Nowadays, using artificial neural networks (ANNs) to perform interesting input/output mappings in various industrial contexts has become almost routine. Indeed, the nonlinear features of this algorithm allow one to deal with real complex systems such as those encountered in the nuclear field.

Here, an ANN algorithm is applied to determine the relationships that exist between some process variables pertaining to the operation of the steam generator of a pressurized water reactor. The exemplars required for the ANN training are obtained from a suitable nonlinear, mathematical model, numerically integrated, whose solution yields pseudo-experimental data that simulate data that would be collected in a real experiment. In the training phase, Ishikawa structural learning that aims at eliminating the unnecessary network connections is performed. After completion of training, without the analyst's intervention, the resulting ANN topology consists of the superposition of three distinct and smaller ANNs. This implies that the network, on the basis of the given exemplars only, without knowledge of the physical laws, is able by itself to decide that the relevant input/output variables could be partitioned in independent groups. The ANNs so identified turn out to be so simple that their mappings could be easily translated into empirical algebraic correlations. Numerical tests validate the correlations thereby obtained.