Determination of Trace Organic Contaminant Concentration via Machine Classification of Surface-Enhanced Raman Spectra

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Data Science

Background:

• Surface-enhanced Raman spectroscopy (SERS) is a sensitive technique for detecting chemical pollutants.
• Quantifying pollutant concentrations with SERS is challenging due to spectral variability.
• Machine learning has been previously applied for SERS-based compound identification.

Purpose of the Study:

• To develop and validate a machine learning approach for predicting chemical pollutant concentrations from SERS spectra.
• To assess the impact of spectral preprocessing techniques on prediction accuracy.
• To evaluate model performance across different data set sizes and sample variations.

Main Methods:

• Collected SERS spectral data for three analytes: rhodamine 6G, chlorpyrifos, and triclosan.
• Applied frequency domain transforms (Fourier and Walsh-Hadamard) to spectral data.
• Trained standard and deep learning machine learning models for concentration prediction.
• Validated model performance using cross-validation across various conditions.

Main Results:

• Machine learning models achieved >80% cross-validation accuracy in predicting pollutant concentrations from raw SERS data.
• Deep learning models reached 85% accuracy on moderately sized datasets and 70-80% on smaller datasets.
• Fourier and Hadamard transforms consistently improved prediction accuracy.
• Model performance remained robust despite variations in sample preparation and environmental media.

Conclusions:

• The developed machine learning approach effectively quantifies pollutant concentrations using SERS spectra.
• Frequency domain transforms enhance the reliability of SERS-based quantitative analysis.
• This method offers a robust solution for chemical pollutant monitoring, overcoming previous limitations in SERS quantification.