Nuclear data evaluation with Bayesian networks

TL;DR

This paper introduces a Bayesian network approach for nuclear data evaluation, enabling flexible, detailed inference of uncertainties and systematic errors, demonstrated through three proof-of-concept examples involving neutron data and $^{56}$Fe.

Contribution

It presents a novel application of Bayesian networks with non-linear relationships and Gaussian processes to improve nuclear data evaluation methods.

Findings

Effective modeling of energy-dependent error components

Successful evaluation of $^{56}$Fe data in challenging energy ranges

Public availability of R scripts and the nucdataBaynet package

Abstract

Bayesian networks are graphical models to represent the probabilistic relationships between variables in the Bayesian framework. The knowledge of all variables can be updated using new information about some of the variables. We show that relying on the Bayesian network interpretation enables large scale inference and gives flexibility in incorporating prior assumptions and constraints into the nuclear data evaluation process, such as sum rules and the non-negativity of cross sections. The latter constraint is accounted for by a non-linear transformation and therefore we also discuss inference in Bayesian networks with non-linear relationships. Using Bayesian networks, the evaluation process yields detailed information, such as posterior estimates and uncertainties of all statistical and systematic errors. We also elaborate on a sparse Gaussian process construction compatible with the Bayesian network framework that can for instance be used as prior on energy-dependent model parameters, model deficiencies and energy-dependent systematic errors of experiments. We present three proof-of-concept examples that emerged in the context of the neutron data standards project and in the ongoing international evaluation efforts of $^{56}$Fe. In the first example we demonstrate the modelization and explicit estimation of relative energy-dependent error components of experimental datasets. Then we show an example evaluation using the outlined Gaussian process construction in an evaluation of $^{56}$Fe in the energy range between one and two MeV, where R-Matrix and nuclear model fits are difficult. Finally, we present a model-based evaluation of $^{56}$Fe between 5 MeV and 30 MeV with a sound treatment of model deficiencies. The R scripts to reproduce the Bayesian network examples and the nucdataBaynet package for Bayesian network modeling and inference have been made publicly available.