How physics equations could fix deep neural networks on graphs

Graph neural networks typically pass messages via diffusion, which causes node features to blur into indistinguishability at depth. This team borrowed ideas from the Navier Stokes equations—adding convection (directional flow) alongside diffusion to create a dynamic velocity field on graphs. This lets information travel more efficiently and adaptively while working across datasets with varying structure. On 12 real-world benchmarks, Graph Navier Stokes Networks beat existing methods and demonstrably reduce the oversmoothing problem that limits depth.