Multi-Modal Foundation · nvidia
StormCast V1
A mesoscale neural model forecasting 99 atmospheric state variables at kilometre scale using generative diffusion.
What StormCast V1 is
StormCast V1 is a mesoscale machine learning model developed by NVIDIA for convection-allowing atmospheric weather forecasting. Convection-allowing models simulate fine-scale meteorological phenomena without parameterising deep moist convection, enabling more realistic depiction of storms and localised weather dynamics.
The model emulates high-resolution numerical weather prediction systems by autoregressively forecasting 99 atmospheric state variables at kilometre scale on an hourly time step. It is designed with dense vertical resolution in the atmospheric boundary layer, capturing critical near-surface interactions.
By combining global atmospheric background conditions from ERA5 reanalysis with regional high-resolution data from the High-Resolution Rapid Refresh (HRRR) system, StormCast V1 provides an AI-driven approach to regional weather emulation.
How it works
StormCast V1 uses a UNet architecture operating within a regression-diffusion generative framework. It takes four-dimensional tensor inputs consisting of 125 variables (99 internal state variables and 26 conditioning variables) alongside a one-dimensional date-time array, generating 99 state variables for the next one-hour step.
The system was trained on paired datasets covering July 2018 to December 2021: the ECMWF ERA5 reanalysis (providing 30 km global hourly estimates across 137 atmospheric levels) and NOAA's HRRR system (a 3 km, hourly updated, cloud-resolving model with radar assimilation).
Key facts
- Architecture
- UNet regression-diffusion generative model
- Model version
- v1
- Temporal resolution
- 1-hour time step
- Input tensor dimensions
- 4D (batch, variable, latitude, longitude)
- Input variables
- 125 total (99 state variables, 26 conditioning variables)
- Output variables
- 99 state variables
- Primary training data
- ERA5 (30 km) and HRRR (3 km), July 2018 to December 2021
- Supported hardware
- NVIDIA Turing, Ampere, and Hopper architectures
- Supported operating system
- Linux
What you can use it for
- Convection-allowing weather forecasting
- Autoregressively emulating regional atmospheric dynamics at kilometre scale over hourly intervals.
- Boundary layer meteorological analysis
- Modelling wind, temperature, pressure, and moisture profiles using dense vertical natural/hybrid model levels in the lower atmosphere.
- Atmospheric variable downscaling and simulation
- Predicting surface and atmospheric state parameters, including composite reflectivity (refc), conditioned on larger-scale atmospheric states.
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
- Predicts 99 surface and hybrid-level atmospheric variables autoregressively at an hourly frequency.
- Provides dense vertical resolution across hybrid model levels within the atmospheric boundary layer.
- Integrates 26 broad-scale atmospheric conditioning fields from ERA5 with high-resolution regional states from HRRR.
Watch out for
- Requires 125 input meteorological variables formatted across specific hybrid and pressure levels, alongside date-time arrays.
- Inference is restricted to Linux operating systems and supported NVIDIA GPU architectures (Ampere, Hopper, Turing).
- Evaluated and tested against HRRR and ERA5 domains, meaning application outside these configurations is unverified.
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 StormCast V1 commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
What variables does StormCast V1 output?
It outputs 99 four-dimensional variables, including 10-metre winds (u10m, v10m), 2-metre temperature (t2m), mean sea-level pressure (mslp), composite reflectivity (refc), and wind, temperature, humidity, geopotential, and pressure across multiple natural/hybrid model levels (hl1 to hl30).
What are the model's conditioning variables?
It accepts 26 conditioning variables from global atmospheric states, including total column water vapour (tcwv), surface pressure (sp), mean sea-level pressure, 2-metre temperature, 10-metre winds, and winds, geopotential, temperature, and specific humidity at 1000, 850, 500, and 250 hPa levels.
Which frameworks and hardware support running inference?
Inference can be executed using PyTorch and PhysicsNeMo on Linux systems equipped with NVIDIA Turing, Ampere, or Hopper GPUs, such as the A100, H100, or L40S.
How to cite it
Kilometer-Scale Convection Allowing Model Emulation using Generative Diffusion Modeling, arXiv:2408.10958
Written from StormCast V1'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.