Smarter Spectra: Texas A&M Researchers Transform Microplastics Identification

Similarity learning enables fast, flexible detection of microplastics in real-world conditions.

October 14, 2025 By Kayley Hart

Similarity learning model generation of embeddings from µFTIR spectra. On the left: architecture of underlying CNN model, consisting of an input layer of the normalized µFTIR absorbance spectra, followed by convolutional layers and then a set of dense layers producing the final vector embedding. On the right: visualization of the embeddings using dimensionality reduction by PaCMAP, showing separated clusters for each polymer composition. — Researchers from Texas A&M University developed a machine learning model that dramatically improves the accuracy and adaptability of identifying microplastics using µFTIR spectroscopy. | Image: Laboratory for Synthetic-Biologic Interactions – Texas A&M

Microplastics — fragments of plastic less than five millimeters in size — have become a pervasive environmental concern, turning up in oceans, rivers, soil and even the human body. As scientists race to understand the scope and impact of microplastic pollution, one of the biggest challenges remains: how to reliably identify these particles in complex, real-world samples.

A team of researchers from Texas A&M University developed a powerful new tool to meet that challenge. The report was published on October 14, 2025 in Proceedings of the National Academy of Sciences of the United States of America. From research conducted in the laboratory of Dr. Karen Wooley, Distinguished Professor of Chemistry and Director of the Laboratory for Synthetic-Biologic Interactions (LSBI), with first author Justin Smolen (Assistant Director, LSBI), Gavin Moore (Department of Chemistry) and Dr. Nicholas Perez (Associate Professor, Department of Geology & Geophysics), the team introduced a machine learning method that dramatically improves the accuracy and adaptability of microplastics identification using micro-Fourier transform infrared (µFTIR) spectroscopy.

Traditional methods for identifying microplastics rely on matching spectral data to databases of known plastics or having experts directly analyze the data, which can both be time-consuming and error-prone, especially when samples are contaminated with organic material or sediment. Even existing machine learning models often struggle when faced with noisy data or unfamiliar plastic types. To overcome these limitations, the research team turned to similarity learning, a form of deep learning that trains models to recognize patterns by comparing how similar or different data points are. Instead of forcing the model to choose from a fixed list of plastic types, similarity learning creates a flexible “embedding space” where spectra from the same polymer cluster together.

This approach allows the model to not only classify known microplastics with high accuracy but also to flag unknown or novel materials — an essential capability for environmental sampling where unexpected contaminants are common. One of the most promising aspects of the model is its extensibility: new plastic types can be added to the system by embedding just a few reference spectra with no need to retrain the entire network. This makes the tool especially valuable for long-term monitoring where the diversity of plastic materials is constantly evolving.

As microplastics research expands, similarity learning could play a vital role in automating and scaling analysis. By integrating the interpretability of conventional spectral matching with the pattern-recognition strength of neural networks, the Texas A&M team has established a foundation for next-generation microplastics detection that operates more rapidly, learns from complex data and adapts to the challenges of real-world samples.

This research was supported by the National Science Foundation (DMR-1905818 and SBE-2343148) and the Robert A. Welch Foundation through the W. T. Doherty-Welch Chair in Chemistry.