Predictability and Risk of Extreme Winter PM2.5 Concentration in Beijing

Jingpeng LIU, Adam A.SCAIFE, Nick DUNSTONE, Hong-Li REN, Doug SMITH,Steven C.HARDIMAN, and Bo WU

1 China Meteorological Administration Key Laboratory for Climate Prediction Studies, National Climate Centre, China Meteorological Administration, Beijing 100081, China

2 Met Office Hadley Centre, Exeter EX1 3PB, United Kingdom

3 College of Engineering, Mathematics, and Physical Sciences, University of Exeter, Exeter EX4 4QG, United Kingdom

4 State Key Laboratory of Severe Weather and Institute of Tibetan Plateau Meteorology, Chinese Academy of Meteorological Sciences, China Meteorological Administration, Beijing 100081, China

5 State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG), Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing 100029, China

ABSTRACT Air pollution remains a serious environmental and social problem in many big cities in the world.How to predict and estimate the risk of extreme air pollution is unsettled yet.This study tries to provide a solution to this challenge by examining the winter PM2.5 concentration in Beijing based on the UNprecedented Simulation of Extremes with ENsembles (UNSEEN) method.The PM2.5 concentration observations in Beijing, Japanese 55-yr reanalysis data, and the Met Office near term climate prediction system (DePreSys3a) large ensemble simulations are used, and 10,000 proxy series are generated with the model fidelity test.It is found that in Beijing, the main meteorological driver of PM2.5 concentration is monthly 850-hPa meridional wind (V850).Although the skill in prediction of V850 is low on seasonal and longer timescales, based on the UNSEEN, we use large ensemble of initialized climate simulations of V850 to estimate the current chance and risk of unprecedented PM2.5 concentration in Beijing.We unravel that there is a 3% (2.1%-3.9%) chance of unprecedented low monthly V850 corresponding to high PM2.5 in each winter, within the 95% range, calculated by bootstrap resampling of the data.Moreover, we use the relationship between air quality and winds to remove the meridional wind influence from the observed record, and find that anthropogenic intervention appears to have reduced the risk of extreme PM2.5 in Beijing in recent years.

Key words: UNprecedented Simulation of Extremes with ENsembles (UNSEEN), climate risk, PM2.5, Beijing

1.Introduction

Severe air pollution is a considerable threat to human health and it affects regional and global climates (Wang et al., 2014; Liu et al., 2017).Related to rapid economic development, severe air pollution has been a notable environmental and social problem in China (Ding and Liu,2014; Huang et al., 2014).Record air pollution occurred in January 2013 (Zhang et al., 2014), February 2014 (Yin et al., 2017), and December 2015 (Zhang et al., 2019)over the North China Plain.Particulate matter 2.5 (PM2.5)exposure in 2010 was estimated to result in about 1 million deaths in China, and PM2.5accounted for approximately 73 deaths per 100,000 population (Wang et al.,2019).

Air quality is strongly dependent on meteorological conditions, climate variabilities, and anthropogenic emissions.Meteorological conditions control the transport and dispersion of air pollutants within the lower atmosphere (Huang et al., 2017; Pei et al., 2020).The meteorological factors related to PM2.5concentration during winter are mainly relative humidity (Ding and Liu, 2014;Zhai et al., 2019), wind speed (He et al., 2001; Zhang et al., 2013; Xiao et al., 2020), temperature (Cai et al.,2017; Jin et al., 2022), and rainfall (Zheng et al., 2014;Zhao et al., 2020).Surface wind speed and surface air temperature have a great impact on PM2.5changes in northern China, while surface air temperature and vegetation dominate the PM2.5in southern China (Jin et al.,2022).Climate variabilities such as El Ni?o-Southern Oscillation (ENSO), East Asian winter monsoon(EAWM), and Pacific Decadal Oscillation (PDO) also have significant effects on PM2.5concentration over China from interannual to decadal timescales (Zhao et al., 2016; Yu et al., 2019; Liu et al., 2022; Wang et al.,2023).On the interannual timescale, the EAWM can cause a seesaw pattern of PM2.5concentration anomalies between Beijing-Tianjin-Hebei (BTH) and the Yangtze River Delta across eastern China in winter (Liu et al.,2022).Over northern China, all kinds of El Ni?o events lead to an increase in severe haze days in winter except for moderate central-Pacific El Ni?o (CP) events (Yu et al., 2019; Wang et al., 2023).On the decadal timescale,decadal variability in the occurrence of wintertime haze in central eastern China is tied to the PDO (Zhao et al.,2016).The anthropogenic emissions mainly influence the long-term trends of PM2.5concentration (Pei et al., 2020;Jeong et al., 2021; Zhai et al., 2023).Recent efforts on emission control account for 34% to 49% of decreased PM2.5in China from 2013 to 2017 (Zhai et al., 2019),and about half of the total variances of the haze intensity can be explained by the changing emissions in the BTH region for the period 1980-2017 (Pei et al., 2020).

Improving the predictability of PM2.5remains a great challenge.Statistical methods such as the simple/multiple linear regression are effective when climate variabilities and meteorological conditions dominate the PM2.5variability (Pei et al., 2020; Jeong et al., 2021).A simple linear regression model shows good performance withr> 0.75 in reproducing PM2.5concentration in Northeast China using the Siberian high indices (Jeong et al.,2021).Utilizing the chemical transport models is another method to provide more detailed spatiotemporal distributions of PM2.5concentration (Mathur et al., 2008; Jeong et al., 2021; Dash et al., 2023).Dash et al.(2023) show that a developed data assimilation system in the Community Multiscale Air Quality Model (CMAQ) helps to improve the predictability of PM2.5in the BTH region.Recently, deep learning and machine learning techniques are widely used to simulate the spatial distribution of PM2.5based on selected variables (Zhao and Song, 2017; Xiao et al., 2020; Chuluunsaikhan et al.,2021; Jin et al., 2022; Kim et al., 2023).Such techniques can mine large quantities of environmental data and show high performance in short-term prediction within 24 h (Lee et al., 2020; Kim et al., 2023).

Here, we use observed daily PM2.5over 2009-2022 to verify different meteorological conditions and climate variabilities, and reveal that the 850-hPa meridional wind is one of the main meteorological factors controlling the winter PM2.5concentration in Beijing.Such meteorological factors can be treated as a proxy measure of PM2.5concentration so that we can use them to make predictions of air quality, and even anticipate the risk of extreme winter haze events.In addition, we use a large ensemble of initialized climate simulations provided by the Met Office third decadal prediction system (DePreSys3)(Dunstone et al., 2016) to assess the predictability of the 850-hPa meridional wind over Beijing and hence PM2.5.

We also assess the current chance of unprecedented PM2.5concentration.It is difficult to assess the current chances of unprecedented climate extremes as, by definition, there are no such events in observational records and the climate is changing fast, making older historical observations unrepresentative of the current climate(Thompson et al., 2019).However, the UNSEEN (UNprecedented Simulation of Extremes with ENsembles)method allows the chance of unprecedented extremes to be estimated from large ensembles of initialized coupled ocean-atmosphere climate simulations (Thompson et al.,2017).The UNSEEN method has also been used to assess the risk of unprecedented hot summer (Thompson et al., 2019) and droughts (Kent et al., 2019) in China.Here, the risk of extreme winter PM2.5concentration in Beijing is estimated by using the UNSEEN method,which can form the basis of new climate services to inform policy planning and decision-making through an improved understanding of the dynamics and chances of extreme winter air quality events.

2.Data and method

2.1 Observations

We use observed time series of daily PM2.5in Beijing(40°N, 117°E) from 2009 to 2017, available from the US Embassy (USE) at the following website (http://www.state air.net/web/historical/1/1.html).This time series has been used in some previous studies (Martini et al., 2015;Zheng et al., 2015; Cai et al., 2017).We also use observed daily PM2.5in Beijing from 2014 to 2022, available from the Ministry of Ecology and Environment(MEE) of the People’s Republic of China at the following website (http://datacenter.mee.gov.cn/websjzx/query Index.vm).We construct the winter daily PM2.5from 2009/2010 winter to 2021/2022 winter by connecting the observations from 2017/2018 winter to 2021/2022 winter provided by the MEE to the observations provided by the USE.These two series show high consistency during the overlapping period from 2015 to 2017.The monthly mean time series of PM2.5from 2010 to 2022 is the average of daily observations in each month.The time series of extreme haze days is the number of days of PM2.5>150 μg m-3in each month.We chose the threshold of 150 μg m-3as the Beijing municipal government issues a“red alert” when the PM2.5concentration is forecast to exceed 150 μg m-3for 72 consecutive hours.

We employ V850 and sea level pressure (SLP) from the Japanese 55-yr reanalysis (JRA-55; Kobayashi et al.,2015) from 1960 to 2022, at a resolution of 1.25° latitude by 1.25° longitude.JRA-55 is designed for high instantaneous accuracy based on the use of as many observations as possible while allowing for the impacts of changes in observing systems.JRA-55 shows good performance in the low-level meridional wind in East Asia(Chen et al., 2014).Satellite data was assimilated in JRA-55 after 1979 (Kobayashi et al., 2015), therefore the monthly data from December to February for 1980-2022 are used to assess the risk of extreme winter haze events in this study.Monthly V850, SLP, air temperature, and 500-hPa zonal wind from the NCEP-NCAR reanalysis 1(NCEP-R1) (Kalnay et al., 1996) from 1960 to 2022, at a resolution of 2.5° latitude by 2.5° longitude are also utilized to verify the relevant analysis.The ENSO index is provided by NOAA/NCEP Climate Prediction Center(CPC) from website: https://psl.noaa.gov/data/correlation/nina34.data.

2.2 Model simulations

Simulations of the current climate are provided by the Met Office near term climate prediction system (De-PreSys3; Dunstone et al., 2016).This system uses the Hadley Centre global climate model, HadGEM3-GC2(Williams et al., 2015), at high resolution compared to most current climate prediction models: 60-km atmosphere and 0.25° ocean.The model is initialized with atmospheric, oceanic, and sea-ice observational data and current anthropogenic and natural forcings, so that the simulations are representative of current climate.Using the large ensemble of simulations from these hindcasts provides many more realizations of the global climate than those from the recent observational period.Monthly model data are taken from 16 month long retrospective forecasts starting every November over 1960-2022.Each hindcast month has 40 ensemble members.We use the full hindcast period to assess the predictability of V850 and the pooled simulations from 1980 to 2022 to estimate the current chance of unprecedented PM2.5concentration.The model output is regridded onto a 1.25°×1.25° grid for this study.

2.3 Model fidelity and the UNSEEN method

The UNprecedented Simulation of Extremes with ENsembles (UNSEEN) method assumes that observations provide only one realization of the plausible climate state and uses initialized climate model ensembles to explore a wider range of possible states.In this study, model fidelity testing follows Thompson et al.(2019) and is assessed by determining if the observations and model data are drawn from the same underlying distribution.10,000 proxy time series with the same length as the observed record are generated from the model data by randomly selecting one ensemble member from each winter month.The mean, standard deviation, skewness, and kurtosis of each of these 10,000 proxy series are calculated, creating distributions of 10,000 values for comparison with the observations.The model is deemed to be indistinguishable from the observations if the statistical characteristics of the observations fall within the 5%-95% range of the model subsamples.In this case we use the UNSEEN method to calculate the chance of extreme and unprecedented events.

3.Results

3.1 Winter PM2.5 in Beijing from 2010 to 2022

Figure 1 shows the characteristics of winter PM2.5in Beijing.The correlation coefficient of daily PM2.5from USE and MEE during the overlapping time period of 2015-2017, is 0.98.Considering the high consistency between these two data sources, we connect the USE data of 2010-2017 with the MEE data of 2018-2022 as a single PM2.5time series (Fig.1a).The correlation between this monthly PM2.5concentration and the frequency of extreme haze days is 0.97, significant at the 99.9% confidence level, suggesting that the monthly PM2.5is an excellent proxy for the number of days with extreme PM2.5(Fig.1b).We also note the lower values after 2016 and shall discuss these later.

Fig.1.(a) The raw data of daily winter PM2.5 observations in Beijing (the black line is the data from 2010 to 2022 provided by the United States Embassy, the red line is the data from 2015 to 2022 provided by the Ministry of Ecological Environment of the People’s Republic of China).(b)the monthly mean PM2.5 time series (black line) and the number of extreme haze days in each month (red line, with the threshold of 150 μg m-3).

3.2 Predictability of the circulation and winter haze events in Beijing

Considering the seriously negative impact of winter haze, it is imperative to predict short-term PM2.5concentration; but given the strong relationship between PM2.5and meteorological conditions, it is first necessary to identify the meteorological factors affecting PM2.5concentration.We use observed monthly PM2.5over 2009-2022 to verify different meteorological conditions and climate variabilities.V850, vertical temperature profile(ΔT), and 500-hPa zonal wind (U500) are chosen meteorological conditions (Cai et al., 2017), and EAWM (Gong et al., 2001) and ENSO (Wang et al., 2023) are key modes of climate variability.The definition of each factor is listed in Table 1.V850 represents near-surface southerly anomalies over 35°-55°N, 105°-125°E that counteract the climatological northerly flow; ΔTshows the impacts of vertical thermal structure between the near-surface (850 hPa) and upper troposphere (250 hPa)over 32.5°-45°N, 112.5°-132.5°E; U500 reflects the impacts of the East Asia trough over 42.5°-52.5°N,110°-137.5°E and 27.5°-37.5°N, 110°-137.5°E; and EAWM over 40°-60°N, 70°-120°E and ENSO over 5°N-5°S, 170°-120°W are two of the most important climate modes affecting winter PM2.5concentration in East Asia (Jeong et al., 2021).

Table 1 shows that except for ENSO, all the other four factors have significant impacts on PM2.5concentration in Beijing.Interestingly, the correlation coefficients have decreased slightly when the 2020-2022 period is included (Table 1).Since 2020, PM2.5concentration dropped abruptly (Fig.1) due to Covid-19 reductions in construction and vehicle traffic (Zhai et al., 2023), which outweigh the impacts of meteorological factors.Table 1 reveals that V850 outperforms these five impact factors,consistent with previous studies (Cai et al., 2017; Pei et al., 2020).Because V850, ΔT, U500, and EAWM are not independent, and meaningful predictions of winter PM2.5concentrations in East Asia can be made using simple linear regression models (Jeong et al., 2021), V850 is an appropriate choice among a suite of relevant meteorological variables.Table 1 shows that V850 is a major factor representing the PM2.5concentration in Beijing because southerly anomalies counteract the climatological north-erly flow, leading to stagnation and build-up of pollution.

Table 1.The correlation coefficient between PM2.5 and possible impact factors

We define a mean wind index (WI) as the mean V850 wind in the black rectangle in Fig.2a and plotted as the red line in Fig.2b.The correlation coefficient between the WI and the PM2.5time series is 0.75 (p= 1.6 × 10-6)for 2010-2019 and 0.56 (p= 2.1 × 10-4) for 2010-2022,indicating that the wind index is a good proxy measure of the winter PM2.5concentration in Beijing.The correlation coefficient was recalculated by utilizing the V850 from the NCEP-R1, which showed high consistency with the JRA-55 dataset (figure not shown).

To assess the PM2.5concentration without the impacts of V850, the residual PM2.5concentration is evaluated by subtracting the product of the wind index and the regression coefficient (RC) of the PM2.5on the WI (Residual =PM2.5- WI × RCPM2.5,WI).This residual is shown in Fig.2c and represents the PM2.5concentration without the impact of V850.The prevailing winter meridional wind in the key region is a northerly wind which dissipates the air pollutants in Beijing and reduces the PM2.5concentration.Figure 2c shows that V850 had stronger impact on the dispersion of PM2.5from 2010 to 2017 than from 2018 to 2022.

Fig.2.(a) The correlation coefficient between near-surface southerlies (V850) from JRA-55 and time series of PM2.5 concentration from 2010 to 2022 (black dots denote areas significant at the 90% confidence level); the black star is the location of Beijing, the black rectangle box is the chosen key region of V850 (35°-55°N, 105°-125°E); (b) the standardized time series of PM2.5 (black line) and mean V850 in the key region (red line) from 2010 to 2022; (c) the PM2.5 concentration without the impact of V850 in the key region.The blue line represents the period during which the V850 plays a weaker role than expected (the start of the blue line is defined as the month when the anomaly of the residual exceeds one standard deviation).

Having identified the driving meteorological pattern for poor air quality in China, we note the recent strong decline in PM2.5in Fig.1.A general decrease of 33% in annual mean PM2.5across China over the 2013-2017 period has been reported (Zhang et al., 2019).There are two possible reasons for the decreasing trend of PM2.5in Beijing: changing meteorological conditions and anthropogenic impacts.Between 2013 and 2017, the anthropogenic impacts are thought to explain the main part of the decline in Beijing and it is possible that the anthropogenic impact is the main factor in the winter of 2018 and 2019(Zhai et al., 2019; Zhang et al., 2022).The PM2.5concentration was significantly lower than normal, with a 56%reduction in 2018 and a 49% reduction in 2019 compared to the climatic mean.Note that the northerly wind is weaker than normal in these two years (Fig.2b), indicative of a scenario with higher PM2.5concentration.The inconsistency shows that the wind is not the most important factor impacting the PM2.5concentration during 2018 and 2019.Figure 2c shows that the residual PM2.5after removing the V850 influence is significantly lower in the winter of 2018 and 2019, confirming this result.One probable reason is that in June 2018, China promulgated the Three-Year Action Plan for winning the Blue Sky Defense Battle to continue its efforts in battling air pollution (Wen et al., 2021).The anthropogenic impact of this action appears to have dominated the V850 impact on air quality in recent years.Since 2020, PM2.5concentration decreased significantly (Fig.2b) due to Covid-19 reductions in anthropogenic emissions (Zhai et al., 2023).However, Fig.2c shows improved impacts of V850 on PM2.5since 2021 when comparing with the 2018-2019 period, indicating the importance of V850 in dispersion of PM2.5under the low emission conditions.

We now assess the predictability of V850 by using the ensemble simulations from 1960 to 2022.Figure 3a shows the correlation coefficient of V850 predicted by DePreSys3 and JRA observations from 1960 to 2022.The correlation coefficient pattern is similar with Fig.3a when V850 provided by NCEP-R1 is utilized (figure not shown).The prediction skill of DJF V850 is high in southern and coastal areas of China (Lockwood et al.,2019), and other studies have shown associated skillful prediction of winter precipitation in China (Lu et al.,2017).However, the skill of predictions of V850 is low in the key region in this prediction system and the correlation coefficient between the ensemble mean V850 in the model and observational analysis is not significant (Fig.3b).This suggests that inherent predictability of PM2.5over Beijing may also be low but this conclusion depends on having an optimal model and forecast system and further developments may increase the skill.

Though the model cannot skillfully predict interannual variability of V850 in the key region, this could simply be due to lack of predictability in the real system and the model simulations may still be able to provide a realistic representation of the climate variability and hence range of states available to the real world.The model could then directly sample more extreme cases than that in the observational record, allowing the identification of unprecedented PM2.5concentration events to assess their likelihood and meteorological drivers.Here, we calculate the current climatological chance of unprecedented V850 by using the UNSEEN method (Thompson et al.,2017).

3.3 Risk of unprecedented PM2.5 concentration in Beijing

In order to accurately calculate the chance of unprecedented monthly V850 in the key region in Fig.2a, the model must be able to provide a realistic representation of the range of states available to the real world.To assess the model fidelity, we use the JRA-55 observational data covering 1980-2019, the same period as the model simulations.

Figure 4a compares the modeled and observed distribution of the monthly V850 for the winter months,December-February.The dots in Fig.4a show that there is good consistency between the occurrence of extreme PM2.5concentration and mean V850 in the key region,consistent with their correlation of 0.75.We then generate 10,000 model proxy time series of equal length to the observed WI record.Distributions of values for the mean, standard deviation, skewness, and kurtosis are plotted in Fig.4.Figures 4b-e show that the observed estimates are all within the 5%-95% range of the modelderived estimates, indicating that the model is statistically indistinguishable from the observations.Hence, we apply the UNSEEN method with no model adjustments necessary.

Fig.3.(a) The correlation coefficient of V850 provided by the DePreSys3 and JRA observations from 1960 to 2022 (black dots denote areas of significance at the 90% confidence level).(b) The red line represents the ensemble mean WI in the model and the black line in JRA reanalysis.The grey lines represent the WI of the 40 ensemble members.

Fig.4.Comparison of the observed (Obs) and modeled (Mod) mean wind index (WI).(a) The observed (blue line) and modeled (red line) probability density distribution (PDF) of December-February monthly mean V850 (m s-1) in the key region over 1980-2019.The black, blue, and red dots denote the values of V850 in the three months with the top largest observed PM2.5 concentration.(b-e) PDFs of the mean, standard deviation, skewness, and kurtosis of 10,000 proxy model time series, compared to the observed values indicated by the black vertical lines.The grey dashed vertical lines in (b-e) show the 5%-95% locations of the model bootstraps.

Fig.5.The relationship between the chance of an air pollution event and unprecedented V850 for (a) winter (December-February), (b) December, (c) January, and (d) February.The uncertainties indicate the 95% range calculated by bootstrap resampling of the data.The upper x-axis coordinates denote the absolute value of the unprecedented V850.

The chance of unprecedented PM2.5concentration in Beijing is now assessed.Figure 5a shows that the chance of an event exceeding the observed maximum V850 in the key region in winter (December to February) in any given year in the model simulations is 3%, indicating a chance of 3% of each coming winter having an unprecedented monthly V850.The uncertainties in Fig.5a ranging from 2.1% to 3.9% indicate the 95% range calculated by bootstrap resampling of the data.As the chance of extreme events is estimated basing on all the given model years, the chance can be regarded as an indication of climate states in a decadal-length period.The probability of extreme severe PM2.5event in each coming winter should be the same regardless of the interannual variability.If more updated model samples are absorbed in, the chance can be reassessed.Figures 5b-d show the chance of an event exceeding the observed maximum V850 in December, January, and February, respectively.The chance of an unprecedented event is 2.6% in December, 2.5% in January, and 1.9% in February.The chances in December and January are comparable, both are larger than in February, and the risk of unprecedented PM2.5concentration in the whole winter is larger than in any winter month.The link between V850 and PM2.5concentration(Fig.2a) further indicates that the risk of V850 can be connected to the risk of unprecedented PM2.5concentration.Note that this is the chance without the anthropogenic intervention, which appears to have reduced the risk in the most recent years.

We now examine the atmospheric circulation related to extreme PM2.5concentration winter months.The observed sea level pressure anomaly (SLPA) patterns from January 2013 and February 2014, the two months with highest PM2.5concentration in observations, are shown in Figs.6a and 6b.Those two patterns are similar with anomalous low pressure to the west of Beijing and anomalous high pressure to the east of Beijing, favoring anomalous southerlies which act to weaken the climatological northerlies, which is unfavorable for the dispersion of air pollutants in Beijing.

Figure 6c shows the composited SLPA of the top 10 model months with extreme wind index.The fixed atmospheric circulation field distribution is that an anomalous high to the northeast of Beijing, and an anomalous low to the northwest of Beijing.Significant west-east sea level pressure gradient exists across Beijing, which can lead to reduced seasonal prevailing surface cold northerlies.Examples of the top three model months with extreme wind index are shown in Figs.6d-f.All of the patterns show high pressure to the east of Beijing.However, there is wide variation in the SLPA patterns leading to extreme haze events.For example, the circulation pattern in Fig.6f is unlike the observed patterns (Figs.6a, b), but still produces anomalous southwesterlies locally over Beijing which are unfavorable for dispersion of air pollutants.These examples serve to highlight that a variety of largescale atmospheric circulation patterns, some of them perhaps not yet realized, can drive regional extreme PM2.5concentration in Beijing.

4.Conclusions and discussion

Using a large ensemble of initialized climate simulations, reanalysis data, and observational PM2.5concentration, the predictability and risk of winter PM2.5concentration in Beijing are investigated.We find that the local monthly 850-hPa meridional wind is highly correlated with the winter PM2.5concentration in Beijing, and can characterize the interannual variability of the latter.However, the prediction skill of this variable is not significant in the key region on seasonal and longer timescales in the prediction system used here.We instead used the large ensemble of initialized climate simulations to estimate the current chance of unprecedented PM2.5concentration in Beijing.We find a 3% chance of an unprecedented V850 month each winter, indicating that there is a significant chance of an unprecedented PM2.5concentration.The risks in December and January are comparable, both are larger than in February, and the risk of unprecedented PM2.5concentration in the whole winter is larger than in any winter month.Despite low deterministic forecasting skills, large sample ensembles are still good at simulating extreme risks.

However, we note that this probability is based upon the relationship over the 2010-2022 period.If the depressed values of PM2.5seen in winters 2018 and 2019,which occurred inconsistently with the observed meteorology, are indeed representative of a new baseline for anthropogenic emissions, the chance of experiencing an unprecedented PM2.5month will be lower than 3%.

To confirm this, we further explore the PM2.5emission in the winters of 2020, 2021, and 2022.The mean PM2.5concentration is 44.5 μg m-3over 2018-2022, with a 59% decrease compared to the mean PM2.5(109 μg m-3) over 2010-2017, showing a significant shift of the mean PM2.5concentration between these two periods.The northerly wind is weaker than normal in the winter of 2020, indicative of a scenario with higher PM2.5concentration.However, the PM2.5concentration remains at a low level, with an average of 55 μg m-3per month.This is good confirmatory evidence that measures to curb anthropogenic emissions are having a significant impact.In consequence, the chance of experiencing an unprecedented PM2.5month will be lower than 3% under the new baseline for anthropogenic emissions.

The large ensemble of model simulations allows us to better estimate the probability of extreme events.However, using a single model is potentially overconfident because the structural uncertainties cannot be fully sampled (Thompson et al., 2019).Further work is therefore needed to repeat the analysis with simulations from other climate models as it was recently done for the Indian monsoon (Jain et al., 2020).Furthermore, it is worth going further and showing the expected anomaly in PM2.5in the future.We would also need to include the uncertainty in the V850-PM2.5relationship to achieve that.Considering the limited predictability of the V850 in the key region, it would also be useful to find other more predictable proxy indices to represent the PM2.5in Beijing, since extreme winter haze events cause significant impacts on society and forecasts of such events would be very useful.Our results could nevertheless quantify the likelihood and intensity of unprecedented PM2.5concentration in the near future.

Fig.6.The large-scale circulation anomalies associated with the extreme events.(a, b) the SLPA anomaly fields (in hPa, relative to 1981-2010)of the two observed months with highest PM2.5 concentration, (c) the composited SLPA of the top 10 extreme simulated months, and (d-f) the SLPA of top three extreme simulated months one of which presents a potential new record PM2.5 concentration scenario in Beijing.The star in(a-f) is the location of Beijing.

Acknowledgments.We thank the Editor and two anonymous reviewers for their constructive comments and suggestions.This work was done while the lead author was visiting Adam Scaife and Nick Dunstone at the Met Office Hadley Centre.