AI framework cuts fusion hardware prediction errors by 80%

Inside DIII-D’s control room in San Diego, the clock between fusion experiments can leave operators roughly 10 minutes to understand what just happened. The machine’s plasma can reach temperatures hotter than the Sun’s core, and the large magnets around it can shift slightly from one “shot,” or experiment, to the next. A new machine-learning framework cut the error in predicting those hardware changes by 80 percent against a conventional static model.

DIII-D is a tokamak, a donut-shaped fusion machine that uses magnetic fields to confine superhot plasma. Its toroidal-field coils — the large magnets surrounding the plasma — are built to tight tolerances, but they are not perfectly still. As plasma stability changes, the coils can behave differently. A model trained once on historical data can miss that drift.

The researchers’ answer is online learning: new information is fed into the system continuously, allowing it to adjust as conditions evolve. They added an ensemble of models trained on different historical time horizons, so one can react to sudden changes while others capture slower movement. Each forecast also includes an uncertainty estimate, giving operators a signal of how much confidence to place in it.

The uncertainty-guided ensemble reduced prediction error by approximately another 10 percent compared with standard single-model online learning. The researchers describe the result as a digital twin — a virtual replica of the coil system that can be supplied with operating parameters to anticipate how the physical machine may respond. They say the framework is ready for deployment on DIII-D and could potentially be adapted to other fusion machines.

So what, concretely? Before a shot begins, operators could use the prediction to modify plasma parameters or investigate maintenance needs, rather than discovering a hardware issue only after it affects an experiment. That could make the short turnaround between shots more useful. The limits are clear: the system still needs to run on years of data, encounter rarer events, and become easier to interpret before researchers can fully judge how reliably it supports operations.