In this study, we retrieved additional Antarctic measurements from two fully intercalibrated affiliated measurement programs. Details of these programs are summarized in Table 1 and Fig. 1. The global terrain model uses ETOPO2v2 developed by the National Geophysical Data Center (NGDC) affiliated with NOAA. Specifically, observational data were obtained from the NOAA network (https://gml.NOAA.gov/aftp/data; last access: 11 April 2025) and the NILU network (https://ebas-data.nilu.no/Default.aspx; last access: 11 April 2025). GC-MS was employed at all stations for analysis, ensuring consistency and comparability across measurements (Montzka et al., 1996, 2015; Montzka and Dutton, 2012; Platt and Lunder, 2024; Lunder, 2008–2024).

Palmer Station, Antarctica (PSA; 64.8° S, 64.1° W; 15 m a.s.l.), is located on Anvers Island, just outside the Antarctic Circle. At this station, air samples are collected in glass flasks at a biweekly frequency. Trollhaugen Station, Antarctica (TRL; 72° S, 2.5° E; 1553 m a.s.l.), operated by NILU, is located approximately 235 km inland from the coast in Jutulsessen, Dronning Maud Land. At Trollhaugen Station, ambient air samples are collected in stainless steel flasks at a biweekly frequency. In addition to these coastal stations, we collected air samples at the South Pole Observatory (SPO; 90° S, 24.8° W, 2837 m a.s.l.), which represents atmospheric conditions over the Antarctic interior. At SPO, paired-flask sampling is conducted at weekly to biweekly frequencies (Montzka et al., 2015; Mahesh et al., 2003).

In addition to data from Antarctic stations, we collected data at other latitudes. Midlatitude baseline measurements were obtained from the Cape Grim Observatory (CGO, 40.7° S, 144.7° E, 164 m a.s.l.) in Tasmania, Australia, which is a global baseline station. The station is located at the northwestern corner of Tasmania and is subject to mainly westerly winds, rendering it ideal for sampling unpolluted air, as the sampled air arrives at Cape Grim after traveling long distances over the Southern Ocean, under conditions described as baseline conditions (Crawford, et al., 2017). Those samples were shipped back to the Boulder laboratory, where they were analyzed via at least 3 GC systems. At the Zeppelin Station (ZEP, 78.9° N, 11.9° W, 474 m a.s.l.), operated by NILU, observations are made in situ using GC–MS instruments and are thus available at a much higher temporal frequency (2 h). The data from ZEP and CGO were averaged to monthly means.

There is a calibration scale difference among AGAGE, NOAA, and NILU observations. However, Prinn et al. (2018) noted that the difference between the NOAA-2016 and SIO-05 scales is smaller than 1 %, and material compatibility tests confirm negligible biases between samples collected in glass and stainless steel containers (Wu et al., 2001). More detailed information on the observations is provided in Table 1.

Air mass backward trajectories were computed using the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model (Stein et al., 2015). HYSPLIT is a complete system for computing simple air parcel trajectories, as well as complex transport, dispersion, chemical transformation, and deposition simulations. The model is widely used to analyze the transport and diffusion of atmospheric pollutants, gases, aerosols, and dust by calculating backward air quality trajectories at various altitudes and times using meteorological data (Ding et al., 2020; Fan et al., 2021; Chen et al., 2023). The model calculation method is a hybrid between the Lagrangian approach and the Eulerian methodology. In the Lagrangian model, air concentrations are obtained by summing virtual air parcels of zero volume, which are advected through grid cells along its trajectory (Escudero et al., 2006; Draxler and Hess, 1998). In the Eulerian model, air concentrations are calculated via the integration of mass fluxes in each grid cell on the basis of their diffusion, advection, and local processes. Gridded meteorological data were sourced from the National Centers for Environmental Prediction (NCEP), namely, from the Global Data Assimilation System (GDAS1, 1° horizontal resolution), which is operated by NOAA and provides 23 vertical levels, from 1000–20 hPa (Draxler and Hess, 1998; Stein et al., 2015) (https://www.ready.noaa.gov/data/archives/gdas1/, last access: 13 April 2025). In this study, the TrajStat tool (MeteoInfoMap plugin for air mass trajectory statistics; Wang, 2014) was used to drive the model.

Because most F-gases are long-lived gases (Table S1), their properties in the atmosphere remain relatively stable. Therefore, they can be transported with air masses across long distances. According to the above characteristics and model operation, the backtracking time was set to 720 h (30 d), and the temporal resolution is 1 h. The boundary layer top in the Antarctic region is low, so the above ground level (a.g.l.) was set to 500 m to represent the top of the boundary layer or the lower portion of the free atmosphere in the area. This height typically reflects the features of air mass movements from regional and long-range sources rather than the direct influence of local sources close to the surface. After calculating backward trajectories over 12 months of 2021, a combined backward trajectory for the entire year was obtained, yielding 8760 trajectories for each station.

To analyze the composition of the backward trajectories of air masses at the various sites and their proportions, trajectory clustering analysis was performed. Trajectories with similar directions and rates were merged and classified to represent a class of air mass transport paths. Ward's least variance method was employed for clustering (Stein et al., 2015; Wang et al., 2015). The basic idea is to classify trajectories into n classes and then reduce each class. It is necessary to satisfy the sum of the Euclidean distances of the pairs of trajectories and the minimum gap between the original distances. Ultimately, each trajectory is reduced to the appropriate class. The Euclidean distance can be calculated as follows: 2d12=∑i=1n(X1i-X2i)2+(Y1i-Y2i)2 where d12 is the Euclidean distance between traces 1 and 2; X1i and Y1i, X2i and Y2i are the positions of point i on trajectory 1 and 2, respectively; and n is the number of trajectory points after 720 h.

To determine a reasonable number of clusters, the total spatial variance (TSV) method (Guan et al., 2019) is generally applied, and the TSV is used to obtain the number of clusters. There is high consistency in TSV-based selection among the four Antarctic stations, all of which exhibit 3 clusters.

To better analyze the anthropogenic source of F-gases at ZOS, a positive matrix factorization (PMF) model was used for source analysis. On the basis of a matrix consisting of the concentrations of diverse chemical species, the objective of the PMF model is to determine the number of F-gas source factors, the chemical composition profile of each factor, and the contribution of each factor to species (Paatero and Tapper, 1994). The PMF model aims to decompose the data matrix X(n×m) into G(n×p), F(p×m) and E(n×m) matrices. The matrix representation is as follows: 3X(n×m)=G(n×p)×F(p×m)+E(n×m) where X(n×m) is the matrix of chemical concentrations measured at the receptor station; G(n×p) denotes the factor contribution matrix; F(p×m) is the factor spectrum matrix; E(n×m) is the residual matrix; and n is the number of samples. m chemical composition species; p is the number of factors (sources of pollution) resolved.

Element eij of the residual matrix E(n×m) is the residual value not explained by the model data value xij. Notably, xij can be calculated as follows: 4xij=∑k=1pgikfkj+eij where xij is the concentration of the jth species in the ith sample; gik is the kth source for the ith sample; fkj is the concentration of the jth element from the kth source; and eij is the residual concentration of the jth species in the ith sample.

The objective function Q is weighted via the least squares method, using residuals and uncertainty (Reff et al., 2007): 5Q(E)=∑i=1n∑j=1meijuij2 where n and m are the numbers of samples and species, respectively, and uij is the uncertainty determined by the detection limit for each species.

The uncertainty can be calculated as follows: 6uij=56×MDLj,C≤MDL(MU×Cij)2+(0.5×MDLj)2,C>MDL where MU is the error fraction, usually between 5 % and 20 %, which is 10 % in this study (Wu et al., 2025); Cij is the concentration of the jth species in the ith sample; and MDLj is the minimum detection limit for the jth species. The measured concentrations of F-gases in this paper are all greater than the MDL of the instrument.

The PMF model is supported by U.S. EPA PMF software version 5.0. The specific process includes data preparation, data input and preliminary inspection, basic calculation, rotation calculation, source analysis result calculation, and assessment of the results.

Backward trajectory analysis via the HYSPLIT model revealed distinct atmospheric transport patterns to the four Antarctic stations (Fig. 4). On the basis of cluster analysis, the air masses arriving at ZOS were categorized into three dominant pathways, with relative contributions of 51.83 %, 37.14 %, and 11.03 %. All the clusters originated from the ocean 30 d prior, forming an approximately circular path due to zonal westerly wind circulation. The primary pathway (88.97 %) exhibited landfall over Princess Elizabeth Land along the Indian Ocean coast before it reached ZOS via westerly advection. Only 11.03 % of the air masses made landfall over East Antarctic at 145° E, moving into Wilkes Land before migrating east of ZOS. Vertical profiling revealed that air parcels initiated at ∼750hPa over oceanic regions experienced orographic uplift followed by katabatic acceleration during eastern ingress into the ZOS area.

TRL exhibited comparable trajectory clustering (46.51 %, 36.03 %, and 17.46 %, respectively) with similar recirculation characteristics, although it was distinguished by landfall dynamics. The dominant cluster (46.51 %) made primary contact with the Antarctic continent in the Atlantic sector (30° E), whereas the secondary cluster (36.03 %) originated from the southern Indian Ocean coast of Enderby Land. Notably, the tertiary trajectory (17.46 %) encompassed a complex transcontinental loop: after initial landfall in Oates Land, it traversed Wilkes Land, re-entered the Indian Ocean, and completed a circumpolar transit before reaching TRL. The vertical displacement patterns revealed gradual subsidence interrupted by intermittent orographic lifting events.

In contrast to ZOS and TRL, SPO demonstrated bifurcated oceanic influences with cluster proportions of 49.03 % (Pacific–Atlantic–SPO), 34.23 % (Indian Ocean–Oates Land–SPO), and 16.75 % (30° E–Antarctic Peninsula–SPO). All trajectories maintained marine origins but diverged in terms of hemispheric routing – approximately half originated from Pacific sectors versus Indian Ocean sources. Vertical development initiated from 700–750 hPa, with notable uplift to 600 hPa through terrain forcing before final subsidence over SPO.

The clustering configurations (40.43 %, 32.89 %, and 26.68 %) at PSA revealed an exclusive Indian Ocean provenance at 750 hPa. Westerly circulation drove progressive subsidence along a west–east axis across the Pacific before ingress into the PSA area. This marine-dominated transport regime, characterized by elevated humidity and enhanced pollutant dispersion capacity levels (Zheng et al., 2023; Wu et al., 2025), is correlated with the observed reduced contaminant concentrations at PSA relative to those at ZOS and TRL. This partly explains the lower concentration at PSA than those at ZOS and TRL.

While backward trajectory analysis can reveal synoptic transport mechanisms, it is not suitable for identifying specific emission sources. All 30 d backward trajectories remained embedded within the Antarctic circumpolar circulation, precluding clear terrestrial source attribution. This suggests that a more detailed atmospheric chemical transport model should be employed to trace and quantify emission sources, as this can provide a clearer understanding of the emission flux and its transport dynamics in the region. Cluster analysis revealed that the concentrations of F-gases from the three air masses transmitted to ZOS were consistent (Table S4 in the Supplement), as were those at TRL, SPO, and PSA. This suggests that the atmospheric background feature of uniform Antarctic mixing may be observed at all four of the Antarctic locations included here.

The potential sources of F-gases and their relative contributions to each category were determined via the PMF model. HFC-236fa, CF4, and PFC-116 were not included because of their low correlation coefficients during concentration fitting, which could affect the accuracy of the model. F-gases were derived from a variety of sources, as detailed in Table 3. Four isolation factors were extracted according to the composition profiles depicted in Fig. 5. The percentages of factors 1–4 were 24.69 %, 23.28 %, 29.95 %, and 22.08 %. Residual analysis of the PMF results was conducted, and the results are shown in Fig. S3 in the Supplement. The histogram of the frequency distribution of the residuals exhibits a symmetrical bell shape. The mean value is close to 0 (0.00079), indicating that there is no significant deviation in the fit of the model to the observed data. The residuals are mainly distributed within the (-0.5∼0.5) range, and the standard deviation is σ=0.12, indicating that the fitting error is small and concentrated. The Gaussian fitting R2 value is 0.99323, indicating that the residual distribution is highly consistent with the normal distribution.

In factor 1, HFC-227ea and HFC-23 exhibited relatively high loadings (53.59 % and 38.02 %, respectively). HFC-227ea is a fire retardant that has replaced halon-1301 and is also used as a propellant in metered-dose inhalers (MDIs) (UNEP, 2022; Vollmer et al., 2011). HFC-23 is used in halon-1301 (CF3Br) production, very low-temperature refrigeration, and specialty fire extinguishers (Oram et al., 1998; Miller et al., 2010, Simmonds et al., 2018), with very low emissions from deliberate use (Takeda et al., 2021). Hence, HFC-23 has mainly been discharged from HCFC-22 production plants into the atmosphere (Montzka et al., 2019). Therefore, factor 1 can be assigned as a potential source of fire extinguishing agents and industrial byproducts.

In factor 2, HFC-32 and HFC-152a attained relatively high loadings (36.7 % and 32.17 %, respectively). HFC-32 is a replacement for HCFC-22 and is applied mainly in the fields of air conditioning and refrigeration. For example, the refrigerant blends R-410A (50 % HFC-32 % and 50 % HFC-125 by weight) and R-407C (23 % HFC-32, 52 % HFC-134a and 25 % HFC-125 by weight) are azeotropes used in stationary air-conditioning systems (Velders et al., 2022). HFC-152a is used mainly as a foaming agent in aerosol inhalers. Therefore, factor 2 can be assigned as a potential source of refrigeration and foaming agents.

In factor 3, HFC-23 and HFC-134a exhibited relatively high loadings (44.61 % and 34.99 %, respectively). HFC-23 exhibits the second-highest radiative forcing among all HFCs and fluorinated gases just after HFC-134a (14.3 mWm-2) (Montzka et al., 2019). HFC-143a is used mostly as a working fluid component in refrigerant blends R404A (52 % HFC-143a, 44 % HFC-125 % and 4 % HFC-134a by weight) and R-507A (an azeotropic blend of 50 % HFC-125 % and 50 % HFC-143a) for low- and medium-temperature commercial refrigeration systems, thereby replacing R-502 (a blend containing HCFC-22 and CFC-115) (Velders et al., 2009). Other minor uses are R-408A, R427A, and R-428A as replacements for R-502 and HCFC-22. HFC-134a is the preferred substitute for CFC-12, primarily in refrigeration and air-conditioning applications (O'Doherty et al., 2014). Therefore, factor 3 can be assigned as a potential source of industrial byproducts and refrigeration gases.

In factor 4, HFC-152a, HFC-32, and NF3 attained relatively high loadings (67.82 %, 63.29 %, and 40.02 %, respectively). NF3 has been used in the semiconductor industry and in the production of photovoltaic cells and flat-panel displays as an etching agent (Arnold et al., 2018). Therefore, factor 4 can be assigned as a potential source of foaming, refrigeration, and electronics products.