Sensitivity study on determining an efficient set of fuel assembly parameters in training data for designing of neural networks in hybrid genetic algorithms

Abstract

Neural Networks (NNs) were applied as a tool for simulating several nuclear reactor physics parameters during core depletion calculations. The main objective was to develop NNs models capable of simulating useful reactor physics parameters to filter out the bad designs created in Genetic Algorithms (GAs) run without the need to perform reactor physics calculations for evaluation of individuals. Applying GAs to optimize both the nuclear reactor Loading Pattern (LP) and Burnable Poison (BP) designs for their respective performance characteristics creates many unwanted results along the way. New population individuals are normally analyzed with a reactor physics code to determine its fitness or applicability for future use. Significant time was required for each reactor physics code calculation and because most of the solution individuals created by GAs result in unusable designs, analyzing every solutions involves prohibitive computational times. Such long computational times can be greatly reduced by applying NNs to filter out most of the unwanted designs. A detailed description of the selection process of the NN architecture, training method, and adequate ranges of data are also presented. Finally, a hybrid GA algorithm is proposed in which two NNs are used to discard most of the worse LP and BP designs.