Highly sensitive superhydrophobic SERS substrate combined with machine learning for precise identification and classification of nanoplastics.

The ubiquitous use of plastics in modern society is accompanied by the pervasive occurrence of plastic pollution. Of particular concern are nanoplastics due to their capability to penetrate biological barriers and cause bioaccumulation, posing an emerging environmental and health threat. Consequently, the development of rapid, straightforward, and sensitive nanoplastic detection methods is urgently required. Surface-enhanced Raman scattering (SERS) stands out as a highly promising solution, offering intrinsic molecular fingerprinting, high sensitivity, and minimal sample preparation. In this study, a superhydrophobic SERS substrate was fabricated via the creation of a microstep/nanosheet hierarchical structure on an Al surface, followed by PDMS spin-coating and Ag deposition. The optimized SERS substrate, benefiting from the combined effect of high-density hot spots and superhydrophobicity, delivers a high enhancement factor (2.89 × 107), along with excellent signal uniformity and outstanding reproducibility (RSD = 5.02%). For real world applications, the substrate demonstrates the capability to identify single nanoplastics (PS, PET, PP). Furthermore, it maintains a sensitive SERS response to complex plastic mixtures, even when spiked into challenging lake water matrices. Given the spectral overlap of some characteristic peaks, conventional machine learning algorithms and 1D-CNN were adopted to reliably distinguish the seven target nanoplastic types spiked in real lake water. The former proved highly effective with over 95% accuracy, whereas the latter achieved a notably high accuracy of 99.88%. Considering the diverse interfering substances present in real lake water, a confidence-based rejection strategy was integrated into the 1D-CNN model to effectively screen out unknown environmental interferences while preserving exceptionally high accuracy. The deep integration of machine learning with SERS in this study provides a high-throughput and highly reliable detection solution for monitoring nanoplastics, demonstrating significant practical application potential.