Abstract
Fourier-transform infrared (FTIR) classifiers can assign familiar polymer labels to spectra outside their training domain. We developed SPARC-FTIR to couple class-conditional conformal prediction with a calibrated distance gate for binary high-density polyethylene (HDPE) and polypropylene (PP) identification with rejection. Five spectral views were combined in a training-derived principal-component representation. The two gates distinguished label-set uncertainty from departure from the known feature domain and produced four evidence states. After auditing sample identifiers, 776 spectra were retained from an open ATR-FTIR collection; 689 HDPE/PP spectra supported training, calibration and testing, while 87 non-target spectra were withheld. On the primary test partition, closed-set balanced accuracy was 100%, 124 of 139 known spectra were accepted correctly, and 37 of 87 unknown spectra were rejected. Rejection was 29.4% for low-density polyethylene, 33.3% for mixtures and 100% for the 14 remaining unknown spectra. Across 30 group-preserving splits, mean known acceptance and unknown rejection were 92.8% and 36.3%, respectively. Matched spectral-library comparisons showed different subgroup trade-offs in acceptance and rejection. Frozen-model evaluation on two additional libraries showed acquisition-dependent acceptance and rejection. The calibrated decision states provide an auditable basis for reviewing spectra within a declared HDPE/PP scope; overlapping polyolefin spectra and cross-library transfer remain the principal constraints.