SERS Mixture Recognition from Pure-Substance Spectra via Component Evidence Learning and Two-Stage Inference.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Chemometrics

Background:

• Surface-enhanced Raman spectroscopy (SERS) is vital for molecular analysis but struggles with mixed spectra due to peak overlap.
• Existing machine learning methods for SERS mixture recognition often require extensive, labeled mixture datasets or component-matching, limiting their applicability.
• The difficulty in acquiring comprehensive mixture-specific data hinders the widespread use of SERS for complex analyses.

Purpose of the Study:

• To develop a novel pure-spectrum-trained framework for SERS mixture recognition.
• To enable accurate identification of chemical mixtures without relying on labeled mixture data.
• To provide a scalable strategy for SERS mixture analysis, particularly when labeled data is scarce.

Main Methods:

• A framework based on component evidence learning and two-stage inference was developed.
• The model learns reusable constituent-level evidence directly from pure-substance spectra.
• Paraquat, thiram, and tricyclazole were used as representative compounds to validate the approach.

Main Results:

• The framework achieved a high mixture recognition accuracy of 98.58% on SERS spectral datasets.
• It successfully recognized binary and ternary mixtures directly within a unified model.
• The approach demonstrated the ability to avoid mixture-specific parameter training.

Conclusions:

• Pure-substance spectral learning is a viable strategy for accurate SERS mixture recognition.
• This method offers a scalable solution for analyzing complex SERS mixtures, especially with limited labeled data.
• The developed framework overcomes the limitations of peak overlap and data acquisition challenges in SERS analysis.