Nuclear Fusion - Deep Learning 12.29.17

Deep Learning Prediction Technology and Nuclear Fusion

Since research began in the 1940s, fusion reactors have been regarded as the future of energy, potentially providing virtually unlimited quantities of cheap electricity. However, after 75 years of development, the creation of a practical reactor remains tantalizingly elusive, with fusion experiments still consuming more energy than they produce. Today, experts believe they are on the verge of a fusion breakthrough, and deep learning prediction technology may play a critical role in this accomplishment.

Scientists at the U.S. Department of Energy Princeton Plasma Physics Laboratory are developing a deep learning prediction system called the Fusion Recurrent Neural Network (FRNN) that can forecast disruptions that can occur within fusion reactors. Such disruptions include the abrupt loss of control of the super-heated plasma used to trigger nuclear fusion. These events can bring fusion reactions to a sudden halt and may even damage the interiors of reactors.

By predicting such disruptions, the deep learning prediction system, the FRNN may be able to mitigate or avoid such disruptions, allowing fusion reactions to be sustained for longer periods of time.

“Deep learning represents an exciting new avenue toward the prediction of disruptions,” said Princeton physicist William Tang. “This capability can now handle multi-dimensional data.”

To train the FRNN, researchers are using the massive database generated by the Joint European Torus (JET) facility in the United Kingdom. The JET is the world’s largest and most powerful tokamak-type fusion reactor now in operation.

The deep learning code is run on graphics processing units (GPUs) that are capable of executing programs in parallel.

The researchers said this approach has allowed the FRNN team to predict disruptions with greater accuracy than previous approaches. They said it has cut down on the number of false positive alarms for the reactor. Eventually, the team hopes to achieve a 95 percent accuracy rate for predictions of disruptions.

The FRNN will be put to the ultimate test in the ITER, an international effort to build an experimental reactor that demonstrates the practicality of fusion technology. Located in France, ITER plans to commence its first plasma by 2025 and will start deuterium-tritium operation in 2035.

ITER is designed to produce 500 megawatts of power for a period of about 20 minutes. This would be approximately enough electrical energy to power 600,000 homes. The reaction would require only 50 megawatts to initiate, yielding a tenfold return on each watt of power invested into the ITER.

Tyler Schulze is vice president, strategy & development at Veritone.  He serves as general manager for developer partnerships, cognitive engine ecosystem, and media ingestion for the Veritone platform. Learn more about our platform and join the Veritone developer ecosystem today.