Computational Model Studio · Model release
NeuralGCM v1 1.4° Stochastic
Simulates atmospheric evolution using a released hybrid machine-learning and physics checkpoint.
Developer: Google Research / NeuralGCM team · Last verified:
What it is designed to do
Advance atmospheric state through time from initial weather fields with required sea-surface temperature and sea-ice forcing.
- Primary contribution
- simulation
- Input
- spatial, time series
- Output
- spatial, time series (simulated atmospheric state trajectory)
Documented uses: Weather forecasts and seasonal-scale atmospheric simulations under documented forcing conditions.
Access and deployment
- NeuralGCM v1 1.4° stochastic checkpoint — available. Google Cloud Storage checkpoint described in the publisher checkpoint list. Artifact documentation Local hardware requirements have not been verified. Creative Commons Attribution-ShareAlike 4.0 International.
Technical profile and review notes
- Construction
- hybrid
- Representation
- neural, numeric
- Normalized architecture tags
- No normalized architecture tag assigned
- Lifecycle
- Unknown from reviewed sources
Inclusion rationale: Adds a bounded hybrid simulation model and explicit external forcing requirement, unlike a general simulation framework.
Limitations
- The model simulates the atmosphere only; users must supply sea-surface-temperature and sea-ice forcing.
- The publisher says this unmodified 1.4° stochastic checkpoint is reliably usable for about six months, not arbitrary multi-decade runs.
- The model requires the NeuralGCM inference runtime and prepared atmospheric state data.
Sources and related Studio records
Sources
Connections
- In areaComputational Models
- Supports functionSimulate atmospheric dynamics
- ArtifactNeuralGCM v1 1.4° stochastic checkpoint