The characterization of radioactive waste in Swedish legacy waste drums is a critical task for ensuring nuclear safety, environmental protection, and regulatory compliance. This project investigates the feasibility of using machine learning (ML) methods to image and localize actinide sources within these legacy waste drums by analyzing correlated neutron and gamma emissions from spontaneous fission. Because of the shielding and heterogeneity of the waste, conventional characterization methods often fail to provide adequate spatial or isotopic information. The proposed approach combines large-scale Monte Carlo simulations using Geant4 with modern data-driven inference techniques to enable non-destructive, remote characterization of nuclear materials. Our preliminary study at PDC based on a small allocation show very promising results in two research directions. Support from PDC for a medium applicaiton is essential to get those two papers published within the near future. Sweden’s renewed interest in nuclear energy underscores the importance of responsible waste management. Modern radioactive waste is handled under strict quality control, employing standardized procedures such as gamma spectroscopy, dose-rate mapping, and material sampling. However, Sweden also possesses substantial amounts of legacy waste, generated during earlier decades of nuclear research and industrial activity. Much of this Swedish legacy waste (SLW) is stored in concrete-filled oil drums originally intended for sea disposal. Documentation about the exact contents of these drums is incomplete or missing, and the concrete shielding complicates internal measurements. As a result, these waste packages cannot be sent for final disposal until they are properly characterized and classified according to modern standards. To minimize radiation exposure and avoid unnecessary handling risks, the waste should ideally be characterized without opening the drums. Historical records indicate that some drums may contain localized concentrations of nuclear material or high-activity sources. Identifying the position and nature of these sources is therefore essential for safe handling and sorting into appropriate disposal streams. The technical approach involves generating extensive simulated datasets using the Geant4 Monte Carlo toolkit. Geant4 will model the transport of neutrons and gamma rays within realistic drum geometries and their interactions with multiple external detectors. We need HPC to produce tens of millions of simulated events under varying source configurations. From these simulations, cumulative time-difference distributions between neutron and gamma detections will be constructed and used to train neural networks capable of predicting source locations. A separate network will be trained on gamma spectral information for isotope identification. The resulting ML models will be validated using experimental data from controlled measurements to assess their predictive accuracy and generalizability. By efficiently generating large, physically accurate training datasets and developing advanced inference algorithms, the project aims to demonstrate how ML-enhanced algorithm can accelerate and improve nuclear-waste characterization. The anticipated outcomes include validated predictive models for source localization and isotope identification, as well as a comprehensive simulated dataset available to the scientific community. The research will contribute to Sweden’s long-term goal of safely managing and disposing of its legacy nuclear waste in compliance with modern safety and environmental standards, ensuring public confidence in the sustainable use of nuclear energy.