CarbonBench | AI4Carbon

The CarbonBench dataset is a benchmark for machine learning emulators of atmospheric CO₂ tracer transport. It was introduced in
Benson et al. (2024): Atmospheric Transport Modeling of CO₂ with Neural Networks.

GitHub Repository

vitusbenson/carbonbench

Official implementation and tools for the CarbonBench dataset - a benchmark for machine learning emulators of atmospheric CO₂ tracer transport.

Stars Forks Python

View on GitHub Download

HuggingFace Repository

vitusbenson/carbonbench

Hosted dataset on Hugging Face with CarbonBench Zarr data and metadata for machine learning transport emulators.

Likes Downloads Dataset

View on Hugging Face Download

Dataset Contents

The dataset contains 3D CO₂ concentrations together with meteorological drivers and flux fields.
All variables are provided as Zarr arrays chunked per timestep.

Variable	Description	Units
`co2massmix`	CO₂ mass mixing ratio	10⁻⁶ kgCO₂ / kgDryAir
`airmass`	Dry air mass	Pg dry air
`gph_bottom`	Geopotential height at lower level	km
`gph_top`	Geopotential height at upper level	km
`p_bottom`	Pressure at lower level	hPa
`p_top`	Pressure at upper level	hPa
`q`	Specific humidity	kg/kg
`t`	Air temperature	K
`u`	Zonal (eastward) wind speed	m/s
`v`	Meridional (northward) wind speed	m/s
`co2flux_anthro`	Anthropogenic surface CO₂ flux	kg CO₂ m⁻² s⁻¹
`co2flux_land`	Land biosphere surface CO₂ flux	kg CO₂ m⁻² s⁻¹
`co2flux_ocean`	Ocean surface CO₂ flux	kg CO₂ m⁻² s⁻¹
`blh`	Planetary boundary layer thickness	m
`tisr`	Incoming solar radiation	J m⁻²
`cell_area`	Surface grid-cell area	m²
`orography`	Surface geopotential	m²/s²

Output variable:

co2massmix at the next timestep (same unit as input)

Baseline Model Results

Performance of the baseline emulators trained on CarbonBench:

Model	Parameters (M)	Decorrelation Time	R²	RMSE
UNet	9.6	>90	0.98	0.52
GraphCast	5.2	>90	0.96	0.86
SFNO	35.7	>90	0.98	0.58
SwinTransformer	37.9	>90	0.99	0.34

Downloading & Preparing the Dataset

You can build the dataset locally using the CarbonBench tools:

# Clone the repo
git clone https://github.com/vitusbenson/carbonbench.git
git clone https://github.com/vitusbenson/neural_transport.git
cd carbonbench

# Install requirements (including neural_transport)
pip install -e ../neural_transport

# Create CarbonTracker-based dataset (example LowRes)
python data/create_carbontracker_dataset.py \
  --save_dir /path/to/data \
  --gridname "latlon5.625" \
  --vertical_levels "l10" \
  --freq "6h"

# Create ObsPack evaluation dataset
python data/create_obspack_dataset.py \
  --save_dir /path/to/data \
  --freq "3h"

Citation

@article{benson2024neuraltransport,
  title = {Atmospheric Transport Modeling of CO2 with Neural Networks},
  author = {Benson, Vitus and Bastos, Ana and Reimers, Christian and Winkler, Alexander J. and Yang, Fanny and Reichstein, Markus},
  year = {2025},
  pages = {e2024MS004655},
  journal = {Journal of Advances in Modeling Earth Systems},
  volume = {17},
  number = {2},
  publisher = {Wiley Online Library},
  url = {https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2024MS004655},
}