Influence of boundary layer structure on air quality in Beijing : Long-term 1 analysis based on self-organizing maps 2

16 Self-organizing maps (SOMs; a feather-extracting technique based on an unsupervised machine learning 17 algorithm) are used to classify the atmospheric boundary layer (ABL) types over Beijing by detecting topological 18 relationships among the 4-yr (2013–2016) radiosonde profiles. The resulting ABL types are then examined in 19 relation to air quality, including surface pollutant concentrations and columnar aerosol properties, to understand 20 the regulating effects of different ABL structures on Beijing’s air quality. The SOM provides nine ABL types (i.e., 21 SOM nodes), and each type is characterized by distinct dynamic and thermodynamic conditions. On average, SO2, 22 NO2, CO, PM10 and PM2.5 increase 120–220 % from a near neutral (i.e., node 1) to strong stable condition (i.e., 23 node 9). The ABL controls on diurnal cycles of pollutants are as follows: (1) elevated inversion enhances the 24 afternoon baseline; and (2) surface inversion improves the evening increment. Comparing the CO/SO2 ratios for 25 the different ABL types demonstrates that the local contribution increases with enhanced static stability near the 26 ground, and it is the stable ABL stratification rather than weak surface wind that confines the regional contribution. 27 Due to regional transport, node 3 (dominated by elevated inversion with high relative humidity) corresponds to 28 the most severe columnar aerosol pollution, characterized by the highest optical depth (1.22) and volume 29 concentration (0.30 μm 3 /μm 2 ). The larger aerosol radiative forcing (ARF) within the atmosphere (> 60 W/m 2 ) in 30 nodes 3, 6 and 9 is likely to strengthen the atmospheric stability and thus induce a positive feedback loop for 31 1 Atmos. Chem. Phys. Discuss., https://doi.org/10.5194/acp-2017-1046 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 13 December 2017 c © Author(s) 2017. CC BY 4.0 License.


Introduction
The atmospheric boundary layer (ABL) is the section of atmosphere that responds directly to the flows of mass, energy and momentum from the earth's surface, characteristically at timescales of an hour or less (Stull, 1988).
Most air pollutants are emitted or chemically produced within this layer and its evolution plays an important role in determining the dispersive and chemical properties of pollutants (Chen et al., 2012;Fan et al., 2008;Whiteman et al., 2014;Platis et al., 2016;Wolf et al., 2014;Wu et al., 2013).Therefore, characterizing typical ABL conditions associated with high pollution levels helps to better understand the role of ABL in governing the transport and distribution of pollutants in the atmosphere.
Beijing, the capital of China, is suffering serious air pollution problems.This city is geographically located at the northwestern border of the Great North China Plain and has three directions that are adjacent to mountains.
The closest coast from the city of Beijing is the Bohai Sea, which is 160 km southeast of the city.Terrain-related circulations can therefore be well developed over Beijing and its surroundings under favorable weather conditions, leading to a complex ABL thermodynamic structure, which is thought to substantially affect Beijing's air quality (Hu et al., 2014;Miao et al., 2017;Ye et al., 2016;Gao et al., 2016;Xu et al., 2016).Several studies used tower-based observations to investigate the interactions between boundary layer dynamics and pollution formation (Sun et al., 2013;Sun et al., 2015;Guinot et al., 2006).However, the results are not ideal because they have a low observational height (325 m).Numerous intensive ABL measures were conducted using other approaches, such as mooring boats, airplane, and ground remote sensing (Tang et al., 2015;Zhu et al., 2016;Zhang et al., 2009;Hua et al., 2016).However, since these approaches are complex, expensive and labor intensive, they are often restricted to the duration of specific research campaigns.Overall, the existing knowledge of linkages between ABL structure and air quality in Beijing is drawn largely from either low observational height or short observational duration.
Due to the lack of long-term effective observations, the influence of ABL on Beijing's air quality remains relatively unclear.For example, many case studies (Jia et al., 2008;Zheng et al., 2015;Hua et al., 2016;Li et al., 2016) claimed that rapid growth of PM 2.5 in Beijing is mainly attributable to the regional transport of the polluted air mass.This view is occasionally questionable, as it is known that the polluted episodes tend to occur with a weak surface wind and stable boundary layer stratification, which are unfavorable for transport (Zhu et al., 2016;Tang et al., 2015).Given these uncertainties, there is an urgent need to investigate and determine the common patterns of ABL structure influence on Beijing's air quality.
On the other hand, the long-term radiosondes are not being fully utilized to investigate urban pollution issues.
The advantage of radiosondes over the other approaches seems to be their length, which usually spans several decades.For a long time, it was challenging to reduce the wealth of radiosonde datasets to characterize the ABL structure, therefore, radiosondes remain in very limited use in case studies (Ji et al., 2012;Zhao et al., 2013;Gao et al., 2016).Recently, self-organizing maps (SOMs; a feather-extracting technique based on an unsupervised machine learning algorithm) (Kohonen, 2001) were introduced to investigate the ABL thermodynamic structure, indicating the capabilities of SOMs in feather extraction from a large dataset of the ABL measurements (Katurji et al., 2015).In fact, the SOM has become increasingly popular in atmospheric and environmental sciences during the past several years (Jensen et al., 2012;Jiang et al., 2017;Gibson et al., 2016;Pearce et al., 2014;Stauffer et al., 2016), including a first application of routine radiosondes in South Africa (Dyson, 2015).However, there is thus far no SOM application in pollution-related ABL structure research.It is expected that such a new analytical approach can tap the potential of routine radiosondes to better understand urban air pollution.
In this study, a long-term analysis regarding the influence of ABL structure on Beijing's air quality is performed based on the application of SOMs to 4 years (2013-2016) of radiosonde measurements.The SOM is first used to classify the vertical temperature profiles for identifying predominant ABL types (see section 3.1).A selection of climatological observations is then subdivided according to the SOM-based ABL classification (see section 3.2).
Finally, we provide a visual insight into air quality variations (including surface pollutant concentrations and columnar aerosol properties) under various ABL conditions and discuss the potential physical mechanisms behind their relationships (see section 3.3-3.5).It is expected that such an association between air quality and ABL type could provide local policy makers with useful information for improving the predictions of urban air quality.

Data preparation and preprocessing
Radiosonde data observed at the Beijing Observatory (39.81 °N, 116.48 °E, WMO station number 54511) were collected from the University of Wyoming (http://weather.uwyo.edu/).The data cover the recent 4-year period from 2013 to 2016.The Beijing Observatory launches a routine radiosonde twice a day (08:00 and 20:00 Beijing Time (BJT), corresponding to the morning and evening, respectively) and provides atmospheric sounding data (profiles of temperature, relative humidity, wind speed, etc.) at the mandatory pressure levels (e.g., surface, 1000, 925, 850, 700 hPa) and additional significant levels.In addition, the hourly near-surface meteorological parameters (including temperature, wind speed and relative humidity) are also collected from the Beijing Meteorological Bureau.
We chose the 2000 m above ground level (AGL) as the upper limit of the ABL based on a number studies investigating the ABL height over Beijing or North China (Tang et al., 2016;Guo et al., 2016;Miao et al., 2017).This height exceeds the ABL height in most cases, and therefore, most ABL processes influencing the near-surface air quality are included in the analysis herein.We classify the daily ABL types using the SOM algorithm.To keep a whole night, the daily vertical profiles are composited from the radiosonde measurements at 20:00 and 08:00 of the next day.
The mass concentrations of atmospheric pollutants (including SO 2 , NO 2 , CO, O 3 , PM 10 and PM 2.5 ) over Beijing during the period from 2013 to 2016 are obtained from the Ministry of Environmental Protection of the People's Republic of China (http://datacenter.mep.gov.cn/).In addition, hourly PM 2.5 measured at the Beijing US Embassy (http://www.stateair.net/)are also used in this study.Hourly concentrations are calculated for the Beijing urban area by averaging concentrations from nine urban sites (including Dongsi, Guanyuan, Tiantan, Wanshouxigong, Aotizhongxin, Nongzhanguan, Gucheng, Haidianwanliu and US Embassy).To maintain consistency with ABL classification, the daily pollutant concentration is performed from noon-to-noon (12:00 h-12:00 h).
In addition to near-surface observations, columnar aerosol parameters (including aerosol optical depth (AOD), Ångström exponent (AE), single scattering albedo (SSA), volume particle size distribution (dV/dlnR), aerosol radiative forcing (ARF) and so on) are also collected from the AERONET Beijing (39.98 °N,116.38 °E) and 116.32 °E) sites.The level-2.0 quality-assured columnar aerosol data from 2013 to 2016 are downloaded from the AERONET data archive (http://aeronet.gsfc.nana.gov).The size distribution is retrieved in 22 logarithmically equidistant bins in a range of sizes from 0.05 to 15 μm through a combined spherical and spheroid particle model (Dubovik and King, 2000;Dubovik et al., 2006).
2.2 Self-organizing maps technique profiles corresponding to the SOM node are plotted in blue.For comparison, the mean and 25th and 75th percentiles for the entire period are plotted in cyan.On the SOM plane, the most notable feather is adjacency of like types (e.g., nodes 1 and 2) and the separation of contrasting types (e.g., nodes 1 and 9).Although the SOM nodes appear to be sorted in a certain order, there is no physical significance associated with this order.Such ordering is a feather of the SOM algorithm (i.e., 'self-organized').This feather allows us to visualize subtle differences between the neighboring clusters of profiles and distinguish the unique characteristics of nodes through the variation of specific features across the SOM plane.
The SOM classification reveals that for the whole study period, the ABL is dominated by near neutral to strong stable conditions, as none of the SOM nodes fall within the unstable category (i.e., super-adiabatic condition).The results are reasonable, considering the daily temperature profile is composited from 20:00 and 08:00 measurements.According to the SOM ordering feather, the SOM nodes in four corners (i.e., nodes 1, 3, 7 and 9) can be thought of as the typical ABL types and the others can be considered transitional ABL types.It is clear from the individual profiles in Fig. 1 that node 1 represents the well-mixed (near-neutral) condition with no temperature inversion, node 3 indicates the ABL type dominated by elevated inversion, node 7 indicates the ABL type dominated by surface inversion, and node 9 represents the ABL type associated with multiple inversions (i.e., including surface and elevated inversions).
Frequency analysis of the nine ABL types indicates that the frequency distribution across the types is quite varied from the expected 11.1 %, with the occurrence frequency showing a 5:1 range from the most frequent type (node 1) to the least frequent type (node 5).The higher-frequency types are presented on the outer portions of the SOM plane, while lesser-frequency types are presented closer towards the center (top-right in Fig. 1).The most dominant types are nodes 1 and 3, and their occurrence frequencies reach 22 % and 20 %, respectively.As synoptic circulations change with the seasons over Beijing, the ABL types are expected to correspond to seasonality.The number of profiles from each season in each ABL type is expressed as a percentage and is shown in Fig. 2. All of the types exhibit strong seasonality.For example, node 1 has the highest occurrence in spring (29.4 %) and the lowest occurrence in autumn (13.7 %); node 9 presents the highest occurrence in winter (16.3 %) and the lowest occurrence in summer (4.9 %).corresponding to each ABL type.As seen in Fig. 3, each of the ABL types is associated with distinct dynamic and thermodynamic conditions.The potential temperature profiles vary from near neutral conditions to strong stable conditions, and this change is closely related to the variance in wind speed, suggesting a strong coupling between the dynamic and thermal effects.The two extreme types (nodes 1 and 9) provide a very useful example.Node 9 is a very strong stable profile, and the wind speeds are very low in the lower ABL.In contrast, node 1 is a well-mixed (near neutral) profile and it corresponds to significantly higher wind speeds throughout the ABL.In addition, when the stability of the atmosphere is strong, vertical mixing is suppressed and winds in the lower ABL become decoupled from the generally stronger wind aloft.This allows moisture, fogs, low clouds and other scalars to build up within the stable layer.As a result, the stable ABL types usually correspond to high RH in the lower ABL.

Evaluation against meteorological data
The near-surface meteorological variables are also examined for each of the ABL types.Fig. 4 shows the diurnal composite plots of surface temperature, wind speed and relative humidity in the four typical ABL types.
As expected, these near-surface variables respond well to the changing ABL structure.Wind speeds are the highest on the days corresponding to near neutral conditions (i.e., node 1).High wind speeds result in a deep, mechanically mixed layer, and these days also exhibited the smallest diurnal amplitude in wind speed, temperature and relative humidity.Such characteristics are likely consistent with the passage of frontal systems.In contrast, the smallest wind speeds are observed on days related to strong stable conditions (i.e., node 9).The stable days also generally exhibit the greatest amplitude of the diurnal signals in temperature and relative humidity.This fact is an indication that stable conditions occur mostly on clear sky days.

Evaluation against surface air quality
The concentrations of gaseous and particulate pollutants in the atmosphere are governed by the rate at which they are emitted from their respective sources, lost by various sink mechanisms, and characteristics of the atmospheric volume into which they mix.While the mixing volume is determined primarily by the boundary layer structure, the chemical transformation also depends on boundary layer meteorology in some cases.In the previous section, it was seen that the SOM technique is an effective tool for classifying boundary layer structures.In this section, we used the classification technique to quantify the influence of the boundary layer structure on near-surface air quality.As expected, the most stable conditions are associated with a dramatic increase in the mass concentrations of air pollutants (except O 3 ).On average, SO 2, NO 2 , CO, PM 10 and PM 2.5 increase by 15.7 μg/m 3 (142 %), 44.3 μg/m 3 (119 %), 1.5 mg/m 3 (202 %), 91.6 μg/m 3 (119 %) and 95.9 μg/m 3 (218 %) from the near neutral ABL condition (i.e., node 1) to strong stable condition (i.e., node 9), respectively.The highest increasing amplitude is related to PM 2.5 , suggesting fine particulate matters are likely accumulated from not only primary emissions but also secondary formation (Zhang and Cao, 2015).As we have shown, the more stable ABL conditions tend to correspond to high relative humidity in the lower ABL (Figs. 3 and 4).Additional enhancement in PM 2.5 can be expected under the humid condition, as it is known that the humidity-related physicochemical formation of particles (such as hygroscopic growth, liquid-phase and heterogeneous reactions) can be intensified by high humidity values (Cheng et al., 2015;Cheng et al., 2016;Zheng et al., 2015).
Interestingly, increasing atmospheric stability has an opposite effect on near-surface O 3 concentrations.Since O 3 is produced by photochemical interactions between NO x (NO + NO 2 ) and volatile organic compounds (VOCs) (Seinfeld and Pandis, 2006), the boundary layer structure alters the O 3 level through modulation of its precursors (NO x and VOCs).The low O 3 level in the stable ABL can be explained by the strong titration reaction.Since O 3 is highly reactive, when trapped in a stable layer, surface titration by the NO emitted from vehicles can cause a rapid reduction in O 3 concentration.In previous studies, persistent low O 3 concentration were observed in the stable boundary layer condition in Beijing (Zhao et al., 2013).Conversely, when near-surface wind speeds are higher (near neutral condition such as node 1), O 3 is mixed downward from the overlying air mass, resulting in higher concentrations.Nevertheless, it is worth noting that the extremely high O 3 values (not shown) were also detected on very stable days (i.e., node 9), suggesting the complexity of O 3 behavior in response to the boundary layer structure (Tong et al., 2011;Haman et al., 2014).
To obtain a more in-depth understanding of the physical mechanisms behind the relationship between air quality and ABL structure, diurnal composite hourly concentrations of atmospheric pollutants are formed for each ABL type.The SOM-based ABL classification scheme provides a consistent, gradual distinction in the diurnal cycles of surface air pollutants from near neutral to strong stable conditions.The composite diurnal evolutions of air pollutants in the four typical ABL types (i.e., nodes 1, 3, 7 and 9) are illustrated in Fig. 6.The diurnal cycles of SO 2 , NO 2 , CO, PM 10 and PM 2.5 are extremely pronounced under the strong stable condition (i.e., node 9), although very reduced under the near neutral condition (i.e., node 1).In contrast, the behavior of O 3 is completely different from other pollutants.The results suggest that the chemical species, which are mainly produced by  et al., 2016).Overall, the diurnal behavior of each pollutant species in each of the ABL types is generally consistent with the existing knowledge for urban areas (Chambers et al., 2015b;Chambers et al., 2015a;Zhang et al., 2012b;Jenner and Abiodun, 2013;Han et al., 2009).
Of particular interest in Fig. 6 is that (1) nodes 3 and 9 have similar magnitudes of concentrations in the afternoon, and (2) nodes 7 and 9 have similar increments in concentrations from afternoon to midnight, although there is a huge distinction in the afternoon concentrations (i.e., afternoon baselines).This sheds some light on the common patterns of the ABL controls on the near-surface air quality in Beijing.Considering the thermal inversion feather in each of ABL types (Fig. 1), the regulating effects of ABL on near-surface concentrations can be concluded as follows: (1) elevated inversion enhances the afternoon baseline; and (2) surface inversion improves the evening increment.Obviously, the high afternoon baselines in nodes 3 and 9 can be attributed to elevated inversion, while high evening increments in nodes 7 and 9 can be attributed to surface inversion.Since surface inversion usually develops shortly before sunset due to radiation cooling, the evening traffic emission peak is counteracted by a stabilizing boundary layer.Consequently, the air pollutants such as NO 2, CO, PM 10 and PM 2.5 often experience an explosive growth from afternoon to midnight.In contrast, elevated inversion usually forms due to synoptic forcing (such as synoptic advection) (Hu et al., 2014;Xu et al., 2016) and can persist for several days; as a result, the daytime mixing volume is also depressed, causing a relatively higher afternoon concentration.
Beijing has relatively little industry but numerous automobiles, and the emissions of SO 2 are small while those of CO, NO x and particles are much larger (Zhao et al., 2012).By comparison, the diurnal behaviors of SO 2 and other pollutants are completely different.For example, in node 9, SO 2 show a lower nighttime concentration but a sharp increase after sunrise, whereas NO 2 , PM 10 and PM 2.5 show a higher nighttime concentration with a slight morning increase associated with the traffic emission.The results largely suggest that the changing ABL structure affects the near-surface observations of locally and remotely sourced pollutants in very different ways.In the evening, since the stable boundary layer (SBL) and the residual layer (RL) are essentially decoupled with each other (Stull, 1988), locally sourced pollutants emitted into the surface layer (such as CO, NO 2 and particulate matters from vehicular emissions) become trapped close to the surface.In contrast, remotely sourced pollutants emitted from chimneystacks above the SBL (such as SO 2 from power plants in the Hebei Province) may be stored within the RL aloft and not penetrate into the SBL.As the daytime convective turbulent mixing developed in the morning, the rapid momentum transfer between the surface and aloft air transported the pollutants stored in RL downward and meanwhile upwardly mixed the pollutants trapped from the previous night in the surface layer (Salmond and McKendry, 2005).It is observed in Fig. 6 that after a stable night, the burst of turbulent activity in the morning coincides with a rapid increase in SO 2 concentration (Fig. 6).Since there is no significant increase in SO 2 emission at the surface at this time, this result strongly suggests that increased SO 2 in the morning resulted from the downward mixing of stored SO 2 in the RL aloft.In a previous case study, Li et al. (2017b) reported that as a result of both turbulent mixing and the advection of high concentrations of air pollutants above the surface layer, the urban area of Beijing experienced a dramatic increase of the PM 2.5 concentration in the morning on 30 November 2015.
Given the importance of local vehicle emissions vs. more-distance power plant and industrial emissions for Beijing's air quality, the ratio of CO/SO 2 can be considered as an indicator of the contribution of local emissions to air pollution, with higher ratios indicating higher local contributions (Tang et al., 2015).Fig. 7 shows the composite diurnal variations of CO/SO 2 ratios in the four typical ABL types (i.e., nodes 1, 3, 7 and 9).The contrasts between CO/SO 2 ratios for the various ABL types are noticeable during the nighttime, whereas differences during the daytime are minimal.During the daytime, when the ABL is well mixed, near-surface pollutant concentrations represent a combination of local and remote sources.In the evening, however, the earth's surface begins to cool, and a stable boundary layer begins to form from the ground up.If sufficiently strong, the nocturnal surface inversion can isolate near-surface observations from the influence of distant sources (Crawford et al., 2016).Consequently, the more stable the nocturnal conditions near the ground, the higher the CO/SO 2 ratios that occur (Fig. 7).The results are consistent with previous studies (Tang et al., 2015;Zhu et al., 2016), indicating local contribution increases with enhanced static stability in the surface layer over Beijing.According to the above analysis, high pollutant loadings in node 9 are mostly attributable to local contributions (the highest CO/SO 2 ratios in node 9); however, high pollutant loadings in node 3 are more likely due to regional contributions (the lowest CO/SO 2 ratios in node 3).Obviously, it is the stable stratification rather than the weak surface wind that confines the regional contribution.

Evaluation against columnar aerosol pollution
For many years, aerosol particles have been the primary pollution problem in Beijing.Atmospheric aerosols play an important role in radiation transfer due to absorption and/or scattering in the atmosphere, and thus could have a great influence on the evolution of the ABL.In recent years, the feedback effect of aerosols on the ABL has drawn much attention (Kajino et al., 2017;Gao et al., 2016;Ding et al., 2016).To further our understanding of aerosol pollution in Beijing, we examine the optical and physical properties and the direct radiative forcing of columnar aerosols in the different ABL types in this section.
Aerosol optical properties can be characterized by three useful parameters: AOD, AE and SSA.Fig. 8 illustrates the AOD 440nm , AE 440nm-870nm and SSA 440nm over Beijing within the nine ABL types.The ABL-type averages of AOD range from 0.52 and 1.22 (Fig. 8a).Comparing with near-surface observations, the greatest difference is that the highest AOD value generally occurs in node 3, rather than in node 9 (the highest surface PM 2.5 and PM 10 concentrations occur in node 9).This may be attributed to the difference in aerosol vertical distribution in these two types.As we have demonstrated in Sect 3.3, node 3 is related to strong regional transport.Since the height of regional transport is usually above the surface layer, such as 200-700 m AGL detected by Li et al. (2017a), more aerosol particles might be suspended above the surface layer in node 3, resulting in the highest AOD value in the atmospheric column.In addition, since high relative humidity also occurs in node 3, the highest AOD value in this ABL type could be partly attributed to the particle hygroscopic growth (Chen et al., 2014;Deng et al., 2016;Zhao et al., 2017).
It is known that high AE values indicate a dominance of fine particles, while low values indicate a dominance of coarse particles.Unlike AOD, the AE shows a relatively low sensitivity to ABL types (Fig. 8b).All type averages of AE are higher than 1.0, suggesting that the proportion of fine particles is always larger than that of coarse particles over Beijing (Yu et al., 2017;Yu et al., 2009).The highest AE occurs in node 6 (1.20) and the lowest is 1.03 in node 1. Node 1 corresponds to the lowest AE value, indicating that under the near neutral ABL condition, the coarse particles contribute a relatively higher proportion of total particles.This could be due to the increasing wind speed with decreasing relative humidity (Figs. 3 and 4).Coarse particles could be from more natural and anthropogenic dust emission under high wind speed conditions.Particularly during the fast northwesterly wind period, dust storms occasionally contribute to the high coarse particle loadings in Beijing (Yu et al., 2016).The long-distance transport of dust particles from northwest China may be the reason for the lowest AE value in node 1.
The SSA is defined as the ratio of the scattering coefficient and the total extinction coefficient.It is mostly dependent on the aerosol size, concentration of absorbing component and its mixture state with non-absorbing components.The daily SSA at 440 nm ranges from 0.82 to 0.98 during the study period, which suggests that there are quite different types of aerosols in the columnar atmosphere over Beijing (varying from strong absorbing aerosols to strong scattering aerosols).It is easy to see that the ABL types associated with a strong surface inversion (i.e., nodes 7, 8 and 9) have lower SSA values (Fig. 8c).The averaged SSA in these nodes is approximately 0.90, which is significantly lower than that in nodes 1, 2 and 3.The low SSA values mean enhancement in the absorbing particles, such as black carbon, which are released from industry, biomass/biofuel burning, diesel vehicle, and coal burning.In contrast, the highest SSA occurring in node 1 can be explained by dust particle transmission and soil aerosol emissions.
The volume particle size distribution retrieved in the sizes between 0.05 and 15 μm is one of the most important parameters for studying the behavior of aerosols (Dubovik and King, 2000).Fig. 9 expresses the mean volume particle size distribution (dV/dlnR) over Beijing in the nine ABL types.Table 1 supplements Fig. 9 with the statistical parameters of aerosol particle size distribution.Clearly, the volume particle size agrees very well with the bimodal lognormal distributions.Both fine (R < 0.6 μm) and coarse (R > 0.6 μm) modes exhibit relative stability with two peaks at a radius of approximately 0.1-0.2μm and 2.0-4.0 μm, which are similar to some previous studies (Eck et al., 2005;Xia et al., 2007;Che et al., 2014).However, the size distribution shows a distinct difference in the changing amplitude for different ABL types.The fine-and coarse-mode particle volumes increase rapidly from left (nodes 1, 4 and 7) to right (nodes 3, 6 and 9) on the SOM plane.This suggests that with the stabilizing boundary layer processes, the atmosphere is more loaded with both fine-and coarse-mode particles over Beijing.In addition, the stabilizing processes are accompanied by the increase of the fine-mode effective radius (R eff -F) and fine-mode volume fraction (Vf/Vt).These results strongly point to the important role of fine-mode particle hygroscopic growth on the days associated with stable nocturnal ABL conditions.
The type-averaged ARF at the surface (BOA), top of atmosphere (TOA), and within the atmosphere (ATM) over Beijing is shown in Fig. 10.The type averages of ARF at the surface range from -47.8 W/m 2 to -110.0 W/m 2 , while at the TOA, they are found to be between -21.1 W/m 2 and -48.0 W/m 2 .Likewise, the ABL type averaged ARF within the atmosphere are between 26.7 W/m 2 and 63.1 W/m 2 .The larger negative ARF at the surface (> 110 W/m 2 ) and positive ARF within the atmosphere (> 60 W/m 2 ) are found in ABL types 3, 6 and 9 over Beijing, implying strong cooling at the surface and warming in the atmosphere.These results are induced by relatively larger aerosol loadings under the stagnant meteorological conditions.The larger ARF within the atmosphere demonstrates that solar radiation is being absorbed within the atmosphere, and as a result, heats the atmosphere Atmos.Chem.Phys.Discuss., https://doi.org/10.5194/acp-2017-1046Manuscript under review for journal Atmos.Chem.Phys.Discussion started: 13 December 2017 c Author(s) 2017.CC BY 4.0 License.and reduces surface temperature.This can change the atmospheric vertical temperature gradient and improve the atmospheric stability (Li et al., 2010;Ge et al., 2010;Zou et al., 2017).Finally, the enhanced stability hinders the vertical diffusion of aerosol particles, leading to the increase of aerosol concentrations and causing a further decrease in the solar radiation and ABL height, which induces a positive feedback loop for causing high surface aerosol concentrations (Quan et al., 2013;Zhong et al., 2017).

Evaluation against heavy polluted episodes
In January of 2013, December of 2015, and December of 2016, heavy aerosol pollution episodes frequently wreaked havoc across Beijing and its surroundings, which resulted in severe damages to the environment and human health.Fig. 11 shows the hourly variations of PM 2.5 and AOD 440nm during the three heavily polluted months.It is observed that the PM 2.5 concentrations were frequently elevated to above 200 μg/m 3 , and the AOD often exceeded 1.0 in Beijing during these three months.The ABL types (shown at the top of each plot) reveal that pollution episodes were generally associated with the control of nodes 3 and 9, and clean episodes were often associated with the control of node 1.For example, the severe pollution episode that occurred from 9-14 January 2013 was due to the alternate control of nodes 3 and 9, and the pollution episode from 15-21 December 2016 was related to the persistent control of node 9.In contrast, multiday control of node 1 caused a clean episode from 14-16 December 2015.The linkages between air quality and the boundary layer structure were consistent with the long-term analyses described in Sects.3.3 and 3.4, indicating that the ABL types are one of the primary drivers of day-to-day variations in air quality over Beijing.
The monthly PM 2.5 concentrations in the Beijing urban area reached up to 180.8 μg/m 3 , 153.9 μg/m 3 and 147.9 μg/m 3 in January 2013, December 2015 and December 2016, respectively.All these values were far larger than the 4-yr winter mean PM 2.5 concentration (110.6 μg/m 3 ).Although the characteristics of PM 2.5 air quality depend on many complex elements, the major contributors are the pollutant emissions and meteorological conditions.In 2013, the Chinese State Council released the "Atmospheric Pollution Prevention and Control Action Plan" to implement a megacity cluster-scale joint prevention and control strategy program.As a result, the PM 2.5 in Beijing decreased from 89.5 μg/m 3 in 2013 to 73.0 μg/m 3 in 2016.However, these meteorology-driven pollution episodes to some degree obscure the true impacts of the emission control strategies implemented by government.Fig. 12 shows a comparison of the occurrence frequency of the nine ABL types to the winter mean frequency (2013)(2014)(2015)(2016) for the three polluted months.Compared with the 4-yr winter mean frequency, the greatest differences are that the occurrences of nodes 3 and 9 (the two most polluted types) increased and node 1 (the clean type) decreased during the three polluted months.Obviously, the elevated PM 2.5 concentrations in the abovementioned months can be mostly attributable to the anomalous boundary layer structures.
Quantitative analysis of the roles of the ABL anomaly in PM 2.5 variations during the pollution months is helpful for the assessment of air pollution prevention and control strategies.In this study, the ABL classification allows for the integrated evaluation of the effects of numerous interrelated ABL meteorological parameters on air quality.
Here, a meteorology-to-environment method (revised from the circulation-to-environment method proposed by Zhang et al. (2012a)) is utilized to evaluate the influence of the ABL anomaly for enhanced PM 2.5 levels during the abovementioned months.We assume the linkages between ABL types and their PM 2.5 loadings in winter are constant in different years.For each polluted month, the total anomaly (C') is defined as the deviation in PM 2.5 from the 4-yr winter mean concentration ( C ).This total anomaly in each month is due to the combined effects of meteorology and emission.The anomaly calculated from the mean PM 2.5 loadings for nine ABL types and their occurrence frequencies during each month can be considered to represent the PM 2.5 change caused by the anomalous boundary layer structure.We refer to this as the "ABL-driven" anomaly.The ABL-driven anomaly , where F i is the occurrence frequency of type-i ABL during a specific period and C i is the corresponding PM 2.5 loading feathering that type.

Summary
The influence of the ABL structure on Beijing's air quality is still unclear due to the lack of long-term observations.On the other hand, the long years of routine radiosondes remain underutilized as a tool for urban pollution studies.In this study, the SOM was applied to 4-yr radiosondes to classify the ABL types over Beijing.
The resulting types were then evaluated in relation to meteorological and environmental variables, with an attempt to understand the roles of different ABL conditions in regulating the air quality in Beijing.The main findings are as follows: 1) The SOM provides a continuum of nine ABL types (i.e., SOM nodes), and each type is characterized with

Fig. 3
Fig.3shows the average vertical profiles of potential temperature, wind speed and relative humidity

Fig. 5
Fig.5examines the daily concentrations of gaseous and particulate pollutants in relation to various ABL types.

Figure 1 .
Figure 1.The 3 × 3 SOM output for radiosonde-based temperature (T) deviation profiles observed at the Beijing Observatory.SOM nodes are shown in red, with the corresponding individual profiles in dark blue.For reference, the overall average temperature profile and 25th and 75th percentile profiles are shown in cyan.The top-right shows the occurrence cases and frequency of each SOM node.

Figure 3 .
Figure 3. Profiles of average potential temperature (θ), wind speed (WS) and relative humidity (RH) corresponding to each SOM node at the Beijing Observatory.The red, green and black labels of the horizontal axis correspond to θ, WS and RH, respectively.

Figure 4 .
Figure 4. Hourly mean diurnal composites of temperature, wind speed and relative humidity in Beijing corresponding to SOM nodes 1, 3, 7 and 9.

Figure 5 .
Figure 5. Daily pollutant concentrations in Beijing corresponding to each SOM node.The solid dots denote 799 the mean.The box and whisker plot presents the median, the first and third quartiles, and the 5th and 95th 800 percentiles, respectively.801 802

Figure 8 .
Figure 8.(a) Aerosol optical depth (AOD 440nm ), (b) Ångström exponent (AE 440nm-870nm ) and (c) single scattering albedo (SSA 440nm ) over Beijing and corresponding to each SOM node.The solid dots denote the mean.The box and whisker plot presents the median, the first and third quartiles, and the 5th and 95th percentiles, respectively.

Figure 9 .
Figure 9. Mean volume particle size distribution over Beijing corresponding to each SOM node.The average volume particle size distribution for each node is shown by the gray line and is repeated on each plot for comparison.The size distribution for each type is highlighted in the black dotted line on the respective plot.

Figure 10 .
Figure 10.Aerosol radiative forcing (ARF) at the surface (BOA), top of atmosphere (TOA), and within the atmosphere (ATM) over Beijing and corresponding to each SOM node.
The ratio of C ABL ' to C' (the difference of C ABL ' to C ) is then used to evaluate the relative (absolute) contribution of the ABL anomaly to the enhanced PM 2.5 level.The results show that the contributions of the frequency anomaly of the ABL type to the