Imbalanced spectral data analysis using data augmentation based on the generative adversarial network.

Area of Science:

• Materials Science
• Chemical Engineering
• Spectroscopy
• Machine Learning

Background:

• Spectroscopic techniques provide unique material fingerprints through frequency-domain peaks.
• Deep neural networks (DNNs) are powerful for spectral data classification.
• Imbalanced spectral data in real-world experiments hinders DNN performance, impacting material phase behavior analysis.

Purpose of the Study:

• To develop and validate a novel data augmentation method for imbalanced spectral data.
• To improve the classification accuracy of material phases, specifically the sol-gel transition in hydrogels.
• To enhance the understanding of soft materials and glycomaterials using spectroscopic data.

Main Methods:

• Application of a generative adversarial network (GAN)-based data augmentation technique.
• Utilizing a three-DNN architecture: generator, discriminator, and classifier.
• Training and testing on imbalanced spectral data from Pluronic F-127 and Alpha-Cyclodextrin hydrogels.

Main Results:

• The proposed GAN-based method significantly improves classification performance over existing augmentation techniques.
• Average improvements of 8.8% in F-score, 6.4% in Precision, and 6.2% in Recall were achieved.
• Generated samples emphasize material characteristic differentiation, leading to balanced and more informative training datasets.

Conclusions:

• The novel GAN-based data augmentation effectively addresses imbalanced spectral data challenges.
• The method enhances the classification of material phases, crucial for understanding soft and glycomaterials.
• This approach holds potential for broad applications in materials science and chemical engineering dealing with imbalanced measurement data.