A data-driven approach for melt flow rate prediction using online rheology

Abstract To ensure stable process conditions and prevent part defects, most polymer processes require viscosity to fall within a narrow range. Since measuring shear viscosity is laborious, slow, and expensive, the polymer industry often uses melt flow rate as a simpler proxy. Inline and online rheometers have been developed in academia and industry to obtain real-time rheological data close to the process. However, regardless of flow geometry and installation, most systems rely on offline MFR calibration, which requires expert knowledge and frequent recalibration and is valid only for a narrow window of material and process conditions. Using a commercial online rheometer coupled with a compounder, we trained machine learning models to accurately predict MFRs measured offline from process data. The experimental dataset was constructed using various blends of virgin and post-consumer polypropylene grades with MFRs ranging from approximately 1 to 80 g 10 min $1\,\text{to}\,80\,\frac{g}{10\,\mathrm{min}}$ . The model that performed best on the test set was a feed-forward neural network with a mean absolute percentage error of ±5.38 % and a mean absolute error of 1.52 g 10 min $1.52\,\frac{g}{10\,\mathrm{min}}$ , and required no offline calibration or user intervention.