Operator Learning and Fourier Neural Operators
Operator-learning methods approximate the solution operators of plasma partial differential equations, generalizing across inputs and resolutions.
Learning operators, not functions
Most surrogates learn a function from finite inputs to finite outputs. Operator learning instead approximates the mapping between functions, for example from an input field to the solution of a partial differential equation. This lets a single trained model solve a whole family of problems and, in some formulations, at resolutions not seen in training.
Fourier neural operators
A Fourier neural operator (FNO) performs its main mixing in the frequency domain: it transforms the input field with a fast Fourier transform, applies learned weights to a set of low-frequency modes, and transforms back. Because convolution in space is multiplication in frequency, this captures global, smooth structure efficiently, which suits many PDE solutions.
- Input and output are fields, such as flux, density, or flux maps
- Learns a resolution-flexible solution operator
- Efficient for smooth, global structure via spectral mixing
- Trained on pairs of inputs and solver outputs
Fusion applications
Operator learning is being explored for equilibrium fields, transport and radiation solution maps, and time evolution of plasma fields. Its promise is one fast model covering many cases at once, replacing repeated solver calls in design scans and, potentially, in fast prediction loops.
Limits
Spectral methods handle smooth structure well but can struggle with sharp features such as separatrices and shock-like fronts, where high frequencies matter. As with all surrogates, reliability holds only over the training distribution, and outputs are validated against reference solvers before use.
For concepts like the Hyperion breeder, operator learning offers a route to fast field-level surrogates for equilibrium and transport across design scans. These are computational tools for a machine under design, anchored to first-principles solvers, with no performance claimed for unbuilt hardware.