Evaluation of Genetic Programming and Neural Networks Techniques for Nuclear Material Identification

Nuclear materials safeguard efforts necessitate the use of non-destructive methods to determine the attributes of fissile samples enclosed in special, non-accessible containers. To this end, a large variety of methods has been developed. Usually, a given set of statistics of the stochastic neutron-photon coupled field, such as sourcedetector, detector-detector cross correlation functions, and multiplicities are measured over a range of known samples to develop calibration algorithms. In this manner, the attributes of unknown samples can be inferred by the use of the calibration results. The sample identification problem, in its most general setting, is then to determine the relationship between the observed features of the measurement and the sample attributes and to combine them for the construction of an optimal identification algorithm. The goal of this paper is to compare a combination of genetic algorithms and neural networks (NN) with genetic programming (GP) for this purpose. To this end, the time-dependent MCNP-DSP Monte Carlo code has been used to simulate the neutron-photon interrogation of sets of uranium metal samples by a 252Cf-source. The resulting sets of source-detector correlation functions, R12(• ) as a function of the time delay, • , served as a data-base for the training and testing of the algorithms.