Mesh-Native Tokenisation: Retrofitting Foundation Models to Unstructured Data
Maierhofer et al.
What it does
Mesh-Native Convolution, a tokenisation layer that encodes fields on unstructured meshes the way a convolutional encoder does on grids, so a physics foundation model pretrained on grids can be moved to arbitrary meshes.
Why it matters
Foundation models trained on grid data could not be used on curved, meshed domains without a different (graph) architecture; with MNC they reach acceptable accuracy often zero-shot and beat baselines after a few epochs of fine-tuning.