REPRESENTING NUCLEAR CRITICALITY EXCURSION EXPERIMENT DATA BY AN ARTIFICIAL NEURAL NETWORK

A feedforward artificial neural network (ANN) is constructed using select nuclear criticality excursion experiment data sets from the French Consequences Radiologiques d'un Accident de Criticite (CRAC) and SILENE reactor campaigns. The ability to represent initial spike characteristics by an ANN provides a new method that is aligned to excursion data more directly and to a wider variable data set than traditional analytic approaches. The ANN is configured, trained, validated, and tested to 85 unique highly enriched uranium (HEU) excursion experiments, considering six input variables and two output variables (specific power and energy). The fidelity of the ANN is enhanced by normalizing the input and output data. The trained ANN is then used to determine output values for 19 select Kinetic Energy Water Boiler experiments and 14 additional CRAC excursions not used in the ANN construction. Furthermore, the same trained ANN is also used for an extensive comparison (80 cases) for a combination of uranium concentrations, concentrations, ramp feed reactivity insertion rates, system volumes, and vertical container sizes. The specific spike energy and power ranges determined are bracketed by published experiment results and are more realistically represented than results derived from well-known analytical methods. The ability to predict initial peak fissions by an ANN does not require determining, a priori, a volume-dependent energy quench parameter ("b") specific to HEU solutions. The results derived from the ANN can aid in designing realistic emergency planning constructs or criticality accident alarm system hardware placements without undue penalty for fission source term uncertainties. Neither excursion characteristics after the initial spike nor explicit time dependencies are modeled by an ANN at this time. The extension of the methods presented is left for further work.