Researchers from the UK Atomic Energy Authority (UKAEA) and the U.S. Department of Energy are developing a plan to link supercomputing facilities across both nations for fusion energy modeling. The project, FUSION-ML, aims to create a federated digital twin by training machine learning models on combined data from the MAST-U tokamak in the UK and the DIII-D National Fusion Facility in the US. This approach allows models to learn from diverse operational scenarios without transferring the petabyte-scale datasets, a process that would be prohibitively slow and expensive. The collaboration involves the Princeton Plasma Physics Laboratory (PPPL), Lawrence Livermore National Laboratory (LLNL), and Oak Ridge National Laboratory (ORNL). Source: PPPL
The core technical strategy is federated learning, a distributed machine learning technique. Instead of centralizing raw experimental data, models are trained locally on the respective supercomputers at each facility. Only the updated model parameters, or gradients, are then exchanged and aggregated to build a single, more robust predictive model. This method preserves data sovereignty and security while enabling the AI to generalize from a wider set of plasma conditions than either machine could provide alone. The resulting 'smarter fusion twin' is expected to yield more accurate predictions of plasma behavior, a critical step for designing and operating future reactors like ITER and DEMO. Source: PPPL
The core technical strategy is federated learning, a distributed machine learning technique.
A primary application for this federated model is the prediction and mitigation of plasma disruptions. These events, where plasma confinement is abruptly lost, can impose extreme thermal and electromagnetic stress on tokamak components, potentially causing significant damage. By training on data from both the conventional-aspect-ratio DIII-D and the spherical MAST-U, the model can learn to identify disruption precursors across different machine geometries and operating regimes. According to Tim Arts, a computational scientist at UKAEA, this enhanced predictive capability is essential for developing reliable control systems that can sustain stable, long-pulse operations in commercial-scale fusion power plants. Source: PPPL
The computational resources involved are among the world's most powerful. The project intends to utilize the exascale supercomputers Frontier at ORNL and the forthcoming El Capitan at LLNL, alongside the UK's ExCALIBUR high-performance computing program. This immense computational power is necessary to process the complex, high-dimensional data from tokamak diagnostics and run the sophisticated simulations that form the basis of the digital twin. The FUSION-ML project is supported by a joint program between the U.S. Department of Energy and the UK's Engineering and Physical Sciences Research Council (EPSRC), highlighting a strategic international focus on applying advanced computing to accelerate fusion science. Source: PPPL