Scientists from the University of Toronto and University College London have proposed a new method for detecting and identifying microplastics via high-frequency ultrasound and deep learning methods. They demonstrated that ultrasound can be used not only for visualization, but also for precise identification of the composition and size of microparticles. The new method opens up possibilities for the creation of compact systems for rapid monitoring of water pollution in near real time.
Microplastics pose an acute problem indeed: today, particles smaller than 5 millimeters in size are found everywhere, from Arctic ice to human tissue. However, the existing analysis methods, such as Raman spectroscopy, infrared spectroscopy and electron microscopy, require complex sample preparation, significant time commitments and laboratory infrastructure. Although these methods deliver high accuracy, they are poorly suited for large-scale and especially field monitoring. A need arose for a method that would be fast, non-destructive and potentially applicable in the natural environment.
The researchers from Canada and the UK turned to high-frequency ultrasound. While frequencies of 2–15 megahertz are typically used in medical diagnostics, their study employed a transducer with a frequency of about 40 megahertz. At these parameters, the sound wavelength becomes comparable to particle sizes of tens of micrometers. This mode is prone to pronounced resonant scattering effects: the reflected signal now depends not only on the size but also on the elasticity and density of the material. In simple terms, materials such as polyethylene, acrylic, glass and steel form different acoustic signatures.
During the experiment, microspheres ranging in diameter from 20 to 330 micrometers were placed in an agarose gel with water-like acoustic properties and scanned in 10-micrometer increments. At each point, the temporal reflection signal was recorded. A three-dimensional data array was generated as a result. The researchers’ key goal was to separate signals related to specific particles from the background response of the environment. To that end, they developed an algorithm to find amplitude maxima with an adaptive threshold. Compared to optical microscopy, its accuracy exceeded 96%.
After that the reflected signals were converted to the frequency domain using a Fourier transform. First, the researchers extracted 15 informative features from the signals, including energy distribution parameters across the spectrum. These features were used to train classic machine learning models. However, the best result was demonstrated by a one-dimensional convolutional neural network, which analyzed the spectra directly and identified meaningful patterns independently. The accuracy of material identification at the individual particle level stood at 97.14%.
Separate models were used to estimate sizes for each material. Particles were distributed across four diameter ranges. Average size classification accuracy reached 99.93%. While the smallest particles proved particularly challenging to analyze due to the lower signal-to-noise ratio, the method demonstrated high stability overall.
The system showed high efficiency from a computational perspective as well, with processing reaching some 1,500 signals per second and the classification of a single particle taking about 16 milliseconds. This makes it possible to operate in near real time.
An advantage of this technology is that it does not require chemical reagents or complex sample preparation. Ultrasound propagates effectively in turbid water and does not depend on the optical transparency of the medium. In the future, this method could be integrated into flow-through microfluidic systems, where particles would be analyzed in the liquid flow without prior filtration or fixation.
