The multi-site observations were synchronously conducted on the North China Plain from 15 November to 31 December 2023 (Fig. S1). Four of the sampling sites are located in urban areas, namely Beijing (BJ), Tianjin (TJ), Handan (HD) and Jinan (JN), which are surrounded by the traffic arteries and dense residential and commercial buildings. And the remaining one, adjacent to Luancheng (LC), serves as a rural station free from significant industrial influences. At each site, the PM_2.5 samples with a 12 h interval were collected onto prebaked (at 450 °C for 6 h) quartz filters using high- (1.13 m³ min⁻¹) or medium-volume (100 L min⁻¹) air samplers; All the samplers located on the rooftops approximately 15–20 m above ground level. After sampling, the filter samples were wrapped in prebaked aluminum foils and stored in a freezer (at -18 °C) prior to analysis.

Hourly concentrations of PM_2.5 and associated pollutants (e.g., NO₂, O₃, CO) at the sampling sites were obtained from the National Urban Air Quality Real-time Release Platform of China (https://air.cnemc.cn:18007/, last access: 6 April 2025), of which monitoring sites are adjacent to ours with a distance of <10 km. The meteorological data including ambient temperature (T) and relative humidity (RH) were downloaded from National Climatic Data Center (https://www.ncei.noaa.gov/, last access: 6 April 2025).

A piece of each filter was extracted with 40 mL ultrapure Milli-Q water (18.2 MΩ) under ultrasonication for 30 min. A part of the extract was used for the detection of water-soluble ions (SO42-, NO3-, Cl⁻, Na⁺, NH4+, K⁺, Mg²⁺ and Ca²⁺) by using an ion chromatography; The remaining part was detected for the water-soluble organic carbon (WSOC) and water-soluble total nitrogen (WSTN) via a total organic carbon (TOC) analyzer (Model TOC-L CPH, Shimadzu, Japan). And the water-soluble organic nitrogen (WSON) can be quantified by deducting the water-soluble inorganic nitrogen (WSIN) from WSTN (i.e., WSON = WSTN - WSIN), thereby revealing variations in the N:C ratios of water-soluble components across different sampling sites (Fig. S2). A DRI-model 2001 thermal–optical carbon analyzer following the IMPROVE-A protocol was applied here to measure organic carbon (OC) and element carbon (EC) of PM_2.5.

The molecular compositions in PM_2.5 samples, including nitro-aromatic compounds (NACs, Table S1), polycyclic aromatic hydrocarbons (PAHs) and others organic tracers, were extracted with a mixture of dichloromethane and methanol (2:1, v/v); Subsequently, the extracts were derivatized via derivatization reagent (a mixture of 50 µL of N,O-bis-(trimethylsilyl)trifluoroacetamide (BSTFA) and 1 % trimethylsilyl chloride and 10 µL of pyridine) prior to GC-MS analyses. More details of the extraction, derivatization and the GC-MS analyses can refer to elsewhere (Wang et al., 2006; Wu et al., 2025a).

Additionally, the aerodyne high-resolution time-of-flight aerosol mass spectrometer (HR-AMS) was applied here to characterize the imidazole-related fragments (e.g., C₃H₃N2+ and C₃H₄N2+) in a typical humid haze event (12 to 31 December of 2023, Fig. S3), following the method reported by Ge et al. (2024). It should be pointed out that above fragments may stem from other nitrogen-containing precursors, which inevitably introduces uncertainties into the quantification. Accordingly, the concentrations derived from these fragment ions cannot represent the actual abundances of the real molecules. Even so, the fragment ion levels can still indicate the variation trends of imidazole N-heterocycles. Because imidazole concentrations measured by orbitrap mass spectrometry in our previous field campaign showed a significant positive correlation with offline-AMS signals of C₃H₃N2+ and C₃H₄N2+ (Fig. S4).

The details of the extraction and derivatization can be found elsewhere (Wu et al., 2024a). Briefly, a part of the PM_2.5 filter was extracted following the procedure similar to that of other water-soluble components; then the extracts were atomized using argon as the carrier gas, dehydrated by a diffusion drier, and ultimately quantified by HR-AMS. Each sample was continuously monitored by HR-AMS for >10 min to obtain stable signal, and above procedure was also applied to the blank samples to account for any potential contamination or background signals. A deep post-processing was conducted for the V-mode data in this study using the Igor-based Aerosol Mass Spectrometer Analysis Toolkit. The mass concentration of imidazole-related fragments was calculated as the follow: 1CIMs=FIMs×WSOM Where CIMs denotes to the concentration of the imidazole-related fragments. FIMs is mass contribution of the imidazole-related fragments to the total fragments measured by HR-AMS; WSOM refers to atmospheric concentration of water-soluble organic matter (µg m⁻³)

The light-absorbing chromophores (i.e., BrC) were also extracted with 10 mL ultrapure Milli-Q water, following the procedure for water-soluble ions. A UV–vis spectrometer was applied here to record the absorption spectra of all the extracts, which were finally converted into absorption coefficient (abs_λ, M m⁻¹) at a given wavelength (λ) using the following equation: 2absλ=(Aλ-A700)vLva×l×ln⁡(10) Where Aλ is the absorbance at wavelength λ (nm), vL and va refer to the volume of the extractant and air through corresponding to filter punches, respectively. The optical path length is 1 cm (i.e, l). From abs_λ, the mass absorption coefficient (MAE_λ, m² g⁻¹) can be characterized as: 3MAEλ=absλC Where C is the mass concentration of water-soluble organic carbon.

Random forest (RF), as a powerful tool for regression and prediction (Hu et al., 2017; Vu et al., 2019), was employed here to elucidate the relationships between secondary WSON (WSONsec) and potential factors (NH4+, NO₂, O₃, ALWC, pH and meteorological factors). The importance of each factor to WSONsec can be determined by a SHAP method originally proposed by Lundberg and Lee (2017). Rooted in cooperative game theory, SHAP framework as a model-agnostic additive feature attribution method enables an in-depth exploration of interaction effects and effectively alleviates poor interpretability of machine learning models. In the RF model design, 70 % of the dataset including all the samples from the five monitoring sites was randomly divided into a training subset to construct the RF model, and the remaining 30 % (i.e., test set) was used to assess the model performance. Following systematic hyperparameter optimization, the decision tree (ntree) was set to 300 to balance model performance and computational efficiency; and number of variables split at each node (nmtry) was limited to 12 to avoid overfitting. Additionally, a 10-fold cross-validation strategy was adopted here to optimize model parameters and estimate model performance. The RF model was constructed using the “randomForest” R package, of which performance was evaluated by coefficient of determination (R2), mean square error (MSE) or root-mean-square error (RMSE) and mean absolute error (MAE). As summarized in Table S2, the simulated WSONsec corrected strongly with observed ones, and the error metric remained at low levels; Thess indicated that the model could reconstruct commendably the variations of WSONsec in this study.

Since gas-phase nitro-aromatic compounds (NACs) were not detected in this study, the particulate-phase fraction (Fp) of NACs was calculated based on absorption equilibrium theory (Pankow et al., 2001; Zuend and Seinfeld, 2012), and the specific calculation method referred to that reported by Chen et al. (2025). As shown in the following equations: 4Fp=CpCp+Cg=Kp×PM1+Kp×PM5Kp=CpCg×PM=760×f×R×T106×PV×M×ξ Herein, Cg and Cp are the mass concentrations of the organic molecule in gas and particle phase, respectively. Kp refers to gas–particle partitioning coefficient, and is determine by many factors in Eq. (5). Specifically, PV is the saturation pressure (Pa), which is 0.00378 torr for 4-nitrophenol (at 298 K). ξ is the activity coefficient of the species, adopting from the values reported by Wang et al. (2019). M refers to the molecular weight of NAC species (g mol⁻¹). Additionally, R is the gas constant (8.314 J (mol K)⁻¹), T is the temperature (K), f denotes the fraction of organic matter in particle, and PM is the mass concentration of PM_2.5.

As shown in Table 1 and Fig. S3, the PM_2.5 concentration across the North China Plain (NCP) exhibited significant spatial heterogeneity, ranging from 4.0 to 223.7 µg m⁻³. Specifically, the highest average PM_2.5 load was recorded in JN (94.1 ± 37.8 µg m⁻³), partially attributed to unfavorable topographic conditions and intensive anthropogenic emissions. The average PM_2.5 level at remaining sampling sites decreased gradually from south to north. Notably, the PM_2.5 concentration in Beijing during the campaign was only a quarter of that recorded a decade ago (158 µg m⁻³), indicating a notable air quality improvement in BJ over recent years. Even so, multi-day haze episodes (PM_2.5 > 75 µg m⁻³) still occurred frequently in BJ (Fig. S3), of which hourly peak concentration even reached up to 203 µg m⁻³, about 3-fold of National Air Quality Standard grade-II (75 µg m⁻³). Additionally, these events were also observed simultaneously at other sites, suggesting that regional PM_2.5 pollution remains a persistent challenge during winter across the NCP, particularly in its southern region.

The spatiotemporal differences in chemical compositions, sources of organic matter (OM) in PM_2.5 among sampling sites are illustrated in Fig. 1. As the major component of PM_2.5, water soluble ions (WSIs) accounted for approximately 30 %–53 % of PM_2.5; Similar to the spatial pattern of PM_2.5, WSIs also exhibited high loads in JN (40.4 ± 23.2 µg m⁻³) and HD (44.1 ± 34.3 µg m⁻³), which were ∼ 1.2–4 fold of those measured in other sites. From Fig. 1, nitrate was the most abundant ion across the NCP, accounting for 14 %–27 % of PM_2.5, followed by sulfate (5 %–9 %) and ammonium (4 %–11 %), respectively. Compared to those in urban sites, relative abundance of SNA (SO42-, NO3- and NH4+) in rural site (LC) moderately enhanced by 10 %–23 %, indicating a significant influence of secondary aerosol formation in the rural area. Furthermore, a relatively high concentration of chloride was detected among the sampling sites, even with a molarity being >1.5-fold of that of sulfate (Fig. S5a). A strong correlation between chloride and levoglucosan+Benzo[k]fluoranthene (BkF) across the sampling sites (R2=0.97) implied that combustion emissions were the primary source for the abundant chloride in the NCP (Fig. S5b). These abundant chlorides could promote heterogeneous formation of sulfate as evidenced by a robust relationship chloride and SOR (SO42-/(SO42-+SO2)) (Fig. S5c). This finding was consistent with the results of our laboratory experiments (Wang et al., 2025b, 2016), which demonstrated that the abundant chloride could significantly enhance the uptake of NO₂ by interfacial electrostatic attraction, subsequently accelerating sulfate formation.

The average concentration of OM ranged from 7.9 ± 5.7 to 18.7 ± 10.7 µg m⁻³, accounting for 17 %–25 % of PM_2.5. Although OM was a predominant species in PM_2.5 at most sites, its spatial distribution differed slightly from that of PM_2.5, with high loads in LC (18.5 ± 11 µg m⁻³) and HD (18.7 ± 10.7 µg m⁻³). To determine the sources of the abundant OM, a PMF was applied here and identified three types OM sources (Fig. S6, Sect. S1). From Fig. 1, a significant fraction of OM (40 %–55 %) at urban sites was associated with secondary formation, comparable to that in northern regions of China (50 %) (Chen et al., 2024b). In contrast, combustion-derived primary OM accounted for less than 40 % of the total OM. This finding was consistent with the field observations, which demonstrated a widespread decline in primary OM in NCP during 2013–2020 due to the significant reduction in residential fuel burning (Chen et al., 2024b). Notably, about 55 % of OM at rural site was derived from combustion, reflecting an urban-rural variation in sources.

Figure 2a displays the average absorption spectra of WSOC across the sampling sites, which exhibits the marked feature of BrC with the reduced light absorption from the ultraviolet to the visible ranges. As illustrated in Table 1 and Fig. 2, the light absorption coefficient at 365 nm (abs₃₆₅) displayed a spatial pattern similar to that of organic matter (OM) across the five sampling sites. Specifically, the highest average abs₃₆₅was observed at LC site, with average of (8.0 ± 4.7 Mm⁻¹), followed by HD (7.9 ± 5.0 Mm⁻¹), JN (6.9 ± 3.0 Mm⁻¹), TJ (5.7 ± 3.8 Mm⁻¹) and BJ (2.3 ± 1.9 Mm⁻¹), respectively. Across the sampling sites, abs₃₆₅ correlated robustly (R2=0.73, p<0.05) with levoglucosan being a known tracer for biomass burning (BB). This implied that water-soluble BrC in the NCP was significantly affected by fresh emissions from BB, which aligns with the prior observations in China (Li et al., 2022b; Desyaterik et al., 2013; Li et al., 2023). Additionally, the BB emission was more pronounced in rural areas, as indicated by a higher mass fraction of levoglucosan to OM (0.67 ± 0.39 %), which was approximately 2.1 to 5 times higher than those in urban areas. These findings highlight the significant impact of anthropogenic combustion on air quality in the rural NCP. From Fig. 2b, the average light-absorption of urban BrC in the NCP exhibited a decline of ∼ 45 % from 10.7 ± 3.0 Mm⁻¹ in 2018 to 5.8 ± 2.2 Mm⁻¹ in 2023. Beijing exhibited a more substantial reduction, where the abs₃₆₅ was only 16 % of that recorded a decade ago (Fig. 2b). Even so, the averaged abs₃₆₅ in the NCP still remains higher than that in Nanjing, Guangzhou and urban areas in developed countries (Table S3).

As a metric for characterizing the light absorptivity of BrC, the MAE at 365 nm (MAE₃₆₅) was quantified by the linear regression slope of abs₃₆₅ against WSOC. As illustrated in Fig. 2c, MAE₃₆₅ exhibits a distinctly different spatial pattern relative to abs₃₆₅. A strikingly high MAE₃₆₅ value was measured in JN (1.4 ± 0.04 m² g⁻¹), on par with those found in severely polluted regions like Xingtai, Xi'an and Delhi, India (1.2–1.6 m² g⁻¹, in the cold season) (Li et al., 2023; Wu et al., 2020; Kirillova et al., 2014); And the MAE₃₆₅ value also falls within the range for documented BrC emitted from residential coal combustion (e.g., 1.20–1.59 m² g⁻¹ for bituminous coal) (Ni et al., 2021), suggesting that coal combustion is one of the potential sources of atmospheric BrC in JN. Additionally, TJ, LC and HD shared a comparable MAE₃₆₅ value of approximately 1.1 m² g⁻¹, which closely matches the values associated with biomass burning-derived BrC (1.2 ± 0.3 m² g⁻¹) (Cao et al., 2021). This finding further indicated a significant effect of biomass burning on the wintertime BrC in the NCP, which can be corroborated by the concentration-weighted trajectory (CWT) analysis. As shown in Fig. 3, high CWT loadings for abs₃₆₅ were predominantly associated with regions characterized by concentrated fire hotspots, especially in LC, HD and their adjacent areas. Of note, the MAE₃₆₅ of the BrC in BJ was the lowest (0.71 ± 0.02 m² g⁻¹), only half that of JN. This relatively weak light absorptivity of BrC in BJ may be attributable to substantial vehicle emissions, as BrC from this source typically exhibits relatively low MAE values, ranging from approximately 0.35 to 0.71 m² g⁻¹ (Tang et al., 2020; Huang et al., 2022). Another possibility is that the aerosol in BJ was more age, because BrC light-absorption would decay in the aerosol aging process as revealed by observational and experimental findings (Hems et al., 2021; Qiu et al., 2024). This can be confirmed by the Benzo[e]pyrene/(Benzo[a]pyrene+Benzo[e]pyrene) ratio (BeP/(BaP+BeP)). Since BaP is easily degraded via atmospheric oxidation while BeP is chemically stable, a higher BeP/(BaP+BeP) value indicates more aged atmospheric aerosols. As shown in Fig. S7a, the BeP/(BaP+BeP) ratio in BJ was relatively higher than those at other sampling sites. Moreover, a negative correlation between BeP/(BaP+BeP) and MAE₃₆₅further supported this hypothesis (Fig. S7b). In Fig. 2d, we can note that unlike the temporal evolution of abs₃₆₅, the MAE₃₆₅ of the water-soluble BrC at most sampling sites undergone indistinctive variations in response to the reduction in anthropogenic emission.

As depicted in Fig. 4a, the optical properties of water-soluble BrC in different periods are plotted in the AAE-MAE₄₀₅ (log scale) space, with reference to the optically defined BrC classes proposed by Saleh (2020). On average, the water-soluble BrC across the sampling sites exhibits optical characteristics consistent with biomass-burning and coal-combustion sources, falling into the “weakly” absorbing BrC class. This observation reinforced remarkable impacts of combustion sources on BrC in the NCP. Of particular note, the BrC light-absorptivity was enhanced from clean (PM2.5< 75 µg m⁻³) to haze (PM_2.5 > 75 µg m⁻³) periods at most sampling sites, coinciding with the increasing N:C ratio (Fig. 4a). And a positive correlation between N:C ratio and abs₃₆₅ was also found across the NCP (Fig. 4b), indicating that NOCs likely serve as the potential chromophores modulating BrC optical properties in this region (Figue S8). Accordingly, the high N:C ratio was also associated with an elevated WSOC/OC ratio (Fig. 4b), implying that the secondary formation was likely the predominant source of the N-containing chromophores in the NCP.

To obtain the contribution of secondary formation to the water-soluble organic nitrogen, a tracer method analogous to that used for estimating secondary organic carbon (SOC) with EC as a tracer was adopted here (Sect. S2 in the Supplement). This approach assumes that combustion-related primary WSON (WSONpri) share the common sources with EC and can be approximated using a representative primary WSON/EC ratio. Based on the sensitivity tests (Table S4), the primary WSON/EC ratios used here can capture temporal variations of the WSONpri, which positively correlated with CO at all sampling sites. Even so, the secondary WSON (WSONsec) may still be overestimated because a fraction of WSONpri may arise from non-combustion sources. Therefore, the derived secondary fractions here should be interpreted as semi-quantitative estimates rather than precise source apportionments. From the results illustrated in Fig. 5a, a dominant role of secondary formation in WSON accumulation was observed across five sites. Specifically, secondary WSON in BJ explained more than approximately 64 ± 21 % of the total WSON, which was comparable to those at other urban sites but was approximately 1.2-fold of that in LC. Such spatial pattern was attribute to the enhanced primary emission (e.g., biomass burning) in rural site. Additionally, the fractional contribution of WSONsec increased with the rising PM_2.5 levels (Fig. S8) across sampling sites, particularly in BJ. Above findings further highlighted the critical role of secondary formation in regulating the abundance of strongly light-absorbing N-containing chromophores during haze periods.