In order to establish the variation between results in mass fractions due to software implementation, as measured by the k₀-method for INAA, the IAEA has organized a software intercomparison. A complete set of test spectra and associated information was assembled. Efficiency curves, neutron spectrum parameters, correction factors and mass fractions were calculated with the participating programs (k0-IPEN, k0-INRIM, k0-DALAT, k0-IAEA and KayWin) using identical peak areas. In this paper, we report on the observed discrepancies, causes, remedies and future software developments. The test data, as well as intermediate results and observed mass fractions of the certified reference material BCR-320R “channel sediment” are available through the IAEA on request. The variations in concentrations attributed to differences between the programs were initially found to be 5.6 and 7.9%, for certified and uncertified concentrations, respectively. After the certified concentrations had been made available to the participants and they had been allowed to improve their programs, the variations found were 2.7 and 3.4%, respectively. The main identified remaining causes of variation are differences in the procedures used for detector efficiency characterisation and neutron spectrum parameter determination.

Artificial intelligence methods such as artificial neural networks, Bayesian networks, genetic algorithms, and others, have shown great potential for application, not only as classification schemes, but also in numerical data analysis. In this work, we explore how, from a limited number of spectra (around 200), an ANN could be efficiently developed, using data augmentation techniques and optimized architecture, and used to analyse neutron activation analysis (NAA) data. The IAEA Collaborating Centre Research Institute Delft (RID), Netherlands, has collected NAA data sets consisting of one single spectrum per sample to determine one single element (selenium), with addition of a marker (caesium) for flux normalization, all irradiated and measured the exact same way and analysed with k0-based software. The problem studied is one of the simplest that can be addressed with NAA; therefore the present work is intended merely as proof of concept that ANNs can perform well in NAA data analysis of simple problems. We present the results and discuss how to extend the present work to more demanding problems in NAA.