Space-Time Galerkin Boundary Element Method for the Wave Equation

What the math gives to ML

The paper provides a constructive way to integrate pairwise kernels that are singular at coincident points and nonsmooth at a finite-radius cutoff. Its transferable asset is not the boundary-element formulation itself, but the decomposition of a Cartesian product of cells into convex-hull pieces whose coordinates isolate the singular or discontinuous behavior into a small number of variables. This can become a geometry-aware interaction layer for graph neural networks or physics-informed losses, replacing unstable point-sampled pair interactions with exponentially convergent cell-averaged interactions. The first useful target is a small mesh-based model where neighboring-cell interactions dominate and naive quadrature produces high-variance gradients.