EAST plasma equilibrium reconstruction based on weight-adaptive physics-informed neural network

This paper introduces a weight-adaptive physics-informed neural network (PINN) for plasma equilibrium reconstruction in the EAST tokamak, overcoming the reliance of conventional approaches on predefined current profile assumptions. The proposed framework incorporates the Grad–Shafranov (G–S) equation as a physical constraint and integrates external magnetic diagnostics—such as flux loops and magnetic probes—to jointly exploit both measured data and underlying physics. Three dedicated subnetworks are designed to predict the plasma flux, current profile functions ( p ′ and f f ′ ), and poloidal field coil currents, substantially improving reconstruction accuracy over earlier PINN-based attempts. A hybrid Adam-LBFGS optimizer combined with an adaptive loss weighting scheme ensures robust convergence and higher precision relative to standard PINN. When applied to experimental EAST data, the method demonstrates its capability to accurately match diagnostic measurements while maintaining a low physics loss residual, underscoring the effectiveness of the weight-adaptive strategy. Furthermore, accurate equilibrium reconstruction throughout three full discharges with different shapes confirms the robustness and versatility of the method across diverse plasma configurations. These results demonstrate that the proposed PINN framework can provide a reliable and effective alternative for analyzing EAST discharges. • A three-subnetwork PINN architecture is proposed for integrated tokamak equilibrium reconstruction. • A weight-adaptive loss balancing strategy is developed to robustly handle multi-task constraints. • The method accurately reconstructs plasma equilibria using experimental data from the EAST tokamak. • The framework demonstrates robust performance across the entire discharge duration.