Development of a machine learning model for neutron spectrum unfolding

Neutron spectrometry presents a significant challenge due to the ill posed nature of the unfolding problem, where the neutron energy spectrum cannot be directly extracted from measurements. Traditional unfolding methods, such as Monte Carlo simulations, iterative techniques, Bayesian inference, and the maximum entropy principle, often suffer from substantial limitations, such as high computational cost, sensitivity to noise, numerical instability, difficulty handling incomplete or uncertain data, and a strong dependence on prior assumptions. These constraints have driven a growing interest in alternative solutions. Innovative approaches, based on machine learning (ML), have emerged as a promising substitute to traditional calculations and conventional techniques. In this study, a ML model was developed and trained using database from the International Atomic Energy Agency neutron spectra compendium to perform spectral unfolding from Bonner sphere spectrometer counts rates. Simulation results reveal that the ML based unfolding approach achieves high accuracy and strong generalization capabilities, with reconstructed spectra closely matching reference benchmarks. These findings highlight the potential of ML as a robust and efficient alternative to traditional neutron spectrum unfolding techniques.