Classification of Amino Acids Using Hybrid Terahertz Spectrum and an Efficient Channel Attention Convolutional Neural...

Area of Science:

• Spectroscopy
• Machine Learning
• Chemical Sensing

Background:

• Terahertz (THz) spectroscopy is crucial for analyzing molecular vibrations and rotational energy levels.
• It serves as a key technology for non-destructive inspection and molecular sensing.
• Existing methods often rely solely on absorption spectra, limiting classification accuracy and speed.

Purpose of the Study:

• To develop an advanced THz spectroscopy method for enhanced molecular classification.
• To improve the accuracy and processing speed of THz-based chemical sensors.
• To leverage high-dimensional spectral features for robust identification of molecules.

Main Methods:

• Extraction of high-dimensional features from a hybrid THz spectrum, incorporating absorption rate and refractive index.
• Development of an Efficient Channel Attention (ECA)-calibrated Convolutional Neural Network (CNN) for feature learning.
• Classification of 20 amino acids using the proposed hybrid spectral feature extraction and CNN model.

Main Results:

• Achieved classification accuracies of 99.9% and 99.2% on two independent test datasets.
• Demonstrated significant accuracy improvements of 12.5% and 23% compared to methods using only absorption spectra.
• Realized an exceptionally high processing speed of 3782.46 frames per second (fps), surpassing existing methods.

Conclusions:

• The proposed method, combining hybrid spectral features and an ECA-CNN, offers superior performance for molecular classification.
• Its high accuracy, compact size, and remarkable speed make it a viable candidate for future THz chemical sensor applications.
• This approach significantly advances the capabilities of non-destructive inspection and molecular analysis using THz spectroscopy.