Fusion energy is the Holy Grail of the energy world: low-radioactivity, low-waste, zero-carbon, high-output nuclear power that can run on hydrogen or lithium. To date, however, nuclear fusion has not been successfully scaled up due to difficulties in maintaining the necessary nuclear reactions. Now, a team of researchers has leveraged supercomputer-powered AI in an effort to address one of the key problems with scaling up fusion energy.
In the quest to control nuclear fusion, researchers need to manage plasma. When containing plasma, they use fusion devices called “tokamaks”: magnetic, donut-shaped fusion devices that hold fusion reactions in place so the plasma doesn’t lose its heat or interact with the surrounding materials. However, instabilities in this process (“disruptions”) allow plasma to escape, reach the walls of the tokamak, stop the reaction and potentially cause irreparable damage to the reactor itself. The problem is also scaling up: the larger the fusion reactor, the lower the surface area, increasing the risk of severe damage from a disruption.
Bill Tang (of the U.S. Department of Energy [DOE] and the Princeton Plasma Physics Laboratory [PPPL]) and a team of researchers set out to investigate these fusion disruptions with supercomputing. “We aim to accurately predict the potential for disruptive events before they occur,” Tang said, “as well as understand the reasons why they happen in the first place.”
The disruptions – which happen near-instantly – need to be detected as early as possible. So far, simulations have been unable to deliver fast enough predictions – so the researchers turned to machine learning, which has shown promising results for disruption prediction. The goal: to meet the 95 percent correct disruption prediction threshold required by the under-construction ITER Tokamak, which will be the larger fusion reactor in the world.
Julian Kates-Harbeck (lead author on the paper published in Nature) answered this challenge by developing the Fusion Recurrent Neural Network (FRNN), an AI disruption prediction tool. FRNN learns from thousands of experimental runs – tracking plasma current, temperature, density and other variables – and attempts to learn which factors signal imminent disruptions.
To meet the level of reliability that ITER will demand, the researchers ran FRNN on powerful machines. After initial runs on Tiger (a cluster at Princeton University), they turned to the (now-decommissioned) Titan supercomputer, where they ran FRNN on 6,000 Nvidia Tesla K20X GPUs. “Titan was invaluable for large scaling tests to see how close we could get to reaching solutions with significantly more computing power,” Kates-Harbeck said.
The team then moved on to the Tokyo Institute of Technology’s TSUBAME 3.0 supercomputer, where the Nvidia P100 GPUs allowed them to advance the warning time of their disruption prediction up to threefold. Encouraged, they moved to Summit – the world’s most powerful supercomputer, according to the most recent TOP500 list – where they replicated the results.
Now, the team is running FRNN on the AI Bridging Cloud Infrastructure (an AI-dedicated supercomputer in Japan) and eyeing the more difficult task: prevention.
“With powerful predictive capabilities, we can move from disruption prediction to control, which is the holy grail in fusion,” Tang said. “It’s just like in medicine—the earlier you can diagnose a problem, the better chance you have of solving it.”
About the research
The research discussed in this article was written by Julian Kates-Harbeck, Alexey Svyatkovskiy and William Tang. It was published in Nature and can be found at this link. The original article discussing the research can be found at OLCF’s website here.