Machine Learning-Enhanced Raman Spectroscopy for Fast Nanoplastic Detection at Low SNR

Raman spectroscopy is widely used to identify substances such as microplastics. As the size decreases to nanoplastics, their numbers are expected to increase, making it crucial to determine how many samples can be analyzed quickly and accurately. However, there is always a trade-off between analysis speed and the signal-to-noise ratio (SNR). Obtaining a low SNR is easy, but such spectra have reduced reliability for identification, while achieving a high SNR necessary for accurate identification inevitably requires longer analysis times. With enhanced artificial intelligence, these challenges can be overcome. In practice, establishing a well-labeled low SNR database is crucial for accurate AI-based substance identification. In this study, we have developed a systematic machine-learning approach to address this issue. To address this, we augmented Raman spectral databases by averaging numerous low-SNR spectra acquired over extremely short exposure times, thereby generating databases with broad SNR ranges. From such a broad database, we effectively distinguished noise from signals by using a combination of unsupervised clustering methods—including autoencoder, PCA, and DBSCAN. Such filtered signals were then used to train Bi-Convolutional Neural Network technique, enhanced with positional encoding and multi-head attention, achieving 99.29% (±0.58%) accuracy in classifying spectra with an SNR close to 2, even with very short Raman measurement times (0.001 s). This methodology enables the rapid analysis of nanoplastics and other low-concentration substances, reduces hardware requirements, and has broad applications in fields such as bioprocessing and food analysis. Additionally, this approach can be extended to other spectroscopic methods, accelerating data processing in data-limited environments.