Machine learning driven trace detection of pesticide mixtures using citrate optimized Au nanoparticles based in-expensive efficient micro-drop SERS with portable spectrometer.
Journal:
Spectrochimica acta. Part A, Molecular and biomolecular spectroscopy
Published Date:
May 4, 2025
Abstract
Machine learning (ML)-based surface-enhanced Raman scattering (SERS) spectra analysis has potential to determine the composition of analyte mixtures. The requirement for this task is to acquire spectral datasets from numerous sample matrices and experimental parameters, highlighting the need for an effective SERS technique. This study focuses on development of a cost-effective, straightforward sample preparation SERS technique using citrate-optimized gold nanoparticles (GNPs) attractive for portable spectrometers. Significance of optimal tri-sodium citrate (TSC) to gold precursor ratio and GNP characteristics, like aggregation and size, for maximum SERS signal enhancement is addressed. Additionally, the concentration dependence of SERS intensity variation was correlated to GNPs number-concentration predicted by Mie theory, and extent of analyte adsorption. Raster-scan measurements showed improvement in signal reproducibility. This approach yielded easy detection of about 250 nM and 1.25 µM for thiram and phosmet in beetroot juice, and 5 and 50 nM in water samples, respectively. Further, the potential benefit of this technique is demonstrated with ML based identification of composition of pesticide mixtures using unsupervised and supervised data dimensional reduction methods. For subtle variation spectral classes, effectiveness of five ML models is analyzed with data augmentation, normalization and different number of features and over 97 % classification accuracy is achieved. The study also addresses the principal component analysis for the identification of concentrations in both single and binary mixtures. This efficient SERS technique leverages 3D hot spots from aggregated GNPs of optimum TSC, provides a user-friendly platform to rapidly generate extensive spectral datasets for a variety of sample matrices and compositions at significantly low-cost. Consequently, it is well-suited for novice users, offering the potential for customized ML-driven application development.
Authors
Keywords
No keywords available for this article.