Learning Vertical Coordinates via Automatic Differentiation of a Dynamical Core
- 1Département des Sciences de la Terre et de l’Atmosphère Centre pour l’étude et la simulation du climat à l’échelle régionale (ESCER) Université du Québec à Montréal Montréal QC Canada
- 2Department of Computer Science University of Saskatchewan Saskatoon SK Canada
- 3Department of Mathematics and Statistics Memorial University of Newfoundland St. John's NL Canada
Abstract
Terrain‐following coordinates in atmospheric models often imprint their grid structure onto the solution, particularly over steep topography, where distorted coordinate layers can generate spurious horizontal and vertical motion. Standard formulations, such as hybrid or SLEVE coordinates, mitigate these errors by using analytic decay functions controlled by heuristic scale parameters that are typically tuned by hand and fixed a priori. In this work, we propose a framework that defines a parametric vertical coordinate system as a learnable component within a differentiable dynamical core. We develop an end‐to‐end differentiable numerical solver for the two‐dimensional non‐hydrostatic Euler equations on an Arakawa C‐grid, and introduce a NEUral Vertical Enhancement (NEUVE) terrain‐following coordinate based on an integral transformed neural network that guarantees monotonicity. By backpropagating simulation errors through the time integration to the coordinate parameters, our formulation finds a grid structure optimized for both the underlying physics and the numerical discretization. Using several standard tests, we demonstrate that these learned coordinates reduce the mean squared error by a factor of 1.4–2 in nonlinear statistical benchmarks, and eliminate spurious vertical velocity striations over steep topography.
Related
- An unstructured-grid, nonhydrostatic, GVC ocean model Part I: Model description and application of vertical hybrid coordinates to internal solitary waves
- GPU-accelerated three-dimensional variational data assimilation with automatic differentiation for WRF-Chem
- Cloud vertical structure across China from a national Ka-band cloud radar network: thermodynamic, dynamical, and land-surface controls
- Automatic Monitoring of Rock-Slope Failures Using Distributed Acoustic Sensing and Semi-Supervised Learning
- Automatic detection of laminated clouds in lidar data
BibTeX
@article{whittaker2026learning,
title = {{Learning Vertical Coordinates via Automatic Differentiation of a Dynamical Core}},
author = {Tim Whittaker and Seth Taylor and Elsa Dos Santos Cardoso‐Bihlo and Alejandro Di Luca and Alexander Bihlo},
journal = {Journal of Advances in Modeling Earth Systems},
year = {2026},
month = aug,
doi = {10.1029/2025ms005711},
url = {https://doi.org/10.1029/2025ms005711},
note = {CC BY-NC},
}
RIS
TY - JOUR
TI - Learning Vertical Coordinates via Automatic Differentiation of a Dynamical Core
AU - Tim Whittaker
AU - Seth Taylor
AU - Elsa Dos Santos Cardoso‐Bihlo
AU - Alejandro Di Luca
AU - Alexander Bihlo
PY - 2026
DA - 2026/08/01
JO - Journal of Advances in Modeling Earth Systems
DO - 10.1029/2025ms005711
UR - https://doi.org/10.1029/2025ms005711
AB - Terrain‐following coordinates in atmospheric models often imprint their grid structure onto the solution, particularly over steep topography, where distorted coordinate layers can generate spurious horizontal and vertical motion. Standard formulations, such as hybrid or SLEVE coordinates, mitigate these errors by using analytic decay functions controlled by heuristic scale parameters that are typically tuned by hand and fixed a priori. In this work, we propose a framework that defines a parametric vertical coordinate system as a learnable component within a differentiable dynamical core. We develop an end‐to‐end differentiable numerical solver for the two‐dimensional non‐hydrostatic Euler equations on an Arakawa C‐grid, and introduce a NEUral Vertical Enhancement (NEUVE) terrain‐following coordinate based on an integral transformed neural network that guarantees monotonicity. By backpropagating simulation errors through the time integration to the coordinate parameters, our formulation finds a grid structure optimized for both the underlying physics and the numerical discretization. Using several standard tests, we demonstrate that these learned coordinates reduce the mean squared error by a factor of 1.4–2 in nonlinear statistical benchmarks, and eliminate spurious vertical velocity striations over steep topography.
SN - 1942-2466
ER -