Explainable Deep Learning and Targeted Spectral Augmentation for Mid-Infrared Classification of Additive-Containing...

Area of Science:

• Materials Science
• Spectroscopy
• Artificial Intelligence

Background:

• Deep learning excels at spectral classification but struggles with interpretability and data heterogeneity.
• Real-world plastic waste contains additives that complicate spectral analysis and are often missing from training datasets.

Purpose of the Study:

• To evaluate a 1D-CNN for classifying polymers using mid-infrared (MIR) spectra, focusing on additive effects.
• To improve polymer classification accuracy by integrating explainable AI and targeted experimental data acquisition.

Main Methods:

• A 1D-CNN was trained on MIR spectra of various polymers, including those with additives.
• 1D Grad-CAM++ was used for spectral feature interpretability, and t-SNE for visualizing spectral overlap.
• Experimental data of polymers with common additives (erucamide, calcium carbonate) were generated and incorporated into training sets.

Main Results:

• Explainable AI identified key spectral features for classifying poly(ethylene terephthalate), polyethylene, polypropylene, and polystyrene.
• Targeted experimental augmentation significantly improved standard machine learning classifier accuracy from 79.4% to 86.6%.
• The integrated workflow demonstrated enhanced classification of additive-containing polymers.

Conclusions:

• An integrated workflow combining explainable deep learning, low-dimensional embedding, and targeted experimental data acquisition improves MIR-based polymer classification.
• This approach provides a foundation for autonomous material classification, particularly for small, chemically heterogeneous spectral datasets.