In the race to harness fusion energy, every second counts. A Princeton-led team has created HEAT-ML, an artificial intelligence surrogate that can spot magnetic shadows inside a fusion device in milliseconds, speeding up calculations that protect the most heat-battered surfaces of a tokamak’s exhaust system.
The work, detailed in the peer-reviewed journal Fusion Engineering and Design, focuses on SPARC, a compact, high-field tokamak now being built by Commonwealth Fusion Systems with the Princeton Plasma Physics Laboratory and Oak Ridge National Laboratory.
The researchers trained a neural network on results from the Heat flux Engineering Analysis Toolkit (HEAT) so it could take over the slowest step in the pipeline, calculating the geometric shadow mask, without losing accuracy. That speed could enable both faster design iterations and between-shot adjustments in devices working toward net energy gain. The study is linked to the DOI 10.1016/j.fusengdes.2025.115010 and builds on PPPL’s collaboration with the U.S. Department of Energy.
Why Magnetic Shadows Matter
In a tokamak, magnetic fields hold a superheated plasma in place, guiding exhaust heat along field lines toward the divertor. If heat strikes unprotected surfaces, it can melt or crack them. Engineers therefore shape components so upstream tiles shield downstream edges, creating magnetically shadowed regions that keep intense heat on robust targets. Updating these maps quickly becomes critical as machines like SPARC increase power and explore new plasma configurations. (For an overview of how a tokamak works, see the DOE’s explainer.)
From HEAT To HEAT-ML
HEAT, the open-source Heat flux Engineering Analysis Toolkit, traces magnetic field lines and checks where they intersect with detailed CAD geometry to produce a shadow mask, a binary map showing shielded and unshielded points. This collision detection step dominates the runtime. HEAT-ML replaces it with a neural network that predicts the mask directly from a compact set of equilibrium parameters, while keeping HEAT’s downstream physics untouched. The current work targets a 15-tile region in SPARC’s divertor, where the exhaust heat is most intense.
How The Model Learns
The team generated about one thousand HEAT cases for SPARC, covering a variety of plasma equilibria. Inputs to the neural network included plasma current, the edge safety factor q95, and the incident field angles at the top and bottom of the divertor carrier. Training on GPUs reduced optimization from minutes to seconds, enabling quick experimentation. On unseen cases, HEAT-ML reproduced shadow masks with high accuracy, letting HEAT calculate heat flux profiles with only small differences from the standard pipeline, yet orders of magnitude faster.
Key Findings
- Shadow masks generated in milliseconds make near real-time heat mapping possible on fixed divertor geometry.
- Accuracy matches HEAT’s traditional collision-detection approach, while reducing per-case runtime to seconds (mostly I/O time).
- Available as an optional module in HEAT, keeping full CAD and physics context for engineers.
- Supports between-shot decision-making now and could enable closed-loop divertor protection in the future.
Limits And Next Steps
Because the current model was trained on a fixed mesh, using it for a new geometry requires generating a fresh HEAT dataset. The team is working to generalize the approach to other plasma-facing components and develop mesh-flexible methods. They note that collision detection is still a major bottleneck even in axisymmetric cases, pointing to opportunities for surrogates that can also handle field-line integration. In the long term, combining HEAT-ML with infrared diagnostics and a plasma control system could give operators live heat maps that update instantly if shape, currents, or power change mid-shot.
A Quote From The Team
“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 of the Princeton Plasma Physics Laboratory.
What It Means For Fusion
High-field, compact tokamaks aim for high performance in smaller spaces, but success depends on keeping heat where the hardware can handle it. Faster, accurate shadow mapping lets teams test tile designs, choose operating points, and avoid vulnerable spots without waiting on long simulations. That moves operations from cautious trial-and-error to confident, data-driven control. If operators could see, in real time, which surfaces are safely shadowed and which are exposed, it could change how far they push performance without crossing material limits.
Where To Learn More
Read the peer-reviewed study via its DOI. Explore SPARC’s vision at Commonwealth Fusion Systems. For broader context on fusion science and U.S. research programs, visit the Department of Energy and the Princeton Plasma Physics Laboratory.
ScienceBlog.com has no paywalls, no sponsored content, and no agenda beyond getting the science right. Every story here is written to inform, not to impress an advertiser or push a point of view.
Good science journalism takes time — reading the papers, checking the claims, finding researchers who can put findings in context. We do that work because we think it matters.
If you find this site useful, consider supporting it with a donation. Even a few dollars a month helps keep the coverage independent and free for everyone.