标题： EFIT-mini：一种嵌入式、多任务神经网络驱动的平衡反演算法 显示英文标题

标题： EFIT-mini: An Embedded, Multi-task Neural Network-driven Equilibrium Inversion Algorithm

摘要： 平衡重构通过诊断和线圈电流数据推断托卡马克中的内部磁场、等离子体电流和压力分布，对于受控磁约束核聚变研究至关重要。 然而，传统数值方法由于计算耗时或迭代收敛问题，往往难以满足实时控制需求。 本文介绍了EFIT-mini，一种将机器学习与数值模拟相结合的新算法。 它采用多任务神经网络来替代数值平衡反演中的复杂步骤，如磁面边界识别，结合了两种方法的优势，同时减轻了各自的缺点。 神经网络处理线圈电流和磁测量数据，直接输出等离子体参数，包括$p'$和$ff'$的多项式系数，为后续的Picard迭代提供高精度初始值。 与现有的人工智能驱动方法相比，EFIT-mini引入了更多的物理先验条件（例如最小二乘约束），以提高反演精度。 在EXL-50U托卡马克放电数据上验证，EFIT-mini在最后闭合通量面面积上与传统方法的重叠度超过98%。 此外，EFIT-mini的神经网络和完整算法分别在129$\times$129分辨率下，仅需0.11ms和0.36ms计算单个时间片，速度提升了三个数量级。 这种创新方法利用了机器学习的速度和数值算法的可解释性，为实时等离子体形状控制提供了稳健的解决方案，并有望扩展到动力学平衡重构。 其高效性和通用性使EFIT-mini成为托卡马克实时监测和控制的有前途工具，以及为其他实时反演算法提供关键输入的工具。 显示英文摘要

摘要： Equilibrium reconstruction, which infers internal magnetic fields, plasmas current, and pressure distributions in tokamaks using diagnostic and coil current data, is crucial for controlled magnetic confinement nuclear fusion research. However, traditional numerical methods often fall short of real-time control needs due to time-consuming computations or iteration convergence issues. This paper introduces EFIT-mini, a novel algorithm blending machine learning with numerical simulation. It employs a multi-task neural network to replace complex steps in numerical equilibrium inversion, such as magnetic surface boundary identification, combining the strengths of both approaches while mitigating their individual drawbacks. The neural network processes coil currents and magnetic measurements to directly output plasmas parameters, including polynomial coefficients for $p'$ and $ff'$, providing high-precision initial values for subsequent Picard iterations. Compared to existing AI-driven methods, EFIT-mini incorporates more physical priors (e.g., least squares constraints) to enhance inversion accuracy. Validated on EXL-50U tokamak discharge data, EFIT-mini achieves over 98% overlap in the last closed flux surface area with traditional methods. Besides, EFIT-mini's neural network and full algorithm compute single time slices in just 0.11ms and 0.36ms at 129$\times$129 resolution, respectively, representing a three-order-of-magnitude speedup. This innovative approach leverages machine learning's speed and numerical algorithms' explainability, offering a robust solution for real-time plasmas shape control and potential extension to kinetic equilibrium reconstruction. Its efficiency and versatility position EFIT-mini as a promising tool for tokamak real-time monitoring and control, as well as for providing key inputs to other real-time inversion algorithms.