Fusion researchers in the United States are turning to artificial intelligence to solve one of the field’s most stubborn bottlenecks: the painfully slow process of designing and testing stellarators, the intricate magnetic cages that can confine ultra‑hot plasma. Instead of waiting hours or days for supercomputers to evaluate a single configuration, new AI‑driven tools promise to generate and screen viable designs in seconds.
By pairing machine learning with high‑performance computing, scientists aim to explore a vastly larger design space, build digital twins of experimental devices, and ultimately connect fusion concepts to the power grid far faster than traditional methods allow. The strategy is straightforward but ambitious: automate the most complex calculations, then let human experts focus on the physics and engineering choices that matter most.
AI, STELLAR‑AI and the new fusion design playbook
The United States is knitting together a national fusion computing strategy that leans heavily on artificial intelligence, with the Department of Energy positioning fusion as a pillar of the country’s future energy mix. On its public roadmap, the Department of Energy highlights fusion as part of a broader effort to expand low‑carbon power, and that policy backdrop is now shaping how laboratories invest in software as much as hardware. Instead of treating AI as an add‑on, researchers are embedding it directly into the design cycle for reactors, magnets and control systems.
At the center of this shift is a new platform called STELLAR‑AI, led by the Princeton Plasma Physics Laboratory, which is described as a computing environment that pairs artificial intelligence with high‑performance computing to remove long‑standing bottlenecks in fusion research. The core platform, detailed by PPPL, is built to handle the extreme complexity of plasma physics calculations while still delivering results quickly enough to guide day‑to‑day experimental decisions.
From weeks to seconds: StellFoundry and the stellarator revolution
Stellarators have always been a kind of extreme sport in fusion engineering, with twisted, three‑dimensional magnetic coils that are notoriously hard to design and optimize. Traditional codes can take hours to evaluate a single configuration, which means exploring the full range of shapes, field strengths and plasma conditions is practically impossible. Earlier work by scientists at Princeton Plasma Physics Laboratory already showed that new numerical approaches could cut this time dramatically, with a US fusion code slashing stellarator design time to under 10 seconds while still accounting for the intricate three‑dimensional magnet geometries that make these devices so powerful and so difficult.
Building on that foundation, a new effort called StellFoundry is now using AI to push stellarator design into an even faster regime, where entire families of configurations can be generated and tested in a fraction of a second. Reporting on the project explains that the StellFoundry system lets scientists rapidly explore a huge number of stellarator configurations, using advances in AI and high‑performance computing supported by the DOE’s Office of Science to evaluate how each design would behave. One account notes that the US system can test stellarator design options within a fraction of a second, while another emphasizes that researchers plan to use advances in AI to scan through many possibilities under the DOE’s Office of umbrella.
Inside STELLAR‑AI: digital twins, NSTX‑U and cross‑lab computing
STELLAR‑AI is not just a single code, it is a coordinated platform that ties together multiple AI models and physics simulations so they can share data and learn from each other. Project leaders describe it as a way to pair speed with precision, with STELLAR‑AI explicitly designed to combine rapid AI‑driven predictions with the accuracy of high‑fidelity plasma models. The goal is to accelerate scientific discovery across DOE laboratories, so that a breakthrough in one code or device can be propagated quickly to others instead of remaining siloed.
One of the flagship efforts inside this ecosystem is the creation of a digital twin of NSTX‑U, the National Spherical Torus Experiment Upgrade, which is a major fusion facility at Princeton Plasma Physics Laboratory. According to a detailed description of the program, one of the key STELLAR‑AI projects is to build this virtual replica so that researchers can test control strategies and plasma scenarios in software before trying them on the real machine. Another project, highlighted in a separate report, is StellFoundry itself, which is described as using AI to sift through enormous amounts of data and design information so that stellarator configurations can be optimized far more quickly, with another account stressing that this will allow researchers to explore a much wider design space and speed scientific discovery across DOE laboratories.
Cutting timelines by orders of magnitude
What makes STELLAR‑AI and StellFoundry so disruptive is not just that they are faster, but that they change what kinds of questions fusion scientists can afford to ask. Instead of hand‑tuning a handful of designs, researchers can now treat stellarator optimization as a search problem over millions of possibilities, with AI models steering the search toward promising regions of parameter space. A detailed overview of the computing initiative notes that this US‑led project is explicitly framed as a way to remove one of the biggest obstacles to fusion progress, namely the sheer time it takes to run the necessary simulations, with the expectation that AI will cut research timelines by orders of magnitude. That ambition is spelled out in coverage of the STELLAR‑AI project, which emphasizes how much faster design and analysis cycles could become.
Earlier work already hinted at what such acceleration might look like in practice. A widely cited report on a US fusion code described how scientists were able to reduce stellarator design time to under 10 seconds while still accounting for key parameters such as plasma pressure, shape and magnetic field strength, all of which are crucial for commercial power generation. STELLAR‑AI and StellFoundry effectively take that concept and scale it up, using AI to orchestrate many such calculations in parallel and to learn from the results so that each new design cycle is smarter than the last.
From plasma control to the power grid
Designing a stellarator in seconds is only part of the story, because fusion has to integrate with the broader energy system if it is going to matter on the grid. Analysts working with the fusion community point out that expanding the nation’s energy portfolio will require not just new reactors but also new tools to understand how fusion plants interact with existing infrastructure. One recent assessment of AI in energy notes that expanding the nation’s energy portfolio by making fusion a viable source of electricity involves designing complex devices and then predicting their behavior in far less time than traditional methods allow, particularly for stellarators whose three‑dimensional fields are difficult to model.
Researchers are therefore using AI not only to shape magnetic coils but also to coordinate the many computer programs that simulate plasma behavior, materials performance and grid dynamics. A detailed technical overview explains that using AI to advance U.S. leadership in fusion science involves getting separate computer programs to work together, so that scientists can quickly identify which designs have the properties they want and how those devices might operate as part of a real‑world power system. That same perspective stresses that coordinating these codes is only one aspect of the challenge, but it is a necessary step if fusion plants are to be dispatched and controlled like other grid assets.
More from Morning Overview