Parallelisation of Discrete Exterior Calculus via Representation Theory on Curved and Three-Dimensional Meshes

What the math gives to ML

The paper gives a reusable way to exploit finite symmetries of discretized geometric domains: every DEC operator commuting with the group action can be transformed into independent representation-theoretic blocks using one mesh-dependent basis. The transferable asset is simultaneous block structure for an entire family of geometry-derived operators, including operators introduced later or learned as linear combinations under the same commutation constraint. The strongest neural-network application is an implicit or unrolled neural PDE model whose graph/DEC linear solves dominate runtime: transform once, solve the isotypic blocks independently, and reuse the factorization across training iterations and operator-splitting substeps.