Local Well-Posedness of the Inviscid Shallow Water Equations with Surface Tension

What the math gives to ML

The paper exposes a structure-preserving decomposition for coupled dynamics: a positive symmetrizer defines the energy metric, while the principal transport and surface-tension operators are arranged so that their contributions are energy-neutral rather than amplifying. In two dimensions, the nonsymmetric velocity coupling is explicitly split into symmetric and skew parts, and the specific vorticity \(\theta=\operatorname{curl}u/h\) is exactly advected. These constructions can be transferred to recurrent, state-space, graph, or sequence-mixing layers by parameterizing the interaction operator as a sum of energy-symmetric damping and energy-skew transport. The resulting layers should permit substantially larger stable step sizes and reduce exploding or vanishing activations without relying only on spectral normalization or small residual scales.