A fusion-AI benchmark delivers a winner and a warning about scaling up

A fusion reactor is, in the most literal sense, a container for plasma hotter than the core of the sun. The software that figures out where that plasma sits, what shape it holds, and whether the magnetic fields are still confining it has to do its math thousands of times per second, because the moment the math falls behind, the plasma falls apart, and the machine eats its own walls.

The Grad-Shafranov equation describes the equilibrium shape of the hot gas inside a tokamak, a donut-shaped reactor that uses magnetic fields to hold fusion fuel as a thin ribbon of plasma. Solving the equation directly takes long enough that today's plasma-control loops cannot run it on the fly. Which is why a new benchmark paper on arXiv puts five neural networks in the operator's seat and asks which one is good enough to do the work in time.

The candidates: a multilayer perceptron (MLP), a convolutional neural network (CNN), a Fourier neural operator (FNO), a Transformer, and a Kolmogorov-Arnold network (KAN). Each was trained on a numerical database of 100,000 in-distribution Grad-Shafranov solutions, the plasma shapes the model was allowed to learn from, plus 10,000 out-of-distribution (OOD) solutions, plasma geometries the model had never seen, the conditions a real reactor will eventually throw at it. In machine-learning terms, in-distribution is the textbook; out-of-distribution is the exam the textbook did not prepare you for.

CNN comes out the most useful for a real-time control loop. The authors report it runs in about 0.7 milliseconds with TensorRT acceleration on the same hardware, fast enough to fit inside a plasma-control cycle, while keeping relative error against the numerical solver in the $10^{-3}$ to $10^{-2}$ range. Transformer wins on raw in-distribution accuracy, but at a heavier compute cost that makes it a better fit for offline analysis, where the speed budget is not measured in milliseconds.

CNN and FNO are also the most stable when the plasma does something new. On out-of-distribution geometries, both held relative $L_2$ error to around 4%-5%. MLP and KAN degraded more sharply, and the gap between architectures widened as the plasma shapes drifted further from the training set. That is the test that matters for a real reactor, where the model will not be shown only the conditions it was trained on.

Scaling data and model capacity sharpens performance on plasma shapes the network has already seen, and does nothing, or actively hurts, performance on the unseen shapes a real reactor will eventually produce. The authors call it a capacity-OOD trade-off, and the device-level validation they ran on the EXL-50U tokamak in Chengdu shows the same pattern. The match between the numerical solver, the surrogate, and the machine's standard reference diagnostic stayed in the $10^{-3}$ range, so the AI predictions did not break the moment they touched real hardware.

For fusion-AI teams, the practical takeaway is concrete. For a real-time control loop today, CNN is the safest pick: accurate, fast, and the most stable when the plasma does something the training set did not cover. For offline analysis, where raw accuracy matters more than latency, Transformer is the better tool. And for anyone betting that throwing more compute at the problem will close the generalization gap, the benchmark has a warning. Bigger is not better. The hardest part of the problem, extrapolating to plasma shapes no one has shown the model, is still the part that does not respond to scale.

The paper is a preprint and has not been peer reviewed. The device-validation results are author-claimed against the EXL-50U machine, not independently corroborated.