Machine Learning-Based Unfolding of Neutron Spectra in the BR1 Research Reactor Cavity Using 4H-SiC Solid State Detectors

This work explores a machine-learning-based methodology for neutron spectrum unfolding using 4H-SiC detectors, which are suitable for diagnostics in harsh reactor environments such as ITER and JET. Within the framework of the VALUE project, a collaborative effort between CEA and SCK CEN aimed at the characterization of the neutron spectrum in the VENUS-F zero power fast reactor, an experimental campaign was conducted in the BR1 reactor cavity using two reference spectrum converters (MARK III and Uranium Shells). 4H-SiC detectors of different thicknesses (20 μm, 60 μm, and 100 μm) were employed in the measurement; the energy deposited in each sensor was extracted and subsequently used as input for an unfolding algorithm based on Gaussian Process Regression. This methodology adopts a flexible probabilistic approach to spectrum reconstruction. The article details the experimental setup, data acquisition and processing steps, and the adopted learning strategy. The results demonstrate the feasibility of leveraging machine learning techniques for neutron spectrum unfolding in fast reactor applications and provide a proof of concept for normalized neutron spectrum reconstruction, i.e., the spectral shape independent of absolute intensity, paving the way toward future applications in fission and fusion facilities.