The Plasma Problem That Physics Couldn't Solve Alone

Nuclear fusion runs on a paradox. To release energy, you need plasma hotter than the sun’s core. To contain plasma that hot, you need a magnetic field strong enough and stable enough to hold it indefinitely. The tokamak, a donut-shaped reactor chamber, is the best design physicists have come up with for doing that. It still isn’t good enough.

The failure point has a name: tearing modes. These are magnetic instabilities that form at specific internal surfaces within the plasma column, places where the forces holding the plasma steady and the forces pulling it apart happen to be roughly equal. When something elsewhere in the rotating plasma shifts abruptly, that balance breaks. Magnetic field lines reorganize. The tokamak’s rotational symmetry, which the entire confinement approach depends on, collapses. In the worst cases, a magnetic bubble expands inside the plasma, kills its rotation, and drives the plasma into the reactor wall.

Researchers Cristina Rea and Stuart Benjamin have been looking at what machine learning can do where physics has stalled. Their work covers three fronts: predicting tearing mode onset before it happens, extracting clearer physical insight from experimental data, and designing plasma controllers that act on AI-generated stability forecasts in real time.

The appeal of machine learning here is structural. Tearing modes are nonlinear, chaotic, and coupled to processes happening across the whole plasma column simultaneously. Classical models can describe the end state well enough: a magnetic bubble grows, rotation stops, confinement fails. What they can’t do reliably is predict when that sequence will begin, or catch the early signals in time to intervene. A well-trained model, fed enough data from large experimental tokamaks, can.

The stakes are straightforward. Tearing modes that go uncontrolled in a research device cost an afternoon. In a commercial fusion power plant, they cost far more. Getting AI into the control loop isn’t a refinement of fusion science. It’s a requirement for fusion power to function at scale.