Multi-Scale Feature Attention Network for Polymer Classification using THz Dual-Comb Spectroscopy

Researchers developed a Multi-Scale Feature Attention Network (MSFAN) that combines Terahertz Dual-Comb Spectroscopy with deep learning to classify 12 types of polymers with 85.2% accuracy. This approach offers a non-destructive, rapid alternative to conventional sorting techniques for recycled plastics, addressing critical quality and safety concerns in plastic recycling industries.

This research addresses a significant bottleneck in the circular economy: reliable polymer identification for recycling. Current sorting methods rely on manual labor, infrared spectroscopy, or density-based separation, all of which struggle with mixed plastics, multilayer films, and contaminated materials. THz-DCS overcomes these limitations by providing high-resolution spectral fingerprints across a broad frequency range that captures unique molecular vibration patterns invisible to conventional techniques. The MSFAN architecture represents a meaningful advance in applying deep learning to spectroscopic data. Rather than treating spectral signals as generic sequences, the model incorporates domain-specific design choices: feature gating recalibrates signal importance, multi-scale convolutions capture patterns across different frequency resolutions, and attention mechanisms automatically identify the most diagnostic THz regions. Achieving 85.2% accuracy across 12 polymer classes—including challenging cases like commercial blends and biopolymers—demonstrates practical viability. For the recycling industry, this combination enables faster, higher-purity material sorting, directly reducing contamination in recycled plastic streams and improving material value. Equipment manufacturers could commercialize THz-DCS sorters with embedded AI models, creating new market opportunities. From an AI perspective, this work exemplifies how combining domain expertise with deep learning yields superior results compared to generic models. The interpretability aspect—identifying which THz regions drive classifications—builds confidence for industrial deployment. Future deployment challenges include scaling production THz-DCS hardware, validating performance on real-world contaminated samples, and competing on cost against incumbent sorting technologies.

• →MSFAN achieves 85.2% accuracy classifying 12 polymer types using THz-DCS, outperforming existing spectroscopic and deep learning approaches.
• →Terahertz spectroscopy provides non-destructive, rapid analysis of complex polymers including multilayer films and commercial blends that challenge conventional sorting.
• →Deep learning architectures with attention mechanisms and multi-scale feature processing significantly improve polymer classification from spectral data.
• →This technology addresses a critical recycling industry need for reliable, automated polymer identification to improve material purity and circular economy viability.
• →Commercialization potential exists for THz-DCS sorting systems integrated with AI models, competing in the multi-billion-dollar plastic sorting equipment market.