A new artificial intelligence framework developed at the Princeton Plasma Physics Laboratory (PPPL) and Princeton University demonstrates the ability to make critical plasma control decisions in milliseconds, a speed essential for preventing damaging disruptions in tokamak fusion devices. The system, named PACMAN for Policy-based Automatic Control of Man-made and Natural systems, uses a novel multi-fidelity approach to balance computational speed with predictive accuracy. The research, led by PPPL scientist Azarakhsh Jalalvand, was detailed in a peer-reviewed paper published in *Nature Communications*. Source: PPPL
PACMAN's core innovation is its dynamic use of computational models. The framework continuously runs a fast, simplified plasma model to monitor for potential instabilities. When this low-fidelity model flags a high-risk state, PACMAN activates a more complex and computationally intensive high-fidelity model. This second model provides a more precise prediction, allowing the system to confirm the threat and determine an appropriate corrective action, such as adjusting heating systems or injecting gas to stabilize the plasma. This hierarchical method avoids the prohibitive computational cost of running a high-fidelity simulation in real-time for the entire duration of a plasma shot. Source: PPPL
The framework continuously runs a fast, simplified plasma model to monitor for potential instabilities.
The system was validated using historical experimental data from the DIII-D National Fusion Facility operated by General Atomics. By feeding PACMAN data from past DIII-D experiments, the researchers demonstrated its ability to foresee and react to plasma instabilities far faster than a human operator. Current control systems in tokamaks often rely on pre-programmed responses and lack the flexibility to manage unexpected events, which can lead to plasma disruptions that release intense heat and electromagnetic forces capable of damaging the reactor's internal components. Autonomous, rapid-response systems are considered a critical enabling technology for future steady-state fusion power plants. Source: PPPL
The development of advanced control systems is a major focus area across the global fusion research community, including for large-scale projects like ITER. The PACMAN framework represents a significant step in applying modern AI techniques to the longstanding challenge of plasma stability. According to lead author Jalalvand, the system's ability to switch between models based on risk assessment is key to its efficiency. The next phase for the research team involves testing the PACMAN software on an operational fusion device, with plans to deploy it at DIII-D. The framework's architecture is also being considered for applications beyond fusion, including the control of drone swarms and the management of electrical power grids. Source: PPPL