ArchesWeatherGen: Skillful and compute-efficient probabilistic weather forecasting with machine learning.

Area of Science:

• Meteorology and Atmospheric Science
• Artificial Intelligence
• Machine Learning

Background:

• Weather forecasting is crucial for various sectors, including agriculture, logistics, renewable energy, and extreme weather preparedness.
• Deep learning models, particularly those trained on ERA5 data with a next-state prediction objective, have demonstrated significant success in weather forecasting.
• Existing numerical global circulation models are computationally intensive.

Purpose of the Study:

• To develop a methodology for creating probabilistic weather models by leveraging deterministic weather models.
• To improve the performance and reduce the computational costs associated with deep learning weather forecasting.
• To democratize the use of generative machine learning models in weather forecasting research.

Main Methods:

• Designed a probabilistic weather model utilizing flow matching, a modern variant of diffusion models.
• Trained the model to project deterministic weather predictions to the distribution of ERA5 weather states.
• Evaluated the model's performance using the WeatherBench benchmark.

Main Results:

• The proposed model, ArchesWeatherGen, surpassed existing models like IFS ENS and NeuralGCM across most WeatherBench headline variables.
• Demonstrated improved performance compared to current deep learning weather forecasting models.
• Indicated potential for reduced computing costs in probabilistic weather modeling.

Conclusions:

• The ArchesWeatherGen model represents a significant advancement in probabilistic weather forecasting using generative machine learning.
• The methodology offers a promising approach to enhance weather prediction accuracy and efficiency.
• This work contributes to making advanced machine learning techniques more accessible for weather forecasting research.