Fusion: E-learning reduces prediction error on DIII-D tokamak by 80%

Data scientists at the Jefferson Lab have designed a digital twin that anticipates the micro-displacements of the DIII-D tokamak's coils between two shots, with 80% less error compared to static models.

Inside the DIII-D tokamak near San Diego, a plasma hotter than the heart of the Sun is compressed and stirred in an intense magnetic field. This environment ends up imperceptibly deforming the machine itself: the toroidal field coils, although designed to be rigid, move in a tiny way from one experimental shot to the next, at a rate of one shot every ten minutes or so.

Data scientists at the Thomas Jefferson National Accelerator Facility, in partnership with the University of Houston, General Atomics and the Pacific Northwest National Laboratory, have built a digital twin based on deep neural networks to predict these movements before they disrupt experiments. The special feature of the device is that it uses online learning: the models are continuously updated as new data arrives, instead of being frozen after training.

A first on drifting data flows

The gain is clear: 80% less prediction error compared to static models, with an additional 10% obtained by integrating uncertainty quantification. The authors present this work, published in Machine Learning with Applications, as the first application of e-learning to drifting data streams in fusion research. For lead author Kishan Rajput, the goal is to enable teams to spot problems before they manifest themselves on the physical machine.