Rapid Microplastic Detection Using High-Throughput Screening Raman Spectroscopy

Abstract

Microplastic pollution poses a significant environmental challenge, with particles ranging from micrometers to millimeters contaminating ecosystems worldwide. Traditional Raman microspectroscopy struggles to balance spatial resolution, field of view, and throughput, especially at low particle concentrations. Here, we present a high-throughput Raman spectroscopy (HTS-RS) platform that overcomes these limitations by combining a 3.15 × 2.10 mm² field of view with a spatial resolution of 1.4 μm, enabling rapid, label-free detection and classification of microplastics across a wide size range. The system integrates automated particle recognition, autofocus correction, and Raman spectral acquisition into a seamless workflow, reducing user intervention and accelerating data acquisition. Validation on reference microplastic mixtures demonstrated precise detection from 7 μm to over 400 μm, with robust morphological and chemical characterization. With its high sensitivity, throughput, and automation, our platform sets a new benchmark for microplastic monitoring and provides a scalable solution for environmental screening applications.

Introduction

For almost a century, plastic goods have benefited society at large and are widely used in packaging, textiles, consumer goods, construction, transportation, electronics, industrial, and medical applications. These items make our lives easier, more convenient, and safer. − While plastics offer a multitude of advantages, their extensive use, particularly in disposable forms, such as packaging materials, results in their accumulation in the environment, with plastic waste accounting for nearly 10% of total municipal garbage globally. Wind, waves, and solar radiation can break down these polymers into microsized plastic particles, known as microplastics (MPs), which range from 1 μm to 5 mm. − Depending on the source, they can be classified as either “primary microplastics” or “secondary microplastics”. The former refers to all microplastic particles that are generated and released, such as raw materials used as precursors for plastic goods, including industrial pellets, fibers, or microbeads, and abrasives added to cosmetics and cleaning products. ,, Through treated and untreated urban and industrial effluent discharges or land applications of sewage sludge, these primary MPs can enter aquatic systems and soils. , When primary MPs are subjected to physical, chemical, or biological factors in the environment, such as ultraviolet radiation or physical abrasion, they degrade, yielding secondary MPs, which have irregular shapes and sizes and are the most common type of plastic debris found in the environment. ,,, The reduction in size, of course, makes the investigation and characterization of MPs extremely difficult.

Analyzing MPs from various environmental samples necessitates several steps including sampling, purification, separation, and identification. Our research focuses on the final step of the process: the identification of microplastics. The literature contains numerous studies proposing different approaches for assessing MPs in environmental samples. , Some of the most common methods for identifying MPs include Fourier-transform infrared spectroscopy (FT-IR), Raman spectroscopy (RS), high-temperature gel-permeation chromatography (HT-GPC) with IR detection, scanning electron microscopy with energy-dispersive X-ray spectroscopy (SEM-EDS), and pyrolysis-gas chromatography coupled with mass spectrometry (Pyr-GC-MS).

Generally, two analytical techniques are primarily used to assess the chemical nature of MPs: thermoanalytical methods and optical spectroscopy methods. Thermoanalytical procedures are destructive and require specialized equipment, such as pyrolyzers, to thermally decompose the material under controlled conditions. The resulting pyrolysis products are then analyzed using gas chromatography coupled with mass spectrometry (GC/MS). − Techniques such as Py-GC/MS , and thermo-extraction desorption GC/MS (TED-GC/MS) − are commonly employed for mass-quantitative MP analysis. Information from Py-GC/MS is represented as mass fraction or mass concentration of the polymers. Furthermore, the destructive nature of this process hinders further investigation of the plastic particles. For instance, burning the sample renders particle counts impossible, while details about the polymer type and their shape/size remain critical for examining the impact of MPs on the environment. , Thus, optical spectroscopy is often preferred for MP analysis and includes methods such as Raman spectroscopy (RS) , and Fourier-transform infrared spectroscopy, which examine the molecular vibrations of a sample and provide a distinct molecular fingerprint. The resulting spectra represent the molecular bonds within a substance, allowing comparison with reference spectra from a library. A key advantage of both modalities is that the information can be obtained label-free and nondestructively, offering a clear and understandable chemical interpretation.

Relative to FTIR spectroscopy, RS provides superior spatial resolution, measuring down to 1 μm compared to FTIR’s 10–20 μm. − Additionally, RS has greater sensitivity to nonpolar functional groups and shows reduced interference from water. While Raman and FTIR are complementary techniques, Raman is particularly strong in detecting nonpolar symmetrical bonds, whereas FTIR excels at identifying polar groups. , The carbon–carbon bonds, both single (C–C) and double (CC), that typically make up the backbone of plastic compounds, are readily identified by Raman spectroscopy. This high sensitivity of Raman spectroscopy to nonpolar molecular entities makes it especially suitable for plastic investigations. Although RS is recognized as an excellent choice for analyzing MPs, prior purification of the microplastic samples is essential due to significant interference from the fluorescence signal of organic and biological materials as well as inorganic contaminants. Once impurities are removed, filtration becomes a critical step in isolating MPs from either water samples or density-separated supernatants. Studies on MPs analysis have utilized filters with pores as small as 0.45 μm. However, filtering large sample volumes may clog the filters’ small pores, leading to a thick layer of natural particles accumulating on the filters. To prevent this, filter cascades can be employed, and sample volumes can be divided into smaller quantities before each subsample is analyzed. This, of course, increases the time required for sample processing, but it is necessary for an accurate representation of MP concentration. Moreover, MPs are often present at very low concentrations of only a few dozen particles per mL, making the localization of these particles on the filter for spectroscopy measurements time-consuming and cumbersome.

Most commonly Raman spectroscopy for MP analysis is based on microscopy systems known as Raman microspectroscopy. These systems are designed to achieve diffraction-limited spatial resolution for a given objective lens. Due to the relatively low intensity of the Raman signal compared to other methods, an objective lens with a high numerical aperture (NA), e.g. NA of 0.8–1.0, is commonly employed. Depending on the specific design of the optical system, the resulting field of view (FOV) for Raman systems with high-NA objective lenses typically ranges in the order of a few hundred micrometers with diffraction-limited spatial resolution. This presents a challenge when trying to identify a few MPs on a filter or substrate that measures several millimeters, a situation that is quite common. The complexity increases with the requirement for automatic detection and data capture of the MPs. We have previously reported on a high-throughput Raman system designed for the automated analysis of eukaryotic cells. This particular system utilizes an objective lens with an NA of 1.0, which, along with its optical design, yields a FOV of 83 × 62 μm. For a typical Raman substrate, measuring 12 × 9 mm, more than 21,000 FOVs will have to be measured to sample the entire substrate and locate all the MPs. This not only time-consuming but requires significant computational time to determine if there is an MP is present in the image. Furthermore, MPs in samples can span from just a few micrometers up to several millimeters, underscoring the need for a single system that combines a large field of view with high resolution. Such an approach can quickly scan wide areas while reliably detecting both the smallest and largest MPs in one pass.

In this work, we present a HTS-Raman system, which is specifically designed and tailored for Raman-based microparticle analysis in a low concentration environment. We have applied well-known optical design criteria based on the space-bandwidth product (SBP) to develop and implement a high-throughput screening Raman spectroscopy system. This system enables the largest possible FOV while maintaining spatial resolution, allowing for the characterization of a wide range of MPs. The proposed system provides a FOV of 3.1 × 2.1 mm, with a spatial resolution of 1.4 μm, facilitating the rapid detection and label-free characterization of MPs, even at very low concentrations over a large size range. By integrating automation with refined optical implementation this system offers a robust and efficient solution for the challenges faced in low-concentration setting for MPs detection and characterization.

Materials and Methods

Component Selection

Raman microspectroscopy combines a Raman spectrometer with an optical microscope, enabling optical visualization and molecular characterization of the sample under diffraction-limited conditions. As readily stated, the considerations for Raman spectroscopy are primarily guided by the requirement for increased signal collection, which is reflected in the use of a high NA of objective lenses. While this is certainly true for biological samples, in the case of MP, where carbon–carbon bonds serve as the backbone of plastics, this consideration can be greatly reduced. Since signal collection is proportional to NA, reducing NA from 1.0 to 0.3 leads to an 11-fold decrease in signal strength. While for biological samples, this reduction becomes challenging. However, for MPs, which have repeating molecular bonds, the difference can be quite acceptable. This suggests that the high-NA requirement can potentially be relaxed in the design.

There are several factors in an optical design, which are directly influenced by the NA of an objective lens, including optical resolution and FOV. Both parameters can be summarized by a single characteristic of the optical system: the space-bandwidth product (SBP). The SBP measures how much information an optical system can transfer. In other words, this dimensionless number indicates how many distinct spots can be optically resolved within a FOV. A higher SBP provides more information, meaning more diffraction-limited spots in a FOV. A detailed analysis of using the SBP to design a two-photon excitation microscope was outlined by Bumstead et al. Generally, the SBP can be approximated by

$$\mathrm{S}\mathrm{B}\mathrm{P}=\frac{2\times {(\mathrm{F}\mathrm{O}\mathrm{V})}^2}{{(0.61\lambda /\mathrm{N}\mathrm{A})}^2}$$

The FOV of a microscope is ultimately limited by the objective lens, the tube lens, camera sensor size, and other factors. Since the objective lens and tube lens are a coupled imaging system, the first part to increase the FOV is by identifying commercially available microscope objectives are suitable for large FOV imaging with good resolution. The full derivation of SBP and the numerical comparison of 21 objectives are provided in Supporting Information Note S1 (Table S1).

From Table S1 it is evident that MVPLAPO 2 × C has the highest SBP, i.e. 523 Mpixels, while also having the largest FOV at a subμm resolution, making this objective lens the best candidate for the high-throughput screening Raman system for MPs analysis. Unfortunately, this objective lens is much larger than a standard microscope objective, with a diameter of 65 mm and a height of 120 mm, closer in appearance to single-lens-reflex camera objective lenses. The next best candidate is the Olympus XFLUOR 4×, having an SBP of 40.65 million pixels with an NA 0.28, magnification 4, and field number (F.N) 22. While the NA is somewhat smaller, the achieved lateral resolution of 1.2 μm and the circular FOV of 5.5 mm has the potential to be an excellent fit for a HTS-RS system.

After evaluating the optical performance and information capacity of various objective lenses using the SBP, the next step is to ensure that this optical information is completely captured by the imaging sensor. Detailed Nyquist-sampling calculations, together with a comprehensive comparison of candidate cameras, are provided in Table S2. We explored available cameras to meet these requirements and found that the Ximea Gpixel GSENSE6060 camera offers a high pixel count of 38 megapixels. However, it comes with a pixel size of 10 μm, which is significantly larger than the 2.44 μm required for optimal resolution. While this camera might provide a larger FOV, it would not achieve the fine resolution needed to capture the detailed information provided by the Olympus XFLUOR 4×. Alternatively, the CMOS camera IDS U3-3800 SE provides 20.44 megapixels with a smaller pixel size of 2.40 μm. Although this camera has a slightly smaller diagonal FOV (approximately 4 mm) compared to the 5.5 mm FOV of the Olympus XFLUOR 4×, it still captures about 53% of the optical informationequivalent to roughly 21 million resolvable spots. This makes the IDS U3-3800 SE an excellent, budget-friendly choice for our high-throughput screening applications, offering a solid balance between performance and affordability.

System Overview

The designed Raman microscope (Figure ) integrates three relevant optical pathsthe Raman excitation path (light red), the Raman signal collection path (red), and the brightfield imaging path (blue)all aligned to enable simultaneous Raman spectral acquisition and microscopic imaging along the same optical axis. Table summarizes every optical and mechanical component used in the instrument.

1.

Schematic of the designed HTS-RS, showing three color-coded optical paths: Raman excitation (light red): a 785 nm laser is fiber-coupled, then collimated and passed through a laser line filter (LLF). The beam is directed by a dichroic beamsplitter (F1) and focused onto the sample by the 4× objective (OBJ). Raman signal collection (red): scattered light returns through the objective and F1, passes a long-pass filter (LP) to remove Rayleigh-scattered light, then encounters a short-pass dichroic (F2) that reflects the Stokes-shifted signals into lens L3, which focuses them into a multimode fiber (MMF). The fiber delivers the light to a spectrograph, where a CCD records the dispersed Raman spectrum. Brightfield imaging (blue): a white LED is collimated by L1 and illuminates the sample. Transmitted light is collected by the same objective, forming a collimated beam that is focused by tube lens L4 onto a brightfield camera.

1. List of Components Used in HTS-RS.

S. No	component	description	part	vendor
1	LED	white LED	MCWHL7g	Thorlabs
2	L1	collimation lens	ACL2520U-A	Thorlabs
3	OBJ	microscope objective	XFLUOR 4X	Olympus
4	M	silver elliptical mirror	PFE10-P01	Thorlabs
5	F1	beam splitter	RT785rdc	Chroma
6	LLF	laser line filter	RET785/6x	Chroma
7	L2	collimation lens	LA1027-B-ML	Thorlabs
8	LAS	diode laser 785 nm	iBEAM SMART 785	Toptica Photonics
9	LP	long pass filter	RET792lp	Chroma
10	BF	bright field camera	U3-3800SE Rev.1.2	IDS Imaging Development Systems GmbH
11	L4	tube lens	TTL180-A	Thorlabs
12	L3	focusing lens	AC254–0100-B-ML	Thorlabs
13	SPEC	spectrometer	Isoplane-160	Princeton
14	SMF	single mode fiber	TOTICA OK 001526	Toptica
15	MMF	multimode fiber	M96L01	Thorlabs
16	STG	stage X–Y	CONEX MFA-Series	Newport, USA
		stage Z	MTS25-Z8	Thorlabs
17	F2	short pass beam splitter	HC 749 SP	Semrock

A 785 nm single-mode laser (LAS, IBEAM-SMART-CD, Toptica, Germany) with a nominal 250 mW output serves as the Raman excitation source. The laser is fiber-coupled through a polarization-maintaining single-mode fiber and collimated by lens L2 (focal length: 35 mm; Thorlabs). A 785 nm laser was selected because this wavelength greatly reduces the excitation of fluorescence background signals from pigments and weathering products commonly present in environmental MPs. To eliminate residual in-fiber Raman signals, the collimated beam subsequently passes through a laser line filter (LLF, 785 ± 3 nm; Semrock), ensuring a spectrally pure excitation line. The filtered excitation light is reflected by a long-pass dichroic beamsplitter F1 (RT 785 rdc) and coupled to a 4× microscope objective (OBJ, NA = 0.28; Olympus), which focuses the laser onto the sample plane, forming an approximately 6 μm spot size. The sample is placed on a custom-built holder, which is attached to two motorized x–y translational stages (CONEX MFA-Series; Newport). These x–y stages are mounted on a motorized z-positioning stage (MTS25-Z8; Thorlabs), allowing for automated vertical motion and precise focusing onto the sample plane.

Raman-scattered photons from the sample are collected by the same objective lens and travel back through filter F1. A long-pass filter (LP, 792 nm) then blocks the Rayleigh-scattered excitation light at 785 nm, allowing only Stokes-shifted photons above 792 nm to pass. These Raman signals are subsequently directed by a short-pass dichroic mirror (F2, HC 749 SP; Semrock), which reflects wavelengths above 792 nm toward the detection path. The reflected light is focused by lens L3 (a 100 mm achromatic doublet; Thorlabs) into a 105 μm multimode fiber (MMF). The fiber guides the collected Raman photons to a spectrograph (IsoPlane160, Princeton Instruments) equipped with a 400 g/mm grating blazed at 750 nm. Finally, the dispersed Raman spectrum is captured by a CCD camera (PIXIS-400BR-eXcelon; Princeton Instruments) with a 400 × 1340 pixel active area, completing the detection pathway.

A white LED provides sample illumination, collimated by lens L1. The transmitted light is collected by 4× objective, producing a collimated beam that is subsequently focused by a tube lens (L4, TTL-180A; focal length = 180 mm) to form a brightfield image on the camera sensor (BF, U3-3800SE Rev. 1.2; 5536 × 3692 pixels). The LED is externally triggered by the CCD camera, ensuring it automatically turns off when Raman spectra are being acquired. This synchronization prevents any interference with the Raman measurements.

Due to the objective lens’s large back aperture, the footprint of the beam on the dichroic mirror surfaces is considerably enlarged, making optical flatness a key requirement. Even minor deviations in reflected wavefront error (RWE) introduce astigmatism or focal shifts, degrading image quality. To minimize aberrations, the short-pass dichroic (HC 749 SP, Semrock) was selected with an RWE < 2λ peak-to-valley at 632.8 nm, ensuring minimal distortion for both the Raman signal and the brightfield image.

Automated Data Acquisition

The system is operated by custom-developed software in LabVIEW (Version 2018), which is designed to automate high-throughput scanning and Raman spectral acquisition of randomly scattered MPs (see Figure for the user interface). Before measurements begin, the software calculates stage tilt in both the x and y directions using a contrast-based autofocus routine. This routine moves the stage through multiple z positions, applying a Sobel filter to each captured image to detect edges. The average edge magnitude is used as a sharpness score, and the z position with the highest score is selected as the optimal focus. To correct stage tilt, we autofocus at the first and last frames in the scanning direction, record their optimal z-positions, and then linearly interpolate z for each intermediate frame. This two-point method produces a tilt slope that ensures consistent focus across all frames without requiring autofocus on every single frame. Once the tilt correction is determined, the high-throughput scanning process begins.

2.

LabVIEW-based “main page” interface for automated microplastic (MP) detection and Raman spectroscopy. (a). File path controlusers specify the main folder, subfolder, and file name, then start or end a measurement program. (b) Brightfield imaginga live/preview window of the sample (here showing microplastic particles on a substrate), with motorized stage controls (X, Y, Z) and field-of-view (FOV) settings. (c) Tilt correction and autofocusthe system performs a Sobel-based autofocusing at the first and last frames along each axis to determine optimal Z-positions. It then linearly interpolates Z across the scan, compensating for sample tilt and ensuring consistent focus in every field of view. (d) MP detectionprocessed image identifying scattered particles; real-time updates include exposure, total MP count, and stage coordinates. (e) Raman spectrum displayPlots the real-time Raman signal amplitude as spectra are acquired. Calibration compound selection and the number of calibration spectra can also be set here for reference measurements.

Initially, the laser spot is already defined in the camera’s field of view. The user specifies the number of frames to acquire, and the software generates an XYZ movement matrix based on the known FOV dimensions and the stage tilt correction. Specifically, X and Y positions are computed from the FOV size along each axis, while Z is adjusted according to the tilt function derived from autofocus during the initial and final frames. This matrix produces a snake frame sequence by reversing the X coordinate progression on alternating rows, thereby minimizing stage travel and overall scan time. During high-throughput scanning, autofocus is conducted only on the first frame; all subsequent frames depend on the precomputed tilt corrections. For each frame, an image is captured and processed to locate particle positions, as explained in the next section, yielding the pixel coordinates of each detected center. The software then calculates the pixel offset between the laser spot and each particle center, converting that offset into real units using a calibrated pixel-to-micron factor. The stage then moves the laser onto each particle sequentially, and a Raman spectrum is acquired on individual particles.

After processing each frame, the frame count is updated, and the loop continues until all specified frames are scanned. Finally, the stage returns to its starting position. At a 0.5 s integration time per particle, this automated procedure can measure roughly 2500 particles per hour (∼1.4 s per particle). Images of each frame are saved, and spectra are recorded with their corresponding positions (in pixels) and their morphological data, e.g. max Feret diameter, perimeter, area, and Heywood circularity factor. This integrated approach ensures precise, automated Raman measurements of randomly scattered particles with minimal user intervention.

Particle Recognition

Different types of samples require distinct particle recognition methods. Here, we employed a series of morphological image operations on the brightfield images. After acquiring the image (see Figure a), Otsu’s thresholding is applied to convert the grayscale image into a binary image (Figure b). Otsu’s method automatically determines an optimal threshold value by analyzing the image histogram and minimizing intraclass variance. This ensures better adaptability to variations in image contrast and improves the separation of the foreground (objects of interest) from the background. This simplification facilitates further image processing and analysis by clearly distinguishing microplastic particles from the backdrop. To ensure that particles are properly connected and solid, the binary images undergo a combined process of dilation and hole filling (Figure c). Dilation connects disjointed particles by expanding their boundaries, while hole filling ensures that any internal voids within particles are filled, resulting in solid representations of the objects. We intentionally chose not to perform any erosion or remove objects from the edges of the image, as our objective was to detect every particle present, even those partially within the frame, so that we do not miss any. Consequently, if a single large particle extends across multiple frames, it may be counted more than once, but this guarantees each particle is detected at least once. The final step involves identifying the pixel coordinates of the centroids and overlaying them on the original brightfield images (Figure d) for clear visualization and accurate localization. Finally, the pixel coordinates of each particle’s centroid along with the corresponding morphological information, providing essential data for subsequent analyses.

3.

Stages of particle recognition in brightfield images through morphological image processing. (a) Original grayscale image showing particles of interest. (b) Binary image after applying Otsu threshold to distinguish foreground (particles) from the background. (c) Processed binary image after dilation and hole filling to connect disjointed particles and ensure solid representation. (d) Final image with centroids of detected particles overlaid (marked by red dots) for clear visualization and accurate localization.

Raman Data Processing

Raman spectra were analyzed in Python, starting with wavenumber calibration using a 4-acetaminophenol reference spectrum. The known Raman peaks of 4-acetaminophenol were utilized to align the CCD pixel positions with their corresponding wavenumbers through a fourth-order polynomial fit, ensuring accurate conversion from pixel to wavenumber. After calibration, a rolling-ball baseline correction (window size = 90 data points) was applied to eliminate any polynomial background contributions common to fluorescence or scattering. The data were then smoothed using a Savitzky–Golay filter (window length = 11, polynomial order = 3) to reduce high-frequency noise while preserving narrow spectral features. Finally, the spectra were area-normalized for consistent comparisons across different samples. This standardized data set served as the foundation for subsequent analyses, such as principal component analysis (PCA) and clustering, performed on the preprocessed spectra.

Samples

In this study, we utilized a range of reference microplastic particles (polyethylene terephthalate (PET), polyethylene (PE), and polyvinyl chloride (PVC)) to characterize the system. All particles were produced and provided by WESSLING Hungary, with sizes spanning approximately from 10 to 400 μm. These particles were prepared using a cryogenic grinding method, where larger plastic foils were first cut into small fragments and then ground under cryogenic conditions to prevent melting, resulting in a range of particle sizes. The details of this production method are described in the study. Additionally, smaller polystyrene (PS) particles with an average diameter of 7 μm were sourced from Microparticles GmbH (product number PS/Q-F-L-2883). The combination of these particles allowed us to achieve a comprehensive representation of the various types of MPs typically encountered in environmental settings. For deposition of microplastic samples we used CaF₂ (12 × 9 mm²).

Characterization

To characterize the design parameters of our HTS-Raman system, we conducted standard measurements to determine spatial resolution, FOV, and laser spot size. Various methods can be used to assess the spatial resolution of an optical system, such as using standard specimens like diatoms, conducting a fluorescent knife-edge test, measuring the point spread function (PSF) with subresolution beads, or utilizing a resolution target. We selected a 1951 USAF resolution chart, which utilizes patterns with known spacings to establish resolution by identifying the smallest line spacing that the microscope can clearly resolve.

Spatial Resolution

Figure a presents the full image of the 1951 USAF resolution target, with an inset highlighting groups 6–9, the region containing the finest features on the chart. In optical systems, the modulation transfer function (MTF) is often used to quantify how contrast is preserved at different spatial frequencies. MTF can be derived from sinusoidal patterns or via the Fourier transform of a system’s point spread function. However, when using a bar-pattern testsuch as the USAF chart’s square-wave linesthe resulting measurement is typically referred to as the contrast transfer function (CTF). For each set of bars, we computed the contrast using the formula (I _max – I _min)/ (I _max + I _min), where I _max and I _min represent the peak (bright) and valley (dark) intensities, respectively. We then plotted these values against spatial frequency (in line pairs per millimeter) to construct the CTF curve, which reflects how effectively the system reproduces image contrast across varying spatial details (Figure b). Based on these CTF curves, we adopted a 20% contrast threshold to define our resolution limit. , According to this criterion, our system resolves 1.4 μm (group 8, element 4) at 20% contrast in both the horizontal and vertical directions.

4.

System characterization: (a) full view of the 1951 USAF resolution target, with a red inset highlighting groups 6–7 and a green inset highlighting groups 8–9. (b) Contrast transfer function curves derived from horizontal and vertical bar patterns, indicating a 1.4 μm resolution limit at 20% contrast, which corresponds to group 8, element 4 on the chart. (c) Image of the laser spot focused on the sample, with intensity profiles extracted in the horizontal (red line) and vertical (yellow line) directions. The intensity profiles are fitted with Gaussian functions, resulting in full width at half-maximum (fwhm) values of approximately 6.4 μm horizontally and 5.8 μm vertically. (d) 2 × 2 stitched frames captured before slope correction. The colored boxes mark the boundaries between frames, revealing misalignment at those edges. (e) 2 × 2 stitched frames captured after slope correction. The colored boxes again mark the frame boundaries, now showing markedly improved alignment at the edges.

Field of View

With the spatial resolution established, we proceeded to calculate the FOV to characterize the system’s overall imaging capability. We selected group 2, element 2 for this calculation due to its well-resolved features. For element 2, the line width was measured to be 194 pixels. Using the known physical line width of 111.36 μm, we calculated a pixel-to-micron conversion factor of 0.57 μm/pixel. Applying this conversion factor to the image dimensions of 5536 × 3692 pixels, we determined the FOV to be 3.15 mm × 2.10 mm².

Laser Spot Size

The brightfield imaging arm and the Raman excitation arm are two distinct optical systems with different optical characteristics. For the Raman excitation arm, the size of the focal spot is critical to ensuring that MPs can be accurately detected and analyzed. This process helps determining the smallest particles that can be effectively identified. To assess the focal spot size, an image of the laser spot focused on the sample was captured using a bright-field camera. To remove unwanted background contributions, a background image without the excitation laser was captured and subtracted from the laser-illuminated image. The intensity profiles in both horizontal and vertical directions were fitted with a Gaussian function, yielding full width at half-maximum (fwhm) values of approximately 5.8 and 6.4 μm respectively, see Figure c. Particles smaller than the spot size can still be detected, however, their Raman signals may overlap with the surrounding material, complicating the spectral analysis.

Stage Alignment and Slope Correction

While our large FOV (3.1 × 2.1 mm²) is advantageous for scanning broad sample regions, it is also crucial that the motorized stage can move precisely to each intended coordinate. Any minor mechanical misalignmentespecially nonorthogonality between the x and y axesbecomes increasingly problematic at this scale, where a small error can translate into a noticeable positional offset. As illustrated in Figure d, when 2 × 2 frames are stitched together, these small but systematic shifts accumulate at the boundaries, resulting in visible misalignments. Furthermore, such stage misalignment does not merely affect stitched images; it also compromises measurement accuracy when the stage travels to different positions within a single frame.

To address this, we quantified the unintended motion in one axis, i.e. y, when commanding displacement in the other, i.e. x, and vice versa. From these measurements, we derived two slopes

$$m_x=\frac{\mathrm{\Delta }y}{\mathrm{\Delta }x}=0.006,m_y=\frac{\mathrm{\Delta }x}{\mathrm{\Delta }y}=0.002$$

where Δx and Δy are the nominal stage moves along each axis. We then applied correction factors whenever moving from (x,y)­to­(x′,y′)

$$X=x\pm m_x\mathrm{\Delta }y,Y=y\pm m_y\mathrm{\Delta }x$$

with Δx = x′ – x and Δy = y′ – y. Incorporating these corrections into our motion commands ensures the stage consistently reaches its intended coordinates. Implementing this slope correction significantly reduces misalignment, which is critical for measurement reliability in a high-throughput workflow. To illustrate this improvement, we stitched a 2 × 2 grid of frames while applying the correction. As shown in Figure e, perfect edge matching is rarely possible. However, adjusting the slope significantly reduces nonorthogonality and greatly improves alignment between adjacent frames. Notably, the X/Y stage’s guaranteed absolute accuracy is ±3 μm (with typical performance near 0.7 μm), placing a practical limit on the overall positioning precision despite these corrections.

Assessment of Measurement Quality

We define “system assessment” as the evaluation of the HTS system’s ability to accurately locate plastic particles and collect their Raman spectra. We examined targeting precision by scanning 7 μm polystyrene (PS) beads and measuring the intensity at a known peak. This approach quantifies how effectively the stage centers on each particle, revealing any positional offsets and their impact on the resulting Raman spectra. These beads were chosen for their uniform size and well-defined Raman spectral features.

For this measurement, we first manually aligned the laser on individual beads and collected 0.5 s spectra at 10 different beads; the average (with standard deviation) serves as a reference (Figure a) and shows a ∼225 counts at 1001.5 cm^–1. This corresponds to the achievable number of counts for the given conditions and a perfect positioning on the bead. We then scanned a 2 × 2 grid of frames, each covering 3.15 × 2.10 mm² (for a total area of 6.3 × 4.2 mm²), detecting 1348 particles. With 0.5 s per particle, the complete scan took ∼1941 s, corresponding to ∼1.44 s per particle.

5.

(a) Reference Raman spectrum of a 7 μm polystyrene (PS) bead, highlighting the 1001.5 cm^–1 peak. The shaded area shows the standard deviation from 10 precise position measurements on the beads. (b) Histogram of the 1001.5 cm^–1 intensities expressed as a percentage of the 225-count reference, spanning 0–115%. The numeric labels above each bar indicate the number of particles in that percentage bin. (c) Representative Raman spectra of 7 μm polystyrene (PS) beads grouped by their 1001.5 cm^–1 intensity relative to the 225-count reference: (0–30)%, (30–60)%, (60–90)%, and (90–115)%. The number of particles in each bin is shown in parentheses. Spectra below 30% exhibit noticeably weaker PS features, while higher-intensity bins yield more distinct peaks. (d) Brightfield image demonstrating the system’s capability to handle a broad particle-size rangefrom a 7 μm PS bead (blue inset) to a 411 μm polyethylene terephthalate (PET) particle (dashed yellow line denotes length). (e) Z-scan intensity profiles of PS and PET, showing a 46 μm offset between their respective optimal focal planes; despite this difference, both yield robust Raman signals at a single focal setting (inset magnifies the maxima).

To evaluate targeting precision, we converted each particle’s 1001.5 cm^–1 intensity into a percentagetaking our 225-count average as 100%and observed values ranging from 0% to ∼115% (see Figure b). Only nine spectra fell below 30%about 0.7% of the totalsuggesting minor off-center hits, likely due to stage positional error. For illustration, we plotted representative spectra from all bins in Figure c, noting that below ∼30%, spectral features become hard to discern, whereas above this typically preserve recognizable PS signatures. Accordingly, over 99% of the scanned PS particles produced strong Raman signals, underscoring the system’s high alignment accuracy and minimal off-center errors.

Since our system is designed to accommodate a wide range of microparticles, we investigated whether a single, autofocus-determined focal plane could yield robust Raman signals for particles of markedly different sizes. To test this, we selected two types of particles, shown in Figure d: 7 μm PS and 411 μm PET. Using a Sobel-based autofocus routine, the software converged on a plane (z ₁) that maximized brightfield sharpness. We then performed z-scans for both particle types and acquired Raman spectra, measuring peak intensities at 1001.5 cm^–1 for PS and 1614 cm^–1 for PET (Figure e). As expected, the PS particle reached its maximum intensity at z ₁, whereas the larger PET particle required a slightly different optimal z (46 μm offset, inset of Figure e). Nonetheless, the PET particle still exhibited ∼98% of its peak intensity near z ₁, indicating that one focus setting can effectively capture Raman signals from both small and large particles in the same FOV. These findings highlight the flexibility of our high-FOV system, allowing varied particle sizes to be measured without repeated refocusinga crucial advantage for high-throughput applications.

Results and Discussion

To demonstrate the efficacy of our HTS system, we conducted an experiment using a mixed MPs sample containing PE, PET, PVC, and PS particles ranging from 7 to 500 μm. These particles were dispersed on a CaF₂ slide and scanned in a 2 × 3 grid of frames (6.3 × 6.3) mm², which were then stitched together to form a composite image (Figure a). In this scan a total of 245 particles were detected.

6.

(a) Stitched image showing the spatial distribution of particles detected across six scanned frames on a CaF₂ slide. Rectangles highlight regions of interest, with zoomed-in views shown on right. Rectangle 1 contains three particles of different sizesPVC (376 μm), PS (7 μm), and PET (14 μm). Rectangle 2 includes a 391 μm PET particle alongside a smaller 7 μm PS particle. Rectangle 3 displays fragments classified as nonplastic, while Rectangle 4 shows 86 μm PE and 123 μm PVC particles. Rectangle 5 highlights a particle that spans four frames, illustrating consistent detection across borders. Each particle’s assignment to a specific material was determined via Raman spectral analysis and PCA–SVM classification. (b) Reference Raman spectra for the four polymer types (PE, PET, PVC, PS), showing key vibrational peaks. (c) Scree plot demonstrating that four principal components account for ∼93% of the total variance. (d) PCA loadings for these four components, highlighting the Raman shifts most responsible for separating PVC, PET, PE and PS. (e) Confusion matrix confirming 100% classification accuracy on the validation data using the PCA–SVM model. (f) Box plot of classification confidence scores for each polymer (and nonplastic), with a 70% threshold used to label uncertain spectra as nonplastic. (g) Pairwise scatter matrix of the PCA-transformed data, illustrating distinct clusters for each polymer and nonplastic group. (h) Average Raman spectra of the identified classes, confirming consistency with their known vibrational features.

Given the large FOV (approximately 6.3 mm by 6.3 mm), we have indicated selected regions of interest within the stitched image to exemplify the variety of detected particles (Figure a). These highlighted areas, detailed in the figure caption, are meant to illustrate representative examples of different particle types, sizes. The particle detection algorithm identified particles even at the border of frames. When a particle spans across two frames, it might be detected in both, leading to multiple Raman measurements for the same particle (highlighted in box 5 of Figure a). This is acceptable, as it ensures that every particle is detected, rather than missing any by rejecting border particles.

7.

Box plots illustrating four morphological descriptors(a) max Feret diameter, (b) perimeter, (c) Heywood circularityfor PE, PS, PET, and PVC. PS remains the smallest in diameter (∼7 μm beads), with near-circular shapes. PE spans a broader interval (8–159 μm in diameter, perimeters up to ∼430 μm, circularity 0.69–1.01), indicating both rounded and elongated fragments. PS has diameter around 7 μm, with Heywood circularity values near unity (0.9–1.0). PVC spans a max Feret diameter of 19–391 μm and perimeter values from 43 to 1101 μm, paired with Heywood circularity ranging from 0.66 to 0.88. This broad size interval and relatively lower circularity suggest generally larger, less circular fragments compared to the other polymers. PET covers a max Feret diameter of 10–398 μm, with perimeter values reaching 26–1114 μm and Heywood circularity ranging from 0.69 to 0.94. This wide interval indicates mostly moderate-sized fragments, yet includes several large outliers showing more variable shapes.

To reliably identify and classify each particle type, we assembled a reference library of Raman spectra. For this, we acquired 50 Raman spectra per polymer type (PE, PET, PVC, and PS) from particles of varying sizes (Figure b). These reference spectra formed our training data set. With this reference library in hand, we next turned to multivariate analyses to extract and leverage the most informative spectral features for classification. After preprocessing these spectra, PCA was used to reduce the dimensionality of the data. The scree plot (Figure c) confirmed that four principal components captured nearly 93% of the total variance, highlighting the adequacy of this reduced set for downstream analysis. To understand which Raman bands are key to this separation, we examined the PCA loadings in the 400–1800 and 2800–3200 cm^–1 ranges (Figure d).

For PC1, positive loadings at 1063, 1129, and 1440 cm^–1associated with C–C and CH₂ bending (scissoring) vibrationsand at 2846 and 2881 cm^–1corresponding to the CH₂ symmetric and asymmetric stretching bands typical of aliphatic polymer chains in PEcontrast with a strong negative loading at 636 cm^–1 (attributed to C–Cl stretching in PVC) and additional negative loadings at 1614 and 1728 cm^–1 (assigned to the CC stretching of aromatic rings and CO stretching in PET). For PC2, positive loadings at 1001 and 1602 cm^–1, dominated by the spectral features of PS, are opposed by negative loadings at 1129, 1440, 2846, and 2881 cm^–1, which are linked to PE vibrations. In PC3, negative loadings at 636 and 1001 cm^–1 highlight contributions from PVC and PS, respectively, while a positive loading at 1294 cm^–1 is characteristic of PE; further, the positive loadings at 1614 and 1728 cm^–1 confirm the influence of PET. Finally, PC4 exhibits negative loadings at 1063 and 1294 cm^–1 (associated with PE), whereas positive loadings at 1614 and 1728 cm^–1 (indicative of PET) along with additional positive contributions at 2846 and 2881 cm^–1 suggest overlapping spectral features between PE and PET. These PCA loadings demonstrate how the different spectral regions contribute to the separation of the polymers.

After confirming that the four principal components sufficiently captured the key spectral differences among PE, PET, PVC, and PS, we used them as input features for our supervised classification. For this, we split the reference data set using stratified sampling (70% for training, 30% for validation) to maintain balanced classes. This gives 140 training spectra and 60 validation spectra. We trained a support vector machine (SVM) with an RBF kernel on the PCA-transformed training data and then applied the same PCA transformation to the validation set.

This model achieved 100% accuracy on the validation data (shown in confusion matrix Figure e), confirming that the SVM-based model effectively distinguished each polymer. This model achieved 100% accuracy on 60 validation spectra (shown in confusion matrix Figure e), confirming that the SVM-based model effectively distinguished each polymer. We then applied the same PCA transformation to the unknown spectra and subsequently classified them with the SVM. For each spectrum, the classifier calculated a predicted probability for each polymer type, and if the highest probability fell below 70%, we labeled the result as nonplastic. Using this criterion, the SVM confidently assigned four clusters to the known reference materials, comprising 79 PE, 18 PS, 8 PET, and 19 PVC particles in total. Their prediction probabilities are summarized in the box plot (Figure f). Next, we projected both data sets into the same PCA space, creating a pairwise scatter plot (Figure g). The average Raman spectra for each cluster are shown in Figure h, with shaded areas representing standard deviations to illustrate intracluster variability.

Having classified each spectrum, we next examined the morphological characteristics of the identified particles. Specifically, we measured max Feret diameter, perimeter, area, and Heywood circularity to gain insight into their size and shape variability. The box plots in Figure illustrate these descriptors across PE, PS, PET, and PVC, showing how polymer type correlates with distinct physical dimensions. The smallest particle detected was a PS bead of roughly 7 μm across, displaying a near-spherical shape (circularity ∼1.0) consistent with commercial 7 μm beads. PVC ranged from 19 μm to about 392 μm in diameter, with perimeters up to 1101 μm, reflecting larger, more elongated fragments (circularities ∼0.66–0.88). PET covered a similarly broad span (10–398 μm in diameter, perimeter to ∼1114 μm), with circularities dropping to 0.69, indicating generally fewer circular shapes. Meanwhile, PE occupied a somewhat narrower interval (8–160 μm in diameter, perimeter ∼21–430 μm, circularity 0.69–1.01), featuring both near-circular and elongated particles. This range in sizes and shapesfrom tiny 7 μm PS beads to ∼400 μm fragmentshighlights the system’s capacity to rapidly characterize diverse MPs by both chemical and morphological criteria.

Our classification provides convincing results for these particles. Real-world MPs often become biofouled or aged, leading to strong fluorescence backgrounds and noticeable attenuation or distortion of the Raman bands. − Future work will systematically incorporate aged and fouled MPs into the training set, validating our HTS-RS platform’s performance in real-world monitoring scenarios. ,

To benchmark HTS-RS against established workflows, Table lists the key characteristics of current microplastic-detection techniques. Pyrolysis–GC–MS delivers rapid bulk-polymer profiles (≈2 samples h^–1) but destroys the sample, so particle counts and sizes are lost. Automated Raman mappingexemplified by the GEPARD workflowprocesses one particle in ≈5 s. Particle-based μFT-IR point mapping is far slower, taking ≈7.5 min per particle because the stage must halt and refocus for every spectrum. HTS-RS, in contrast, scans a 3.1 × 2.1 mm² FOV, detects particles down to ∼7 μm, and completes the measurement for each particle in ≈1.4 s. Thus, HTS-RS is 4× faster than automated Raman and >300× faster than particle-based μFT-IR, while remaining completely nondestructive.

2. Comparison of Microplastic-Analysis Techniques by Field of View, Spatial Resolution, Throughput, and Destructiveness.

S. No	technique	FOV (mm2)	size range (μm)	time	destructive
1	HTS-Raman (this work)	3.1 × 2.1	7–400	1.5 s/particle	No
2	automated Raman microspectroscopy	0.5 × 0.3	1–500	5 s/particle	No
3	particle based μ-FTIR imaging ,	0.6 × 0.6	25–300	7.5 min/particle	No
4	Py-GC-MS			2 samples per hour	Yes

Conclusion

In this work, we developed and validated a high-throughput Raman spectroscopy (HTS-RS) platform specifically designed for the efficient detection and chemical characterization of microplastics across a wide range of particle sizes and shapes. By leveraging optical design principles based on the space-bandwidth product, we achieved a field of view of 3.1 × 2.1 mm² while maintaining a spatial resolution of 1.4 μm, which enables the system to rapidly localize and analyze microplastic particles as small as 7 μm. The system integrates fully automated particle detection, autofocus correction, stage control, and Raman spectral acquisition into a seamless workflow. During our validation measurements, we successfully scanned over 1300 particles in less than 35 min, corresponding to an average measurement time of ∼1.4 s per particle. The integrated software, developed in LabVIEW, automatically performs tilt correction, particle localization, and morphological analysis (including size, perimeter, and circularity), thereby significantly reducing operator intervention. Chemometric analysis using PCA and SVM classifiers demonstrated 100% classification accuracy on the reference data set for four common polymer types (PE, PET, PVC, PS). Importantly, the system reliably handled a broad particle size spectrum, from small spherical beads (7 μm PS particles) to large irregular fragments (>400 μm), without the need for repeated refocusing. These results underscore the platform’s potential for scalable, high-throughput microplastic analysis in environmental monitoring applications. Looking ahead, integrating real-time spectral evaluation directly into the acquisition software will further enhance workflow efficiency, enabling near-instant classification. Additionally, future work will address performance on environmental samples containing weathered and biofouled particles to demonstrate robustness under real-world conditions. Overall, the presented HTS-RS system represents a significant advancement over conventional Raman microspectroscopy, offering a powerful combination of wide-area screening, fine spatial resolution, fast acquisition speed, and automated spectral analysis. This makes it a promising tool for routine environmental monitoring, regulatory compliance testing, and large-scale microplastic pollution assessment.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.5c04522.

• Detailed SBP derivation, objective-lens comparison, and camera-sensor evaluation (PDF)

The authors declare no competing financial interest.