Inferring Reactor Parameters from Nuclear Waste with Bayesian Inference

Future nuclear disarmament regimes would benefit from a verifiable account of the fis sile material that emerges from the dismantled warheads as well as the other stockpiles not under safeguards. Verifying the completeness of such an account is a challenging task that would be facilitated by applying nuclear archaeology methods to reconstruct the production history of nuclear facilities and corroborate fissile material stockpile de clarations. Nuclear archaeology is a field of research that studies and develops methods to use samples of irradiated material from a nuclear facility to derive its operational his tory. The research primarily focuses on methods applicable to nuclear reactors, where the principal sources of irradiated material are structural material from the core and spent fuel. In the context of military applications of nuclear technology, most spent fuel will have been reprocessed, and highlevel reprocessing waste may be one of the few sources of evidence. However, reprocessing waste stemming from multiple fuel batches and potentially from different reactors is likely to have been mixed, which adds signif icant complexity to the problem. Previous research has demonstrated the potential of a Bayesian inference framework to reconstruct burnup and time since irradiation of a waste mixture stemming from two batches of fuel using isotopic ratio measurements from re processing waste. Here, we present ongoing research to further explore the potential of this framework in a simulationbased study. We investigate models for reconstructing four reactor parameters (burnup, time since irradiation, initial enrichment, and average power density), for distinguishing between three different reactor types (Magnox, PWR, and PHWR), and for reconstructing the parameters of mixed waste samples. Each model is applied to a set of synthetic test data consisting of simulated isotope ratio values with associated “true” parameter values. We quantify the results of each individual inference run in terms of the highest posterior density region (HDR) and evaluate the performance of the models over the entire test dataset by comparing the HDR with the true parameter values. Reconstructing four operating parameters and distinguishing the reactor types performs well on the test dataset. Inference with mixed samples yields promising results, demonstrating the framework’s potential while highlighting the need for further research to develop a robust method.