Three-dimensional fluorescence spectroscopy recognition and component analysis based on machine learning.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Three-dimensional excitation-emission matrix (3D EEM) fluorescence spectroscopy is valuable for identifying fluorescent substances.
• Spectral contamination from impurities can compromise identification accuracy.
• Developing robust methods for contamination detection and component analysis is crucial.

Purpose of the Study:

• To present an integrated machine learning and 3D EEM spectroscopy approach for contamination detection and component analysis.
• To evaluate the performance of various machine learning algorithms in identifying contaminants.
• To develop an advanced deep learning model for analyzing complex fluorescent mixtures.

Main Methods:

• Collected 3D EEM fluorescence spectral data in a simulated contaminated environment.
• Assessed contaminant detection using K-Nearest Neighbors (KNN), Random Forest (RF), and Convolutional Neural Network (CNN).
• Developed and trained a Shared Encoder U-Net (SE-UNet) model using PARAFAC-derived spectral profiles for component analysis.

Main Results:

• All tested deep learning models showed comparable accuracy in simple binary-component scenarios.
• The optimized SE-UNet demonstrated superior performance and generalization in complex mixtures compared to CNN and VGG architectures.
• SE-UNet enabled rapid single-sample inference, outperforming the iterative Parallel Factor Analysis (PARAFAC) method.

Conclusions:

• The integrated framework offers a practical and scalable solution for contamination analysis.
• The SE-UNet model provides a robust tool for identifying fluorescent impurities and analyzing complex mixtures.
• This approach enhances the reliability of 3D EEM spectroscopy in laboratory and environmental monitoring.