SpecRecFormer: Deep Learning-Driven Adaptive Component Identification of PAH Mixtures Based on Single-Component Raman...

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Identifying components in mixed spectra is challenging due to peak overlap, shifts, and matrix effects.
• Deep learning shows promise but is hindered by limited labeled data and fixed-threshold model issues.
• Polycyclic Aromatic Hydrocarbons (PAHs) are significant environmental pollutants requiring accurate identification.

Purpose of the Study:

• To develop a deep learning model for rapid identification of individual components in mixed PAH Raman spectra.
• To address limitations of existing methods, including data scarcity and adaptive thresholding.
• To evaluate the model's generalization capabilities on real-world datasets.

Main Methods:

• Developed SpecRecFormer, integrating a dual-channel CNN for local features and a Transformer for global representation.
• Trained the model on a reference database of single-component spectra, augmented with simulated mixed spectra.
• Implemented an adaptive threshold strategy for dynamic decision-making to improve recognition accuracy.

Main Results:

• SpecRecFormer achieved high accuracies (93.75%, 89.21%, 93.63%) on three real-world PAH datasets.
• The model demonstrated effective generalization from only four single-component reference spectra.
• Performance significantly surpassed conventional neural network models.

Conclusions:

• SpecRecFormer offers an innovative and effective approach for mixed spectral analysis.
• The model shows substantial potential for advancing environmental science and chemical analysis applications.
• This method overcomes key challenges in spectral component identification, particularly for PAHs.