Erlangen 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
AGA: Arbeitsgruppe Physik und Abrüstung
AGA 7: An Overview of Physics and Disarmament Research in Germany – Social Event & Poster Session
AGA 7.1: Poster
Thursday, March 19, 2026, 17:30–18:30, KH 00.016
Nuclear archaeology with normalizing flows: Learning about past plutonium production from reprocessing waste — •Fabian Unruh1 and Malte Göttsche1,2 — 1Peace Research Institute Frankfurt — 2TU Darmstadt
Plutonium, a material used in nuclear weapons, is produced in nuclear reactors. In reprocessing, plutonium is separated from the fuel rod—a crucial procedure for weaponization—producing highly radioactive waste. Being very informative about operational reactor parameters, an analysis of the waste could be an asset for verifying declarations on plutonium production for an arms control agreement.
In this nuclear archaeology approach, the isotopic composition of the waste is analyzed with normalizing flow-based invertible neural networks, which unlike many Deep Learning (DL) techniques output probability density functions rather than point-estimates. The obtained posterior distributions for burnup and time since irradiation can represent multiple modes corresponding to mixtures of waste from several irradiation campaigns. Knowing only true parameter values, but not true posterior distributions, for test data renders validation challenging and requires specific approaches. First, coverages of posterior distributions, i.e. the fraction of instances where a confidence region covers the true value, are compared to expectations. Second, clustering algorithms partition multimodal distributions and the probability mass of individual modes is related to the corresponding campaign’s share in the total waste. The presented results contribute to developing sound and trustworthy DL techniques for nuclear archaeology.
Keywords: Nuclear Archaeology; Nuclear Verification; Deep Learning; Inverse Problems; Normalizing Flows