Semi-Mechanistic Prediction and Optimization of Residence Time Metrics of a Starve-Fed Extruder via a Hybrid Machine-Learning Convection-Diffusion Model

This study addresses the challenge of determining mixing dynamics, such as the residence time distribution (RTD) and relative standard deviation (RSD) of a system, which usually require tedious experimental setups or computationally intensive models. A reduced order model (ROM) in the form of a machine-learning-based convection-diffusion model was developed for the computationally efficient prediction of mixing metrics. A two-dimensional compartmental convection-diffusion model was applied to a starve-fed single screw extruder, capturing non-uniform axial velocity fluxes. The model was enhanced through machine learning algorithms to establish correlations between process parameters, velocity fluxes, and diffusion coefficients. The use of a convection-diffusion model provided a mechanistic prediction of RTD and RSD. The model accuracy was demonstrated via an R-squared value of above 0.9 for the prediction of different metrics evaluated, such as mean residence time and variance. The calibrated model was then optimized to identify improved input conditions that led to the desired mixing metrics. The developed ROM provided an efficient alternative to the full computational fluid dynamics (CFD) model by substantially reducing the computational time from 44 h to 2.5 s while retaining model accuracy. Source codes for the model are available on GitHub repositories at https://github.com/adan626/ROM_CFD and https://github.com/patilurjit/CFD-ML.