Deep-learning based real-time KSTAR fusion neutron energy spectrum unfolding with TRIASSIC and MCNP simulation

This paper describes deep-learning based real-time neutron energy spectrum unfolding method that handles fusion neutrons from KSTAR tokamak entering stilbene scintillator. Neutron energy spectrum unfolding with organic scintillators such as stilbene requires knowledge of the neutrons response function which correlates the neutrons energy with pulse heights. Despite the presence of likelihood maximization-based neutron unfolding techniques, necessity of deep-learning appears in that calculation time is reduced by time complexity order which leads to the feasibility of real-time neutron energy spectrum unfolding. Convolution-based deep learning model structure was constructed with pairs of fusion neutron energy distribution of KSTAR plasma and neutron light output energy spectrum acquired from end-to-end simulation: from generation of fusion neutron in KSTAR tokamak to stilbene scintillator-based detection system. Through TRIASSIC (Tokamak reactor integrated automated suite for simulation and computation), calculation of temporal fusion plasma power which corresponds to the temporal probability of neutron emission was executed. MCNP (Monte-Carlo N-particle transport) code simulates the transport of fusion neutrons entering the stilbene scintillator. Stilbene scintillator is widely utilized for fast-neutron spectroscopy in many applications due to its decent property for real-time unfolding. Through real-time neutron energy spectrum unfolding, immediate recovery feedback of unstabilized plasma would be possible.