Why trust nuclear data evaluations with Bayesian networks

Abstract

Bayesian networks were recently suggested as a framework for nuclear data evaluation. Their theory was to some extent described in a recent preprint and some example evaluations were presented. However, due to their newness in the context of nuclear data evaluation and consequently the lack of experience with them within the community makes it dicult to develop trust in the underlying methodology and consequently also the results produced by it. In this contribution, we aim to make a case why evaluators can trust this methodology in principle but will also elaborate on the fact that Bayesian networks are not a silver bullet for evaluation work. On the contrary, evaluators must assess and quantify essential assumptions about nuclear models and experiments with the same dilligence that is already necessary for the application of the wellestablished Generalized Least Squares (GLS) method. We also explain that the increased ease and flexibility to introduce assumptions regarding nuclear models, experiments and their relationships can help an evaluator to rigorously account for assumptions that are very often neglected in evaluations with the GLS method, such as the non-negativity of cross sections, relative experimental normalization uncertainties and the non-linearity in ratios of cross sections. We believe that adopting the Bayesian network paradigm can help both humans to produce evaluations with clearly traceable assumptions and machines to deal with nuclear data more eciently in terms of execution speed and storage size requirements.