AMNO surrogate cuts ICRH full-wave solve time to 0.25 s on EAST with 7.5 % reference points
Preprint describes a physics-constrained neural operator that delivers sub-second parametric ICRH field solutions for EAST with far less supervision than dense full-wave runs. Accuracy holds for interpolation within the trained minority-fraction range but remains untested outside it. The approach targets the computational barrier to rapid scenario exploration in tokamak heating design.
The model learns a shared operator from spatially varying anisotropic dielectric tensors to frequency-domain Maxwell solutions. Spectral layers capture global coupling while a physics residual enforces the curl equations, allowing stable interpolation on unseen X_H values. Compared with a sparsely supervised Fourier neural operator, AMNO reduced relative L2 error by 66-90 % against COMSOL references on the same EAST geometry.
EAST ICRH scenarios are central to minority heating schemes that will be scaled to ITER and CFETR; repeated full-wave solves currently limit real-time parameter studies. By replacing repeated matrix assembly with a single trained operator, AMNO directly addresses the bottleneck of exploring minority-fraction dependence in antenna-plasma coupling efficiency.
The work extends Fourier neural operator methods previously demonstrated on simpler wave problems to vector Maxwell systems with tensor coefficients. No peer-reviewed follow-up yet exists, and the training domain remains narrow (fixed frequency, fixed geometry). Wider adoption will require demonstration on multi-parameter families and experimental validation against measured loading resistance.
Next steps include coupling the surrogate to ray-tracing or Fokker-Planck codes for self-consistent minority distribution functions and testing generalization to off-axis heating geometries planned for EAST upgrades in 2026-2027.
Zhang et al.: AMNO error will stay below 5 % L2 on a 20 % wider X_H interval when retrained with EAST 2026 discharge data by end of 2027
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2608.27490)
- [2]Supporting Source(https://arxiv.org/abs/2010.08895)
- [3]Supporting Source(https://doi.org/10.1088/1741-4326/ac3e3f)