Rapid changes in air pollution have occurred in China over the last few decades because of dramatic transformations in economic development and air pollutant emissions. Following substantial increases in emissions in the 1990s and 2000s, nationwide pollutant emission controls since 2013 have led to remarkable reductions in fine particulate matter (PM2.5), with national-population-weighted annual mean concentrations decreasing from 62 to 42 µgm-3 during 2013–2017 . However, surface ozone (O3) pollution is becoming increasingly prevalent in China despite these emission controls, as recent emission policies have primarily targeted fine particles . Reductions in the emissions of nitrogen oxides (NOx), a precursor of both O3 and fine particles, may lead to increased O3 concentrations due to non-linear O3 chemistry and to strengthened incoming solar radiation . In addition, anthropogenic emissions of other O3 precursors that contribute to O3 formation, e.g. volatile organic compounds (VOCs) and methane (CH4), are less well regulated . Observed summertime surface maximum 8 h average (MDA8) O3 concentrations in China showed a consistent annual increase of 1.9 ppb between 2013 and 2019 , and this increase is greater in high-emission regions, reaching 3.3 ppbyr-1 on the North China Plain. Given that O3 production in these regions tends to be VOC-limited , reducing emissions of NOx and VOCs simultaneously has become crucial. There are also significant natural sources of O3 precursors from vegetation and soils that may increase due to a warmer climate . Since surface O3 is detrimental for human health, plant growth, and crop yields , robust and effective emission controls on O3 precursors are needed.

The Intergovernmental Panel on Climate Change (IPCC) calls for cutting global greenhouse emissions to close to zero to reduce the risks of climate change . Many countries have recently adopted such net-zero policies to reduce net greenhouse gas emissions to zero by 2050, and China has also implemented emission policies that aim to achieve a carbon peak before 2030 and carbon neutrality by 2060 . These low-carbon policies alongside reductions in anthropogenic air pollutant emissions will have co-benefits for both global climate and air quality . However, since surface O3 changes are not directly proportional to emission changes, it is challenging to quantify the benefits for O3 accurately. Future O3 is also influenced by climate change through changes in atmospheric stagnation, natural emission sources, chemical reaction rates, and deposition rates . Regional surface O3 changes also depend on emission pathways in other parts of the world, which influence the long-range transport of O3 and its precursors across continents . The combination of these factors shapes the changes in the future O3 but imposes large uncertainties in O3 projections , which poses a challenge to assess the underlying impacts of net-zero policies on future air quality.

While the general relationships between O3, its precursor emissions, and climate change are known well , the relative importance of these drivers remains very uncertain. Challenges remain in the capability of chemistry–climate models to simulate O3 changes accurately because processes occurring at small scales cannot be resolved adequately. show that there are systematic biases in the simulation of present-day O3 concentrations in current chemistry–climate models, and this raises questions regarding their skill in representing long-term O3 changes . Averaging output from a number of different models is a common way to obtain more robust results but does not eliminate the O3 biases that are shown to be systematic . In addition, the models tend to use different parameterisations to represent different processes and may misrepresent the importance of local emission controls or the risks caused by climate change. It is hence valuable to correct model simulations to produce more robust O3 projections.

A practical way to address this is to apply deep learning models. Deep learning approaches have developed quickly in the last decade due to advances in computational speed that allow intensive training, and they have been applied widely in scientific fields . Deep learning models have been shown to be a universal approximator and can thus be applied to compensate for discrepancies between physical model simulations and observations. We have demonstrated a successful application of deep learning to correct the biases in surface O3 simulations from a global chemistry–climate model and found that changes in surface O3 in high-emission regions across the world may be overestimated with the process-based model. This bias correction approach allows us to provide a more robust and reliable assessment of future surface O3 projections under the effect of different emission policies and facilitates an examination of their effectiveness.

The aim of this study is to produce reliable estimates of future O3 changes associated with changing emissions and climate under a net-zero pathway in China and to determine how well strict emission controls can tackle the increasing frequency of high-O3 episodes. We introduce the chemistry–climate model used in Sect. 2 along with different emission and climate scenarios, and we describe the deep learning model that we have implemented to correct surface O3 biases. In order to highlight the value of bias correction, we show the results of version 1 of the United Kingdom Earth System Model (UKESM1) before showing the corrected results. We first investigate surface O3 changes in China from the present day to the future under a net-zero emission pathway simulated with UKESM1 in Sect. 3. The influences of emission changes outside China, changes in global CH4 concentrations, and climate change are examined in Sect. 4. We demonstrate the capability of the deep learning model in simulating the biases in surface O3 and apply this bias correction technique to estimate future O3 changes and high-O3 episodes in Sect. 5. Conclusions are presented in Sect. 6.

Seasonal mean surface O3 mixing ratios in China simulated with UKESM1 are shown in Fig. 1 for the present day and the net-zero pathway, without bias correction. There is a clear seasonal variation in surface O3, with high summertime O3 and low wintertime O3 (Fig. 1a and d). However, this variation is reduced under net zero (Fig. 1b and e) due to O3 decreases in summer (Fig. 1c) and O3 increases in parts of eastern China in winter (Fig. 1f) in the future. Surface O3 mixing ratios decrease by about 16 ppb in summer, demonstrating the great benefits of emission controls in mitigating summertime O3 pollution. Other studies show similar results, with 18 ppb decreases in MDA8 O3 mixing ratios achieved from net-zero policies . However, smaller changes are seen in the most polluted industrial areas of China, namely the North China Plain, the Yangtze River Delta, and the Pearl River Delta, even though reductions in anthropogenic emissions in these areas are substantially larger than other regions (Fig. S1a and b in the Supplement). This is principally due to VOC-limited O3 formation regimes in which decreased NOx emissions increase O3 mixing ratios . Much greater reductions in NOx emissions or further reductions in VOC emissions are needed to reduce surface O3 mixing ratios in these high-emission regions. In contrast, higher emissions following SSP3-7.0 will greatly increase summertime O3 (Fig. S2a–c), and the transport sector is shown to have the largest impact with 10 ppb O3 increases.

In wintertime, surface O3 mixing ratios generally decrease by 1 ppb in mainland China but increase in eastern China by up to 20 ppb in heavily populated industrial regions. This results in a reduced latitudinal gradient of O3 mixing ratios in China in wintertime under the net-zero scenario. These contrasting responses further demonstrate regional differences in the chemical environment for O3 production. Polluted urban environments are dominated by VOC-limited O3 formation, particularly in winter when weak boundary layer mixing permits greater NOx accumulation at the surface and rapid local O3 destruction. Therefore, increased NOx emissions from the main emission sectors such as power plants, industry, and transport under SSP3-7.0 cause notable decreases in O3 mixing ratios of up to 3 ppb in winter (Fig. S2e–g), although the effect of the residential sector is relatively small (Fig. S2h) as small changes in NOx emissions are accompanied by substantial changes in VOC emissions .

While local emission changes directly influence surface O3 changes in the future, there are a number of other important drivers that govern surface O3. We investigate four independent drivers: changes in emissions inside China (local emissions) and outside China (external emissions), changes in atmospheric CH4 concentrations (high CH4), and a warmer climate (warmer climate) relative to the net-zero pathway; see Fig. 2. Local anthropogenic emission changes in China are shown to have the largest impact in both seasons (Fig. 2a and e), but other drivers also contribute to surface O3 changes and show substantial regional and seasonal differences.

The effect of changes in emissions outside China reflects the importance of transport of O3 from other countries and higher background O3 concentrations. If the rest of the world did not follow a net-zero emission pathway, surface O3 mixing ratios would be more than 10 ppb higher in summer (Fig. 2b). The contribution to O3 in winter is generally smaller, estimated here as 4 ppb (Fig. 2f). The contribution of external emissions is much larger near the country's borders than in central China. Changes in atmospheric CH4 abundance have a relatively uniform influence on surface O3 in eastern China, with slightly greater effects in western China where altitudes are higher. A 4 ppb O3 increase due to higher CH4 is seen for both seasons (Fig. 2c and g). The O3 changes due to CH4 are comparable to those across central China due to higher emissions outside China. In contrast, a warmer climate under the SSP3-7.0 scenario compared to the net-zero pathway has minor impacts on surface O3 changes (<1 ppb). In general, surface O3 mixing ratios decrease likely due to increased humidity under a warm climate but may increase locally due to higher temperatures, natural emissions, and reduced O3 deposition rates . There are increased natural BVOC emissions in China under both net-zero and SSP3-7.0 scenarios (Fig. S1c and f), particularly in southern China where vegetation is more abundant than in the north. Regional surface O3 responds differently to different future climates (Fig. 2d and h), with O3 increases in the south and O3 decreases in the north under a warmer climate. The regional differences are consistent with those found under the effects of changing BVOC emissions in the future . These O3 increases occur in both seasons, but although they are more pronounced in summer, they remain much smaller than the changes due to anthropogenic emissions. The relative impacts of climate change on O3 may become larger in the future as anthropogenic emissions reduce towards net-zero targets. Overall, we show that while local emissions are critical to O3 pollution, emissions outside China and global CH4 concentrations are also important drivers of concern.

The seasonality of surface O3 changes in China and globally is shown in Fig. 3. In summer, local emissions dominate surface O3 increases, while in winter and spring, O3 transport from other countries and O3 increases due to elevated CH4 concentrations are more important. Strong NO titration of O3 leads to substantial O3 decreases in winter, but its effects are suppressed by more efficient O3 production over summer (Fig. 3a). Emissions outside China increase O3 mixing ratios throughout the year, with the greatest impact in late spring and early summer when intercontinental transport is strongest. The seasonal variation in the influence of local and external emissions is relatively small on a global scale, reflecting a limited sensitivity of global O3 changes to emissions (Fig. 3b). The uniform influence of changes in CH4 concentration is comparable both in China and globally. The warmer climate under SSP3-7.0 leads to slightly larger O3 decreases on a global scale relative to the net-zero scenario. We emphasise that seasonal O3 responses to emission changes are more pronounced at a regional scale and become weaker in winter and that O3 continental transport and background O3 concentrations may still contribute to O3 pollution.

To examine how the occurrence of high-O3 episodes may change in the future, we show the frequency distributions of daily mean surface O3 mixing ratios for all grid cells over China under different scenarios in Fig. 4. We find that surface O3 mixing ratios under the net-zero pathway follow an approximate normal distribution, with a mean O3 of about 20 ppb (Fig. 4a). The frequency of high O3 greater than 40 ppb can be greatly reduced under net zero. This is substantially different from the present-day and SSP3-7.0 scenarios. SSP3-7.0 assumes that there are no emission controls in China, leading to a higher frequency of high O3 mixing ratios (>50 ppb). However, the faster NO titration on O3 with higher NOx emissions also increases the frequency of low O3 mixing ratios (<10 ppb). In Fig. 4b, we show that the O3 distribution shifts to higher values of O3 under the high-internal-emission scenario and is substantially different from the other scenarios shown here, indicating that there is a large change in local O3 production due to local emission changes. The frequency of O3 mixing ratios between 30 and 50 ppb is highest in the scenarios of high external emissions and high CH4 concentrations, demonstrating that these factors can lead to an overall increase in daily mean O3. In addition, we do not find significant changes in O3 mixing ratios due to a warmer climate under SSP3-7.0.

Since there are systematic biases in surface O3 simulations with UKESM1 (Fig. S3a and b), the reliability of future O3 projections remains uncertain. We estimate the biases in surface O3 through the deep learning model and apply this to generate a more robust assessment of O3 changes under the net-zero pathway. A fully independent evaluation for the deep learning model is conducted using a testing dataset; see Fig. 5. We show that the magnitudes and distributions of biases in the UKESM1 simulations are reproduced well by the deep learning model, with a correlation coefficient of 0.96, a mean bias error of 0.1 ppb, and a root mean square error (RMSE) of 4.0 ppb, which demonstrates the robustness of this approach. We also subtract the biases from UKESM1 and examine the spatial and temporal distribution of O3 mixing ratios in China in Fig. 6. Spatial distributions of surface O3 in China over 2013–2017 can also be captured well (Fig. 6a, b, d, and e), with the highest summertime O3 and the lowest wintertime O3 in the North China Plain. The magnitudes of surface O3 mixing ratios with bias correction are in close agreement with the observations. The time series of daily mean O3 can also be simulated well in Beijing and Guangzhou (Fig. 6c and f), which represent two different locations in northern and southern China with rather different chemical and meteorological environments. The evaluation demonstrates the capability of the deep learning model in correcting the seasonal and daily UKESM1 simulation of surface O3. This approach shows great promise in reducing current model errors and hence has the potential to improve simulations of surface O3 under future scenarios.

Spatial distributions of future bias-corrected surface O3 under the net-zero pathways are shown in Fig. 7 to compare and contrast with UKESM1 outputs (Fig. 1) and to assess the effectiveness of emission controls. With bias correction, summertime O3 mixing ratios generally decrease under net zero (Fig. 7a and b), which is consistent with UKESM1 results (Fig. 1c). We find that there are larger O3 decreases in summer in the North China Plain and the Yangtze River Delta (Fig. 7c) than in other less-polluted regions. However, the overall magnitudes of surface O3 decreases are not as large as simulated with UKESM1. There are noticeable differences in the latitudinal mean surface O3 decreases, with the maximum changes estimated as 10 ppb in the bias-corrected simulation, smaller than 20 ppb predicted with UKESM1 (Fig. 7d). This indicates that the underlying impacts of emission controls on O3 may not be as large as the model suggests and that the O3 responses to changing emissions may be overestimated. This is also reflected in the overestimation of O3 changes in southern China in the SSP3-7.0 scenario (Fig. S4a–c).

In wintertime, while surface O3 mixing ratios increase in high-emission regions under net zero, as seen in both UKESM1 and the bias-corrected results, areas with O3 increases are smaller than those predicted by UKESM1 (Fig. 7). This again suggests that the magnitude and spatial extent of O3 titration by NO may be overestimated in UKESM1. The same effect is seen in the bias-corrected wintertime O3 under SSP3-7.0 (Fig. S4). In general, biases in O3 simulations from UKESM1 are smaller in the net-zero scenario but still remain large in the SSP3-7.0 scenario (Fig. S3b–d). These two scenarios correspond to low- and high-emission pathways, respectively, which indicates that the accuracy of O3 simulations in UKESM1 may decrease when emission changes become larger. The bias-corrected results show that only industrial regions with high NOx emissions in China show substantial O3 increases under net zero, while surface O3 mixing ratios decrease in less-polluted regions in winter. This leads to a general decrease in latitudinal surface O3 mixing ratios in wintertime (Fig. 7h).

With bias correction, the average surface O3 mixing ratios are estimated to decrease in both seasons in the eastern part of China in the future under the net-zero pathway. O3 decreases of 5 ppb are predicted to occur in summer, which are slightly larger than the 4 ppb decreases predicted in winter. This demonstrates the overall advantages of net-zero policies in achieving a surface ozone air quality co-benefit. Furthermore, in high-emission regions, the directions of surface O3 changes are different in summer and winter, as shown in both UKESM1 and the corrected UKESM1, indicating that VOC-limited O3 formation still dominates there in winter.

We also calculate the annual average number of days with daily mean O3 over 50 ppb as a measure to quantify high-O3 pollution episodes; see Fig. 8. The number of days per year with high-O3 episodes under present-day conditions can be reproduced well following bias correction (Fig. 8a and b, Table 3), with intensive areas of high O3 pollution in the North China Plain (60 d), particularly in summertime, and relatively low occurrence in the Pearl River Delta (31 d). There is an average of 33 dyr-1 with high O3 pollution over China. We find that the net-zero policies will succeed in reducing the number of high O3 pollution days markedly by 65 % in the future. In contrast, following higher emission control policies will increase high-O3 episodes by almost a factor of 4 (Table 3).

Following net-zero emission controls, the Yangtze River Delta and the Pearl River Delta only have high-O3 episodes for a few days each year. However, high-O3 episodes still occur for almost 1 month (30 d) on the North China Plain and parts of central China in the future, demonstrating that O3 pollution cannot be fully eliminated in this region. The Sichuan Basin is also a region where high O3 pollution cannot be fully addressed, likely due to the favourable meteorological conditions leading to O3 formation associated with the complex topography. Nevertheless, net-zero policies are expected to deliver great benefits in mitigating O3 pollution in China. Indeed, O3 pollution is likely to become much worse if emissions continue to rise (Fig. 8d; Table 3). Even stricter controls on anthropogenic emissions than proposed to meet net zero may be required to avoid high O3 pollution in the North China Plain.

Net-zero emission polices are important for reducing regional surface O3 pollution as well as for mitigating climate change. We use a chemistry–climate model to quantify the O3 changes in China under a net-zero pathway and investigate the relative importance of different drivers of these changes. We also place our results in context by comparing to a scenario, SSP3-7.0, in which weak climate mitigation leads to continued increases in precursor emissions. Surface O3 responses to net-zero emission control policies in China are distinctly different in different seasons, with substantial O3 decreases in summer and O3 increases in winter in high-emission regions due to decreased O3 titration by NO. This demonstrates the large benefits of emission controls in reducing summertime average O3 pollution in China by as much as 16 ppb.

Local emission changes are shown to be the most important driver influencing regional O3 changes, which generally outweighs other drivers such as transport of O3 from other countries, increased background O3 formation through rising CH4 abundance, and a warmer climate. We do not find substantial changes in surface O3 in China between net-zero and SSP3-7.0 scenarios due to a warmer climate, but there are surface O3 increases in southern China. Impacts of future local and external emissions on surface O3 show strong seasonal variation, while increasing future CH4 concentrations have a relatively uniform effect on O3 throughout the year. In winter and spring, future external emissions outside China and higher CH4 concentrations are more important than local emissions at a regional average scale.

We further demonstrate the capability of deep learning approaches to correct the biases in simulated daily mean O3. UKESM1 shows systematic biases in simulated O3 like many other chemistry–climate models; these are expected to influence their projections of future O3. Deep learning can provide improved assessment of the impacts of net-zero policies on surface O3. We find that surface O3 changes are overestimated by UKESM1 in summertime, and therefore the benefits of emission controls may be overestimated by chemistry–climate models. UKESM1 estimates that the mean latitudinal surface O3 decreases due to emission controls could be up to 20 ppb in summer, but bias correction shows that these may only be up to 10 ppb.

We acknowledge that there are uncertainties associated with the choice of the deep learning model used and with the variables and parameters it is trained on, but the biases are sufficiently well predicted here that we are confident in the robustness of our results. The prediction might be further improved by employing more advanced deep learning architectures and considering a wider range of variables. The prediction of future surface O3 biases may be slightly different under these conditions, but we believe that our principal results are likely to remain robust. The driving variables under the net-zero scenario typically lie in the ranges associated with the present-day conditions that were used to train the model, suggesting that the relationships between inputs and outputs derived from the deep learning model are suitable for predicting future situations.

However, net-zero emission policies succeed in reducing the number of days of high O3 pollution by 65 % in China per year, with the number dropping from 33 d under present-day conditions to 11 d each year under net zero. The North China Plain will still be affected by high O3 pollution in the future, meaning that stricter emission policies are needed in this region. In the Yangtze River Delta and the Pearl River Delta, O3 pollution is likely to be less of a concern in the future as there are only a few days with high O3 pollution under net zero. It is also clear that if emissions continue to rise, air quality in China will be substantially worse than at present, and therefore emission controls are essential. However, it is clear from these studies that emission controls can be very effective in reducing surface O3 pollution and that net-zero emission policies can substantially mitigate O3 pollution in China.