In this work a feedforward neural network, trained and validated using experimental values of freezing and thawing times of foods and test substances of different geometries, is developed. A total of 796 experimental times of both processes were collected from reported works. The database used covered a wide range of operative conditions as well as size, shape and type of material. The input layer had seven elements: shape factor, characteristic dimension, Biot number, thermal diffusivity, initial, ambient and final temperatures. The output layer had one element: the process time. The total number of hidden layers and the number of neurons in each hidden layer were chosen by trial and error. For each topology, a simple based genetic algorithm search technique was applied to obtain the initial training parameters of the neural network that improve its generalization capacity. Three particular networks were evaluated: one for freezing times, another one for thawing times, and a third one for both freezing and thawing times. The final topologies has one or two hidden layers with 4 nodes in each one. Our results show that the neural network had an average absolute relative error of less than 10%, suggesting that ANN provide a simple and accurate prediction method for freezing and thawing times, valid for wide ranges of food types, sizes, shapes and working conditions. (C) 2007 Elsevier Ltd. All rights reserved.