Bonn 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

AGA: Arbeitsgruppe Physik und Abrüstung

AGA 2: Nuclear Archeology

AGA 2.1: Vortrag

Mittwoch, 12. März 2025, 16:00–16:30, HS HISKP

Applying neural networks in nuclear archaeology with nuclear reprocessing waste — •Fabian Unruh¹ and Malte Göttsche^1,2 — ¹PRIF -- Leibniz-Institut für Friedens- und Konfliktforschung — ²Technische Universität Darmstadt

Nuclear reprocessing waste is obtained by the separation of plutonium---potentially used for nuclear weapons---from nuclear reactor fuel. An analysis of the isotopic waste composition reveals information about the irradiation process and is useful for verifying fissile material declarations in arms control agreements.

For the analysis of waste, we simulate the nuclear reactor at Yongbyon, DPRK, and apply neural networks for the reconstruction of burnup and cooling time of several irradiation campaigns ("batches"). First, fully-connected neural networks are trained on isotopic ratios of reprocessing waste from a single batch for the reconstruction. To reduce laboratory costs in an application, the most important isotopic ratios for the reconstruction are selected by applying a gradient-based metric. Finally, posterior distributions are obtained by conditionally invertible neural networks using the reduced isotopic ratio set and the prospect of resolving different batches in the waste is explored.

The methodology successfully reduces the set of isotopic ratios without deteriorating the reconstruction capabilities. The applied neural networks yield multi-modal posterior distributions for the irradiation parameters of different batches. The techniques contribute to making nuclear archaeology for reprocessing waste suitable for application by addressing challenges of costly measurements and mixtures of waste.

Keywords: Nuclear Archaeology; Nuclear Verification; Neural Networks; Inverse Problems; Fissile Material