Reclassifying neutron resonance spins with Machine Learning

Neutron resonances are sharp fluctuations seen in neutron transmission and capture experiments at low-energy neutron-induced reactions. Properties of neutron resonances are some of the few experimental constraints to nuclear level densities and gamma strength functions (crucial for modeling many nuclear applications). Resonances are characterized by their angular momenta quantum numbers, which are normally assigned through fits often done in a not fully reproducible manner. Comprehensive compilations of evaluated resonances often contain incorrectly assigned spins. To address these issues, we developed and successfully applied a Machine-Learning method [1] to train a multi-label classifier to identify resonances with incorrect spin assignments. Model training can be done either on synthetic data built to simulate statistical properties of resonances seen in real nuclei, or on ranges of real experimental data known to have reliable assignments . The trained classifier can be applied to resonances sequences from compiled, evaluated, or experimental data. We will show results on 52Cr and 238U using synthetic and/or real data to train, cross-validate and predict spin assignments on evaluated data. We will also discuss future developments.

[1] G.P.A. Nobre et al., Physical Review C 107, 034612 (2023)