ABSTRACT The melt flow index (MFI) is a fundamental indicator of polymer processability, directly related to molecular weight and melt viscosity, particularly during industrial granulation with twin‐screw extruders. Conventional laboratory measurement of MFI is offline, time‐consuming, and unsuitable for real‐time quality control. To address this limitation, a machine‐learning–based soft sensor was developed to predict MFI in‐line using multivariate process‐variable data from an industrial extrusion–granulation system. Key process parameters were identified based on field knowledge, followed by data preprocessing and feature engineering. Minority transition regions were augmented using the k‐Nearest Neighbour Synthetic Minority Oversampling Technique (KNN‐SMOTE), thereby improving model robustness in high‐deviation regimes and across grade transition boundaries. Also, eXtreme Gradient Boosting (XGBoost) models, which employ an ensemble gradient‐boosting framework, were evaluated via hyperparameter optimization. The results indicated that augmentation with smaller k values and lower thresholds for detecting high‐deviation responses achieved superior performance, with an R 2 of 0.99 and an RMSE of 0.59. In addition, SHapley Additive exPlanations (SHAP) analysis confirmed model interpretability by identifying the dominant process relevant features and their consistent directional influence on MFI predictions. Overall, the proposed framework enables real‐time monitoring of MFI in polyethylene granulation, reducing the frequency of laboratory testing and improving product quality control.
Jani et al. (Wed,) studied this question.