Real-time organic scintillator neutron spectrum unfolding using a deep learning approach and data generated from TRIASSIC and MCNP

Under harsh operating conditions of a demonstration fusion power plant (DEMO) with high neutron fluxes, neutron diagnostics can provide a map of core plasma conditions. In this study, we present FISTA-Net, a novel real-time neutron spectrum unfolding methodologies for organic scintillator that discriminates unscattered neutrons—a methodology for space-resolved neutron diagnostics. Previous inverse problem solving methodologies like Maximum Likelihood Estimation Method (MLEM) are limited in either inapplicability in real-time or estimation accuracy in the presence of noise. Dataset utilized for FISTA-Net’s training was generated by integrating Tokamak Reactor Integrated Automated Suite for Simulation and Computation (TRIASSIC) and Monte-Carlo N-particle transport code (MCNP) simulation that implements KSTAR fusion neutrons. MCNP simulation for light output spectrum was validated with Deuteron-Deuteron (D-D) fusion neutron generator experiment. FISTA-Net architecture integrating Fast Iterative Shrinkage Thresholding Algorithm (FISTA) and convolutional neural network modules was developed to solve the inverse problem of spectrum unfolding. Even with noise, FISTA-Net achieved a mean relative error of 3.72% at 2.45 MeV peak, completing unfolding in 6.19 milliseconds, whereas naive FISTA showed a mean relative error of 18.8% and MLEM failed to converge. This achievement marks a significant step toward integrating deep learning-based neutron diagnostics into future DEMO operation.