Reactor-Grade Non-Additive Polypropylene Architecture Defines the Physicochemical Limits of Melt Flow Index Model Transferability: An Industrial Chemometric and Machine Learning Study.

Rapid estimation of the melt flow index (MFI) is essential for the timely control of industrial polypropylene polymerization because conventional plastometer measurements require offline sampling and introduce analytical delays. This work investigates the physicochemical limits of MFI model transferability using 425 reactor-grade polypropylene production runs spanning homopolymer, random copolymer, and impact copolymer architectures. Principal component analysis (PCA), partial least squares (PLS), PCR-Ridge, Random Forest, Extra Trees, and Gradient Boosting were evaluated using within-pool, repeated five-fold, and product group validation. The latent structure, predictive performance, and dominant process descriptors were strongly architecture-dependent. Hydrogen-related variables remained central to molecular weight control, whereas comonomer descriptors, catalyst and donor variables, hydrodynamic conditions, and second reactor variables gained importance as compositional and morphological complexity increased. Nonlinear ensembles improved prediction within heterogeneous pools, but product group validation still revealed transferability losses when models crossed architecture-dependent process-structure-property domains. Repeated cross-validation and preprocessing sensitivity analyses confirmed that the principal model rankings and architecture-dependent conclusions were stable under resampling and correlation filtering. These results show that the polymer architecture defines a practical applicability boundary for industrial MFI soft sensors and supports architecture-specific or architecture-routed calibration when universal models fail cross-family validation.