Temporally Validated and Laboratory-Reconciled Soft Sensing of Xylene-Soluble Fraction in Industrial Polypropylene Manufacturing

The xylene-soluble fraction (XS) is a critical quality attribute in polypropylene manufacturing because it is associated with polymer architecture, catalyst stereoselectivity, comonomer incorporation, and polymerization conditions. Its conventional determination by ASTM D5492 is laboratory-based and cannot provide continuous information on the evolving quality state of an industrial reactor. This study addresses how much XS predictability originates from the previously known analytical state and how much is attributable to the contemporaneous reactor state. Of the 425 original industrial records included, one physically invalid feed condition was excluded, leaving 425 curated observations. These observations were evaluated chronologically rather than via random row-wise partitioning, and product code and downstream quality descriptors that were unavailable at prediction time were excluded from the primary formulation. The process-only model achieved a final temporal-test R2 of 0.666, whereas historical analytical/recipe information alone yielded R2 = 0.843. Integration of both information sources increased performance to R2 = 0.902, RMSE = 0.640 wt.%, and MAE = 0.306 wt.%. Expanding-window temporal validation yielded a mean R2 of 0.875 ± 0.115. The strongest independently supported domain was Family H, for which 50 future observations were predicted with R2 = 0.930, RMSE = 0.304 wt.%, and MAE = 0.188 wt.%. The results support a laboratory-reconciled soft-sensor architecture in which ASTM D5492 serves as the certification reference, while the model provides intermediate estimates between laboratory measurements.