We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
A modular convolutional neural network framework for Raman spectra-based identification of environmental microplastics
Summary
Scientists trained computer models (using AI) to automatically identify tiny plastic particles found in lakes, rivers, and drinking water, making this detection process much faster than having human experts check every sample by hand. This matters because finding microplastics in our water is the first step to understanding how much we're exposed to and what risks they might pose to our health—and faster, more reliable detection tools could help researchers monitor contamination at a much larger scale.
Raman spectroscopy is widely employed in microplastic (MP) research for polymer identification. However, when applied to spectra obtained from complex environmental samples, the conventional spectral interpretation approach based on Hit Quality Index (HQI) is limited with respect to sensitivity and precision. While manual interpretation by experts provides improved accuracy, it is time-consuming thus restricting its application. To address this challenge, a series of modular convolutional neural network (CNN) models which target seven different polymer types (originating from lakes, rivers, and drinking water) were optimized by incorporating approxiamtely 16,000 expert-verified Raman spectra. Optimized CNN models achieved high sensitivity (97%-100%) across all target polymers with precision exceeding 50% for most. It was determined that approximately 500-1500 unique spectra per polymer type were required for effective training of CNN models; oversampling via use of replication did not improve performance. Targeted inclusion of specific types of spectra was identified as a viable approach to reduce model bias and improve sensitivity and precision. When compared to multilayer perceptron (MLP) models, CNNs achieved higher sensitivity and precision when abundant training data was available, whereas MLPs were beneficial when training data was limited or imbalanced (between positive and negative). These results highlight the importance of evaluating models using spectra which include noise and non-target signals, in order to avoid overestimation of performance. A hybrid workflow which combined preliminary machine learning (ML) screening with expert verification was found to substantially reduce manual workload while maintaining data quality. The modular ML-based framework developed in this study provides a flexible, scalable, and robust strategy for identification of MPs using Raman spectra, offering a pathway toward high-throughput analysis of microplastics in complex environmental matrices.