Scientific Ml

Neural Operators

Neural operators learn mappings between function spaces, so one trained model can solve a whole family of PDEs at any resolution.

From functions to functions

A conventional network maps a fixed-size vector to another vector. A neural operator instead maps an entire input function to an entire output function. For a differential equation this means learning the operator that takes a coefficient field, boundary data, or initial condition to the corresponding solution field. One trained operator can then solve every instance in a family of problems, not just one.

Why this is powerful

Kronos motion — family decades

Because the target is an operator between infinite-dimensional spaces, a properly built neural operator is discretization-invariant: it can be trained on one grid resolution and evaluated on another without retraining. This separates the learned mapping from the numerical mesh, a property finite-dimensional networks do not have.

The general recipe

Most neural operators lift the input function into a higher-dimensional latent representation, apply several layers that mix information globally across the domain, then project back to the output function. The global mixing is the key ingredient: solving a PDE couples every point to every other point, so a purely local architecture cannot capture the operator.

Leading architectures

DeepONet, which splits the operator into a branch network for the input function and a trunk network for the query location
The Fourier neural operator, which performs global mixing as a multiplication in the frequency domain
Graph neural operators, which handle irregular meshes through message passing
Wavelet and low-rank variants that trade global reach for efficiency

Where they fit

Neural operators shine when the same class of equation must be solved thousands of times with varying inputs, as in design sweeps, uncertainty propagation, or real-time control. They are surrogates: fast approximations trained against a trusted solver, validated on held-out cases before being trusted for new inputs.