Researchers have developed an artificial intelligence surrogate model capable of predicting turbulent ion heat flux in stellarators in milliseconds, a task that typically requires substantial high-performance computing resources. The model, detailed in a preprint, was trained on a dataset of over 200,000 nonlinear, adiabatic electron gyrokinetic simulations performed with the gyrokinetic flux-tube code GX. This dataset spanned approximately 23,000 unique stellarator magnetic configurations, covering a wide range of plasma positions and gradient scale lengths. The surrogate's speed enables its direct integration into stellarator design optimization loops, a computationally prohibitive step with traditional simulation methods. Source: arXiv
The primary challenge in designing high-performance stellarators is mitigating turbulent transport, a dominant mechanism of heat loss that degrades plasma confinement. Traditionally, designers rely on proxy physics heuristics, such as optimizing for neoclassical transport or magnetohydrodynamic stability, with the assumption that these will indirectly lead to low turbulence. However, these proxies do not always correlate well with the turbulent transport that ultimately governs performance. Direct simulation of turbulence using gyrokinetic codes like GX is the most accurate method, but the computational cost makes it impractical for the thousands of iterative evaluations required in a typical design optimization workflow. Source: arXiv
However, these proxies do not always correlate well with the turbulent transport that ultimately governs performance.
The AI surrogate directly addresses this computational bottleneck. By replacing the full GX simulation within the optimization loop, the model allows for the direct calculation of turbulent transport metrics across the plasma's full radial profile. This enables novel optimization objectives that were previously unattainable, such as maximizing the turbulent critical gradient—the point at which turbulence is suppressed—at multiple locations. This direct approach provides a more physically accurate optimization target compared to indirect proxies, potentially leading to magnetic confinement designs with superior performance characteristics. The model's rapid evaluation time of approximately one millisecond per query is a key enabler for this shift in design philosophy. Source: arXiv
This work represents a significant step in applying machine learning to accelerate the complex physics simulations central to fusion energy development. While this paper focuses on stellarator optimization, the methodology of creating fast, accurate surrogates for computationally expensive physics codes has broad applicability across different fusion confinement concepts. The next steps will involve validating the surrogate's predictions against experimental data and integrating it into comprehensive transport solvers for scenario planning. The ability to rapidly and directly optimize for turbulence suppression could significantly shorten the design-build-test cycle for next-generation stellarator experiments and power plants. Source: arXiv