Numerical methods, approximate formulas and artificial neural networks have equivalent accuracy for the prediction of food freezing and thawing times

S.M. GOÑI; L.A. CAMPAÑONE; V.O. SALVADORI; R.H. MASCHERONI

Sitges

Congreso; 31 EFFOST INTERNATIONAL CONFERENCE; 2017

EFFOST

Three different methods (numerical models, approximate formulas and Artificial Neural Networks (ANN)) are extensively tested for the prediction of freezing and thawing times of foods. The numerical model includes a detailed description of the phase changes involved in freezing or thawing, considering variable thermal properties as a function of temperature and composition. For one-dimensional geometries an implicit finite difference method was employed, while for bi- or three-dimensional geometries a finite element commercial package was used. The approximate prediction formulas were previously developed covering a wide range of food shapes and sizes and working conditions. There are specific formulas for each regular one?dimensional shape (slab, infinite cylinder and sphere), for both freezing and thawing. Their validity is extended to multidimensional shapes via the use of shape factors. ANNs were developed and trained based on our extensive database of experimental values of freezing and thawing times of foods and test substances of different geometries. It includes a total of 787 experimental times of both processes. ANNs were of the feed-forward type. The input layer had seven elements: characteristic dimension, shape factor, Biot Number, thermal diffusivity, initial, ambient and final temperatures. The output layer had one element: process time. Our results show that the three prediction methods have equivalent accuracy. So, for practical uses, ANNs or approximate formulas are recommended.