One of the biggest problems with the development of nuclear fusion reactors based on the tokamak design has to do with the appearance of major disruptions in the control of the plasma in the donut shaped chamber. When these disruptions occur, the fusion reaction can stop and the walls of the containment vessel can be damaged.
Researchers at the U.S. Department of Energy’s Princeton Plasma Physics Laboratory (PPPL) and Princeton University are working on using artificial intelligence to predict the occurrence of disruptions in the plasma. The research group is developing these predictive computer models for the ITER project in France, an international cooperative project intended to demonstrate the practicality of nuclear fusion for energy production.
The new predictive software is called the Fusion Recurrent Neural Network (FRNN). It is based on the AI deep learning process which is a new powerful means of training computers in pattern recognition. The research team at Princeton is the first to apply a deep learning technique to the problem of predicting disruptions in tokamak fusion plasmas. FRNN is the best way to analyze sequential data with long-range patterns that has been developed to date.
The U.K. is host to the Joint European Torus (JET) project which is the biggest operational tokamak fusion reactor in the world. It is managed by EUROfusion, the European Consortium for the Development of Fusion Energy. The Princeton team was able to make use of the huge database at JET to improve predictions of disruptions and reduce the number of false alarms.
The next goal of the research team is to develop their system to meet the difficult challenges that face the ITER project. One of those challenges is to develop a monitoring system that can generate ninety-five percent correct predictions of real disruptions while producing less than three percent false alarms. One researcher at the PPPL lab said, “On the test data sets examined, the FRNN has improved the curve for predicting true positives while reducing false positives.” “We are working on bringing in more training data to do even better.”
A big data expert at Princeton said, “Training deep neural networks is a computationally intensive task that requires engagement of high-performance computing hardware.” “That is why a large part of what we do is developing and distributing new algorithms across many processors to achieve highly efficient parallel computing. Such computing will handle the increasing size of problems drawn from the disruption-relevant data base from JET and other tokamaks.”
The new software was initially developed and tested on a massive parallel array of graphic processor units (GPUs) at Princeton called the Tiger cluster. It has been run successfully on other bigger GPU clusters in the U.S. and abroad. The performance of the software scales well as the number of GPUs in the cluster increases.
The research team hopes to demonstrate that their software can run successfully on operational tokamaks here and around the world. They are also working on increasing the speed of disruption analysis for the huge data sets they will be working with on the operational tokamaks.
