Operator Learning
Definition
The umbrella term for machine learning that targets mappings between infinite-dimensional function spaces; neural operators are its dominant implementation. What separates it from ordinary supervised learning is that both inputs and outputs are functions, not vectors.
What it is
Classical supervised learning maps finite-dimensional vectors R^n -> R^m; operator learning maps A -> U where both A and U are function spaces such as L2(D). The model must therefore hold under any discretization, which is why discretization invariance is emphasized so much.
Key points
- Different evaluation criteria: besides ordinary error, you must test resolution generalization (train coarse, test fine) and parameter generalization (coefficient ranges outside the training set).
- Theory: universal approximation theorems exist for operators (for Lipschitz-continuous operators), but practical convergence rates still depend on problem structure.
- Overlap with ROM: both seek a low-dimensional representation, but operator learning learns the map from data while ROM projects the equations.
- Training data usually comes from classical solvers, so data-generation cost is the most practical limit on adoption.
Common pitfalls / boundaries
- In-distribution test error alone overstates capability; always keep a separate out-of-distribution evaluation set.
- Operator learning is not universal acceleration: when every new case requires thousands of fresh training samples, optimizing the solver itself is cheaper.