DeepMIR: A Hybrid Convolutional Neural Network-Transformer Framework for Accurate Identification of Target Components...

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Accurate component identification in mid-infrared (MIR) spectra of mixtures is challenging due to spectral overlap and instrumental variability.
• Existing methods often struggle with diverse spectral acquisition techniques.

Purpose of the Study:

• To develop a robust deep learning framework, DeepMIR, for targeted component identification in complex mixtures.
• To enable accurate identification across different spectral acquisition methods (e.g., transmission, attenuated total reflectance).

Main Methods:

• DeepMIR integrates a convolutional neural network (CNN) for local spectral features and a transformer encoder for global dependencies.
• The model was trained and validated on over 67,000 synthetically augmented spectral pairs.
• A hybrid deep learning architecture was employed.

Main Results:

• DeepMIR achieved 99.5% accuracy on a test set of spectral pairs.
• High real-world validation accuracies were obtained: 94.8% for liquid solvents, 93.9% for solid pigments, and 99.1% for textiles.
• Reliable detection limits were established at 20% v/v in liquids and 10% w/w in solids.

Conclusions:

• DeepMIR offers a significant advancement over traditional library search methods for MIR spectral analysis.
• The open-access web server provides a practical and accessible tool for the scientific community.
• The framework demonstrates high accuracy and reliability in diverse real-world applications.