Reproduce “Learning vertical coordinates via automatic differentiation of a dynamical core” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tim Whittaker, Seth Taylor, Elsa Cardoso-Bihlo, Alejandro Di Luca, Alex Bihlo
Researchers from Université du Québec à Montréal, University of Saskatchewan, and Memorial University of Newfoundland introduced NEUVE, a neural network-based vertical coordinate system, optimized within a differentiable dynamical core. The learned coordinate system reduced L2-errors by factors of 1.4 to 2 in advection tests and allowed for a four-fold coarsening of horizontal grid resolution while maintaining accuracy compared to traditional methods.
Key insights
- Differentiable programming provides a viable method for optimizing fundamental numerical components, such as vertical coordinate systems, within atmospheric dynamical cores.
- An objectively optimized neural network parameterization can adapt coordinate structures to specific flow regimes, outperforming rigid, hand-tuned heuristic designs.
- Exact computation of geometric metric terms through automatic differentiation enhances numerical consistency and reduces spurious forcing over complex terrain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.