Navid Zarrabi, Eric M. Strohm, Hadi Rezvani, Matthew Lisondra, Nariman Yousefi, Sajad Saeedi,
Michael C. Kolios
Published in Nature, npj Emerging Contaminants
Standard microplastic identification methods, such as Raman and FTIR spectroscopy, pyrolysis-GC, and electron microscopy, are slow, manual, and hard to scale. This study shows that high-frequency ultrasound offers a fast, non-destructive, label-free alternative. Echoes from individual particles carry material- and size-dependent signatures, and deep learning can read those signatures to identify what a particle is made of and how big it is.
Particles are immobilized within an agar gel in a Petri dish. The transducer performs scanning in X–Y plane to acquire backscattered signals. These measurements are organized into a 3D tensor, with X–Y representing spatial dimensions and Z corresponding to the temporal axis.
High-frequency mechanical waves in the MHz range are transmitted into the medium, where the incident wave interacts with particles and generates scattered waves. A portion of this scattered energy returns to the transducer as backscatter. The characteristics of the backscattered signal are governed by factors such as the acoustic impedance mismatch between the particle and surrounding medium, the particle size relative to the wavelength, and the excitation frequency. These signal patterns are subsequently analyzed using machine learning models to extract discriminative features for particle identification and characterization.
Temporal and spectral features are extracted from each particle’s echo. Spectral spread, spectral entropy, and spectral centroid carry the most information, and each material shows its own feature pattern. A subset of 13 features gives the best accuracy for classical models.
Several machine learning models were compared with a 1D CNN that learns directly from each echo’s frequency spectrum. The CNN performs best at both the signal and particle level, reaching 97.14% ± 3.11% particle-level accuracy over 10 random splits. Combining the predictions from all signals belonging to the same particle makes the result more robust to noisy echoes. PE and PMMA are identified almost perfectly, and most remaining errors are between glass and steel, whose spectra overlap.
Echo spectra change predictably with particle diameter. A separate neural network for each material sorts particles into four size ranges: 20–28 µm, 38–57 µm, 67–90 µm, and 300–330 µm. Accuracy is 100% for PE, PMMA, and steel and 99.01% for glass. Each particle’s signals are kept within a single data split, so the models are always tested on particles they haven’t seen.
Detection becomes harder in crowded regions, and a strongly reflecting particle can hide a weaker neighbour. Very small particles produce weak echoes, so at 40 MHz the practical lower size limit comes from signal strength, not resolution. Higher-frequency transducers could push this limit down. The spheres used here are also more uniform than real environmental microplastics. A natural next step is acoustic flow cytometry, which analyzes particles as they flow past the transducer instead of scanning a fixed sample.
Cite
@article{zarrabi2026high,
title={High-frequency ultrasound combined with deep learning enables identification and size estimation of microplastics},
author={Zarrabi, Navid and Strohm, Eric M and Rezvani, Hadi and Lisondra, Matthew and Yousefi, Nariman and Saeedi, Sajad and Kolios, Michael C},
journal={npj Emerging Contaminants},
volume={2},
number={1},
pages={9},
year={2026},
publisher={Nature Publishing Group UK, London}
}