The collaboration between the ALCF and the DIII-D National Fusion Facility integrates experiments, AI, data, and supercomputing to accelerate fusion research. (Images: Argonne National Laboratory and DIII-D National Fusion Facility)
Using AI and supercomputers to streamline fusion experiments
By integrating DIII-D experiments with ALCF computing resources, researchers are enabling faster data analysis and developing AI-powered digital twins to help inform decisions between plasma shots.
To advance fusion energy research, scientists at the DIII-D National Fusion Facility study plasmas in short experimental shots that last only seconds. After each plasma shot, a cooling period offers a brief window to analyze the results and prepare for the next shot in the experiment.
“The shots occur about every 10 to 20 minutes,” said David Schissel, computer systems and science coordinator at DIII-D. “The idea is to do analysis between shots, figure out what’s happening in the plasma, and then decide what to change for the next shot to progress your experiment.”
Researchers use the DIII-D tokamak to study fusion plasmas and investigate key science and technology challenges on the path to practical fusion energy. (Image: DIII-D National Fusion Facility)
These experiments are helping scientists better understand and control the behavior of plasmas, a key challenge in developing fusion as a practical energy source. But turning data into decisions quickly enough to guide the next experiment requires high-performance computing (HPC) and AI.
Through a long-standing collaboration between DIII-D and the U.S. Department of Energy's Argonne National Laboratory, researchers are bringing advanced computing capabilities into the experiment cycle. The effort links two DOE Office of Science user facilities: DIII-D and the Argonne Leadership Computing Facility (ALCF). DIII-D is the nation's largest magnetic fusion experiment, operated by General Atomics on behalf of DOE. The ALCF is home to some of the world's most powerful supercomputing and AI systems available for scientific research.
The work enabled by this partnership is now helping advance broader DOE efforts under the Genesis Mission and its American Science Cloud platform. Many of the capabilities being developed through the DIII-D–ALCF collaboration support the goals of the American Science Cloud. These include integrated workflows, automated data movement, and on-demand computing access. Together, they help connect experimental facilities, supercomputers, data systems, and AI services into a unified ecosystem.
In parallel, DIII-D researchers are developing an AI-enabled digital twin of the facility to further enhance the impact of each experiment. The project uses supercomputers at the ALCF and the National Energy Research Scientific Computing Center (NERSC) to develop and train AI models for the digital twin. NERSC is a DOE Office of Science user facility at Lawrence Berkeley National Laboratory.
“Having a digital twin in the control room would allow scientists to explore countless different plasma scenarios in the digital world without using any experimental shots,” Schissel said. “Then, once you have figured out what you want to try, you can feed those parameters into our control system and run the experiment in the real world. That would be a huge step forward, helping scientists make the best use of our experimental resources and accelerate progress in fusion energy science and technology.”
Building the foundation for integrated science
The collaboration between Argonne and DIII-D dates back 25 years. Researchers from both institutions were part of one of DOE's first Scientific Discovery through Advanced Computing (SciDAC) awards, the National Fusion Collaboratory Project. The project focused on developing software tools to enhance collaboration and data sharing across the fusion research community.
“What we started with the Fusion Collaboratory was really an early vision of integrated research infrastructure,” said ALCF Director Michael Papka. “Today, efforts like the American Science Cloud are taking those same ideas further by integrating AI, data platforms, experimental facilities, and exascale computing into a more seamless and capable scientific ecosystem.”
Back in 2001, however, bringing advanced computing into fusion experiments was still more vision than reality.
“When we started 25 years ago with the Collaboratory, we ran very few things between shots because we couldn't do it. Instead, we ran a lot of analysis overnight,” Schissel said. “You often would come in the next day and look at your data, spend a day or two analyzing it, and realize you didn't understand something that was going on.”
Over time, advances in HPC have turned overnight workflows into calculations that can now be completed between plasma shots. But many of these analyses require capabilities beyond the facility's in-house computing resources.
Researchers use the ALCF's Polaris supercomputer to run computationally intensive fusion simulations and analysis. (Image: Argonne National Laboratory)
“We have local computing resources, but we’re working with analysis workflows that dwarf the capabilities of anything we have on site,” Schissel said. “You increase the fidelity, the spatial resolution, the time resolution, reduce assumptions in the code, and suddenly you need a supercomputer.”
Making this possible requires technologies that allow experiments, data systems, and supercomputers to work together seamlessly. ALCF researchers, including Thomas Uram and Christine Simpson, have worked closely with the DIII-D team to develop several key capabilities. These include on-demand queues for time-sensitive workflows, service accounts for automated job execution, and Globus services for secure data movement during live experiments.
Accelerating analysis with supercomputers
Much of the current work is being carried out through a DOE award of supercomputing time at the ALCF and NERSC. The project is led by Mark Kostuk, leader of DIII-D's Advanced Computing Group. A major focus of this effort is running advanced analysis workflows on remote DOE supercomputers and returning results quickly enough to support live experiments.
Two of the workflows targeted through the project are Consistent Automatic Kinetic Equilibria (CAKE) and IonOrb. CAKE models the plasma's magnetic equilibrium self-consistently across multiple diagnostics. These results provide a starting point for many other analyses. IonOrb tracks the paths of energetic particles inside the tokamak. This enables researchers to identify where particles strike the vessel wall, providing information that can help protect the machine.
The IonOrb code generates a heat map of deposited high energy particles onto the walls of the DIII-D tokamak (blue). This information is crucial for the protection of sensitive equipment and the prevention of excessive impurity generation. (Image: DIII-D National Fusion Facility)
By optimizing these workflows to run on remote DOE supercomputers, the team is demonstrating how rapid analysis can support live experiments where timely results are critical.
"ALCF and NERSC allow us to routinely run more demanding analyses and get results back faster," Kostuk said. "That ultimately leads to better experiments."
Developing an AI-powered digital twin
The workflows used to support experiments between plasma shots are also helping researchers build the AI-enabled digital twin of DIII-D. This effort brings together General Atomics and NVIDIA, with support from the ALCF, NERSC, and the San Diego Supercomputer Center. The goal is to create a virtual version of the facility to help scientists predict plasma behavior and evaluate different scenarios before testing the best options experimentally.
To develop the digital twin, the team is combining experimental data with large amounts of simulation data generated on supercomputers to train AI surrogate models. By running CAKE, IonOrb, and other codes routinely on DOE supercomputers, the team is generating high-fidelity data that can be used to train AI surrogate models. These models approximate the output of more time-consuming simulations, allowing the digital twin to provide real-time feedback that supports interactive use in the control room.
That speed is critical during live operations. For example, a full CAKE workflow can take roughly 30 minutes to an hour to complete on a supercomputer, while a surrogate model can return results almost immediately.
Researchers review experimental data and coordinate plasma operations in the DIII-D control room. (Image: DIII-D National Fusion Facility)
“By bringing the digital twin into the control room, researchers can use models that execute in milliseconds rather than the hours or days required for traditional simulations to help guide experiments,” said Raffi Nazikian, fusion data science lead at General Atomics.
The DIII-D team recently presented a demo of the digital twin at Confab26, a conference jointly organized by the Energy Sciences Network (ESnet) and the American Science Cloud. The demo showed how researchers can use data from an experimental shot to run virtual scenarios and compare different control options before deciding what to try next.
“If you can run these virtual experiments from your desk and see what would happen, you're effectively multiplying the amount of experimental time you have,” Kostuk said.
As work continues under the Genesis Mission and the American Science Cloud, the DIII-D–ALCF collaboration provides an example of how experiments, AI, data, and advanced computing can be integrated across facilities to accelerate discovery. Building on successes with workflows such as CAKE and IonOrb, the team aims to optimize additional analysis code for DOE supercomputers while continuing to develop and validate the digital twin and its underlying models. As new scenarios and simulation results become available, the team can retrain the models to improve their performance over time.
“Validating the digital twin on an operating fusion facility is a key step toward making this vision a reality,” Nazikian said. “Once we have this in hand, we’ll have a powerful tool that can advance fusion research, improve facility operations, and accelerate the design and operation of future fusion reactors so we can bring about fusion energy much faster.”
This research uses resources and expertise from the ALCF, NERSC, and the DIII-D National Fusion Facility, with support from DOE’s Advanced Scientific Computing Research (ASCR) Leadership Computing Challenge (ALCC) program, the American Science Cloud (AmSC) project and Integrated Research Infrastructure (IRI) program.
To learn more about this work, join us on July 16 for the ALCF Service-Enabled Science webinar, “Fusing Experiments, Data, and Computing at DIII-D and ALCF,” which will cover automated remote analysis workflows, the DIII-D digital twin, and the Fusion Data Platform.
