions of percentage differences (%) in near-surface O₃ concentrations between 2025 (2020–2029 mean) and 2095 (2091–2100 mean) driven by climate change under four scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5, from left to right) averaged in MAM (March–April–May), JJA (June–July–August), SON (September–October–November), and DJF (December–January–February) (from top to bottom). No overlaying hatch pattern indicates statistical significance with 95 % confidence from a two-tailed t test.

3.3 The seasonality of future climate-driven O₃ variations

Climate over Asia has obvious seasonal variation related to the Asian monsoon system. Figure 9 shows the spatial distributions of percentage changes in projected climate-driven O₃ concentrations in spring (March–April–May, MAM), summer (June–July–August, JJA), autumn (September–October–November, SON), and winter (December–January–February, DJF) between 2025 and 2095 under the four scenarios. In general, air quality in many regions of Asia will deteriorate in all seasons associated with intensified O₃ pollution under climate change.

In eastern China, O₃ pollution occurs most frequently in summer and is more severe in NC than SC currently (Li et al., 2019). Under future climate warming, JJA O₃ concentrations will increase by 5 %–20 % in SC under the high-forcing scenarios, while the changes in NC are less than 5 %. This suggests that future climate change will expand the summertime O₃ pollution from NC to SC over eastern China. Another feature is the strong increases in O₃ concentrations by 10 %–20 % throughout eastern China and exceeding 20 % over Sichuan Basin in SON, which relate to the significant increases in temperature (Fig. S8) and solar radiation (Fig. S9) in this season over central-eastern China under the high-forcing scenarios. It further indicates that future climate change will extend the O₃ pollution from summer into autumn.

In South Asia, the climate-driven increases in O₃ concentrations vary from JJA over SI to DJF over GP. Relative to 2025, in the summer of 2095, anomalous high pressure (Fig. S10) along with an anticyclone (Figs. S11 and S12) dominates South Asia, which is not conducive to O₃ diffusion, leading to increases in JJA O₃ concentrations over SI. The intensified O₃ pollution across GP in DJF under climate change is related to the strong surface warming (Fig. S8), decreases in relative humidity (Fig. S13), cloud cover (Fig. S14), and rainfall (Fig. S15), as well as increases in solar radiation at the surface (Fig. S9), favoring the photochemical production of O₃. In the northern part of Southeast Asia, JJA has the largest O₃ rise via the same mechanism as SI, while O₃ increases by the same magnitude in all seasons in the southern part of Southeast Asia driven by future climate change.

4 Conclusions and discussion

O₃ pollution has been increasing over Asia in recent decades, which harms human health and vegetation. In the future warmer climate, O₃ pollution over Asia can be modulated by changes in meteorological fields. In this study, to examine the variations in O₃ concentrations over Asia due to the future climate change, monthly near-surface O₃ concentrations from 2020 to 2100 under four climate scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) are predicted using an ML model with input data from assimilated O₃ combining GEOS-Chem simulations and real-time observations, future meteorological parameters from CMIP6 multi-model simulations, emissions of O₃ precursors, land use, topography, population density, and spatiotemporal information. Our results suggest that future O₃ pollution over Asia will be significantly exacerbated in a warming climate, especially under high-forcing scenarios.

Trained by the assimilated O₃ concentrations and reanalysis data, the ML model can well predict O₃ over Asia with the coefficient determination of 0.92 between assimilated and predicted O₃ concentrations and relative error of 9 %. Then the future Asian O₃ concentrations from 2020 to 2100 driven by climate change are projected in the ML model with varying meteorological fields from 18 CMIP6 models under four future climate scenarios.

The climate penalty on O₃ is robust over most regions of Asia. The annual mean O₃ levels in 2095 are projected to increase by 5 %–20 % relative to 2025 under the high-forcing scenarios over South China, Southeast Asia, and South India and less than 10 % over North China and the Gangetic Plains, due to more favorable meteorological conditions for O₃ photochemical production, while there is a decrease of <5 % over the Tibetan Plateau. The climate-driven changes in O₃ concentrations are smaller under the scenarios of less warming, suggesting that future climate following low emissions and sustainable pathways would be more effective in the mitigation of O₃ pollution in Asia than the high-forcing scenarios. Seasonal variation analysis reveals that summertime O₃ pollution over eastern China will expand from North China to South China and extend into the cold season under future climate change. In addition, South Asian O₃ pollution will increase over South India in summer and over the Gangetic Plains in winter.

Zanis et al. (2022) analyzed the global climate change benefit and penalty of O₃ based on sensitivity simulations from five CMIP6 models under the SSP3-7.0 scenario. They showed positive changes in JJA O₃ concentrations by less than 1 ppb from 2010 to 2095 over East Asia and South Asia driven by climate change but with large uncertainties due to the model diversity. The ML method in this study gives similar positive changes in O₃ as Zanis et al. (2022). Pommier et al. (2018) applied the EMEP (European Monitoring and Evaluation Programme) chemical transport model driven by the downscaled meteorological data from NorESM1-M to investigate the impacts of regional climate change on near-surface O₃ over India. They showed that near-surface O₃ would increase by up to 4 % over northern India and decrease by 3 % over southern India from 2050 to 2100 under the RCP8.5 scenario. We show that the climate-driven O₃ in this study would increase over both the Gangetic Plains (0.2 %) and South India (3 %) under the SSP5-8.5 scenario in 2050 relative to 2016 (2014–2019 mean). The discrepancies may rise from that the results of Pommier et al. (2018) were based on NorESM1-M simulated climate alone, while the climate change predicted by 18 CMIP6 models were applied in this study and the ensemble mean O₃ concentrations were shown here.

There are a few uncertainties and limitations in the projected near-surface O₃ concentrations over Asia in terms of input data, GEOS-Chem simulations, CMIP6 multi-model simulations, and the ML model. First, only observational data over 2014–2019 across China were used for the O₃ assimilation. Longer-term measurements with broader spatial coverage are more desirable to improve the model performance. Land use data and population density are fixed at present-day conditions when predicting the future O₃ since we focus on the variations in meteorological parameters under climate change, which will vary in the future. In addition, natural O₃ precursor emissions such as biogenic emissions of NMVOCs and NO_x from soil and lightning sources are fixed at levels from the year 2016 in the future estimates, which can induce biases in the O₃ projections since climate change can strongly influence natural emissions of O₃ precursors (Liu et al., 2019). Although the methane concentrations in the GEOS-Chem model are prescribed and its role in the O₃ production is not considered in the ML model, the climate influence of methane is included in the CMIP6 multi-model simulations. Consequently, the impact of future changes in methane on O₃ concentrations via climate change are considered in the future projections.

Second, the GEOS-Chem model has been demonstrated to well capture the magnitude of and spatiotemporal variations in O₃, with an average bias of about 10 % over China (Lou et al., 2014) and Southeast Asia (Marvin et al., 2021) and less than 20 % over India (David et al., 2019). The future decrease in relative humidity will cause stomatal closure and also increase near-surface O₃. The O₃–vegetation interactions are not represented in the default GEOS-Chem model. A newly coupled global atmospheric chemistry–vegetation model (Lei et al., 2020) could be applied in the future study.

Third, the meteorological parameters characterizing future climate change from the CMIP6 multi-model simulations can also give rise to uncertainties in this study (Xu et al., 2021). Moreover, the spatial autocorrelation in a random split of training data for cross-validation would lead to the overly optimistic statistics of ML model predictive power (Ploton et al., 2020). Additionally, the overall importance scores of the features in this study can only reflect that from the whole study domain. Further investigations are required to identify and quantify the importance score of each local variable contribution to the near-surface O₃ predictions in different specific regions. Also, the good ability of the ML model for the present-day condition may not imply a satisfactorily extrapolation under the future warming condition, which can bias our results and deserves further investigation in future studies.

Last but not least, near-surface O₃ has increased rapidly in China since 2013 owing to both precursor emission changes and atmospheric warming (Li et al., 2021b), which significantly affect human health (Lu et al., 2020) and also require further studies.

Overall, our study provides a framework of combining real-time observations, chemical transport model simulations, and multi-climate model predictions with data assimilation and machine learning methods to estimate future climate-driven near-surface O₃ concentrations. The emphasis of this work is to quantify the impacts of future climate change on O₃ pollution in Asia, which is of great significance for future O₃ pollution mitigation strategies.