Held–Suarez Dry Atmosphere (dry-atmosphere-3D)

One line description of the data: 162 3D Held–Suarez dry-atmosphere simulations on 8 pressure levels.

Longer description of the data: Held–Suarez is the standard benchmark for the dynamical cores of atmospheric general circulation models. It is a dry atmosphere with no topography, forced by relaxing temperature toward a zonally symmetric equilibrium and damped by Rayleigh friction near the surface. Baroclinic instability produces eddy-driven jets and storm tracks. We vary the rotation rate $\mathrm{\Omega }\backslash Omega$, the equator-to-pole temperature contrast $\mathrm{\Delta }{T}_y\backslash Delta\; T\_y$, and the static stability $\mathrm{\Delta }{\theta }_z\backslash Delta\backslash theta\_z$, which changes the number, latitude, and width of the jets. This dataset is of particular relevance since it serves as a simple proxy for the complex atmospheric dynamics that drive weather forecasting.

Code or software used to generate the data: ClimaAtmos.jl, the atmospheric component of the CliMA software stack.

Visualization:

About the data

Dimension of discretized data: 1461 snapshots per trajectory (365 days, 6-hourly) on an $8\times 144\times 2888\; \backslash times\; 144\; \backslash times\; 288$ (level × lat × lon) grid, plus surface pressure.

Fields available in the data:

HDF5 path	Shape	Units	Description
t0_fields/T	(1, 1461, 8, 144, 288)	K	Air temperature
t0_fields/ps	(1, 1461, 1, 144, 288)	Pa	Surface pressure
t1_fields/velocity	(1, 1461, 8, 144, 288, 2)	m/s	Horizontal wind; last axis is [u, v] (zonal, meridional)
scalars/{omega_factor, delta_T_y, delta_theta_z, seed}	()	—	Run parameters (seed is stored as a float)
dimensions/{time, level, lat, lon}	—	s, hPa, °, °	Coordinates

Number of trajectories: 162: $3\times 3\times 33\; \backslash times\; 3\; \backslash times\; 3$ parameter combinations, each with 6 seeds.

Estimated size of the ensemble of all simulations: 667 GB (compressed HDF5).

Grid type: $8\times 144\times 2888\; \backslash times\; 144\; \backslash times\; 288$ level-lat-lon grid (equiangular). Levels are given by pressure from 50 hPa to 1000 hPa. If the surface pressure is below 1000 hPa, ClimaAtmos.jl fills the 1000 hPa level with the values at the lowest z-level of the simulation.

Initial conditions: An atmosphere at rest with a decaying temperature profile, plus a small seeded temperature perturbation below 5 km. The first 200 days are discarded as spin-up.

Boundary conditions: Spherical; flat and homogeneous lower boundary (no topography, no surface fluxes).

Simulation time-step $(\mathrm{\Delta }t)(\backslash Delta\; t)$: 100 s.

Data are stored separated by $(\delta t)(\backslash delta\; t)$: 6 hours.

Total time range $(t_{min}(t\_\{min\}$ to $t_{max})t\_\{max\})$: Day 200 to day 565 of each run. (time coordinate counts seconds from the start of the simulation, so the runs begin at 17,280,000 s).

Set of coefficients or non-dimensional parameters evaluated:

• omega_factor $=\mathrm{\Omega }\mathrm{/}{\mathrm{\Omega }}_{\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{h}}\in \{0.5,1.0,2.0\}=\; \backslash Omega/\backslash Omega\_\{\backslash rm\; earth\}\; \backslash in\; \backslash \{0.5,\; 1.0,\; 2.0\backslash \}$, where ${\mathrm{\Omega }}_{\mathrm{e}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{h}}=7.292\times {10}^{-5}{\mathrm{s}}^{-1}\backslash Omega\_\{\backslash rm\; earth\}\; =\; 7.292\; \backslash times\; 10^\{-5\}\backslash ,\backslash mathrm\{s\}^\{-1\}$
• delta_T_y $=\mathrm{\Delta }{T}_y\in \{40,60,80\}=\; \backslash Delta\; T\_y\; \backslash in\; \backslash \{40,\; 60,\; 80\backslash \}$ K, the equator-to-pole contrast in equilibrium temperature
• delta_theta_z $=\mathrm{\Delta }{\theta }_z\in \{5,10,20\}=\; \backslash Delta\backslash theta\_z\; \backslash in\; \backslash \{5,\; 10,\; 20\backslash \}$ K, the vertical potential-temperature contrast (sets the static stability)
• seed $\in \{0,\dots ,5\}\backslash in\; \backslash \{0,\; \backslash ldots,\; 5\backslash \}$, the initial perturbation

The canonical Held–Suarez setting is at the center point (1.0, 60, 10).

Approximate time to generate the data: Per-run wall-time on the order of six hours.

Hardware used to generate the data: GPU-accelerated generation, using one GPU per run (either an NVIDIA RTX 4090, NVIDIA L40S, or NVIDIA A100).

Loading

It's important to load hdf5plugin before h5py to make sure the filters are registered properly.

Download a single run (≈4.3 GB) and read it with h5py:

import hdf5plugin  # registers the LZ4 filter, must come before reading
import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download("sada-group/dry-atmosphere-3D", "data/test/run_0014.h5", repo_type="dataset")
with h5py.File(path, "r") as f:
    T = f["t0_fields/T"][0, :8]              # (8 steps, 8 levels, 144, 288)
    u, v = f["t1_fields/velocity"][0, :8].transpose(4, 0, 1, 2, 3)
    params = {k: f["scalars"][k][()] for k in f["scalars"]}

The full dataset is 667 GB. To download a subset, e.g. the 16 test runs (≈70 GB), use snapshot_download with allow_patterns:

from huggingface_hub import snapshot_download

root = snapshot_download(
    "sada-group/dry-atmosphere-3D",
    repo_type="dataset",
    allow_patterns=["data/test/*", "stats.yaml"],  # drop allow_patterns for everything
)

Please cite the associated paper if you use this data in your research:

@misc{muser2026dandelionsphericalflowerneural,
      title={Dandelion: A Spherical Flower for Neural Simulation of Planetary Dynamics}, 
      author={Till Muser and Giovanni Abati and Ivan Dokmanić},
      year={2026},
      eprint={2608.27521},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2608.27521}, 
}

as well as the original benchmark and the solver:

@article{held1994proposal,
  title={A proposal for the intercomparison of the dynamical cores of atmospheric general circulation models},
  author={Held, Isaac M and Suarez, Max J},
  journal={Bulletin of the American Meteorological society},
  volume={75},
  number={10},
  pages={1825--1830},
  year={1994},
  publisher={American Meteorological Society}

}

@article{yatunin2026climate,
  title={The climate modeling alliance atmosphere dynamical core: Concepts, numerics, and scaling},
  author={Yatunin, Dennis and Byrne, Simon and Kawczynski, Charles and Kandala, Sriharsha and Bozzola, Gabriele and Sridhar, Akshay and Shen, Zhaoyi and Jaruga, Anna and Sloan, Julia and He, Jia and others},
  journal={Journal of Advances in Modeling Earth Systems},
  volume={18},
  number={3},
  pages={e2025MS005014},
  year={2026},
  publisher={Wiley Online Library}
}