Multi-Modal Foundation · nvidia
cBottle
A cascaded diffusion model for generating kilometer-scale Earth atmospheric states on HEALPix geometry.
What cBottle is
Climate in a Bottle (cBottle) is a diffusion-based generative model developed by NVIDIA for Earth system science and climate modelling. It synthesises atmospheric states at kilometer-scale resolution, addressing the spatial and temporal limitations of traditional coarse-resolution reanalysis products.
While reanalysis products such as ERA5 offer global atmospheric states at approximately 25-kilometre resolution, they cannot explicitly resolve mesoscale phenomena like deep convection. cBottle bridges this gap by combining historical reanalysis observations with high-resolution global climate simulations, generating detailed representations of atmospheric dynamics.
How it works
cBottle uses a cascaded diffusion architecture built on a Song-UNet convolutional neural network mapped to a HEALPix spherical geometry. The system comprises multiple variants: cBottle-3d (150 million parameters) for coarse spatial states, cBottle-video (282 million parameters), cBottle-SR (330 million parameters) for super-resolution upscaling, and cBottle-3d-tc (150 million parameters), which incorporates two extra convolutional layers at the UNet base to enable tropical cyclone guidance.
The model conditions on one-hot dataset identifiers, temporal indices (day of year and second of day), monthly mean sea surface temperatures at HEALPix Nside=64, and optional tropical cyclone location maps at HEALPix Nside=8. Training employs four primary data sources: ERA5 reanalysis (1980–2017) regridded to HEALPix, coupled 5-kilometre ICON Cycle 3 simulations (2020–2024), inputs4MIPs sea surface boundary conditions, and IBTrACS cyclone tracks. Coarse stages output fields on an Nside=64 HEALPix grid, whilst the super-resolution component scales outputs to an Nside=1024 grid.
Key facts
- Architecture
- Cascaded diffusion model with Song-UNet on HEALPix geometry
- Model variants
- cBottle-3d, cBottle-video, cBottle-SR, cBottle-3d-tc
- Parameter count (cBottle-3d)
- 150M
- Parameter count (cBottle-video)
- 282M
- Parameter count (cBottle-SR)
- 330M
- Parameter count (cBottle-3d-tc)
- 150M
- Model version
- v1.2
- Release date
- 12 May 2025
- Software framework
- PyTorch
- Training data
- ERA5, ICON Cycle 3, inputs4MIPs, IBTrACS
What you can use it for
- Kilometer-scale atmospheric state generation
- Generates random global atmospheric fields at kilometer resolution conditioned on seasonal timing and sea surface temperatures.
- Atmospheric super-resolution
- Upscales coarse-resolution atmospheric fields from HEALPix Nside=64 to fine-scale Nside=1024 grids using cBottle-SR.
- Cyclone-guided atmospheric synthesis
- Generates atmospheric states steered by designated tropical cyclone positions using cBottle-3d-tc with HPX8 location guidance.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Generates atmospheric states at kilometer resolution where mesoscale convective processes are resolved rather than parameterized.
- Operates natively on spherical HEALPix geometry across both coarse (Nside=64) and high-resolution (Nside=1024) grids.
- Combines observation-assimilated reanalysis data (ERA5) with high-fidelity, physically coupled global fluid simulation data (ICON).
Watch out for
- No independent ICON simulation data was withheld for model evaluation.
- Below-surface atmospheric pressure levels require synthetic extrapolation using hydrostatic assumptions and a constant temperature lapse rate.
- Inference is restricted to Linux environments running NVIDIA GPU architectures (Turing, Ampere, Hopper, or Blackwell).
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use cBottle commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
What grid representations do the model inputs and outputs use?
Coarse inputs and baseline outputs use the HEALPix grid at Nside=64, optional tropical cyclone guidance inputs use Nside=8, and the super-resolution module (cBottle-SR) produces output fields at Nside=1024.
What inputs are required to run inference?
The model expects PyTorch tensors specifying a dataset label (ERA5 or ICON), day of the year, second of the day, and monthly mean sea surface temperature on an HPX64 grid, along with an optional HPX8 tropical cyclone location map.
How do cBottle-3d and cBottle-3d-tc differ?
cBottle-3d-tc adds two convolutional layers to the lowest level of the Song-UNet architecture present in cBottle-3d, allowing conditioning on tropical cyclone locations while retaining a 150-million-parameter footprint.
Written from cBottle's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.