A Novel Sampling and Analytical Framework for Airborne Carbon Nanotubes

Abstract

Airborne carbon nanotubes (CNTs) present growing environmental and occupational health concerns. Like asbestos, CNT can deposit deep in the lungs and induce toxic effects, underscoring the need for accurate exposure metrics amid increasing global production. Current exposure assessment methods by the National Institute for Occupational Safety and Health (NIOSH) rely on mass-based measurements and indirect transmission electron microscopy (TEM), which often underestimate exposure, particularly for nanoscale fibers. To overcome these limitations, the objectives of this study were to develop a CNT-specific sampling and analysis method to improve fiber detection and characterization and validate its performance by comparing the novel diffusion-based sampler, Tsai Diffusion Sampler (TDS), with conventional approaches. Samplers were operated side-by-side under controlled conditions to ensure comparability. Furthermore, polycarbonate (PC) and mixed cellulose ester (MCE) filters were evaluated for all of the sampling configurations. CNTs collected onto filters and TEM grids were examined for their corresponding direct or indirect analysis. An automated image segmentation algorithm was applied to CNT images to ensure consistent fiber sizing and to support a count-based exposure metric. Results demonstrated that the TDS (which utilizes low flow rate) paired with PC filters most effectively captured individualized nanoscale fibers, enabling recovery of CNT structures of less than 100 nm that were underrepresented in compared methods. Aerosolized hydrophilic CNTs exhibited higher number concentrations than hydrophobic CNTs, while open-face samplers with MCE filters produced the highest mass concentrations. Importantly, these findings establish a reproducible, integrated sampling, and analytical framework that enhances CNT exposure assessment and supports refinement of occupational health guidelines.

1. Introduction

Rapid growth of nanotechnology has driven a significant increase in the global production and application of carbon nanotubes (CNTs), with key sectors including aerospace, electronics, construction, and medicine. Due to their desirable physicochemical properties, such as tensile strength and conductivity, global CNT production is projected to surpass $100 billion by 2030. − However, their structural similarities to asbestos fibers, including a high aspect ratio (ranging from 500 to 100,000 according to Wang et al.) and biopersistence, make them a respiratory hazard when inhaled, with studies in animals and human cells demonstrating pulmonary inflammation and fibrosis. − Reflecting these hazards, the International Agency for Research on Cancer (IARC) has classified multiwalled CNTs type 7 (MWCNT-7) as “possibly carcinogenic to humans” (Group 2B). ,

While particles across a broad size range can become airborne, the behavior of nanoparticles differs fundamentally from that of larger particles and is largely governed by size-dependent aerosol physics, which is often underappreciated in exposure assessment. Evidence from a real-life particle emission study across material life cycles shows that nanoparticle release and airborne transport vary with mechanical and processing conditions, underscoring the need for exposure assessment approaches that reflect realistic emission scenarios.

Within the risk analysis framework developed by NIOSH, the likelihood and magnitude of CNT exposure are determined by both material form and work activity. Dry CNT powders handled during dumping or sieving present a high potential for airborne release, while spraying or sonication of CNT suspensions may generate respirable droplets. Even when CNTs are incorporated into composite materials, mechanical processes, such as cutting, sanding, or grinding, can liberate respirable fragments. These release scenarios demonstrate that airborne CNT exposure is task-dependent and may vary in the particle size and concentration, reinforcing the need for robust sampling and analytical methods capable of accurately characterizing fibers across realistic occupational conditions.

Additionally, in 2013, the National Institute for Occupational Safety and Health (NIOSH) set a Recommended Exposure Limit (REL) for respirable elemental carbon (EC) of 1.0 μg/m³, over an 8-h time-weighted average (TWA) as a means of protecting against pulmonary fibrosis. , Despite this REL, CNTs remain unregulated under the Occupational Safety and Health Administration (OSHA) Permissible Exposure Limit (PEL) framework and Environmental Protection Agency (EPA) in the U.S., leaving employers and the public without enforceable regulatory benchmarks for controlling CNT exposures beyond General Industry or Construction standards, or general duty clause obligations.

A 2020 systematic review reported that 85% of field studies assessing occupational CNT exposure found EC concentrations exceeding the REL, highlighting the need for regulations and improving the current sampling and analytical methods. Additionally, a European expert panel concerned with occupational CNT exposure proposed a health-based nano-reference value (HNRVs) for engineered nanomaterials recommending the use of fiber count, rather than mass concentration, as it was deemed the more relevant dose metric for representing CNT exposure. The proposed occupational exposure limit (OEL) based on HNRVs is 0.01 fibers/cm³ for multiwalled CNTs.

While CNTs exist with dimensions that fall within the nanoscale, current sampling and analysis methods for exposure in the U.S. are based on gravimetric analysis of elemental carbon (EC) (NMAM 5040), a metric that relies on detectable mass and does not account for the unique physiochemical properties of CNT fibers and their ability to deposit deep into the lungs. Recognizing the limitations of gravimetric analysis for CNTs, NIOSH adapted a method, NMAM 7402, that was originally developed for micrometer-scale asbestos fibers. This revised method involves collecting airborne particles on mixed cellulose ester (MCE) filters using 25 mm open-face cassettes to approximate the inhalable fraction. After sampling, particles collected on the top of the filter are transferred to a transmission electron microscopy (TEM) grid through indirect analysis. However, this indirect analytical approach is technically challenging when applied to nanoscale fibers since analytes get trapped within the filter and can be lost during the transfer process, resulting in misclassification of fiber size and counts. Direct collection of airborne nanoparticles onto TEM grids for subsequent analysis has been previously established in the literature as an effective method for nanoscale particle characterization. Given the current and previously highlighted limitations surrounding CNT sampling and analysis, in this study, it was hypothesized that there is a systematic underestimation of CNT fiber exposure, especially those within the submicrometer and nanometer ranges. To address these technical challenges, a protocol for sampling and analyzing a subcategory of CNTs called multiwalled CNTs (MWCNTs) was developed. MWNCTs are composed of multiple concentric, cylindrical, graphene sheets, which are often used in exposure studies. ,,

The performance of three 25 mm samplers was evaluated in this study: open-face cassettes (targeting the inhalable fraction of airborne particles), closed-face cassettes equipped with cyclones (targeting the respirable fraction), and the Tsai Diffusion Sampler (TDS), a diffusion-based sampler whose novel design integrates a controlled inlet probe diameter, internal geometry, and low-flow airflow profile to achieve a defined 50% cutoff aerodynamic diameter (D ₅₀ = 3.8 μm) designed for collecting nanometer-sized particles via enhanced diffusion-dominant mechanisms.

In parallel, the influence of filter material, specifically MCE and PC, on both particle capture and downstream analytical performance was examined. PC filters with smaller pore size (0.22 μm vs 0.45 μm for MCE), featuring a smoother surface morphology and enhanced collection due to small pores, have historically offered superior contrast in SEM imaging directly after sampling, , thereby eliminating the transfer steps required for MCE filters and reducing the potential for fiber loss due to larger porous structure and solvent-involved transfer required in using MCE filters.

In this study, both hydrophobic and OH-functionalized hydrophilic MWCNTs were aerosolized and sampled to evaluate how CNT physicochemical properties interact with various sampler–filter configurations and their performance in evaluating airborne CNT across different particle size fractions relevant to respiratory deposition. To improve analytical efficiency and reproducibility, a custom image segmentation pipeline based on Meta’s Segment Anything Model (SAM) was integrated to automate the detection, sizing, and enumeration of CNT fibers in electron microscopy (EM) images. This computational tool enhances both the speed and consistency of fiber analysis, offering a scalable approach to high-throughput image processing. In this study, the combined effects of sampler design, filter media, and analytical workflow on CNT fiber detection, size resolution, and sample integrity were investigated. The goal was to develop a validated, CNT-specific sampling and analysis protocol that improves exposure assessment accuracy and informs future exposure guidelines for nanomaterials.

2. Materials and Methods

2.1. Experimental Configurations

Figure presents the experimental setup that a glovebox contains a beaker filled with bulk CNT powder, which is stirred to generate aerosols. Direct-reading instruments, NanoScan scanning mobility particle sizer (SMPS), and optical particle sizer (OPS) were attached to conductive tubing placed inside the glovebox. Samplers attached to pumps via conductive tubing were also placed inside the glovebox.

1.

Experimental setup.

Six sampling configurations were tested (Table ), combining three 25 mm sampler typesopen-face cassettes representing the inhalable fraction, closed-face cassettes with cyclones targeting the respirable fraction, and the TDS designed to capture both respirable and nanoscale particleswith two filter media. The first filter type was a PC Isopore filter (0.22 μm pore size, 25 mm; MilliporeSigma, MA, USA) and the second was an MCE membrane filter (0.45 μm pore size, 25 mm; SKC Inc., PA, USA).

1. Sampler–Filter Configurations for This Study (A–F) .

filter and sampler type	open face (inhalable)	closed face (respirable)	TDS (respirable)	sample analysis (2 MWCNT types)	analytical lab
grids + PC	A × 3	B × 3	C × 3	direct TEM (18 grids)	internal campus facility
				direct SEM (18 filters)
grids + MCE	D × 3	E × 3	F × 3	direct TEM (18 grids)	internal campus facility
				direct SEM (18 filters)
				indirect TEM (9 MCE filters → 27 grids)	external vendor

^aThree samplers: open face, closed face, and Tsai diffusion sampler (TDS) were combined and compared with two types of filters: mixed cellulose ester (MCE) and polycarbonate (PC). Each configuration was repeated three times, then doubled for each type of multiwalled carbon nanotube (hydrophilic and hydrophobic). Indirect (D–F) and direct TEM (A–F) analyses were compared for fiber retention. Direct SEM (A–F) analysis was conducted across all configurations.

For each experiment, all three samplers were operated simultaneously with the same filter type. One set of experiments was performed with PC filters, and another set was performed with MCE filters. Each set was repeated three times for both hydrophobic and hydrophilic MWCNTs, representing research- and industrial-grade products commonly used. This design resulted in six configurations (Table : A–C for PC, D–F for MCE) and a total of 36 samples (6 configurations × 3 replicates × 2 MWCNT types).

2.2. MWCNT Types

The outer diameter (OD) of industry-grade MWCNTs is 15 ± 5 nm, with a length range of 5–20 μm, and the specific surface area (SSA) is between 200 and 400 m²/g. The hydrophobic MWCNTs had a purity > 85 wt %, with residual impurities including iron and sulfur. Hydrophilic MWCNTs, which were functionalized with hydroxyl (−OH) groups were selected to represent commonly used research-grade CNT. It featured a similar OD of 10–20 nm and slightly longer fiber lengths ranging 10–30 μm. The hydrophilic MWCNTs exhibited a high purity of >99.9 wt % and an SSA of approximately 100 m²/g. Both MWCNT bulk powders were used in the condition they were received, without further modification or size classification prior to aerosolization.

2.3. Grid and Filter Preparation

Prior to sampling, TEM grids (pure carbon film 400-mesh copper grid or carbon Type B, 300-mesh, nickel grid from Ted Pella, Inc., Redding, CA, USA) were mounted directly onto both MCE and PC filters both at the center and 50% radius (see Figure S1) to enable direct analysis of deposited particles, thereby eliminating the filter-to-grid transfer step required in indirect analysis. This approach minimizes particle loss and preserves the arrangement of CNT structures as they are sampled. For hydrophobic CNTs, pure carbon film 400 mesh copper grids were selected as the higher grid density provides greater mechanical stability, which is advantageous given the tendency for CNTs to form agglomerates. In contrast, carbon-filmed 300-mesh nickel grids were used for hydrophilic CNTs to provide an increased open viewing area. The use of both mesh types ensured an appropriate balance between grid stability and the viewable surface area for different CNT types. Filters with affixed grids were weighed on a microbalance with an antistatic kit (model XPR2U, material no. 30279196 and antistatic kit, material no. 30499859; METTLER TOLEDO, Greifensee, Switzerland) to eliminate electrostatic charge and were conditioned in a Ruggard Electronic Dry Cabinet (50L) at 30% relative humidity before and after sampling for gravimetric analysis.

2.4. Sampling Setup and Procedure

Experiments were conducted inside a fully enclosed glovebox with a built-in ultralow particulate air (ULPA) filtration system to eliminate ambient aerosol interference and to ensure no direct contact to the researchers. All three samplers were prepared with either PC or MCE filters and were mounted on stands at the same horizontal plane, equidistant from the aerosol source, and connected to calibrated GilAir Plus air sampling pumps (Sensidyne, St. Petersburg, FL, USA) (Figure ). Samplers were simultaneously operated to ensure uniform exposure conditions. Flow rates were set at 2.0 L/min for open-face and closed-face samplers. The TDS as described earlier was set to 0.3 L/min to promote a low air velocity (1 cm/s) through the filter face, thus enhancing nanoparticle collection efficiency approaching 100% via Brownian motion with a large root-mean-square distance (X _rms) of particles moving on the filter face. The differences in sampler design and flow rate will inherently impact fiber collection, thus the resulting mass concentrations.

During separate experiments, hydrophilic or hydrophobic CNT powders were aerosolized by mechanically stirring bulk material in an open beaker (Figure ) using a digital overhead stirrer (Model 50006–01, Cole–Parmer Instrument Company, Vernon Hills, IL, USA). Before hydrophilic CNTs were aerosolized, discharge treatment (PELCO easiGlow Glow Discharge Cleaning System; Ted Pella, Inc., Redding, CA, USA) was applied to both the filters and affixed grids immediately before sampling to enhance hydrophilicity of the substrate surfaces to improve particle adhesion. Particle size distributions and number concentrations of aerosolized MWCNTs (Figure ) were measured using two different direct-reading instruments, NanoScan SMPS, abbreviated to SMPS herein (Model 3910, TSI, Shoreview, MN, USA) and an OPS (Model 3330, TSI, Shoreview, MN, USA). It was concluded from our previous testing that stirring at 300–500 rpm allowed us to consistently control the number concentration of aerosols generated for experiments. The ULPA filter was turned off just prior to stirring to allow for accurate measure of size and concentrations generated during sampling. When stirring started, all pumps were turned on for a sampling duration of 120 min based on other trials and guidance reported by Birch et al. to avoid filter overloading. Additionally, 10 min background concentrations (before and after sampling) were subtracted from the number concentrations measured during the sampling period to reveal true concentrations generated solely from CNTs during the 120 min sampling period. To further investigate the effects of the filter material on mass loading and fiber collection, additional side-by-side sampling experiments were conducted. Identical samplers (either two TDS or two open-face cassettes) were positioned next to each other and equidistant from the aerosol source. One was fitted with an MCE filter and the other with a PC filter. This arrangement ensured that both samplers were exposed to identical aerosol conditions.

2.5. Statistical Analysis

The data set for statistical analysis was categorized based on mass concentrations determined for each sample across all configurations (Table : A–F). A three-way analysis of variance (ANOVA) was used to evaluate the main effects and interactions of the sampler type, filter medium, and CNT type on measured mass concentrations. A three-way ANOVA was selected due to its suitability for detecting statistically significant differences among multiple factor levels and their interactions. Additionally, a multiple linear regression model was applied to quantify the magnitude and direction of effects attributable to the sampler type and filter medium, and to assess whether filter effects differed across sampler configurations. The linear model specified open-face cassettes with MCE filters as the reference condition.

The model took the form below in eq

$$Y_i={\beta }_0+{\beta }_1{X}_{\{\mathrm{PC}\}}+{\beta }_2{X}_{\{\mathrm{TDS}\}}+{\beta }_3{X}_{\{\mathrm{closedface}\}}+{\beta }_4(X_{\{\mathrm{PC}\}}\times X_{\{\mathrm{TDS}\}})+{\beta }_5(X_{\{\mathrm{PC}\}}\times X_{\{\mathrm{closedface}\}})$$ 1

In this model, Y _i represents the measured mass concentration in mg/m³ for observation i. X variables are binary indicators (0 or 1), representing the presence of a particular filter or sampler type.

• X _{PC} = 1 if the filter type is PC and 0 if MCE.

• X _{TDS} = 1 if the sampler type is TDS and 0 otherwise.

• X _{closedface} = 1 if the sampler type is closed face and 0 otherwise.

Coefficient interpretations:

• β₀: The mean mass concentration for the reference group (open face + MCE),

• β₁: The effect of switching from MCE to PC filter (with open-face sampler),

• β₂: The effect of switching from open-face to TDS sampler (with MCE filter),

• β₃: The effect of switching from open-face to closed-face sampler (with MCE filter),

• β₄: The additional effect of using a PC filter with a TDS sampler (beyond additive main effects), and

• β₅: The additional effect of using a PC filter with a closed-face sampler (beyond additive main effects).

2.6. Analytical Method

Three EM techniques outlined below were employed and compared in this study for their suitability to characterize all CNT structures, including fibers, for each configuration (Table ): direct SEM, direct TEM, and indirect TEM analyses.

2.6.1. Direct SEM

After sampling, all PC and MCE filters were cut into radial pie-shaped segments extending from the center to the edge. A smaller subsection from the central region (approximately 5 mm from the center) was trimmed from the segment and mounted onto an aluminum SEM stub using conductive carbon tape to support conductivity during imaging. The filter section was cut from a region distinct from the TEM grid location to prevent disturbance of particles collected on the grids. Prior to analysis, the samples were sputter-coated with a thin layer of gold (3–4 nm) for 60 s using a Pelco SC-7 sputter-coater (Ted Pella, Inc., Redding, CA, USA). This coating is required to enhance the surface conductivity and image quality. The SEM was operated at an accelerating voltage of 10 kV, enabling visualization of both fine and coarse particle fractions on the filters. This was performed at a university facilityElectron Imaging Center for Nanosystems (EICN) at the California NanoSystems Institute (CNSI) using the Teneo LoVac SEM model number 1068566 (Thermo Fisher Scientific, Massachusetts, USA).

In SEM, incident electrons interact with the surface of the sample, generating secondary electrons that are detected to produce images with a strong topographic contrast. This provides a three-dimensional appearance, making it well-suited for observing the surface morphology and spatial arrangement of CNT structures.

In conventional bright-field TEM, electrons are transmitted through an ultrathin grid-supported sample, producing a two-dimensional projection in which contrast reflects differences in densitydenser regions appear darker. This makes TEM particularly well-suited for counting estimations and for the smallest CNT structures, often down to 20 nm in width, that may not be visible with SEM. With its higher magnification and resolution, TEM can reveal internal structural details and fine features of individual CNTs, providing nanoscale information that complements the surface morphology captured by SEM.

Filter samples were systematically scanned, and representative images were taken at multiple magnifications: (1) low magnification to provide an overview of particle distribution across the filter; (2) intermediate magnification to capture a variety of particle morphologies and sizes, and (3) high magnification to examine specific features of agglomerated structures and fine fibers. These images were used to qualitatively assess the morphology and spatial distribution of particles captured on the filters and later serve for quantitative sizing and counting purposes.

2.6.2. Direct TEM

After sampling, TEM grids affixed to the filters at approximately 50% radial location were carefully removed and directly analyzed without additional preparation using a T12 TEM instrument (Thermo Fisher Scientific, Massachusetts, USA) at CNSI. The pure carbon-coated grids do not have the holey texture characteristic of filter substrates and therefore retained the smallest CNTs. At higher magnifications, TEM provides sufficient resolution to observe fine structural details of CNTs, including walls and internal morphology, making it particularly suited for visualizing small CNTs with diameters on the order of 20 nm. Each grid was initially examined at low magnification to perform a systematic sweep of the surface and assess the overall particle loading. Areas containing visible agglomerates were subsequently investigated at higher magnifications to determine whether individualized nanotubes or smaller CNT structures were present within or around the agglomerates. In parallel, four randomly selected grid cells were imaged at high magnification to capture representative fields independent of the agglomerate location. This combined approach ensured that both the agglomerated and dispersed CNT structures were evaluated.

2.6.3. Indirect TEM

Nine MCE filters from the hydrophilic MWCNT configurations (Table , D–F) were sent to an external licensed vendor for particle transfer from the filters onto TEM grids. Each filter produced three grids, resulting in a total of 27 grids along with 3 blanks prepared for indirect analysis following the NMAM 7402 protocol, allowing comparison with the newly developed analytical method in this study. The vendor prepared the samples using 100-mesh carbon-coated copper grids, which were then returned to our research team for imaging. These grids were analyzed at CNSI using the same TEM and following the same imaging procedures described for direct analysis. Overall, the study comprised direct analysis of 36 TEM grids and 36 SEM filters at CNSI, in addition to 27 TEM grids prepared through the vendor’s MCE filter transfer process.

2.7. Segmentation Algorithm

To quantify particle morphology and size from EM images, an automated segmentation workflow was developed based on the Segment Anything Model (SAM), a recently released foundation model for computer vision developed by Meta AI. Version 1 of SAM was used with the “vit_h” transformer-based model checkpoint, pretrained on the SA-1B data set containing over a billion segmentation masks. Although SAM was originally trained on natural image data sets, its zero-shot generalization makes it highly suitable for scientific imagery. In the workflow, each image from SEM and TEM analysis was processed using a bounding box prompt covering the entire image, allowing SAM to identify and isolate all particle-like features without requiring point prompts or manual annotation.

To prepare the images for segmentation, a preprocessing step was performed to standardize the bit depth and enhance image contrast, ensuring compatibility with the model. It was found that SAM performed particularly well on high-magnification images where particles were visible against a clean, uniform backgroundsuch as PC filter surfaces without visible pores. In these cases, the model produced accurate masks of agglomerates and individual fibers, enabling straightforward downstream measurement of the projected area and equivalent circular diameter of a particle.

In some cases, particularly on PC filters at lower magnifications, the filter background exhibited a patterned texture due to visible pores, leading to what is effectively a “dual background.” This presented challenges for SAM, which occasionally mistook pores for particle edges or returned fragmented masks. To address this issue, an adaptive background thresholding step was applied to images with visible pores. This preprocessing enhanced the contrast between particles and the porous filter texture, helping SAM to better distinguish true particle boundaries from background features. While simple, this addition significantly improved the segmentation quality across complex images.

All image processing, mask analysis, and data extraction were implemented using Python. To automate the conversion of pixel-based measurements to real-world units, a scale calibration pipeline using EasyOCR, a deep-learning-based optical character recognition library, was developed. The OCR system identified scale bar text embedded within the images, which was then parsed using regular expressions to extract physical distances and pixel lengths. These values were used to calibrate size measurements such as equivalent diameter and projected area across the data set.

3. Results

3.1. Number Concentration Measured by Real-Time Instrument

Figure shows that both CNT types produced unimodal size distributions in the 10–420 nm range measured by the SMPS (Figure A), and right-skewed distributions in the 0.3–10 μm range measured by the OPS (Figure B). Hydrophilic CNTs consistently exhibited higher number concentrations than hydrophobic CNTs across experiments. These results demonstrate that hydrophilic CNTs disperse more effectively, generating higher airborne concentrations throughout the ultrafine and submicron fractions. These aerosol characteristics provide important context for interpreting sampler and filter performance in subsequent sections.

2.

Average number concentration and size distributions of hydrophobic and hydrophilic multiwalled carbon nanotubes (MWCNTs) measured across all experiments. (A) SMPS distribution (10–420 nm, electrical mobility diameter) and (B) OPS distributions (0.3–10 μm, optical diameter).

Figure shows the concentration and size distribution of aerosolized hydrophobic (blue) and hydrophilic (orange) MWCNTs, with background levels subtracted to reflect aerosols generated during active stirring. These data sets represent the average concentrations for all configurations (Table : A–F) across each 2-h sampling period. Concentrations in the 10–420 nm size range measured by the SMPS (A) were orders of magnitude higher than those in the 0.3–10 μm range measured by the OPS (B), indicating that aerosolized CNTs were dominated by submicron particles within the measured size range.

In the 10–420 nm size range, both materials exhibited unimodal distributions centered around 86.6 nm. Hydrophilic CNTs peaked around 80–90 nm at 3032 particles/cm³, approximately double the peak concentration for hydrophobic CNTs (1510 particles/cm³). The 0.3–10 μm size range measured by OPS showed a similar relationship, with hydrophilic CNTs (maximum of 226 particles/cm³) consistently outnumbering hydrophobic CNTs (maximum of 48.1 particles/cm³) at 0.37 μm. Temporal patterns also differed (Figure S2): Hydrophilic CNTs gradually increased over a longer duration, reaching peak levels later in the trial, while the hydrophobic ones rose quickly but plateaued earlier and at lower concentrations. These findings experimentally validate previous findings that “surface modification is necessary to overcome CNT agglomeration and enhance dispersibility”. Figure S3 demonstrates number concentrations for hydrophilic MWCNTs experiments to demonstrate the patterns of aerosolization across six separate experiments under the same conditions.

Functionalization of CNTs through the addition of hydroxy (−OH) groups to the nanotube surface reduces van der Waals cohesion and introduces polarity, promoting hydrogen bonding with ambient water molecules. , Functionalization can also introduce structural defects (e.g., conversion of sp² to sp³ carbon) that reduce mechanical strength, making nanotubes more susceptible to fragmentation under shear stress, such as stirring in this study. The combined effects of enhanced dispersion and possible fragmentation contributed to the elevated airborne particle counts measured for hydrophilic MWCNTs across all sizes measured by SMPS and OPS instruments. These findings highlight an important implication: Airborne CNT exposure may be substantially higher with small fibers in laboratory or workplace settings where functionalized materials are used compared to hydrophobic CNTs.

3.2. Gravimetric Analysis of CNTs Collection

Gravimetric analysis was performed to quantify the average airborne mass concentrations of CNTs collected across different sampler and filter configurations. This approach reflects the current method recommended by NIOSH and provides a baseline for evaluating how the sampler design, filter material, and CNT type influence total particle collection. On the basis of mass concentrations calculated, closed-face samplers systematically underperformed relative to open face and TDS.

Gravimetric analysis revealed consistent patterns across sampler types, filter media, and CNT types (Figure ). On average, MCE filters collected higher mass concentrations than PC filters in every configuration, indicating a greater particle retention. Among sampler types, TDS paired with MCE filters produced the highest mass concentrations, particularly for hydrophobic CNTs (2.3 mg/m³), while closed face consistently yielded the lowest values, especially paired with the PC filter (0.01 mg/m³). Mass concentrations of hydrophobic CNTs were comparable between open-face and TDS samplers when paired with PC filters (0.63 vs 0.64 mg/m³, respectively). Overall, hydrophobic CNTs produced higher mass concentrations than did hydrophilic CNTs.

3.

Average mass concentrations of MWCNTs by the sampler type, filter media, and CNT type. Error bars represent standard error of the mean.

Statistical testing confirmed that the filter type was the primary factor influencing mass concentration (Table S1). The three-way ANOVA showed a significant main effect of filter type (p = 0.02), with MCE filters outperforming PC. No significant main effects were observed for the sampler type (p = 0.12) or CNT type (p = 0.81), and all interaction terms were nonsignificant (p > 0.19), indicating that the filter effect was consistent across sampler and CNT types.

The linear model (Table S2, with data from Figure S4) used open face + MCE (mean = 1.88 mg/m³) as the reference group, allowing direct contrasts of each sampler–filter configuration. Switching to PC filters in the open-face sampler reduced the mass concentration by 1.43 mg/m³, a marginally significant effect (p = 0.056). TDS with MCE yielded nearly identical results to open face + MCE (β = −0.03, p = 0.97), indicating no meaningful difference.

Closed-face samplers showed the strongest reductions. When paired with MCE filters, mass concentration was significantly lower than the reference (β = −1.58, p = 0.039). When paired with PC filters, closed face also collected less mass (−1.79 mg/m³, relative to the reference), though this difference was not statistically significant (p = 0.26). TDS with PC filters produced a 1.29 mg/m³ reduction compared to the reference, but this effect was also not statistically significant (p = 0.87) (Figure S4).

Mass-based concentrations, the foundation of NMAM 5040, provided useful but limited insights. As expected, MCE filters consistently retained higher mass concentrations than PC filters across the sampler types. This outcome reflects differences in the pore structure and filter thickness, where MCE produced a higher pressure drop compared to PC filters (Table S3; Figure S5). This also illustrates a methodological drawback as stated in the introduction: dense deposition on MCE filters obscured smaller fibers and increased the risk of overloading. By contrast, PC filters provided less loading and clearer particle recovery, particularly of individualized fibers.

3.3. Direct Scanning Electron Microscopy

3.3.1. MWCNT Type: Hydrophobic vs Hydrophilic

Hydrophobic MWCNTs (Figure A1–C2) were predominantly found as dense agglomerates, particularly with open-face sampling (Figure B1,B1). This agglomeration is due to strong van der Waals interactions found in hydrophobic MWCNTs. Hydrophilic MWCNTs (Figure D1–F2) are more dispersed across filter surfaces, with individualized fibers and smaller clusters especially evident on PC filters (Figure D2,E2). Fragmented fibers found in these images reflect the structural fragility and tendency to break during stirring or upon impaction with filter media. These findings are consistent with prior results showing that surface functionalization improves CNT dispersibility by weakening van der Waal forces (responsible for agglomeration) and promoting interaction with the surrounding environment. ,

4.

SEM images showing the morphology and agglomeration behavior of MWCNTs collected on different filter types and sampler configurations under hydrophobic and hydrophilic conditions. Hydrophobic MWCNT samples include panels (A1–C2). Hydrophilic MWCNT samples include panels (D1–F2).

3.3.2. Filter Type: MCE vs PC

PC filters captured more individual fibers and smaller clusters (Figure A2,C2,D2,E2) due to their small pore size, whereas MCE filters retained larger, entangled masses (A1–C1,D1,E1), resulting in higher mass concentrations (Figure ). Small fibers either embed into the filter matrix or pass through, reducing the recovery of individual fibers. This was particularly evident when comparing MCE to PC (A1 to A2; D1 to D2; and E1 to E2).

3.3.3. Sampler Type: Open Face vs TDS vs Closed Face

Among sampler types, open face (inhalable fraction) consistently yielded the largest number of agglomerates (Figure B1,B2,E1,E2). Closed-face samplers consistently yielded the lowest particle loading (Figure C1,C2,F2) due to the cutoff diameter and particle loss on the internal wall of the cyclone. The TDS retained the most individual fibers, especially when combined with PC; however, agglomerates were found. SEM images such as TDS with PC (Figure A2) demonstrate that both sampler–filter configuration and CNT surface structure (Figure E2) strongly influence the morphology and deposition patterns of the captured structures.

All configurations contained different structures (agglomerates, matrix, and individual fibers); however, the distribution of these structures reflected whether the sampler targeted inhalable, respirable, and nanosize fractions.

3.4. Direct and Indirect Transmission Electron Microscopy (TEM) Analysis

As shown in Figure , G1–G3 provide high-magnification views (200 nm scale) of individual CNTs obtained through direct TEM analysis, while H1–J3 present lower-magnification overviews (10 μm scale) illustrating particle distribution and density across grids for each sampler type (TDS, open face, and closed face) with arrows pointing out small fibers. These images highlight the influence of sampler configuration and analysis method on particle recovery, agglomerations, and visibility, providing essential insight into CNT capture and fiber resolution.

5.

TEM images of hydrophilic MWCNTs collected using three sampler types (TDS, open face, and closed face) under indirect and direct analysis configurations.

Agglomerates and individual fibers were observed across all configurations (Figure ), with their relative abundance and spatial patterns dependent on the sampler type and analysis workflow. Across the top row (Figure G1–G3; 200 nm), high magnification was used to compare indirect (G1) and direct (G2 and G3) analysis methods for grids sampled with the TDS. Results show that direct analysis captured fibers in both samples where grids were placed on MCE (G2) and PC (G3) compared to indirect analysis, which mostly shows an agglomerate (G1). At a lower magnification (Figure H1–J3;10 μm), both analysis methods were compared across all samplers: TDS (H1–H3); open face (I1–I3); and closed face (J1–J3). Images from direct analysis show clear distinction of structures with lower particle density except for samples collected on open face. Overall, TDS had fewer agglomerates compared to open face and closed face, with fibers being the predominant structures.

One important limitation of EM analysis is that it measures projected-area size in two dimensions, which does not directly represent aerodynamic behaviordetermined by geometric diameter, density, shape factor, and orientation in airflow. The hydrophilic MWCNTs in this study had a true density of 2.1 g/cm³ and a bulk density of 0.22 g/cm³, illustrating a large difference in effective mass between individual fibers and porous agglomerates. Previous research has quantified the shape factor of MWCNTs. Chen et al. and Ku and Kulkarni found that fibers typically have shape factors ranging from 1.5 to 2.6, depending on their orientation and aspect ratio, while agglomerates (especially those with fractal-like structures) can exhibit even higher shape factors, (up to 3.5) due to loosely packed morphology and internal voids. , Additionally, Baron et al. found that the aerodynamic diameter of airborne SWCNT agglomerates was up to 10 times smaller than their geometric size, highlighting how the project area alone can greatly overestimate a particle’s aerodynamic behavior. Orientation also plays a critical role in the aerodynamic diameter. A long fiber oriented perpendicular to airflow experiences greater drag and a larger aerodynamic diameter than one aligned parallel to flow. Conversely, agglomerates may appear spherical in 2D images but can have complex internal voids and high drag, resulting in elevated shape factors and unpredictable aerodynamic behavior. These factors are important when considering how particles, especially those fibrous structures such as CNTs containing substances, will behave in the environment and its route to the human’s lungs.

The calculated aerodynamic diameters of CNTs using eq and demonstrated in eq S1 range from 0.18 μm for individual CNT fibers (d _pa = 0.1 μm) to 17.6 μm for large agglomerates (d _pa = 20 μm), highlighting the discrepancy between projected-area measurements and the aerodynamic behavior

$$d_{\mathrm{a}}=d_{\mathrm{pa}}\sqrt{\left(\frac{{\rho }_{\mathrm{p}}\chi }{{\rho }_0}\right)}$$ 2

where

d _a = aerodynamic diameter (μm),

d _pa = projected-area diameter (μm),

ρ_p = effective particle density (g/cm³),

χ = dynamic shape factor, and

ρ₀ = 1.0 g/cm³

In the calculation, the projected-area diameter was 20 μm, with a bulk density of 0.22 g/cm³, a shape factor of 3.5, with a resulting aerodynamic diameter of 17.6 μm. This reduction reflects the influence of internal voids and irregular packing, which lower effective density and increase drag compared to a solid sphere of the same size. Conversely, for an individual CNT fiber, the projected-area diameter given was 0.1 μm, with a true density of 2.1 g/cm³, and a lower shape factor of 1.5. This led to a slight increase of the aerodynamic diameter resulting in 0.18 μm. Although the fiber appeared much smaller in two-dimensional imaging, its higher density amplified its effective aerodynamic size. Together, these comparisons show that projected area alone can either overestimate or underestimate the aerodynamic behavior of CNTs, depending on whether they are present as porous agglomerates or fibers.

3.5. Implementing Segmentation Algorithm and Size Distribution Analysis

The segmentation algorithm designed for this study automatically identified and isolated particle features in each direct EM image (Figure S6), enabling high-throughput measurement of equivalent diameter. Scale bars were interpreted using EasyOCR to convert pixel-based outputs into real-world units. Figure presents violin plots, showing the distribution of equivalent particle diameters (on a log scale) segmented from images for each sampler type.

6.

Violin plot of probability densities of equivalent diameter for hydrophilic MWCNTs by sampler type. Particles were segmented, sized, and counted using a custom algorithm applied to EM images for each sampler type. Particle counts: TDS, 279; open face, 186; closed face, 127.

Violin plots illustrate the probability density of equivalent particle diameters calculated on CNTs collected from SEM and TEM images of hydrophilic MWCNTs, spanning a range from 10 nm to 20 μm. All particles observed from SEM images were collected on PC filters, and TEM images were taken from a grid placed at either the center or 50% radial position on the filters. While SEM was the primary imaging modality, TEM was used to supplement individual fiber counts and resolve finer structures with both imaging techniques being useful to capture the wide range of particle sizes collected by each sampler.

The TDS sampler exhibited the broadest size distribution and was the only sampler to capture particles below 100 nm, confirming its enhanced performance for ultrafine CNT collection.

In contrast, the open-face sampler showed a narrower distribution with no particles below 100 nm observed. Although open-face samplers capture a broad size range of particles, this distribution was not fully represented in the EM data. Surface agglomerates sitting loosely on the open-face filter detach during handling and were often lost due to filter loading. Additionally, under a microscope, there is an inherent bias toward imaging smaller particles, further reducing the representation of coarse particles. As a result, the size distributions reported from open-face samples reflect only a fraction of the particles collected, emphasizing finer structures while underestimating larger agglomerates. The closed-face sampler had an even narrower spread and also lacked representation in the ultrafine range. Together, this data shows that TDS provides the most comprehensive size range, including nanoscale fibers and larger agglomerates.

The sampler design strongly influenced how CNTs were represented in the collected EM samples. Closed-face samplers consistently yielded the lowest recoveries, consistent with particle losses along cyclone walls. Open-face samplers produced higher mass concentration but were dominated by moderate-sized agglomerates. The TDS demonstrated a clear advantage, collecting a broader particle size range and uniquely capturing fibers <100 nm (Figure ). Its diffusion-based design promotes spatial separation of agglomerates and individualized fibers across the filter surface, which enhanced visualization and reduced overlap under microscopy. This separation is critical given that individualized nanoscale fibersthough low in masspose the greatest inhalation hazard compared to larger particles, which get caught in the upper airways. Results for particle counts and size distribution across different samplers are shown in Figure (open face, A; closed face, B; and TDS, C) with cumulative probability (Figure D) and probability density (Figure E) curves.

7.

Histograms: (A) open face, (B) closed face, and (C) TDS, particle counts, and size distribution collected across all sampler types with the largest observed particles at 20 μm. (D) Cumulative probability curves (log scale) showing count median diameters (CMDs). (E) Lognormal probability density curves illustrating the lower bound of the 50% cutoff for particles within the respirable size range (4 μm).

Distinct differences in particle size distributions were observed across the three sampler types (Figure A–C). Calculated aerodynamic size ranges from 0.18 to 17.6 μm. The open-face sampler collected 186 particles with equivalent diameters under 20 μm with a mean particle diameter of 1.8 μm (σ = 3.4 μm). Although most particles were found to be less than 2 μm, no particles less than 100 nm were found. The submicron region shows a right-skewed distribution. The closed-face sampler collected fewer particles (n = 127), with an overall right-skewed distribution and the same mean diameter as open face of 1.8 μm (σ = 2.4 μm). In the submicron range, there’s a bimodal distribution, and no particles less than 100 nm were found. In contrast, the TDS collected the most particles (n = 279), showing the widest size distribution with a mean diameter of 4.2 μm (σ = 4.8 μm), which represents the respirable cutoff size. Notably, the TDS was the only sampler to detect particles <100 nm (approximately 8%), highlighting its effectiveness for ultrafine particle collection through diffusion-based mechanisms.

Figure D shows the cumulative probability distributions of equivalent diameter for CNTs collected by the three sampler types with the count median diameter (CMD) marked for each sampler. The open-face sampler (orange) exhibited a CMD of approximately 520 nm. In theory, open-face samplers should yield a larger CMD because they collect the inhalable fraction including coarse agglomerates. However, large agglomerates are prone to loss during filter handling and removal, which likely shifts the observed CMD downward. The closed-face sampler (green) had a CMD near 820 nm. The TDS (blue) yielded the largest CMD at about 1 030 nm, consistent with its broader collection profile that included both ultrafine fibers and larger agglomerates.

According to the International Organization for Standardization (ISO), American Conference of Governmental Industrial Hygienists (ACGIH), and the European Committee for Standardization (CEN) standard, the 50% cutoff for particle collection is defined at 4 μm aerodynamic size for the respirable fraction. Probability density functions describe the relative frequency of particles within specific size intervals. The data from the probability density curve (Figure E) represents the lower half of this cutoff to best represent the small sizes dominated in the sampling. The curve for TDS shows the highest peak at the lowest size compared to the open face and closed face, which both have a wider distribution (Figure E).

4. Discussion

While engineered nanomaterials like CNT offer important functional benefits to various sectors, current exposure assessment approaches may still underestimate smaller airborne fractions generated in occupational settings. One study investigated nanoparticle exposure from abrasion testing of TiO₂ nanoparticle coatings. Submicron particles (100–1000 nm) had the highest number concentrations, with some nanoparticles as small as 30 nm found.

Although prior studies have used diverse methods to investigate aerosolization and sampling efficiency across different particle size ranges, − the present work addresses this methodological fragmentation by integrating diffusion-based sampling with direct SEM and TEM analyses to capture a large particle size range while incorporating an automated image segmentation algorithm that enables consistent, repeatable particle sizing and counting.

This study set out to validate and address limitations of existing sampling and analytical methods for CNTs. The findings indicate that current gravimetric and asbestos-adapted methods do not adequately represent airborne CNT exposure, particularly for submicron structures and nanosized fibers. Mass concentration measurements were useful for identifying differences between sampler and filter configurations as well as between CNT types. However, mass concentration-based metrics fail to capture particle morphology or detect individualized nanoscale fibers, underscoring their limitations for occupational and environmental exposure assessment. In contrast, the integrated protocol developed in the study, combining the TDS with PC filters, direct SEM and TEM analysis, and automated segmentation, successfully recovered individualized fibers < 100 nm, addressing the limitations of conventional approaches.

Statistical analysis confirmed that the filter material was the primary factor influencing mass concentration, with MCE filters collecting significantly more mass than PC filters, F(1, 55) = 5.36, p = 0.024. This effect is attributed to the thickness and the characteristics of having a tortuous pore path, leading to a higher pressure drop in the MCE filter.

In the three-way ANOVA, the sampler type (open face, TDS, closed face) did not have a statistically significant main effect on mass concentration, F(2, 55) = 2.21, p = 0.120. This indicates that when averaged across filter and CNT types, the three sampler designs did not differ significantly. Furthermore, none of the interaction terms involving the sampler type were significant (all p values >0.19), indicating that the effect of sampler type on mass concentration did not depend on the filter or CNT type. Finally, the linear model revealed that the closed-face sampler with an MCE filter collected significantly less mass than the reference configuration (open face + MCE), p = 0.039.

Although MCE filters consistently yield higher mass concentrations across configurations, this apparent advantage carries analytical drawbacks that the heavy particle loading on MCE obscures individualized fibers. PC filters, while collecting lower mass concentrations, did not overload as easily under the same conditions and, therefore, preserved deposition surfaces that allowed individualized CNTs to remain visible, particularly when paired with the TDS.

Sampler configuration further influenced fiber recovery. Closed-face cassettes systematically underperformed due to particle loss in the cyclone walls, while open-face cassettes collected more mass concentration but were dominated by large agglomerates. The TDS provided the clearest evidence in support of the study aim, uniquely capturing more individual fibers and promoting the spatial separation of agglomerates and fibers across the filter surface. This capability demonstrates that individualized nanoscale CNTs are indeed present in the environment but are underrepresented in conventional configurations.

Analytical workflow comparisons also underscore the need for protocol refinement. Indirect TEM introduced additional background interference from treated grids and loss of fine fibers during the solvent-involved transfer process, confirming the limitations of NMAM 7402 for nanoscale CNTs. Premounted grids for direct TEM analysis preserved CNT morphology, consistently revealing individualized fibers that would otherwise be lost or obscured. Direct SEM provided a complementary addition to the workflow by enabling broader field-of-view imaging and particle sizing at lower magnifications, which helped distinguish deposition patterns and agglomeration across filters.

The integration of an automated segmentation algorithm further strengthens the protocol by enabling reproducible, high-throughput analysis. Manual enumeration remains a critical bottleneck in structure-based CNT assessment, but this algorithm generalized effectively across imaging conditions and successfully extracted distributions spanning from 10 nm fibers to 20 μm agglomerates. This innovation reduces bias and provides a scalable alternative to indirect analysis, supporting the broader shift toward count- and morphology-based exposure metrics recommended by international parties. As far as the authors are aware, this is the first documented application of a foundation model like SAM to the quantification of CNT structures in electron microscopy. The ability to segment and analyze CNTs across varied imaging conditions with no retraining or fine-tuning represents a major advancement for high-throughput, standardized analysis. This work lays the groundwork for future studies on automated nanomaterial characterization to environmental CNT exposure and can easily expand on the technical design of this pipeline and its performance across a broader range of nanostructures. Given the generality and robustness of the approach, this framework could also have broad utility in the analysis of particulate matter in air-quality research and beyond.

Taken together, these findings demonstrate that current recommended methods systematically underestimate nanoscale CNTs, while the TDS combined with PC filters and direct EM analysis overcomes these barriers by recovering individualized fibers across a broad size spectrum. Although limitations remainincluding reliance on the projected area rather than aerodynamic diameters and the need for validation across additional CNT types and relevant environmentsthe framework developed here represents a robust foundation for advancing exposure assessment. By explicitly addressing the underrepresentation of nanoscale CNTs in conventional methods, this protocol provides a relevant framework supporting the environmental exposure assessment to engineered nanomaterials and further develops protective standards.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.analchem.5c06701.

• Additional experimental details, materials, and methods, as presented in Figures S1–S6, Equation S1, and Tables S1–S3, including photographs of grid placement and fiber segments, SMPS data, mass concentrations, relationships of pressure drop and volumetric flow rate, statistical analysis data, and applicable equations (PDF)

This study is supported by the National Institute for Occupational Safety and Health (NIOSH) under R21 Grant 5R21OH012397. The financial support to H.D. in part was provided by the Centers for Disease Control and Prevention (CDC)/NIOSH through Grant Number 5T42OH008412-16 to the Southern California Education Research Center, and the Department of Health and Human Services, National Institutes of Health, National Institute of Environmental Health Sciences (NIEHS) through Grant Number 1R25ES033043-01 to the Southern California Superfund Research Program at University of California Los Angeles. Its contents are solely the responsibility of the authors and do not necessarily represent the official view of CDC, NIOSH, or NIEHS.

The workflow was implemented in Python, using (among other components) easyOCR for automated scale extraction and scikit-image for postprocessing and connected-component quantification. The segmentation code used for analysis is publicly available at GitHub: “SAM-SEM-Segmentation”. The arXiv preprint describing the algorithm and its validation is available at doi: 10.48550/arXiv.2601.06673.

The authors declare no competing financial interest.