Plant-based protein extrusion optimization: Comparison between machine learning and conventional experimental design

High-moisture extrusion (HME) is a promising technique for developing fibrous plant-based meat analogues. In HME, protein-water formulations are passed through a heated twin-screw barrel before solidifying in a cooling die, where complex physicochemical transformations occur, making process optimization challenging. Traditional approaches like Response Surface Methodology (RSM) require extensive trials and rely on predefined polynomial models, limiting predictive power. In contrast, Bayesian Optimization (BO), a machine learning technique, uses probabilistic surrogate models to efficiently explore parameter spaces and optimize black-box functions with fewer experiments. This study compares RSM and BO for optimizing the mechanical properties of twin-screw extruded meat analogues to replicate chicken breast by varying barrel temperature, water content, and cooling die temperature. To facilitate a direct comparison, BO was constrained to explore within the dataset employed by RSM, although this restriction may limit BO's full optimization potential. Tensile strength was identified as a key property that improved model fitting and predictive accuracy for both RSM and BO. Compared to the 15 experimental trials required by the RSM approach, BO converged on an optimal set of parameters using only 11 of the 15 RSM trials without tensile strength. When tensile strength was included, the output of only 10 trials was needed before convergence was observed. Experimental validation showed BO predictions had lower a prediction error (≤24.5 %) compared to RSM (up to 61.0 %). These findings highlight the potential of superior predictive accuracy and efficiency in optimizing complex pilot-scale food processing systems such as HME through BO.