Rapid Identification and Quantification of Aqueous Antibiotics over a Machine Learning-Integrated Raman Sensor Array

Surface-enhanced Raman spectroscopy (SERS) is a promising technique for on-site detection of aqueous pollutants, whereas single-substrate SERS suffers from low separability of analogous compounds and poor accuracy. Here, we propose a multisubstrate SERS fusion strategy to elevate the dimension of SERS information based on machine learning (ML) integration. An array of SERS substrates fabricated through modifying Au nanoparticles with polyvinylpyrrolidone, l-cysteine, and 3-mercaptopropionic acid exhibits two responsive modes (peak superposition and signal succession) toward three sulfonamide antibiotics. Moreover, ML models developed on multi-interface concatenated spectra (MICS) enable simultaneous identification and quantification of the sulfonamide antibiotics, outperforming any of the single substrates. The classification accuracy increased from 95-96% to 99.6% for single pollutants and 82.2-92.2% to 96.7% for mixtures, respectively, while the quantitative prediction achieved R² up to 0.99 with markedly reduced prediction errors. The method is validated in river water, lake water, and municipal wastewater secondary effluent, yielding recoveries of 75.5-116% (median 97.5%), with 23/27 measurements within 80-110%, and an overall turnaround of ∼30 min from sample preparation to readout. Through coupling the dual-mode responsive Raman sensor array with data-driven modeling, this work provides a paradigm of rapid and multianalyte SERS analysis of structurally similar pollutants for on-site environmental monitoring.