Known as HEAT-ML, the new AI could lay the foundation for software that significantly speeds up the design of future fusion systems. Such software could also enable good decision-making during fusion operations by adjusting the plasma so that potential problems are thwarted before they start.
“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,” said Michael Churchill, co-author of a paper in Fusion Engineering and Design about HEAT-ML and head of digital engineering at PPPL.
Fusion, the reaction that fuels the sun and stars, could provide potentially limitless amounts of electricity on Earth. To harness it, researchers need to overcome key scientific and engineering challenges. One such challenge is handling the intense heat coming from the plasma, which reaches temperatures hotter than the sun’s core when confined using magnetic fields in a fusion vessel known as a tokamak. Speeding up the calculations that predict where this heat will hit and what parts of the tokamak will be safe in the shadows of other parts is key to bringing fusion power to the grid.
“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,” said Domenica Corona Rivera, an associate research physicist at PPPL and first author on the paper on HEAT-ML. “The worst thing that can happen is that you would have to stop operations.”
Simulating how heat impacts SPARC’s interior is central to this goal and a big computing challenge. To break down the challenge into something manageable, the team focused on a section of SPARC where the most intense plasma heat exhaust intersects with the material wall. This particular part of the tokamak, representing 15 tiles near the bottom of the machine, is the part of the machine’s exhaust system that will be subjected to the most heat.
To create such a simulation, researchers generate what they call shadow masks. Shadow masks are 3D maps of magnetic shadows, which are specific areas on the surfaces of a fusion system’s internal components that are shielded from direct heat. The location of these shadows depends on the shape of the parts inside the tokamak and how they interact with the magnetic field lines that confine the plasma.
HEAT-ML traces magnetic field lines from the surface of a component to see if the line intersects other internal parts of the tokamak. If it does, that region is marked as “shadowed.” However, tracing these lines and finding where they intersect the detailed 3D machine geometry was a significant bottleneck in the process. It could take around 30 minutes for a single simulation and even longer for some complex geometries.
HEAT-ML overcomes this bottleneck, accelerating the calculations to a few milliseconds. It uses a deep neural network: a type of AI that has hidden layers of mathematical operations and parameters that it applies to the data to learn how to do a specific task by looking for patterns. HEAT-ML’s deep neural network was trained using a database of approximately 1,000 SPARC simulations from HEAT to learn how to calculate shadow masks.
HEAT-ML is currently tied to the specific design of SPARC’s exhaust system; it only works for that small part of that particular tokamak and is an optional setting in the HEAT code. However, the research team hopes to expand its capabilities to generalize the calculation of shadow masks for exhaust systems of any shape and size, as well as the rest of the plasma-facing components inside a tokamak.
