Expert System for Fourier Transform Infrared Spectra Recognition Based on a Convolutional Neural Network With...

Area of Science:

• Spectroscopy
• Computational Chemistry
• Machine Learning

Background:

• Fourier transform infrared spectroscopy (FT-IR) is crucial for analyzing chemical compounds.
• Manual spectral interpretation is time-consuming and requires expertise.
• Identifying functional groups and bonds is key to understanding molecular structures.

Purpose of the Study:

• To develop deep learning models for automated FT-IR spectral analysis.
• To identify 17 classes of functional groups and 72 classes of coupling oscillations.
• To create visualization tools for interpreting model predictions.

Main Methods:

• Convolutional neural networks (CNNs) were employed for spectral analysis.
• A dataset of 14,361 FT-IR spectra of organic molecules was compiled.
• Shapley additive explanations (SHAP) and GradCAM were used for visualization.

Main Results:

• The models achieved F1-weighted scores of 93% for 17 classes and 88% for 72 classes.
• Incorporating absorption maxima positions improved model performance.
• Visualization tools effectively highlighted relevant spectral regions.

Conclusions:

• Deep learning offers an efficient method for FT-IR spectral interpretation.
• The developed models can accelerate routine analysis in chemistry, materials science, and biology.
• This approach aids in data preparation for scientific publications.