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

Abstract

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.