Fusion, the nuclear reaction that powers the Sun and the stars, has incredible potential as a source of safe, carbon-free and essentially limitless energy. But making it a practical and economic reality has bedeviled scientists since the 1950s. Now fusion researchers are getting ready to leverage the power of exascale computing to unravel the mysteries of what may be the ultimate renewable energy source.
“To really understand what’s going on and what’s going to happen in the next experiment, you need big codes and big computers,” says Dr. Choongseok “CS” Chang, lead PI of a multi-institutional multi-disciplinary U.S. SciDAC Partnership Center for High-fidelity Boundary Plasma Simulation, headquartered at Princeton Plasma Physics Laboratory, Princeton University.
Exascale supercomputers are exactly what current fusion research needs, Chang explains. One of the biggest current challenges is making accurate predictions about the processes that occur inside tokamak reactors, which use giant magnetic fields to confine plasma fuel in a torus shape to achieve the conditions necessary for fusion. To advance this science, Chang’s team is preparing to use the Aurora Exascale supercomputer, the country’s first Intel-architecture-based exascale HPC system that will be deployed at the U.S. Department of Energy’s (DOE) Argonne National Laboratory.
It’s critical to be able to predict and control any disturbances that occur inside a tokomak that could bring the ultra-hot plasma into contact with the reactor wall. Engineers have windows measured in milliseconds to control instabilities before the plasma erupts from its magnetic confinement and potentially damages the reactor.
The walls of the reactor must be made out of materials that can bear the incredibly high heat and pressure of the plasma. Tungsten has the highest tensile strength of any pure metal, which is why it is being incorporated into the tokamak at the International Thermonuclear Experimental Reactor (ITER), an international fusion research and engineering project in southern France. Begun in 2013 and projected to have its first plasma in 2026 and begin full operation in 2035, ITER is the world’s largest tokamak, and is intended to prove that large-scale fusion energy is possible.
Experiments at JET (Joint European Tokamak) found that using tungsten in the tokamak walls resulted in a plasma confinement that was lower than expected. “It was totally unexpected,” Chang says, “so they’re really worried. If that was true, then ITER may have difficulty in producing 10 times more fusion energy than the input energy in its present design condition.”
But experiments have also showed that injecting a very light impurity material such as nitrogen or neon gas made the confinement levels recover. Nevertheless, even without injecting light impurity particles, researchers at the Joint European Torus (JET) fusion project in Oxfordshire, UK, were able to produce a groundbreaking 59 megajoules of energy during a five-second pulse in December 2021, more than doubling the previous world record. “This was a historic event, enough to claim that yes, fusion is actually practical,” Chang offers. However, the energy yield rate was not yet up to the level required by ITER since they did not inject light impurity particles into the tokamak edge.
Since JET uses the same wall material to ITER, Chang says “our early science on Aurora is to understand this tungsten-wall experiment and how this will be extrapolated to ITER. We need to do first-principles-based high fidelity simulation and fundamentally verify the physics.” Their questions include why tungsten is degrading the fusion performance so much, why light impurity particles would bring back the performance, and how best to incorporate them into the reactor design.
The increased processing powers of exascale makes much higher fidelity scientific predictions and offers the potential to train more specialized surrogate models that can be shared in real-time with experimental facilities. “By using large scale HPCs optimized for AI and ML, there can be daily communication and progress between exascale computers running large scale simulations and large scale experiments,” Chang says, comparing it to the current trial-and-error process that can take years. “Aurora is expected to be 2 exaFLOPs peak dual-precision – that will be perfect.”
Chang’s team is using XGC Gyrokinetic code, a modern particle-in-cell code built and optimized for extreme scale computers, especially GPU machines. It’s highly scalable and open source to the U.S. community. “It’s a big code designed to take advantage of big modern HPCs – when I see the specs of Aurora, I get excited,” he laughs. He hopes to have the code ready for Aurora in 2022 or early 2023.
Aurora will integrate Intel’s upcoming HPC and AI hardware and software innovations, including future generation Intel Xeon Scalable processors (codenamed Sapphire Rapids HBM), and accelerated by future Intel Data Center GPUs (codenamed Ponte Vecchio). It is based on Slingshot 11 fabric and the HPE Cray EX supercomputer platform. It will support ten petabytes of memory, and will leverage Intel Distributed Asynchronous Object Storage (DAOS) technology, supported on Intel Optane Persistent Memory. “The oneAPI unified programming model will simplify development across diverse architectures,” Chang says.
Chang has faith that Aurora will shorten the timeframe for the development of commercial fusion energy, which he acknowledges always seemed to be just a few decades away. In the meantime, he says, “there’s still a lot of work to do.”
Julian Smith is an award-winning green tech, conservation and travel writer based in Portland, Oregon, whose work has appeared in Wired, Smithsonian, New Scientist and the Washington Post among others.