Predicting techno-functional properties of food hydrocolloids, such as emulsification, foaming and gelling, remains a major challenge. While powerful models exist, they are only as good as the data they are trained on. Too often, data is of low quality and quantity, due to non-standardised, fragmented measurement methods and closed databases. These data-related challenges limit our ability to build accurate predictive tools for hydrocolloid functionality.

Therefore, we propose an approach that will advance techno-functionality prediction, which requires both better quality data and smarter models. On the data side, globally harmonised protocols and systematic reporting of data is essential to ensure comparability and reproducibility across laboratories in the whole food science community. We will describe our recently started COST Action INFOTECH-DATA, which aims to connect the food colloids community by defining shared methodologies and build open-acceess databases.

On the modelling side, we aim to build models that can handle data scarcity. We present a hybrid modelling approach, which combines physics and machine learning, in so-called physics-encoded neural networks (PeNN) . We will showcase this approach for oil-in-water emulsions by predicting viscosity based on oil volume fraction and the type of protein used. We encode a neural network with a Quemada model, and compared this PeNN model with a standard neural network. The PeNN gave mean squared errors that was always (in a order of a factor thousand) smaller than a standard neural network. This means that we need a factor of 5-10 times less data to predict emulsion viscosity when encoding physics into a neural network. Encoding physics also adds causality to the model, reducing the black-box-nature of neural networks.

Together, better data and smarter model pave the way for predictive techno-functionality. This dual strategy will allow us to move into new scientific directions, and at the same time accelerate sustainable food structure design by enabling more rational ingredient & formulation choices.