发表机构
Princeton Plasma Physics Laboratory; University of Maryland, College Park; Oak Ridge National Laboratory; Type One Energy; California Institute of Technology; Argonne National Laboratory; Princeton University; University of Wisconsin, Madison（普林斯顿等离子体物理实验室; 马里兰大学帕克分校; 橡树岭国家实验室; Type One Energy; 加州理工学院; 阿贡国家实验室; 普林斯顿大学; 威斯康星大学麦迪逊分校）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文利用AI替代模型加速仿星器湍流输运预测，实现磁平衡优化与输运求解加速，并通过AI智能体自动探索目标与超参数，发现改进的湍流优化磁平衡。
AI 中文摘要
先前的工作为非线性回旋动理学模拟代码构建了基于AI的替代模型，其目标是将其用于仿星器设计优化和实验情景规划中湍流离子热通量的快速直接计算。这些AI替代模型在超过20万次非线性、绝热电子回旋动理学模拟的数据上进行了训练，这些模拟使用回旋动理学通量管代码GX，涵盖了广泛的仿星器磁位形（约2.3万种）、等离子体中的位置以及梯度尺度长度。在本文中，我们展示了基于AI的湍流替代模型在仿星器磁平衡优化和加速仿星器输运求解器中的应用。由于其速度（约毫秒级），基于AI的替代模型使得以前无法实现的目标成为可能，例如完整的离子湍流热通量径向剖面，或直接优化以最大化等离子体多个位置的湍流临界梯度。这些直接计算提供了可能更准确的优化目标，并减少了对可能无法准确捕捉湍流输运随磁位形变化的临时启发式方法的依赖。通过将基于AI的湍流热通量替代模型纳入输运求解器，我们可以快速进行后处理并确认优化平衡所导致的改进的离子温度。最后，我们展示了使用具有强大推理能力的AI模型的AI智能体来自动化外层循环，利用这种基于AI的湍流替代模型探索许多目标和超参数配置，以发现改进的湍流优化的磁平衡。
英文摘要
Previous work built AI-based surrogates for a nonlinear gyrokinetic simulation code with the goal of using them for fast, direct calculations of turbulent ion heat flux in stellarator design optimizations and scenario planning for experiments. These AI surrogates were trained on data from >200k nonlinear, adiabatic electron gyrokinetic simulations with the gyrokinetic flux-tube code GX, using a wide range of stellarator magnetic configurations ($\sim$23k), positions in the plasma, and gradient scale lengths. In this paper, we demonstrate the use of the AI-based turbulence surrogate in the optimization of stellarator magnetic equilibrium and to speed up stellarator transport solvers. Due to its speed ($\sim$ms), the AI-based surrogate enables previously unattainable optimization objectives, such as full radial profiles of ion turbulent heat flux, or directly optimizing to maximize the turbulent critical gradient at multiple locations across the plasma. These direct calculations provide a potentially more accurate optimization target and reduce reliance on ad-hoc heuristics that may not accurately capture the variation of turbulent transport with magnetic configuration. By including the AI-based surrogate for turbulent heat flux in a transport solver, we can quickly postprocess and confirm the improved ion temperature resulting from the optimized equilibrium. Finally, we demonstrate the use of AI agents with strong reasoning AI models to automate the outer loop, exploring many objective and hyperparameter configurations with this AI-based turbulence surrogate to discover improved turbulence optimized magnetic equilibria.
