A high-dimensional Bayesian approach for spectrum unfolding and uncertainty quantification in spectrometric measurements

Neutron spectrum unfolding and uncertainty quantification face inherent challenges due to high dimensionality and ill-posed characteristics. We propose a high-dimensional Bayesian approach leveraging local probability decomposition, which comprises a three-phase framework: (1) generating global initial solutions using the GRAVEL algorithm to constrain parameter spaces; (2) conducting localized probability analysis for targeted energy groups via dynamic sliding windows to construct marginal distributions; (3) performing Markov Chain Monte Carlo (MCMC) sampling with optimized initial points derived from marginal distributions. This innovation significantly enhances the performance of Bayesian spectral unfolding methods. Validated against IAEA-403 Cf-source and ²⁴¹Am-Be experimental measurements, the method achieves precise spectral reconstruction across 13–53 energy groups, demonstrating <15 % spectral relative deviation from the ground truth and 92.3 % uncertainty coverage probability for true values. It outperforms non-informative Bayesian, GRAVEL-informed Bayesian, and conventional GRAVEL methods by reducing relative deviation by up to 54 % and improving coverage probability by 34 percentage points. This study establishes an efficient approach for high-dimensional spectrum unfolding and uncertainty analysis, providing a critical tool for precise radiation dose assessment, shielding optimization, and safety assurance in next-generation advanced nuclear reactors. © 2001 Elsevier Science. All rights reserved.