Plastic waste in our lakes and oceans breaks down into microplastics and nanoplastics, particles small enough to spread through aquatic environments, carry pollutants up the food chain, and pass for natural debris. Identifying them quickly and accurately remains an unsolved problem. Our research uses artificial intelligence, scanning electron microscopy, and ultrasound to detect, classify, and measure these particles, and to study how they form from everyday plastic products.
Deep learning models are trained to detect and segment plastic particles in SEM images and to characterize each particle’s morphology. Their trained weights are then used to prompt foundation models, reducing the need for manual annotation.
A pipeline that localizes particles in ultrasound scans, extracts features from their high-frequency ultrasound RF signals, and uses deep learning to identify each particle’s material and estimate its size.
Particles are focused into a single stream and pass an ultrasound transducer, where short RF signals are captured for identification. Transformers adapted to these signals prove effective, and the study also examines model interpretability and throughput.
H-SPAR is an open-source ROS 2 and Gazebo simulator that combines current-driven particle transport with autonomous surface vessel missions.
Which sheds more microplastics, a plastic bag or a water bottle? Particles released by everyday plastics are compared in SEM images, with FastSAM used to accelerate annotation.