Aerosol measurements were conducted at the Mace Head Atmospheric Research Station (53°19^′ N, 9°54^′ W) on the west coast of Ireland from 2009 to 2018. The station, regularly exposed to clean marine air masses from the North Atlantic, has been a representative site for studying clean marine aerosols for several decades (O'Dowd et al., 2014).

We employed a high-resolution time-of-flight AMS (HR-ToF-AMS; DeCarlo et al., 2004) at Mace Head (Ovadnevaite et al., 2014) to measure the PM₁ (particulate matter with a diameter smaller than 1 µm) chemical composition, including organic aerosol (OA), non-sea-salt sulfate (nss-SO₄), sea salt, methanesulfonic acid (MSA) (Ovadnevaite et al., 2014), ammonium (NH₄), and nitrate (NO₃). Additionally, black carbon (BC) was measured using a multi-angle absorption photometer (MAAP) to trace anthropogenic emissions. Meteorological conditions were also recorded, including air temperature, pressure, precipitation, relative humidity (RH), wind speed, and wind direction.

A humidified tandem differential mobility analyser (HTDMA) (Swietlicki et al., 2000) was used to measure aerosol hygroscopic growth at a fixed relative humidity of 90 % for aerosol with selected dried sizes of 35, 50, 75, 110, and 165 nm. The growth factors measured by the HTDMA were inverted using a piecewise linear function (Gysel et al., 2009). 1GFmeana,b=1nfa,b∫abGFcGF,DdGF The arithmetic mean GF (GF_mean) was calculated as 2GFmean=∫GFcGF,DdGF. To quantify the mixing state, the GF spread factor (SF), defined as the standard deviation of the GF-PDF divided by the GF_mean, was calculated as 3SF=∫0∞GF-GFmean2c(GF,D)dGF1/2GFmean. The GF was measured at 90 % RH; however, the RH of the second differential mobility analyser (DMA) fluctuated slightly with the ambient temperature. The data between 88 %–92 % RH were corrected using the κ-Köhler theory according to the formula derived from Petters and Kreidenweis (2007), 4κ=GF3-1(1-aw)aw→GFaw,κ=1+κaw1-aw1/3, where a_w is the water activity, and were obtained by Köhler theory: 5aw=RHexp(4σsvwRTD), where σs is the surface tension of the droplet, vw is the partial molar volume of water, R is the universal gas constant, T is the temperature, and D is the diameter of the droplet. The surface tension is assumed to be 0.072 mN m⁻¹.

Data from the AMS and MAAP were averaged to a 10 min resolution, while the meteorological records were initially recorded hourly and later downscaled to 10 min intervals using linear interpolation to enlarge the dataset's availability. Any gaps in the AMS or MAAP data that contain invalid measurements were removed. Hourly boundary layer heights were obtained from ERA5 (Hersbach et al., 2020) and downscaled to 10 min resolution using linear interpolation.

To differentiate between marine POA and SOA at the Mace Head Atmospheric Research Station (MHD), this study employs an ML model to predict the mass concentration of marine SOA. Subsequently, POA concentrations are obtained by subtracting SOA from the measured total OA. Marine SOA was chosen as the ML model predictor because SOA is expected to be impacted by environmental factors and because less is known about POA, emission strengthening, and source regions. To ensure the high quality of the SOA production period, we implemented multiple screening criteria to minimize the influence of anthropogenic and marine primary signatures. Firstly, we applied clean sector criteria, limiting BC concentration to less than 15 ng m⁻³ and selecting wind direction between 190 and 300° (O'Dowd et al., 2014) to exclude continental outflows and ship plumes. The clean sector criteria, which have been established and applied in various MHD research studies (O'Dowd et al., 2014; Ovadnevaite et al., 2014; Xu et al., 2022), were employed to isolate marine air masses from anthropogenic influences.

We then applied additional filtering processes to reduce the impact of POA production. We further refined the data by keeping instances with wind speed below 6 m s⁻¹ and sea salt mass concentration under 0.03 µg m⁻³ to minimize the concentration of marine POA. Finally, we retained only those data points where nss-SO₄ was the dominant component (nss-SO₄/OA > 4) to ensure a predominantly secondary source. We presumed that, in these selected data, OA was predominantly SOA, with the contribution of POA represented by a minor and constant background concentration (POA_bg).

Subsequently, we employed a support vector regression (SVR) (Awad et al., 2015) trained on these rigorously selected SOA production periods. SVR was chosen for its generalizability in handling small datasets and its resistance to overfitting (Ghimire et al., 2022; Juang and Hsieh, 2009). Unlike tree-based models like random forest (Breiman, 2001), SVR model can predict continuous values (Ma et al., 2003; Tang et al., 2024). The hyperparameters, including the penalty coefficient (C) and gamma (γ) of the radial basis function (RBF) kernel, were tuned via grid research. The model targeted OA concentration, using nss-SO₄, MSA, NH₄, and meteorological parameters (temperature, relative humidity, boundary layer height, wind direction, and pressure) as predictors. We also included hours of the day to capture diurnal variations. Predictors not directly linked to secondary production, e.g. sea salt, NO₃, and BC, are excluded to avoid over-fitting and ensure generalizability, even though including these might have enhanced the model performance in the training dataset. Wind speed was used to select the SOA production period; therefore, it was not suitable as a predictor. A summary of the variables employed as predictors is shown in Table 1. The SVR was trained using the “tidymodel 1.3.0” framework using R programming software (version 4.4.3).

The selected SOA production periods were split by the year 2015: data before 2015 were used to train the model. The training dataset covered a significant fraction of the variabilities in predictors, as illustrated in Fig. S1 in the Supplement. The hyperparameters of the SOA-SVR model were optimized using grid search and 5-fold cross-validation. The data after 2015, which were unseen for the training process, were used to challenge the model's generalizability. Overall, there are 1700 h of SOA production periods and 477 h (27.8 %) after 2015.

A schematic diagram of the proposed methodology is shown in Fig. 1. Initially, the clean marine dataset was extracted by applying the clean sector criteria, followed by additional filtering processes to minimize the influence of POA. To assess the representativeness of the selected data as secondary sources, the Fuzzy C-Means (FCM) clustering method (Bezdek et al., 1984) was utilized. Subsequently, data were divided into training, validation, and test sets for ML, with SOA (including minor POA_bg) as the predictive variable. To mitigate experimental uncertainties, cross-validation and Monte Carlo simulations were performed. The study further investigated the magnitude of the POA_bg values. Finally, the ML method was applied on the clean marine air mass data to predict marine SOA concentrations, enabling the differentiation of marine POA concentrations within the total marine OA.

FCM is a clustering algorithm that enables the grouping of data points into multiple clusters with varying degrees of membership. Unlike traditional hard clustering techniques, where each data point is assigned to a single cluster, FCM assigns membership levels to each data point, indicating the degree to which it belongs to each cluster. This soft clustering approach is particularly useful when dealing with complex datasets where boundaries between clusters are not well defined or overlap significantly. By optimizing an objective function that minimizes the weighted sum of squared errors, FCM iteratively updates the cluster centres and membership degrees, providing a flexible and robust means of uncovering underlying patterns and structures within data. In this study, the input variables for the FCM model included chemical components (sea salt, OA, NO₃, SO₄, NH₄, MSA, and BC) and meteorological parameters (temperature (temp), relative humidity (rhum), wind speed (wdsp), and wind direction (wddir)). We randomly selected 10 000 samples and retained only those with positive concentrations for all chemical components. The data were then log₁₀-transformed and Z-score-standardized. The FCM model was configured with five clusters, a fuzziness exponent of 1.2, and Euclidean distance as the distance metric.

We firstly examined the aerosol composition in the selected secondary marine aerosol dataset. In these SOA production periods, nss-SO₄ constituted 68.9 ± 8.7 % of the PM₁ mass, followed by OA at 12.1 ± 3.4 % and MSA at 6.8 ± 4.0 %. The average concentrations of sea salt and wind speed were 0.015 ± 0.009 µg m⁻³ and 4.6 ± 1.2 m s⁻¹, respectively. The MSA–nss-SO₄ ratio was 0.10 ± 0.06, aligning with previous findings at the same site (Ovadnevaite et al., 2014). BC concentrations remained well below the 15 ng m⁻³ threshold, averaging at 6.0 ± 3.7 ng m⁻³. Taken together, this chemical composition indicates the selected data mainly originated from secondary sources.

To ensure the representativeness of the selected training data as secondary sources, we applied the FCM method on the clean marine dataset to identify the characteristics of chemical and meteorological parameters from typical sources. Compared to more conventionally used k-means clustering, FCM allows data instances to belong to multiple clusters with varying degrees of membership (or probability). The membership is used to determine how strongly each data instance belongs to each cluster. The FCM clustering, which is independent of the selection of SOA production periods, provides further validation and examination of the data selection for model training. We selected those with any cluster membership higher than 80 % to show the clustering centre of each cluster. As shown in Fig. S2a, the second factor, which is featured by low sea salt, low wind speed, and high temperature, is most likely to be of the secondary origin. This factor also showed high MSA and nss-SO_4, supporting the selection criteria for the SOA production period. Indeed, the highest possibility of the selected training data is found to be the second factor (Fig. S2b), reaffirming the SOA production characteristics.

Subsequently, we employed the SVR model, leveraging the clean marine dataset, to predict OA (SOA + POA_bg) concentrations. Cross-validation yields Pearson's R of 0.97 for training and validation datasets, demonstrating the model's accuracy in predicting total OA (SOA + POA_bg) concentration using the selected predictors. The slopes between estimated SOA + POA_bg and measured OA were 0.97 for both the training and validation data, indicating robust pattern recognition across most concentration ranges. The model also exhibited great generalizability and performed consistently well on an unseen dataset (Fig. 2c), with a Pearson's R value of 0.94 and a slope of 0.98 between observed OA and estimated SOA + POA_bg, reaffirming the model's efficacy in modelling the complex dynamics of SOA.

Permutation importance analysis highlighted nss-SO₄ and MSA as the most influential variables, followed by NH₄ (Fig. 3a). Partial dependence plots (Fig. 3b) indicate a nonlinear relationship between SOA and increasing levels of nss-SO₄ and MSA. While these plots do not imply a causal relationship, they highlight the complexity of the interactions and underscore the importance of employing a machine learning model to effectively capture such intricate patterns. Various meteorological parameters were also found to influence SOA concentration, especially relative humidity and precipitation, but to a lower extent. It should be noted that the nss-SO₄, although being found in marine secondary species, exhibits different formation dynamics and timescales with marine SOA, which might induce some extent of uncertainty. However, both marine SOA and nss-SO₄ might originate from marine biological activities. Furthermore, marine air masses arriving in MHD are expected to be advected over the Northeast Atlantic for several days. The use of nss-SO₄ can also be supported by the high correlation between nss-SO₄ and MSA, which exhibit different atmospheric formation dynamics.

In this study, one of the major assumptions of this approach is to assume that the OA in the selected secondary marine data for training the model is dominated by SOA. To evaluate the potential influence of POA contributions in those secondary marine data cases on the entire clean marine dataset, we conducted a sensitivity analysis. We assumed that POA_bg constitutes 5 % to 30 % of OA. As shown in Fig. S3, compared to the original assumption, the monthly averaged concentration of POA systematically increased throughout most of the year, except during winter, when alternative assumption predicted POA concentrations lower than 0, which is, of course, nonphysical. Then, we applied the model to the entire clean marine dataset and tried different fixed POA_bg values iteratively. As shown in Fig. S4, the POA_bg of 0.01 µg m⁻³ yields the fewest nonphysical predictions (either POA or SOA lower than 0). Therefore, the POA_bg of 0.01 µg m⁻³ was used in the following calculation. Based on this strategy, SOA concentrations were predicted, enabling the estimation of POA concentrations. Furthermore, if the POA production period is defined as a period with POA concentrations exceeding 0.1 µg m⁻³ for more than 12 h, there were more than 60 such POA production periods during the 10-year period (Fig. S5). Detecting these POA production periods allows detailed characterization, potentially enhancing its parameterization.

Note that the measurement itself contains relatively large uncertainty, compared to summertime measurement. Indeed, we manually tuned the POA_bg value to minimize the negative values. This is supported by previous studies that the marine area is a large organic pool (Quinn et al., 2014). Given the inherent uncertainties in aerosol measurements, which are well documented in previous studies (Ovadnevaite et al., 2014), we further quantified the uncertainties associated with the ML model using Monte Carlo simulations. To do this, we performed a robustness test by randomly validating the model 1000 times, each time excluding 20 % of the data from the training set. The lower and upper limits of the estimated POA seasonality are shown in Fig. S6, which is similar to the original model. The Monte Carlo ensembles demonstrated negligible differences in the contribution of POA, indicating stable model performance across different scenarios. While it should be noted that the Monte Carlo bootstrapping is used to assess the random uncertainties, potential uncertainties associated with possible systematic uncertainties require further investigation.

To validate the ML-based POA concentrations, we further compared them with PMF-based POA concentrations from Chevassus et al. (2025). The PMF-based source apportionment was conducted for about 1 month. The Pearson's correlation coefficient between ML-based POA and PMF-based POA was about 0.91, indicating strong agreement between the two methods (Fig. 4). Compared to the conventional AMS-based OA source apportionment techniques, e.g. PMF, this ML approach requires significantly fewer computational resources and is less dependent on detailed knowledge of the mass spectra signatures of marine POA and SOA. For example, the model performed equally well even after removing MSA as a predictor (Fig. S7). Given that MSA can only be resolved by high-resolution AMS, this suggests that our approach could be extended to Aerosol Chemical Speciation Monitor (ACSM) data, which are more affordable and widely used but have a lower mass resolution. This would enable broader applications of our method, offering a more comprehensive understanding of marine POA over global oceans. Finally, although secondary production of OA and nss-SO₄ relies on similar meteorological conditions, it should be noted that many marine VOC species do not share the same sources and oxidization pathways as DMS and its derivatives. For example, some VOCs are produced by different organisms or abiotically from the sea surface microlayer (Ciuraru et al., 2015; Mungall et al., 2017), which could introduce additional uncertainty in the SOA quantification based on periods of high nss-SO₄, which adds to the uncertainty for POA attribution as well.

The evaluation of the model's performance has shown a clear relationship between SOA and the predictor variables in the chosen SOA production period. Assuming there is little and constant contribution from POA_bg, the estimated SOA concentration should be similar to the measured OA. We then evaluated the model's performance in well-defined cases over finer timescales. During a typical SOA production period from 11 to 14 August 2011 (Fig. 4a), which was not included in the training dataset because of the slightly elevated sea salt above the threshold of 0.03 µg m⁻³, the measured OA closely followed the variation in nss-SO₄ and MSA, with Pearson's R values of over 0.80 and 0.84, respectively. These high correlations indicate a predominant secondary source of OA. The model's estimates of SOA concentrations were very close to the measured total OA (Fig. 4b), with an OA/SOA ratio of 1.03 ± 0.04. This further suggests that the ML model is able to predict the variability in SOA.

In contrast, during a well-documented marine POA plume, from 13 to 18 August 2009 (Fig. 5c), the OA dominated the PM₁ concentration and showed little correlation with nss-SO₄ or MSA (Pearson's R of 0.05 and 0.08, respectively). Notably, the estimated SOA deviates significantly from the observed OA (Fig. 5d). In this instance, the SOA estimated by the ML model accounted for only about 20 % of the total OA during the plume, underscoring the significant contribution of marine POA. The different performance, presented in Fig. 5, is expected. In Fig. 5a, the model is mostly influenced by the nss-SO₄, a marker for secondary species, leading to a great agreement between modelled SOA and observed OA. Conversely, Fig. 5d shows a significant discrepancy between modelled SOA and observed OA. The difference is largely attributed to the contribution of POA. This case was reported by Ovadnevaite et al. (2011a), in which the marine POA was identified using HR-ToF-AMS mass spectra and the SOA during the period was not quantified.

The cases of SOA and POA production periods indicate the model's capability to predict the SOA and POA levels. As illustrated in Fig. 6, both POA and SOA reached their peak in June and dropped to their lowest during the winter months. Typically, median concentrations of SOA were higher than those of POA across most months, except for May. However, POA concentrations spiked periodically, highlighted by outliers. The mean relative contribution of POA to total OA, detailed in Fig. 6b, shows the lowest contribution in winter and the highest contribution from May to July, peaking at approximately 50 % and slightly later in summer than SOA. The pattern corresponds with the enhanced marine activities in late spring and early summer of the North Atlantic, involving extensive phytoplankton proliferation and other marine organisms that release organic matter into the atmosphere through wave breaking and bubble bursting.

It is important to note that the total marine OA concentrations during winter at Mace Head are very low, introducing substantial uncertainty in OA separation during this season. The minimal POA concentration observed in winter suggests a distinct relationship between total OA, secondary species, and environmental factors. This relationship closely mirrors the dynamics observed in SOA. As shown in Fig. S8, the estimated SOA closely aligns with measured OA throughout the winter, pointing to low contribution from POA. Conversely, during the summer, numerous data points shown in Fig. S8 deviate to the right of the 1:1 lines, indicating a substantial contribution from POA.

The differentiation of POA from SOA is further substantiated by additional correlation analysis. As shown in Fig. S9, the correlation between OA and nss-SO₄ across the entire clean marine dataset is relatively low, at approximately 0.17, suggesting that nss-SO₄ explains less than 3 % of the variability in marine OA. Upon decoupling the OA into POA and SOA, we observed distinct correlation patterns: the correlation between SOA and nss-SO₄ increased to 0.88, reflecting a strong linkage, whereas, for POA, it decreased to 0.08. This stark contrast underscores the different sources and atmospheric behaviours of marine POA and SOA. This different correlation analysis provides a clear delineation of how POA and SOA contribute to marine OA and emphasizes the capacity of advanced modelling techniques and long-term observations to unravel complex atmospheric processes.

The potential sources of POA and SOA were investigated using the potential source contribution function (PSCF) combined with air mass backward trajectories (Mansour et al., 2020). As shown in Fig. 7, POA likely originates from the Northeast Atlantic polar marine regions (Fig. 7a), which are recognized as biologically active waters. In contrast, SOA sources are traced to tropical marine regions (Fig. 7b). The identified POA sources align with previous studies suggesting that regional marine biological activity is a key driver of POA production (O'Dowd et al., 2004; Sellegri et al., 2021). For SOA, enhanced photochemical reactions in lower-latitude waters likely promote the formation of secondary species. The distinct source regions of marine POA and SOA underscore the need for models to incorporate specific parameterization schemes that account for these spatial and mechanistic differences.

In contrast to prior studies that relied on filter-based measurements with limited temporal resolution, this study introduces an ML framework to systematically differentiate and quantify marine POA and SOA. While seasonal variations in POA/SOA have been reported previously, our decade-long dataset – the most extensive of its kind to date – provides unprecedented resolution to constrain and develop POA and SOA parameterization for climate models. Furthermore, the distinct source regions identified for POA (polar marine zones) and SOA (tropical waters) underscore their divergent formation mechanisms. Current model estimates of global emissions of POA span 6.9–76 Tg yr⁻¹ for < 1 µm emissions, and the global source of SOA was thought to be substantially smaller than marine POA. Our measurements found similar contributions of POA and SOA, while it has to be noted that some fraction of SOA was transformed via atmospheric ageing from POA, which is difficult to quantify. A recent study found the photochemical reactions in the sea–air interface produced substantial VOC as the precursors of marine SOA (Brüggemann et al., 2018); the complex sources of SOA highlight the need for field observational data to challenge the models. The combined sensitivity to marine biological activities and photochemistry of marine OA was also supported by Sanchez et al. (2021), who found high correlation with downward shortwave flux and net primary production, although they did not separate POA and SOA.

The long-term marine POA and SOA time series now enable an assessment of their influence on aerosol hygroscopicity – a relationship which was previously uncertain. As shown in Fig. 8a, increasing POA concentrations from 0.1 to 1 µg m⁻³ reduces the hygroscopicity parameter (κ) across all particle sizes, dropping values from ∼ 0.5 to below 0.25. This finding contrasts with earlier experimental work suggesting POA production has negligible effects on aerosol hygroscopicity and CCN activity (Quinn et al., 2014) but aligns with our prior case study highlighting the inherently low hygroscopicity of POA (Ovadnevaite et al., 2011b). Quinn et al. (2014) investigated the impact of marine POA on its cloud condensation nuclei using a sea sweep device; they found no significant difference in POA contribution at different oceanic regions with diverse ranges of chlorophyll a, and they attributed the POA to the ocean carbon pool. However, based on our field measurements, it is unlikely that the ocean carbon pool induces such large variations in the observed POA, underscoring the importance of oceanic biological activities.

As for the SOA, the increasing SOA concentrations from 0.1 to 0.3 µg m⁻³ only slightly reduce κHTDMA values (from 0.5 to 0.45) for particles between 50 and 165 nm, with no significant change observed for 35 nm particles. This muted response may arise from co-varying increases in secondary species, such as nss-SO42- or MSA, which help maintain hygroscopicity at relatively high levels. This is consistent with our recent high-temporal-resolution online measurement, which shows a simultaneous increase in SOA and nss-SO42- during particle growth (Xu et al., 2024; Zheng et al., 2020). These results underscore the distinct roles of POA and SOA in modulating aerosol water uptake and cloud-forming potential, emphasizing the need to explicitly represent OA composition and sources in climate models.

The influence of POA and SOA on the aerosol mixing state was further investigated using the spread factor, calculated from growth factor probability density functions (Xu et al., 2020b). A spread factor of 0 indicates a theoretically internal mixture, while higher values reflect increasing external mixing. Based on established thresholds (Swietlicki et al., 2008; Xu et al., 2020a), a spread factor of ≤ 0.05 is classified as an internal mixture, whereas values ≥ 0.2 signify external mixing. POA and SOA exhibit divergent impacts on the aerosol mixing state. As shown in Fig. 9a, increasing the POA contribution from 0 % to 100 % elevates the spread factor, suggesting POA production promotes external mixing. Conversely, SOA accumulation drives the system toward a more internally mixed state. This aligns with aerosol ageing processes, where particles tend to homogenize over time. Accurately representing the hygroscopicity and mixing-state dynamics of POA and SOA is critical for assessing their climatic impacts, as these properties directly influence aerosol–cloud interactions and radiative forcing.