Complex extreme learning machine applications in terahertz pulsed signals feature sets.

Area of Science:

• Spectroscopy
• Machine Learning
• Data Science

Background:

• Terahertz (THz) spectroscopy generates large datasets with potential applications in various scientific fields.
• Classifying complex THz spectra, especially those with subtle features, presents a significant challenge.
• Existing methods may not fully leverage the amplitude and phase information present in THz signals.

Purpose of the Study:

• To develop and evaluate a novel approach for automatic classification of large terahertz pulse transient signal datasets.
• To compare the performance of a complex-valued extreme learning machine (ELM) algorithm against a support vector machine (SVM) for THz spectral classification.
• To investigate the utility of amplitude and phase signatures in THz spectra for classification tasks.

Main Methods:

• Utilized a complex-valued extreme learning machine (ELM) algorithm for spectral classification.
• Considered both amplitude and phase features of THz spectra as input vectors.
• Performed binary classification (poly-A vs. poly-C RNA) and multi-class classification (six powder samples).
• Compared ELM performance with a support vector machine (SVM) using different Gaussian kernels.

Main Results:

• The complex-valued ELM demonstrated high accuracy and speed in classifying large THz datasets.
• The ELM effectively utilized both amplitude and phase information for improved classification.
• The algorithm showed robustness in handling noisy data and identifying subtle spectral features.
• Performance was systematically compared across different Gaussian kernels for amplitude and phase signatures.

Conclusions:

• Complex-valued ELM algorithms are effective chemometric tools for classifying large THz datasets.
• This approach supports the broader adoption of THz sensing technology in chemical sensing, quality control, security, and diagnostics.
• The method is robust for heterogeneous materials and noisy data, applicable to tomographic settings and other large-dataset classification problems.