Neutron spectrum unfolding

Accurate neutron spectrum characterization is critical for reactor dosimetry, shielding, and material damage assessment. Since direct measurement is often infeasible, activation foils are used, yielding only energy-integrated reaction rates. Unfolding this discrete data into a continuous spectrum constitutes a highly underdetermined, ill-posed inverse problem. Traditional legacy codes like SAND-II address this but suffer from strong structural prior bias and lack uncertainty quantification. To address these limitations, a probabilistic Bayesian framework utilizing Gaussian Process (GP) regression is proposed for nonparametric unfolding. Evaluated with synthetic multi-foil data from MCNP-simulated reference spectra of the BR1 reactor, three GP kernels (Squared Exponential, Rational Quadratic, Gibbs) were compared against a parameterized Bayesian model and SAND-II. Results indicate that nonparametric GP models successfully capture physical features—such as epithermal 1/E down-scattering and the fast fission tail—without relying on deterministic structural guesses, thereby avoiding the restrictive assumptions of parameterized priors. The GP framework yields defensible uncertainty estimates, collapsing credible intervals at dominant resonance peaks while broadening them in data-poor regions. Conversely, SAND-II demonstrated severe vulnerability to structural artifacts when supplied with an inaccurate prior. Ultimately, this GP framework enables robust spectral recovery with rigorous, data-driven uncertainty quantification.