The partnership between Commonwealth Fusion Systems (CFS), the Princeton Plasma Physics Laboratory of the U.S. Department of Energy (PPPL), and Oak Ridge National Laboratory has led to the development of an innovative artificial intelligence (AI)-based approach for the rapid detection of so-called "magnetic shadows" in fusion reactors. These areas are shielded from intense plasma heat, which is crucial for the longevity and safety of reactor equipment.

A New AI Tool for Fusion Systems

The AI system, named HEAT-ML, could become the foundation for software that accelerates the design of future fusion devices. This tool is capable not only of optimizing engineering solutions but also of supporting real-time reactor operations by adjusting plasma parameters to prevent potential issues.

HEAT-ML was created to model a small section of the SPARC tokamak, which is currently being built by CFS. By 2027, the company aims to demonstrate a positive energy balance—meaning SPARC will generate more energy than it consumes.

Modeling Thermal Loads

Assessing the impact of heat on SPARC’s internal components is a key challenge that requires significant computational resources. To streamline the process, the team focused on the area where the most intense plasma heat flux contacts the wall material. This part of the tokamak consists of 15 tiles at the bottom of the machine and is subjected to the highest thermal loads.

For such simulations, researchers create so-called "shadow masks"—three-dimensional maps of magnetic shadows, or regions on the internal surfaces of components that are protected from direct heat exposure. The location of these shadows depends on the shape of the internal parts of the tokamak and their interaction with the magnetic field lines that confine the plasma.

Advantages of HEAT-ML

Initially, the open-source HEAT (Heat flux Engineering Analysis Toolkit) program was used to calculate shadow masks. HEAT was first applied to heat exhaust systems on the PPPL experimental facility. HEAT-ML tracks magnetic field lines from the surface of a component to determine whether they intersect other internal parts of the tokamak. If they do, that area is considered "shadowed." However, this analysis could take up to 30 minutes per simulation, and even longer for complex geometries.

HEAT-ML eliminates this bottleneck, reducing computation time to just a few milliseconds. The system uses a deep neural network trained on about 1,000 SPARC simulations performed with HEAT, enabling it to quickly calculate shadow masks.

Future Prospects

Currently, HEAT-ML works only for a specific part of SPARC’s heat exhaust system and is an optional feature within the HEAT code. The research team plans to expand the system’s capabilities to calculate shadow masks for any shape and size of heat exhaust systems, as well as for other components that come into contact with plasma inside the tokamak.

This work is supported by the U.S. Department of Energy under contracts DE-AC02-09CH11466 and DE-AC05-00OR22725, as well as by CFS.