Field observations were conducted at a national air-quality monitoring station in Yangzhou (32.41° N, 119.40° E), a representative urban center within the YRD of China. The monitoring site is situated 20 m above ground level in a mixed residential, commercial, and educational district, representing typical urban background conditions in the region. To examine atmospheric responses to abrupt emission changes during the COVID-19 pandemic, we selected four rain-free five-day episodes with broadly comparable meteorological conditions, comprising two emission control periods (ECPs) and their corresponding normal periods (NPs): a winter ECP (W-ECP, 29 January–2 February 2020) and its winter NP (W-NP, 20–24 February 2020), as well as a summer ECP (S-ECP, 17–21 August 2021) and its summer NP (S-NP, 17–21 September 2021).

Single-particle chemical composition, vacuum aerodynamic diameter (Dva), and mixing state were measured in real time using a single-particle aerosol mass spectrometer (SPA-MS; Hexin Analytical Instrument Co., Ltd., China). The instrument primarily detected particles with Dva ranging from 0.2 to 2.0 µm, as detailed previously (Dai et al., 2024). Concurrently, PM_2.5 mass concentrations were measured by a particulate matter monitor (XHPM2000E, Xianhe, China), while trace gases (NO_x, SO₂, CO, and O₃) were quantified using Thermo Scientific analyzers (Models 42i, 43i, 48i, and 49i) (Fig. S1 in the Supplement). All instruments were co-located to ensure the temporal and spatial consistency of the aerosol and precursor datasets.

Meteorological parameters were retrieved via Google Earth Engine (GEE) from the NASA Goddard Earth Observing System Composition Forecast system (GEOS-CF) and the ERA5-Land hourly dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF). Ten variables were extracted for the grid cell nearest to the monitoring site (Table S2 in the Supplement): 2 m air temperature (T2m), 10 m air temperature (T10m), eastward and northward wind components at 2 m (U2m, V2m), surface net solar radiation (SSR), planetary boundary layer height (BLH), sea level pressure (SLP), total precipitation (TPREC), 2 m dew point temperature (DP2m), and relative humidity (RH). These variables were selected for their potential roles in influencing atmospheric stability, pollutant transport, and chemical transformation kinetics. All datasets were temporally synchronized to an hourly resolution to facilitate integrated analysis with the chemical measurements.

A total of 324 090 ambient particles were detected and classified based on their SPA-MS mass spectra. Particles were clustered using the ART-2a algorithm in MATLAB R2024a (MathWorks, Inc.) via the Computational Continuation Core toolkit (COCO v1.4). The algorithm was executed for 20 iterations with a vigilance factor of 0.8 and a learning rate of 0.05, parameters consistent with established single-particle aerosol studies (Huang et al., 2023; Zhang et al., 2021a; Dai et al., 2024). Following the initial clustering, particle types were aggregated into five operational principal groups according to their dominant mass-spectral signatures. Black Carbon (BC) particles were identified by carbon-cluster ions (e.g., m/z ±12[C]^+∕−, ±36[C₃]^+∕−, ±48[C₄]^+∕−, and ±60[C₅]^+∕−). BC with organic carbon (BCOC) particles exhibited the same BC signatures together with characteristic organic-carbon (OC) fragments (e.g., m/z 27[C₂H₃]⁺, 29[C₂H₅]⁺, 37[C₃H]⁺, and 43[C₂H₃O]⁺). BC-free inorganic (BF) particles lacked BC cluster signatures and were dominated by inorganic species. BC-free OC-rich (BFOC) particles also lacked BC clusters but showed a high relative abundance of OC fragments. Residual subclasses with mixed or ambiguous signatures were defined as “Other”.

To better resolve S-N partitioning, we further divided these principal groups into 19 subclasses. Subclasses enriched in secondary inorganic species (SIS) were identified by sulfate (-97[HSO4-], -96[SO4-]), nitrate (-62[NO3-], -46[NO2-]), or both, and were labeled with the suffixes “-S”, “-N” and “-SN”, respectively. Particles exhibiting strong cyanide-related ions (-26[CN⁻] and -42[CNO⁻]) were labeled with “-CN”. Number fractions and representative mass spectra are summarized in Table S1 and Fig. S2.

Aerosols were further grouped into fresh-like (-CN) and aged-like (-sec) based on diagnostic SPA-MS fragments and Dva (Figs. 1f and 2). Strong cyanide-related ions together with carbonaceous fragments are commonly observed in relatively fresh combustion particles from biomass burning, coal combustion, and vehicle exhaust (Huang et al., 2023; Sun et al., 2017; Zhang et al., 2020a). We therefore treated CN-containing subclasses within each principal group as operational fresh-like states. For instance, BCOC-CN particles exhibit abundant carbon-cluster and hydrocarbon fragments accompanied by cyanide signatures (Fig. 2h), occurring predominantly at smaller Dva (100–600 nm), consistent with freshly emitted combustion aerosols (Sharma et al., 2018). Similarly, BFOC-CN particles display +39[K⁺], -26[CN⁻], and levoglucosan-derived fragments (-45[CHO2-], -59[C₂H₃O2-], -73[C₃H₅O2-]) and are predominantly found at small sizes, representing primary organic aerosol from biomass burning (Fig. 2c) (Bi et al., 2011; Huang et al., 2023).

Aged-like types (-sec) were constructed by regrouping sulfate-rich (-S), nitrate-rich (-N) and mixed (-SN) subclasses into composite SIS-enriched categories within each principal group, reflecting extensive secondary processing (Xie et al., 2020). BCOC-sec types display coexisting BC, OC and strong SIS fragments (Fig. 2g), indicating substantial internal mixing of carbonaceous and secondary inorganics, in line with previous observations (Dai et al., 2024; Xie et al., 2020). BC-sec (BC-S, BC-N, BC-SN; Fig. 2f showed negligible OC signals but elevated SIS ions, while BF-sec and BFOC-sec (BF-S, BC-N, BC-SN; BFOC-S, BC-N, BC-SN) exhibit elevated sulfate and nitrate ions but lack BC fragments (Fig. 2b, d).

The consistency of this fresh-aged framework is supported by both compositional and temporal variations. The relative contribution of SIS signal is systematically lower in “-CN” than in “-sec” types for BC, BCOC, BF and BFOC groups (Fig. S3a). In addition, hourly number concentrations of each “-sec” type correlate most strongly with the “-CN” type within the same principal group (Fig. S3b). Together with the systematic increase in Dva from “-CN” to “-sec”, these diagnostics support interpreting “-CN” and “-sec” as fresh-like and aged-like compositional states within each principal group. We therefore define four dominant fresh-like/aged-like pairs: BC-CN/BC-sec, BCOC-CN/BCOC-sec, BF-CN/BF-sec, and BFOC-CN/BFOC-sec.

To quantify S-N partitioning across gas, particles, and number domains, we define three complementary SNR metrics: gas-phase SNR (gSNR), particle-phase SNR (pSNR), and number-based SNR (nSNR). The gSNR is calculated as the molar ratio of SO₂ to NO_x from hourly observations: 1gSNR=[SO2][NOx]. The pSNR is derived from SPA-MS spectra by calculating the ratio of total peak areas (PA) of sulfate ions (-97[HSO4-], -96[SO4-]) to nitrate ions (-62[NO3-], -46[NO2-]) for a selected particle or particle class: 2pSNR=PAsulfatePAnitrate. Although SPA-MS is semi-quantitative, uniform acquisition and processing of sulfate- and nitrate-related ion signals provide a consistent basis for comparing relative variations in pSNR. Because ion responses and matrix effects may vary, pSNR is interpreted as a relative spectral indicator rather than an absolute sulfate-to-nitrate molar or mass ratio, consistent with prior SPA-MS studies (Healy et al., 2013; Zhang et al., 2021a).

The nSNR represents the ratio of the number concentrations (N) of sulfur-containing particles to nitrogen-containing particles, serving as a population-level proxy for external mixing tendency of sulfate- versus nitrate-containing particles: 3nSNR=Nsulfur-containingNnitrogen-containing.

To characterize particle-type-resolved hygroscopicity, we inferred an effective hygroscopicity parameter (κeff) for each particle class at hourly resolution using aggregated SPA-MS ion signals. Major ion peaks were assigned to several chemical components (e.g., sulfate, nitrate, organics, black carbon, and potassium salts). Sulfate and nitrate were represented by ammonium sulfate and ammonium nitrate, respectively. For each component i, the hourly summed peak area Ai was normalized by its assumed density ρ_i (e.g., ammonium nitrate for nitrate, ammonium sulfate for sulfate; Table S3) to obtain an equivalent volume. The volume fraction of component i was then approximated as 4ϵi≈Ai/ρi∑j(Aj/ρj). The effective hygroscopicity of each particle class was calculated using κ-Köhler theory and the Zdanovskii-Stokes-Robinson (ZSR) mixing rule (Petters and Kreidenweis, 2007): 5κeff=∑i(ϵi×κi), where κi is the literature-derived hygroscopicity parameter of component i. Table S3 lists the lower, mean, and upper κi values for ammonium sulfate and ammonium nitrate and the prescribed values for the other components. The mean values were used in the baseline calculation. Sensitivity was evaluated using four combinations of the lower and upper values for ammonium sulfate and ammonium nitrate (lower-lower, lower-upper, upper-lower, and upper-upper), while the component volume fractions, densities, κi values of the other components, and ZSR mixing rule were held constant.

The resulting κeff represents an hourly, particle-type-resolved estimate derived from aggregated SPA-MS spectra. SPA-MS ion signals are semi-quantitative and may be affected by species-dependent ionization efficiencies and matrix effects. The estimates also assume ammonium salts for sulfate and nitrate and volume-additive mixing under the ZSR rule. Accordingly, κeff is interpreted as a relative indicator for comparisons among particle classes, periods, and aging states, rather than as a direct hygroscopicity measurement.

Causal networks, expressed as directed acyclic graphs (DAGs), characterize conditional dependence structures and identify directed relationships compatible with potential causal pathways among variables (Koller, 2009). In this study, we applied the Causal Learner toolbox in MATLAB R2024a (MathWorks, Inc.) (Ling et al., 2022) to explore potential relationships among SNR metrics, meteorological parameters, and gaseous pollutants.

For each target SNR metric (gSNR, pSNR, nSNR), the Grow-Shrink (GS) algorithm was first applied to determine its Markov blanket (MB), the minimal variable set directly linked to the target. The GS procedure alternates between adding conditionally dependent variables (grow phase) and removing conditionally independent variables (shrink phase) until convergence (Margaritis, 1999). The MB-derived subset was further refined using the Greedy Equivalence Search (GES) algorithm, which iteratively improves the global causal structure via adding edges (forward phase) and removing redundant links to prevent overfitting (backward phase) (Chickering, 2003). The resulting DAGs identified direct and indirect dependency pathways linking SNR evolution to meteorological and chemical drivers. The statistical support and physical plausibility of these pathways provide a process-consistent structural basis for subsequent predictive modeling.

Building on the DAG-derived causal structure, we used the eXtreme Gradient Boosting (XGBoost) algorithm to construct separate models for each aged-like particle class (BCOC-sec, BC-sec, BFOC-sec, BF-sec), with hourly pSNR and nSNR modeled in turn as target variables, to quantify the relative influence of the identified drivers on their variability. All predictors were standardized by the z-score method before model fitting. The dataset was split into six folds: five folds for iterative training and validation, and one held-out fold reserved exclusively for interpretation to prevent information leakage. Model hyperparameters were tuned via the Sparrow Search Algorithm (SSA), a swarm-intelligence metaheuristic that balances global exploration (discoverer phase) and local exploitation (joiner phase) to prevent premature convergence (Yang et al., 2025). SSA was run with a population size of 10 and 50 iterations. Model performance was assessed using root-mean-square error (RMSE) and the coefficient of determination (R2), averaged over five validation folds (Table S4). The SSA-XGBoost model achieved high R2 (0.71–0.91) for major particle types, ensuring the robustness of the predicted SNR values.

We interpreted the trained models using SHapley Additive exPlanations (SHAP), a cooperative game-theoretic approach that decomposes individual predictions into additive feature contributions (Ekanayake et al., 2022). SHAP values quantify both the magnitude and direction of each predictor's effect on pSNR or nSNR. This approach provides transparent and physically meaningful insights into how gSNR, meteorology, and gaseous pollutants modulate SNR variability across aged-like particle types.

Using the four episodes defined in Sect. 2.1, we assessed atmospheric responses to abrupt anthropogenic emission shifts across two COVID-19 ECP and NP cycles in Yangzhou (Figs. 1 and S1). During the ECPs, surface NO_x concentrations decreased sharply relative to the corresponding NPs, by 39 % in winter and 34 % in summer (Fig. 1a). In contrast, SO₂ changed little and remained within 3.5 to 3.9 ppb. This asymmetric response significantly enhanced gSNR, with seasonal averages increasing from 0.21 to 0.36 (+71 %) in winter and from 0.35 to 0.56 (+60 %) in summer. O₃ was about 15 % higher during ECPs, consistent with weaker NO titration (Wang et al., 2022). PM_2.5 decreased modestly during ECPs, by 13 % in winter and 9 % in summer (Fig. 1a).

Time series show that gSNR, pSNR, and nSNR varied synchronously, and their fluctuations were more pronounced during ECPs than during NPs (Fig. 1c). This co-variation indicates that changes in precursor composition were accompanied by coherent responses in particle composition and in the population-level abundance of sulfur- and nitrogen-containing aerosols. Total particle number concentration (N) increased during NPs while PM_2.5 mass remained similar, suggesting a shift toward smaller particles that contribute more strongly to number than to mass after controls were relaxed (Fig. 1d).

As shown in Fig. 1b, BC-containing particles dominated the aerosol population across all episodes, accounting for an average of 58 % of total detections and peaking at 69 % during S-ECP. Freshly emitted particles accounted for approximately 16 % of detections during ECPs and rose to ∼ 26 % in the NPs, reflecting the rebound in primary emissions. Relative to NPs, ECPs were associated with more sulfate-dominated secondary inorganic species (SIS) and pronounced particle growth across particle classes. Averaged across all categories, pSNR was about 57 % higher during ECPs than during NPs (Fig. 1e), with a stronger shift in BC-containing than in BC-free particles and a larger contrast in summer than in winter. Mean Dva was increased by about 27 % during ECPs, with the largest increases in aged-like types (Fig. 1f). Across all principal groups, aged-like categories consistently showed higher pSNR and broader size distributions than their fresh-like counterparts, consistent with enhanced oxidation under elevated gSNR, favoring sulfate-rich growth.

To quantify compositional effects on pSNR variation, representative aged-like particles were decomposed into BC, OC, and SIS components (Fig. 2). The BF-sec group was used as the non-carbonaceous reference. To quantify the relative sulfate-to-nitrate enrichment in carbonaceous particles, we defined a normalized pSNR enhancement factor (λ) as follows: 6λi=pSNRipSNRBF-sec, where i corresponds to BFOC-sec, BC-sec, and BCOC-sec particle types, yielding λOC, λBC, and λBCOC, respectively.

Figure 3 shows pronounced diurnal and seasonal contrasts in λ, with larger amplitudes and greater variability in summer. A distinct daytime transition is evident in summer, with λBC rising from 1–3 in the morning to 6–9 near local solar noon, exceeding λOC (typically 3–6) and remaining dominant until evening (Fig. 3d). This behavior is consistent with the hypothesis of enhanced sulfate formation on BC-containing particles during photochemically active periods, potentially linked to BC-mediated HONO/NO2 release (Liang et al., 2021; Ye et al., 2017) and photo-induced formation of reactive oxygen species (ROS) that accelerate the conversion of SO₂ to sulfate (Zhang et al., 2022; Zhu et al., 2020). At night, when solar radiation is negligible and RH is high (> 80 %), λOC exceeds λBC, aligning with aqueous-phase sulfate production through H₂O₂ or O₃ oxidation and with possible organosulfate formation (Cheng et al., 2016; Wang et al., 2021b). Increased organic-phase viscosity may further limit gas-particle partitioning (Vaden et al., 2011), water uptake (Rickards et al., 2015), and multiphase reactions (Shiraiwa et al., 2011), which could prolong sulfate retention within OC-rich particles. Laboratory and modeling studies also suggest that BC can catalyze SO₂ oxidation even in darkness (He et al., 2018; He and He, 2020), although this likely represents a minor nighttime pathway compared with OC-driven aqueous processes.

Across the summer diurnal cycle, λBCOC consistently exceeds both λBC and λOC, peaking in the afternoon (Fig. 3d). This persistent enhancement suggests a synergistic role of BCOC, BC provides reactive or catalytic surfaces, while OC facilitates sulfate accommodation or retention within the particle matrix. Such synergy likely reflects the complex surface chemistry of BC, involving polycyclic aromatic hydrocarbons (PAHs) and aromatic oxygenated species, which may promote SO₂ uptake and facilitate heterogeneous oxidation under ultraviolet (UV) irradiation (Riva et al., 2016). Light-absorbing organic compounds, such as brown carbon and humic-like substances, may further act as photosensitizers and accelerate SO₂-to-sulfate conversion (Wang et al., 2020b).

In winter (Fig. 3a, b), all λ values are smaller, but the compositional contrasts remain. λOC generally exceeds λBC throughout the day, consistent with a larger relative contribution of OC-related aqueous pathways under weaker photochemistry. Nevertheless, λBCOC remains the largest, indicating that BCOC mixtures still provide the most favorable environment for sulfate enrichment even under low-radiation, low-temperature conditions. This persistent BCOC advantage may reflect not only enhanced photochemical/heterogeneous processing, but also microphysical effects, such as larger accessible surface associated with soot morphology (Beeler et al., 2025) and coating softening or water uptake (Li et al., 2021) that promotes sulfate partitioning and retention.

These observations show that the diurnal variation of pSNR is strongly composition-dependent. BC-containing particles show the strongest enhancement during photochemically active hours, OC-rich particles become relatively more important under humid nighttime and winter conditions, and mixed BCOC particles show the largest enhancement in both seasons, likely driven by combined chemical and microphysical effects.

To characterize particle-type-dependent aging, we jointly examined changes in inferred κeff, pSNR, and Dva and developed a conceptual framework in the κeff-pSNR space (Fig. 4). As defined in Sect. 2.3, each fresh-like subtype was paired with its dominant aged-like counterpart. Stage I compares fresh-like and aged-like populations during NPs, whereas Stage II compares aged-like populations between NPs and ECPs. These comparisons represent population-level contrasts rather than trajectories of individual particles. To evaluate the robustness of inferred κeff, calculations were repeated using alternative hygroscopicity values for ammonium sulfate and ammonium nitrate (Table S3; Fig. S4). The resulting ranges represent parameter sensitivity within the adopted mixing rule rather than the full uncertainty in κeff. Because relative ionization efficiencies and matrix effects can affect SPA-MS signals, pSNR changes are interpreted as relative ion-signal shifts rather than absolute composition changes.

For OC-rich fresh particles (BCOC-CN and BFOC-CN), Stage I aged-like populations had an approximately 5 % lower mean Dva and an 83 % higher pSNR than their fresh-like counterparts, whereas κeff changes were generally small or parameter dependent. No consistent κeff increase or decrease was obtained for BCOC in either season or BFOC in winter, while the summer BFOC response exhibited changes of 2.5 %–17.0 %. This pattern may reflect compaction and the formation of an organic-rich shell that retains sulfate while limiting additional water uptake (Mikhailov et al., 2009; Riemer et al., 2019). During Stage II, stronger oxidation is associated with an increase in Dva of about 25 % and a further rise in pSNR of about 80 %, again with minimal change in κeff, consistent with continued sulfate accumulation and coating growth on a compact and relatively hydrophobic matrix.

For BC particles with weaker organic signatures, the transition from BC-CN to BC-sec exhibited the largest κeff increase (∼ 31 %). This response remained positive across all tested parameter combinations, ranging from 16.8 % to 27.5 % in winter and from 28.5 % to 47.1 % in summer (Fig. S4). Together with increases in Dva (∼ 24 %) and pSNR (∼ 87 %), this pattern indicates dominant inorganic uptake during NPs. During Stage II, pSNR increased markedly (∼ 105 %) and Dva increased by ∼ 36 % to ∼ 700 nm, while the inferred κeff increase remained small across the tested parameter combinations (1.0 %–4.6 %). This shift indicates enhanced sulfate-rich coating growth as gSNR and oxidant levels rise.

For BC-free particles (BF-CN to BF-sec), pSNR remained relatively low and changed only slightly. Stage I κeff changes remained positive but small in both winter (1.3 %–5.6 %) and summer (1.3 %–7.3 %). The Stage II κeff response was also modest in winter (0.9 %–7.3 %) but stronger in summer (11.6 %–18.2 %). Together with the approximately 22 % increase in Dva, these changes are consistent with increased secondary inorganic contributions, particularly during summer ECPs.

Overall, the coupling between inferred κeffand pSNR strengthened under higher gSNR and stronger oxidative conditions, although the response remained particle-type dependent. BC-containing particles showed pronounced sulfate enrichment, consistent with the accumulation of secondary sulfate coatings on soot particles. OC-rich particles also exhibited strong pSNR enhancement but comparatively limited changes in inferred hygroscopicity, whereas BC-free particles showed weaker compositional adjustment and more gradual particle growth. Thus, the two-stage framework captures distinct composition-dependent responses rather than a uniform aging pathway, reflecting differences in core type, coating chemistry, and mixing state.

Ammonium sulfate (AS) and ammonium nitrate (AN) are the dominant SIS in PM_2.5 (Fu et al., 2016; Lin et al., 2020). Their deliquescence relative humidity (DRH), together with the mutual deliquescence RH (MDRH) of their mixtures, provide reference points for interpreting particle phase transitions and aqueous processing (Liu et al., 2021; Sun et al., 2018). These transitions may shift in the presence of organic constituents and other components in ambient particles (Li et al., 2021). As RH increases, the associated rise in aerosol liquid water (ALW) reduces matrix viscosity, enhances molecular transport, and promotes morphological restructuring and internal mixing (Mikhailov et al., 2009; Shiraiwa et al., 2011). In our observations, increasing RH is accompanied by generally declining and progressively converging gSNR, pSNR, and nSNR, indicating a shift toward more balanced S-N partitioning across gas, particle, and number domains (Fig. S5a, e). In parallel, BCOC-sec consistently exhibited the highest pSNR and nSNR, while BF-sec remained the lowest, but these contrasts weaken as RH increases, suggesting that elevated humidity reduces particle-type compositional and population-level contrasts of S-N partitioning (Fig.  S5b, f). Guided by these physicochemical thresholds and the observed SNR evolution, we divide the observations into four characteristic RH regimes (Fig. 5). Because particle phase state and ALW were not directly measured, these regimes are operational ranges rather than fixed phase-transition boundaries.