Researchers from Harvard University, Princeton University’s Plasma Physics Laboratory, and the U.S. Department of Energy are working towards replicating the energy of the sun in the form of fusion energy.
The process involves heating nuclei of light atoms to create plasma comprised of ionized particles to millions of degrees, however, the plasma created is simply too hot and unstable for known materials to contain it. To help solve the problem, researchers designed a tokamak fusion reactor capable of confining the hot plasma to the center of a donut-shaped device comprised of powerful magnets
In a new Nature paper published this week, the team describes how they were able to use the cuDNN-accelerated TensorFlow deep learning framework and NVIDIA V100 GPUs to forecast sudden disruptions that can halt fusion reactions and damage the tokamaks that house the reactions.
The deep learning code called the Fusion Recurrent Neural Network was trained on more than two terabytes of data.
“Nuclear fusion power delivered by magnetic-confinement tokamak reactors holds the promise of sustainable and clean energy,” the researchers wrote in their paper.
The scientists are leveraging the power of NVIDIA GPUs on the Summit and Titan supercomputers at the Oak Ridge National Laboratory in Tennessee.
“Training deep neural networks is a computationally intensive problem that requires the engagement of high-performance computing clusters,” said Alexey Svyatkovskiy, a coauthor of the Nature paper who helped convert the algorithms into production code and now is at Microsoft. “We put a copy of our entire neural network across many processors to achieve highly efficient parallel processing,” he said.
The software can predict disruptions within the 30-millisecond time frame the new multibillion-dollar international Thermonuclear Experimental Reactor (ITER) will require.
Related Video: Fusion Research at the Summit Supercomputer
“ITER aims to be the first reactor that produces more power from fusion than is injected to heat the plasma.,” the researchers said. “These initial results illustrate the potential for deep learning to accelerate progress in fusion-energy science and, more generally, in the understanding and prediction of complex physical systems.”