A collaboration between Commonwealth Fusion Systems (CFS), the US Department of Energy’s Princeton Plasma Physics Laboratory (PPPL), and Oak Ridge National Laboratory has resulted in a new AI technique that speeds up calculations critical to protecting fusion vessels from the extreme heat generated by plasma.
The AI system, known as HEAT-ML, rapidly identifies ‘magnetic shadows’ within a fusion device – regions shielded from direct plasma heat. This approach could form the basis for software that accelerates the design of future fusion systems and enables real-time decision-making during operations, allowing adjustments to prevent potential issues before they occur.
Michael Churchill, Head of Digital Engineering at PPPL and co-author of a paper on HEAT-ML published in Fusion Engineering and Design, explained: “This research shows that you can take an existing code and create an AI surrogate that will speed up your ability to get useful answers, and it opens up interesting avenues in terms of control and scenario planning.”
Fusion, the same process that powers the sun, has long been seen as a potential source of abundant, clean electricity. However, harnessing it on Earth requires controlling plasma that reaches temperatures hotter than the solar core. One of the greatest challenges lies in predicting how this heat interacts with the inner walls of a fusion vessel, or tokamak. Accurate, fast calculations of heat impact zones and protected regions are essential to designing and operating these machines efficiently.
Doménica Corona Rivera, Associate Research Physicist at PPPL and lead author of the HEAT-ML paper, highlighted the stakes: “The plasma-facing components of the tokamak might come in contact with the plasma, which is very hot and can melt or damage these elements. The worst thing that can happen is that you would have to stop operations.”
From SPARC simulations to smarter systems
HEAT-ML was developed to simulate a small section of SPARC, the tokamak currently being built by CFS in Massachusetts. SPARC aims to demonstrate a net energy gain – producing more energy than it consumes – by 2027. To achieve this, researchers must precisely model how plasma heat will affect the reactor’s interior.
The team focused on a critical area of SPARC’s exhaust system: 15 tiles near the bottom of the machine expected to experience the most intense plasma heat load. To simulate heat distribution, researchers create shadow masks – 3D maps showing where magnetic field lines shield components from direct plasma exposure. These depend on the vessel’s geometry and magnetic configuration.
Originally, the open-source HEAT (Heat flux Engineering Analysis Toolkit) program handled these simulations. Developed by CFS Manager Tom Looby during his doctoral research with Matt Reinke, now SPARC Diagnostic Team Leader, HEAT was first applied to PPPL’s National Spherical Torus Experiment-Upgrade project.
However, the process was computationally demanding. Each HEAT simulation could take up to thirty minutes or longer for complex geometries, as it involved tracing magnetic field lines and calculating their intersections with intricate 3D surfaces.
AI reduces calculation time from minutes to milliseconds
HEAT-ML removes this bottleneck using a deep neural network trained on around 1,000 HEAT-generated SPARC simulations. The AI learned to predict shadow masks within milliseconds – a dramatic improvement that enables rapid exploration of design options and faster operational responses.
Currently, HEAT-ML remains specific to SPARC’s exhaust system, functioning as an optional setting within the HEAT code. Nonetheless, the research team plans to expand its scope to handle other plasma-facing components and diverse exhaust geometries across future fusion devices.
By combining advanced AI with plasma physics, the partnership between PPPL, Oak Ridge, and CFS has demonstrated how digital tools could accelerate fusion research. Faster simulations not only streamline design but may also one day allow fusion reactors to self-adjust in real time, protecting their components and maintaining stable operation.
