Machine Learning-Enhanced SERS Sensor Using Microgroove Structures for Enriching and Confining Nanoplastics in...

Trace detection of nanoplastics (NPs) is limited by random analyte deposition and poor contact with hotspots. A major hurdle for on-site monitoring is the precise localization of enriched analytes. To address these challenges, we developed a 3D Ag/In₂O₃/fluorine-doped tin oxide microcavity array featuring microgrooves. In this design, the pinning effect of the microgrooves inhibits the stick-slip motion of the three-phase contact line, thereby suppressing the multiring deposition of NPs. This pinning mechanism transforms random analyte precipitation into directional enrichment within the 3D hotspots of microcavities, while simultaneously acting as a structural guide for the precise localization of NP-enriched zones. Consequently, the sensor achieved a limit of detection (LOD) of 2.4 × 10^-13 M for 4-MBA, with excellent spatial uniformity (RSD = 7.55%). Furthermore, the sensor successfully detected NPs of different sizes and types, including polystyrene (PS), poly(methyl methacrylate), and polyethylene terephthalate, with an LOD of 20 ng/mL for 100 nm PS. In practical analysis, the sensor achieved LODs of 360 ng/mL in river water and 2.94 μg/g in the fish matrix for PS. Additionally, ML-assisted surface-enhanced Raman spectroscopy (SERS) analysis using K-nearest neighbor, gradient boosting decision trees, convolutional neural networks (CNN), and Transformer models enabled precise classification; despite being trained on only 800 SERS spectra from deionized water, the CNN achieves 100% accuracy in river water and 99% accuracy in the fish matrix. The strategic harmonization of microgroove-induced confinement enrichment, precise localization, and 3D hotspots within microcavities, combined with small-data set machine learning, provides a field-ready solution for the detection of trace pollutants in real-world scenarios.