Debris transportation based on AI modelling for nuclear site decommissioning - AI-STARS: AI for Investigating Sludge Transport And RadWaste Removal

The safe and efficient retrieval and transport of fuel debris and radioactive sludge are pressing challenges in nuclear decommissioning, particularly in legacy facilities at Sellafield and post-accident environments such as Fukushima. These materials exhibit complex, evolving multiphysics behaviours that are not well captured by current simulation tools. This proposal aims to develop a transformative, AI-based modelling framework capable of simulating and optimising fuel debris and radioactive sludge transport processes at unprecedented scales and resolutions, with an emphasis on operational safety and feasibility. State-of-the-art computational methods such as Computational Fluid Dynamics (CFD) - Discrete Element Method (DEM) deliver high-fidelity simulations but remain computationally intensive, often requiring weeks of HPC runtime. They also neglect critical phenomena, including corrosion, radiolytic gas evolution, and heat generation. Traditional Eulerian approaches oversimplify particle characteristics, reducing realism, while Lagrangian models become infeasible at operational scales. Data centric AI-based models, on the other hand, often suffer from poor generalisability due to sparse training data in radioactive settings. To overcome these limitations, this project proposes a hybrid physics–AI strategy that blends mechanistic fidelity with scalable efficiency. The methodology centres on the co-development of benchmark experimental datasets and advanced AI-based simulation tools. Experimental campaigns in Japan will create fuel debris and sludge surrogates with controlled composition, porosity, and morphology. High-resolution diagnostics—including high-speed imaging, velocimetry, and tomography—will provide rich, reproducible datasets for AI training and AI physics based model calibration. These experiments will anchor the models in physical reality and ensure generalisability to real-world conditions. The modelling strategy builds on Imperial College London’s recent advances in neural network simulation frameworks. Neural Networks for Partial Differential Equations (NN4PDEs) accelerates PDE solvers, enabling faster and more accurate fluid and reactive transport simulations. In addition, a similar approach for particle modelling called NN4DEM (Neural Networks for Discrete Element Modelling) reduces the computational cost of discrete element modelling while maintaining detailed inter-particle physics. This project introduces NN4DEM as well as (Neural Networks for Finite-Discrete Element Method) NN4FDEM—a novel GPU-accelerated Combined Finite–Discrete Element Method enhanced with neural networks—to model interactions of arbitrary-shaped particles with fluids at large scale. Coupling NN4FDEM with NN4PDEs yields a unified framework for simulating multiphase solid–fluid systems at particle-level resolution. To further improve scalability and inference speed, the project introduces Scalable Computational AI for Learned Engineering Dynamics (SCALED) - a domain-invariant surrogate model built on diffusion AI methods and U-Net architecture. SCALED can infer fluid-particle dynamics across arbitrarily large, complex domains, offering rapid predictions suitable for real-time applications. The framework will be validated through criticality and reactivity assessments in partnership with Amentum and the University of Tokyo. Reactivity mapping with MONK and FETCH codes will constrain model outputs and ensure compliance with nuclear safety requirements. Ultimately, the validated models will feed into a digital twin platform integrating sludge and RadWaste transport mechanics and real-time safety analytics for nuclear decommissioning. This UK–Japan collaboration leverages complementary strengths. Imperial College London leads AI-accelerated modelling and surrogate development (NN4PDE, NN4FDEM, SCALED), while the University of Tokyo (UoT) contributes state-of-the-art CFD–DEM expertise and reduced-order models tailored to post-accident scenarios like Fukushima. UoT will also provide simulant fabrication, radiological analogues, and experimental validation of debris flow dynamics. Together, the partners will establish a robust platform for validated, AI-based simulation tools to support safe, efficient radioactive waste retrieval and long-term storage across international contexts.