Keeping a fusion plasma under control requires decisions to be made extraordinarily quickly. Inside a tokamak, disturbances can develop within milliseconds, giving human operators little chance to react before an instability begins affecting the experiment.
Researchers at Princeton University and the US Department of Energy’s Princeton Plasma Physics Laboratory (PPPL) have now tested an artificial intelligence framework designed to operate at those timescales. Called PACMAN, the system can continuously monitor a fusion plasma, predict how it is likely to behave and make adjustments in milliseconds.
The technology has already been tested on a real fusion machine rather than remaining purely in simulation. Researchers ran PACMAN during five experiments at the DIII-D National Fusion Facility tokamak in San Diego, where it was used for tasks ranging from controlling plasma heating to predicting and preventing an instability before it appeared.
Fusion machines such as tokamaks use powerful magnetic fields to confine plasma at extremely high temperatures. For a fusion reaction to be maintained, that plasma needs to remain hot, dense and stable, requiring constant adjustments to equipment including magnets, heating systems and gas injectors.
Advertisement
Traditional computer simulations can model what the plasma is doing in considerable detail, but some take days or even months to complete. That makes them useful for planning experiments, but not for responding to an instability developing in milliseconds.
PACMAN — short for Prediction And Control using MAchiNe learning — instead combines several machine-learning models inside a repeating control loop. The framework typically completes that loop in around 20 milliseconds, compared with a response time measured in seconds for even a highly focused human operator.
The process begins by collecting live information from the tokamak, including its temperature, density and magnetic signals. AI models analyse those measurements to estimate what the plasma is doing and predict what could happen next. Controllers then determine the appropriate response before safety checks are applied and commands are sent back to the machine.
One experiment demonstrated why that speed could matter. PACMAN predicted a potentially damaging plasma instability known as a tearing mode about 200 milliseconds before it developed. Rather than attempting to suppress the instability after it had already begun, the system altered the plasma in advance to prevent it from forming.
During other experiments, the framework controlled heating systems using reinforcement learning, predicted bursts of energy from the plasma edge and adjusted plasma density and rotation. It also coordinated all six of DIII-D’s gyrotrons, which use powerful microwave beams to heat the plasma, adjusting their power and mirror positions in real time to reach targets set by researchers.
Advertisement
Despite its speed and autonomy, PACMAN is not intended to remove people from the control room. Researchers still set the objectives, while the framework applies hardware safety limits regardless of what an individual AI model recommends. Scientists then analyse each experiment and refine the controllers before subsequent tests.
The wider significance may be PACMAN’s modular design. Different AI models can be added, replaced or operated simultaneously without rebuilding the entire control system, potentially allowing the framework to be adapted for other tokamaks and future fusion machines.
Fusion still faces enormous scientific and engineering challenges before it can become a practical source of electricity. But as experiments become increasingly complex, controlling them may require decisions on timescales that humans simply cannot match. PACMAN demonstrates how AI could begin filling that gap, reacting to problems in the plasma before a person has even had time to see them.