System and Method for Continuous Microplastic Detection and Polymer-Type Classification in Municipal Water Distribution Systems Using Inline Multi-Angle Light Scattering and On-Device Neural Network Particle Identification

Abstract

Disclosed is a system and method for continuous, real-time detection, counting, and polymer-type classification of microplastic particles in municipal drinking water distribution systems. The system employs an inline flow-through optical cell containing a 635 nm laser diode and an array of eight silicon photodetectors positioned at fixed scattering angles (10°, 30°, 45°, 60°, 90°, 120°, 150°, and 170°). Each particle transiting the cell produces a multi-angle scattering intensity profile governed by Mie theory, where the angular distribution encodes information about particle diameter (1–500 μm), refractive index (which differs by polymer type: polyethylene 1.50, polypropylene 1.49, polystyrene 1.59, PET 1.58, nylon 1.53), and morphology (sphere, fiber, fragment, film). An on-device one-dimensional convolutional neural network (1D-CNN), quantized to INT8 and running on an ARM Cortex-M7 microcontroller, classifies each particle event in under 5 ms. The system reports particle counts by polymer type and size bin at 1-minute intervals over LoRaWAN or cellular backhaul to a central monitoring dashboard. Distributed deployment across treatment plant outlets, distribution mains, and point-of-use taps enables source attribution, treatment efficacy verification, and regulatory compliance monitoring at a per-sensor bill-of-materials cost below $120.

Field of the Invention

This invention relates to environmental monitoring of municipal water infrastructure, specifically to automated inline detection and classification of microplastic contaminants using optical scattering sensors and edge-deployed machine learning inference.

Background

Microplastic contamination of drinking water is now documented worldwide. A 2024 study by Qian et al. in Nature Nanotechnology found approximately 240,000 detectable plastic particles per liter in bottled water using stimulated Raman scattering microscopy, with 90% in the nanoplastic range (< 1 μm). Tap water studies report lower but still significant concentrations: Kosuth et al. (2018) found 5.45 particles/L in treated tap water across 159 samples from 14 countries, while Mintenig et al. (2019) reported 0.7 particles/L in German drinking water treatment plants. The WHO (2019) concluded that microplastics in drinking water do not appear to pose a health risk at current levels but acknowledged substantial uncertainty due to limited data and inconsistent measurement methods.

Current microplastic analysis methods are laboratory-bound, expensive, and slow:

• Fourier-Transform Infrared Spectroscopy (FTIR): The reference method for polymer identification. Requires sample filtration, visual sorting under a stereomicroscope, and per-particle spectral acquisition. Cost: $300–800 per sample. Turnaround: 2–6 weeks. Löder et al. (2017) demonstrated automated focal-plane array FTIR, but instrumentation costs exceed $150,000.
• Raman Microspectroscopy: Higher spatial resolution than FTIR (down to ~1 μm). Araujo et al. (2018) benchmarked Raman against FTIR and found 87% concordance for polymer identification. Equipment cost: $100,000–300,000. Measurement time: 10–30 minutes per particle.
• Pyrolysis-GC/MS: Destructive technique that identifies polymer mass rather than particle count. Dekiff et al. (2023) applied pyrolysis-GC/MS to drinking water treatment chains. Equipment cost: $80,000–120,000. Cannot provide per-particle counts or morphology data.
• Flow Cytometry: Sgier et al. (2025) demonstrated microplastic and nanoplastic detection using flow cytometry with fluorescent Nile Red dye staining, achieving particle counting but limited polymer-type discrimination.

The gap in the art is a low-cost, inline, continuous monitoring system that: (a) detects microplastic particles in flowing water without sample collection, (b) classifies polymer type without spectroscopic instrumentation, (c) operates at the edge without cloud connectivity for primary classification, and (d) enables distributed network deployment across water distribution systems at price points accessible to small utilities.

Detailed Description

1. Inline Optical Flow Cell

The sensor core is a cylindrical flow-through cell machined from anodized aluminum with quartz optical windows. Internal dimensions: 4 mm bore diameter, 20 mm optical path length. The cell accepts standard 1/4" NPT fittings for inline installation on water distribution pipes via a sampling tee. Flow rate through the cell: 50–200 mL/min, regulated by an upstream needle valve. At 100 mL/min, the transit time through the optical interrogation zone (~2 mm length) is approximately 600 μs for particles in the flow center, providing sufficient dwell time for photodetector sampling at 100 kHz.

A 635 nm laser diode (5 mW, Thorlabs CPS635R or equivalent, unit cost ~$25) provides the illumination beam, focused to a 100 μm waist at the cell center via an aspheric collimating lens. The beam geometry defines the optical interrogation volume: approximately 0.03 μL, ensuring that at expected microplastic concentrations (1–100 particles/L), the probability of two particles occupying the interrogation volume simultaneously is negligible (< 10⁻⁶).

2. Multi-Angle Scattering Detection

Eight silicon photodiodes (Hamamatsu S1227-1010BQ or equivalent, active area 10 mm², unit cost ~$3 each) are mounted on a precision-machined ring at fixed angular positions relative to the incident beam: 10° (near-forward), 30°, 45°, 60°, 90° (side-scatter), 120°, 150°, and 170° (near-backward). Each photodiode is fitted with a 635 ± 5 nm bandpass filter to reject ambient and fluorescence signals. Transimpedance amplifier circuits (gain: 10⁷ V/A) convert photocurrent to voltage signals digitized by the microcontroller's 12-bit ADC at 100 kHz per channel.

When a particle transits the laser beam, each photodiode records a pulse whose amplitude depends on the differential scattering cross-section at that angle. According to Mie theory, the angular scattering pattern from a spherical particle is determined by the size parameter x = πd/λ (where d is diameter and λ is wavelength) and the complex refractive index ratio m = n_particle/n_medium. For microplastics in water (n_medium = 1.33), the refractive index ratios differ meaningfully by polymer: PE gives m = 1.128, PP gives m = 1.120, PS gives m = 1.195, PET gives m = 1.188, and nylon gives m = 1.150. These differences produce measurably distinct angular scattering signatures, particularly in the 60°–150° range where the sensitivity to refractive index dominates over size effects.

3. Particle Event Detection and Feature Extraction

A hardware-level trigger fires when any photodiode channel exceeds a configurable threshold (default: 3× the running baseline noise level, computed over a 10-second sliding window). Upon trigger, the system captures a 2 ms window of data across all eight channels (200 samples per channel at 100 kHz). From this raw capture, the following features are extracted per event:

• Peak amplitude vector: 8 values representing the maximum scattering intensity at each angle, normalized to the 90° channel to remove laser power fluctuations.
• Pulse width vector: 8 values representing the full-width-at-half-maximum (FWHM) of the scattering pulse at each angle. Pulse width encodes particle size: a 10 μm particle at 100 mL/min flow produces a ~60 μs pulse; a 500 μm particle produces a ~3 ms pulse.
• Asymmetry ratio: The ratio of forward-hemisphere scattering (sum of 10°, 30°, 45°, 60° channels) to backward-hemisphere scattering (sum of 120°, 150°, 170° channels). Larger particles and higher-refractive-index particles scatter more strongly forward.
• Depolarization index: If an optional cross-polarized photodiode is installed at 90°, the ratio of cross-polarized to co-polarized scattering at 90° encodes particle asphericity. Fibers produce depolarization ratios of 0.3–0.6; spherical beads produce < 0.05.

4. On-Device Neural Network Classification

A 1D-CNN classifier processes the concatenated feature vector (8 peak amplitudes + 8 pulse widths + 1 asymmetry ratio + 1 depolarization index = 18 features). Architecture: input layer (18), two 1D convolutional layers (32 and 64 filters, kernel size 3, ReLU activation), global average pooling, a 64-unit dense layer with dropout (0.3), and a dual-head softmax output. The first head classifies polymer type across seven classes: polyethylene (PE), polypropylene (PP), polystyrene (PS), polyethylene terephthalate (PET), polyamide (nylon), mineral/inorganic, and organic/biofilm. The second head classifies size bin across five classes: 1–10 μm, 10–50 μm, 50–100 μm, 100–500 μm, and > 500 μm.

The model is trained on synthetic Mie scattering calculations (using the PyMieScatt library) covering 500,000 parameter combinations spanning the full range of sizes, refractive indices, and morphologies, supplemented by 20,000 laboratory-validated measurements of reference microplastic standards (Cospheric CPMS microspheres, PolyAn GmbH fiber standards). Data augmentation includes Gaussian noise injection (σ = 5% of signal), flow-rate variation (±30%), and laser power drift (±10%). The model is quantized to INT8 via TensorFlow Lite for Microcontrollers and runs on an ARM Cortex-M7 (e.g., STM32H743, 480 MHz, unit cost ~$12). Inference time: < 5 ms per particle event. Model size: 42 KB.

Per-class confidence thresholds are configurable (defaults: PE 0.65, PP 0.70, PS 0.60, PET 0.60, nylon 0.75, mineral 0.50, organic 0.50). Events below threshold are logged as "unclassified" with the raw feature vector retained for later cloud-based reanalysis.

5. Self-Calibration Protocol

Each sensor contains a small reservoir (5 mL) of NIST-traceable polystyrene latex microsphere suspension (Duke Standards 4000 Series, nominal diameters: 10 μm, 50 μm, 100 μm, concentration 1,000 particles/mL). A solenoid-actuated injection port introduces calibration particles into the flow stream on a configurable schedule (default: once per 24 hours) or on manual trigger. The known size and refractive index of polystyrene (n = 1.59 at 635 nm) provides a reference scattering pattern. The system computes the deviation between observed and expected scattering profiles, generating correction factors for each photodiode channel to compensate for window fouling, laser degradation, and photodetector drift. If any correction factor exceeds ±20%, the sensor generates a maintenance alert.

6. Network Architecture and Data Aggregation

Each sensor transmits a compressed data packet at configurable intervals (default: 1 minute) containing: polymer-type count vector (7 classes × 16-bit count = 14 bytes), size-bin count vector (5 bins × 16-bit count = 10 bytes), mean confidence score per class (7 bytes), sensor health metrics (flow rate, laser power, calibration status: 6 bytes), and up to 10 high-interest raw feature vectors for unclassified events (180 bytes). Total packet size: < 220 bytes per transmission.

Communication options include LoRaWAN (for utilities with existing LPWAN infrastructure), NB-IoT cellular (for standalone deployment), or RS-485 Modbus (for integration with existing SCADA systems). A central aggregation server computes distribution-system-wide microplastic maps, identifies concentration anomalies, performs source attribution by correlating upstream and downstream sensor readings, and generates regulatory compliance reports against emerging standards such as California SB 1422 (the first U.S. state mandate for microplastics monitoring in drinking water).

7. Applications

• Treatment efficacy monitoring: Compare microplastic counts upstream and downstream of treatment processes (coagulation, sand filtration, membrane filtration, activated carbon) to quantify removal efficiency by polymer type and size. Identify treatment breakthrough events in real time rather than through periodic grab samples.
• Distribution system integrity: Track microplastic counts along distribution mains to identify pipe segments contributing plastic particles (PVC and HDPE pipe degradation, gasket deterioration, biofilm entrainment). A sustained increase at a downstream sensor without a corresponding upstream increase localizes the source.
• Point-of-use verification: Deploy under-sink sensors to quantify the effectiveness of consumer filtration systems (activated carbon, reverse osmosis, ultrafiltration) against microplastics, providing data for consumer product certification.
• Regulatory compliance: Automate the monitoring and reporting requirements of microplastics regulations. California's SB 1422 mandated a standardized testing methodology by 2024; continuous inline monitoring eliminates the cost and delay of laboratory grab sampling.

8. Figures Description

• Figure 1: Cross-sectional diagram of the inline flow-through optical cell showing laser diode, collimating lens, quartz windows, eight photodetector positions at specified angles, and NPT pipe fittings.
• Figure 2: Simulated Mie scattering angular profiles (scattering intensity vs. angle, 10°–170°) for 50 μm spherical particles of PE, PP, PS, PET, and nylon in water at 635 nm, demonstrating the polymer-discriminating angular features exploited by the classifier.
• Figure 3: System architecture showing distributed sensor deployment across a municipal water distribution network, with LoRaWAN/cellular backhaul, central aggregation server, and compliance reporting dashboard.
• Figure 4: Confusion matrix from laboratory validation of the 1D-CNN classifier against 2,000 reference microplastic particles (Cospheric standards), showing per-class precision and recall.

Claims

1. A system for continuous detection and classification of microplastic particles in municipal water, comprising: an inline flow-through optical cell with a laser diode and a plurality of photodetectors arranged at fixed scattering angles; wherein each particle transiting the cell produces a multi-angle scattering intensity profile; and an on-device neural network classifier that determines particle polymer type and size from the scattering profile without spectroscopic analysis.
2. The system of claim 1, wherein the photodetector array comprises at least six photodetectors positioned at angles spanning forward-scatter (< 45°), side-scatter (45°–135°), and back-scatter (> 135°) to capture the full angular scattering distribution.
3. The system of claim 1, wherein polymer-type classification exploits differences in refractive index among common microplastic polymers (PE, PP, PS, PET, polyamide) as encoded in the Mie scattering angular distribution at the operating wavelength.
4. The system of claim 1, wherein the neural network classifier is a one-dimensional convolutional neural network quantized to INT8 and executed on a microcontroller with ARM Cortex-M class processing capability, achieving per-event inference in under 10 milliseconds.
5. The system of claim 1, further comprising a self-calibration module that periodically injects reference microspheres of known size and refractive index into the flow stream and computes per-channel correction factors to compensate for optical window fouling, laser degradation, and photodetector drift.
6. A method for monitoring microplastic contamination in a water distribution system, comprising: installing a plurality of inline optical scattering sensors at distributed points in the distribution network; continuously classifying particles by polymer type and size at each sensor using on-device inference; aggregating classified particle counts at a central server; and performing source attribution by correlating concentration changes across upstream and downstream sensor positions.
7. The method of claim 6, further comprising computing treatment process efficacy by comparing polymer-specific particle counts upstream and downstream of water treatment unit processes, enabling real-time detection of treatment breakthrough events.
8. The method of claim 6, further comprising localizing pipe-segment microplastic sources within the distribution network by identifying sensor positions where downstream counts exceed upstream counts without a corresponding increase at adjacent upstream sensors.
9. The system of claim 1, further comprising a particle morphology estimation module that derives shape classification (sphere, fiber, fragment, film) from the pulse-width profile and depolarization ratio of the scattering event across the photodetector array.
10. The system of claim 1, wherein each sensor unit has a bill-of-materials cost below $150 and communicates via LoRaWAN, NB-IoT, or RS-485 Modbus for integration with municipal SCADA infrastructure.

Prior Art References

1. Qian et al., PNAS 2024 — ~240,000 plastic particles/L in bottled water via stimulated Raman scattering
2. Kosuth et al., PLOS ONE 2018 — 5.45 particles/L in treated tap water across 14 countries
3. Mintenig et al., Water Research 2019 — 0.7 particles/L in German drinking water treatment plants
4. WHO, 2019 — Microplastics in drinking water risk assessment
5. Löder et al., Analytical and Bioanalytical Chemistry 2017 — Automated focal-plane array FTIR for microplastic identification
6. Araujo et al., Science of the Total Environment 2018 — FTIR vs. Raman microplastic identification comparison
7. Sgier et al., 2025 — Flow cytometry with Nile Red for microplastic/nanoplastic detection
8. Dekiff et al., 2023 — Pyrolysis-GC/MS for drinking water microplastics from source to tap
9. PyMieScatt — Open-source Python Mie scattering calculation library
10. California SB 1422 — First U.S. state mandate for microplastics in drinking water monitoring
11. TensorFlow Lite for Microcontrollers — On-device ML runtime for embedded classification
12. STM32H743 — ARM Cortex-M7 microcontroller (480 MHz, DSP, FPU)
13. Thorlabs CPS635R — 635 nm collimated laser diode module
14. Hamamatsu S1227-1010BQ — Silicon photodiode for scattering detection