Identification of Polymers with a Small Data Set of Mid-infrared Spectra: A Comparison between Machine Learning and Deep Learning Models

Identifying environmental polymers and micro -plastics is crucial for the scientific world, environmental agencies, and water authorities to estimate their environmental impact and increase efforts to decrease emissions. On the basis of different spectroscopy techniques, e.g., laser-directed infrared imaging and Raman spectroscopy, polymers can be observed and represented as spectroscopic signals. The latter can be further analyzed and classified by data science, in particular, machine learning (ML). Past studies applied a variety of ML models to identify polymers from small or large data sets. However, a comprehensive comparison of multiple models across different data set sizes is still needed, which is presented in this study. Furthermore, we also provide a practical data augmentation technique to generate synthetic samples when only a limited number of samples are available. Our results show that the ensemble ML model, compared to neural network models, takes the least training time to achieve the best performance, i.e., a classification accuracy of 99.5%. This study provides a generic framework for selecting ML models and boosting model performance to accurately identify polymers.

Automated summary

Identifying environmental polymers and micro-plastics is crucial, and various spectroscopy techniques and machine learning models can be used for accurate identification and classification. This study compares different machine learning models on different data set sizes and proposes a data augmentation technique for generating synthetic samples. The results show that an ensemble ML model achieves the best performance with a classification accuracy of 99.5%.