Researchers have developed a data-driven model to predict the loss of enhanced energy confinement in pellet-fueled discharges within the Wendelstein 7-X stellarator. According to a preprint published on arXiv, the neural network model can forecast the remaining time before a plasma's enhanced confinement state deteriorates. The model demonstrated high accuracy, with at least 90% of its predictions correct to within 51 milliseconds. This predictive capability is designed to inform a real-time control system, enabling optimized timing for subsequent pellet injections to sustain high-performance plasma conditions. Source: arXiv

The energy confinement time is a critical performance metric in magnetic fusion devices, directly influencing the path to ignition. In both tokamaks and stellarators, injecting frozen fuel pellets can trigger a transition to a state of improved confinement. However, this enhancement is transient and typically decays over time. The new model addresses this by calculating a real-time deadline for the next pellet injection needed to maintain the high-confinement mode. By predicting the decay window, the system can avoid both premature injections that may be inefficient and late injections that would allow the plasma to revert to a lower-confinement state. Source: arXiv

The energy confinement time is a critical performance metric in magnetic fusion devices, directly influencing the path to ignition.

The model's predictive accuracy of 51 ms is a significant result because this interval is shorter than both the typical energy confinement time in Wendelstein 7-X and the minimum hardware-limited time separation between consecutive pellet injections. This ensures the model's predictions are actionable within the operational constraints of the device's fueling system. The rapid evaluation speed of the neural network further supports its suitability for integration into the plasma control loop, where low-latency decisions are paramount for maintaining plasma stability and performance. Source: arXiv

This work represents a step forward in applying machine learning techniques to complex control problems in fusion science. While previous efforts have focused on areas like disruption prediction in tokamaks, this model provides a specific tool for actively managing and sustaining an enhanced operational scenario in a stellarator. The successful application at W7-X could inform similar control strategies for other devices, including ITER and future power plants, where maintaining steady-state, high-performance operations will be essential for achieving a high engineering gain, or Q_engineering. The next phase will likely involve testing the model's performance in a live feedback loop during W7-X experimental campaigns. Source: arXiv