A modular convolutional neural network framework for Raman spectra-based identification of environmental...

Area of Science:

• Environmental Science
• Analytical Chemistry
• Materials Science

Background:

• Raman spectroscopy is crucial for microplastic (MP) identification, but conventional methods like Hit Quality Index (HQI) lack sensitivity and precision in complex samples.
• Expert manual interpretation improves accuracy but is inefficient, limiting high-throughput analysis of environmental MPs.

Purpose of the Study:

• To develop and optimize modular convolutional neural network (CNN) models for sensitive and precise identification of seven polymer types in environmental samples using Raman spectroscopy.
• To compare the performance of CNN models against multilayer perceptron (MLP) models and establish optimal training data requirements.

Main Methods:

• Optimization of modular CNN models using approximately 16,000 expert-verified Raman spectra from various water sources.
• Evaluation of model performance using spectra with noise and non-target signals.
• Comparison of CNNs with MLPs, particularly concerning training data quantity and balance.

Main Results:

• Optimized CNN models achieved 97%-100% sensitivity and >50% precision for target polymers.
• Effective CNN training required 500-1500 unique spectra per polymer; oversampling did not enhance performance.
• Targeted spectral inclusion reduced model bias, improving sensitivity and precision.
• CNNs outperformed MLPs with abundant data; MLPs were better with limited or imbalanced data.

Conclusions:

• Modular CNNs provide a flexible, scalable, and robust framework for high-throughput microplastic identification from Raman spectra in complex environmental matrices.
• A hybrid workflow combining machine learning screening with expert verification significantly reduces manual workload while maintaining data quality.
• Careful model evaluation using realistic spectral data (including noise) is essential to avoid overestimating performance.