2',7'-Dichlorodihydrofluorescein diacetate (DCFH-DA, ≥97 %) was purchased from Aladdin, and horseradish peroxidase (HRP, ≥200 unitsmg-1) was obtained from MREDA. Potassium dihydrogen phosphate (KH2PO4, 99.5 %), dipotassium hydrogen phosphate trihydrate (K2HPO4⋅3H2O, 99.0 %), sodium hydroxide (NaOH, 97.0 %), ethanol (CH3CH2OH, 99.8 %), and hydrogen peroxide (H2O2, 1000 µgmL-1) were purchased from Macklin. All chemicals were of analytical grade, and deionized (DI) water was used for solution preparation. The phosphate buffer solution (PBS, pH 7.0) was prepared by mixing KH2PO4 and K2HPO4⋅3H2O in DI water at appropriate ratios. To obtain the fluorescent reaction solution, DCFH-DA was dissolved in ethanol, hydrolyzed with NaOH for 30 min in the dark, and subsequently diluted with PBS to yield 10 µmolL-1 2',7'-dichlorodihydrofluorescein (DCFH). For the enzyme reaction solution, HRP was dissolved in PBS to a final activity of 2 unitsmL-1. All reagents were stored at low temperature and protected from light.

In this study, an online analyzer was developed to simultaneously quantify ROS_g and ROS_p using wet-chemical collection coupled with a fluorescence-probe assay. The instrument features two parallel systems, each composed of integrated sampling, delivery, reactor, and detection units, for real-time, in situ monitoring of both phases (Fig. 1). For ROS_g, ambient air was first passed through a membrane filter to remove particles; the gaseous fraction was then absorbed into solution using a glass spiral absorption tube, followed by a gas-liquid separation chamber. The collected solution was mixed with the fluorescent probe and enzyme reagent in a premixing chamber and was subsequently derivatized at constant temperature in a reaction chamber prior to fluorescence detection. For ROS_p, PM_2.5 was size-selected using a cyclone, and residual ROS_g was removed by a rotating wet annular denuder (WAD). The remaining particles were collected in a spray growth collection chamber and transferred to the liquid phase for analysis, after which the same derivatization and detection steps as in the gaseous system were applied. A WAD was not used as a single-step phase separator because its large liquid holdup volume would lower the solution renewal rate and sensitivity, thereby degrading temporal resolution and smoothing short-term concentration variability.

The core principle of ROS quantification relies on the HRP-catalyzed oxidation of DCFH by ROS. As illustrated in Fig. 2, ROS_g and ROS_p samples were first mixed with HRP. HRP was oxidized by ROS to form the active intermediate compound I (HRP-I). This intermediate then oxidized two DCFH molecules with weak native fluorescence to the highly fluorescent product DCF. Upon excitation at 470 nm, DCF emits fluorescence at 520 nm, and the signal intensity was proportional to the ROS concentration in the sample. Thus, ROS_g and ROS_p were quantified from the measured fluorescence signal. It should be noted that the DCFH-HRP assay is not equally sensitive to all ROS species. Previous characterization of DCFH-based atmospheric ROS measurements showed that peracetic acid produced a response close to that of H2O2, whereas sterically hindered organic peroxides, such as tert-butyl hydroperoxide, benzoyl peroxide, lauroyl peroxide, and 2-butanone peroxide, exhibited much lower relative sensitivities (Zhou et al., 2018). Therefore, the measured signal represents an operationally defined fraction of water-soluble, DCFH-reactive oxidants expressed as H2O2 equivalents, rather than the total abundance of all atmospheric ROS. Consequently, ROS species with low DCFH-HRP reactivity or limited aqueous stability may contribute less efficiently to the fluorescence signal, resulting in species-dependent response biases.

The instrument operates via three integrated stages: sample collection, chemical derivatization, and optical detection. ROS in gas and particle phases are separately captured using wet absorption and spray growth collection chamber, then mixed online with DCFH and HRP to generate fluorescent products, enabling real-time, simultaneous measurement through a compact LED-PMT detection module.

Calibration procedures were conducted using high-purity N2 at flow rates exceeding sampling conditions to ensure system stability. A series of H2O2 standard solutions and blank controls were introduced into the reaction chamber, and real-time signals were recorded via LabVIEW. The fluorescence response typically reached equilibrium within 10 min, after which data were collected for an additional 10 min to obtain steady-state averages. All measured ROS concentrations are expressed as H2O2 equivalents and should therefore be regarded as operationally defined ROS responses rather than absolute or molecule-specific quantification of total atmospheric ROS.

The concentrations of ROS_g and ROS_p in the sampling solutions were determined from voltage signals acquired in real-time using a LabVIEW program, following the relationship between fluorescence intensity and concentration: 1Ia=KCROS-I0 where Ia is the measured fluorescence signal intensity, I0 is the baseline intensity, and CROS is the concentration of ROS in the sampling solution (µgH2O2L-1). The coefficient K was derived from the fluorescence intensities corresponding to standard H2O2 solutions. A blank-corrected through-origin working curve was used for quantification; the blank signal was subtracted from the measured fluorescence signal prior to concentration calculation, thereby defining zero ROS concentration as zero net fluorescence response.

The atmospheric concentration of ROS_g was calculated as: 2CROSg(µgm-3)=CROSFIFgγg where CROSg (µgm-3) is the ROS_g concentration, FI is the absorption liquid flow rate in the glass spiral absorber (1.0 mLmin-1), Fg is the gas sample flow rate (1.0 Lmin-1), and γg is the collection efficiency of the glass spiral absorber for ROS_g. As the collection efficiency exceeded 99 %, sampling losses were considered negligible.

The mass concentration was further converted to volumetric mixing ratios (ppbv) using: 3CROSg(ppbv)=CROS(µgm-3)VmM(T0/T)(P/P0) where Vm is the molar volume at standard conditions (22.4 Lmol-1), M is the molecular weight of H2O2 (34 gmol-1), T0 and P0 are the standard temperature (273.15 K) and pressure (101.325 kPa), T and P are the ambient temperature and pressure during measurements.

The atmospheric concentration of ROS_p was calculated as: 4CROSp(µgm-3)=CROSF0Fgγp where CROSp (µgm-3) is the ROS_p concentration, F0 is the collection liquid flow rate in the spray growth collection chamber (1.0 mLmin-1), Fg is the aerosol sampling flow rate (16.7 Lmin-1), and γp is the collection efficiency of the spray growth collection chamber. The value of γp was determined from recovery experiments based on three fractions: the chamber-collected liquid (Mcol), the wall-rinse solution obtained by washing the chamber and associated tubing to recover deposited material (Mwall), and the downstream backup filter-rinse solution used to quantify particle breakthrough (Mfilter). For each fraction, the fluorescence response was measured under the same DCFH-HRP detection conditions and converted to an H2O2-equivalent amount using the corresponding calibration curve. The collection efficiency was defined as: 5γp=McolMcol+Mwall+Mfilter

The recovery experiment was conducted in triplicate under actual ambient aerosol sampling conditions. The mean recovered fractions were 81.72 % in the chamber-collected liquid, 9.46 % in the wall-rinse solution, and 8.80 % on the downstream backup filter. The wall-rinse and backup-filter fractions were used to evaluate wall deposition and particle breakthrough, respectively. Since only the chamber-collected liquid was directly delivered to the second premixing chamber during routine online operation, γp=81.72 % was used as the effective online collection efficiency for ROS_p quantification in Eq. (4). This value represents an operational mean collection efficiency for the present field deployment. Potential variations associated with particle loading, hygroscopic growth, aerosol chemical composition, and long-term operation should be further evaluated in future applications.

To enhance detection performance, systematic optimization of key parameters was conducted. This included reagent concentrations (DCFH and HRP), reaction temperature, and PMT/LED settings. Orthogonal experiments and systematic testing were performed to determine the optimal combination that maximizes fluorescence sensitivity, stability, and signal-to-noise ratio for reliable ROS quantification.

To ensure accurate and reliable quantification of atmospheric ROS in both gas and particle phases, a comprehensive performance assessment was conducted (Table 2). This included tests of baseline stability, detection limits, reproducibility, sensitivity, response time, and linear working range. These evaluations confirm the instrument's suitability for long-term field deployment under complex atmospheric conditions.

Potential interferences in the DCFH-HRP assay may arise from direct probe oxidation, aqueous consumption of H2O2, or side reactions affecting DCF formation. For oxidizing gases, previous studies have shown that O3 interference is generally weak under ambient-relevant conditions. The dissolved O3 level estimated at ∼40 ppbv is insufficient to produce measurable DCFH-based artifacts, and laboratory tests at 60–80 ppbv O3 showed negligible responses (Huang et al., 2016; King and Weber, 2013). Field evaluations further indicated that even 100 ppbv O3 produced a maximum H2O2 quantification error below 0.03 ppbv, whereas positive artifacts became important only under extremely elevated O3 levels up to several hundred ppbv due to secondary oxidant formation (Lazrus et al., 1986; Montesinos et al., 2015). Since the O3 levels observed in the present campaign were within the ambient-relevant range evaluated in these studies, O3-related interference was considered a minor positive artifact in the gas-phase channel.

Reducing gases and soluble transition metals were more directly relevant to potential negative biases in the present DCFH-HRP configuration. NO has been reported to cause only weak H2O2 loss, whereas SO2 can suppress H2O2 detection through aqueous S(IV) chemistry with strong dependence on concentration and pH (Hua et al., 2008; Lazrus et al., 1986; Komazaki et al., 2001). Fe2+ can also consume H2O2 through Fenton-type reactions, while Fe3+ shows limited interference under comparable conditions (Kolthoff and Medalia, 1949; Zhou et al., 2018). Therefore, direct laboratory tests were conducted for SO2, NO, Fe2+, and their mixtures under the same reaction conditions as the analyzer. The interference tests used a 3 µgL-1 H2O2 standard as the reference solution, and the H2O2-only solution was used as the control. The tested atmospheric-equivalent levels included 0.5, 1, 10, and 25 ppbv for SO2; 10, 25, 50, and 100 ppbv for NO; and 20, 80, 160, and 400 ngm-3 for Fe2+. Two mixed conditions were further examined: Mix1, an environmentally representative mixed condition composed of 1 ppbv SO2, 25 ppbv NO, and 80 ngm-3 Fe2+, and Mix2, a high-level sensitivity condition composed of 10 ppbv SO2, 100 ppbv NO, and 160 ngm-3 Fe2+. As shown in Fig. 4, NO produced negligible interference across the tested range, with biases from -1.02 % to 0.04 %. SO2 showed a clear concentration-dependent negative effect. The biases were minor at 0.5 and 1 ppbv SO2 (-0.20 % and -0.23 %) but increased to -10.00 % and -29.97 % at 10 and 25 ppbv SO2, respectively, consistent with H2O2 consumption by dissolved S(IV). Formaldehyde was not used in the present flow configuration; therefore, SO2-related effects were retained as a potential negative interference rather than assumed to be chemically suppressed.

Fe2+ also caused a systematic negative bias, increasing from -0.05 % at 20 ngm-3 to -2.73 %, -4.69 %, and -7.53 % at 80, 160, and 400 ngm-3, respectively. This trend agrees with Fe(II)-driven Fenton-type H2O2 consumption. The mixed-interferent tests provided a direct estimate of total interference under coexisting soluble species. The total bias was -3.23 % for Mix1 and -11.03 % for Mix2, indicating that the combined effect was mainly negative and governed by SO2 and Fe2+, without evidence of additional synergistic amplification.

Based on the calibration conversion used in this study, the Mix1 bias corresponds to an atmospheric-scale uncertainty of approximately -0.056 ppbv for ROS_g and -0.006 µgm-3 for ROS_p, which is below the instrumental detection limits. The Mix2 bias corresponds to approximately -0.19 ppbv for ROS_g and -0.022 µgm-3 for ROS_p and represents a conservative high-level sensitivity scenario. Therefore, the quantified interference was unlikely to affect the seasonal pattern or pollution-regime interpretation of the present field observations, although high-SO2 or Fe-rich environments may lead to underestimation of ROS and should be further evaluated in future applications.

Field observations of ROS_g and ROS_p were conducted at the Peking University Computing Center in Haidian District, northwestern Beijing, China (39.99° N, 116.31° E). The site is located in the Zhongguancun area, a densely populated urban district characterized by intensive educational, commercial, residential, and traffic activities. It is therefore affected by mixed urban emissions, including vehicle exhaust, residential and commercial activities, and regional transport from surrounding areas of the North China Plain. These features provide a complex urban atmospheric environment for evaluating the field applicability of the analyzer under seasonally varying photochemical and particulate pollution conditions. The field campaign covered four seasons: autumn (29 October–15 November 2024), winter (10–31 December 2024), spring (24 April–15 May 2025), and summer (13 June–5 July 2025).

In addition to ROS measurements, routine atmospheric pollutants were continuously monitored in real time. O3 was monitored using a Model 49i ozone analyzer (Thermo Fisher Scientific, USA), while NOx and SO2 were measured with Model 42i and Model 43i-TLE analyzers, respectively (Thermo Fisher Scientific, USA). PM_2.5 mass concentration was determined using a TH-2000Z1 monitor (Tianhong, China). Photolysis frequencies, including j(O1D), j(HONO), and j(H2O2), were derived from actinic flux spectra recorded over 280–650 nm using a UF-CCD spectroradiometer (MetCon, Germany) followed by spectral inversion. HONO, NH3, and HNO3 were quantified using a gas aerosol collector-ion chromatography (GAC-IC) system (Peking University, China). VOCs were measured by ZF-PKU-VOC1007 system (Pengyu Changya, China), and non-refractory components in submicron aerosol particles were characterized by ToF-ACSM (Aerodyne Research, USA).

Observation days were classified as polluted days when the daily maximum 8 h average O3 concentration exceeded 160 µgm-3, corresponding to a mixing ratio of ∼80 ppbv, or when the 24 h average concentration of PM_2.5 and PM₁₀ exceeded 75 and 150 µgm-3, respectively. Days that did not meet any of these criteria were defined as clean days.

Figure 5 presents the seasonal and pollution-dependent variations of ROS and associated atmospheric species, and the corresponding seasonal statistics under clean days (CDs) and polluted days (PDs) are summarized in Table S2. Notably, ROS_g and ROS_pexhibited a consistent seasonal pattern, with maximum values in spring (2.28 ppbv and 0.50 µgm-3, respectively) and minimum values in autumn (1.03 ppbv and 0.17 µgm-3, respectively). The lowest seasonal mean concentrations were within the verified linear response range of the instrument, which covered the corresponding atmospheric-equivalent levels of the field observations. These values also exceeded the corresponding 3σ detection limits by factors of approximately 15 and 24 for ROS_g and ROS_p, respectively, confirming that the seasonal-average signals were analytically resolved rather than artifacts of instrumental blank noise or short-term instrumental drift. A broader comparison with previous field observations is provided in Fig. 6 and Table S3.

As summarized there, reported atmospheric ROS levels span a wide range across regions and seasons, although part of the gas-phase dataset used for comparison is derived from measurements of gaseous H2O2 alone. Within this observational context, the ROS_g levels measured in this study are generally at the upper end of previously reported urban observations, whereas ROS_p falls within the range of earlier measurements and remains at a moderate level relative to the most elevated reported values. Specifically, the seasonal-mean ROS_g (1.03–2.28 ppbv) exceeds most reported values for urban Beijing and several other Asian sites, while remaining comparable to observations at photochemically active continental locations. In contrast, the seasonal-mean ROS_p (0.17–0.50 µgm-3) is broadly comparable to previous urban measurements, but lower than the highest values reported for highly oxidized particulate environments. Overall, these comparisons indicate that the present observations were characterized by relatively elevated ROS_g but moderate ROS_p, highlighting a distinct phase-dependent distribution of atmospheric oxidative burden in urban Beijing.

In autumn, pollution was primarily characterized by PM_2.5 accumulation under humid and weakly dispersive conditions. Compared with clean days, polluted days exhibited a near threefold increase in PM_2.5 (94.88 vs. 33.45 µgm-3), accompanied by higher RH (78 % vs. 67 %) and lower wind speed (1.19 vs. 1.68 ms-1) (Table S2). These meteorological features favor pollutant accumulation and enhance multiphase processing, consistent with the elevated ROS_g from 0.89 ppbv on CDs to 1.34 ppbv on PDs. By contrast, ROS_p showed no corresponding increase (0.18 vs. 0.15 µgm-3), suggesting that particle-phase oxidative activity in autumn was not directly proportional to PM_2.5 mass, but was instead constrained by the competing effects of ROS formation and depletion under humid, stagnant conditions (Campbell et al., 2021).

During winter, ROS_g (1.22 ppbv) and ROS_p (0.26 µgm-3) exceeded autumn despite weak photolysis, consistent with the seasonally highest SO2 and strengthened combustion-related emissions. Enhanced residential coal burning for heating can increase primary oxidant inputs and elevate the fractions of black carbon and redox-active metals in PM_2.5. These components promote secondary inorganic production and provide abundant reactive surface area and condensed-phase microenvironments that facilitate heterogeneous and multiphase processing, thereby sustaining ROS_p and indirectly supporting ROS_g via gas-particle partitioning and multiphase recycling even under weak winter radiation (An et al., 2019; Song et al., 2024).

In spring, ROS_g increased from 2.10 ppbv on CDs to 3.02 ppbv on PDs, accompanied by higher O3 (60.60–104.74 ppbv) and j(O1D) (4.52×10-6–7.13×10-6 s-1), indicating an O3-driven photochemical regime. This seasonal maximum contrasts with previous rural Beijing observations, where spring ROS was lower than winter ROS under relatively clean, haze-free spring conditions, which limited haze-related precursor accumulation and radical formation (Huang et al., 2016). In the present urban observations, stronger spring photolysis and elevated O3 instead favored gas-phase ROS production, explaining why spring ROS exceeded winter levels. ROS_p remained comparably high on CDs and PDs (0.50 µgm-3), suggesting that, once photochemistry was sufficiently active, particle oxidative activity could be sustained despite lower aerosol loading through continued production and uptake of peroxides and other semi-volatile oxidants and in-particle transformation pathways (Huang et al., 2016; Zhou et al., 2019).

Summer featured the strongest photochemical environment, with O3 approaching 80.35 ppbv and j(O1D) reaching about 1.03×10-5 s-1. ROS_g increased modestly from 1.04 to 1.29 ppbv from CDs to PDs, while ROS_p showed a pronounced enhancement (0.27–0.46 µgm-3). High temperatures (∼302 K) enhance biogenic isoprene abundance (1.45 ppbv) and accelerate its oxidation kinetics, leading to increased ROx production (Wennberg et al., 2018). Meanwhile, isoprene-driven SOA formation supplies peroxide-rich and low-volatility products to particles, boosting ROS_p concentration and yielding a larger ROS_p response than ROS_g (Zhou et al., 2019; Enami, 2021; Kroll et al., 2006).

Across all seasons, the diurnal evolution of ROS reflected the combined influence of photochemical intensity, precursor availability, and boundary-layer dynamics, with clear contrasts between gas-phase and particle-phase behavior under clean (Fig. 7) and polluted conditions (Fig. 8). Overall, ROS_g was primarily regulated by daytime photochemical production and nighttime regeneration, whereas ROS_p integrated the cumulative effects of gas-particle partitioning of oxidized products and multiphase oxidation, leading to distinct phase-dependent diurnal responses. Figure 9 summarizes the covariation between ROS and key atmospheric species, with correlations of |r|<0.3 classified as weak associations.

From late morning to early afternoon (08:00–15:00 local time (LT)), photochemical processes dominated ROS evolution. ROS_g increased markedly with rising O3 and photolysis activity, reflecting rapid expansion of the daytime oxidative pool involving HOx-ROx radicals and peroxides (Stone et al., 2012; Nosaka and Nosaka, 2017; Liu et al., 2023). Under clean conditions, ROS_g and ROS_p exhibited distinct responses, particularly in autumn (Fig. 7a). ROS_g showed only a weak negative association with O3 in clean autumn air (r=-0.15) (Fig. 9), indicating that O3 alone did not explain ROS_g variability. The deviation between ROS_g and O3 is consistent with additional influences from titration and precursor limitation in relatively clean air masses. Meanwhile, reactions between O3 and alkenes may generate Criegee intermediates, promoting SOA formation and secondary ROS production (Chen et al., 2011; Yao et al., 2014). In contrast, ROS_p generally peaked earlier than ROS_g and showed a strong positive association with NO under clean autumn conditions (r=0.60) (Fig. 9), indicating close coupling with fresh NOx-influenced air masses. Under polluted conditions, daytime ROS_p was more clearly decoupled from bulk particle loading. In polluted autumn (Fig. 8a), ROS_p was only weakly associated with PM_2.5 (r=0.28), indicating that bulk particle mass alone could not account for ROS_p variability. This pattern points to a stronger influence of aerosol composition and secondary processing than by total particle mass, consistent with previous studies (Liu et al., 2023; Zhou et al., 2019; Huang et al., 2016). At the same time, the positive association between ROS_p and NO weakened relative to clean conditions (r=0.46) (Fig. 9), indicating a reduced influence of fresh emissions and a greater contribution from secondary formation processes.

During the late afternoon and early evening (15:00–20:00 LT), seasonal contrasts became more pronounced. In spring polluted conditions (Fig. 8b), ROS_g continued to increase until late afternoon, whereas ROS_p declined after its midday maximum. In summer, ROS_g generally peaked later in the day, whereas ROS_p peaked earlier under polluted conditions (Fig. 8c) but remained relatively stable under clean conditions (Fig. 7d). Under clean summer conditions (Fig. 9), ROS_p showed moderate negative correlations with NO (r=-0.40) and NO2 (r=-0.39), indicating suppression of ROS_p accumulation by fresh NOx, while ROS_g was positively correlated with NO (r=0.36) and negatively correlated with O3 (r=-0.22), reflecting competition between titration and photochemical regeneration (Zhou et al., 2019; He et al., 2022). In autumn, the delayed ROS_g maxima relative to O3, together with a weak positive association with PM_2.5 under polluted conditions (r=0.28) (Fig. 9), indicate that ROS_g variability was not controlled by photochemistry alone. Gas-particle coupling, boundary-layer evolution, and O3-alkene reactions may have jointly contributed to sustaining oxidative capacity after peak photochemistry (Wang et al., 2023b). In winter, declining ROS_g and weakly rebounding ROS_p in the evening, together with a weak positive association between ROS_p and PM_2.5 (r=0.22) (Fig. 9), are consistent with an increasing contribution from heterogeneous and multiphase processes as photochemical activity wanes (Xue et al., 2022).

From evening through early morning (20:00–08:00 LT), boundary-layer stabilization favored nocturnal accumulation and phase repartitioning, while oxidant production was increasingly regulated by nighttime chemistry and post-sunrise reactivation. In autumn polluted conditions (Fig. 8a), ROS_p typically reached an early-evening maximum and then remained elevated, whereas ROS_g declined after dusk but rebounded toward midnight; this nocturnal increase is plausibly linked to NO3-driven oxidation, peroxide decomposition, and ozonolysis of alkenes, which can sustain nighttime radical recycling and replenish the gas-phase oxidative pool (Venkatachari and Hopke, 2008). Under clean winter and spring conditions (Fig. 7b and c), ROS_g showed weak nocturnal variability, while ROS_p generally decreased in colder seasons but persisted or increased in summer despite weak PM_2.5 dependence (r=0.12) (Fig. 9), suggesting that bulk particle loading was not the primary driver of ROS_p variability. This behavior is compatible with continued multiphase ROS formation and aerosol aging under weak nocturnal photochemical forcing (Brown and Stutz, 2012; Wang et al., 2023a). After midnight, ROS_g in autumn continued to decrease toward sunrise under both clean (Fig. 7a) and polluted conditions (Fig. 8a), and under clean conditions it tracked NO (r=0.36) (Fig. 9), implying precursor-driven background control under weak radiation (King and Weber, 2013; Nan et al., 2017). Under clean winter conditions, ROS_g showed limited variability and only weak associations with NO and NO2 (r=0.06–0.11), while ROS_p also exhibited a weak positive association with PM_2.5 (r=0.22) (Fig. 9). Together with the nocturnal persistence of ROS_p, this pattern is consistent with a contribution from heterogeneous and aqueous-phase oxidation (Liu et al., 2023; Campbell et al., 2021). In spring and summer, ROS_g more frequently rebounded or increased in the early morning, particularly under polluted conditions, suggesting enhanced oxidative activity after sunrise.

Based on the time series (Fig. 10) and clustered back trajectories (Fig. S5), the three representative episodes can be classified into two contrasting ROS regimes: an aerosol water-driven multiphase regime in autumn and photochemically driven oxidation regimes in spring and summer. Across all episodes, ROS_g generally increased during pollution development, whereas ROS_p showed episode-dependent behavior. Aerosol liquid water content (ALWC) and particle pH estimated by ISORROPIA-II provide a consistent thermodynamic framework for interpreting the coupled evolution of aerosol water, acidity, and semi-volatile partitioning (Fountoukis and Nenes, 2007). The HYSPLIT back trajectories further suggest seasonally distinct transport patterns that modulated precursor supply and ventilation efficiency, thereby influencing ROS formation and removal (Stein et al., 2015; Draxler and Hess, 1998, 1997).

The autumn episode was characterized by PM_2.5-dominated pollution, with the highest ALWC and the strongest coupling of PM_2.5 with aerosol water and secondary inorganic aerosols, especially nitrate (Fig. 10). During this period, ROS_g increased modestly from 0.89 to 1.34 ppbv, whereas ROS_p slightly decreased from 0.18 to 0.15 µgm-3, and both remained lower than in spring and summer (Table S2). This indicates that the event was governed mainly by hygroscopic growth and nitrate accumulation, rather than by strong oxidative ROS production. High RH and ALWC favor heterogeneous N2O5 hydrolysis and HNO3 partitioning, promoting rapid nitrate build-up and aerosol mass enhancement (Liu et al., 2020; Zang et al., 2022). However, weak autumn photochemistry and efficient radical termination under high-NOx/HNO3 conditions likely limited peroxide formation and suppressed ROS_g accumulation (Tan et al., 2018; Ye et al., 2025). Meanwhile, elevated aerosol water may also accelerate multiphase peroxide loss and organic peroxide decomposition, preventing ROS_p from increasing despite substantial particle growth (Xuan et al., 2020). Thus, autumn represents a high-mass but weakly oxidative regime, where nitrate-water amplification outweighed ROS production. Trajectory clustering indicates that polluted days were dominated by westerly air masses (∼41 %), with additional southerly (∼28 %) and easterly (∼16 %) influence (Fig. S5a), consistent with regionally confined transport and limited ventilation over the Beijing-Hebei region. During the clean-up stage, the airflow shifted toward more ventilated northerly (∼43 %) and southwesterly (∼24 %) pathways (Fig. S5b), favoring pollutant dispersion and the decline of ROS.

The spring pollution episode was an O3-dominated event, with mean O3 reaching 104.74 ppbv, while ALWC and secondary inorganic aerosol levels were lower than in autumn (Fig. 10). In contrast to autumn, ROS_g increased markedly from 2.10 to 3.02 ppbv, whereas ROS_p remained relatively stable at about 0.50 µgm-3 (Table S2). This pattern indicates that spring pollution was controlled mainly by intensified gas-phase photochemistry rather than aerosol aqueous processing. High-O3 conditions generally reflect strong atmospheric oxidation capacity, sustained by O3 photolysis, HONO photolysis, and active HOx cycling, all of which favor the formation of peroxides and other ROS_g (Tan et al., 2019; Ye et al., 2025). The limited change in ROS_p indicates that ROS_p was less sensitive to oxidant abundance alone and more constrained by aerosol composition and SOA aging processes (Zhou et al., 2019). Therefore, spring can be interpreted as a photochemically active gas-phase oxidation regime, characterized by efficient ROS_g production but only moderate ROS_p enhancement. Spring trajectories were dominated by southerly to southeasterly transport (∼81 %), with minor northeasterly (∼9 %) and westerly (∼8 %) contributions (Fig. S5c), indicating persistent precursor import from the North China Plain. During the clean-up stage, the contribution of northwesterly and northerly inflow increased (∼24 % and ∼23 %), while the southern sector decreased to ∼28 % (Fig. S5d), consistent with weaker precursor supply, stronger ventilation, and the coherent decline of O3 and ROS.

Summer showed the clearest decoupling between aerosol mass and ROS behavior (Fig. 10). Although PM_2.5 varied only slightly (23.5–27.5 µgm-3), and both ALWC and secondary inorganic ions were the lowest of the three seasons, ROS_p increased substantially from 0.27 to 0.46 µgm-3, whereas ROS_g rose only slightly from 1.04 to 1.29 ppbv (Table S2). This pattern suggests that summertime ROS chemistry was controlled mainly by photochemical aging of organic aerosol rather than by bulk particle loading. Strong radiation and high temperature enhance VOCs oxidation and O3 formation, promoting the generation of more oxidized organic aerosol that can sustain elevated ROS_p even under relatively low PM and inorganic ion levels (Zhou et al., 2019; Zhu et al., 2020). More acidic summer aerosols may further facilitate acid-catalyzed SOA processing and associated ROS_p formation (Wei et al., 2022). In addition, intense solar radiation can activate photo-initiated heterogeneous pathways that further enhance particle oxidative activity. Laboratory studies have demonstrated that visible-light irradiation of soot microstructures can directly generate ROS on particle surfaces, providing an additional photochemical source of ROS_p in strongly illuminated environments (Zhu et al., 2021). By contrast, the modest increase in ROS_g indicates rapid turnover rather than substantial accumulation of ROS_g in summer. With strong photochemistry and relatively low PM_2.5, oxidation was manifested more in efficient O3 production and continued aerosol aging than in a pronounced enhancement of ROS_g. In this sense, summer represents an O3-rich and organic-oxidation-dominated regime, where aerosol composition, together with enhanced acidity, rather than aerosol mass, played the key role in determining ROS_p. Trajectory clusters show that polluted days were dominated by northerly transport (Fig. S5e), suggesting the influence of photochemically aged, O3-enriched air masses under weak NO titration. During the clean-up stage, southeasterly inflow became dominant, led by a maritime branch (∼32 %) and additional southeast pathways (∼23 %) (Fig. S5f), providing cleaner ventilation and driving the rapid decline of O3 and ROS.