Confidence estimation in the prediction of epithermal neutron resonance self-shielding factors in irradiation samples using an ensemble neural network
Ian M. Wilkinson,
Ritwik R. Bhattacharjee,
Jenifer C. Shafer,
Andrew G. Osborne
Affiliations
Ian M. Wilkinson
Nuclear Science & Engineering, The Colorado School of Mines, 1012 14th St., Golden, CO, 80401, USA; Department of Chemistry, The Colorado School of Mines, 1012 14th St., Golden, CO, 80401, USA
Ritwik R. Bhattacharjee
Nuclear Science & Engineering, The Colorado School of Mines, 1012 14th St., Golden, CO, 80401, USA; Department of Mechanical Engineering, The Colorado School of Mines, 1610 Illinois St., Golden, CO, 80401, USA
Jenifer C. Shafer
Nuclear Science & Engineering, The Colorado School of Mines, 1012 14th St., Golden, CO, 80401, USA; Department of Chemistry, The Colorado School of Mines, 1012 14th St., Golden, CO, 80401, USA
Andrew G. Osborne
Nuclear Science & Engineering, The Colorado School of Mines, 1012 14th St., Golden, CO, 80401, USA; Department of Mechanical Engineering, The Colorado School of Mines, 1610 Illinois St., Golden, CO, 80401, USA; Corresponding author.
Large neutron absorption resonances in the nuclides present in irradiation samples reduce the irradiating neutron flux at energies close to a resonance. In neutron activation analysis of optically thick samples with resonant isotopes, this self-shielding effect can be significant, and must be accounted for to ensure accurate measurements. Here we show that an ensemble artificial neural network can be used to accurately predict the epithermal self-shielding factors in wires composed of up to 57 nuclides. Importantly, the neural network can account for resonance interference that affects the self-shielding in samples containing nuclides with large overlapping resonances. We use Monte Carlo simulations of sample wires irradiated in a thermal neutron spectrum to create the data for training the neural network and validate its predictions. A Gaussian negative log likelihood loss function is combined with the ensemble to estimate the confidence in the predicted self-shielding factors when ground-truth data are unavailable.
