Identification of recycled PET fibers using TMAH-assisted reactive pyrolysis-GC/MS and machine learning

Recycled poly(ethylene terephthalate) (rPET) is increasingly used in polyester fibers, but reliable molecular-level identification remains challenging because chemical differences between virgin poly(ethylene terephthalate) (vPET) and rPET are often weak and distributed across multiple trace features. In this study, a tetramethylammonium hydroxide (TMAH)-assisted reactive pyrolysis-gas chromatography/mass spectrometry (Py-GC/MS) workflow coupled with machine learning was developed to distinguish rPET from vPET fibers. Poly(ethylene terephthalate) (PET) fiber samples were analyzed by full-scan TMAH-assisted Py-GC/MS, and peak-level features were extracted using MS-DIAL software. To reduce the influence of background peaks and unstable integrations, candidate variables were selected within the training set using partial least squares-discriminant analysis (PLS-DA) variable importance in projection (VIP) ranking and were further curated by manual inspection of peak shape, retention behavior, quantifier ions, and electron ionization (EI) mass spectra. Six classifiers, including LDA, RF, SVM, XGBoost, ElasticNet, and a lightweight neural network, were compared using repeated cross-validation. The SVM model showed the best overall performance, with accuracy, balanced accuracy, and receiver operating characteristic area under the curve (ROC AUC) of 0.875, 0.887, and 0.956, respectively. For the final SVM model, training and held-out test accuracies were 0.925 and 0.909, respectively. Candidate discriminant peaks mainly included low-abundance aromatic fragments, fatty acid methyl esters, hydrophobic accompanying compounds, and reproducible unknown features, suggesting that rPET/vPET differences arise from combined molecular fingerprints rather than a single marker. This workflow provides a chemically interpretable approach for the identification of mechanically recycled polyester fibers.