Fourier Neural Operator FNO ★
Definition
A neural operator that convolves in the Fourier domain: lift the input to frequency space, truncate to low-frequency modes, apply a learned linear transform, then invert back to physical space. Extremely efficient on regular grids and the most widely used neural operator in practice.
What it is
Each FNO layer does three things: FFT, a learned complex linear transform on the truncated low-frequency modes, then inverse FFT. Because physical fields concentrate their energy at low frequencies, truncating high modes saves compute and acts as implicit regularization. The network stacks several such spectral convolution layers with pointwise nonlinearities.
Key points
- Complexity: FFT is
O(N log N), far cheaper than a spatial convolution — this is why FNO is fast. - Resolution-agnostic: low-mode weights do not depend on grid size, so a model trained at 64x64 can in principle run at 256x256.
- Requires a regular grid: FFT assumes a structured grid. Unstructured meshes need interpolation onto a regular grid (which adds error) or a graph neural operator instead.
- Performs best under periodic boundary conditions; padding or extension handles non-periodic cases.
Common pitfalls / boundaries
- "Resolution invariant" is not "accuracy invariant": at higher resolution the missing high-frequency detail still shows up as error.
- High-Reynolds / turbulent flows need many modes — the mode truncation is a key hyperparameter to tune.
- For complex geometries (e.g. flow around a holed airfoil) the regular-grid assumption becomes the bottleneck; evaluate graph neural operators there.