Interactive Training of Machine Learning Models for Nuclear Safeguards Surveillance

Volume: 68, December 2026, pages 15-23

Authors: Alvaro Casado-Coscolla^a,b , Carlos Sanchez-Belenguer^a , Erik Wolfart^a , Vitor Sequeira^a , Carlos
Angorrilla-Bustamante^a , Eduardo Vendrell-Vidal^b , Juha Pekkarinen^c , João Rocha^c , Kai Ruuska^c , Elena
Bellido-Vera^c , Maikael Thomas^d , Alex Pollack^d , Simone Rocchi^d , Nivetha Balasankaran^d , Melvin John^d , and
Martin Möslinger^d

^aEuropean Commission - Joint Research Centre, Ispra, Italy.
^b Instituto Universitario de Automática e Informática Industrial, Universitat Politècnica de València, València, Spain
^cEuropean Commission – Directorate-General for Energy, Luxembourg, Luxembourg
^d International Atomic Energy Agency, Vienna, Austria
 

Abstract: We introduce Next Generation Model Trainer (NGMT), a novel software tool designed for interactive training of machine learning models to support the review of surveillance images in nuclear safeguards. This domain poses unique challenges for machine learning applications, including the lack of pretrained models for relevant objects, high variability in objects and activities of interest between facilities, and the time-consuming process of creating labelled training data. To address these challenges, our approach involves training a dedicated model for each surveillance camera and area of interest, combining a pretrained deep vision model with a lightweight binary classifier trained ad-hoc for the specific task. The system leverages active learning and label propagation to enable efficient model training. This paper provides an overview of our methodology and presents the results of a validation campaign conducted at the nuclear inspectorates.

Reference guideline: Casado-Coscolla, A., et al. (2026). Interactive Training of Machine Learning Models for Nuclear Safeguards
Surveillance, ESARDA Bulletin - The International Journal of Nuclear Safeguards and Non-proliferation,
68,15-23.  https://doi.org /10.3011/ESARDA.IJNSNP.2026.3