Paperlane

no. 16 · Tuesday 17 November 2026

Physics-Spawned Neural Integration of Rotating Shallow-Water Dynamics

Kaushal Gianchandani, Luka Todorovic and Sai Ravela

math.NA · via physics.comp-ph

Abstract

Equation-derived neural integration can execute prescribed dynamics without training, while informative ensemble learning can estimate uncertain parameters and identify governing interactions without requiring differentiability. Together, these approaches motivate Physics-Spawned Informative Networks ($Φ$SpIN), a framework for prescribed and adaptive hybrid dynamical systems. As a first step toward flood modeling applications, we present $Φ$SpIN-SW, its prescribed-physics realization for the rotating shallow-water equations. Compatible C-grid operators on spherical Voronoi meshes define a computational graph that retains non-polynomial operations and advances the state through recurrent neural integration compiled with JAX. Analytical benchmarks characterize numerical errors, and global and coastal experiments closely reproduce reference solutions generated using the Model Prediction Across Scales (MPAS) model. At approximately 14~km global resolution, A100, H200 and B200 GPUs attain 9-, 19- and 34-fold higher timestep throughput than a 96-core MPAS-SW CPU reference under matched workloads. These results extend training-free neural integration beyond polynomial dynamics to complex geophysical simulation, establishing a computational foundation for subsequent adjoint-free ensemble estimation of physical and neural parameters and informative adaptation of operator structure.

fields
Numerical Analysis
pdf ↗ page arXiv:2611.20054 ↗