Quantifying techno-functional properties of ingredients from multiple crops using machine learning

Abstract

Food ingredients with a low degree of refining consist of multiple components. Therefore, it is essential to formulate food products based on techno-functional properties rather than composition. We assessed the potential of quantifying techno-functional properties of ingredient blends from multiple crops as opposed to single crops. The properties quantified were gelation, viscosity, emulsion stability, and foaming capacity of ingredients from yellow pea and lupine seeds. The relationships were quantified using spline regression, random forest, and neural networks. Suitable models were picked based on model accuracy and physical feasibility of model predictions. A single model to quantify the properties of both crops could be created for each techno-functional property, albeit with a trade-off of higher prediction errors as compared to models based on individual crops. A reflection on the number of observations in each dataset showed that they could be reduced for some properties.