The PACMAN system operates on a pipeline consisting of four stations: raw diagnostic data collection (temperature, density, magnetic parameters), measurement error filtering, ML-based fuel behavior prediction, and the conversion of forecasts into hardware commands with conflict resolution and safety limit compliance.
The modular design allows individual algorithms to be modified or replaced within days without affecting other parts of the solution. Previously, each ML project for nuclear fusion was developed from scratch.
While traditional plasma physics modeling can take several months, ML models generate predictions in fractions of a second. Human operators still oversee the process by setting goals, adjusting parameters between pulses, and relying on strict limits to prevent equipment damage.
The developers state that the platform is easily portable to other tokamaks, providing an adaptable control standard for future commercial installations.
