Invariant-domain-preserving limiting with Adaptive Mesh Refinement for Legendre-Gauss-Lobatto Discontinuous Galerkin Spectral Element Methods

arXiv:2607.06045 2026 Architecture 1 ideas extracted · analyzed Aug 30, 2026

What the math gives to ML

The paper develops a concrete conservative coupling for interfaces where one representation has a different resolution or number of nodes than the other. Its transferable asset is the combination of nonnegative partition-of-unity weights, exact weighted conservation, and sparsification based on local subcell supports. This suggests a stable multiresolution neural module in which coarse and fine tokens, graph nodes, or latent cells communicate through a sparse transport operator that preserves weighted feature averages. A convex limiter can then blend this safe low-order coupling with an expressive high-order neural update while enforcing a prescribed convex feature domain.

Ideas from this paper

✗ Failed on benchmark 2026

Conservative Sparse Mortar Cross-Attention

Replace dense coarse-to-fine cross-attention at multiresolution interfaces with a sparse, nonnegative overlap operator whose weighted feature average is exactly conserved between the two resolutions. Use this operator as a low-order path and blend it with an unrestricted neural cross-attention path through a convex limiter that keeps features inside a box or simplex domain. The construction is especially suitable for adaptive token grids, hierarchical graph neural networks, neural operators…

Useful7/10

Difficulty5/10

Novelty8/10

Architecture Stability Scalability Accuracy

Paper: Invariant-domain-preserving limiting with Adaptive Mesh Refinement for Legendre-Gauss-Lobatto Discontinuous Galerkin Spectral Element Methods arXiv:2607.06045