WeatherNext models are global medium-range forecasting models with broad applications across diverse use cases and industries. If your work currently relies on models such as ECMWF HRES/ENS or NOAA GFS, or if you utilize ERA5/HRES Analysis as ground truth, WeatherNext offers a compelling alternative for enhanced forecasting capabilities.
Key use cases for WeatherNext models and forecasts
- Global Medium-Range Forecasting: WeatherNex models predict hundreds of atmospheric variables globally at high accuracy and resolution.
- Severe Event Prediction: WeatherNext models can provide improved predictions for certain severe events compared to ECMWF HRES/ENS. Some examples that have already been already studied are cyclones, atmospheric rivers, and extreme temperatures. See our research for detailed methodology.
- Input for Downstream Applications: WeatherNext outputs can be utilized as inputs into other downstream applications, systems and models. Examples include ensemble forecasting systems and flood forecasting models or using the forecasts as inputs for specialized regional models.
- Uncertainty Quantification: Probabilistic models such as WeatherNext 2 and WeatherNext Gen can generate large ensembles of possible weather trajectories (up to 64 in forecast dataset, larger ensembles available using Vertex AI) to better characterize predictive uncertainty and estimate risk, which is vital for critical decision-making.
- Fast Inference Speed: WeatherNext models boast significantly faster inference speeds, capable of generating a 15-day forecast in just a few minutes on single TPU. This can enable new applications, especially with the integration of observational data. For more info on forecast latency, see the Dissemination schedule.
Limitations
- Targeting operational analysis: Existing WeatherNext models target global operational analysis as ground truth, which has limited resolution, and other types of biases (e.g. wouldn't typically exactly match what one would measure on the ground). For some applications that require on the ground accuracy for measurements that don't match ERA5/HRES-fc-0 very well in the first place, it may be necessary to apply bias-correction techniques to the outputs.
- Blurring: Deterministic ML models like WeatherNext Graph tend to produce forecasts that are blurred, particularly at longer lead times.
- Specific Variable Limitations: Core WeatherNext model outputs don't predict variables like precipitation rate, 2-meter dew point temperature, irradiance, or cloud fraction at this time.
- Precipitation Data Quality: WeatherNext model precipitation outputs target ERA5 precipitation data, which has certain limitations and biases and is often excluded from main evaluations (see papers for details). The same limitations and biases may apply to the precipitation forecasts.
- Artifacts in Predictions: WeatherNext 2 can exhibit subtle artifacts in the forecast states, sometimes appearing as 'honeycomb' patterns corresponding to the processor's mesh structure, particularly in higher frequency variables. Some minor build-up in high-frequency signal at longer lead times is also observed.