Development of a machine-learning model for microplastic analysis in an FT-IR microscopy image

Abstract

The escalating concern regarding microplastics (MPs) in the environment has recently accentuated the need for comprehensive analyses across various matrices. Fourier Transfrom Infrared (FT-IR) microscopy is widely used method for MP identification, but challenges arise due to the presence of secondary materials on real samples, causing inaccuracies in spectral matching. To tackle this issue, we propose a solution: a 1D-convolution neural network (1D-CNN) machine-learning model classifying FT-IR spectra into 16 polymer species. Using a dataset of 5413 spectra, with 80% (4330) for training and 20% (1083) for external testing, our method achieved 98.59% accuracy for cross-validation and 92.34% for external validation. This study underscores the efficacy of machine learning in discerning polymer types among MPs, even in real samples tainted by secondary materials. The implementation of our 1D-CNN model marks a significant leap in overcoming conventional method limitations, providing a robust tool for accurately unraveling MPs intricacies in environmental matrices.