Testing population initialisation schemes for differential evolution applied to a nuclear reactor core design

Abstract

In the process of nuclear core design, reactor cell parameters such as dimensions, enrichment and materials must be adjusted considering restrictions such as the average thermal flux, criticality and sub-moderation. This problem may be formulated using global optimisation methods in order to generate sets of design parameters to be analysed by programs that simulate the neutron interactions in the reactor core. This problem is highly multimodal, requiring techniques that overcome local optima, which can be done by promoting a greater populational diversity. An approach that has been overlooked is the selection of an initial set of solutions of the populational algorithms. In this work, we use the differential evolution algorithm to test two different generation schemes besides pseudorandom generation: pseudorandom generation followed by the application of opposition-based learning and the Sobol quasi-random generator. The results show the potential of each scheme for application in other nuclear science and engineering problems.