Bayesian Neural Networks for the Neutron Spectrum Unfolding in the EAST Tokamak

Neutron spectra, as the key parameters of neutron fields around tokamaks, play a central role in evaluating the radiation harm to the staff and the ambient diagnostics for the normal operation of a tokamak. Several unfolding methods have been proposed for deriving neutron spectra from Bonner sphere spectrometer (BSS) counts. However, the main challenge in addressing this problem lies in the extreme sensitivity of the final solution to measurement data. The Bayesian neural network (BNN), which integrates knowledge-driven and data-driven merits, is suggested here to unfold the neutron spectra with a wide energy range and complex structure in the tokamak environment. In this study, a Monte Carlo model of Experimental Advanced Superconducting Tokamak (EAST) is constructed to generate the neutron spectra. The corresponding counts of BSS are also obtained using the experiment-evaluated response functions. These neutron spectra and BSS counts were used to train the BNN which could well predict the neutron spectra and their derived quantities, even when uncertainties of the BSS response functions are considered in the unfolding process. The comparison reveals that the BNNs outperform the other existing methods in terms of spectrum accuracy and robustness. Furthermore, verified by the BSS counts from two exposed positions collected during the EAST operations, this approach significantly enhances our ability to measure neutron energy spectra accurately.