The daily PM₁₀ sampling was performed every third day from 2 January 2013 to 28 May 2023, on 150 mm-diameter quartz fibre filter (Tissu-quartz PALL QAT-UP 2500 diameter 150 mm) using high-volume samplers (Digitel DA80, 30 m³ h⁻¹). A weekly PM₁₀ sampling was conducted during the same period using a low-volume sampler (Partisol, 1 m³ h⁻¹) onto 47 mm diameter quartz fibre filters (Tissuquartz PALL QAT-UP 2500 diameter 47 mm). The processes of preparation, handling, and storing filters, in order to guarantee optimum quality for chemical analyses were presented in Borlaza et al. (2021a). Field blank filters were also collected (about 8 %–10 % of the total samples) to estimate the detection limits and evaluate the filter contamination during the overall handling and analysis processes.

The daily PM₁₀ samples (n=1570) and field blanks were analyzed for elemental carbon (EC) and organic carbon (OC), major ions (Cl⁻, NO3-, SO42-, Na⁺, NH4+, K⁺, Mg²⁺, Ca²⁺), methanesulfonic acid (MSA), anhydrous sugar and saccharides (levoglucosan, mannosan, arabitol, mannitol), and trace elements (As, Ba, Cd, Cr, Cu, Mn, Ni, Pb, Sb, V, Zn). However, the concentrations of the daily trace elements were analyzed only in 3 periods, including: (1) from 2 January 2013 to 31 December 2013 (n=122), (2) from 28 February 2017 to 13 March 2018 (n=125), (3) from 30 June 2020 to 18 June 2021 (n=115). The weekly samples and blanks were analyzed for trace metal concentrations for the whole sampling period (n=842).

All analyses were previously described in detail (Borlaza et al., 2021a). In brief, EC and OC analysis was performed using a Sunset Lab analyser with the EUSAAR2 thermo-optical protocol. The eight major ionic components and MSA were analyzed, after aqueous extraction of the filters using orbital shacking, by ionic chromatography using an ICS3000 dual-channel chromatograph (Thermo-Fisher) with a CS16 column for cation analysis and an AS11 HC column for anion analysis. The anhydrous-sugar and saccharides analyses were performed on the same water extract by an HPLC method using PAD (Pulsed Amperometric Detection) (model Dionex DX500 + ED40) with Metrosep columns (Carb 1-Guard+A Supp15-150+Carb1-150) in the period before the year 2017. From 2017 to the present, the measurement with ICS 5000 with pulsed amperometric detection (HPLC-PAD) was performed following the CEN method (European committee for standardization, 2024). The analysis is isocratic with 15 % eluent of sodium hydroxide (200 mM), sodium acetate (4 mM), and 85 % water at 1 mL min⁻¹.

The daily and weekly metals were measured by Inductively coupled plasma mass spectroscopy (ICP-MS) (ELAN 6100 DRC II PerkinElmer or NEXION PerkinElmer). The measurement was performed on the mineralization of a 38 mm diameter punch of each filter, using 5 mL of HNO3 (70 %) and 1.25 mL of H₂O₂ for 30 min at 180 °C in a microwave.

Two complementary OP assays, including the two probes ascorbic acid (AA) and dithiothreitol (DTT) were performed on the same filters with PM₁₀ components analysis (from 2 January 2013 to 28 May 2023, n=1570). Filter samples are extracted using a simulated lung fluid which is the mixing of Gamble and DPPC (dipalmitoylphosphatidylcholine) solutions, during 1 h 15 min at 37 °C, pH 7.4 , creating a physiological environment for the extraction (Calas et al., 2017) . The AA method quantifies the consumption of ascorbic acid, an endogenous antioxidant in the lung, by PM and was described in Calas et al. (2017, 2018). The reaction mixture (extract + AA) was transferred to UV-transparent 96-well plates (CELLSTAR, Greiner-Bio), and the residual AA was measured at 265 nm with a TECAN Infinite M200 Pro spectrophotometer. The AA consumption rate (nmolmin-1) reflects the capacity of PM₁₀ to catalyze electron transfer from AA to molecular oxygen.

DTT assay relies on dithiothreitol, a chemical surrogate for cellular reducing agents, allowing for emulation of in vivo interaction among PM₁₀ and biological reducing agents (for example, nicotinamide adenine dinucleotide (NADH), nicotinamide adenine dinucleotide phosphate oxidase (NADPH)). After incubation of the PM suspension within the lining fluid with DTT, the remaining DTT was titrated with 5,5^′-dithiobis-(2-nitrobenzoic acid) (DTNB) to form 5-mercapto-2-nitrobenzoic acid (TNB). Absorbance at 412 nm (TECAN Infinite M200 Pro) in 96-well plates provided the concentration of unconsumed DTT, from which the DTT consumption rate (nmolmin-1) was calculated. The batches were standardized using a common control (lab's rooftop filter analysis for every batch) to ensure consistency between batches.

After analysis, the OP activities were blank subtracted and then normalized using the PM₁₀ mass concentration and the sampling air volumes. The mass-normalized OP (OP^m, nmolmin-1µg-1) represents the intrinsic OP of 1 µg PM, while the volume-normalized OP (OP^v, nmol min⁻¹ m⁻³) represents PM-derived OP per m³ of air. Each sample is analyzed in triplicate for AA and triplicate for DTT, respectively. Consequently, the OP values presented in the study are the mean and the standard deviation of these replicates.

Vertical temperature and humidity were measured every 30 min from November 2017 to May 2023 using Tinytag TGP-4500 from Gemini Data Loggers. A Stevenson Type Screen protects each Tinytag loggers from radiant heat (direct sunlight). Sensors are installed at a minimum of 3 m from the ground. The measurements have been performed at different elevations of the Bastille hill, located a few hundred meters from the city center (5°43^′37.0^′′ E, 45°11^′40.8^′′ N), including z=230, 309, 496, 916 m altitudes.

Further, measurement of the PM₁₀ mass was conducted (hourly) using tapered element oscillating microbalances equipped with filter dynamics measurement systems (TEOM-FDMS) at the same site as the filter collection. The PM concentration used in this study is the 24 h average concentration, which is associated with the days of filter-based sample measurement (from 2 January 2013 to 28 May 2023).

EPA PMF 5.0 (Norris et al., 2014) was used to identify and quantify the PM₁₀ sources based on the observed concentrations and their related uncertainties. The concept of PMF is to use the weighted least square fit method and apply a non-negative constraint with the weight calculated based on analysis uncertainties (Paatero and Tappert, 1994) (Eq. S1, Sect. S2). In this study, the input matrix of the PMF comprises 25 chemical species, including PM₁₀ (set as the total variable), carbonaceous fractions (OC*, EC), ions (Cl⁻, NO3-, SO42-, Na⁺, NH4+, K⁺, Mg²⁺, Ca²⁺), organic tracers (MSA, levoglucosan, mannosan, polyols) and trace metals (As, Ba, Cd, Cr, Cu, Ni, Pb, Sb, V, Zn). The trace metals were the daily measured metals in some periods (2013, 2017–2018, 2020–2021) and the daily imputed metals. The OC* (= OC minus the sum of the carbon mass from the organic tracers used in the input variables) was used to avoid considering twice the mass of C atoms in organic markers. Polyols were calculated as the sum of arabitol and mannitol, supposing that their origin is similar (Samaké et al., 2019b). The input uncertainties were calculated based on the concentrations and the uncertainties in the analysis (Gianini et al., 2012; Waked et al., 2014). Details on the calculation of OC* and uncertainties of PMF input are presented in Sect. S3 in the Supplement. The selection of the input variables is guided by our previous yearly PMF studies at this site (Borlaza et al., 2021a; Srivastava et al., 2018; Weber et al., 2019).

The application of PMF constraints is recommended in the European guide on air pollution source apportionment with receptor models (Belis et al., 2014) to avoid mixing between some factors and reduce the uncertainty of the rotational ambiguity. The constraints used in this study are also based on the previous PMF studies in Grenoble (Borlaza et al., 2021a; Srivastava et al., 2018; Weber et al., 2019) and are detailed in Table S3.

Several solutions, including those from 4 to 11 factors, were investigated to determine the optimal output. This selection is based on the ratio of Qtrue/Qrobust (evaluating the outlier's effect), the clarity of the chemical profile, the contribution of factors to PM₁₀, the correlation between measured and predicted concentration, and the stability of the solution. This stability was evaluated using the bootstrapping (BS) and displacement (DISP) methods. BS analysis randomly resamples the data observation matrix and uses it to run a new PMF. The base-run and boot-run factors are matched if their correlation exceeds the threshold (generally chosen at 0.6). DISP estimates each species' uncertainty in the factor profile by fitting the model many times until this variable turns displaced (upper or lower) from its fitted value. The details of the set criteria for validation are presented in Sect. S4.

To evaluate the stability of the PMF solution over time (including possible changes in the chemical profiles of the sources), we also implemented separated PMF SA for every successive period of 3 years (2013–2016, 2017–2020, 2021–2023) and then we investigated the homogeneity of the chemical profiles by using the Pearson distance (PD) and standardized identity distance (SID) metrics (Belis et al., 2015). The chemical profiles of PMF solutions every 3 years and 11 years, and those published in Borlaza et al. (2021a) are compared to assess the homogeneity of the chemical profiles.

Using a unique chemical profile for each of the sources for such a long-term period can potentially limit the assessment of its evolution (Borlaza et al., 2022). To evaluate such a phenomenon in our case, we investigated the chemical profile and contribution of PM₁₀ sources for three distinct periods (2013–2016, 2017–2020, 2021–2023) and compared the results with those for the full 11-year period, as well as to the results presented in Borlaza et al. (2021a) for the year 2017. Particularly, we checked the similarity of the chemical profiles of these PMF solutions using PD and SID metrics (Belis et al., 2015).

For each SA, the PMF solution was tested from 4 to 11 factors and validated by the criteria presented in Sect. S4. The results of these validations (Qtrue/Qrobust, bootstrap run, displacement run, and statistical validation) are presented in Sect. S5, Tables S4, S5 and S6. The runs of 4 to 9 factors returned at least one merging factor, and the solution with 11 factors led to a factor without geochemical identity. Finally, for each PMF tested (11 years, 2013–2016, 2017–2021, 2022–2023), the best solution includes 10 PM₁₀ sources, with mineral dust, sulfate-rich, primary traffic, biomass burning, primary biogenic, nitrate-rich, MSA-rich, aged sea salt, industrial, and chloride-rich.

The similarity of the chemical profiles is presented in Fig. 3. Most of the factors (i.e., aged sea salt, mineral dust, primary biogenic, biomass burning, primary traffic, industrial, nitrate-rich, and sulfate-rich) present quite homogenous chemical profiles over the 3 successive periods, indicating that these source profiles are quite stable during the full 11-year period and similar compared to sources reported in Borlaza et al. (2021a). The MSA-rich and chloride-rich sources are the most divergent but are still within the limit of the accepted PD and SID range; however, their standard deviations for PD are slightly higher than for the other sources (Fig. 3). This is due to differences in the contributions of SO42- in the chemical profile of MSA-rich, which varied from 6 % to 17 %, and that of Cl⁻ (73 %–83 %) in the chloride-rich factor. In a previous study, Weber et al. (2019) also reported that the proportion of SO42- in the MSA-rich source can significantly vary across French sites, from 6 % to 24 %. The chloride-rich source in this study (previously named sea/road salt in Borlaza et al. (2021a) is essentially composed of a high proportion of Cl⁻, with less than 10 % of Na⁺ and some metals (Cu, Mn, Ni, V). This source is detected in other alpine valley environments (Glojek et al., 2024), with a similar temporal evolution as here. Since chloride depletion from the particulate phase can greatly depend on solar radiation, relative humidity, and temperature, the chemical profile of this factor can vary on different time scales. This source was also observed to be heterogeneous in the three neighboring sites investigated within 15 km in the previous study in Grenoble (Borlaza et al., 2021a). Nevertheless, it should be noted that it represents only a very minor fraction of the PM₁₀ total mass (about 1 %).

With these stabilities of the chemical profiles over the years, the solution for the 11-year SA is considered suitable for further data analyses in this paper. In the next section (Sect. 3.3.2), we investigate how the contribution of these sources to total PM₁₀ loadings changed over time.

As presented in Fig. 4, the optimal PMF solution for the 11-year time series identifies 10 PM₁₀ sources, with the contributions of mineral dust (20.9 %), sulfate-rich (19.7 %), traffic (16.0 %), biomass burning (13.5 %), primary biogenic (10.7 %), nitrate-rich (7.2 %), MSA-rich (6.2 %), industrial (2.2 %), aged sea salt (2.5 %), and chloride-rich (1.0 %). The chemical profile and contribution of each source are shown in Figs. S3 and S4, respectively. Even though the chemical profiles are homogenous, the contributions of these sources show minor differences from those reported for this same site by Borlaza et al. (2021a) and Srivastava et al. (2018), partly because of the differences in the respective periods of the studies. However, the main sources are similar, i.e., SIA (nitrate and sulfate-rich), mineral dust, biomass burning, and primary traffic. Similar general results are also presented for Swiss Alpine (Ducret-stich and Tsai, 2013), French Alpine (Weber et al., 2018), and Slovenian Alpine areas (Glojek et al., 2024), showing biomass burning and secondary inorganic aerosols being the most abundant contributions to PM mass. Primary biogenic and MSA-rich sources are the biogenic sources rarely reported in the literature; however, they account together for 17 % of total PM₁₀ mass on average in this study, which is in line with those reported in urban background sites in France (Samaké et al., 2019a; Weber et al., 2019). The absolute PM₁₀ source contributions are also compared to the average annual concentration of PM₁₀ mass to demonstrate the ability of the PMF model to reconstruct the PM₁₀ mass. The difference between observed and reconstructed PM₁₀ concentrations on the 11-year average is about 1 µgm-3 (5 %), with no more than 2 µgm-3 for any single year, demonstrating that the PMF model performs well at reconstructing the PM₁₀ concentrations.

Significant trends in source contributions over this 11-year period are detected (and discussed in Sect. 3.3); nevertheless, the main contributors to the total PM₁₀ mass do not change, with mineral dust, biomass burning, sulphate-rich, nitrate-rich, and primary traffic being the main contributors to PM₁₀. The highest PM₁₀ concentrations (observed in winter/spring 2013 and 2015) are associated with the highest contribution of SIA and biomass burning sources. On the other hand, the relative contribution of SIA and biomass burning showed a negligible difference (varied from 0.3 % to 4 %) between these years compared to 2014 and 2016 (Fig. S5). The lowest PM₁₀ annual concentration was detected in 2021, notably when the third COVID-19 pandemic lockdown restrictions applied in France. In addition, the relative contributions (see Fig. S5) showed only small changes compared to those in other years, with an increasing contribution of primary biogenic sources in 2021 (4 % compared to 2020), and only a very light decrease in the anthropogenic sources.

The decrease in PM₁₀ annual average concentrations observed since 2017 is associated with decreases in the contribution of some of the anthropogenic PM₁₀ sources. However, using yearly averages for trend analysis may prevent a proper understanding of the variation in time and of the estimation of the trends based on monthly averages, which might be more informative, as discussed in Sect. 3.3.

The source contribution trend analysis was achieved through STL deconvolution (see Sect. 2.6). These trends for all sources over the full period of the study are presented in Table 1. In this table, the part labeled “Rest” represents the difference between the total PM₁₀ measured mass and the sum of the mass of all PMF-derived factors in order to assess any trend of the unresolved part of PM₁₀ within our SA study.

PM₁₀ concentrations present a downward trend from 2013 to 2023, with an average diminution of 0.73 µgm-3yr-1 (3.9 %) (Sect. S6, Fig. S6). Such a downward trend of PM₁₀ in Grenoble is in line with that observed in other urban sites in Europe (Aas et al., 2024; Borlaza et al., 2022; Caporale et al., 2021; Colette et al., 2020; Gama et al., 2018; Li et al., 2018; Pandolfi et al., 2016).

The reduction of PM₁₀ in Grenoble during this period is significantly larger than that in 30 rural sites of the European Monitoring and Evaluation Programme (EMEP) from 2000 to 2017, which show reductions of PM₁₀ from -1.5 % to -2.5 % (-0.008 to -0.58 µgm-3) (Colette et al., 2020). However, the results of this study are highly coherent with results from Aas et al. (2024), presenting a reduction of PM₁₀ in 2 rural sites in France (La Tardière and Revin) of -3.5 % yr⁻¹ between 2005 and 2019. The reduction of PM in this Grenoble site, as an urban site, being higher than those at the rural sites, is due to the changes in specific emission activities at the site. While in the rural sites, the PM emission are influenced by long range transport activities, the PM at the urban site is usually largely impacted by different local activities (Borlaza et al., 2022). Further, France is amongst the EU countries with the highest reduction trend, as presented by Aas et al. (2024).

The anthropogenic sources, such as primary traffic, sulfate-rich, and biomass burning, display the highest decrease between 2013 and 2023 in Grenoble, with a reduction of 12.9 %, 6.9 %, and 5.5 % (0.37, 0.25, and 0.13 µgm-3yr-1), respectively. The other anthropogenic sources also present significant decreasing trends; however, these trends are much lower (nitrate-rich: -0.11 µgm-3yr-1, industrial: -0.02 µgm-3yr-1). The downward trends of these anthropogenic sources (mainly traffic, SIA, and industrial) were also underlined for other European urban sites (Colette et al., 2020; Diapouli et al., 2017; Pandolfi et al., 2016) with various approaches. For instance, a similar approach using PMF (albeit without organic markers) was followed by Pandolfi et al. (2016), investigating the Mann-Kendall trend of PMF-derived sources, and reported an almost equivalent downward trend of the sulfate-rich factor of 53 % (i.e., 0.53 % yr⁻¹) between 2004 and 2014 in Spain. The decreasing trends of primary traffic, domestic biomass burning, and industrial emissions are potentially influenced by the reduction in primary emissions due to various abatement strategies (as discussed in the following subsections, notably in Sect. 3.3.3 and 3.3.4).

Conversely, natural sources such as mineral dust and chloride-rich factors do not show any significant trend or follow a very weak one (aged sea salt, primary biogenic). MSA-rich is the only source that displays a significant upward trend, with an increase of 0.08 µgm-3yr-1; further studies would be needed to relate this last increase to changes in precursor emissions or reactivity during transport. Finally, the low evolutions in the contributions of the natural sources demonstrate that the reduction in PM₁₀ in Grenoble is essentially related to the reduction of anthropogenic activities, especially sources related to traffic and domestic biomass burning activities.

The STL deconvolution is inherently constructed to separate the yearly and seasonal variations from the long-term trends. While we discuss the long-term trends of the sources in other sections (Sect. 3.3.1, 3.3.3, and 3.3.4), it is also interesting to evaluate the impact of the meteorology on the seasonal variations of the concentrations. It is well known that inversion layers in the lower atmosphere are extremely important for the modulation of the concentrations at the ground, particularly in the context of Alpine valleys during winter (Carbone et al., 2010; Glojek et al., 2022). In this section, we tried to better evaluate these impacts on the concentrations from the sources of PM in the case of our time series.

This was considered with the measurements of temperature along the slopes of the mountains very close to the city center (as described in Sect. 2.2.3), for the winter periods of 2017–2023. It has been previously shown by Allard et al. (2019) that such measurements are representative of the temperature in the valley, despite the potential influence of wind slopes. We particularly considered the temperature gradient over the first 700 m above ground and the number of days with persistent inversion, as defined in Sect. 2.2.3.

The analysis of the relationship between the PM₁₀ and bulk temperature vertical gradients (ΔT/Δz) in winter (November, December, January, February, March), summer (May, June, July, August), and transition season (remaining months) reveals that thermal inversion events and high PM₁₀ concentration are mainly occurring in winter time (Sect. S7, Fig. S8) during the 5 years of the study. Periods of persistent temperature inversion were assessed based on the condition in Eq. (1), which detected 79 persistent inversion days in series from 4 to 22 consecutive days, for the winter periods 2017–2023. A meaningful correlation is obtained between PM₁₀ concentrations and bulk temperature vertical gradient (r reaching 0.60, p≪0.001) for these winter months and even better when considering only the persistent inversion periods (r reaching 0.67, p≪0.001) for individual years (Table S7).

The distribution between the daily PM₁₀ concentration and daily average ΔT/Δz in winter months revealed that the majority of PM₁₀ concentration peaks (in excess of 40 µgm-3) occur during the persistent inversion days (Fig. S9). However, it also shows that a few high PM₁₀ concentrations could be found on the days without persistent inversion; meanwhile, the days with persistent inversion do not always have high PM₁₀ concentrations. This result is not surprising since the concentration of PM₁₀ is not only associated with thermal inversion events but also depends on other meteorological conditions (precipitation, heat deficit) and the variation of pollutant emissions (Carbone et al., 2010; Largeron and Staquet, 2016).

Interestingly, the impact of persistent inversion days on PM₁₀ concentrations from the residential biomass burning source is larger than that for other sources or total PM₁₀ (Fig. 5), with a higher correlation (0.63). In addition, the contribution of this source is systematically lower during non-inversion days, and large concentrations are essentially made during persistent days. The large impact of the inversions on the local sources is confirmed when comparing the source contribution of the inversion days vs. non-inversion days (Fig. 6). This figure shows both the large increase in average PM₁₀ concentrations and also the contributions of the local sources (emissions from residential biomass burning, traffic, industries, mineral dust probably from resuspension) in the cases of inversion days during winter. Conversely, long-range transport sources (sulfate-rich, nitrate-rich) tend to be less important during these inversion days. A similar pattern is observed for the relative contribution of sources to PM (Fig. S10), in which the significant contribution of biomass burning, dust, industrial, and primary traffic is detected during inversion events. The trends of the two most important local anthropogenic sources (domestic biomass burning and traffic) are further discussed in the next sections.

The trend of the domestic biomass burning PM₁₀ concentrations is investigated via an STL decomposition analysis on this PMF-derived source (Fig. 7), indicating a statistically significant decreasing trend from 2013 to 2023 (p values ≪0.01). The seasonal estimate shows the highest values in the winter season (November, December, January), with a visual trend to a smoothing of the peak concentrations; conversely, from Mar to Sept, the seasonal variations showed constantly lowest values. Extreme residual values were detected in the winter months of 2016, 2017, and 2021, explained by high-concentration episodes of PM₁₀, where the concentration exceeded the European standard for PM₁₀ concentration in 24 h (PM₁₀ concentration varied from 50 to 78 µgm-3). The linear fit line of the trend is highly significant with R2=0.97, with a reduction of 134 ng m⁻³ yr⁻¹ (-5.5 % yr⁻¹).

This reduction of biomass burning concentrations in Grenoble is 4 times higher than the results from a long-term study (2012 to 2020) in a French rural site – (Observatoire Pérenne de l'Environnement, OPE) (Borlaza et al., 2022) – estimated at 33 ng m⁻³ yr⁻¹ over the same period. Besides the study of Borlaza et al. (2022), there are no previous PMF studies describing any trend of biomass burning factors. Nevertheless, similar trends were found for concentrations of biomass burning tracers. In particular, Font et al. (2022) presented a downward trend of PM₁₀ concentration from wood burning (a reduction from 1.5 % yr⁻¹ to 3.8 % yr⁻¹) in urban sites in the United Kingdom from 2010 to 2021, by calculating the emission of wood burning from aethalometer measurement. Similarly, from 2002 to 2018 in Norway, a downward trend of 2.8 % yr⁻¹ was also detected for levoglucosan (Yttri et al., 2021). Additionally, Colette et al. (2020) modeled the trend of the emissions from different activities in Europe, showing that the trend of PM₁₀ heating emissions was decreasing in the period 2000–2017, with mean rate values varying from 0.8 % yr⁻¹ to 3.3 % yr⁻¹ for 30 European countries (EMEP monitoring sites). Even though the chemicals and the period of these studies differ, a decreasing trend is generally observed among European cities, including the one investigated here. Interestingly, the biomass burning source in Grenoble shows the strongest decreasing trend, with a reduction of 5.5 % yr⁻¹.

Since the biomass burning sources in Grenoble are related to residential heating, the observed reduction of the concentrations from this source could be linked to household behaviors (including appliance renovation) on top of the changes in meteorological conditions, lowering the overall heating demand. The average annual biomass burning sources PMF-derived is compared to the local PM₁₀ emission inventory for residential heating (tonnes) in the Grenoble metropolis, estimated by the regional air quality monitoring agency (Atmo AuRA), to confirm the trend of biomass burning (Fig. 8). This emission inventory has been available until 2022.

Except for the year 2020, the annual average of biomass burning agreed with the emission inventory, demonstrating the consistency between the sources observed by the PMF model and the local inventory emission data. Since 2015, the Grenoble metropolis has set up an air-wood bonus to encourage households to renew their individual wood-burning appliance (fireplace or stove). It aims to replace all open fireplaces with closed appliances in October 2024. The downward trend of biomass burning concentration could then be considered as partly due to the implementation of dedicated action plans at the regional scale.

Similar to the time series of biomass burning concentrations, the traffic contribution was subjected to specific STL analysis (Fig. 9). A significant downward trend of the concentrations of PM from traffic emission is detected with a reduction of 374 ng m⁻³ yr⁻¹ (12.9 % yr⁻¹) (p value ≪0.01). This reduction is almost 3 times larger than that of the biomass burning concentrations. Traffic concentration before 2017 also showed a clear seasonality with maxima in winter, which nearly disappeared from 2018 onward. It is striking that the same behaviors (strong downward trend and smoothing) are also observed for NO_x concentration, another indicator of traffic exhaust emission, which is also observed for NO_x seasonal patterns (see Sect. S6 and Fig. S7). Residuals show extreme values in the same month as biomass burning in 2016 and 2017, matching the PM₁₀ episode. The traffic trend closely follows a linear regression fit line, with R2=0.94.

The downward traffic trend observed in this study is consistent with another long-term study (2012–2020) of a rural site in France, which showed a traffic trend of -6.5 % yr⁻¹ (58 % total reduction) (Borlaza et al., 2022). This is aligned with other results of fossil fuel black carbon in several rural sites in France (Font et al., 2025), or EC over many rural sites in Europe (Aas et al., 2024). Additionally, our result also agrees with other studies, like that by Pandolfi et al. (2016), which indicated a downward trend of traffic sources in an urban site in Spain, with a reduction of 5.6 % yr⁻¹ (56 % total reduction), which is lower than that of this study. Finally, the trend of traffic emission to PM₁₀ in 30 European countries was modeled as reported by Colette et al. (2020), showing a downward trend with a reduction from 2.3 % yr⁻¹ to 3.5 % yr⁻¹ from 2000 to 2017. As for biomass burning, the Grenoble supersite seems then experiencing faster reductions in primary traffic PM loadings than most of others European cities.

Furthermore, the PMF-derived traffic factor was compared to the local PM₁₀ traffic emission inventory by fuel type (provided by Atmo AuRA), revealing very similar trends (Fig. 10). In addition, this source is also compared to the PM₁₀ emission by the transport sector (kilotonnes) over France, which was assessed from the emission inventory data of CITEPA (Fig. S11), also confirming the concomitant reductions of traffic emissions and contributions to PM₁₀ in ambient air.

This traffic trend may be separated into three parts. Between 2014 and 2016 with a slow decrease trend of -3 % yr⁻¹; from 2016 to 2021, with an average reduction of 10 % yr⁻¹, and a mild increasing trend of approximately 3 % yr⁻¹ in the last three years of the study. The beginning of this increase coincides with the post-lockdown period, when transportation activities were back to normal, resulting in a fairly similar contribution of traffic sources compared to that in the pre-lockdown period.

Besides the implementation of the two versions of the Euro 6 emission standards (introduced in 2015 and 2018, respectively), local emission abatement strategies decided by Grenoble municipality from 2016 onwards might be the main drivers for the observed decreasing trends (City's low emission zone https://zfe.grenoblealpesmetropole.fr/, last assess: 21 May 2025).

Following the methodology exposed in Ngoc Thuy et al. (2024), the characteristics of the dataset, including collinearity and heteroscedasticity, are tested in order to select a satisfactory inversion model for OP_DTT source apportionment (SA) and OP_AA SA (Table S8). The OP SA can be applied for the 11-year PMF solution since the source profiles have been demonstrated to be homogenous over the years. Consequently, the OP^m should be substantially homogenous over the years (Ngoc Thuy et al., 2024), and it is unnecessary to perform the OP SA for each year separately. The characteristic tests indicate that the weighted positive least squares (wPLS) and weighted least squares (WLS) could be suitable models for both OP_AA and OP_DTT SA. The average accuracy metrics of the testing dataset in 500 iteration runs (including R2, RMSE, MAE) of wPLS and WLS were compared to select the most appropriate model (Table S9). Finally, WLS was chosen due to the highest R2 and lowest error for both OP_AA and OP_DTT prediction. The comparison between observed and predicted OP_AA and OP_DTT showed a good correlation between measured OP and WLS predicted OP, with R2=0.80 and 0.70 for OP_AA and OP_DTT, respectively (Figs. S12 and S13), with n=1570 for OP_AA and OP_DTT.

In addition, the study revealed good performance of Mutiple Layer Perceptron (MLP) and Random Forest (RF) for the training and testing datasets (Table S10). These neural network models were overfitting the results of OP SA for the 6 French sites tested in Ngoc Thuy et al. (2024) since the number of samples was lower than 200 for individual sites. The present study confirmed the conclusion of Ngoc Thuy et al. (2024), demonstrating that a higher number of samples improved the performance of the neural network model. However, such non-linear models do not provide values for the intrinsic OP, which is basically the regression slope of the regression. Since the objectives of MLP and RF are not to define a “slope” but to better predict OP, therefore, the “slopes” of such models actually constantly vary with the input data to ensure the best performance of the model. Since the OP intrinsic is not defined, these models cannot be selected for the final results at this stage.

The intrinsic OP of 1 µg PM₁₀ source (OP^m nmolmin-1µg-1) is investigated thanks to the WLS technique, resulting in 500 values of OP^m for each source (Tables 2 and S11). The anthropogenic sources, including biomass burning, industrial, and traffic, have the dominant intrinsic OP_DTT and OP_AA, which is consistent with the study in 2017–2018 in Grenoble (Borlaza et al., 2021b) and results obtained at other French sites (Ngoc Thuy et al., 2024; Weber et al., 2021) and EU sites (Fadel et al., 2023; Veld et al., 2023). The different ranking of the intrinsic OP of the sources according to the two assays is also aligned with previous results (Weber et al., 2021). While intrinsic OP_AA of biomass burning is highest (0.76 nmolmin-1µg-1), followed by industrial (0.48 nmolmin-1µg-1) and traffic (0.38 nmolmin-1µg-1), the order of intrinsic OP_DTT is industrial (0.52 nmolmin-1µg-1), traffic (0.38 nmolmin-1µg-1) and biomass burning (0.14 nmolmin-1µg-1). The intrinsic OP_DTT of biomass burning is also lower than that of OP_AA, as reported by Borlaza et al. (2021b), suggesting the synergistic and antagonistic effects between some elements, quinones, or bioaerosols, decreasing the overall intrinsic OP_DTT of this source (Pietrogrande et al., 2022; Samake et al., 2017; Xiong et al., 2017).

The other anthropogenic sources, including nitrate-rich and sulfate-rich, have lower intrinsic OP than anthropogenic sources associated with combustion (traffic and biomass burning), as reported by Daellenbach et al. (2020). The natural sources have a negligible intrinsic OP (lower than 0.03 nmolmin-1µg-1 for OP_DTT and 0.2 nmolmin-1µg-1 for OP_AA). These findings highlight the high impact of the anthropogenic sources, verified for the overall period 2013–2023.

The trend of OP is first presented by the yearly average contribution of sources to OP_AA and OP_DTT (Fig. 4), indicating a reduction of OP values over the years. Overall, the yearly average of the OPAAv and OPDTTv is decreasing and reached its lowest values in 2021 (2.41 and 1.17 nmolmin-1m-3 for OP_AA and OP_DTT, respectively). From 2018 onward, both assays consistently exhibited lower OP^v values than in preceding years. Although OP_v is calculated using PM₁₀ concentration, implying that a decrease in PM₁₀ concentration generally reduces OPv, the contribution of sources to OP is different from that of PM₁₀. While dust and sulfate-rich are dominantly contribute to PM₁₀, biomass burning is the most important contributor to OP_AA (1.87±2.7 nmol min⁻¹ m⁻³), and primary traffic is commonly assessed as the largest contributor to OP_DTT (0.71±0.70 nmol min⁻¹ m⁻³). The industrial mass contribution is 10 times lower than that of the sulfate-rich. However, industrial emissions appear to contribute much more to OP_AA and equally to OP_DTT than the sulfate-rich factor. This finding was also observed in 2017–2018 at the same site in Grenoble (Borlaza et al., 2021b). This significant contribution of traffic and biomass burning over the years is more evident when considering relative contribution (Fig. S15). These results again emphasize the importance of considering not only the mass concentration but also its redox activity in evaluating the potential adverse health effects of a source of PM.

In addition, the temporal evolution of OP_AA and OP_DTT did not exactly follow PM₁₀ trends, especially for the period of 2016–2017 and 2019–2020. Regarding the period between 2016 and 2017, a dramatic increase in PM₁₀ concentration is observed, principally due to the higher contribution of nitrate and sulfate-rich. On the other hand, OP_AA and OP_DTT values remained fairly unchanged between 2016 and 2017. Focus on 2019 and 2020, the PM concentration and OP_v values are identical (less than 0.001 µgm-3 and nmolmin-1m-3 of difference, respectively), while OPAAv presents a remarkable difference (Δ=0.8 nmolmin-1m-3). Indeed, the discrepancy between 2019 and 2020 in OPAAv is principally attributable to a higher contribution to biomass burning, which is the dominant driver of OPAAv. Overall, the downward trend of OP_AA and OP_DTT is different from PM₁₀, since the driven sources of OP and PM are different.

The yearly average may not be properly representative of the trends of OP; therefore, a STL deconvolution was performed for OPAAm and OPDTTm (Figs. S16, S17, respectively) to investigate the trend of OP^m over the 11 years of the study. Indeed, considering the trend of the intrinsic OP^m confirms that the downward trend of some sources leads to a change in the trend of OPAAm and OPDTTm.

An insignificant linear trend is observed for OPAAm (fit line: R2=0.4, p values ≪0.01), yet its average intrinsic activity still exhibits a decreasing value, with the annual mean falling by approximately 0.002 nmolmin-1m-3 (2.5 %) across the study period. Interestingly, the seasonality of OPAAm exactly matches the seasonality of biomass-burning concentrations, pointing out that the high values of OPAAm in winter align with biomass-burning activities. The trend line of OPAAm did not match the trend of biomass burning nor that of the traffic or industrial emissions, suggesting the synergistic effect between sources, as well as the influence of the other sources outside of the winter season, such as MSA-rich and primary biogenic, which get a high ranking of OPAAm (Table 2).

Conversely, the OPDTTm showed a significant downward trend (R2=0.6, p value ≪0.01), with a reduction of 0.005 nmolmin-1µg-1 (6.5 %) across 11 years. The seasonality of OP_DTT is different from that of biomass burning and OPAAm, since biomass burning is not the main driver of OP_DTT (only ranked third), indicating a lower influence of this source on OPDTTm compared to OPAAm. Interestingly, a slight increase in OPDTTm from 2021 onward is also observed, which is associated with PM₁₀ and traffic, suggesting that traffic emission could be the main driver for increasing PM₁₀ concentration and OPDTTm from 2021. Overall, the relative decrease of OPDTTm exceeds that of OPAAm could be explained by the 4th most important contributor to these OPs. All four leading contributors to OPDTTm show significant reductions, whereas MSA-rich factor, one of the top four contributors to OPAAm, has an increasing trend. These findings again underscore that trends in OP^m are governed by the evolution of the sources most active in each assay. Thus, the decrease in the magnitude of the OP_m depends on how its dominant redox-active sources evolve over time.

Considering the volume-based metrics (OP_v), a downward trend is detected for OP_AA and OP_DTT. PM₁₀ decreased by 3.9 % over the decade, which is consistently comparable to OPAAv (4.9 %) and OPDTTv (5.3 %). This good agreement partially reflects the influence of the PM mass concentration since these OP_v values are calculated using PM₁₀ concentration. However, the slight difference in the relative downward trend could be related to the most driven sources of OP and PM, as discussed above.

Finally, the impact of persistent inversion days on the OP^v is also investigated to assess the association between the redox activity of PM sources and thermal inversion. A comparison of the source's contribution to OP^v (for both AA and DTT) between the period with and without persistent inversions is carried out and shown in Fig. S14. The comparison confirms the larger increases in average OP_AA (85.1 %) and OP_DTT (63.8 %) compared to that of PM₁₀ (39.6 %) for the persistent inversion periods. The higher values of OP_AA and OP_DTT are related to the larger increases in the contribution of local anthropogenic sources, with BB impacting most the OP_AA values while traffic significantly influences OP_DTT. This result again highlights the potential effect of persistent inversion on the PM₁₀ source's contribution, but all the more of their redox-active properties, which could be associated with the health-relevant metrics (Tassel et al., 2025).

Over the decade, anthropogenic sources have driven OP, with biomass burning impacting OP_AA and traffic/industrial sources dominating OP_DTT. Frequent thermal inversion in Alpine valley strongly amplifies OP, which is more significant than the mass of PM₁₀ itself. Finally, OP_v and intrinsic OP trends over the decade do not align with that of PM₁₀ mass, emphasizing the need to prioritize redox-active components over the bulk PM concentration in air quality policy.