We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Mid-IR ATR-FTIR with Simple ML Framework for PVC Screening and Polyethylene Subclassification with Perfect External Validation
Summary
Mid-infrared ATR-FTIR spectroscopy combined with simple SVM and neural network models achieves 100% classification accuracy across seven common plastics — including PVC and three polyethylene subclasses — even on post-consumer materials. Precise, low-cost plastic identification is foundational for improving recycling sortation and reducing misidentified polymer waste streams that ultimately fragment into environmentally persistent microplastics.
Reliable and chemistry-specific presorting of waste plastics is a prerequisite for safe and efficient recycling and upcycling. In this study, we show that mid-infrared ATR-FTIR spectra in the fingerprint region are sufficient to perfectly classify seven common plastics including polyvinylchloride (PVC) and to robustly resolve the three PE subclasses when combined with simple machine-learning models. We collected 770 ATR-FTIR spectra (2500-500 cm⁻¹) from LDPE, LLDPE, HDPE, PP, PS, PET and PVC, using commercial pellets together with PET from a post-consumer bottle and PVC from discarded laboratory gloves. Each pellet or cut piece was placed so that a reasonably flat face contacted an ATR crystal and was measured under a fixed clamp setting; spectra were recorded with about 2 cm⁻¹ nominal spacing and normalized before analysis. After truncation to the fingerprint region and normalization, spectra were compressed by principal component analysis (PCA), and the resulting scores were used to train support-vector machine (SVM) and shallow deep neural network (DNN). For the baseline dataset, all four models achieved 100 % accuracy on both internal test splits and a strictly independent external validation set; PVC false-negative and false-positive rates were 0 %, and LDPE, LLDPE and HDPE were perfectly resolved. To assess robustness to reduced spectral quality, we numerically degraded the spectra to effective resolutions of about 8 and 16 cm⁻¹ by convolving with a broadened instrument function and rebinning to coarser wavenumber grids. Even under the most degraded condition (16 cm⁻¹), SVM and the shallow DNN retained 100 % internal and external accuracy for PVC screening, PE-grade subclassification and full seven-class discrimination. These results demonstrate that ATR-FTIR fingerprints combined with straightforward machine-learning classifiers can deliver perfectly reliable PVC detection and PE subclassification across a realistic range of spectral resolutions, and they provide a general framework that can be extended to other grindable solid feeds.