Machine learning-driven multidimensional tea profiling from a single SERS spectrum: toward practical application.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Chemometrics

Background:

• Multidimensional tea profiling is essential for ensuring consumer satisfaction, brand integrity, and market fairness.
• Current authentication methods can be time-consuming and may not capture the full spectrum of tea characteristics.

Purpose of the Study:

• To develop a rapid, systematic, and comprehensive strategy for multidimensional tea authentication.
• To integrate surface-enhanced Raman spectroscopy (SERS) with machine learning for single-measurement analysis.

Main Methods:

• Utilized machine learning algorithms (Partial Least Squares Discriminant Analysis and Support Vector Machine Regression) to analyze SERS spectra.
• Developed a two-tier framework for integrating predictions from individual spectral variables.
• Implemented the framework in a user-friendly application for real-time analysis.

Main Results:

• Single SERS spectra were found to inherently contain multidimensional information on tea category, grade, storage quality, and pesticide residues.
• High accuracies were achieved in classifying tea categories (≥98.9%) and grades (≥98.9%).
• Precise predictions for storage quality (R² > 0.99) and pesticide residues (R² > 0.99) were obtained. The integrated framework achieved an overall accuracy of 98.2%.

Conclusions:

• Surface-enhanced Raman spectroscopy (SERS) combined with machine learning offers an efficient, cost-effective, and scalable solution for multidimensional tea authentication.
• The developed framework enables comprehensive, real-time tea analysis from a single measurement.
• This approach provides a foundation for industry-wide data sharing and collaborative quality improvement.