Predicting rheological parameters of food biopolymer mixtures using machine learning

Predicting the properties of foods prepared with plant protein ingredients through hydrothermal processing remains challenging. This study uses compositional data to predict rheological properties of plant-based biopolymer mixes using machine learning. Samples containing protein concentrations ranging from 14 to 43 % were prepared using a range of formulations, based on yellow pea and faba bean protein ingredients. The formulations were mixed with 0–13 % polysaccharides, namely maize starch, pectin, cellulose and carrageenan, to a final moisture ranging between 40 and 72 %. These mixtures were relevant for high moisture extrusion processing. Rheological data were collected in a closed cavity rheometer, applying small, medium, and large amplitude oscillatory shear. Data from 140 unique formulations were subjected to cluster analysis to identify patterns in the dataset and variable importance analysis to identify key input features and relevant output rheological parameters. Following, multiple supervised machine learning regression models were evaluated, with single-output Random Forest regression effectively predicting parameters in the linear viscoelastic regime, from compositional inputs, but not parameters in the non-linear regime. Accurate predictions of parameters in the non-linear regime could be obtained using multi-output Random Forest regression, with large deformation parameters as input. The results highlighted the interdependencies existing among rheological parameters, and clearly brought evidence of the strength of using machine learning as a tool to predict the rheological properties of plant-based biopolymer mixes, and to highlight trends in the data which may lead to an increased mechanistic understanding of the effect of composition on the structure formation during high moisture extrusion. • Closed cavity rheometer measures viscoelastic properties of biopolymer mixes. • Cluster and variable analysis identify key input and output parameters. • Machine learning predicts rheological parameters in the linear viscoelastic range. • Random Forest Regression models were best suited for predictions. • Machine learning models fail to predict LAOS parameters from composition data.