Model selection for the optimization of the sensory attributes of mayonnaise

A novel optimization framework is presented which aims at optmizing the sensory properties of consumer products by changing the product formulation and the processing conditions. A crucial first step is to find a suitable model which allows predictions of the scores on certain sensory product attributes. Extensive panel testing resulted in the sensory profile of mayonnaise. The dataset was based on the average scores of various sensory attributes. Two different tasting panels were employed. The physicochemical properties of all the mayonnaise samples were also characterized. The physicochemical properties consisted of rheological variables, flavor chemical concentrations and properties related to the microstructure of mayonnaise. A statistical correlation was obtained between the physicochemical properties (independent) and the sensory attributes (dependent). The following multivariate models were tested: linear, second order polynomial and single layer artificial neural networks with 1, 2, 4 and 8 neurons. These models were fitted using one dataset containing the sensory data of 122 different types of mayonnaise obtained by one tasting panel. Then the models were validated using the sensory scores from another panel which assessed 40 types of mayonnaise. The artificial neural network with two neurons performed the best on the validation test, with less than half the number of free parameters and 3 % more accuracy this model is preferred over multiple linear regression. The Akaike Information Criterion gave indications which model would perform the best on the validation data.