Machine Learning-Assisted SERS Sensor for Fast and Ultrasensitive Analysis of Multiplex Hazardous Dyes in Natural Products

Abstract

The adulteration of natural products with multiple azo dyes has become a serious public health concern. Thus, on-site trace additive detection is demanded. Herein, we developed a gold-nanorod-based surface-enhanced Raman scattering (SERS) sensor to detect trace amounts of azo dyes, including lemon yellow, sunset yellow, golden orange II, acid red 73, coccine, and azorubine. After optimizing pre-processing steps, the additives were separated and identified through visual observation. The stable and sensitive SERS sensor developed enabled accurate detection of the added colorants. Density Functional Theory confirmed that the characteristic SERS peaks of the six dyes were accurate and credible. The optimized SERS sensor achieved a detection limit of 50 mg of dye per kilogram of raw material. A SERS spectral dataset comprising 960 replicates from all 64 potential dye combinations was generated, forming robust training sets. The K-Nearest Neighbor model exhibited best performance, identifying dye additives in real samples with a 91% success rate. This model was further validated by screening nine randomly collected safflower batches, identifying three with illegal dye additives, which were subsequently confirmed by HPLC. Summarily, the developed SERS sensor and classification model offer an ultrasensitive, and reliable approach for on-site detection of hazardous dyes in natural products.