papersTODAY 04:00 UTC
Attention-Based Flux Scheme Targets Spurious Oscillations in Finite-Volume Solvers
A new arXiv preprint proposes a conservative finite-volume method on fixed grids in which an attention mechanism selects upstream information to build the numerical flux. The scheme is conditioned on the CFL constraint so that a shock can propagate across multiple cells in a single update without being smeared or broken. The work sits at the intersection of numerical PDE solvers and machine learning attention architectures.