Revisiting microplastic pollution: A novel method for detecting small-sized microplastics in natural waters

Microplastics (MPs) are persistent particles from 1 to 5000 µm¹ reported everywhere² from pole to pole^3,4,5,6,7. Small microplastics represent a larger environmental issue due to their increased bioavailability, enabling ingestion by more organisms^8,9,10 and transfer throughout the food web; however, little is known about their environmental concentration and distribution¹⁰. Most of MPs environmental reports have a threshold above 100 µm, with fewer than 20 studies reporting MPs below 20 µm; conversely, most of the toxicity assessments were conducted with MPs below 100 µm¹¹. This dichotomy between MPs in-situ levels and toxicity limits realistic ecological assessments^10,11,12,13.

MPs reports are often limited by the sampling procedure and analytical technique used^2,13,14,15. A critical review of MPs data¹³, suggested that using 300–350 µm mesh may underestimate MPs pollution reports^3,5,14,16 by up to four orders of magnitude¹³. Limited studies using small-sized mesh showed much greater MPs levels^6,7,17, with 60–80% of the MPs reported in the smaller size-class (5–20 µm), suggesting a greater, mostly undocumented, abundance of MPs.

Currently, only limited analytical techniques measure microplastic type and size below 100 µm simultaneously^2,15. We lack robust high-throughput analytical capabilities essential for routine monitoring of small MPs in industrial processes and in the environment to face emerging legal frameworks², as well as for toxicological assays or studies on MPs behavior under laboratory and natural conditions. For example, Raman spectroscopy can identify MPs type and size as small as 1 µm¹⁸, yet interference with pigments prevents the detection of heavily stained MPs. Fourier-transform infrared (FT-IR) and laser direct infrared (LDIR) spectroscopy, capable of identifying MPs composition and size, require further developments to statistically detect MPs below 15–20 µm^18,19. Additionally, atomic force microscopy coupled with infrared spectroscopy also called infrared nano spectroscopy (AFM-IR²⁰) and optical photothermal infrared spectroscopy (O-PTIR²¹) can also be used to detect and characterize microplastics below 100 µm. Pyrolysis coupled with gas chromatography tandem mass spectrometry (Py-GC/MS) can detect the mass of low-sized MPs at reasonably high throughput, yet no information on size and particle concentration can be derived^2,22. Whereas additional developments^{20,21,22,23,24} are underway, and combining different analytical techniques^15,22,23,24 seems a promising way forward, consistent techniques for environmental monitoring at scale are still missing.

Here, we present a novel rapid quantitative analytical method using a machine learning-assisted spectral flow cytometry (MLA-SFC) to classify plastic microparticles in the 5–70 µm size range. By filling the gap between field MPs observations and toxicity assessments, our study responds to an essential need in MPs research for sound policy decisions^2,25.

Validation of the MLA-SFC

Our technique relies on using a spectral flow cytometer with 64 detectors covering the spectra between 373 and 812 nm, coupled with two machine learning algorithms to differentiate (i) MPs from non-microplastics (NMPs) particles and (ii) different microplastics polymers (Figs. S1–S9). To distinguish MPs from NMPs, the machine learning model was trained using microplastic particles and relevant natural samples (see Table S1). Model training with 20,000 particles per plastic sample (i.e. 235,955 MPs) and 258,663 NMPs, resulted in high algorithms’ performance (Fig. 1) ensuring efficient microplastics detection and distinction of PET, PC, PU, PVC; however, poor distinction of PE and PP led to grouping them in a single category (PE-PP). To further differentiate what resembles MPs standards from what is classified as MPs with a very high level of confidence, a score indicator was implemented, corresponding to scores of 0 and 90, respectively. An intermediate score of 70 was considered unless otherwise stated.

To validate MPs detection using MLA-SFC, standards of heterogeneous sizes from PE, PP, PET, PU, PC, and PVC were spiked as single (Fig. 2A, B) and mixed (Fig. 2C) polymers in ultra-pure water and different natural water types. MPs standards were analyzed directly (Fig. 2A, C) or following filtrations steps to recover MPs from natural waters prior to analysis (Fig. 2B). Direct analysis detected PU, PET, PE-PP, PC, and PVC with an accuracy exceeding 95% and an average of 104 ± 1.1% MPs recovery in most natural water types (Fig. 2A). Heavily particle-charged water (>300,000 particles/mL), such as Rhone River and glacial’s lake, resulted in excess PE (Fig. S10, p 0.003), likely associated with lithogenic particles resembling PE, stressing the need to further implement a separation protocol reducing NMPs.

We have developed a new protocol suited for small-sized-MPs recovery combining a H₂O₂ digestion step with a centrifugation density separation (sodium polytungstate solution, 1.56–1.6 density) to separate MPs from NMPs (Fig. S11, Tables S1–S4). An average 87% MPs recovery rate was obtained for all polymers (Fig. 2D), demonstrating the good performance of this separation protocol.

Procedural average blank value of 48.6 MPs (n = 7) with a detection limit of 52.8 MPs was obtained. Therefore, after blank subtraction, MPs below 4.1 are below the instrumental detection limit. Blanks according to each polymer type (Table S5) and size classes were also measured (Table S6).

Analysis of subsequent 4 L samples—a volume providing a representative microplastics measurement in Lake Geneva (see “Methods”, Fig. S12)—resulted in an average of 483.9 ± 187.0 MPs (Fig. 3A, e.g., 121 MPs/L, Tables S7 and S8, Data S1). The 38.7% standard deviation obtained was deemed satisfactory for an environmental sample, especially as MPs behave heterogeneously in water. Similar variabilities were reported for different polymer types (Fig. 3B, PC (22.4%), PE-PP (32.9%), PU (10%), PVC (24.4)) and size classes (Figs. 3C, 5–15 µm (3.6%), 15–30 µm (12.9%), 30–53 µm (39.3%)). A greater standard deviation was observed for PET (68.6%) and particles greater than 53 µm (66.4%), likely due to very low average contributions with respect to total MPs (0.8 and 0.7%, respectively) (Fig. 3B, C).

To further demonstrate MLA-SFC accuracy, MPs standards and natural samples were analyzed in parallel using LDIR (Fig. 4, Tables S9 and S10). On average, considering the same size ranges for both techniques, 2.2-fold greater MPs concentrations were measured by LDIR compared to MLA-SFC following single PU or MPs mixture additions in pure water and in ethanol (Fig. 4A). For lake samples, more PE-PP, PET, PU but less PC and PVC for the two lake samples were measured by LDIR, yet MPs measured by both techniques were always in the same order of magnitude (Fig. 4B–D). The largest differences observed for PE-PP were 5.2 and 2.8 MPs/L, for PVC 0.05 and 0.95 MPs/L, and for PU 24.3 and 2.3 MPs/L for the LDIR and the MLA-SFC, respectively (Table S10).

Total MPs (assigned to PE-PP, PET, PC, PU, PVC) measured by LDIR and MLA-SFC within the same size range gave very good agreement (lake inshore: 250.5 and 66.3 MPs, lake offshore: 40.9 and 32.6 MPs, and Rhone River: 17.5 and 65.0 MPs, respectively), considering an inter-comparative study led by the European commissions²⁶ using PET in a simple synthetic matrix. Laboratories using µFTIR (n = 32) reported an 1800-fold variation in MPs numbers, LDIR analysis (n = 3) obtained the lowest variability (1.8-fold) reported for any single analytical technique used³⁰⁰. Our results were very satisfying, considering they were performed in a complex matrix. Further, considering MPs heterogeneous behavior during subsampling after sample preparation, different databases for MPs identification, and different accuracy in size identification, MPs counts were indeed quite similar for both techniques, demonstrating MLA-SFC accuracy to detect MPs in natural waters.

LDIR can identify more polymer types than MLA-SFC (Fig. 5, Table S9) trained with the current model. Amongst our samples, the 5 polymers detected by MLA-SFC constitute 44–77% of the total MPs detected by LDIR, showing that MLA-SFC captured a significant fraction of the total MPs. LDIR showed that PP-PE accounted for 15–31%, whereas PU accounted for 5–52%, and rubber for 9–26% of the MPs pollution within the size range 15–100 µm (Table S10). Both LDIR and MLA-SFC showed more important MPs in small-sized fractions (Fig. 3C, Table S10).

Small-sized MPs call for a reassessment of pollution impacts

Twelve samples collected at the surface from the shore of Lake Geneva, offshore surface and deep samples (0–100 m) on the LéXPLORE platform, and from 4 surrounding river surface water samples were analyzed (Fig. 6, Data S1). Microplastics concentrations (5–70 µm) ranged from 32.9 to 687.1 MPs/L for surface waters and from 6.9 to 154.6 MPs/L within the water column (Fig. 6A). No statistical differences were observed amongst MPs concentrations in surface waters, rivers, and deep lake water (p = 0.171 to 0.225), likely due to high variability and limited observations in river samples. High MPs concentrations were observed for the top 10 m of the vertical profile, coinciding with the mixed layer depth (Fig. S13), yet concentration rapidly decreased at depth down to 6.9 MPs/L. At depth, PET and PVC contributions to MPs increased up to 1.7–5% and 30–62%, respectively, while contributions from PC and PU sharply decreased. However, the distribution of PE-PP did not show any trend with depth.

Size distribution was similar across all surface samples analyzed (Fig. 6E–G), with MPs concentrations increasing exponentially for smaller sizes. In all cases, MPs concentrations in the size class 5–15 µm were greater than larger-sized MPs (p < 0.001), and 15–30 µm MPs were greater than MPs above 30 µm (p < 0.001 and p = 0.022 for river samples). In contrast, 30–53 µm and >53 µm MPs concentrations were similar in the river and deep samples (p = 0.067). Lake inshore surface and river waters showed similar size distribution (Fig. 6E, G), yet the contribution of small MPs (5–15 µm) is greater offshore and at depth (Fig. 6F).

Using mid-sizes from the different MPs size classes detected (Fig. 6E–G), an exponential decay successfully described relative MPs contribution with respect to size (R² > 0.937, Fig. 7). Using equations for inshore lake and river surface waters, as well as for offshore and deep lake waters (Fig. 7), one can project the impact of larger-sized (70–100 µm) and most importantly smaller-sized MPs from 1–5 µm to extend findings in the size range 1–100 µm. Because the relative MPs contribution for the mid-size range 10–61 µm (corresponding to our 5–70 µm analytical range) corresponds to 100% of the MPs detected, extending this regression to 1 µm results in values exceeding 100% being 118.1% and 472.1% for inshore lake and river samples and for offshore and deep lake samples, respectively. Therefore, to explore the relative estimated MPs contribution in the 1–100 µm range, one needs to divide MPs contributions by 1.181 and 4.721, respectively. Applying this approach, one can estimate that 43–78% of the MPs are within 1–5 µm, escaping direct quantification by MLA-SFC. Using equations from Fig. 7 and the total % for the 1–100 µm range, one can estimate that the analytical technique measuring MPs polymer and size down to 20 µm, misses from 73 to 97% of MPs, resulting in a significant underestimation of MPs pollution in natural waters.

Further considering the average recovery efficacy for different MPs polymers (87.1%, Fig. 2D), and the average representativeness of the five polymers detected (62.8%, Table S10) compared to total MPs detection by LDIR, one can estimate the total 1–100 µm MPs pollution for Lake Geneva and nearby rivers. By doing so, our 1–100 µm MPs projection showed that analyses by MLA-SFC would underestimate MPs pollution by 3.5-fold in inshore lake and river waters and by 10.8-fold in offshore and deep lake waters. Taking this into consideration resulted in average MPs projected concentrations (1–100 µm, score 70) of 338.6 MPs/L (112.9–702.4) in surface inshore lake waters, 680.7 MPs/ L (74.2–1662.6) in deep lake waters, and 983.4 MPs/L (134.0–2415.7) in river surface waters (Data S1). Being less (score 90) or more (score 0) permissive with the algorithms used in machine learning gives us a good confidence in the expected range for projected MPs above 1 µm in our study region; being from 97.7 to 1211.8 for inshore lake waters, 296.4 to 8954.3 MPs/L for offshore and deep lake waters, from 348.2 to 2492.7 for rivers.

Environmental risk assessment

We use risk characterization ratios (RCRs), defined as the ratio between the exposure concentration and the predicted no-effect concentration (PNEC), to assess risks associated with small-sized microplastics (1–100 µm, score 70). Across the sampling sites, 6 out of 20 showed an indication of possible risk (≥1%) (Fig. S14, Table S11). When the exposure distributions (Table S12) derived from the average MPs concentrations were used, considering the standard deviation from the corresponding sampling points, the probability of risk was calculated to be 0.1%, and 22% for Lake Geneva and nearby rivers, respectively. Risk in rivers was higher due to the outlier data in Sion (73.3% probability of risk). When Sion was excluded, the risk probability decreased to 2.2%. We used the seven subsequent samplings at LéXPLORE as representative to identify the associated risk variability. The variability in MPs concentrations of 38.7% results in risk probability ranging from 1.6 to 59.6% with an overall risk of 19.1% and a standard deviation of 20.9% for at LéXPLORE. Our assessment suggested that environmental risk might be expected in some sampling sites when considering the 1-100µm MPs projections.

Microplastic recovery and polymer class assignment were satisfactory for five polymer classes (PE-PP, PU, PC, PET, PVC) using MLA-SFC. The MLA-SFC results showed good agreement with LDIR, a benchmark technique for microplastic analysis in the same size ranges, for both synthetic plastic-amended (15–30 and 30–100 µm) and natural samples (15–30 and 30–70 µm). Overall, MPs trends by depth, offshore/inshore²⁷ and lake/rivers¹⁷ sites align with previous studies, reinforcing confidence in MLA-SFC’s capability to detect microplastics in natural waters. While the machine learning algorithm underlying MLA-SFC was trained specifically on these five polymers due to the limited availability of standards, it is designed to work with a broader range of plastics. Polymers beyond the training set can still be recognized by the algorithm, though they may be misclassified into incorrect categories; however, they will generally still be identified as plastics. Currently, MLA-SFC can detect PE-PP, PC, PU, PET, and PVC, which represent a significant portion of total MPs in natural waters. Preliminary results suggest that further model training could extend the detection to PS, PA, PTFE, and rubber particles. These were not included here because the limited number of standards would create an unbalanced model, potentially biasing polymer classification. This study demonstrates that MLA-SFC is a valuable tool for microplastic detection, and further machine learning training with naturally altered polymers is needed to improve classification for other polymers relevant to environmental samples, such as weathered MPs. Although MLA-SFC can be influenced by the sample matrix, the absence of interference from humic substances (Figs. S10 and S15), automatic exclusion of highly fluorescent particles like algae, and implementation of our separation protocol help mitigate such an impact.

Here, three out of the four rivers’ samples have greater MPs than average lake water, suggesting overall higher MPs pollution in rivers as previously reported¹⁷. High turbulences and flow rate potentially resuspending MPs from sediment and stabilizing MPs in surface waters might explain this¹⁷. A rapid attenuation of MPs concentrations with depth was also reported for coastal²⁷ and lake waters²⁷. As for our study, the microplastics profile clearly showed low-density polymers primarily found in surface waters, and high-density polymers in deep water, as previously reported^17,27,28,29.

Total MPs concentrations measured (33–687 MPs/L, 5–70 µm, score 70) were similar to studies reporting small-sized MPs (500 MPs/L in snow from remote areas⁶ and 218–319 MPs/L in South Korea lake and river systems¹⁷). As for our data, these studies showed that most MPs are within the small size fractions, with 61–63% in the size class 5–20 µm and a little above 50 µm (2.6–3.5%) in lake and river samples¹⁷. Here, on average, 5–15 µm MPs accounted for 63.9%, 71.2%, and 90.0% of the MPs in the lake surface, river, and offshore to deep waters, respectively (Data S1). Therefore, it is not surprising that our measurements report MPs concentrations 65-fold above the average MPs reported in lakes¹⁶ (1500 MPs/m³) using mesh size between 100 and 300 µm. Indeed, small-sized MPs were typically omitted in these previous reports.

Our projection of MPs concentrations within a 1–100 µm size range is an estimation based on (i) the relationship between MPs size and concentration observed using our MLA-SFC data, (ii) the average MPs recovery from nominal additions, and (iii) the polymers missed from our comparison with LDIR—a technique able to identify most plastic polymers. Our approach provides a good first estimation of small-sized MPs pollution and size distribution in freshwaters. In accordance with predictions from-ref. ¹³, our estimated concentrations of MPs in the 1–100 µm size range exceed previous MPs reports using a manta trawl by 226–656-fold¹⁶. This stresses the need for analytical developments to determine MPs in the 1–100 µm size range that also include information on microplastic size. Currently, no analytical technique is available to simultaneously measure MPs concentrations and size below 10 µm with strong statistical confidence, and our MLA-SFC is currently limited to an identification threshold of 5 µm. However, the statistically robust size-dependency of MPs has allowed estimation of MPs concentrations within the 1–100 µm size range.

This greater repartition of smaller-sized MPs and vertical distribution of polymer according to their density, as observed here and elsewhere^17,27, is relevant to MPs settling and transport in aquatic systems. It can also come with different potential effects according to depth and location within a single lake. Using species sensitivity analysis based on the ToMEx database and no observed effect concentration (NOEC), a calculated hazard concentration of 11,400 MPs/m³ was obtained for Lake Taihu²⁷. In our case, average MPs concentrations are exceeding this threshold for 5–70 µm MPs by 8- and 25-fold for surface inshore lake and river waters. When projecting MPs concentrations down to 1 µm (score 70), MPs’ average concentrations exceed this threshold by 30- and 60-fold for surface and offshore/deep lake waters and by 86-fold in rivers.

The environmental risk assessment is associated with several limitations in that PNEC distribution of weathered MPs (i) was performed with limited data, but still includes studies from at least three trophic levels, as in ref. ³⁰; (ii) covered PS, PET, PVC type of MPs, thus only representing 2 out of the 5 MPs polymers measured; (iii) derived from limited number of species, some of which not native to our study region, and (iv) covered sizes up to 75 μm due to limited data, while MPs predictions were made for sizes up to 100 μm in this study. In addition, as this study focuses on smaller-sized MPs, both the exposure and hazard assessments considered smaller-sized MPs. Therefore, results should be interpreted as applying specifically to smaller size fractions of MPs rather than to the entire microplastic size spectrum. Ideally, more data would be required on toxicity tests with different types of weathered MPs, and more exposure concentration from the sampling points to limit the uncertainty to make risk assessment more robust in the future.

Another limitation may be due to the use of MPs projections (1–100 μm) in exposure assessment. To improve the reliability of the projections, only datasets with a quality score of 70 were used; however, uncertainty associated with the extrapolation remains. The probabilistic risk assessment provides a range of possible risk outcomes rather than a single point estimate, and, where exposure distributions were used, the Monte Carlo simulations accounted for variability in the exposure distributions. The resulting projected MPs pollution is up to 656-fold higher than previously reported, highlighting the urgent need to advance the detection, identification, and monitoring of small-sized MPs in parallel with environmental risk assessment.

MLS-SFC provides insightful information on MPs concentration, nature, and size in the environment, as well as for industrial monitoring. Consistent and documented quality assurance and control, required for high-quality environmental or legal MPs reports^25,31,32, are easily implemented using certified beads (NIST Traceable Size Standards) as well as MPs that mimic MPs likely found in natural systems¹⁹. Further developments of MLA-SFC are needed to include other polymer classes, such as PS, PTFE, PA, rubber, and environmentally altered microplastics through additional ML models training. The method we describe offers a higher throughput (18 analyses daily) compared to LDIR (2 daily), and because MLA-SFC does not rely on mass spectroscopy, it also presents promising opportunities for coupling with existing techniques such as µRaman and µFT-IR mass spectroscopy, LDIR, or even GC-MS pyrolysis. Due to the difficulty in reconciling techniques that use different databases for MPs identification, or measure mass versus size distribution, techniques capable of analyzing microplastics down to or below 1 µm^18,20,22 were not extensively discussed here.

Machine learning and algorithms performance

We propose a process (Fig. S1) to identify and classify microplastics polymer and their colors using machine learning. Known plastic standards spiked in milli-Q water were used as a reference to test the algorithms and train the models.

Seven high-performance machine learning algorithms have been considered as candidates: Logistic Regression, Linear Discriminant Analysis, Decision Tree Classifier, Gaussian Naïve Bayes, Random Forest, XGboost, and Multi-layer Perceptron (MLP). The Deep Learning MLP algorithm was selected for all the classification tasks, as it proved to be highly accurate (Figs. 1, S4, S7). This algorithm showed a good performance on the detection of microplastics (F1 = 94.6%, Fig. S3) and a good generalization ability, by capturing plastic types that were not used for the training. Additionally, the MLP has a larger margin for improvement due to several parameters that can be further optimized. Furthermore, using the positive unlabeled (PU) strategy (Fig. S4), the model was able to train with MPs contaminated samples. The MLP algorithm could identify each plastic through its unique global spectrum pattern (F1 = 84%, Figs. S7 and S9), while also dealing with the intrinsic noise in the data. This showed the usefulness of the model on spectral flow cytometry data to search for contamination in water samples with specific plastics. The algorithm was also very accurate in the identification of plastic types (F1 = 94.4%) and plastic colors (F1 = 88%) for the used standards. However, if not enough standards are used for training, the algorithms are not able to generalize and consequently classify new plastics poorly. The ability to detect color was not used here, but it might be relevant to further investigate pollution in MPs contaminated environments.

The MLP algorithm is a highly flexible nonlinear algorithm capable of learning a near infinite number of mapping functions. With this flexibility comes the disadvantage of high variance in a final model. During the testing exercise of the algorithm, an important variance was detected in both the algorithm and the dataset. Two strategies of variance mitigation were successfully tested and implemented: increasing the training set and using ensemble methods to reduce the variance and increase the classification accuracy. The machine learning code was simplified and made open source on GitHub.

Preparation of MPs standards and algorithms training

To ensure relevance to environmental conditions, most MP standards were derived from everyday consumer goods, reflecting the types of plastics potentially found in natural waters. The plastics that were not already in powder or bead form (smaller than about 60 µm) were ground with a Freezer/Mill (model 6870D, SPEX SamplePrep) equipped with stainless-steel grinding vials. The program used for plastic grinding was the following: 6 cycles, 15 min precool, 2 min run, 2 min cool, 15 cps rate. The powders generated from the grinding process were collected in individual glass tubes. Between the grinding of each sample, the stainless-steel vials were washed with water and soap and rinsed with acetone to remove any plastic residue to avoid cross-contamination.

After several rounds of validation in ultra-pure water, the model for environmental samples was trained using up to 41 home-made and commercial MPs standards (Table S1), as well as zinc, calcium, and magnesium stearates particles (which have a spectral signature resembling MPs), natural samples, soil samples, procedural blanks, and source water for NMPs. To ensure a balanced model, a total of 235,955 MPs and 258,663 NMPs spectra were used, of which 66,333 for PE-PP, 48,366 for PET, 37,902 for PU, 38,092 for PVC, and 45,262 for PC (Table S1). For all consumer-derived samples, polymer identity was confirmed using FT-IR spectroscopy on larger MP fragments.

At least 20,000 particles per MP type or NMP sample were analyzed with a spectral flow cytometer Cytek Aurora (software Cytek Spectroflo v. 3.3.0) equipped with 5 lasers (355 nm, 405 nm, 488 nm, 561 nm, 640 nm), and with an enhanced small particle (ESP) detection option that enables a higher resolution for nanoparticles down to 70 nm. Thresholds were set at SSC = 5000, FSC, SSC, and SSC-B gains set to 1, and all other gains increased (+400% for UV and Violet detectors, +300% for Blue detectors, +800% for Yellow-Green and Red detectors). The raw FCS files produced by the instrument were used for machine learning training or analysis. Sizing of the particles has been estimated using reference Polystyrene beads (Fig. S16).

Data output and scoring

Results are provided in separate files for batch, size, and score. Batch files enumerate particles by type (MPs, NMPs) and polymer class, along with the volume analyzed. Size files categorize particles into four size classes: 5–15 µm, 15–30 µm, 30–53 µm, and >53 µm. Score files quantify the algorithm’s confidence in identifying individual scores to MPs and their associated polymer types.

To improve analytical resolution in complex natural matrices, we used the algorithm confidence score to distinguish weakly and strongly recognized MPs. Machine learning has previously been applied to MP detection²³, but here we extend its utility by assigning confidence scores ranging from 0 to 90. For heavily charged samples (e.g., Rhone River), scores helped differentiate polyethylene (PE) from lithogenic particles. A score of 70 was selected as a compromise between broad matching (score 0) and strict spectral matching (score 90). Data for scores 0, 70, and 90 are available in the supplementary information.

Sample collection and preparation

MP standard solutions were prepared using powdered household and commercial materials diluted in LC/MS-grade water (1153331000, Merck) containing 1% Triton X-100 (T8787, Sigma-Aldrich). Solutions were passed through a 100 µm Corning nylon strainer (CLS431752, Merck) and stored at 4 °C for up to 4 months, with reanalysis prior to each use to ensure consistency of results. This also served as a quality control step. Aliquots were spiked into test solutions at known concentrations, and validation showed good agreement between nominal and measured MPs concentrations (Fig. 2A), confirming Triton X-100’s effectiveness in maintaining sample homogeneity. The solution was vortexed prior to each analysis.

To minimize sediment interference near shorelines, natural samples were collected from bridges (rivers) or transport platforms (CGN, lakes) using an 8 L stainless-steel bucket (28 cm diameter) deployed with a sisal rope. Bottles (1–5 L Schott glass (VWR) or 2.5 L amber glass) were rinsed three times with 100 µm-meshed water (stainless steel mesh, VWR) before being filled. Samples were stored in the dark until further processing. Coordinates for sampling sites are shown in Table S7 together with sampling dates for natural sample analysis. These sites were also sampled for technical validation at a prior time.

Each bottle was shaken for 10 seconds, then filtered using a 0.2 µm PES 500 mL vacuum filter (TPP “rapid”-Filtermax, Merck Millipore). Bottles were rinsed twice with LC/MS-grade water, which was also filtered. Retained particles were resuspended with 700 µL of LC/MS grade water containing 1% Triton X, followed by 700 µL of LC/MS grade alone. Each solution was resuspended repeatedly using a 1 mL pipettor 6-times to rinse the whole filter surface. The resulting particle suspension was collected from the filter surface and transferred to a glass tube, and treated with 30% H₂O₂ (1:3 volume ratio) to digest organic matter overnight at 60 °C at 150 rpm.

To separate MPs from lithogenic particles, we used density-based separation. Although ZnCl₂ (d = 1.8) is commonly used, it caused fluorescence loss in our samples and was deemed unsuitable. Instead, we used freshly prepared sodium polytungstate (3Na₂WO₄ · 9WO₃ · H₂O, 71913, Sigma-Aldrich) solutions adjusted to a final density of 1.56–1.6, verified using a densitometer. This density effectively separates most plastics, including high-density polymers like PET and PVC, is less toxic than ZnCl₂, and does not alter the fluorescence of MPs standards.

Small-sized MPs can stick to the tube walls and gather at the bottom or the top part of the solution. To help with our protocol development, we used blue PS beads (6.92 µm, Spherotech, PPB-60-5) to visualize MPs’ behavior. Because small MPs do not settle in a separation column overnight, centrifugation was used to separate MPs (supernatant) from NMPs (pellet) at 3200g for 1 h. The supernatant was recovered by gently inverting the tube onto a 100 µm mesh (Fig. S11) and collecting supernatant in a 25 mL Pyrex tube. A significant portion of the PS blue beads stuck to the tube wall at this step. MPs adhering to the tube walls were recovered after adding 400 µL of 3% agar solution directly above the pellet (NMPs) with care not to touch the tube wall. Agar was kept liquid in a 90 °C water bath. The agar cap deposited in the glass tube was cooled on ice for 5–10 min, resulting in a solidified agar “plug” that effectively separated MPs from NMPs in subsequent rinsing steps. Tubes were rinsed three times with 0.1% Triton X-100 and once with LC/MS grade water, each time collecting the supernatant by gently inverting the tube onto the 25 mL Pyrex tube to collect rinses. The tube with the agar cap was then discarded, and work proceeded with the solution collected in the Pyrex tube. Following a 10 s vortexing step, the MPs suspension in the Pyrex was filtered on a 0.2 µm PES 500 mL vacuum filter. The Pyrex tube was further rinsed with 0.1% Triton X-100, which was vortexed prior to filtration on the same filter. This operation was repeated four times to ensure complete recovery of MPs on the filter. Particles retained on the filter were collected and resuspended in 1.4 mL Triton 0.5% solution as explained above. The final suspension was screened through a 70 µm mesh to prevent clogging of the flow cytometer capillary. As a result, MPs between 5 and 70 µm were analyzed in natural samples, while 5–100 µm particles were analyzed for synthetic solutions and direct analysis. The total recovered volume ranged from 1.5 to 2.5 mL and was directly measured using spectral flow cytometry at a flow rate of about 120 µL/min. Excess volume is due to liquid remaining on the filter that is sampled when resuspending and collecting particles into glass tubes for subsequent analysis. The solution was vortexed prior to analysis by MLA-SFC.

Quality control and assurance

The Cytek Aurora QC was performed daily before use with SpectroFlo QC Beads (Cytek, B7-10001) for routine performance tracking. Blanks for the solutions used (LC/MS grade water and LC/MS grade water with Triton X-100 1%) and recovery of MPs standard are routinely analyzed. All glassware is washed in an industrial washer and rinsed once with LC/MS-grade water prior to its use in the laboratory. To limit MPs contamination from the surrounding environment and the operator, manipulation was made without gloves (unless a hazardous chemical is manipulated, where nitrile gloves were used) and using cotton or wool clothes or a lab coat under a class II Biological Safety Cabinet. Triton X-100 was used to enhance MPs’ homogeneity, and the sample was vortexed for 3 s at high speed prior to subsampling or analysis. When not under the Biosafety Cabinet, glass tubes were covered by aluminum foil to prevent contamination from airborne MPs.

Blanks were made at every step for sample manipulation to exclude significant contamination. At least three procedural blanks were made for each filter batch. Procedural blanks showed that the more we rinsed the filter, the more plastic we put in suspension, yet the filters used were selected for the low plastic contribution. A reproducible resuspension step was thus implemented, consisting of 6 rinses to cover the whole filter surface for each solution (700 µL LC/MS grade water + Triton X-100 1%, then 700 µL LC/MS grade water). The detection limit for MPs in natural waters was calculated as three times the standard deviation of seven procedural blanks. Procedural blanks were obtained by filtering 1 L certified LC/MS grade water and following the sample preparation as for natural samples. For direct MPs measurements in solution, the blanks were obtained from 0.5% Triton solution analysis made daily.

Technical validation in natural waters

The analytical performance of the spectral flow cytometry for MPs analysis was verified using MPs addition to different natural water solutions. Solutions were analyzed freshly after MPs additions or following sampling preparation used for natural waters. The MPs used were heterogenous in sizes (Fig. S17) and made of PE-PP, PET, PC, PU, and PVC. Both single MPs and mixed PE, PET, and PU additions were used.

Several natural samples were screened to test the applicability of this method to different water types (Figs. S10 and S15); namely, filtered (0.2 μm, PALL Acropak cartridge) lake (Saint Sulpice) and sea waters (Southwestern Greenland fjords), unfiltered Lake Geneva (Saint Sulpice), Glacial Lake at the bottom of the Rhone Glacier (GL), Rhone River (Porte de Scex) and Venoge River. MPs additions were analyzed by comparing each nominal polymer addition with its measured value. Both polymer assignments and MPs levels were calculated. In this case, if the % of the assignment is satisfactory, then the polymer is efficiently recovered and identified.

Separation protocol was tested using ultra-pure water and Lake Geneva surface water (Saint-Sulpice, Table S2) with and without H₂O₂ treatment. Standard addition of PU was made, and its recovery as well as size distribution were measured before and after the density separation procedure. Without H₂O₂ treatment, PU recovery was poor even before the separation protocol, indicating that interactions between natural organic matter and PU alter its detection (Table S2), which was clearly visible at scores 90 and 0. We therefore chose to apply an H₂O₂ digestion step. The separation protocol has no impact on PU size distribution (Table S3). Recovery rates for different polymers were calculated from MPs standards used in machine learning by comparing the concentration following the addition of MPs in 0.5% Triton X-100 and after the separation protocol. Impact on MPs size distribution was only observed for selected standards (PET1, PU1, and PE1), with an aggregation increase of MPs between 15–30 µm (Table S4). However, the average increase in MPs within this size class was only 3.4% among the nine MPs tested and thus deemed negligible.

The evaluation of MPs in the environment is challenging. The theory of sampling^33,34 revealed that for a homogeneous MPs distribution, a sample of 0.25 L would correctly represent MPs concentrations of 1000 MPs/L³³. On the other hand, for a random clustered distribution as expected for MPs, a volume of 20 L would be required³⁴. Clearly, MPs concentrations reported in natural waters^3,4,5,16 are far too low for direct analysis (3 mL) being representative^33,34, yet could be relevant for industrial or polluted site monitoring. We tested sample volume suitability using a 5 L lake water (Saint-Sulpice) spiked with 2000 PU particles/L (Fig. S12). The minimal representative sampling volume was the volume at which the PU recovered concentration was stable and close to the nominal addition and was representative of 2000 MPs/L in situ. Stable MPs concentrations were obtained for samples ≥ 250 mL, suggesting, that for MPs between 5 and 70 µm, an intermediate case scenario between homogeneous and random cluster distribution applies, where 0.5 L and 5 L samples might be representative for a level of 1000 and 100 MPs/L, respectively. Considering the range of MPs reported in preliminary results (110–190 MPs/L, n = 6), we opted for a 4 L sampling as being (i) representative for MPs concentrations and size distributions and (ii) easy to implement. Analysis of seven 4 L lake samples collected simultaneously at offshore site L’éXPLORE (Tables S8 and S9) was used to further verify sampling volume representativity. Volume filtered for analysis was between 3.75 and 4 L for surface river and lake waters (except for the Rhone River, where it was reduced to 2 L as it is heavily charged with particles resulting in filter clogging); volume filtered was increased for the sample at depth up to 8.5 L (Table S7) due to an expected decreased MPs concentration compared to surface waters.

Comparative analyses with LDIR

As a final step for the technical validation of the use of MLA-SFC to detect MPs, we performed comparative analysis with LDIR for synthetic solutions (blanks, PU, PE-PET-PU mixture) as well as 3 natural samples from different locations (inshore and offshore lake water and river water, respectively, Saint Sulpice, LéXPLORE, and Porte de Scex, Tables S7, S9, S10). To improve comparability, the same sample was prepared as explained above and then divided into two subsamples for LDIR and MLA-SFC analyses. As samples cannot be recollected after MLA-SFC or LDIR analysis, one cannot analyze the same particle population with both techniques. The particle size measured by LDIR was classified in the same size class as the MLA-SFC to allow for quantitative comparison. We compared polymer and total MPs data for different size classes, 15–30 µm and 30–70 µm (30–100 µm for synthetic solutions) for both techniques. We also reported the lower size MPs (5–15 µm) detected only by spectral flow cytometry. With a 5.5 µm per pixel resolution for LDIR in reflection mode, an arbitrary cut-off at 3 (i.e., >15 µm) and 5 (i.e., >30 µm) pixels was considered minimal to robust for MPs identification, respectively.

The subsample for LDIR was filtered on a 3.0-μm pore size 25-mm diameter gold-coated PET filter membrane (Lab/Pharma i3 TrackPor P—PET Gold 100/0). The filtration device was rinsed 4 times to recover all MPs that could have adsorbed to the walls. The filter was then left to dry under laminar flow. Particles were analyzed using an Agilent 8700 LDIR. Particle spectra were collected and matched to an IR database using Clarity software version 1.6.83 and the spectral library Microplastics Starter 2.0. The particle analysis was set to autoscan mode with the default particle sensitivity, and particle diameters were set to 10 and 100 μm for the lower and upper limits, respectively. The scan area for analysis was drawn around the entire area where filtration occurred. The spectral range of the LDIR is 975–1800 cm⁻¹, and final spectra were acquired in reflection mode with a spectral resolution of 8 cm⁻¹. Matches for IR spectra to the library use a hit quality index (HQI) as a parameter for confidence in polymer identity. This parameter was set such that any spectra that had an HQI less than 0.65 were classified as unidentified. Identified MPs particles were then classed by polymer types and sizes to allow for comparison using the same size classes and polymer types as for the MLA-SFC.

Environmental risk assessment

An environmental risk assessment focusing on freshwater was conducted by considering the exposure concentrations of projected microplastics from 1 to 100 µm (score 70), and the predicted no-effect concentration (PNEC) for freshwater organisms. To characterize the risk, risk characterization ratios (RCR) were derived by using Eq. 1. A risk can be expected when the RCR is greater than or equal to 1 (RCR ≥ 1), indicating that the exposure is higher than or equal to the PNEC.

where PEC represents the predicted environmental concentration, PNEC represents the predicted no-effect concentration, both are expressed as part/L in this study, and the RCR is unitless.

For the PEC, two different analyses were conducted by using different exposure values presented in this paper: (i) Total projected MP concentration for each sampling, (ii) average projected MPs concentrations considering the standard deviations of the corresponding sampling points for Lake Geneva, rivers, and L’éXPLORE 2*. The first analysis (i) used single exposure values for each site obtained from the projected MP values, while the second one (ii) formed an exposure distribution using Monte Carlo simulations. Data points produced by iteration that were higher than measured concentrations (plus their standard deviation) were removed from the distribution to reflect the measured data.

The PNEC distribution for weathered MPs (up to 75 µm) was derived from³⁵, since weathered MPs are highly relevant for this study. We have chosen, for our assessment, the PNEC distribution excluding the highest no observed concentration values³⁵. The probabilistic species sensitivity distributions method developed by ref. ³⁰ and modified by ref. ³⁶ was used³⁵. This method accounts for the uncertainty and the variability of the inter-laboratory toxicity data.

For the risk calculation, a probabilistic risk assessment approach³⁷ was followed, considering Eq. 1. This approach enables accounting for the uncertainties in experimental measurements and provides a probable range of concentrations. In all analyses, the PNEC distribution derived from ref. ³⁵ was used, and the two exposure approaches described above were evaluated separately. For analysis (i), each projected exposure value was combined with the PNEC distribution to generate an RCR distribution for each sampling site. For analysis (ii), the Monte Carlo-generated exposure distributions were combined with the PNEC distribution to generate RCR distributions. %RCRs indicate the proportion of cases where the exposure is equal to or exceeds the corresponding values in the PNEC distribution, referred to as the probability of environmental risk.

All calculations were made in R software³⁸ using the following packages: “xlxs”³⁹, “msm”⁴⁰, “stringr”⁴¹, “ggplot2”⁴², “mc2d”⁴³, “FSA”⁴⁴, “Perc”⁴⁵, “dplyr”⁴⁶, “reshape2”⁴⁷.

Statistical analysis and figures

Student t-test (level 0.05) was done to compare microplastics size distribution and concentration in the different water types using Sigma Plot (v. 16.0). Figures were prepared using Sigma Plot (v. 16.0), De Novo FCS Express v. 7.22.0006, and R software³⁷ using R Studio version (2024.12.1)⁴⁸.