By Wayne Parker

Microplastic pollution is a growing concern that has led to the need for analytical tools that can accurately and rapidly identify and quantify plastics in environmental samples, such as soil, water, and air. A recent article published by members of the Microplastics Fingerprinting research team introduces a new AI tool they developed and tested called PlasticNet. The tool uses a deep learning convolutional neural network architecture to identify microplastics using images obtained from infrared spectroscopy. The team tested PlasticNet’s ability to recognize microplastics with a range of thickness, additives and surface modifications using a number of datasets from real wastewater treatment plant samples. The model was initially trained using 8,000 spectra from virgin plastics and was then retrained using spectra from “non virgin” plastic datasets. Finally, the performance of PlasticNet was compared to that of an existing “library” search method which relies heavily on expert knowledge to interpret spectral signals. 

Once trained, PlasticNet successfully classified 11 types of common plastic with a 95 per cent level of accuracy. Errors were caused by edge effects, molecular similarity of plastics, and the contamination of standards that were employed to train the model. PlasticNet also demonstrated good performance (more than 92 per cent) in recognizing spectra with increased complexity, such as when samples included additives or experienced weathering. When the team compared the accuracy of PlasticNet to the library search approach, the identification of Polyethylene (PE) improved by 17.3 per cent while both approaches recognized Polypropylene (PP) similarly. However, PlasticNet was 46 per cent faster at recognizing plastics than the library method. Overall, these results suggest that deep learning approaches, such as those used in PlasticNet, have considerable potential to improve the quantity and quality of data generated in microplastic research. 

More information about this work can be found in the research paper, “Leveraging deep learning for automatic recognition of microplastics (MPs) via focal plane array (FPA) micro-FT-IR imaging”, published in Environmental Pollution.