PLASMA CONTROL / SIMULATION GAME

Hold the edge.

Raise pressure for a higher score. Stay below the stability limit. Hold the shot for six simulated seconds.

Modelled on the 2026 PACMAN experiments at DIII-D, where a machine-learning controller ran the heating systems and one model saw a tearing instability coming 200 ms before it formed.

Start with the slow shot. Move the heating slider right, then pull it left when the risk forecast rises.

Custom pilot and speed
Pilot
Speed
  1. Start the shot. In practice, six simulated seconds take about 60 real seconds.
  2. Move the heating slider. Right raises pressure and earns points faster. Left reduces pressure.
  3. Move the shape slider. Right raises the stability limit, but acts slowly and reduces heating efficiency.
  4. Watch the risk forecast. Reduce heating when risk rises or your stability margin shrinks. A disruption ends the round early.
3D toroidal schematicReady
ILLUSTRATIVE GEOMETRY / SIMULATED SIGNALS
Your shot is ready.Start the shot, then move the heating and shape sliders.
Reduce heating when the risk forecast rises.
0.000 sof 6.000
Nobody at the controls
Plasma pressure βN
0.00
Stability limit — · margin —
Tearing forecast, next 200 ms0%
quietcautionat the limit
Tearing mode amplitude0%
noise floorlocks at 100%

Right: more pressure & points. Left: reduce pressure.

Right: more triangular, higher stability limit. Responds slowly.

Score0.0βN·s
Drag vertically on the plasma view to adjust heating.
Keys: ↑↓ heating, ←→ shape, space start or end the shot.

Shot length—
Peak βN—
Performance—
Pressure βNThe dashed line is the stability limit as the forecast model sees it, 200 ms ahead.
Tearing modeWhat the magnetic probes on the wall pick up. Locks at the red line.
ActuatorsBeam power and triangularity as a share of their range. Ticks along the bottom mark controller cycles.

Shot log

No shots yet. Try a practice shot at 1/10 speed, then real time, then the controller. Guided stages repeat seed 11; custom shots vary.

ShotPilotSpeedResultLengthPeak βNβN·sYour latency
Nothing flown yet.

The clock

  • A person's simple reaction to a visual cue takes about a quarter of a second. Choosing between actions takes longer.
  • PACMAN completes a control cycle in about 20 ms. The reinforcement-learning heating controller in the paper ran at 50 ms.
  • In one experiment a model saw a tearing mode coming about 200 ms ahead, and the plasma was changed so it never formed.
  • Plasma pressure in DIII-D responds to heating on the energy confinement time, roughly 100 ms. Cutting power late does not cut pressure quickly.
  • Once a tearing mode starts growing it can lock and end the discharge within tens of milliseconds.

By the time a person has noticed the forecast, decided, and moved a hand, the window in which pressure could still have been brought down has usually closed.

What is real

DIII-D is the US Department of Energy's tokamak in San Diego. PACMAN, built at the Princeton Plasma Physics Laboratory and Princeton University, ran five experiments there: a reinforcement-learning model took full control of the heating systems, other models held density and rotation at targets, one predicted edge bursts, and one predicted and avoided a tearing mode. Hardware safety limits were enforced in every case and people set the goals.

The tearing-avoidance idea comes from a 2024 Nature paper by the same group, where a reinforcement-learning controller adjusted beam power and plasma shape to keep the tearing likelihood under a threshold, including in the ITER baseline scenario.

Normalised pressure βN, triangularity, the 100 ms confinement time, and sawtooth crashes seeding tearing modes are all real vocabulary and real scales.

What is a toy

Four equations, not a plasma. The stability limit here is a scripted curve with dips the forecast can see coming; a real predictor is a neural network reading dozens of diagnostics, and it is sometimes wrong.

The controller is a model-predictive rule that reads the same forecast you do and acts on it every 20 ms. It stands in for the reinforcement-learning policy, which is a trained network, not a rule.

The mode's rotation is slowed roughly a thousand times so you can see it. Performance is measured as βN·seconds, a proxy for confined energy, not fusion power.