Read, Write, Relax: Why Neural PDE Surrogates Need Both Global and Local Processing
Kumar et al.
What it does
Interleaves latent-token attention, which carries information across the whole mesh, with message-passing relaxation along mesh edges, like the two halves of a multigrid cycle.
Why it matters
Explains why global (latent) and local (graph) surrogates each fail on large industrial meshes: attention acts as a spatial low-pass filter, message passing lacks reach. Combining them was the most accurate model in nearly every comparison reported.