We used a Multiparameter Bioaerosol Spectrometer (MBS; University of Herefordshire, UK), a UV-LIF spectroscopy device developed for detecting bioaerosols in ambient air . The MBS operates at a fixed instrument flow rate set by the user, typically between 1–2.5 Lmin-1, but particle measurement frequency is constrained by coded triggering sequences. A fixed fraction (0.165) of the total flow is drawn as sample flow with the remainder serving as sheath flow. In the measurement stage, each particle passes sequentially through a low-power laser beam for sizing and a high-power pulsed laser beam for diffraction pattern measurement, and is then exposed to a xenon flash lamp emitting 280 nm light to excite fluorescence. Light scattered from the sizing laser is detected using a photomultiplier, and the particle's size is estimated using Mie scattering theory. With its optical configuration, the effective particle size range for measurement with the MBS is 0.5 to ∼ 20 µm. The high-power laser emits a 10 µs pulse of light for each particle; light scattered by the particle is directed through a beam splitter to two 512-pixel complementary metal oxide semiconductor (CMOS) arrays positioned on the left (L) and right (R) sides of the beam. These arrays measure chords across the particle's diffraction pattern, providing optical scattering information, serving as proxies for morphology. The arrays are offset from the beam centerline by a distance set for the instrument's target particle size range. Consequently, the morphology metrics derived from diffraction are inherently size-dependent because diffraction pattern scales vary with particle size, whereas the array spacing is fixed. Fluorescence emitted after excitation is collected by a hemispheric mirror, limited to exclude < 305 nm light with a long-pass filter, dispersed using a diffraction grating, and measured in eight acceptance intervals (channels, labeled A-H) with roughly equidistant central wavelengths (315, 364, 414, 461, 508, 552, 595, and 640 nm, respectively) spanning a detection region of approximately 305–655 nm (Fig. S1 in the Supplement). Fluorescence intensity is not measured in physical units; we report fluorescence intensities as fractions normalized by detector signals. Fluorescence signals in individual channels may saturate if intensities exceed the detector's upper sensitivity limit. The maximum particle measurement rate is also limited by the xenon lamp's recharge time, yielding an upper rate of about 160 particless-1 , enough to characterize most particles in even high ambient coarse-mode particle concentrations.

The MBS periodically measures background intensities sensed by each detector, i.e. that of the optical stage after a pulse of excitation light without sample flow, with a series of 42 xenon lamp flashes once when beginning the measurement and then typically after 30 000 samples . This provides an estimate of the instrument's baseline detector intensity spectrum, which is subtracted from each subsequently measured particle's measured spectrum. Fluorescence intensity can then be expressed in terms of standard deviations (σ, determined by the spread in background measurements) above the background mean. Because the MBS's sampling efficiency drops below 50 % for sizes < 0.8 µm , we exclude all particles smaller than this from our analysis. Particle morphology parameters are calculated by the MBS software from the two CMOS array scattering signals. These parameters distill characteristics of particle scattering properties into statistical and shape descriptors. For each array, the software calculates the mean and variance of signal intensity, as well as the skew and kurtosis (treating each detected curve as a normal distribution). It also locates peaks using a user-defined local peak-trough threshold and derives peak width (width at half maximum of the strongest peak), peak-to-mean ratio (PTMR; also based on the strongest peak), and peak count for each array. Mirror symmetry, defined as the similarity between the bottom and top halves of the signal, is calculated for each array and given as a percentage. Because particles align with the flow and have arbitrary rotational orientations in the detection stage, we report array-specific parameters using an arbitrary array of choice (right; R) for visualization. Two additional asymmetry parameters are provided: left-right (L-R) pixel-wise asymmetry, and pixel-wise asymmetry with one array inverted, both calculated as a percentage.

developed a heuristic decision tree (DT)-based classification of particles measured by the MBS based on fluorescence, which was later applied and validated in subsequent studies . In this classification scheme (Fig. 2 in ), particles are classified as fluorescent (3–9σ fluorescence in any channel), highly fluorescent particles (HFPs; > 9σ signal in any channel), and fluorescent PBAPs (fPBAPs), defined as HFPs with a fluorescence maximum in the second (B) channel. It is well documented that PBAPs, especially bacteria, commonly show pronounced fluorescence around 350–400 nm due to fluorophores such as tryptophan . This feature underpins the MBS's design and is widely used in UV-LIF-based PBAP detection (e.g. ). In this study, we used this DT-based method as a benchmark for PBAP identification. As an additional metric of the shapes of fluorescence spectra, we calculated a fluorescence ratio defined as the sum of channels A and B divided by the sum of channels C-H, with higher values indicating a stronger peak in the earlier channels.

The samples characterized with the MBS in this study were selected to represent common sources of aerosols potentially relevant in many environments, including tree pollen, marine bacteria, dust, and synthetic contaminants (Table ). Cellulose, a major component of plant cell walls, was chosen as an analog to plant matter and as a constituent of pollen (e.g. ). We used pure, synthesized crystalline cellulose (Sigma Aldrich, cat. no. 435236).

Pollen were sourced from various locations in the Czech Republic by Pharmallerga CZ s.r.o. (Lišov, Czech Republic). The species of trees from which pollen was sourced for this study were ash (Fraxinus excelsior), black alder (Alnus glutinosa), birch (Betula pendula), hazel (Corylus avellana), juniper (Juniperus communis), Scotch pine (Pinus sylvestris), and willow (Salix caprea). Pharmallerga assessed pollen quality by microscopy and with acetic blue staining; pollen processing and size ranges they reported are listed in Table S1 in the Supplement. Pollen was collected by hand, with the exception of juniper, collected with vacuum, the spring/summer before the experiments took place (January 2022). All samples were processed dry (sifting and drying), with the exception of willow, which was defatted with acetone. Samples were stored cold (< 8 °C) in centrifuge tubes at the supplier prior to dispatch and in a freezer (∼ -20 °C) thereafter; experiments were conducted over the next two months. To investigate whether aerosolization method affected MBS-observed properties, we used both dry and wet nebulization.

Bacterial samples were mixed cultures of cyanobacteria from Baltic seawater. We selected three cultures with distinct natural pigment profiles (phycobiliproteins and carotenoids) to capture diversity, but we did not identify individual species. These pigments – phycobiliproteins (phycoerythrin, phycocyanin, and allophycocyanin) and other compounds (e.g. carotenoids) – absorb in different spectra of light and differ from chlorophyll A absorption , allowing them to be differentiated from algae.

Ground polyethylene (PE) was chosen as a representative microplastic. Because environmental microplastics undergo weathering processes that may alter their optical and fluorescence properties, we measured freshly ground PE particles, with one subsample exposed to UV light to simulate photolytic aging. While we compare the relative effects of aging in our experiments, we do not attempt to quantify physico-chemical changes resulting from UV aging as the MBS is not suitable to this end. Measured quantities reported here are thus not expected to be directly comparable with those found in other experiments using more exact methods.

Natural dust samples collected in the field from four sources, Dyngjusandur and Myrdalssandur in Iceland, Svalbard, and Sakurajima in Japan, were included as potential high-latitude dust sources relevant to the Arctic. Dyngjusandur and Myrdalssandur samples contain primarily volcanic silt and sand , whereas the Sakurajima sample is primarily volcanic ash. The Svalbard sample, previously characterized for its ice nucleation activity in (sample BR1607b), originates from glacial outwash sediment. In addition, kaolinite clay was selected as a sample of mineral dust reference expected to contain minimal organic matter.

For dry, solid materials, 5 mL of the sample was placed in a centrifuge tube mounted on a speaker (Fig. S2b in the Supplement) and vibrated at a frequency selected using a function generator. The optimal frequency depended on the sample type and nebulization efficiency, but stable, high particle concentrations were typically achieved at 15–30 Hz. A commercial nebulizer (Topas GmbH, Germany, model ATM228) with a fixed nebulization pressure of 100 hPa (gauge) was used for wet and liquid samples, including pollen solutions. For wet nebulization using the TOPAS nebulizer, the sample air stream was dried using a Nafion dryer, while the sample air from the dry (speaker) nebulization was not actively dried (see Fig. S2). The sample air relative humidity was monitored using an inline RH/temperature sensor (Hytelog USB, B+B sensors, Germany) immediately upstream of the MBS to ensure particles were measured dry. Each sample was nebulized and measured for at least 10 min. Background tests using empty containers (shaken empty centrifuge tubes for dry samples, nebulized pure Milli-Q water for wet) showed negligible background concentrations, far below those seen in the characterization experiments (∼ 10⁰ L-1). Samples were measured in sequence from those least to most likely to adhere to the instrument and plumbing to minimize carryover between experiments.

To compare our characterization of specific aerosol types to previous work, we include laboratory data from , , and , as well as data from known pollution events in and . The pollution events were observed during an Arctic Ocean research cruise when the ship's exhaust plume was transported directly to the inlet from which the MBS was sampling. Because the ship operated in very remote locations (> 80° N) and particle concentrations during the pollution events exceeded background concentrations of other fluorescent particles (e.g., PBAPs) by several orders of magnitude, we consider these observations representative of fossil fuel combustion sources, despite being measured in ambient air. The sea spray aerosol (SSA) measured by were generated with a plunging-jet sea spray simulation chamber of the type described by and . We include data from our characterization experiments and these external sources in our ML training and classification algorithm (see Sect. ). This extends the classes of particles to include fungal spores and SSA , and increases diversity in bacteria, pollen, and dust samples . In addition, we include data from polystyrene latex spheres (PSLs) measured for calibration in ; these represent particles with well-constrained optical and morphological properties (spherical and opaque) and provide another sample to the microplastic class. Table  lists source type, used nebulization methods, and citations for all data considered in our study, including our own characterization experiments.

Information from the MBS can be used to train particle identification algorithms on laboratory characterization data; such approaches using ML have previously been applied to measurements made with the MBS and other UV-LIF spectrometers . Here, we developed a new and improved classification algorithm for MBS data with three tasks, in order of priority: 1.

flag fluorescent particles that are likely interferents, so they can be excluded from PBAP concentration estimates;

The methods used for each component model are described below, with their implications discussed in the results (Sect. ). Table S2 in the Supplement summarizes the MBS data input variables and their roles in the classification algorithm.

We treated tasks 1 and 2 as binary classification tasks and used logistic regression modeling (LRM; ) for both. LRMs fit logistic functions to training data to estimate the probability that a new observation belongs to class 0 or 1, and can accommodate many input variables. This approach is well suited to task 1, flagging observations as interferents or non-interferents in a pollution model, and to the dust vs. non-dust decision in a dust model for task 3. The model output is a probability estimate between 0 and 1, obtained by averaging logistic function outputs across input features, which we name pollution-likelihood and dust-likelihood for the pollution and dust models, respectively. Positive identification can be determined by rounding this probability or by choosing a threshold, which can adjust identification confidence. These LRMs were trained using least squares error (L2) loss functions and the limited-memory Broyden–Fletcher–Goldfarb–Shanno (LBFGS) solver .

Dimensionality reduction is useful for quantifying and presenting differences between observations with a high number of features , such as the many fluorescence and morphology metrics measured by the MBS. Incorporating morphology metrics into particle classification can help prevent overly strong influence by fluorescence information, which may not be enough to distinguish particle types. Dimensionality reduction can be used to rapidly map similarities with known particles and as a pre-processing step for other ML methods, or as its own classification method (e.g., ) in certain cases. For task 2, we used uniform manifold approximation and projection (UMAP), a technique for approximating a topological manifold on which data with any number of dimensions are located and quantifying their relative proximities in a lower dimensional space. UMAP approximators are created by iteratively considering n neighbors in the dataset and estimating proximities between members, projecting the graph in e.g. two-dimensional space, and ensuring that connections between individual points in the original data are preserved . Our UMAP approximators were made using the UMAP-learn Python library; details about the package's methods and underlying theory are described in . For training, we used 100 neighbors, set a minimum separation between transformed points of 0.15, and provided training class labels for clustering. The approximator's outputs are the transformed data positions in two arbitrary dimensions, here defined as UMAP1 and UMAP2, similar to other non-metric multidimensional scaling methods (e.g. ).

After UMAP transformation, we used k-nearest neighbors (kNN) to classify data according to their positions in the transformed space relative to the positions of particles of known types from source characterization data. In kNN, observations are compared to the k closest neighbors in the training data and classified by votes defined as the number of neighbors belonging to each class. The kNN outputs a class identification corresponding to the class with the most votes and the percentage of votes among the k neighbors cast for each class. Because kNN classifiers generally struggle with high-dimensional data, the use of UMAP both simplifies class labeling (compared to other multiclass methods such as random forest) and provides an additional clustering step for more robust identification.

To assess our classification algorithm, we applied it to MBS observations from Zeppelin Observatory, Svalbard made over 2020 during the Ny-Ålesund Aerosol Cloud Experiment (NASCENT) campaign . This reproduces the analysis of , who applied and further validated the fluorescence-only DT-based approach that has since proven successful in other studies made with the MBS . Repeating this published analysis with our new classification algorithm allows a direct assessment of its strengths and weaknesses. Additional data for assessment were obtained from 24 h to weekly resolved filter sampling , except for equivalent black carbon (eBC), which was obtained with the Multiangle Aerosol Absorption Photometer (MAAP) at minutely resolution averaged to daily values . Sampling frequencies and references for the species (tracers) used for the assessment here and in are summarized in Table S3 in the Supplement. Both classification algorithms were applied to MBS data collocated with the tracer sampling windows. Particle class concentrations were calculated as the number of observed particles assigned to each class divided by the MBS sample air volume during that period. Sample volume was determined from the sampling duration and the constant sample flow rate of 0.33 Lmin-1 reported in .

We used chemical tracers for dust (silicon, iron, potassium) , PBAP (fructose, glucose, arabitol, and mannitol) , SSA (sodium, chloride, magnesium, calcium) , anthropogenic aerosol (eBC), and biomass burning (eBC, levoglucosan) . Fructose and glucose are monosaccharides used as PBAP tracers . Glucose is ubiquitous across organisms, whereas fructose is a photosynthetic product rapidly metabolized by non-plant organisms and is thus more characteristic of pollen and plant matter . Arabitol and mannitol are sugar alcohols that serve as energy reserves and, owing to their abundance in fungi, are considered as tracers for fungal spores , and have been used to verify UV-LIF-based fungal spore detection . Levoglucosan forms from the pyrolysis of cellulose and is an established tracer for biomass burning (BB) . eBC is emitted by both fossil fuel and biomass combustion and is thus used to trace the influence of all combustion aerosols .

Having simultaneous knowledge about particle composition from detailed tracer data also affords the possibility of directly implementing lessons from the assessment in tuning the classification algorithm, a process known as transfer learning . Domain adaptation is a transfer learning technique for ML wherein a model trained in one domain (here, in laboratory settings) is adapted for use in another (field observations), which helps overcome challenges introduced by differences between the domains (e.g. differences in particle properties between naturally and laboratory-generated aerosol). We implemented domain adaptation in our classification algorithm based on the results of the assessment by using tracer concentrations to construct probability estimates of observing targeted particle types, which were used to further train the model in a controlled manner. The details of the domain adaptation implementation are further explained alongside our assessment of the initial classification algorithm in Sect.  and in the SI (Sect. S1 in the Supplement).

Emission spectra and size distributions measured in dry and wet pollen experiments are shown in Fig. . The particles we detected were far smaller than the whole pollen grain sizes reported by the supplier (Table S1); our results therefore represent fragments produced by the rupture or fragmentation of intact pollen grains. Moisture and impaction are known to induce fragmentation in laboratory experiments and in ambient air, potentially providing a secondary source of PBAPs with a greater transport range and residence time than intact pollen that may affect key atmospheric processes . Although their ambient detection remains difficult, observational evidence of strong variability in pollen fragment emission and concentrations under humid conditions has been found using UV-LIF . While our data should aid pollen fragment identification in field observations, it is important to note that intact pollen grains lie outside the MBS's detection size range and cannot be measured with this instrument. Additionally, rupture and/or fragmentation may have been promoted during sample transport and storage.

Most dry pollen fragments peak in the C channel (∼ 414 nm), with ash, juniper, and pine having slightly broader spectra (Fig. o–u). Wet nebulization consistently broadens the emission spectra of pollen fragments for willow pollen (Fig. u) and shifts fluorescence peaks to channel D (∼ 461 nm) for birch, hazel, juniper, pine, and willow pollen (Fig. q–u). In general, dry pollen fragment fluorescence is dominated by fluorophores on the grain exterior, composed mainly of sporopollenin , whereas pollen fragments can include interior fluorophores, such as phenols, carotenoids, azulene, and anthocyanin, all with different emission characteristics . Wet nebulization produced more numerous, smaller particles (Fig. h–n) than dry nebulization (Fig. a–g), indicating a different emission mode, likely including soluble material re-formed as residue after drying. These particles show more regular, symmetric, droplet-like scattering signatures (Sect. ). Thus, the aerosolization method clearly affects both fluorescence and morphology, with implications for ambient detection. The especially strong wet-dry contrast in fluorescence seen with willow pollen (Fig. g and n) may be due to acetone de-fatting (Table S1), as removing surface lipids likely increases water uptake and solubility, promoting rupture and altering fluorescence.

Figure a–c show that fluorescence spectra recorded for bacteria are dominated by a B channel peak (∼ 315 nm), consistent with the known region of fluorescence emission by tryptophan when excited at this wavelength , the findings of , and characterizations of bacteria with other UV-LIF instruments (e.g. ). This feature is unique among our samples and provides a robust marker for fPBAP identification. In contrast, PE, both fresh and UV-aged, and cellulose (Fig. d–f) constitute potential interferents. PE spectra (Fig. j and k) closely resemble those of dry pollen fragments with C channel maxima (Fig. o–u), and cellulose shows similar pollen-like fluorescence (Fig. l). While atmospheric coarse-mode microplastic concentrations are likely only high enough to significantly bias fPBAP detection in specific locations with strong anthropogenic activity, such as roads or urban centers , these results show that synthetics can mimic the fluorescence properties of fPBAPs. Cellulose-fPBAP similarity (Fig. f and l), however, is relevant for the atmosphere because cellulose-containing plant debris may be abundant . Being PBAPs themselves, they may be less of a complicating factor in UV-LIF detection, although this depends on classification precision requirements. Our pure cellulose spectra agree with published measurements of cellulose , and the C channel peak may contribute to pollen fluorescence signals because the pollen intine is cellulose-rich and can be exposed during fragmentation . Among all fluorescent samples measured here (Figs.  and ), fluorescence intensity increases with size, as expected from the larger amount of emitting material, implying that fluorescence contrasts among smaller particles are inherently muted compared to larger particles.

Here we present the results of all source characterization experiments and MBS observations used in this study (Table ), processed as described in Sect. . Figure  summarizes fluorescence emission spectra for all fluorescent materials, and Fig.  presents class contributions and select morphology parameters. Pollen fragments measured in our experiments (Fig. b–o) and those of (Fig. p–s) show broader spectra and higher intensities than bacteria (Fig. t–x). Despite containing many HFPs (Fig. a), only a small fraction (∼ 0.01 %–0.1 %) of particles from pollen samples are classified as fPBAP by the DT method. This indicates that a substantial share of pollen-derived particles would be missed by fluorescence-only schemes, but the variable peak channels (Figs.  and b–s) make it difficult to define a robust single-channel classification criterion. Natural dust measured in our experiments and the filtered seawater measurements in also contained low fractions of HFPs (< 0.1 %), indicating some organics or potential PBAPs present in these samples.

Bacteria samples from and exhibit the same B channel peak and similar fluorescence intensities as our bacteria samples, reinforcing the robustness of the DT-based fPBAP criterion of . Fungal spores fluoresce more strongly than bacteria and peak in either channels B or C, meaning some fungal spores will be missed by the DT method. We find that PSLs (Fig. cc) closely resemble bacterial emission spectra and all highly fluorescent PSLs meet the DT definition of fPBAP (Fig. a), demonstrating that synthetic particles can mimic bioaerosols in UV-LIF measurements. We also find that the diesel combustion particles from and exhibit broad, intense fluorescence (Fig. dd) and that about 10 % of them meet the DT-based fPBAP criterion (Fig. a), consistent with interference by combustion particles noted in . Because combustion particles from fossil fuel and BB are abundant in the atmosphere, their fluorescence similarity to bioaerosols is a primary concern for PBAP quantification. Using multiple excitation wavelengths can add discriminating power to UV-LIF instruments; in studies employing the WIBS, particles excited at a different wavelength than that of the MBS were shown to correlate with combustion tracers (e.g., ). Despite higher spectral resolution in fluorescence emission detection, the MBS is limited in this regard, motivating a combined fluorescence-morphology classification approach.

Beyond fluorescence, the MBS provides extensive optical scattering-based morphology information (Fig. b–f), which offers additional leverage against non-biological interferents. Since optical morphology signals are stronger with larger particles, we visualize morphology parameter distributions for < 3 and ≥ 3 µm particles. Importantly, interpretation must account for different emission mechanisms for various bioaerosols. Bacteria and many fungal spores fall within the MBS's size range and may be detected intact, whereas pollen rupture and plant debris emission generate irregular fragments with highly variable shapes. Fungal spores can also fragment under prolonged exposure to high humidity . Accordingly, pollen fragments, and to some extent fungal spores, may be expected to have more irregular morphologies whereas bacteria and intact spores may be expected to exhibit more regular scattering signatures; these behaviors are reflected in their measured properties (Fig. c–f). Combustion particles, aside from their high peak counts, exhibit scattering signatures indicative of irregular morphologies. This may help to distinguish them from bacteria and fungal spores but not necessarily from pollen fragments. Consistent with the discussion in Sect. , wet nebulized pollen fragments have more sphere-like morphologies than dry nebulized pollen (seen in e.g. their asymmetries; Fig. c), indicating that many particles are dried residues of dissolved constituents.

Figures c–f and S5 in the Supplement compare dust, nascent SSA , and PSLs . Dust and SSA particles clearly differ in mean signal (Fig. d), peak count (Fig. f), variance (Fig. S5e), and peak width (Fig. S5i), even at smaller sizes. L-R asymmetry is broader and more extreme for dust, although overlap remains (Fig. c). Surface roughness is reflected in signal mean, PTMR, and peak count (Fig. d–f, respectively), as well as in variance, kurtosis, and peak width (Fig. S5e, g, and i, respectively). Rough surfaces smear diffraction patterns, raising scattering signal mean and variance, reducing the number of distinct peaks, and increasing peak width and PTMR. Our dust samples show lower peak counts and higher signal mean and variance, PTMR, and peak widths than SSA and PSLs, indicating rougher, more irregular surfaces. Dust also shows higher and broader skewness distributions (Fig. S5f), consistent with irregular shapes. Strong smearing by diffraction on rough surfaces can increase mirror symmetry, which may explain the relatively high mirror values seen in dust (Fig. S5c). Optically, pollen fragments and fungal spores have irregular scattering signatures resembling dust with low peak counts and high asymmetries, mean signals, and PTMRs. Meanwhile, bacteria optically resemble SSA and PSLs, consistent with smooth surfaces.

In summary, dust exhibits high surface roughness and shape irregularity, whereas SSA shows sharper diffraction features and lower asymmetry. These features are consistent among all dust samples measured here, with small differences in size distributions (Fig. b). Dust sample fluorescence varies (Fig. a), possibly indicating the presence of fluorophores such as humic substances . Some of the consistency in measured dust features may derive from the generation method, which can bias certain particle sizes and alter their properties . SSA morphology is known to vary with size , source conditions such as productivity and organic content , wind speed at emission , and atmospheric processing . However, the SSA samples considered here are freshly generated, representing nascent particles; how aging affects MBS scattering and fluorescence signatures thus remains an open question.

To summarize the comparisons described here and illustrated in Fig. , common coarse-mode aerosol sources can produce particles with distinct, consistent sets of features when measured by the MBS. Among biological particles, bacteria displayed prominent tryptophan-like early channel-heavy fluorescence spectra; while fungal spores shared early channel-skewed fluorescence spectra, they exhibited stronger fluorescence in later channels and variable peak channels. Pollen fragments and pure cellulose had broader fluorescence spectra easily distinguishable from bacteria. The fluorescence spectra of freshly emitted combustion and microplastic particles shared similarities with those of all biological particles measured; although microplastic particles fluoresced at similar intensities to most biological particles, the combustion particles fluoresced more strongly despite their small sizes. Morphologically, bacteria stood out from fungal spores with smaller sizes and weaker scattering signals with stronger peaks. While this can help distinguish between them, combustion particle scattering signals resembled those of bacteria. Dry-nebulized pollen fragments were generally larger and had scattering signals indicative of rougher surfaces and higher asymmetry than those of wet-nebulized pollen fragments. Among non-biological particles, the lab-nebulized dust and SSA particles analyzed here had distinctly different optical morphologies, with SSA particles imparting strong diffraction peaks indicative of crisp, sharp edges and dust particles having stronger, non-peaked scattering signals indicative of rough surfaces.

Using the source characterization data presented in Sect. , we built classification models as a basis for an algorithm for (1) distinguishing fPBAPs from interferents among fluorescent particles, (2) attributing them to broad fPBAP classes, and (3) distinguishing between non-fluorescent dust- and SSA-like coarse-mode particles. For (1), we considered only combustion particles as interferents due to limited synthetic polymer sample diversity in our training data and because they are likely far more abundant than microplastic particles in the atmosphere. This simplifies the task to binary classification (pollution or not). For (2), we divided fPBAPs into three classes: pollen fragments, bacteria, and fungal spores; cellulose is not represented because real plant debris was not included in the source characterization data. Binary classification was also used for (3) to estimate dust- or SSA-likeness among non-fluorescent particles. These components were implemented in the classification algorithm presented in Fig.  (given in written form in Table S4 in the Supplement), which also presents the DT method of for comparison.

In our algorithm, particles are first filtered according to fluorescence; highly fluorescent (≥ 9σ) particles of all sizes are passed to the pollution/fPBAP identification branch, while the remaining particles with sizes ≥ 2.5 µm are passed to the dust/SSA branch. Following the discussion on Fig. , we trained the dust model only on dust and SSA particles ≥ 2.5 µm to ensure more distinct separation between their scattering signatures, as significant overlap in morphology signal distributions among particles of different classes are seen below this size range. Particles in the fluorescent branch are first assigned a pollution-likelihood (ppol) by the pollution model, and are then fed to the UMAP approximator. UMAP outputs are passed to the kNN classifier, which provides class likelihoods for four target classes (pollution, pollen fragments, bacteria, and fungal spores). A fluorescent particle's ppol must not exceed a confidence threshold (αpol), which may be chosen depending on confidence requirements and auxiliary data in each application setting, for it to be classified as a fPBAP. Thus, the algorithm uses a two-step identification process for fPBAPs where a fluorescent particle must be both not flagged as an interferent by the pollution model and positively identified as one of the fPBAP classes by the multiclass model. The multiclass model's output also allows a further confidence threshold for fPBAP identification to be chosen based on the kNN's probability estimates. Particles ≥ 2.5 µm in the non-fluorescent branch are assigned a dust-likelihood (pdust) by the dust model. This may be used for identification based on a confidence threshold (αdust) or as a metric itself (“dust-likeness”). In this assessment, we used the most likely class (the rounded probability for LRMs, i.e. αpol,αdust = 0.5, and the class of highest probability for the kNN) for identification.

As inputs for the pollution LRM and UMAP approximator, we used 34 features (Table S2), including: normalized fluorescence spectra, fluorescence ratio, total fluorescence, size, L-R asymmetries (top-to-bottom and inverse), and scattering parameters for L and R arrays (nine each). The UMAP approximator was fitted only on highly fluorescent combustion, pollen fragment, bacteria, and fungal spore particle data, and excluded cellulose and microplastics; the limited diversity represented in our data for these latter two classes would likely affect the generalizability of a ML-based approach. However, UMAP transformations were calculated for microplastic and cellulose particles to quantify their similarities with the other classes. After transformation, fluorescent particles can be classified according to their proximity in UMAP phase space to clusters seen in the training data. We trained a kNN classifier on UMAP-projected source characterization data, which uses a brute force global search of training data positions in UMAP space to select the 100 closest neighbors (k=100) according to Euclidean distance. The probability that the unknown particle belongs to each class is the fraction of neighbors belonging to them, weighted inversely by distance.

Table  presents confusion matrices for each ML model included in the algorithm. The pollution LRM performed well on testing data, identifying combustion particles as pollution with 92 % precision. Figure a shows UMAP projections for fluorescent particles using only pollen fragments, bacteria, fungal spores, and combustion particles for training, illustrating clear separation between the combustion particles and all fPBAP classes. The positions of particles in classes unknown to the UMAP (microplastics and cellulose) are distributed among the three fPBAP classes, indicating similarities shared with each of them. Few microplastic particles appear near the combustion particle cluster, and cellulose particles are mostly positioned near the pollen cluster. These relations between particle clusters in UMAP space reflect their similarities discussed in Sect. , indicating that the UMAP approximator effectively separated clusters of broad fluorescent particle classes and captures similarities between them.

Figure b shows UMAP transformations of the testing data and illustrates that the approximator also produced the general distributions and locations of class clusters seen in Fig. b with unknown particles. Lower certainty can be expected where significant overlap exists between clusters, manifested in lower kNN prediction probabilities at midpoints between clusters. Most fPBAP incorrectly classified as pollution were pollen fragments and fungal spores, and combustion particles that were wrongly classified as fPBAP were exclusively labeled as pollen fragments; the UMAP-kNN classifier model misclassified combustion particles as pollen fragments ∼ 11 % of the time (Table ). This reflects similarities between combustion and pollen fragments discussed in Sect.  (i.e., their broad fluorescence curves and irregular morphologies). Among fPBAP classes, the UMAP-kNN classifier was weakest at identifying bacteria, and significant confusion between certain fungal spores and bacteria were seen; ∼ 16 % of bacteria particles were misclassified as fungal spores. This may be expected given their similarities in spectral maxima, and because the UMAP-kNN method incorporates morphology information that also determines classification.

For the dust model, we trained a LRM on 22 morphology parameters (top-to-bottom and inverse L-R asymmetries and the nine chord shape parameters for each side; see Table S2). We disregarded particle size in the dust LRM to reduce the influence of biases introduced to the training data by laboratory aerosolization methods via their size distributions, and focused the method of distinction on distributions of optical scattering properties. Equal samples of 6000 particles each were randomly drawn from this size range. The LRM converged quickly (23 iterations) and testing resulted in 99 % precision for both classes (Table ), indicating that the distributions of optical morphology parameters are distinct enough to accurately distinguish them. Training the same LRM using all particles ≥ 0.8 µm resulted in a slight decrease in the model's overall precision (95 %). Despite this, we argue that training particles in a size range where the instrument better resolves optical scattering signatures produces a more physically sound model and elected to use the LRM trained only on ≥ 2.5 µm particles in our classification algorithm.

Figure  presents the results of our classification algorithm applied to MBS observations made by at Zeppelin Observatory during 2020, again using simple majority classification thresholds for pollution and dust (αpol,αdust = 0.5). The algorithm in its raw, untuned form, trained only on source characterization data, struggled to recognize the types of interferents seen at Zeppelin Observatory, as evidenced by the high correlations between fPBAP concentrations and levoglucosan and eBC (Fig. S6b in the Supplement). That fPBAP concentrations derived from the untuned model correlated more strongly with levoglucosan than with eBC may suggest an influence by specific combustion particle types, i.e., BB- rather than fossil fuel-derived. Biological tracer and DT-based fPBAP concentrations indicate a clear seasonal cycle of biological aerosol influence (Fig. a), which our algorithm initially failed to capture in its untuned form (Fig. b). This resulted in likely overestimated fPBAP concentrations, especially evident between January–April when eBC concentrations were high.

To address this, we applied domain adaptation to tune the pollution LRM to HFPs observed at Zeppelin Observatory using tracer-based soft labels describing the degree of pollution vs. biological influence. This is accomplished by choosing a selection of sample periods with overlapping tracer data for both pollution and biological aerosols, attributing “soft” (i.e. continuous and non-discrete) pollution labels for HFPs observed during these periods based on tracer concentrations, and further training the LRM on these data (the tuning block) with lower weights than those of the initial training data to limit its influence on model fit. Details of this step are explained in the SI (Sect. S1 in the Supplement). The pollution influence metric used for soft labeling was calculated using normalized eBC and arabitol concentrations. Individual weights were calculated for tuning data based on our confidence in the soft labels, determined by the pollution influence metric, and a down-weighting term. To reduce the risk of overfitting to uncertain pseudo-labeled particles while preserving the classifier's fit to the laboratory reference data, the total contribution of the soft-labeled tuning set was limited to 20 % of the weight assigned to the laboratory-derived source training data. This follows the common domain adaptation strategy of down-weighting uncertain samples and treating their influence as a regularization term rather than as fully trusted labels (e.g., ). A sensitivity test determined that 20 % influence by the tuning samples was conservative enough to preserve accuracy in in-sample testing with the training data while leading to significant changes in classification behavior; hence, we chose this degree of tuning to investigate the effects of domain adaptation on the classification algorithm.

The tuned fluorescent particle classifier reproduced the DT-based fPBAP concentration time series , with few (∼ 10⁻³–10⁻² L-1) fPBAP during winter and early spring, a peak in July (∼ 10¹ L-1), and a subsequent decrease during winter onset (Fig. a and b). This cycle is also in agreement with , who determined that most carbonaceous coarse-mode particles sampled at Zeppelin Observatory in July 2020 and analyzed with scanning electron microscopy were likely PBAPs (such as microorganisms and plant debris), while none were detected in March 2020. Compared to the DT-based method, the ML-based classifier also better captured a second peak in PBAP tracers during October that coincided with a significant increase in warm-activating INPs (Fig. 1c of ). The low proportion of HFPs identified as fPBAP between January–April (Fig. a) illustrates that the tuned classification algorithm robustly reduced the influence of combustion particles on fPBAP concentrations. This is reflected in improved correlations between fPBAP concentrations with biological tracers arabitol and mannitol (Figs. e and S6); however, these correlations do not constitute fully independent validation of the adapted model and should be interpreted cautiously due to their use in the tuning process. In August, when biological tracer concentrations were highest, our algorithm detected fPBAP concentrations nearly two orders of magnitude higher than the DT method did. This is expected because the DT-based fPBAP criterion was validated for one type of fPBAP (marine bacteria; ), while our classification algorithm targets multiple fPBAP types with diverse fluorescence properties. The low (high) proportion of HFPs classified as fPBAPs during March (July) reflects a similar seasonal fPBAP cycle as that described in . Both our algorithm's and DT-based fPBAP concentrations correlate significantly with BB tracers (Fig. e), in agreement with . Although levoglucosan is formed only through pyrolysis, incomplete biomass combustion can release monosaccharides (e.g., glucose, fructose) when wood carbohydrates (e.g., cellulose) are broken down . This has implications for the interpretation of fructose and glucose as PBAP tracers, as illustrated by their correlations with levoglucosan (Spearman ρ = 0.56 and 0.59, respectively; Fig. S7 in the Supplement).

The pollen class contributed the largest share to particles identified as fPBAP by our algorithm throughout the year but more consistently in summer, followed by fungal spores and bacteria (Fig. c). Bacteria made up a small proportion (∼ 0.1 %–1 %) of fPBAP identified by our classifier. It is unlikely that pollen fragments contribute such a large share of PBAPs in Svalbard; earlier identifications of PBAP classes in Ny-Ålesund in two summers found very low concentrations (∼ 10⁻³–10⁻¹ L-1 in summer) of intact pollen , while the mid-July peak concentration of pollen-class particles identified by our classifier is 3.30 L-1. Considering the similarities between fPBAP classes, cellulose, and combustion particles seen in Sects.  and , plant debris or combustion particles may be misidentified as pollen- or fungal spore-like by our classifier due to their broad fluorescence spectra (as in Fig. a). This is corroborated by the persistent presence of pollen-like and fungal spore-class particles identified throughout winter (Fig. c). Fungal spores misclassified as pollen fragments may possibly also contribute to the dominance of the pollen class in fPBAP identifications, as suggested by the pollen-like class correlating more strongly with fungal tracers arabitol and mannitol (ρ = 0.69 and 0.65, respectively) than the fungal spore class does (ρ = 0.49 and 0.43, respectively; Fig. e).

Figure d reveals that the untuned dust LRM was also insensitive to variability in SSA and dust tracers (sodium and silicon, respectively), indicated by the near-constant, high fraction (∼ 80 %) of particles labeled as dust. This high a dust fraction is unlikely, since total coarse-mode particle concentrations measured by the MBS are strongly correlated with SSA tracers sodium, chloride, and magnesium, and do not correlate with mineral dust tracers (Fig. e). The modest but significant correlation between calcium and potassium with total coarse-mode particle concentrations may reflect that both marine and terrestrial sources can influence calcium aerosol content , and that potassium may be emitted by mineral dust, marine, and BB sources . Inspecting non-fluorescent particles measured at Zeppelin Observatory during a sampling period with minimal dust influence (the wintertime minimum of dust influence relative to SSA tracers; see Sect. S1) reveals that they resembled dust more than SSA when compared to laboratory characterization data (Fig. S8 in the Supplement), suggesting that natural processes (e.g., aging of SSA) may have modified particle morphology. found similar difficulties in correctly identifying out-of-sample lab-generated SSA particles using conformal prediction. Laboratory-generated SSA properties vary markedly depending on generation method (e.g. and ), introducing biases to inductive approaches based on laboratory data.

To address this weakness, we applied domain adaptation on the dust LRM, this time constructing soft labels using mineral constituents silicon, aluminum, iron, manganese, titanium, nickel, and chromium for mineral dust mass and the sum of major SSA ions (sodium, chloride, magnesium, and potassium) for SSA mass; details of our dust influence metric's construction are further explained in the SI (Sect. S1). This step's aim is to relax the dust LRM's fit on the highly regular scattering signatures of nascent laboratory-generated SSA described in Sect.  around Fig. . The tuning block's influence was again weighted according to our confidence in the field label based on a dust influence metric and on a scaling factor limiting the total influence of the tuning block on the LRM's training to 20 % of that of the laboratory training data. The model's overall precision in identifying dust and SSA in repeated in-sample testing with laboratory data was not significantly affected by the tuning (Fig. S9 in the Supplement). Tuning the dust model significantly reduced the fraction of particles observed at Zeppelin Observatory classified as dust and increased dust fraction variability (Fig. d). The tuned model's resulting dust fraction was lowest when silicon concentrations were lowest between March–April, and increased episodically to maxima of ∼ 0.6 during summer when sodium concentrations were lowest. Although correlations do not independently validate the model, this adaptation results in a significant positive correlation with silicon and anti-correlation with SSA tracers sodium, chloride, and magnesium, which are strengthened relative to the untuned model's output (Fig. S7b). These fractions are consistent with those determined in this size range by via electron microscopy (their Fig. 5a), who found that mineral dust particles comprised about half of particles > 2 µm in March and July. A significant proportion of the remainder were mixed sea salt-mineral particles ; what scattering signatures these mixed particles would exhibit is unknown.