Scientific Ml

Graph Neural Operators

Graph neural operators learn PDE solution maps on irregular meshes by passing messages between neighboring nodes and edges.

When the grid is not a grid

The Fourier neural operator relies on a regular grid for its fast transform. Many engineering domains, however, are meshed with irregular, unstructured elements that conform to complex geometry. Graph neural operators handle these by treating the mesh as a graph: nodes are mesh points, edges connect neighbors, and information flows along the edges.

Message passing as integration

Kronos motion — neural operator

Each layer updates a node's latent state by aggregating messages from its neighbors, weighted by learned functions of the edge geometry. Stacking layers widens the receptive field, letting distant parts of the domain influence one another, which approximates the global coupling that PDE solutions require. This mirrors how a numerical stencil spreads information across a mesh over successive steps.

Multi-scale graphs

A single-resolution graph spreads information slowly, one hop per layer. Multi-level constructions add coarse graphs that connect distant nodes directly, so long-range interactions propagate in few layers. This is the graph analogue of multigrid and is essential for problems where boundaries influence the whole interior.

Trade-offs

Handles arbitrary geometry and adaptive meshes naturally
Discretization-aware: can be evaluated on refined or coarsened meshes
More expensive per step than grid-based spectral methods
Long-range coupling needs many layers or explicit multi-scale structure

Where they fit

Graph neural operators suit structural, fluid, and electromagnetic problems posed on the kind of conforming meshes used in engineering analysis. For a shaped device such as the Hyperion breeder, whose spherical-tokamak geometry and negative triangularity of -0.30 demand body-fitted meshes, mesh-native surrogates are a natural complement to grid-based operators. As always, they are trained against and validated by trusted solvers, not treated as ground truth themselves.