Neutron spectra reconstruction based on an artificial neural network trained with a large built dataset - Archive ouverte HAL Access content directly
Conference Papers Year : 2022

Neutron spectra reconstruction based on an artificial neural network trained with a large built dataset

(1) , (1) , (2) , (1) , (2)
Maha Bouhadida
Mariya Brovenchenko
  • Function : Author
  • PersonId : 1064119
Thibaut Vinchon
  • Function : Author
  • PersonId : 1201903
Wilfried Monange
  • Function : Author
  • PersonId : 1112824
Francois Trompier


Neutron spectrometry is of great significance in different fields as reactors design, nuclear safety and radiationprotection. However, determining neutron spectra is a heavy task due to the complexity of neutron interactions and the wide range of neutron energy. Moreover, there is no direct measurement method or neutron fluence energy distribution covering the whole range of energy of neutrons. Bonner spheres spectroscopy (BSS) and activation methods remain the most used approaches for providing accurate determination of the neutron spectrum, but, the measured data needs to be analyzed with suitable spectrum unfolding program. There are several unfolding algorithms to reconstruct neutron spectra. We can cite algorithms based on iteration, maximum entropy, genetic algorithms etc… These approaches have limitations especially the requirement of a prior spectrum. To overcome this, new unfolding methods based on artificial neural networks ANNs has become of interest and different techniques were proposed in recent years to solve the related unfolding problems. Neural network models are algorithms based on concepts derived from research on the nature of the brain. Contrary to the other unfolding methods needing prior information about spectra, the neural networks explore a training process, which rules out the requirement of pre-known data. However, the proper use of ANN requires the availability of a sufficient size training dataset and an optimization of the data processing (scaling, feature engineering, normalization…). From the recent literature, it is difficult to evaluate the pertinence and the efficiency of ANN compared to other methods of neutron unfolding literature, since the two aspects mentioned above are insufficiently addressed. In our project, we aim to evaluate properly the performance of ANN for neurons spectra unfolding and to compare with other unfolding methods. We also want to go further by combining them in order to create a robust and a reliable solution. In this paper, we present the first step towards our goals which is the dataset building technique allowing the generation of a large number of physical neutron spectra. We also detail the dataset processing and we describe our first implemented architecture for the spectra unfolding
Fichier principal
Vignette du fichier
MBouhadida_ICRS14RPSD2022_VF.pdf (337.43 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-03900950 , version 1 (10-01-2023)


Attribution - CC BY 4.0


  • HAL Id : hal-03900950 , version 1


Maha Bouhadida, Mariya Brovenchenko, Thibaut Vinchon, Wilfried Monange, Francois Trompier. Neutron spectra reconstruction based on an artificial neural network trained with a large built dataset. ICRS 14/RPSD 2022 (14th International Conference on Radiation Shielding and 21st Topical Meeting of the Radiation Protection and Shielding Division), ANS, Sep 2022, Seattle, United States. ⟨hal-03900950⟩
0 View
0 Download


Gmail Facebook Twitter LinkedIn More