Future Co-Occurrences of Hot Days and Ozone-Polluted Days Over China Under Scenarios of Shared Socioeconomic Pathways Predicted Through a Machine-Learning Approach

The warming climate increases the probability of hot days, which leads to a penalty effect increasing ozone (O 3 )-polluted days in polluted regions. Here, we established a random forest algorithm to predict future probabilities of O 3 exceedance ( P ) during hot days and further examined the future co-occurrences of O 3 -polluted days and hot days under two different scenarios of Shared Socioeconomic Pathways (SSPs) 1-2.6 and 5-8.5 for 2030–2050s. Ground-level observations, simulated O 3 -temperature sensitivities in the GEOS-Chem model, multimodel seasonal-mean O 3 concentrations and daily maximum temperature from the sixth Coupled Model Intercomparison Project (CMIP6), leaf area index from reanalyzed data, and local geographical information were comprehensively utilized. Evaluations showed that the algorithm captured the spatial patterns of present-day P values well with a correlation coefficient of 0.92 over China. Results showed that the strong reductions in anthropogenic emissions under SSP 1-2.6 significantly reduced the risks of O 3 exceedance during hot days nationwide from 3.7 days in the 2030s to 3.0 days in the 2050s. However, the SSP 5-8.5 scenario witnessed more frequent co-occurrences of O 3 -polluted days and hot days over the 2030–2050s with nationally averaged values from 4.5 to 6.4 days. Our results highlight the co-benefits

summertime mean O 3 concentrations as well as the frequencies of O 3 -polluted days (defined as average maximum daily 8-hr (MDA8) O 3 concentrations larger than 80 ppbv in this study based on Gong et al. (2020)) have been increasing over eastern China (Gong et al., 2020;Li et al., 2019;Lu et al., 2020). Understanding how global changes, especially changes in anthropogenic emissions and climate, influence the frequency of O 3 -polluted days is essential to execute proper control strategies.
Temperature was identified to have the strongest impacts on O 3 concentrations among all meteorological factors in most regions with high anthropogenic emissions (Dawson et al., 2007;Shi et al., 2020;Steiner et al., 2006). High temperature increases O 3 concentrations by processes that include enhancing biogenic emissions of volatile organic carbon (BVOCs) Ma et al., 2019) and accelerating chemical reactions . Specifically, O 3 -polluted days tend to occur on hot days (defined as daily maximum temperature (Tmax) exceeding 35°C in this study) (Pu et al., 2017;J. L. Schnell & Prather, 2017). For example, Solberg et al. (2008) examined the strong heat waves in June-August 2003 over Europe and showed that extremely high temperatures led to maximum O 3 concentrations exceeding 230 μg m −3 at most observed sites. Analitis et al. (2014) analyzed data in nine European cities from 1990 to 2004 and showed that high O 3 concentrations during hot days increased daily deaths by 54% among people aged 75-84 years compared to the deaths induced by hot days alone, suggesting composite health risks associated with co-occurrence of hot days and O 3 -polluted days.
The Sixth Assessment Report by the Intergovernmental Panel on Climate Change (IPCC) (https://www.ipcc. ch/report/ar6/wg1/) showed that global temperature will continue to increase until 2050s even with the strictest emission control measures under scenarios of Shared Socioeconomic Pathways (SSPs) 1-1.9. A warmer climate has a high likelihood of bringing a "penalty effect" on O 3 in polluted regions (Bloomer et al., 2009;Fu & Tian, 2019). For instance, Colette et al. (2015) composited projections of 25 models from 11 published studies and found that future climate scenarios have penalty effects on summertime surface O 3 of at most 5 ppbv over Europe from 2071 to 2100. Schnell et al. (2016) assembled projected O 3 from four global models with fixed anthropogenic ozone precursors. They showed that warming under the RCP8.5 scenario increased the summermean O 3 of about 4-10 ppbv in eastern U.S, 1-5 ppbv in southern Europe, and 3-6 ppbv in northeastern Asia in the 2100s compared to the 2000s. To better indicate the intensity of the "climate penalty on O 3 ," Wu et al. (2008) proposed a "climate penalty factor," which is represented by the sensitivity of O 3 to Tmax (dO 3 /dT) and has been widely used (Bloomer et al., 2009;Fu et al., 2015;Gu et al., 2020;Rasmussen et al., 2013;Zhao et al., 2013). In addition, a warmer climate increases the probability of hot days (Ding et al., 2010;Perkins-Kirkpatrick & Lewis, 2020), which further increases the probability of the occurrences of severe O 3 -polluted days. However, to our knowledge, the future co-occurrence of O 3 -polluted days and hot days under different emission scenarios has rarely been examined.
Simulations of future hot days and future O 3 -polluted days usually rely on Earth System Models (ESMs) with fully coupled NO x -O 3 -hydrocarbon-aerosol chemical mechanisms (Meehl et al., 2018). However, current ESMs generally use a simplified O 3 chemistry scheme (Lamarque et al., 2013) to reduce the computational burden, which would cause uncertainties in the simulated future O 3 . Furthermore, a quite limited number of ESMs (to our knowledge, only one model, namely UKESM1-0-LL) in the recent Sixth Coupled Model Intercomparison Project (CMIP6) simultaneously provide both future daily temperature and hourly O 3 concentrations under different projected emission scenarios (e.g., SSP1-2.6 and SSP5-8.5) (https://esgf-node.llnl.gov/projects/cmip6/). Instead, the multimodel-predicted monthly mean Tmax and O 3 concentrations, which potentially represent the risks of co-occurrence of hot days and O 3 -polluted days, are assessable. As a result, we aim to fully utilize the multimodel results to avoid the one-model dependence problem and better predict future co-occurrence of hot days and O 3 -polluted days.
In this study, we established a machine-learning approach by using a random forest algorithm to quantify the climatological risk level of O 3 pollution during hot days and then further predicted the future co-occurrence of O 3 -polluted days and hot days under two different projected emission scenarios (SSP1-2.6 and SSP5-8.5). Section 2 describes the framework of this study based on the random forest algorithm and further introduces the data, including site-level observed data, reanalyzed meteorological data, simulations from the 3D global chemical transport model (GEOS-Chem), and multimodel results from CMIP6. In the results section, the current risk level of O 3 pollution during hot days over China since 2013 is shown. Then, we establish and evaluate the machine-learning approach and predict the future co-occurrence of O 3 -polluted days and hot days. Finally, we draw conclusions and discuss the uncertainties and implications of this study.

Framework of Predicting the Future Co-Occurrences of O 3 -Polluted Days and Hot Days
In this study, we introduced the probability of O 3 exceedance (P) during hot days to reflect the climatological risk level of co-occurrence of O 3 -polluted days and hot days. Present-day P can be represented by the following equation: where HD is the number of hot days, and OD H is the number of O 3 -polluted days that occurred on all hot days at one specific site or grid. The P value is calculated only when the HD is larger than 3 days over May-October in 1 year.
We aimed to predict future P values through a framework of the random forest algorithm with the help of multisource data ( Figure 1). To begin with, the random forest algorithm was trained and evaluated to predict the present-day P values in China over May-October in 2014-2019 based on observed data. Then, we applied the established random forest algorithm to predict future P values in the 2030, 2040s, and 2050s under SSP 1-2.6 and SSP 5-8.5 scenarios with the help of GEOS-Chem simulations (providing O 3 -Tmax sensitivities) and CMIP6 multimodel predictions (providing daily mean Tmax and monthly mean O 3 concentrations). Finally, we derived future co-occurrences of O 3 -polluted days and hot days by multiplying future P values from the random forest algorithm and projected hot days in CMIP6 ensembles.
In particular, we applied the O 3 -Tmax sensitivities (S) following the definition in Wu et al. (2008) to better achieve the prediction of P: Figure 1. Summary of the procedures to predict future co-occurrences of O 3 -polluted days and hot days. The red parameters in the present day input indicate the selected parameters used in the random forest algorithm (see Table 1 and Section 3.2).
where [MDA8 O 3 ] and Tmax are the daily series of MDA8 O 3 concentrations and Tmax over May-October at a specific site or grid, respectively; and S and b are the annual sensitivity of MDA8 O 3 concentrations to temperature and the residuals, respectively, both of which are obtained by least square fitting. The present-day S values were derived from site-level observations (see details in Section 2.3), while the future S values under different scenarios were from GEOS-Chem simulations (see details in Section 2.4). Since S represents the enhancement of O 3 concentrations with an increase in temperature of 1°C, high S values implied a high probability of the occurrence of O 3 -polluted days during hot days.

Random Forest Algorithm
As a traditional machine-learning approach with the capability to solve complex regressions and classification problems, the random forest algorithm has been widely used in the field of atmospheric chemistry in recent years (Araki et al., 2018;Chen et al., 2018;Kaminska, 2019;Wang et al., 2019;Wei et al., 2019;Zhan et al., 2018). A random forest consists of a large number of individual decision trees, which are grown on dependent bootstrap samples from training data sets. Each decision tree divides bootstrap samples into branches through decision nodes until a leaf node is attained. Leaf nodes represent predicted estimates of one simple decision tree, and the mean prediction of the individual trees is determined as the regression output of the random forest algorithm (Breiman, 2001).
In this study, the function RandomForestRegressor of the python package scikit-learn was applied (Pedregosa et al., 2011). To predict the probability of O 3 exceedance (P in Equation 1) during hot days, we composited a series of related predictors, including O 3 -Tmax sensitivity (S in Equation 2), latitude, longitude, seasonal-mean meteorological factors, O 3 concentrations, reanalyzed leaf area index (LAI), and anthropogenic NO x and VOCs emissions, to test the random forest algorithm. The number of decision trees was 200. To be consistent with the future grid-level input data in the random forest model, we used the 1° × 1° (rather than site-level) predictors for 2014-2017 from the observations and reanalysis as the training data and evaluated the random forest by using data in 2018-2019 with two indicators: correlation coefficient (R) and root mean square error (RMSE).

Observed O 3 and Meteorological Data
The ground-level hourly observed O 3 concentrations were obtained from the monitoring network established by the Ministry of Ecology and Environment (MEE) of China (http://www.cnemc.cn). We then derived the MDA8 O 3 concentrations for all 1,582 sites ( Figure S1 in Supporting Information S1) over May-October in 2014-2019 and interpolated site-level data to 1° × 1° for further application in the random forest algorithm.
Four observed daily meteorological factors, namely the Tmax, daily mean relative humidity (RH), pressure (PRS), and wind speed (WS), were obtained from ground-level weather monitoring stations maintained by the China Meteorological Administration (http://data.cma.cn/). The meteorological data observed at more than 600 sites ( Figure S1 in Supporting Information S1) were interpolated into O 3 observed sites by averaging the meteorological values at the nearest five sites within a 2-degree square domain.

GEOS-Chem-Simulated O 3 -Tmax Sensitivities
The GEOS-Chem model is a state-of-art 3D global chemistry transport model with good capabilities in simulating the spatiotemporal variations in gas phase pollutants and aerosols (Dang & Liao, 2019;Li et al., 2019;Porter & Heald, 2019). It is driven by MERRA2 reanalyzed data from the Goddard Earth Observing System (GEOS) of the NASA Global Modeling and Assimilation Office (GMAO) and employs a fully coupled NO x -O x -hydrocarbon-aerosol chemistry mechanism. The vertical resolution of GEOS-Chem is 47 layers from the surface to 0.1 hPa. Photolysis rates are computed by the Fast-JX scheme (Bey et al., 2001;Park et al., 2003;Pye et al., 2009). Stratospheric O 3 is simulated by linearized ozone parameterization (LINOZ) following McLinden et al. (2000). A nonlocal scheme in planetary boundary layers is applied to describe the vertical mixing process (Lin & McElroy, 2010). The Model of Emissions of Gases and Aerosols from Nature (MEGAN v2.1) is utilized to dynamically calculate biogenic VOCs emissions (Guenther et al., 2012).
In this study, we first performed simulations over May-October during 2014-2019 in the Asian nested domain (11°S-55°N, 60°-150°E) with horizontal resolutions of latitude 0.5° × longitude 0.625°. The lateral boundary layers of all chemical tracers were provided by global simulations at a resolution of 2° latitude × 2.5° longitude. Anthropogenic emissions over China were obtained from the annually updated Multi-resolution Emission Inventory (MEIC, http://www.meicmodel.org) (Zheng et al., 2018). It should be noted that the present-day simulated results were not used in the random forest algorithm but were designed to evaluate model performance in simulating O 3 -Tmax sensitivities and to further revise future simulated S values.
Since many previous studies have shown that the decadal variations in S values in polluted regions were determined by the changes in anthropogenic emissions (Bloomer et al., 2009;He et al., 2013;Jing et al., 2017;Rasmussen et al., 2013;Wu et al., 2008), we further derived future S values from GEOS-Chem simulation with fixed meteorology in 2015 and future anthropogenic emissions (years 2030, 2040, and 2050) under SSP 1-2.6 and SSP 5-8.5 scenarios. Furthermore, those simulations were only performed at a global resolution of 2° latitude × 2.5° longitude to better integrate data from climate-system models in CMIP6 with coarse horizontal resolutions.

CMIP6 Multimodel Outputs
To perform a reasonable prediction of future changes in temperature and surface O 3 , we applied the daily Tmax data from 18 climate models in CMIP6 and the monthly mean surface O 3 data from 11 chemistry climate coupling models (Table S1 in Supporting Information S1) (downloaded from https://esgf-node.llnl.gov/projects/cmip6/). Model outputs with the "r1i1p1f1" variant in two scenarios (SSP 1-2.6 and SSP 5-8.5) were selected. The horizontal resolutions of variables from different models were integrated to a resolution of latitude 2° × longitude 2.5° by bilinear interpolation. Seasonal-mean (May-October) values (such as Tmax and O 3 concentrations) as well as the number of hot days were further derived for the present-day (2015), 2030s (averaged over 2028-2032), 2040s (averaged over 2038-2042), and 2050s (averaged over 2048-2052).

Auxiliary Data
Two sets of data were applied as an auxiliary to establish the random forest algorithm. One data set was the daily grid-level LAI from the MERRA2 reanalyzed data set with horizontal resolutions of latitude 0.5° × longitude 0.625°, considering the significant impacts of vegetation on O 3 as revealed by many previous studies Lin et al., 2020). The other data set was the monthly mean anthropogenic NO    Figure S3 in Supporting Information S1) are likely important factors, and the changes in meteorology and anthropogenic emissions may also be influential. To quantify the complex and nonlinear relationships between P values and those potential predictors, the random forest algorithm was established and evaluated to predict the future risks of the co-occurrence of O 3 -polluted days and hot days.

Establishment and Evaluation of the Random Forest Algorithm
We used the 1° × 1° grid-level data over 2014-2017 as the training data and designed eight random forest experiments (EX1-EX8) to select the most effective and essential predictors to predict P values. All experiments included the predictor of O 3 -Tmax sensitivity S because high S values implied a high probability of the occurrence of O 3 -polluted days during hot days. In EX1, only S was considered, and we further added latitude-longitude geographic information, seasonal-mean O 3 concentrations, Tmax, LAI, other meteorological factors (RH, PRS, and WS), and anthropogenic NO x and VOCs emissions step-by-step for EX2-8. A detailed description of each predictor is shown in Table S2 in Supporting Information S1. Table 1 shows the random-forest score (the coefficient of determination) and RMSE over all grids in China between the observed and random-forest-predicted P values during the warm season in 2018-2019. The variables S, latitude, longitude, and seasonal-mean O 3 concentrations were the most effective predictors of P, and the score reached 0.72 and the RMSE decreased to 15.5% in EX3. Including the seasonal mean Tmax and LAI slightly improved the performance of random forest although the other three meteorological factors as well as the anthropogenic NO x and VOCs emissions had opposite effects. As a result, the configuration of EX5 was determined as the final random forest algorithm.   Figure S4a in Supporting Information S1 shows high spatial correlations between site-level observed and GEOS-Chem-simulated S values with R = 0.75. However, the simulated S values were significantly underestimated, especially in severely polluted Note. Detailed descriptions of each predictor are shown in Table S2 in Supporting Information S1. Score (the coefficient of determination) and RMSE were derived between the observed and random-forest-predicted P values on 1° × 1° grids in China averaged over May-October during 2018-2019 following these two formulas:

GEOS-Chem-Simulated S Values
where P obs,i,j and P rf,i,j indicate the observed and random-forest-predicted P values in the gird with i latitude and j longitude. are observed P values averaged over all grids in China. n represents the total number of 1° × 1° grids in China.

Table 1 Eight Experiments of the Random Forest Algorithm Using Different Predictors
regions such as North China, which were likely to be induced by the underestimates of extremely high O 3 concentrations in GEOS-Chem when hot days occurred (Gong & Liao, 2019;Gong et al., 2020).
To alleviate the underestimates, we further revised the simulated S values in the present-day simulations and applied the revision in future GEOS-Chem simulations. To begin with, we found that the underestimates were mainly induced by regions with high S. For all grids with the observed S larger than 2 ppbv K −1 , the mean observed S was 2.7 ppbv K −1 , while the mean simulated S only reached 1.8 ppbv K −1 . As a result, we applied a ratio of 1.5 (2.7/1.8) to revise the S values for all simulated girds and then kept S that is lower than 2 ppbv K −1 unchanged as the original simulated values. In this way, the normalized mean bias was significantly reduced from −33.8% to −9.8% nationwide ( Figure S4b in Supporting Information S1).

Selection of CMIP6 Models
We selected 5 model outputs of Tmax from 18 CMIP6 models, including AWI-CM, CanESM5, CMCC-ESM2, INM-CM4-8, and MPI-ESM1-HR, by comparing the daily Tmax under the SSP 5-8.5 scenario in 2015 to MERRA2 reanalyzed data over China following the three criteria: (a) correlation coefficients of seasonal mean Tmax between simulations and MERRA2 at all grids over China should exceed 0.5 ( Figure S5 in Supporting Information S1); (b) the mean bias of seasonal-mean Tmax over China should be lower than 1°C ( Figure S6 in Supporting Information S1); and (c) the mean bias of hot days over China should be lower than 5 days ( Figure S7 in Supporting Information S1). The comparison of MERRA2 to CMIP6 model outputs under the SSP 1-2.6 scenario is not shown here because similar patterns with SSP 5-8.5 were induced by the same anthropogenic climate forcings of 2015.
The patterns of seasonal-mean O 3 concentrations from 11 chemistry climate models in 2015 are shown in Figure S8 in Supporting Information S1. Only 5 models showed valid O 3 concentrations over the mainland, and MPI-ESM-2-HR and MPI-ESM-2-LR (also NorESM2-LM and NorEMS2-MM) were not independent of each other. To obtain as much independent data from different models as possible, we averaged the simulated future O 3 concentrations in BCC-CSM2-MR, MPI-ESM1-2-HR, NorESM2-LM, and the simulated O 3 concentrations in GEOS-Chem with future emissions but fixed the meteorology as the future seasonal-mean O 3 fields to predict future P values in the random forest algorithm. The GEOS-Chem simulations were also used because the global transport chemistry model generally has more comprehensive chemical mechanisms than most of the climate models (Lamarque et al., 2013). Although the climate conditions were fixed in 2015, the GEOS-Chem model applying future anthropogenic emissions was still able to describe reasonable patterns of seasonal-mean O 3 concentrations ( Figure S8 in Supporting Information S1) and further balance the bias in climate models.

Predicted Future Co-Occurrences of O 3 -Polluted Days and Hot Days
To better understand the predicted results from the random forest algorithm, we first compared two different emission scenarios (SSP 1-2.6 and SSP 5-8.5). Since it is well known that SSP 1-2.6 and SSP 5-8.5 represent low and high levels of future greenhouse gas emissions, respectively (Meinshausen et al., 2020); here, we focus on the anthropogenic emissions of air pollutants (such as ammonia (NH 3 ), NO x , VOCs, sulfur dioxide (SO 2 ), and carbon monoxide (CO)). In the SSP 1-2.6 scenario, almost all air pollutants in China show marvelous emission reductions in the future, except for NH 3 (Figure 3a). Especially, the two most important O 3 precursors, NO x and VOCs, reduced from 31.8 Tg to 16.1 TgC in 2015 to 10.1 Tg and 2.4 TgC in 2050, respectively. Such drastic changes are mainly contributed by the reductions in eastern China (Figures S9 and S10 in Supporting Information S1), implying where the future O 3 pollution would be significantly alleviated under the SSP 1-2.6 scenario. In comparison, although the emissions of SO 2 and CO in the SSP 5-8.5 scenario also show a significant reduction from 2015 to 2050, other species, including NO x and VOC, change within a much smaller magnitude compared to SSP 1-2.6 over China (Figure 3b). Furthermore, the emissions of O 3 precursors in eastern China remain at a high level in the future (Figures S9 and S10 in Supporting Information S1).
We applied the random forest algorithm (Section 3.2) with S derived from the GEOS-Chem simulation (described in Section 2.4), seasonal-mean Tmax and O 3 concentrations from CMIP6 models (Section 3.3.2) and the LAI in 2015 to examine future P values under SSP1-2.6 and SSP 5-8.5 scenarios. Future S value, seasonal-mean O 3 concentrations, and Tmax are shown in Figures S11-S13 in Supporting Information S1. The LAI values from MERRA2 were fixed in 2015 since the carbon cycles in most of the CMIP6 models remained large uncertainties (Arora et al., 2020). Figure 4 shows that the nationally averaged P values decreased to 29.9% in the 2050s following the SSP 1-2.6 scenario. The risk of the co-occurrences of O 3 -polluted days and hot days was at a low level in most regions over China although the P values in North China remained at 48.7% in the 2050s. However, the P values under the SSP 5-8.5 scenario were predicted to remain high during the 2030-2050s with values ranging from 37.3% to 37.7% nationwide and 70.1%-70.5% averaged over North China. North Xinjiang Province, central eastern China, and northeast China all presented a high composited risk of O 3 exceedance during hot days. Figure S13 in Supporting Information S1 shows the future number of hot days averaged from the 5 selected CMIP models under different scenarios. Following both SSP 1-2.6 and SSP 5-8.5 scenarios, the frequencies of hot days were highest in Northwest China from May to October. Since the strict reductions in anthropogenic carbon emissions (SSP 1-2.6) still failed to stop climate warming ( Figure S14 in Supporting Information S1), the frequencies of hot days showed a slightly increasing trend over central eastern China from the 2030-2050s. SSP 5-8.5 witnessed a more rapid increase in the frequency of hot days compared to SSP 1-2.6. In particular, the number of hot days in North China increased from 12.1 days in the 2030s to 15.3 days in the 2050s, implying a higher risk of the co-occurrence of O 3 -polluted days and hot days.
By multiplying the P values in Figure 4 and the number of hot days in Figure S13 in Supporting Information S1, we were able to predict the future frequencies of the co-occurrences of O 3 -polluted days and hot days ( Figure 5). Under the SSP 1-2.6 scenario, the composited risk of O 3 pollution during hot days significantly decreased, especially in North China, where the co-occurrence of O 3 -polluted days and hot days decreased from 7.6 days  2030, 2040, and 2050, respectively. in the 2030s to 4.9 days in the 2050s. However, SSP 5-8.5 predicted a worse future in which the co-occurrences of O 3 -polluted days and hot days increased in North China from the 2030-2050s. These results implied large benefits in alleviating the composite risk of extreme weather events and severe air-polluted days through anthropogenic emission reductions.

Comparison Between the Chemistry Climate Model Predictions and the Random Forest Algorithm
We compared our predicted future co-occurrences of hot days and O 3 -polluted days with the chemistry-climate model results from the UKESM1-0-LL, which to our knowledge was the only one model in CMIP6 that provide future daily temperature and hourly O 3 concentration data under both SSP 1-2.6 and SSP 5-8.5 scenarios. Note that the UKESM1-0-LL results here were with the "r1i1p1f2" variant so they were not included in our random forest algorithm (see the CMIP6 model description in Section 2.5).
To begin with, we evaluated the UKESM1-0-LL model performance by comparing the observed and simulated seasonal-mean Tmax and MDA8 O 3 concentrations over May-October in 2015 ( Figure S15 in Supporting Information S1). The site-level observations and the origin model outputs with a resolution of 1.25° latitude × 1.875° longitude were both interpolated into 2° latitude × 2.5° longitude. Although the UKESM1-0-LL model-simulated reasonable patterns of seasonal-mean Tmax, the O 3 concentrations were strongly overestimated (normalized mean bias exceeding 50% over China), which would definitely bring in large uncertainties in future prediction. As a result, we revised the future simulated daily Tmax and MDA8 O 3 concentrations by applying the linear regression relationships (Figures S15d and S15h in Supporting Information S1) between observations and simulations in 2015 under different SSP scenarios, and then obtained the UKESM1-0-LL-predicted future co-occurrences of hot days and O 3 -polluted days over China.
The UKESM1-0-LL predictions (after revision) in the 2050s showed high risks of co-occurrences of hot days and O 3 -polluted days not only in North China but also in northwest China ( Figure S16 in Supporting Information S1), especially under the SSP 5-8.5 scenario, which was similar to our results from the random forest algorithm ( Figure 5). Furthermore, the UKESM1-0-LL model-predicted national averaged days of co-occurrences decreased from 3.8 days in 2030s to 1.9 days 2050s under the SSP 1-2.6 scenario, but increased from 6.9 days in the 2030s to 13.7 days in the 2050s under the SSP 5-8.5 scenario. Although the specific numbers were different from our predictions ( Figure 5), which might led by the underestimates of P values in the random forest algorithm, especially in North China or the overestimates in O 3 concentrations in UKESM1-0-LL, the UKESM1-0-LL model generally predicted similar future spatiotemporal variations of co-occurrences of hot days and O 3 -polluted days as our results, suggesting that our random forest algorithm was capable to achieve the prediction within the reasonable bias.

Conclusions and Discussions
In summary, we predicted the future climate penalty on O 3 pollution from the perspective of extreme cases, which, to our knowledge, has rarely been examined. Since multimodel simulations of future daily O 3 concentrations were unavailable, we aimed to derive the climatological risk factor P, which was the probability of O 3 exceedance during hot days, through a random forest algorithm and further predicted the future co-occurrence of O 3 -polluted days and hot days. Evaluations showed that the random forest algorithm was capable to derive a reasonable pattern of P values with a mean bias lower than 10% in most regions over China despite the underestimates in North China. By combining GEOS-Chem simulations and CMIP6 multimodel outputs, our results showed that both the SSP 1-2.6 and SSP 5-8.5 scenarios increased the future hot days nationwide. The SSP 1-2.6 scenario significantly reduced the composite risks of O 3 exceedance during hot days with P values reduced from 64.7% in the 2030s to 48.7% in the 2050s due to the strict reductions in anthropogenic emissions. However, the P values remained high under the SSP 5-8.5 scenario in eastern China, thus leading to increases in the co-occurrences of O 3 -polluted days and hot days until the 2050s.
Uncertainties in the predicted co-occurrences of O 3 -polluted days and hot days may be caused by three factors: algorithm uncertainties, model data uncertainties, and exclusion of variations in terrestrial vegetation. First, the random forest algorithm showed insufficient ability to predict extremely high P values (Figure 2). Even though we have tried almost all related parameters as predictors in EX8 (Table 1), such underestimates remained. As a result, the predicted P values under the SSP 5-8.5 scenario shown in Figure 4 might be underestimated, further underestimating the future co-occurrence of O 3 -polluted days and hot days. Second, although we applied the ensemble mean from the multimodel outputs to avoid the problem of model dependence, uncertainties could still be induced by large discrepancies in variables among different CMIP6 models, especially for the prediction of future hot days ( Figure S17 in Supporting Information S1). Meanwhile, all 5 selected climate models overestimated the number of hot days in northwest China ( Figure S7 in Supporting Information S1), which might exaggerate the co-occurrence days in Xinjiang Province as shown in Figure 5. Furthermore, the variations in terrestrial vegetation could influence O 3 concentrations through BVOCs emissions and stomatal dry depositions on decadal time scales (Lin et al., 2020;Wang et al., 2020), which were not considered in this study because most of the CMIP6 models lacked a dynamic vegetation module (Arora et al., 2020); thus, large uncertainties remained in predicting future vegetation.
Our approach applied the O 3 -Tmax sensitivity rather than the direct NO x and VOCs emissions (Table 1) because it seems difficult for the random forest to well represent the complex nonlinear chemical reactions with only emission inputs (Keller & Evans, 2019). We tried to replace the S value with the NO x and VOCs emissions in EX5 but found out random forest score of 0.64 and an RMSE value of 17.2% between observed and random-forest-predicted P values over China (not shown in Table 1), indicating a poorer performance compared to the origin EX5 configuration. Nevertheless, using the GEOS-Chem model to describe future S values still brought in new uncertainties considering the model bias ( Figure S4 in Supporting Information S1) and fixed meteorological fields. A dilemma was confronted that either suffering the low performance of random forest with exact emissions or admitting the simulated bias in a global transport chemistry model. However, it should be clarified that despite the direct NO x and VOCs emissions were omitted (Table 1), future changes in anthropogenic emissions under different SSP scenarios were comprehensively considered in our approach through the GEOS-Chem-simulated S values.
Despite the uncertainties, we established an effective machine learning method to predict the composite risks of the co-occurrences of extreme weather events and air-polluted days by considering future changes in both anthropogenic emissions and climate. The algorithm avoided the model dependence problem by taking full advantage of the CMIP6 multimodel outputs. Our results highlight that the reductions in anthropogenic emissions can significantly alleviate O 3 pollution during hot days with the implications of benefits of reducing the related health risks.