Machine learning enabled surface-enhanced Raman spectroscopy quantitative analysis for food safety monitoring.

Crystal violet (CV), a prohibited dye frequently misused in aquaculture, presents notable toxicological concerns and continues to threaten food safety. In this work, a fast and highly sensitive strategy was developed for the determination of CV residues in fish matrices by coupling surface-enhanced Raman spectroscopy (SERS) with a tailored machine learning approach. Silver nanocubes were employed as the SERS-active platform, offering abundant electromagnetic enhancement sites that amplify spectral signals. To achieve accurate quantification, three regression algorithms-Partial Least Squares (PLS), Support Vector Machine regression (SVM), and Random Forest (RF)-were comparatively evaluated. The findings show that PLS performance is limited by its linear assumptions and susceptibility to overfitting (R2 = 0.9529), whereas SVM demonstrates robustness mainly at lower concentration levels in the presence of noise. In contrast, the RF model exhibits the most reliable predictive capability, yielding a test-set R2 of 0.9926 alongside a low mean absolute error of 0.0046 ppm. Further interpretation of feature contributions indicates that RF effectively integrates spectral information across a wide frequency range, from low-frequency structural vibrations to high-frequency stretching modes of conjugated systems, thereby maintaining prediction stability even under conditions approaching signal saturation.