# GenCast: Diffusion-based ensemble forecasting for medium-range weather

**Authors:** Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R. Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, Matthew Willson  
**arXiv:** [2312.15796](https://arxiv.org/abs/2312.15796)  
**Date:** 2023-12-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.15796v2/

## Key points

- Weather forecasts are crucial for making critical decisions
- Traditional weather forecasting relies on numerical weather prediction (NWP) algorithms
- Ensemble forecasts like ECMWF's ENS model are used to account for uncertainty in weather predictions
- GenCast introduces a machine learning weather prediction (MLWP) method that generates probabilistic 15-day global forecasts with greater skill and speed than existing NWP-based models
- GenCast produces an ensemble of stochastic forecasts for over 80 surface and atmospheric variables in just 8 minutes
- GenCast outperforms ENS on 97.4% of evaluated targets, excels in predicting extreme weather events, tropical cyclones, and wind power production
- GenCast significantly improves predictions for events like high temperature exceedance and cyclone presence across various lead times
- GenCast's enhanced accuracy in predicting tropical cyclones can help mitigate the impacts of these destructive storms
- Overall, GenCast represents a groundbreaking advancement in operational weather forecasting by providing more accurate and efficient predictions

## Summary

Weather forecasts are crucial for making critical decisions. They impact everyday activities like deciding whether to bring an umbrella and significant events such as preparing for a tropical cyclone or planning renewable energy use. However, due to the inherent uncertainty in weather predictions, it is essential to consider a range of probable scenarios and their likelihood of occurrence. Traditional weather forecasting relies on numerical weather prediction (NWP) algorithms that approximate atmospheric dynamics to generate deterministic forecasts. To account for this uncertainty, weather agencies increasingly use ensemble forecasts like the European Centre for Medium Range Weather Forecasting (ECMWF)'s ENS model. While effective at modeling different weather scenarios, NWP-based ensemble forecasts can still be prone to errors and require significant computational resources. In contrast, GenCast introduces a machine learning weather prediction (MLWP) method that leverages decades of reanalysis data to generate probabilistic 15-day global forecasts with greater skill and speed than existing NWP-based models. By utilizing a diffusion-based approach, GenCast produces an ensemble of stochastic forecasts for over 80 surface and atmospheric variables in just 8 minutes. This innovative method outperforms ENS on 97.4% of evaluated targets and excels in predicting extreme weather events, tropical cyclones, and wind power production. Furthermore, GenCast's superior performance in assessing extreme-weather prediction is evident through Brier skill scores comparison with ENS for high and low percentiles of climatological distributions. The model significantly improves predictions for events like high temperature exceedance and cyclone presence across various lead times. Tropical cyclones pose significant risks globally, causing fatalities and economic damages annually. GenCast's enhanced accuracy in predicting tropical cyclones underscores its potential impact on mitigating the impacts of these destructive storms. Overall, GenCast represents a groundbreaking advancement in operational weather forecasting by providing more accurate and efficient predictions. These predictions empower decision-makers with reliable information for better preparedness and response strategies.