Physics-Informed Neural Network Solves Quasi-static Tokamak MHD

A new paper submitted to arXiv on 22 Apr 2026 demonstrates a physics-informed neural network, `PINN`, trained without experimental or synthetic data to solve the time-dependent quasi-static magnetohydrodynamic equations, `MHD`, in axisymmetric tokamak geometry. The authors applied the method to an ITER-like tokamak and report that the PINN reproduces a vertically displacing plasma with general agreement against a ground-truth simulation.

The PINN learns MHD dynamics purely from the governing equations and boundary conditions, avoiding supervised data. Key implementation choices include careful treatment of the quasi-static approximation and axisymmetric coordinates. The paper emphasizes:

• •physics-only training via PDE residual and boundary-condition losses
• •axisymmetric tokamak geometry and ITER-like parameters
• •time-dependent prediction of plasma vertical displacement events
• •validation against conventional numerical simulation showing qualitative agreement

This is a targeted advance for computational plasma physics and fusion engineering. Traditional MHD solvers remain computationally expensive and require fine meshes and implicit timestepping for stability. A PINN that captures essential dynamics could act as a fast surrogate for parameter scans, control design, or real-time state estimation. At the same time, PINNs historically struggle with stiff, multi-scale PDEs and enforcing complex boundary conditions; showing a viable PINN for tokamak MHD is an important proof of concept, not yet a drop-in replacement for established solvers.

Next steps to validate utility are quantitative benchmarks on convergence, robustness across operating regimes, and scaling to full 3D MHD and disruption physics. Availability of code and reproducible experiments will determine uptake by fusion modelers and control engineers.