Researchers have developed machine learning surrogate models to predict local, linear gyrokinetic stability in the pedestal region of spherical tokamak plasmas, according to a new preprint. The work, focused on a parameter space relevant to the MAST-U device, aims to create computationally inexpensive tools that can replace or augment reduced transport assumptions in integrated pedestal modeling workflows. By training models on data from the high-fidelity gyrokinetic code GENE, the effort seeks to provide faster, physics-based inputs for predictive models like EPED, which currently rely on simpler assumptions such as the ballooning-critical pedestal limit. Source: arXiv
The performance of high-confinement-mode (H-mode) plasmas is strongly influenced by the pressure gradient achievable in the pedestal, a narrow region of steep gradients at the plasma edge. This region is susceptible to microinstabilities that drive turbulent transport, limiting pedestal height and fusion performance. While gyrokinetic codes can accurately model this transport, their computational expense prohibits their routine use for large-scale parameter scans or in integrated modeling frameworks that require rapid iteration. This computational bottleneck has led to the widespread use of reduced models that often simplify the underlying physics. Source: arXiv
This region is susceptible to microinstabilities that drive turbulent transport, limiting pedestal height and fusion performance.
A key aspect of the new methodology is a novel data generation workflow designed to reduce the dimensionality of the problem while maintaining physical consistency. Instead of sampling a wide range of local gyrokinetic inputs directly, the team varied higher-level pedestal profile parameters within experimentally motivated bounds. For each sample, a physically self-consistent Grad-Shafranov equilibrium was generated. This ensures that the combinations of plasma profiles, geometry, and local stability parameters used for training the surrogate models are physically plausible, a critical step for creating robust and reliable predictive tools for fusion plasma modeling. Source: arXiv
The resulting surrogate models are trained to predict the linear growth rates and frequencies of the dominant microinstabilities as a function of the input profile parameters. By capturing the output of the complex GENE simulations in a fast-executing model, this approach could significantly accelerate the exploration of pedestal stability. Such a tool would enable more comprehensive analysis of fusion science by allowing for rapid assessment of how different operational scenarios and plasma shapes impact pedestal turbulence, moving beyond the limitations of current simplified models. The ultimate goal is to integrate these faster, gyrokinetic-based predictions into operational and design workflows for future fusion devices. Source: arXiv