Physically Informed Multi-Magnification Learning for Scalable, Non-Destructive Assessment of Polypropylene Degradation in Recycled Materials

Reliable quality control of recycled polypropylene (PP) is essential for circular plastics management, but conventional degradation tests are often destructive or unsuitable for large-scale screening. This study introduces a physically informed, multi-magnification learning framework for non-destructive and scalable assessment of PP degradation from optical micrographs. A three-channel feature representation captures baseline-referenced yellowing, defect-associated contrast, and microstructural anisotropy. To address multi-scale variability, magnification-aware image patches are integrated via a hierarchical attention-based multiple instance learning model. The model jointly predicts continuous weathering cycles and ordinal degradation stages, enabling quantitative evaluation of heterogeneous material conditions. On the validation set, the approach achieved a balanced accuracy of 0.747, a stage-decoded mean absolute error of 5.6 cycles, and a quadratic weighted kappa of 0.908, with an accuracy within ±1 stage of 0.97. Notably, the model identified 10-cycle-related image signals despite the absence of a clear increase in attenuated total reflection infrared carbonyl index under the same conditions. These results demonstrate the feasibility of AI-enabled optical micrograph analysis as a rapid and non-destructive screening approach for weathering-related surface-state assessment under the controlled post-industrial recycled (PIR) PP conditions examined in this study. The framework can therefore serve as a basis for high-throughput image-based screening of weathering-related surface states, although extension to broader recycled PP streams and practical quality attributes requires further validation.