Real-time mid-infrared characterization of post-consumer plastics for in-line quality control using physics-informed deep learning

Reliable characterization of post-consumer plastic waste is essential for cleaner production and the circular economy of post-consumer polymers, yet most mid-infrared (MIR) spectroscopy studies are evaluated under stationary laboratory conditions that do not capture conveyor motion, contamination, optical variability, or carbon-black pigmentation. This study presents a real-time MIR characterization framework based on physics-informed progressive data augmentation and compact deep learning, extending a previously validated and published static-laboratory foundation—acquisition protocol, dataset design, preprocessing pipeline, and compact network architecture—to conveyor-relevant operation. Six acquisition-relevant spectral disturbances, including temperature-induced wavelength drift, detector noise, illumination variation, motion blur, wavelength-dependent gain, and vibration jitter, are incrementally introduced during training to improve robustness under conveyor-relevant conditions. The resulting 374k-parameter convolutional neural network improves characterization accuracy on moving, contaminated samples from 86% to 94%, while maintaining clean-sample accuracy above 98% across seven polymers and 96% accuracy for carbon-black plastics. Identically trained chemometric baselines (PLS-DA, SVM-RBF) reach at most 88% on the same real conveyor data, with class-level failures, so the network’s margin emerges precisely under real acquisition conditions. Across three independent runs, mean clean accuracy reaches 98.73% ± 0.14%, reducing run-to-run variability by 84% and narrowing the clean–contaminated performance gap. A confidence-aware prototype provides real-time in-line quality control, with a computational processing throughput of approximately 1250 spectra/s, sub-second latency, a live operator dashboard, and 99.73% uptime. The results demonstrate that physics-informed learning can improve MIR-based polymer characterization under realistic acquisition disturbances, supporting more reliable decision-making in circular-economy recovery and sustainable-manufacturing (closed-loop recycling) workflows.