Hyperspectral imaging for chemicals identification: a human-inspired machine learning approach.

Area of Science:

• Spectroscopy and Analytical Chemistry
• Machine Learning and Artificial Intelligence
• Materials Science

Background:

• Machine learning (ML) algorithms lack domain flexibility due to their reliance on mathematical models without physical understanding.
• Identifying thin organic layers with hyperspectral imaging (HSI) is challenging due to nonlinear spectral mixing.

Purpose of the Study:

• To develop a machine-educating approach for accurate identification of thin organic materials using HSI.
• To overcome the limitations of classical ML in spectral analysis and improve generalization capabilities.

Main Methods:

• A machine-educating framework was developed, integrating a physical model, universal building blocks, and unlabeled HSI data.
• The machine was trained to resolve nonlinear spectral mixtures and identify target material signatures.
• Inputs included a nonlinear mixing model, pure material spectra, and unlabeled HSI data.

Main Results:

• The educated machine demonstrated superior accuracy and generalization compared to classical ML approaches.
• Falsely identified samples were reduced approximately 100-fold with the educated machine.
• The probability of detection increased to 96% from 90% compared to classical methods.

Conclusions:

• Machine education offers a robust approach for analyzing complex spectral data in HSI.
• This method enhances the identification of thin organic layers, outperforming traditional machine learning.
• The physical model integration provides better accuracy and reliability in spectral analysis.