Supporting Information for: On the chemical nature of the Oxygenated Organic Aerosol: implication in the formation and aging of α-pinene SOA in a Mediterranean Environment, Marseille

a Aix-Marseille Univ, Laboratoire Chimie Environneme nt, 13331, Marseille cedex 03, France b Université Lyon 1, Lyon, F-69626, France; CNRS, UM R5256, IRCELYON, Institut de Recherches sur la Catalyse et l’Environnement de Lyon, Villeurbanne, F-69626, France c Universités Joseph Fourier-Grenoble 1-CNRS, UMR 51 83, Laboratoire de Glaciologie et Géophysique de l’Environnement, Saint Martin d’Hères, 38402, Franc e. d Laboratoire des Sciences du Climat et de l’Environ nement, CEA-CNRS-UVSQ-IPSL, Gif sur Yvette, 91191, France.


S.1 Site description and sampling strategy
Description of the sampling site and meteorological conditions encountered during the study are thoroughly detailed in El Haddad et al., 2011a and b, and only a brief outline follows.Field measurements were conducted in summer 2008 (30 June-14 July), at an urban background site located in a downtown park, in Marseille (43°18′20″N, 5°23′40″E, 64 m.a.s.l.).Marseille, the second most populated city in France, comprises the most important port of the Mediterranean Sea.It is also in the vicinity of a large petrochemical and industrial area, located 40 km northwest of the metropolitan area.The main industries include petroleum refining, shipbuilding, steel facilities and coke production.
Owing to a particular air mass circulation in this region, industrial emissions can directly impact the metropolitan area of Marseille, particularly in sea breeze conditions.This area is also well known for its photochemical pollution, especially regarding ozone (Flaounas et al., 2009), and evidence of rapid formation of secondary organic aerosol has been pointed out within the frameworks of ESCOMPTE experiment (Cachier et al., 2005) and BOND project (Petäjä et al., 2007).Fig. S1 illustrates the air masses impacting the sampling site during the measurement period, showing that these are mostly associated with long range transport from the Atlantic and the Mediterranean.(Rolph, 2010) illustrating the overall air masses circulation occurring during the entire measurement campaign.Backward trajectories are confirmed by both MM5 modeling and local wind measurements.PM 2.5 collection was performed using high volume samplers (HiVol,Digitel DA80) operating at a flow rate of 30 m 3 h -1 .The samples were collected continuously on a 12hour basis (5:30 to 17:30 UT, and 17:30 to 05:30 UT, total number of 30 samples) onto 150mm-diameter quartz fiber filters (Whatman QMA), pre-heated at 500 °C during 3 h.Samples were then stored at -18 °C in aluminium foil, sealed in polyethylene bags until analysis.Six field blank samples were also prepared following the same procedure.PM 2.5 were also collected on a 24-h timescale onto pre-heated 25 mm-diameter quartz filters using a Dekati 13-stage low pressure cascade impactor (LPI) at a flow rate of 30 l min −1 , for size resolved EC/OC analysis.
The chemical composition of fine PM was also investigated every 2 min using a compact time-of-flight (c-TOF, Tofwerk) Aerodyne Aerosol Mass Spectrometer (AMS, Aerodyne).This instrument allows real-time measurements of PM 1 non-refractory components (OA, NH 4 , NO 3 and SO 4 ) combining thermal vaporization and electron ionization (Drewnick et al., 2005).Aerosol size distribution (mobility diameters from 11 to 1083 nm), was investigated using a Scanning Mobility Particle Sizer (SMPS, L-DMA and CPC5403, GRIMM).Semi-continuous hourly concentrations of elemental carbon (EC) and organic carbon (OC) in PM 2.5 were obtained in the field from an OC/EC Sunset field instrument (Sunset Laboratory, Forest Grove, OR, USA; Bae et al., 2004) running at 8l min -1 .AMS results are all corrected for the collection efficiency, using a common factor of 0.65±0.14,estimated based on SMPS and EC data.Fig. S2:.Time series of the main PM 1 components (EC, OA, NH 4 , NO 3 and SO 4 ) during the period of study.Due to technical issues, AMS measurements are not available between the 9th and 10 th of July.
In addition to aerosol sampling, HS-PTRMS (High Sensitivity Proton Transfer Reaction Mass Spectrometer, Ionicon Analytic, (Lindinger et al., 1998) was deployed in order to quantify volatile organic compounds (VOCs) and oxygenated VOCs.In this paper, a special focus is placed on the isoprene oxidation products (Methacroleine, MACR and Methyl Vinyl Ketone, MVK), tracers of aged biogenic air masses impacting the site.
Finally, 15 min-averaged NO X , O 3 , SO 2 and PM 2.5 concentrations were also measured with the standard equipment of the air quality monitoring network.

S.2 Offline chemical analyses
PM 2.5 collected onto 150mm-diameter filters was comprehensively characterized.
Technical description of the analysis techniques can be found in El Haddad et al., 2011a and b and only a brief outline follows.
EC/OC, ions, WSOC, HULIS WS and elements: The carbonaceous content was analyzed for EC and OC using a Thermo-Optical Transmission method on a Sunset Lab analyzer (Birch and Cary, 1996), following both NIOSH (Schmid et al., 2001) and EUSAAR-2 (Cavalli et al., 2010) protocols.It is well established that different protocols result in very different values for EC (Schmid et al., 2001).We based our analysis (i.e.Chemical Mass Balance analysis and multiple regression analysis, see below sections S.4 and S.8) on concentrations determined following NIOSH protocol, as source profiles were determined based on this protocol.Biases arising from discrepancies between the two protocols are all discussed in S.8.2 of the supporting material.
HULIS analysis was performed following the method described in Baduel et al. (2009Baduel et al. ( , 2010)).This method involves extraction of HULIS by adsorption onto DEAE resin (GE Healthcare®, HiTrapTM DEAE FF, 0.7 cm ID×2.5 cm length) and its subsequent quantification with an OI Analytical 700 total organic carbon analyzer.
Finally, fifty elements were measured using ICP-MS (Agilent 7500ce) following complete dissolution of filter aliquots in a mixture of high-purity concentrated HF and HNO 3 .Element concentrations were then calculated using the rock reference material BR (Chauvel et al., 2010).
Radiocarbon measurements: Radiocarbon ( 14 C) measurements were conducted on high volume quartz filter fractions (~40 cm 2 ) using ARTEMIS Accelerator Mass Spectrometry.Each sample was first packed into a prefired quartz tube containing CuO and Ag powder to be combusted at 850°C in a muffle furnace for 4 hours.Carbon dioxide was collected and purified before its conversion into graphite by hydrogen reduction at 600°C using Fe catalyst.The modern fraction (f m ) was determined as the ratio of 14 C/ 12 C in aerosol sample to 14 C/ 12 C in the NBS Oxalic Acid standard (NIST-SRM-4990B).
In order to account for the thermonuclear weapon tests of the late 1950s and early 1960s (Levin et al., 1985), the modern fraction (f m ) is divided by a ratio of 1.1 to get a corrected non-fossil fraction (f nf ) (Levin and Hesshaimer, 2000).This value is subsequently subtracted from one to obtain the fossil fraction (f f ).
Organic speciation: A chemical derivatisation/gas chromatography-mass spectrometry (GC-MS) approach was used to quantify primary and secondary organic markers, including α-pinene oxidation products, a major focus of this study.The approach is fully described in El Haddad et al., 2011b and will be only outlined in the following.
Prior to extraction, filters were spiked with known amounts of two isotope-labelled standards: tetracosane-d50 and cholesterol-d6.Organic species were extracted from filters with a dichloromethane/acetone mix (1/1 v:v) using an accelerated pressurized solvent extraction device (ASE, Dionex 300).Extracts were then reduced to a volume of 500µL using a Turbo Vap II concentrator.The remainder was split into two fractions.
The first fraction was directly injected, whilst the second fraction was subjected to derivation for 2 h at 70 °C before GC-MS analysis, using N,O-Bis(trimethylsilyl)trifluoroacetamide containing 10% trimethyl-chlorosilane.The two fractions were analyzed following the same GC-MS conditions: Aliquots of 2 µL were analyzed using a Thermo Trace GC chromatograph interfaced to a Polaris Q ion trap mass spectrometer fitted with an external electron ionization source.The chromatographic separation was accomplished on a TR-5MS capillary column (Thermo Electron, 30 m × 0.25 mm i.d.× 0.25 µm film thickness).Field blank filters were also treated with the same procedure and none of the target compounds were detected in these field blanks.

S.3 Comparison between offline and online measurements
The aim of this section is to evaluate biases and artefacts associated with the offline and online measurements of OA (e.g.AMS particle collection efficiency, adsorption artefacts onto filters).Fig. S3 conveys the comparison between AMS (PM 1 ) and filter based (PM 2.5 ) measurements for the two major aerosol components: SO 4 and OA.SO 4 is expected to primarily occur in the PM 1 fraction and to be mostly associated with ammonium sulfate and bisulfate (very little influence from sea salt), and thus to be quantitatively analysed by the AMS.A very good agreement can be observed between the AMS-SO 4 and the HiVol-SO 4 (s~1, i~0 and R 2 >0.9;Fig. S3), substantiating our AMS measurements and the particle collection efficiency factor, CE=0.Fig. S3: Comparison between AMS and offline measurements for SO 4 (A) and OA (B).HiVol-OA was corrected for differences in the diameter cut-offs between the AMS and the HiVol sampler (see text); it refers to the PM 1 fraction.Also, shown are the 1:1 line and the slopes (s), intercepts (i) and coefficients of determination (R 2 ) obtained by linear fits of the data.

S.4 CMB analysis
Available data used here also include source contributions to OC, apportioned using a Chemical Mass Balance analysis (CMB) in conjunction with organic marker concentrations, as fully described in El Haddad et al. (2011a).CMB model is based on the mass conservation of individual organic markers.In the mass conservation equations, known concentrations (C ik ) of specific markers of primary sources at receptor site k are written as the product of known source profiles a ij and unknown primary source contributions s jk (Watson et al., 1998) as expressed in equation 1: where m denotes the total number of emission sources and a ij is represented by the fractional abundances of chemical species in the source emissions, expressed as markerto-OC ratios.The set of linear equations generated by equation 1 is solved with an effective variance weighted least square method using Environmental Protection agency EPA-CMB8.2software.
Primary markers and source profiles selection is detailed in El Haddad et al., 2011a.
In this study, emissions from the three industrial processes are lumped together under the term "industrial OA".Biomass burning, vegetative detritus and natural gas combustion contributed very little OC during the period of measurements (El Haddad et al., 2011a) and were not considered in the comparison between CMB and AMS/PMF results (see section S.7).CMB technique is unable to directly apportion secondary sources; however, the fraction of OC non-attributed to primary sources is considered as an upper bound estimate of secondary OC (SOC).
In order to compare CMB and AMS/PMF results, primary OA associated with vehicular and industrial emissions were calculated applying an OM-to-OC ratio of 1.2 (based on Aiken et al., 2008).SOA is considered as the difference between the total OA, determined by scaling the total OC by an OM-to-OC ratio of 1.67 ‡ , and the primary OA.
An OM-to-OC ratio of 1.82 can be inferred for SOA (i.e.SOA-to-SOC), consistent with an overwhelmingly secondary origin of this fraction (Aiken et al., 2008).

S.5 estimation of α-pinene SOA contributions
Contributions of α-pinene SOA to ambient OA is estimated following the marker-based approach developed by Kleindienst et al. (2007).This approach consists of converting the measured concentrations of marker compounds derived from a given precursor hydrocarbon (hydrocarbon, HC) into a SOA concentration in µg m -3 ( [ ] HC SOA ), using the laboratory-generated mass fractions of the same markers ( HC SOA f , ) determined by Kleindienst et al. (2007).[ ] HC SOA can be subsequently calculated as follows: where [ ] i M is the concentration of the marker i and J is the total number of markers derived from the hydrocarbon HC.For the apportionment of α-pinene SOA contributions, pinonic acid, pinic acid and A1-A7 were considered as markers and a HC SOA f , factor of 0.168±0.08 was used (Kleindienst et al., 2007).

S.6.1 General principle
The AMS OA dataset was analyzed applying Positive Matrix Factorization (PMF2) (Ulbrich et al. 2009).A complete description of PMF2 and the subsequent evaluation tools can be found in Paatero and Tapper (1994), Lanz et al. (2007) and Ulbrich et al. (2009).Time series of organic mass spectra, arranged as a matrix (X), are deconvolved into a linear combination of smaller matrices such that (3) ‡ OM-to-OC ratio of 1.67 is calculated by comparing AMS OA with LPI OC measurements, see section S.3.
where ij X represents the elements of matrix X, p the number of factors in the solution, g ip and f pj the elements of matrices G and F representing respectively time series and mass spectra of each factor, and e ij the elements of matrix E of residuals not fitted by the model for each data point.This model includes both the data matrix of organic fragments and an instrumental error matrix, both obtained from the AMS-data-analysis Squirrel software.
In this study, the data matrix is composed of 4043 data points (time series) of 276 m/z.We first examined a 3 factor solution, apportioning the OA to hydrocarbon like OA (HOA), related to traffic emissions, and to two oxygenated fractions related to semivolatile moderately oxidized (SV-OOA) and low volatility highly oxidized (LV-OOA) OA. Adding one more factor enabled revealing the industrial influence, evidenced by the correlation of the corresponding factor (termed F4) with heavy metals (e.g.Pb, Cs, Mo, Fe, La, V, Zn, Ni and Co measured on filter samples) and polycyclic aromatic compounds (PAH, measured online by the AMS and offline).This factor exhibited remarkable variations similar to that of industrial OA apportioned using CMB, both characterised by episodic ten-fold enhancements in their contributions (Fig. 2 in the manuscript).The wide variability in the contribution of industrial emissions is due to the fact that it is a local point source, whose influence is strongly dependent on local meteorology (El Haddad et al., 2011a).The mass spectrum of F4 exhibits a similar pattern as HOA, with high contribution from m/z 55, 57, and 69 and higher molecular weight fragments.F4 mass spectrum is also associated with high contribution from oxygenated fragments, e.g.m/z44 (CO 2 + ), suggesting to some extent the aging of the emissions while transported to the site.Unfortunately, the low resolution of the C-TOF precludes an enhanced chemical characterisation of this factor.Overall, for this solution, the factor's mass spectra exhibit typical patterns, very similar to those obtained in previous studies (Tab.S1).SV-OOA correlates fairly well with volatile and semi-volatile secondary components such as nitrate and isoprene gas phase oxidation products (MVK+MACR, see Fig. 2).In contrast, LV-OOA correlates with less volatile secondary components, including HULIS and sulfate (Fig. 2), which clearly relates this fraction with low volatile highly aged SOA.HOA is characterised by a prominent diurnal pattern, with increasing contributions during rush hours (Fig. S5), correlating with vehicular tracers (EC and NO X , Fig. 2).

S.6.3 Robustness of the selected solution
To assess the robustness of the 4 factor solution, rotational ambiguity has been investigated by varying FPEAK from -2 to 2 with 0.1 steps.Two main groups of solutions are identified, the first one corresponding to "FPEAK" values below 0, for which unrealistic zero time series values are observed for LVOOA, and the other one corresponding to "FPEAK" above 0. Robust solutions are found between 0 to 1 FPEAKs, with very little variability in the factor's time series and mass spectra (Fig. S7).June to 13 th July 2008 period, with "FPEAK" = 0 and "seed" = 0.
The influence of the initial conditions "SEEDS" (corresponding to pseudorandom starting-points of the PMF2 algorithm) ranging from 0 to 60 (with steps of 1) was also verified (Fig. S8).No influence of different SEEDS is observed, an evidence of the robustness of the chosen solution.Fig. S9: Inter-comparison of AMS and CMB results.For the AMS/PMF2, Traffic OA, industrial OA and SOA denote the HOA, F4 and the sum of SV-OOA and LV-OOA.For the CMB model, industrial OA represents the aggregate contribution from three processes: coke production, metal smelting and shipping/oil burning and SOA is the fraction un-apportioned to the primary sources.
Regardless a great number of uncertainties, artefacts and assumptions underlying the apportionments delivered by AMS/PMF and CMB § , the inter-comparison between both § Uncertainties/errors include: uncertainties in the measurement of markers and mass spectra, uncertainties in the PMF model (e.g.number of factors considered) and source profile selection in CMB (e.g.nonrepresentative or non-considered profiles), uncertainties in OM/OC ratios, reactivity of organic markers, adsorption artefacts onto filters and differences in size cut-offs between HiVol samplers (PM 2.5 ) and AMS (PM 1 ).
models is excellent (Fig. S9).Further, as shown in Fig. 2 in the manuscript for the case of industrial emissions, both models capture the same variability in the factor's contributions.However, AMS/PMF2 tends to provide less contribution from SOA (sum of OOAs), as unlike POA, this component is likely to be enhanced in the PM 2.5 -PM 1 fraction and thus would not be measured by the AMS: ~20% of the carbon mass is estimated to occur in the PM 2.5 -PM 1 fraction, mostly related to secondary matter and primary dust and biological particles.On the other hand, the CMB model underapportions the industrial source compared to the AMS/PMF2, owing most probably to omitted profiles for fugitive industrial emissions in the CMB and to the occurrence of aged OA in the industrial plumes.These emissions are often concomitant with new particle formation events, when the site is impacted by land and sea breezes allowing the aging of the emissions (El Haddad et al., 2011a).

S.8.1 Multiple regression model
AMS/PMF apportionments and 14 C measurements are combined using a multiple regression model to estimate the fossil and the non-fossil contributions to both SVOOA and LVOOA.It is worthwhile to note that such a combination is not straightforward, involving a certain number of assumptions resulting in considerable uncertainties.
First of all, 14 C measurements are conducted on PM 2.5 OC onto filter samples that are subjected to well-known but non systematic adsorption artefacts of gas phase organic compounds.In contrast, AMS provides real-time measurements of PM 1 OA with little interference from gas phase organics.However, particle collection efficiency (CE) of the AMS estimated, here, (CE=0.65)can be highly variable, depending on the aerosol chemical nature and mixing state.
Second of all, 14 C measurement conducted in this study relates to the total carbon (TC) mass that can be oxidized at 850 °C under oxygen, i.e. organic carbon (OC) and elemental carbon (EC), whereas AMS quantifies OA that consists of OC and the associated heteroatom (H, N, O, S …).This fundamental difference engenders two major limitations for the assessment of fossil and non-fossil contributions of the OOA fractions.
First, as the TC apportioned by 14 C measurements also includes EC, assumptions related to the origins of the latter must be made.Furthermore, since the separation between EC and OC measured using a Sunset analyser is method-dependent, biases associated with EC determination can impact the estimations.Second, the AMS measurement of OA includes heteroatom that can be unevenly distributed between the fossil carbon and the non-fossil carbon.Here also, this distribution is not empirically accessible and assumptions have to be considered to achieve our estimations.Further uncertainties can arise from PMF calculations and residuals and from variability in the biomass 14 C/ 12 C ratio.The assumptions made to achieve the apportionment are explicitly presented in this section and the resulting biases and uncertainties are thoroughly discussed in section S.8.2.2.
The procedure goes as follows: First, in order to estimate the fossil and non-fossil fractions of OC, EC was assumed to be entirely related to fossil carbon (assumption founded on Chemical Mass Balance calculations reported in El Haddad et al., 2011a).This is described in equations ( 4) and ( 5): ( ) Where f TC , ( ) whereα is the ratio between AMS PM 1 OA and filter PM 2.5 OC. α is variable depending on the considered sample, but has an average of 0.92±0.21.The assumption underlying this scaling is that PM 2.5 OC and PM 1 OA are associated with the same proportions of fossil and non-fossil mass (i.e. the ratio fossil/total is the same for PM 2.5 OC and PM 1 OA).
Biases arising form this assumption are discussed in the section S.8.2.2.
where l and m are the total numbers of where i a and i b denote the relative share of fossil and non-fossil fractions to i OA factors, b for nf SVOOA and 3 b for nf LVOOA ).This system of linear equations can be visualised as the following matrix equation: ), the parameter 1 a can be assumed to equal 1, implying that 1 b is equal to 0 (i.e.0 = nf POA , the sensitivity of the results to this assumption is assessed in section S.8.2.2).Equation 12can be then simplified as follows: . The equations of the resulting linear system (equation 13) are not independent and thus the system has no solution.For that reason, a multiple linear regression analysis was applied instead to solve equation ( 13) finding the average values for i a and i b that fit best the equation, in the sense of solving the quadratic minimization problem.In equation ( 13) AMS/PMF i OA vectors are included as independent variables and f OA and nf OA as dependant variables.
It should be noted that the apportionment procedure followed here is not exclusive.As this study focuses mainly on AMS measurements, the method chosen here is AMS data oriented in that the resulting apportionments would exhibit the same variability as the AMS/PMF factors (e.g.
).This is the result of the equation ( 13), in which AMS/PMF i OA vectors were chosen as independent variables.As this equation is a self-consistent system (i.e.
another approach is also possible giving the same average result but orienting the variability towards 14 C measurements, by including f OA and nf OA as independent variables.

S.8.2.1 Output quality control and residual analyses
One of the major drawbacks of the multiple regression analysis applied here is that it considers a constant contribution of fossil and non-fossil sources to each of the i OA factors (i.e.constant i a and i b ratios), while these contributions may significantly vary over the course of the measurements.Accordingly, i a and i b ratios should be regarded as average contributions of fossil and non-fossil sources to i OA .These ratios are reported in Tab.S2. Tab.S2:

S.8.2.2 General assessment of uncertainties and biases
It is worthwhile to note that a great part of the uncertainties assessed for the absolute concentrations of fossil and non-fossil OA arises from discrepancies between AMS and filter measurements and hence not representative of the statistical significance of each of the fractions.The statistical significance of the relative contributions of HOA, F4, SVOOA f , LVOOA f , SVOOA nf and LVOOA nf are assessed through a sensitivity test, using a random selection technique.Inputs to the calculation are the PMF factor mass concentrations, 14 C data and OC/EC measurements.The calculation is performed based on equations (4-13); it proceeds as follows: o For each of the input parameters a range is assigned, within which these can vary (see BoxTab.S1).The criteria on which we based our assessment of this range are developed below, in the TextBox.S1.
o The parameters are then allowed to randomly vary within the range predetermined in the previous step, assuming a normal distribution.This approach is somewhat similar to Monte Carlo calculations and allows vast numbers of combinations of input parameters to be computed.A Monte Carlo simulation would involve testing all possible combinations of input parameters, which is proven prohibitive in terms of processing time.In contrast, random sampling is much more effective and for our purposes provides essentially the same results as a full Monte Carlo analysis.
o Following the approach described above, fifty sets of parameters are generated randomly and used subsequently in the equations 4-13 to calculate the inputs for the multiple linear regression analysis (i.e.A great advantage of this approach is that combinations of parameters which are very unlikely (e.g. that only the minimum-possible values from each parameter were used) will represent only a small percentage of the output.The obtained 500 solutions are presented graphically as a probability density (frequency distribution) of possible solutions to the source apportionment problem we have set up, as shown in Fig. S12.
This analysis provides strong support to our results, allowing the assessment of the uncertainties underlying our measurements and assumptions and offering a measure to our ability in separating the different components (statistical significance of each component and the likelihood between them).It shows that the uncertainties on our estimations depend on the component considered.Depending strictly on the PMF analysis errors, the uncertainties associated with the contributions of POA (HOA and F4) are less 10%.Conversely, for OOA components the uncertainties are less homogenous.
For non-fossil OOAs the uncertainties are around 10% as these are well resolved by the regression model, whereas for fossil OOAs uncertainties are higher (~36% and ~58% for SVOOA f and LVOOA f , respectively), as these are poorly resolved by the regression model and strongly dependant on the EC measurements and the assumptions on POA.All 6 fractions are statistically significant with contributions higher than 0 (Z equal 23, 7.4, 2.8, 1.7, 9.2, 12 for HOA, F4, SVOOA f , LVOOA f , SVOOA nf and LVOOA nf respectively, with Z=average/uncertainty).The sensitivity test offers as well the assessment of the biases on the apportionments presented in the manuscript.It suggest that we might underestimate the contributions of F4, HOA, SVOOA nf and LVOOA f by 9%, 6%, 5% and 23%, respectively and overestimate the SVOOA f and LVOOA nf by 34% and 5%, respectively.The main conclusion to be derived from this analysis is the robustness of the results presented and discussed in the manuscript.For example, OOA nf is clearly the biggest contributor to OA.
It is also clear that LVOOA derives predominately from non-fossil precursors (LVOOA nf /LVOOA=89±7%), whereas SVOOA encompasses a bigger fraction of fossil SOA (SVOOA nf /SVOOA=75±8%).Given the wide range of uncertainties used in the sensitivity test, these results demonstrate that in general we can clearly identify the contribution from different components.
TextBox.S1: Calculations of the different parameters entered as inputs to the uncertainty calculation For the parameters in BoxTab.S1, the ranges were established as follows: For EC/OC measurements, the range is designed to encompass biases and uncertainties associated with the separation between EC and OC.This range is bounded by measurements determined following NIOSH and EUSAAR2 protocols, respectively.For EC and OC, a constant bias between the 2 protocols is obtained as 40±8% and 6±5%.
The average uncertainty for the discrimination between fossil and non-fossil TC is 4%, including uncertainties on 14 C measurements and errors on the correction for 14 C inputs from the bomb testing.
An assumption made in equations 4 and 5 relates to the origin of EC, estimated to only pertain to the fossil fraction.The transgression of this assumption would bias high the contributions of fossil sources to the secondary OC fractions.As there was too little influence from biomass burning, we assumed an upper limit contribution of non-fossil sources to EC of 15%, (based on Minguillón et al., 2011 and references), and varied this contribution between 0 and 15%.The conversion from PM 2.5 OC measured on filter samples to PM 1 OA determined by the AMS was performed in equations 6 and 7, using a common factor α (with α = PM 1 OA/PM 2.5 OC) for both fossil and non-fossil OC.This factor encompasses three key corrections related to differences between the two measurement techniques, including differences in diameter cut-offs between AMS and filter sampling (referred to as OC PM OC PM 5 . 2 1 ratio), the adsorption artefacts on the filters and the OM/OC ratio.
The assumption underlying the PM 2.5 OC to PM 1 OA conversion is that the aggregate of the aforementioned corrections is similar for both fossil and non-fossil OC and well represented byα .
TextBox.S1: Continues.Indeed, these corrections can vary greatly between the primary and the secondary fractions, which exhibit variable contributions to the fossil and non-fossil OC.To address the biases resulting from the application of a single conversion factorα , the latter is deconvolved in equation ( 14 In this equation, ij α denotes the factor used for a conversion (j) applied to an i OC fraction.n is the total number of OC fractions; in our case, it is limited to two fractions representing the primary and the secondary OC.p is the total number of i α conversions applied to i OC ; in our case, p is equal 3, accounting for the OC PM OC PM 5 . 2 1 ratio (j=1), for the sampling artefacts (j=2) and for the OM/OC ratio (j=3).The following is an example performed for the campaign average value, representing the matrix of ij α factors for primary and secondary OC (BoxTab.S2).Similar calculations were performed for all the data set to achieve the sensitivity test.In this calculation, ij α were measured for the total -OC 2 α is the ratio allowing the correction for the sampling artefacts retrieved from Fig. S3, with an average value of 0.72.Artefacts are assumed to be evenly distributed between the primary and the secondary fractions, i.e. -OC 3 α is the OM/OC ratio, obtained from the comparison between PM 1 AMS and LPI measurements.In the calculation of equations 10 and 11, we considered for the apportionments the AMS/PMF2 results, including POA, SVOOA and LVOOA.One approach to assess the uncertainties on the AMS/PMF2 apportionments consists of varying FPEAK within a reasonable range, in our case between FPEAK=0 and FPEAK=1.
In the calculation of equations 12 and 13, we considered that POA pertains only to the fossil fraction, while we observed some evidence of some inputs from cooking (a non fossil primary source) to HOA.
To take this into account in the uncertainty calculations, we considered that these inputs contribute between 10% and 25% of total HOA, based on Fig. S6.
Fig.S1: 24h HYSPLIT air masses backward trajectories at 100m above sea level(Rolph, 2010) illustrating the overall air masses circulation occurring during the entire measurement campaign.Backward trajectories are confirmed by both MM5 modeling and local wind measurements.
Fig.S2 displays EC, OA, NH 4 , NO 3 and SO 4 time series recorded over the period of study.

Fig. S4 :
Fig.S4: Factor spectral profiles derived from the 4 factor solution PMF2 analysis and their average mass contribution over the period of study for "FPEAK"=0 and "seed"=0.The number of factors has been evaluated taking into account the physical meaning of the factors, good correlations with external time series (Fig.2 in the manuscript) and with mass spectra from laboratory and ambient samples (see for example Tab.S1 for the 4 factor solution).Diurnal patterns of the factors were examined as well.Ultimately, we retained a factor solution; the mass spectra, time series and diurnal variations of these factors a represented in Fig.S4, Fig.2 and Fig.S5, respectively.
Fig.S5:Diurnal patterns of the OA four factors identified during FORMES campaign in Marseille (solution for FPEAK=0 and SEED=0).
Fig.S6: HOA vs. EC.Color scale: hour of the day Increasing further the numbers of factors resulted in a splitting/mixing of SVOOA, Fig.S7: Spectra calculated for four factorial PMF2factors and their mass contribution for the 29 th non-fossil TC, OC and EC, respectively.The second step involves the transition from PM 2.5 OC measured on filter samples to PM 1 OA determined by the AMS.This conversion is achieved for each of the samples by scaling f OC and nf OC to fossil OA ( ) f OA and non-fossil OA ( ) nf OA , respectively, by a factorα : directly accessible, equations 7 and 8 can be written in terms of the OA factors ( ) i OA determined by AMS/PMF analysis and the respective share of fossil and non-fossil fractions to these factors such that fossil and non-fossil fractions of SOA, respectively, calculated as: Fig.S10compares the measured and the modeled concentrations for the total fossil and non-fossil fractions.It shows that the model capture quite well the amounts and the variability of the measured concentrations, especially in the case of the fossil fraction (Fig.S10-A).In the case of the non-fossil fraction, the model tends to slightly underestimate (overestimate) the measured levels at low (high) concentrations (Fig.S10-B).It should be noted though that most of the variability observed in panels A and B of Fig.S10, arises from discrepancies between the PM 1 OA and filter-based PM 2.5 OC, as shown in panel C of the same figure.Differences between the 2 measurement techniques were accounted for in equations 6 and 7 prior to the multiple regression analyses by the coefficient α that encompasses various conversion factors.Overall, these comparisons validate the representativeness of i a and i b obtained in the multiple regression analyses.
and LVOOA ).This provides for each set of parameters average values for a i and b i plus the corresponding uncertainties.oFor each set of the coefficients a i and b i previously generated, the average values of these coefficients are varied assuming a binomial distribution, derived based on the corresponding uncertainties provided by the multiple linear regression analyses.In this step and for each set of parameters, ten values are generated for a i and b i and used to compute the contributions of SVOOA f , LVOOA f , SVOOA nf and LVOOA nf .This gave in total 500 different solutions.
Fig.S12: Frequency distributions for the contribution of the different components assessed through the sensitivity test.The plot gives access to the average contribution and the associated standard deviation for the different components, indicated in the figure legend (Avg±1σ).
using size resolved OC measurements, i.e. 0.82±0.06%.POC was assumed to pertain entirely to the PM 1 fraction ( ) 77 (i.e.77% of the PM 2.5 SOC are comprised in the PM 1 fraction).

α
can be inferred for POC, SOC and total OC, which are equal to 0.84, 0.98 and 0.95, respectively.Estimated ∏ 95) is comparable to the averageα empirically determined and used in equations (6) and (7) to convert from PM 2.5 OC measured on filter samples to PM 1 OA determined by the AMS ( ) one can apply different conversion factors to POA and SOA.In the sensitivity test such a calculation has been made by considering a range of OC 1

2 Number of factors: Assigning factors to emissions and processes
Ulbrich et al. (2009)ulated in Squirrel software was modified following the recommendations ofUlbrich et al. (2009)and references therein. S.6.
Pearson correlation coefficients (R) between Factor's mass spectra and reference mass spectra drawn from previous ambient, emission and chamber studies. Tab.S1: