CFD-based and machine learning-assisted optimization of screw geometry in food extrusion processes

This study presents an integrated computational framework for optimizing screw geometry in a compact food extruder by coupling computational fluid dynamics (CFD) simulations with machine learning (ML) and multi-objective optimization. Three-dimensional CFD simulations are performed using COMSOL Multiphysics to capture the complex thermo-fluid behavior within the extruder and to quantify key process responses. A Design of Experiments (DOE) based on Latin Hypercube Sampling (LHS) is employed to efficiently explore the high-dimensional design space of screw geometric parameters. Based on the CFD-generated dataset, a Random Forest (RF) regression model is developed as a surrogate to approximate the nonlinear relationships between screw geometry and performance indicators, achieving average relative prediction errors of approximately 8% for the axial surface force ( F S x ) and 5% for the specific mechanical energy ( SME ). Multi-objective optimization is subsequently conducted using three evolutionary algorithms, namely the Non-dominated Sorting Genetic Algorithm II (NSGA-II), the Non-dominated Sorting Genetic Algorithm III (NSGA-III), and the Multiobjective Evolutionary Algorithm based on Decomposition (MOEA/D), to identify optimal trade-offs between minimizing F S x and SME . The resulting Pareto fronts obtained from the three algorithms are systematically compared to evaluate their convergence and solution diversity. The Pareto-optimal solutions highlight the competing effects of screw geometry on process performance and provide quantitative guidance for screw design improvement. Overall, the proposed CFD-driven, surrogate-assisted multi-objective optimization framework shows strong potential for smart manufacturing and intelligent design of advanced food extrusion systems.