Neural Network-Based Free Boundary Equilibrium Solver for Control-Oriented Transport Simulations

Z. Wang, X. Song, T. Rafiq, E. Schuster

30th IEEE Symposium on Fusion Energy

Oxford, UK, July 9-13, 2023

Abstract

A neural network-based free-boundary equilibrium solver has been developed with the ultimate goal of being integrated into COTSIM (Control Oriented T ransport SIMulator). Such integration would enable fast performance assessment of integrated equilibrium-scenario feedback control algorithms as well as fast equilibrium-scenario planning by model-based feedforward-control optimization. To evolve transport equations such as the poloidal magnetic flux, heat, particle, and momentum transport equations, a 2D equilibrium solver must provide in advance flux-averaged parameters associated with the equilibrium configuration. A numerical free-boundary equilibrium (FBE) solver, based on finite-difference and Picard-iteration methods, has been developed on a rectangular grid to compute the poloidal-magnetic-flux distribution and pass the needed flux- surface-averaged quantities to the transport solver in COTSIM. An accelerated version of this computationally intensive, numerical solver has been developed by leveraging deep learning techniques. The neural network-based FBE, which is referred to as FBE-net, uses a partially-connected multi-layer perceptron (MLP) architecture. FBE-net is trained on a dataset generated by the numerical solver, which serves as a source of ground truth. The inputs for the FBE-net are the plasma current, poloidal beta, and the poloidal-field coil currents, while the outputs are the 2D flux map and associated flux-averaged parameters. In comparison to the numerical solver, the neural network-based solver displays a significant increase in computational efficiency without sacrificing computational accuracy.

*Supported by the US DOE under DE-SC0010537.