Neutron Tomography of a Fuel Cell: Statistical Learning Implementation of a Penalized Likelihood Method

Published

October 15, 2013

Author(s)

Kevin J. Coakley, Dominic F. Vecchia, Daniel S. Hussey, David L. Jacobson

Abstract

At the NIST Neutron Imaging Facility, we collect neutron projection data for both the dry and wet states of a Proton-Exchange-Membrane (PEM) fuel cell. Transmitted neutrons captured in a scintillator doped with Lithium-6 produce scintillation light that is detected by an amorphous silicon detector. Based on joint analysis of the dry and wet state projection data, we reconstruct a residual neutron attenuation image with a penalized likelihood method. We use a Huber penalty function and select its regularization parameters by two-fold cross-validation. Before reconstruction, we transform the projection data so that the variance-to-mean ratio is approximately one. For the measured projection data, the Penalized likelihood method reconstruction is visually sharper than a reconstruction yielded by Filtered Back Projection. In an idealized simulation experiment, we demonstrate that the cross-validation procedure selects regularization parameters that yield a reconstruction that is nearly optimal according to a root mean-square prediction error criterion.

Citation

IEEE Transactions on Nuclear Science

Volume

60

Pub Type

Journals

Keywords

Bayesian, cross-validation, fuel cell, Huber penalty, method of surrogates, neutron attenuation, neutron transmission tomography, penalized likelihood, PEM fuel cell, statistical learning, water density

Citation

Coakley, K. , Vecchia, D. , Hussey, D. and Jacobson, D. (2013), Neutron Tomography of a Fuel Cell: Statistical Learning Implementation of a Penalized Likelihood Method, IEEE Transactions on Nuclear Science (Accessed April 3, 2025)