Machine Learning for Fusion: An Overview
How data-driven models complement physics simulation across control, prediction, reconstruction, and design in magnetic-confinement fusion.
Why fusion invites machine learning
Magnetic-confinement plasmas are high-dimensional, nonlinear, and only partially observed. First-principles simulation of turbulence and transport is expensive, and many operational decisions must be made in milliseconds. Machine learning (ML) fills gaps where physics models are too slow, too incomplete, or too costly to evaluate in real time. It does not replace physics; it interpolates within regimes where data are dense and flags where extrapolation is unsafe.
Where ML is applied
- Disruption prediction and avoidance from diagnostic time series
- Plasma profile and equilibrium reconstruction
- Surrogate models for transport, turbulence, and heating
- Reinforcement-learning control of coils and actuators
- Anomaly detection on diagnostic streams
- Data-driven scaling laws and regime classification
- Accelerating or emulating expensive simulations
The honest caveats
Fusion datasets are small by modern ML standards, imbalanced (disruptions are rare), and drawn from a handful of devices. Models trained on one machine rarely transfer cleanly to another, and future reactor regimes lie outside any existing dataset. Good practice treats ML predictions as decision support with quantified uncertainty, validated against held-out shots and, wherever possible, against physics.
Design versus hardware
At Kronos the breeder (Hyperion, a D-T spherical tokamak) and the burner (a D-3He tandem-mirror generator, in the Aegis and MetroVolt housings) are design and simulation efforts. Any ML discussed here informs modeling and control studies, not claims about built hardware. No hardware net-gain claim is made before first-of-a-kind first tritium.
The pages in this section describe general, published, and reproducible techniques. Read them as method, not as performance guarantees on any specific device.