Neural Operator ★
Definition
A deep learning paradigm that learns a function-to-function mapping: feed in a parameter field or boundary condition and get the whole solution field out. Unlike a PINN, it learns the solution operator for a family of PDEs and can be reused after one training run.
What it is
Neural operators treat solving as a mapping between function spaces: given an input function (coefficient field, initial field, boundary condition, geometry descriptor) the network outputs the solution function. It learns the operator G: a -> u, not the solution of one particular problem. After training, a new input function yields its solution in a single forward pass — no retraining.
Key points
- Discretization invariance is the headline property: the same model accepts inputs at any resolution and the output can be evaluated anywhere.
- Two dominant designs: FNO uses spectral convolutions (needs a regular grid), DeepONet uses a branch/trunk pair (naturally mesh-free).
- Inference runs in milliseconds to seconds — orders of magnitude faster than a classical solver, which matters when you must solve thousands of times (design optimization, control, inversion).
- The cost is data hunger: training pairs are generated by classical solvers, so offline cost is real.
Common pitfalls / boundaries
- Neural operators do not replace PINNs — the problem settings differ. They can also be combined (physics-constrained operator learning).
- "Zero-shot generalization" is relative. Outside the training distribution of geometries or parameters, error grows fast; always run an extrapolation test before production use.
- Reported speedups are often measured against an unoptimized serial solver — check what the baseline actually is.