1 of 1

INTRODUCTION:

IMPROVING DEGRADED MICROPLASTIC IDENTIFICATION USING FTIR-BASED CNN MODELING

Helen Ly a , Maria Geidel b , Liyuan Gong c , Md Iftakhar Khan d , Erfan Eskandari d , Yang Lin d

a Computer Engineering, University of Rhode Island, Kingston, RI c Chemistry, Humboldt-Universität zu Berlin, Berlin, Germany

b Chemistry, Humboldt-Universität zu Berlin, Berlin, Germany d Mechanical Engineering, University of Rhode Island, Kingston, RI

Polystyrene (PS)

Polypropylene (PP)

Samples are degraded in a UV weathering chamber, at an intensity of 340 UVA. Samples are taken out every 4 days

Samples are prepared into 2-5mm sample sizes and then grinded down in a Planetary Ball Mill PM 100 1 .

Each sample is extracted and diluted in

two beakers with 300mL of deionized water.

Control

Test

Fisher Scientific Accuspin 8C Clinical Centrifuge

Data

Collection

Sample

Preparation

UV-light

Degradation

QUV Accelerated Weather Tester

RESULTS:

METHODOLOGY:

  1. Retsch Planetary Ball Mill PM 100. Profilab24. https://profilab24.com/en/laboratory/mills/retsch-planetary-ball-mill-pm-100

​

  • De Frond, H., Rubinovitz, R., & Rochman, C. M. (2021). μATR-FTIR Spectral Libraries of Plastic Particles (FLOPP and FLOPP-e) for the Analysis of Microplastics. Analytical Chemistry, 93(48), 15878–15885. https://doi.org/10.1021/acs.analchem.1c02549

​

  • Gong, L., Martinez, O., Mesquita, P., Kurtz, K., Xu, Y., & Lin, Y. (2023). A microfluidic approach for label-free identification of small-sized microplastics in seawater. Scientific Reports, 13(1), 11011. https://doi.org/10.1038/s41598-023-37900-9

​

  • Campanale, C., Savino, I., Massarelli, C., & Uricchio, V. F. (2023). Fourier Transform Infrared Spectroscopy to Assess the Degree of Alteration of Artificially Aged and Environmentally Weathered Microplastics. Polymers, 15(4), Article 4. https://doi.org/10.3390/polym15040911

This data is used to train a 1-D CNN model. The CNN structure used is shown above 3.

A dataset focused on these eight plastics and polyurethane is created with FTIR spectra data from the FLOPP and FLOPP-e library 2 and sample scans.

​

    • # of spectrums used: 363
    • Spectra range used:

675-4000 cm-1

​

RESULTS:

FUTURE WORK:

  • Conduct data augmentation to further improve accuracy
  • Collect more data on different plastics (polyurethane)
  • Further degrade existing samples and different plastics
  • Use a saline solution (3.5% sodium chloride in a water solvent) and study the UV degradation of plastic samples in saltwater environments (left).

and undergo centrifugation at

3000-3500 rpm. 3-10µL of concentrated

sample is dried

and scanned

using FTIR

Spectroscopy.

Model Training

Identifying plastics in the environment is difficult as their chemical compositions change due to ultra-violet (UV) exposure. This work captures the degradation of plastics in a controlled environment and trains a Convolutional Neural Network (CNN) model to better identify these plastics.

  • Fourier Transform Infrared (FTIR) Spectroscopy: Uses infrared light which is absorbed at different frequencies in the atomic bonds of a sample. This absorption is observed in spectra data and helps to identify different bonds.
    • Is more sensitive to the oxidation process polymers undergo during the degradation process.

Stage of the FTIR

Additional samples are scanned including:

  • polystyrene (PS)
  • polypropylene (PP)
  • polycarbonate (PC)
  • polyvinyl chloride (PVC)
  • polyethylene terephthalate (PET)
  • polyethylene (PE)
  • nylon (PA)
  • polyester

Below showcases the split in validation (20%) and training sets.

The peak at ~1725 cm-1 indicates a carbonyl bond is present, a sign of oxidization which is not present in the control group. This agrees with other literature where carbon-oxygen bonds occur during UV exposure 4 and validates the methodology used.

FTIR spectra of polystyrene UV degradation of both groups during a 12-day period.

This work is supported by the U.S. National Science Foundation under Award #2348968, REU Site: URI Plastic Initiative at the University of Rhode Island, and the Donald R. Wilson Jr. Family Foundation. This research is supported by the National Science Foundation (NSF, Grant 2347408). Partial support for professional development activities was provided by EPSCoR Cooperative Agreement #OIA-2433276 and the RI Commerce Corporation through the Science and Technology Advisory Committee. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the funding partners.

ACKNOWLEDGEMENTS:

METHODOLOGY:

Training Accuracy: 94.6%

Validation Accuracy: 83.6%

​

Training: How well the model performs on training data.

Validation: How well the model performs on new data.

Epochs: How many times the model looks at the data.

Below are the training and validation accuracy curves graphed.

REFERENCES: