High-order finite element method for perfect conductivity and linear elasticity with nearly touching inclusions

What the math gives to ML

The paper develops a constructive, distance-robust discretization for PDE solutions whose derivatives blow up in narrow gaps between nearly touching inclusions. Its transferable asset is the explicit graded-resolution law: computational samples are concentrated according to distance from the singular gap, with a separate minimum scale inside the narrow core. This suggests singularity-aware collocation for PINNs and neural operators solving elliptic interface problems, where uniform sampling often under-resolves high-gradient boundary layers. The idea is directly falsifiable by comparing equal-budget training against uniform and residual-adaptive sampling.