A Comparative Study of Kernel Ridge Regression and Gaussian Process Regression for Neutron Spectrum Reconstruction with CVD Diamond Sensor

Accurate reconstruction of the incident neutron energy spectrum is essential in various fields, including radiation protection, nuclear reactor physics, and criticality safety. Traditional neutron unfolding techniques, particularly those based on Bayesian inference, often require an a priori estimate of the spectrum, which can heavily influence the final result.In this study, we introduce an innovative approach leveraging machine learning methodologies to perform neutron spectrum unfolding based on the energy deposition signals measured by chemical vapor deposition diamond detector. The theoretical study is conducted using the simulated deposited energy in the diamond sensor modeled with the PHITS code. Specifically, we compare the predictive capabilities of two machine learning models: kernel ridge regression model and Gaussian process regression model.We assess their performance using a combined score that integrates both the coefficient of determination and cosine similarity, applied to a set of benchmark spectra from the International Atomic Energy Agency Compendium Database. Our results demonstrate promising accuracy across both models. Furthermore, the models were employed to predict the neutron spectrum 252Cf neutron source using experimental data collected at the TOTEM facility at CEA Cadarache, yielding encouraging agreement with the reference spectra.