Inside a tokamak, a giant magnetic donut, plasma is heated to hundreds of millions of degrees. It doesn't touch the walls; it floats in a vacuum, held by invisible magnetic fields. But plasma is finicky: a slight change in current or pressure and its shape starts to wobble like jelly on a plate. To keep it in place, you need to know its exact position every millisecond.
Conventional programs are too slow for this. They solve complex equations, like an accountant recalculating every number by hand. Scientists found a faster way: they taught the computer to "feel" the plasma indirectly, through magnetic fields outside the chamber, and instantly infer its shape. For this, they used off-the-shelf machine learning tools, PyTorch and TensorRT, which speed up computations on graphics cards.
The new algorithm is called PTEFIT. It runs on the EXL-50U tokamak and takes just 0.268 milliseconds per frame — faster than a fly's wingbeat. That's enough to control the magnetic coils in real time: keep the plasma centered in the chamber and maintain the desired shape of the magnetic field lines.
This approach is like switching from manual driving to autopilot. Instead of solving equations from scratch every time, the program learns from examples and then acts instantly. This paves the way for more stable operation of future fusion reactors, where plasma must be controlled continuously, without pauses.
as written by the authors arXiv:2601.12378
This work proposes for the first time a novel optimization method for numerical algorithms, which takes advantages of machine learning frameworks PyTorch and TensorRT, leveraging their modularity, low development threshold, and automatic tuning characteristics to achieve a real-time plasmas reconstruction algorithm called PTEFIT as an application in tokamak-based controlled fusion that combines performance, flexibility, and usability. The algorithm has been deployed and routinely operated on the EXL-50U spherical tokamak, with an average inference time of only 0.268ms per time slice at $$129\times 129$$ resolution, and has successfully driven feedback control of the maximum radial position of plasmas and isoflux control. We believe that its design philosophy has sufficient potential to accelerate development and optimization in GPU parallel computing, and is expected to be extended to other numerical algorithms.
The temperature of plasma in a tokamak is several times higher than at the center of the Sun, but it is held not by walls but by a magnetic field, like an invisible bottle.