Deep Learning Meets Spectroscopy to Transform Plastic Recycling Accuracy

A recent study published in the journal Resources, Conservation and Recycling discusses how the convergence of spectroscopy and advanced machine learning (ML) could dramatically improve the efficiency and accuracy of plastic recycling systems (1). This study, which was led by researchers from Washington State University Tri-Cities, advances sustainable materials management and reducing the environmental footprint of plastics.

Global plastic production has been increasing each year. In 2020, global plastic production was approximately 435 metric tons (2). Based on current trends, it is expected that global plastic production will hit 736 metric tons by 2040 (2). As a result, the global plastics market is witnessing steady growth of approximately mid-single-digit growth, with an expected CAGR of 5.1% between 2025 and 2032 (3).

Currently, the vast majority of consumer plastics, which include polyethylene terephthalate (PET), high-density polyethylene (HDPE), polyvinyl chloride (PVC), low-density polyethylene (LDPE), polypropylene (PP), and polystyrene (PS), ultimately enter municipal waste streams and landfills, where they pose long-term environmental challenges.

In this study, the researchers looked to solve the problem of downcycling, which is the process of reusing materials to create something of lesser value (4). Although downcycling is better than treating plastics as pure waste, it does not lead to true material recovery. Current sorting methods often struggle with contaminated, dyed, or weathered plastics, and the researchers developed a new method that could resolve these issues.

Their method involved developing a ML framework rooted in convolutional neural networks (CNNs), leveraging their strength in analyzing high-dimensional data such as spectroscopic signals. The bulk of the study focuses on vibrational spectroscopy, particularly Raman scattering and attenuated total reflectance–Fourier transform infrared (ATR-FTIR) spectroscopy (1). Raman spectra collected from real recycling samples were used to train a six-class CNN capable of distinguishing among the most common consumer plastics.

As a result, the CNN achieved 100% classification accuracy using Raman data, demonstrating that deep learning can extract chemically meaningful features from complex spectra without the need for manual preprocessing or feature selection (1).

To enhance real-world applicability, the researchers expanded the approach to ATR-FTIR spectroscopy, which is widely used in field and industrial settings. Using a similar CNN architecture, the ATR-FTIR-based model achieved 95% accuracy, confirming that the approach is robust across complementary spectroscopic techniques (1). In addition, the models maintained low computational cost, a critical factor for deployment in automated, high-throughput recycling environments.

The study also investigated how four analytical techniques, (ATR-FTIR, near-infrared reflectance [NIR], laser-induced breakdown spectroscopy [LIBS], and X-ray fluorescence [XRF]) can identify and classify plastics. The results indicated that ATR-FTIR, NIR, and LIBS were effective for consumer plastics, with success rates of 99%, 91%, and 97%, respectively (1). When applied to marine plastic debris, which is often more degraded and contaminated, ATR-FTIR again led the field with a 99% success rate, while NIR, LIBS, and XRF achieved 81%, 76%, and 66%, respectively (1).

The main takeaway from this study is that the researchers used CNN for plastic identification. Although CNNs are well established in fields such as image recognition, their application to spectroscopic data sets for plastic classification remains relatively novel (1). The CNN was able to autonomously learn relevant spectral features, which eliminates the need for expert-driven spectral interpretation and manual feature engineering. This not only improves efficiency, but that also enhances scalability across different instruments and operating conditions.

Equally important is the model’s resilience to real-world variability. The researchers found that CNN successfully handled plastics with dyes, additives, and signs of environmental degradation, which are factors that routinely confound conventional sorting technologies (1). According to the researchers, this robustness represents a significant step toward field-ready systems capable of operating reliably in recycling facilities, material recovery facilities, and potentially even in marine debris monitoring efforts (1).

“The model effectively handles dyed plastics, plastics containing additives, and plastics subjected to real-world environmental conditions and deterioration,” the authors wrote in their study (1), “representing a significant practical advancement for real-world applications."

References

1. Garcia Tovar, M. P.; Villarreal Blanco, M. A.; Primera-Pedrozo, O. M.; et al. Identification of Common Types of Plastics by Vibrational Spectroscopic Techniques. Res. Con. Rec. 2026, 227, 108767. DOI: 10.1016/j.resconrec.2025.108767
2. Dabo, M. Global Plastic Demand Shows No Signs of Slowing. Packaging Gateway. Available at: https://www.packaging-gateway.com/features/global-plastic-demand-shows-no-signs-of-slowing/ (accessed 2026-01-05).
3. rePurpose Team, Plastics Market 2025 Outlook: What Will Regulation, Demand Shifts, and a UN Treaty Mean? rePurpose. Available at: https://www.repurpose.global/blog/plastics-market-2025-outlook-what-will-regulation-demand-shifts-and-a-un-treaty-mean (accessed 2026-01-05).
4. Cary Company, What Is Downcycling? Cary Company. Available at: https://www.thecarycompany.com/insights/articles/what-is-down-cycling?srsltid=AfmBOoqpImDvlsFCrLfrYHkynyV8AvKXXEumnhpwoQ-pkCcOuDsGss5p (accessed 2026-01-05).