Speaker
Description
The different processes known in astrophysics to govern the successions of nuclear formation and decay, r-process and s-process, depend strongly on nuclear properties such as the intrinsic density of levels of a given nucleus and how they behave when emitting or absorbing particles such as neutrons, protons and/or photons. There are few experimental constraints that we can use to narrow down such properties and most of the information we know comes from experimental measurements of resonance states seen in compound nuclei formed by neutron-induced reactions. These resonances are seen as sharp increases in measured neutron transmission and capture due to the proximity of the neutron energy with excited levels in the compound nucleus.
Therefore, a proper and reliable account of all resonances is crucial for the description of nuclear reactions. We have developed a Machine-Learning method, the Bayesian Resonance Reclassifier, to make use of the statistical properties of resonances to train an algorithm capable to identify missing and misassigned resonances. We are able to train the model on synthetic data and use transfer learning to assess resonance sequences from the Atlas of Neutron Resonances, evaluated files or experimental data. We performed an extensive optimization through multiple grid search of hyper-parameter values for each of the classifiers studied. We were also able to use experimental resonance data obtained through measurements with polarized neutron beams as a validation tool to assess the quality of the method, with encouraging results.