An unfitted finite element discrete fracture model for low-permeability barriers via local stiffness matrix modification

What the math gives to ML

The paper constructs discontinuities without enriching the global finite-element space by introducing local auxiliary jump variables and eliminating them through stationarity. The transferable asset is the resulting Schur-complement correction: a small set of latent region offsets produces a structured, low-rank modification of a baseline coupling matrix while preserving the original visible variables. This suggests a graph-neural-network message-passing layer that learns sparse barriers between token or node groups, reducing oversmoothing while keeping the same node representation size and making the barrier effect analytically stable.