Density profile reconstruction with PIDP-KAN model training based on polarimeter–interferometer measurement on EAST

Abstract Plasma density and current profile are key parameters for studying plasma confinement and transport in tokamak devices. In the Experimental Advanced Superconducting Tokamak (EAST), the polarimeter–interferometer (POINT) system has been used to provide simultaneous and effective measurement of density and current for plasma discharges. The traditional way to reconstruct profiles is based on the POINT diagnostics model and magnetic surfaces. In this paper, a novel Kolmogorov–Arnold network (KAN)-based deep neural network (NN) model capturing nonlinear relationships using spline functions is proposed to reconstruct plasma density, the point density prediction KAN (PIDP-KAN). Unlike traditional fully connected networks multilayer perceptrons (MLPs) with fixed activation functions, PIDP-KAN employs learnable spline functions inspired by the Kolmogorov–Arnold theorem, enabling adaptive fitting of nonlinear relationships through B-spline parameterization. This architecture eliminates linear weight matrices, reducing parameter counts while achieving higher accuracy and real-time processing. To our knowledge, this is the first application of the KAN to plasma diagnostics, demonstrating its potential for complex physical systems. The PIDP-KAN network exhibits outstanding performance in reconstructing plasma density profiles in EAST, achieving errors less than 5% and a processing time response up to 2 ms, which demonstrates the model’s accuracy and reliability. In the future, it is expected to achieve real-time distribution inversion, providing a stable diagnostic foundation for EAST and future fusion reactors.