(333a) Semi-Mechanistic Prediction of Residence Time Metrics and Mixing Dynamics in Single-Screw Extrusion Via a 2-D Convection-Diffusion Model Combined with Machine Learning

In an extrusion process, mixing dynamics is an important factor that correlates with the quality of extrudates, such as viscosity. The degree of mixing is related to the distribution of materials in the extruder, thus affecting the resultant flow pattern and system residence time distribution (RTD). Generally, RTD measurements have been used to quantify axial mixing, as it provides the mean residence time (MRT) and can be used to estimate the amount of dispersive mixing [1]. A greater MRT within the extruder barrel promotes efficient mixing, which affects the viscosity of the extrudates in the output. RTDs are influenced by process parameters (screw speed, feed rate, temperature profiles), equipment parameters (screw design), and material properties (feed viscosity) [2].

A starved-fed extruder, as compared to a flooded extruder, is highly applicable and industrially relevant as it eliminates issues related to bridging and funneling in the feed hopper or slippage on the barrels in the extruders [3]. In a starved extruder, another important metric that influences RTDs is the fill length, which represents the length of the extruder where barrels are fully filled with materials. Olofsson et al. [4] developed an advanced CFD model simulating starved extruders, which predicted residence time based on the fill length. These models were validated using experimental data of pressure drop observed at the extruder outlet, and the RTD curves were also compared to the analytical solutions. However, these models are computationally expensive, taking approximately 48 hours to obtain a single RTD curve. Therefore, there is a need to develop a computationally efficient model that can be used to predict RTDs and mixing dynamics

For this work, a hybrid model is developed, whereby a ML component is integrated with a 2-D convection-diffusion model, to accurately predict RTDs and mixing dynamics. Here, mixing dynamics are represented by the relative standard deviation (RSD) across the cross-section of the extruder. A ML-based model will be trained as a surrogate to replicate the CFD simulations, providing information on velocity fluxes across discretized radial and axial grids along the extruder, which are required inputs to the convection-diffusion model. CFD simulated data corresponding to a 3x3x3 factorial design DOE with independent parameters, including screw speed, feed rate, and feed viscosity, will be generated to train and validate a model that can account for these variations as model inputs. The dispersion coefficient will act as a tuning parameter in the model and will be estimated through model calibration minimizing errors between the predicted and CFD simulated RTD and RSD results. Finally, the model validation will be carried out using the estimated dispersion coefficient to test the prediction accuracy of the model outputs. The proposed semi-empirical model, involving data-driven predictions of velocity fluxes trained using data from first principles CFD models, coupled with a 2-D convection-diffusion model, will be able to predict RTDs and axial RSDs as a function of process parameters and feed properties.

1. Kumar, A., Experimental and model-based analysis of twin-screw wet granulation in pharmaceutical processes. 2015, Ghent University.
2. Yeh, A.-I. and Y.-M. Jaw, Modeling residence time distributions for single screw extrusion process. Journal of Food Engineering, 1998. 35(2): p. 211-232.
3. Giles Jr, H.F., E.M. Mount III, and J.R. Wagner Jr, Extrusion: the definitive processing guide and handbook. 2004: William Andrew.
4. Olofsson, E.H., Roland, M., Spangenberg, J., Jokil, N.H., Hattel, J.H., A CFD-model with free surface tracking: Predicting fill level and residence time in a starve-fed single screw extruder. The International Journal of Advanced Manufacturing Technology.