Speaker
Description
The Laboratory for micro-irradiation, neutron metrology and dosimetry (LMDN) at IRSN is responsible for characterizing various neutrons fields that necessitate the utilization of advanced unfolding techniques. Usually, the LMDN and the laboratory of ionizing Radiation dosimetry (LDRI) use well known unfolding methods among Bayesian inference and GRAVEL algorithm. However, those methods require an initial and arbitrary guess of the spectrum solution, which may significantly impact the unfolding result. To address this bottleneck, we propose with the Neutronic Laboratory (LN) at IRSN, an innovative method for neutron spectrum reconstruction using machine learning techniques trained on a large dataset of spectra built by using a dynamic genetic algorithm and the response function of the studied detector. This genetic algorithm has been developed during the project in order to build a balanced dataset in term of variety. A new convolutional neural network architecture is then employed and is specially adapted for unfolding neutron spectra. This methodology is applied to a neutron activation spectrometer comprising of foils of different materials (the SNAC) with a response function recomputed by the MCNP code using the IRDFF library. This choice was motivated by recent results measured in the AMANDE facility that were in better agreement with this library. For a set of criticity experiments performed at the Silene reactor in Valduc, the different neutron spectra are inferred from the reaction rates measured by the activation of the foils composing the SNAC. These experimental neutron spectra are compared with those resulting from a Bayesian approach or from Monte Carlo simulations of the Silene platform as well as from other reference measurements performed on this platform in the same conditions in terms of screening of the reactor, position of the SNAC and mode of the criticity accident.
The transposability of this approach to other neutron activation detectors needing unfolding is then discussed. We will present at ANIMMA 2025 the methodology, the preliminary results obtained for activation detector measurements performed on the Silene installation in different configuration. A large discussion about generalization of that method will then be presented.
