Machine learning to quantify techno-functional properties - A case study for gel stiffness with pea ingredients

Abstract

Mildly refined ingredients are included more easily in food products when selected based on techno-functional properties instead of composition. We assess different machine learning methods that quantitatively link relevant techno-functional properties to the composition and processing history of the ingredient in a case study using the gel stiffness (Young's modulus) by conventionally and mildly refined ingredients of yellow pea. Linear (multiple, log transformed and polynomial) and non-linear models (spline regression, decision trees, and neural networks) were explored. The final model selection was based on 1) the statistical model metrics (RMSE, R2, and MAE) of the training and independent test set and 2) expert knowledge to evaluate the plausibility of the model predictions. In this case, neural networks can describe the gel stiffness of yellow pea ingredients most accurately. The approach that we follow can be applied to other techno-functional properties to improve the chain sustainability while ensuring the full functionality of the products.