Mid-infrared spectroscopy and deep learning for robust classification of post-consumer plastics: A domain-tailored framework

Document Type

Article

Publication Date

4-1-2026

Department

Department of Mechanical and Aerospace Engineering

Abstract

Accurate classification of post-consumer plastics, particularly black polymers, remains a challenge for automated recycling. Mid-infrared (MIR) spectroscopy provides chemically rich signals that overcome limitations of near-infrared sensing, yet deployment is hindered by noise, baseline drift, and contamination. This paper presents a five-stage framework integrating dataset quality assessment, preprocessing optimization, tailored 1D convolutional neural networks (CNNs), explainable AI, and external validation. Using over 320,000 spectra across eight polymer classes, we introduce a ten-metric dataset quality protocol and benchmark 24 preprocessing pipelines. An optimized Savitzky–Golay smoothing with Standard Normal Variate normalization (SG+SNV), paired with a compact 374k-parameter CNN, achieves 99.71% accuracy on clean data and 98.50% on post-consumer plastics, including (Formula presented) accuracy for black polymers. Interpretability confirms reliance on chemically meaningful MIR absorption bands, while external testing demonstrates robustness to industrial noise. With an 8.4 MB model footprint and GPU inference exceeding 1250 spectra/s, the results establish technical feasibility for MIR-based plastic classification and indicate potential scalability with future system-level optimization.

Publication Title

Resources Conservation and Recycling

Link to Full Text

Share

COinS