Comparative assessment of feature extraction for fast neutron spectra prediction with machine learning algorithms using a CVD diamond detector

Machine learning algorithms are recognized as effective tools for addressing challenges in nuclear methodologies and instrumentation, particularly when obtaining experimental data or precise mathematical models is difficult. One notable challenge in fast neutron spectroscopy involves unfolding neutron spectra from solid-state detectors, where conventional deconvolution techniques often suffer from instability. This study uses the PHITS code to simulate the behavior of a CVD diamond detector subjected to various neutron spectra. The simulated data serve as input to a regression model aimed at reconstructing the original spectra. In particular, we investigate the impact of four different signal processing and feature extraction techniques—Fourier transform, nodal basis projection, wavelet decomposition and agglomerative clustering-based compression—on the performance of spectrum reconstruction. These transformations are applied as preprocessing steps, and the resulting features are then used to train a supervised regression model (linear or kernel ridge, depending on the method). Model performance is evaluated using the R 2 metric against reference spectra from the IAEA Compendium database. The most effective approach is then applied to experimental data collected with a CVD diamond detector at the DANAIDES irradiation facility, part of the TOTEM platform in Cadarache. This final test, performed on 14 MeV and AmBe neutron sources, demonstrates the practical feasibility of the proposed methodology for real-world neutron spectrum reconstruction.