Fourier Neural Operator (Original Paper)

The paper proposes parameterizing the integral kernel directly in Fourier space and composing linear transforms with nonlinear activations, yielding an operator that maps between function spaces rather than finite vectors.

Key result: Trained on a low resolution, the model generalizes to higher resolutions at inference - zero-shot super-resolution - with large speedups over classical solvers on Burgers, Darcy and Navier-Stokes.