Mesh-Native Convolution Enables Tokenization for Unstructured Physical Data

Mesh-Native Convolution (MNC) introduces a locality-aware tokenization layer designed for scientific computing on unstructured meshes and curved boundaries. This approach allows physics foundation models to process complex geometric data more efficiently than standard graph neural networks.