End-to-end data-driven weather prediction

Area of Science:

• Artificial Intelligence
• Atmospheric Sciences
• Computational Science

Background:

• Accurate weather prediction is vital for numerous sectors and public safety.
• Machine learning (ML) has enhanced numerical weather prediction (NWP) but often still relies on NWP components.
• Existing ML models face limitations in speed and accuracy due to dependence on NWP.

Purpose of the Study:

• To develop and evaluate an end-to-end, data-driven machine learning system for weather prediction.
• To demonstrate that a single ML model can replace the entire NWP pipeline.
• To assess the performance of this system against established NWP baselines.

Main Methods:

• Developed Aardvark Weather, an end-to-end data-driven system ingesting observational data.
• Generated global gridded forecasts and local station forecasts using the ML model.
• Compared Aardvark Weather's forecasts against an operational NWP baseline and a state-of-the-art system.

Main Results:

• Aardvark Weather's global forecasts surpassed the operational NWP baseline for multiple variables and lead times.
• Local station forecasts demonstrated skill up to ten days, rivaling post-processed NWP and human-assisted systems.
• End-to-end tuning further enhanced the accuracy of local forecasts.

Conclusions:

• Skilful weather forecasting is achievable without reliance on NWP at deployment.
• Data-driven, end-to-end ML models offer significant speed and accuracy benefits.
• Aardvark Weather represents a new generation of weather prediction models, reducing costs and enabling customization.