Expanding the nation's energy portfolio by making fusion a viable source of electricity is challenging and involves designing complex fusion devices by sifting through enormous amounts of data to determine which device configurations are best. Now, a project led by a U.S. Department of Energy (DOE) national laboratory is pioneering ways to speed up the design of twisty fusion facilities known as stellarators by using artificial intelligence (AI) to sift through data more quickly.
Known as StellFoundry, the project focuses on replacing lengthy calculations in the design process with types of digital models, or surrogates, that provide rough calculation estimates. These fast approximations allow scientists to quickly test a large number of possible stellarator configurations more quickly than if they used calculations that were more precise but needed more time to perform. StellFoundry will also enable separate stellarator computer programs to work together, producing more complete simulations of stellarator behavior in less time. StellFoundry is led by the DOE's Princeton Plasma Physics Laboratory (PPPL) and comprises approximately 30 researchers from 10 organizations, including universities, software companies and national laboratories.
StellFoundry is supported by the DOE's Advanced Scientific Computing Research program and Fusion Energy Sciences. It is also part of the DOE's Scientific Discovery through Advanced Computing program.
The project is one of the latest ways that PPPL has used computation to advance fusion science. Since its founding 75 years ago, the Lab has continuously used advanced computer models to increase understanding of fusion plasma and design innovative fusion systems.
Scientists predict that StellFoundry innovations could accomplish in milliseconds what now takes hours or days. They plan to use advances produced by the project to more efficiently design stellarators that can produce the temperatures, pressures and stability needed for fusion reactions, all as economically as possible.
"When designing stellarators, computation has an outsized return on investment," said Michael Churchill, the head of AI for Science and digital engineering at PPPL and principal investigator of the project. "I'm excited about the possibility of giving stellarator designers the tools they need to make the process faster."
This project aligns with the DOE's new Genesis Mission, a major DOE initiative to accelerate scientific discovery and enhance national security using AI, as well as with the DOE's recently released Fusion Science and Technology Roadmap, which outlines a national strategy to accelerate the development of commercial fusion energy over the next decade by aligning public and private investment. The roadmap highlights critical gaps that innovative partnerships can fill to deliver the public infrastructure needed for the private sector's expansion into fusion in the 2030s.
StellFoundry specifically fits in with the Roadmap's AI-Fusion Digital Convergence Platform, an effort in which the DOE's investments in advanced computational modeling are boosting the development of both AI and fusion technology in partnership with the Office of Science's Advanced Scientific Computing Research program. It will also be crucial to help private stellarator companies participating in the DOE Milestone-Based Fusion Development Program in testing different system configurations.
"StellFoundry directly advances the DOE Fusion Science and Technology Roadmap by accelerating the integration of advanced computation, AI and digital engineering into fusion system design," said Jean Paul Allain, associate director for Fusion Energy Sciences at the DOE's Office of Science. "By reducing design timelines, improving whole-device optimization and enabling tighter coupling between physics and engineering models, this work supports the Roadmap's priorities in predictive capability, innovative confinement concepts and bridging foundational science to practical fusion energy systems."
Building on a legacy of computing leadership to advance fusion research
The StellFoundry team aims to decrease the amount of time required to design a stellarator by finding ways to allow a variety of individual computer programs to perform their calculations simultaneously, rather than one after another, as currently done. This process would be quicker and more accurate than the current procedure, in which scientists use programs to design different features of the stellarator or the plasma independently. Such features include the shape and placement of the magnets; the amount of unwanted plasma movement, or turbulence, that can cause heat to escape; the movement of plasma particles; and how many neutrons the plasma produces.
The challenge with this procedure is that all of the features are linked. If one of the calculations changes, or if the design of one of the parts has to be modified, all of the other parts of the stellarator must now be redesigned to mesh with this new parameter.
"But the more the tools are combined and the more they can interact, the more accurate the calculations will be at the very beginning," said Robert Hager, a PPPL research physicist and member of the StellFoundry team. "That's what you get from an effective framework - software tools talking to each other and producing better results with fewer steps."
Using AI to advance U.S. leadership in fusion science
Getting the separate computer programs to work together is only one aspect of StellFoundry. Another is speeding up the calculations of each individual step. Even with the most powerful supercomputers, calculating a feature like plasma turbulence - eddies and whirls in plasma that remove fusion-sustaining heat - is difficult and can take a large amount of time. But by training computer programs with large amounts of data from past simulations and experiments, scientists can get a probable answer much more quickly, reducing the calculation time from days to milliseconds.
"We can quickly determine how much turbulence a particular stellarator design might have by asking an AI program," said Churchill. "That means we can sift through lots of possible configurations quickly and find the few that have the properties we want."
Using decades of expertise to push the limits of science and engineering
Creating surrogates is not the same, though, as creating digital twins, a software-based model of a system that allows for easier manipulation of that system for greater efficiency. Typically, a digital twin involves a physical machine linked to a computer model of that machine. The computer model is constantly updated by data from the real machine, allowing researchers to test ideas on the twin before making any changes to the real device.
"People have asked about the possibility of creating digital twins for fusion power plants, but no one has developed such a twin because no fusion power plants yet exist," Churchill said. "Ultimately, though, once you have lots of the digital models that StellFoundry is developing, you start to have the basis for a possible digital twin just waiting for someone to build a physical plant."
VIDEO: What is a digital twin?
Developing AI systems for fusion facility design can only happen at America's national laboratories
Innovative programs like StellFoundry that require sustained periods of research happen readily in the DOE's national laboratories. "If you want to do something like StellFoundry, but you're a private company and only have funding for, say, five years, you might not be able to risk making that expense," Hager said. "And that's something we hear from the private companies themselves. They don't have the necessary time horizons. That's why we need publicly funded research. This is precisely the function it performs the best."
Looking to the future: Strengthening energy resiliency
The StellFoundry team is making progress, including plans to create a digital surrogate for a divertor, the part of a fusion system that removes unwanted plasma particles and waste heat. The team is also looking forward to overcoming some of the fundamental hurdles along the way to completion.
One hurdle involves getting different codes to jell, despite the fact that they are written using various techniques and incorporating different physics assumptions. "The task of bringing together lots of different codes is not a simple one," said Churchill. "It's not like connecting a bunch of wires and poof, you're ready to go."
"You have to pick the right codes, the right tools, to combine so that you have one set of assumptions," Hager said. "Not every combination of tools makes sense when you put them together, sometimes because the equations solved by the different tools are not always compatible."
Moreover, creating a stellarator computer code design framework might require answering questions in theoretical physics. "There is still a lot of unexplored territory that might require focused research or more effort," Hager said. "There may still be fundamental research that is needed to actually make such a framework in the first place."
At the moment, the team is thrilled by what might come in the future. "These AI tools and advanced optimization techniques could allow researchers to use their tools in ways they couldn't before," Churchill said. "And that's really exciting."
