Neural networks have known an explosive development, due to their abil
ity to cope with a vast amount of data, to learn them and to extract,
at request, the desired information. Also, the neural networks are cap
able of generalization that means giving a correct answer to a questio
n outside the learning set. However, the generalization capacity dimin
ishes if the question is off the range of the input variables of the d
ata set. With this drawback kept in mind, the generalization capacity
is used with the forward mutilayer net to design a fin heat exchanger.
The training data set was obtained with a previously developed mathem
atical model for a fin Cube heat exchanger, which is self-adaptive in
respect of the topology of the exchanger.(1) As it was shown in the ci
ted paper, for a given thermal charge there are several operating cond
itions and exchanger topologies which permit its accomplishment. Thus,
the neural network must be able to classify all these mathematical so
lutions. To do this, tile training data set is split into two parts: 3
/4 of all data were used in the learning Phase, while the rest was use
d. as a stopping criterion.(2) The learning strategy was tile back-pro
pagation algorithm.(3) The accuracy of the answer was up to 80% - 90%
for the test set, which encourages the authors to believe that the neu
ral network is a reliable designing tool for the fin heat exchanger.