Neural-Network-Based Free-Boundary Equilibrium Solver to Enable Fast Scenario Simulations
Z. Wang, X. Song, T. Rafiq, E. Schuster
IEEE Transactions on Plasma Science, vol. 52, no. 9, pp. 4147-4153, Sept. 2024
Abstract
|
|
A numerical free-boundary equilibrium (FBE)
solver, based on finite-difference and Picard-iteration methods,
has been recently developed on a rectangular grid to compute
the poloidal-flux distribution in tokamaks. An accelerated version
of this computationally intensive numerical solver, named
FBE-Net, has been developed in this work by leveraging the
physics-informed neural network (PINN) method. Within this
framework, the neural-network (NN) component employs a
fully connected multilayer perceptron (MLP) architecture. Critically,
the underlying physical constraints are defined by the
Grad–Shafranov (G-S) equation, ensuring the NN-based solver
adheres to essential governing principles. FBE-net is trained on
a dataset generated by the numerical solver, which serves as a
source of ground truth. The inputs for FBE-Net are the plasma
current, the normalized beta, and the coil currents, while the
outputs are the poloidal-flux map and a set of flux-averaged equilibrium
parameters. When compared to the numerical solver, the
NN-based solver displays a significant increase in computational
efficiency without notably sacrificing accuracy.