Chapter 3
Confining a 100-million-degree plasma is one of the most complex control challenges in engineering. AI and machine learning are revolutionizing how we stabilize fusion plasmas.
A fusion plasma is a turbulent, nonlinear system with thousands of coupled variables evolving on timescales from microseconds to minutes. The plasma must be kept hot enough for fusion (100 million °C), dense enough for reactions (10²⁰ particles/m³), and confined long enough for net energy production (seconds). Simultaneously, it must be kept away from the reactor walls, which would be destroyed by direct contact with the plasma.
Traditional control systems struggle with this complexity. The plasma responds to control inputs in ways that depend on its current state, past history, and subtle interactions between different physical phenomena. What works in one discharge may fail in the next.
Plasmas are naturally unstable. Like a ball balanced on a hilltop, small perturbations can grow rapidly. Understanding and controlling these instabilities is central to fusion research.
Magnetohydrodynamic (MHD) instabilities occur when the plasma pressure or current distribution exceeds stability limits. They are characterized by mode numbers (m,n) describing their spatial structure.
| Mode | Description | Consequence |
|---|---|---|
| Kink (m=1, n=1) | Helical deformation of plasma column | Internal disruption, sawtooth crashes |
| Tearing (m=2, n=1) | Magnetic island formation | Confinement degradation, locked modes |
| Ballooning | Pressure-driven bulging at outer edge | ELMs, transport |
| Vertical Displacement | Plasma moves up or down | Wall contact, disruption |
ELMs are quasi-periodic bursts at the plasma edge that expel particles and energy. While they help remove impurities, large ELMs can damage plasma-facing components. ITER requires ELM mitigation to protect its divertor.
A disruption is a sudden, uncontrolled loss of plasma confinement. In milliseconds, all the plasma's stored energy (up to 350 MJ in ITER) is dumped onto the reactor walls and into electromagnetic forces on the structure. Disruptions can:
For ITER and future power plants, disruptions must be either avoided entirely or mitigated to prevent damage. This is where AI becomes essential.
Machine learning is transforming fusion plasma control. Deep neural networks can learn complex relationships between plasma parameters that are impossible to model from first principles. Real-time AI systems now operate on tokamaks worldwide.
AI models trained on thousands of past plasma shots can predict disruptions hundreds of milliseconds in advance—enough time to initiate mitigation. Modern systems achieve >95% detection rates with <5% false positives.
AI systems now control plasma shape in real-time, adjusting magnetic coil currents to maintain the desired equilibrium. DeepMind's work with the TCV tokamak demonstrated that reinforcement learning can discover novel plasma configurations that human operators had not considered.
Machine learning optimizes the mix of heating systems (NBI, ECRH, ICRH) to maximize performance while respecting constraints on wall heat loads and stability limits.
When a disruption cannot be avoided, mitigation systems limit the damage:
| System | Mechanism | Purpose |
|---|---|---|
| Massive Gas Injection (MGI) | Inject 10²³-10²⁴ atoms of Ar, Ne, or D₂ | Radiate energy, reduce wall heat load |
| Shattered Pellet Injection | Launch frozen pellets that shatter on entry | Faster mixing, better radiation distribution |
| Runaway Electron Mitigation | Inject high-Z materials to scatter runaways | Prevent localized wall damage |
A modern tokamak control system operates at multiple timescales:
AI systems are being integrated at all levels, with the greatest impact currently in prediction (disruption forecasting) and optimization (heating/fueling). Future systems will use AI for direct closed-loop control.