Signatures from the spent fuel: simulations and interpretation of the data with neural network analysis

In the last years, the safeguards verification of spent fuel assemblies by NDA has received increased interest also due to upcoming programmes for the geological disposal. During safeguards inspections one aims at verifying the completeness and correctness of operator declared data. One should then be able to draw conclusions on the fuel integrity and diversion of pins, as well as checking the consistency of operator declarations on initial enrichment, fuel type, burnup and cooling time. The verification of spent fuel is also important for safety aspects related to the storage of spent fuel. The experimental observables associated to NDA of spent fuel assemblies are often a complex function of the characteristics of the fuel, its irradiation history and other variables related to the used measurement setup and devices; nowadays one often assumes that some of the variables are known to interpret the data and draw conclusions. To facilitate the interpretation of the data and draw more robust safeguards conclusions, an R&D effort is on-going at SCK•CEN and its results are given in this paper. This work reports first about the efforts done at SCK•CENon simulating detector response functions for different types of NDA instruments such as the Fork detector, the Forkball detector and SINRD detectors. These responses are obtained from Monte Carlo model of the fuel and measurement setup. The spent fuel composition and radiation characteristics are taken from a spent fuel reference library developed in recent years. A database of detector responses corresponding to 8400 cases with different fuel characteristics and irradiation parameters was then obtained. We explore the use of these simulated observables as input for data analysis algorithms aimed at uniquely characterizing the spent fuel and drawing safeguards conclusions. More specifically, we focus on the application of artificial neural networks due to their ability to generalize non-linear relationships. As a first step, cooling times smaller than 100 years were selected from the database, and several network configurations and training schemes were investigated.