Identification of surface-enhanced Raman spectroscopy using hybrid transformer network

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Surface-enhanced Raman Spectroscopy (SERS) is a sensitive technique for drug detection.
• Convolutional Neural Networks (CNNs) are used for SERS spectra identification but have limitations in sequential data analysis.
• The local receptive field of CNNs restricts comprehensive feature extraction from spectral data.

Purpose of the Study:

• To develop an advanced deep learning model for accurate SERS spectra identification.
• To overcome the limitations of CNNs in analyzing sequential spectral data.
• To enhance drug detection capabilities using SERS technology.

Main Methods:

• A hybrid Transformer network, TMNet, was developed by integrating a Transformer encoder and a multi-layer perceptron.
• The Transformer encoder utilizes self-attention for precise feature representation of sequential spectra.
• The multi-layer perceptron efficiently transforms these representations for final identification.

Main Results:

• TMNet achieved high identification accuracies: 99.07% for hair spectra and 97.12% for urine spectra.
• The model demonstrated superior performance compared to other methods, even with various noise types (Gaussian, baseline, mixed).
• TMNet exhibited excellent noise resistance and generalization capabilities.

Conclusions:

• The proposed TMNet method accurately identifies SERS spectra, offering robust noise resistance and generalization.
• This hybrid Transformer network shows significant potential for drug analysis and other spectroscopy applications.
• TMNet advances the application of deep learning in SERS-based detection and analysis.