Neural Operators for PDEs
Networks that learn mappings between functions, approximating the solution operators of the equations governing plasmas.
From functions to functions
Many plasma quantities obey partial differential equations (PDEs): fields, profiles, and flows evolve according to differential operators. A neural operator learns the mapping from an input function (say, initial conditions or coefficients) to an output function (the solution), rather than mapping fixed-size vectors.
Why this matters
A standard network fixed to one grid must be retrained for another resolution. Neural operators aim to be discretization-invariant: trained on one grid, they can be queried at another. This suits multi-scale plasma problems where resolution needs vary.
- Fourier neural operators, which learn in the frequency domain
- Graph-based operators for irregular meshes
- DeepONet-style architectures with separate branch and trunk networks
Uses in fusion modeling
Neural operators are studied as fast emulators for field evolution, equilibrium solves, and transport, where the object of interest is a whole profile or field, not a scalar. Once trained on solver output, they can produce approximate solutions far faster.
Limits and honesty
Neural operators learn from data generated by a solver and inherit its physics and its domain. They can violate conservation laws unless constrained, and they extrapolate poorly to inputs unlike the training set. Reported speedups are meaningful only alongside accuracy on held-out cases and a clear statement of the input range where the operator is valid.
They are a promising tool for accelerating PDE solves, not a replacement for the physics that defines the problem.