Speaker
Description
Optimizing divertor systems and controlling plasma exhaust are critical challenges for reactor-grade magnetic fusion devices such as ITER and future fusion power plants. Achieving these goals requires rapid, accurate modeling of boundary plasma behavior. Traditional high-fidelity edge plasma simulations, while scientifically valuable, are computationally intensive and were not designed for routine engineering tasks like real-time discharge control, scenario development, or divertor optimization. To overcome these limitations and accelerate fusion energy development, there is an urgent need for engineering-oriented modeling tools. In this work, we present two such models: a machine learning-based surrogate for real-time divertor detachment control, and a novel boundary transport model for rapid divertor optimization.
The first model, DivControlNN [1], is a data-driven surrogate designed for fast and reliable prediction of boundary and divertor plasma states using real-time control parameters and diagnostics. Trained on over 70,000 two-dimensional UEDGE simulations from KSTAR tokamak equilibria, DivControlNN employs latent space mapping [2] to efficiently capture the complex physics of divertor plasmas. This approach delivers a computational speed-up exceeding eight orders of magnitude over conventional simulations, while maintaining a relative error below 20% for key plasma parameters. As a result, it enables quasi-real-time predictions, reducing simulation times from hours to less than 0.2 ms per prediction within the Plasma Control System. A prototype detachment control system powered by DivControlNN was successfully deployed during the 2024 KSTAR experimental campaign [3]. It demonstrated robust detachment control on its first attempt, even with a new tungsten divertor configuration and without parameter fine-tuning, highlighting the model’s generalizability and immediate impact in experimental settings. The model’s predictions can be tailored to specific applications. For instance, DivControlNN forecasts plasma conditions at both the outboard midplane and divertor targets, as well as radiation information, providing key metrics for particle and heat exhaust and core plasma performance. These capabilities enable advanced integrated control systems and between-shot scenario development.
The second model aims to advance boundary transport codes suited for engineering applications by incorporating state-of-the-art numerical methods. By combining a flux-coordinate independent approach with immersed boundary techniques, this model can simulate the entire boundary plasma, from inside the separatrix to the first wall, providing critical heat and particle fluence data across all surfaces, independent of magnetic configurations. This flexibility enhances its utility for detailed divertor and blanket design, as well as integrated plasma-material modeling. Additionally, by leveraging differentiable programming, the model enables efficient sensitivity analysis and uncertainty quantification, which are essential for robust scenario development and divertor geometry optimization.
In summary, integrating machine learning surrogates like DivControlNN with new transport models marks a transformative step toward meeting the dual requirements of speed and accuracy in exhaust control and divertor design. These tools promise to accelerate the development of optimized, reliable solutions for plasma-facing components in future fusion reactors.
[1] B. Zhu, M. Zhao, et al., Physics of Plasmas, 32, 062508 (2025)
[2] B. Zhu, M. Zhao, et al., Journal of Plasma Physics, 88, 895880504 (2022)
[3] A. Gupta, D. Eldon, et al., Plasma Physics and Controlled Fusion (under review, 2025)
This work was performed for USDOE by LLNL under DE-AC52-07NA27344. LLNL-ABS-2007686
| Speaker's title | Mr |
|---|---|
| Speaker's Affiliation | Lawrence Livermore National Laboratory, Livermore, CA |
| Member State or IGO | United States of America |
