|
Description
|
Achieving sustainable fusion energy critically depends on accurately modeling complex plasma dynamics within tokamak reactors, particularly in the Scrape-off Layer (SOL), where heat and particles directly interact with reactor walls, influencing reactor performance and component longevity. Particle-in-Cell (PIC) simulations, although highly accurate, are computationally expensive and time-consuming, limiting their use for iterative design and real-time control. We employ an Extreme Gradient Boosting (XGBoost)-based surrogate model to efficiently predict plasma potential along the tokamak SOL using data from PIC simulations under varying operating conditions. The machine learning (ML) approach integrates physics-informed segmentation of the spatial modeling domain, distinguishing sharply between sheath regions and the quasineutral bulk plasma. This segmentation substantially enhances the surrogate model’s predictive accuracy to localized physical phenomena, a marked improvement over traditional global modeling strategies. We utilize XGBoost regression with hyperparameter optimization achieved through a tailored leave-one-curve-out (LOCO) validation method, ensuring robust generalization to previously unseen plasma conditions. We found that a global model with segmented consideration of the spatial dimension based on boundary plasma physics captures localized behaviors more accurately. This segmented approach leads to a mean absolute percentage error (MAPE) of 3.2%, outperforming other methods. The main engineering application of this approach is the significant reduction in computational resources and simulation time required for fusion reactor design and real-time plasma control. This allows rapid iterative design, improved operational decision-making, and potentially extends the operational lifetime of reactor components. (2026-02-27)
|
|
Keyword
|
Tokamak, Surrogate model, Hyperparameter, Extreme learning machine, Dimensionality reduction, Plasma, Segmentation, Boosting (machine learning) |
