Melt viscosity control in polymer extrusion using nonlinear model predictive control with neural state space modelling and soft sensor feedback

Melt viscosity is a key quality indicator in polymer extrusion processes, as it directly influences the mechanical properties, dimensional stability, and surface finish of the final product. However, real-time melt viscosity monitoring and control remain a major challenge in industrial polymer extrusion, due to the limitations of physical viscosity monitoring techniques, such as disturbances to the melt flow, reduced throughputs, and measurement delays. To address this issue, this study proposes a nonlinear model predictive control framework that enables direct, real-time control of melt viscosity using non-invasive feedback from a deep neural network-based soft sensor. A neural state-space model is trained on real experimental data to learn the underlying process dynamics and serves as the internal model of the controller. The soft sensor provides melt viscosity estimates based on readily available process variables (i.e., screw speed and barrel temperatures). These estimates are used by an extended Kalman filter with state augmentation to correct the internal state predictions. The proposed control system was rigorously evaluated via simulation across a variety of setpoint changes and disturbance scenarios. Results show that the controller maintains the melt viscosity within ± 2 % of the setpoint, irrespective of the initial conditions used, with settling times below 20 s. Under step and ramp disturbances applied to the output variable, screw speed, and barrel temperatures, the controller exhibited strong disturbance rejection capabilities. Notably, under step disturbances of ± 100 Pa⋅s acting on the melt viscosity output, the controller quickly restored the viscosity to the setpoint with settling times under 18 s. The real-time closed-loop melt viscosity control framework proposed in this study should be invaluable for advancing process monitoring and control of polymer extrusion processes. • Melt viscosity is a key indicator of melt quality in polymer extrusion processes. • A nonlinear model predictive controller is proposed to control the melt viscosity. • State-space representation of the system is learnt by a neural state-space model. • A soft sensing technique is used to provide feedback to the controller. • Simulation results showed good setpoint tracking and disturbance rejection ability.