Neural Operators

keras_climate.operators — neural operator learning (PDE surrogate modeling), scoped to the one operator architecture with a direct climate application in this repo. Unlike a standard CNN, it is resolution-invariant: the same learned weights apply regardless of the input grid's spatial size.

AFNO

Adaptive Fourier Neural Operator (Guibas et al. 2021) — a token-mixing block for ViT-style backbones that replaces self-attention with a 2D FFT over the spatial token grid, a block-diagonal complex-valued 2-layer MLP applied per retained frequency mode, soft-thresholding sparsification, and an inverse FFT — all at O(N log N) instead of self-attention's O(N²). This is the token-mixer used inside FourCastNet. Only hard_thresholding_fraction=1.0 (every frequency mode kept) is supported — the setting FourCastNet's own released checkpoint actually uses.

No standalone, task-agnostic pretrained AFNO checkpoint exists anywhere (the only released AFNO weights are FourCastNet's own) — see Pretrained Weights.

Note

Generic, climate-agnostic operator architectures (FNO, DeepONet, UNO — typically trained/benchmarked on fluid-dynamics or Darcy-flow PDEs, not Earth-system data) have been removed from this repo's scope. AFNO stays because it's the actual building block a real climate model (FourCastNet) uses.