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

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

Problem Statement
Reliable polymer identification is critical for the quality and safety of recycled plastics, particularly in food packaging and medical devices. However, industrial recycling faces bottlenecks due to the difficulty of accurately classifying polymers within heterogeneous waste streams containing blends, multilayer films, and biopolymers. Conventional sorting methods (manual inspection, chemical testing) are often destructive or limited to surface properties. While spectroscopic techniques like Infrared (IR) and Raman offer non-invasive alternatives, they suffer from environmental noise, weak signals, or fluorescence interference. Terahertz (THz) spectroscopy presents a promising solution due to its high penetration depth and ability to measure both amplitude and phase. Yet, conventional THz Time-Domain Spectroscopy (THz-TDS) is limited by acquisition speed and spectral resolution, hindering high-throughput industrial application. Furthermore, existing deep learning approaches for polymer classification often rely on datasets composed exclusively of pure polymers or synthetic mixtures, failing to reflect the complex material variability (blends, multilayers) found in real-world scenarios.

Methodology
The authors propose a framework utilizing Terahertz Dual-Comb Spectroscopy (THz-DCS) combined with a novel deep learning architecture, the Multi-Scale Feature Attention Network (MSFAN), to classify 12 distinct polymer types.

1. Data Acquisition and Preprocessing:

   • Dataset: Measurements were collected over five days for 12 polymer categories, including pure polymers, multilayer films, commercial blends, and biopolymers. Signals were acquired in the 100–590 GHz range with 10 GHz steps (50 frequency points) using both low-gain (LG) and high-gain (HG) configurations.
   • Physical Invariance: To address the dependence of spectral magnitude on sample thickness (governed by the Beer-Lambert law), the authors implement a thickness-invariant normalization strategy. Raw spectral amplitudes are divided by the sample thickness ($d$), creating a feature space $\tilde{X}$ that relies solely on intrinsic spectral signatures.
   • Signal Integration: Experiments compared single-source inputs (LG, HG) against fusion strategies (averaging, maximization, concatenation, and late fusion). Results indicated that the isolated high-gain (HG) configuration yielded the best performance, while low-gain data introduced noise.

2. MSFAN Architecture:
   The proposed network processes the thickness-invariant spectrum through four key stages:

   • Spectral Recalibration (Gating): A learnable dense layer with a hard-sigmoid activation generates an attention mask. This mask is applied element-wise to the input spectrum to dynamically reweight spectral bands, suppressing background noise and emphasizing discriminative interference features.
   • Multi-Scale Feature Extraction: An inception-inspired block employs three parallel 1D convolutional branches with kernel sizes of 3, 5, and 7. This allows the simultaneous capture of high-frequency local peaks and broad spectral trends.
   • Global Contextualization (MHSA): A Multi-Head Self-Attention (MHSA) mechanism treats the feature map as a sequence to capture long-range dependencies and global spectral context.
   • Attention Pooling and Classification: A content-aware attention pooling mechanism aggregates the sequence into a fixed-length embedding by computing scalar importance scores for each frequency. This embedding is fed into a Multilayer Perceptron (MLP) with dropout regularization for final classification.
   • Optimization: The model is trained using a weighted combination of cross-entropy loss (with label smoothing) and an $L_1$ sparsity penalty on the gating activation vector. This sparsity constraint forces the network to focus computational resources on discriminative frequencies while suppressing irrelevant bands.

Key Contributions

• THz-DCS Validation: The study demonstrates the suitability of THz-DCS for accurate polymer classification using high-resolution spectra across a diverse set of 12 materials, including complex multilayer films and blends.
• Novel Architecture (MSFAN): The authors introduce a deep learning framework integrating spectral selection, multi-scale convolutions, cross-feature attention, and attention-based pooling. The architecture is designed to inherently identify the most informative THz frequency regions.
• Comprehensive Evaluation: Unlike prior works focusing on pure polymers, this framework is evaluated on a heterogeneous dataset reflecting real-world industrial variability.

Results

• Performance: MSFAN achieved a classification accuracy of 85.2% and an F1-score of 82.4%, outperforming state-of-the-art benchmarks including Improved CNN PM (83.8% accuracy), Transformer PM, and standard ANN/CNN baselines.
• Signal Processing: Thickness-invariant normalization improved performance by an average of 16.05% compared to raw spectral inputs. The high-gain (HG) signal configuration proved superior to low-gain or fused configurations.
• Ablation Studies:
  • Regularization: An optimal sparsity weight ($\lambda = 0.1$) was identified, improving accuracy from 82.9% (unregularized) to 85.2%.
  • Component Analysis: The full integration of all modules (Gating, Convolutions, Attention, Pooling) provided a 13.9% absolute improvement in accuracy over a baseline pooling-only model.
  • Specificity: The model significantly improved the discrimination of difficult pairs, such as distinguishing Polymer H (PP/tie/EVOH/tie/PP) from Polymer J (LDPE), raising pairwise accuracy from 61.1% (Improved CNN PM) to 75.1%.
• Interpretability: Spectral saliency analysis revealed that the model adopts distinct recognition strategies for different polymers, prioritizing specific frequency ranges (e.g., 300–400 GHz for some, >400 GHz for others) based on their structural properties.

Significance and Claims
The paper claims that combining THz-DCS with the proposed deep learning techniques offers a pathway toward effective, scalable, and interpretable polymer classification. By addressing the complexity of spectral signals through multi-scale feature attention and physical invariance preprocessing, the study establishes a new benchmark for accuracy in this domain. The authors note that while the current work operates under a closed-set assumption on uncontaminated samples, the demonstrated ability to isolate discriminative frequency bands lays the groundwork for future work involving open-set recognition and robustness against industrial contaminants. The research is funded by the Generalitat Valenciana under project CIGE/2024/147 (CLAIRE).