A Bayesian escalation model to predict nuclear accidents and risk

This paper designs prediction models to estimate the chance of the most severe nuclear accidents (such as complete core melts) for the population of U.S. and worldwide nuclear reactors. The authors formally introduce the notion of random escalation of incident severity. They then develop a class of models that views accidents of high severity as members of a subpopulation of incidents of lower severity; a random escalation model (REM) uses Bayesian methods to update unobservable failure rates and other model parameters. The priors for failure rates are based on extensive engineering judgment about the probabilities of core melt. Predictive distributions for time to time to core melt are calculated from the model, based on operational experience and accident data accumulated to data; the results are compared with those of N.C. Rasmussen, H.W. Lewis, P.C. Groer and others. The paper includes three theorems that reveal the structure of separable densities for parameter updating, the invariance of REMs under severity level classification and the reproducibility of Poisson-Binomial REMs. In an appendix, the authors examine the special assumptions that are required to specify the current U.S. Nuclear Regulatory Commission risk model.