Deep learning-based Raman spectroscopy qualitative analysis algorithm: A convolutional neural network and transformer...

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Raman spectroscopy provides detailed chemical structure information non-destructively.
• Traditional chemometric methods struggle with complex spectral overlap and noise in mixture analysis.
• Deep learning offers a novel approach for qualitative analysis of mixed Raman spectra.

Purpose of the Study:

• To develop a deep learning-based algorithm for qualitative analysis of mixed Raman spectra.
• To enhance component identification accuracy and robustness in complex mixtures.
• To improve the interpretability of spectroscopic learning patterns.

Main Methods:

• Proposed a deep learning algorithm (RST) integrating convolutional neural network and Transformer architectures.
• Transformed Raman spectra into 64 word vectors to determine component contribution weights.
• Validated the algorithm using 75 spectral datasets.

Main Results:

• Achieved 100.00% positive identification rate and 99.3% recall rate on validation data.
• Attained an average identification score of 9.51.
• Demonstrated superior accuracy and robustness compared to traditional Convolutional Neural Network (CNN) models in complex mixtures.

Conclusions:

• The RST algorithm effectively analyzes mixed Raman spectra, overcoming limitations of traditional methods.
• The model shows high performance and interpretability, applicable to Raman and surface-enhanced Raman spectroscopy.
• Enhanced understanding of spectroscopic learning patterns facilitates future analysis of more complex mixtures.