24AI Hindcasting with Google DeepMind's GraphCast: A 45-Year Retrospective Analysis from 1979 to 2024

This study presents a comprehensive hindcast dataset generated using the Google DeepMind Graphcast model, covering the period from 1979 to 2024. This effort represents a significant advancement in leveraging artificial intelligence for atmospheric science, utilizing a state-of-the-art deep learning model to produce daily 15-day forecasts initialized at 00 UTC. The hindcasts provide valuable insights into the model's performance across different climate regimes and periods, allowing for an extensive evaluation of long-term forecast skills and biases.

The dataset was generated by initializing the Graphcast model daily at 00 UTC, producing 15-day forecasts for the entire 45-year period at a high resolution of 0.25° x 0.25° globally. Notably, the model demonstrates remarkable computational efficiency, requiring less than 4 minutes to complete a 15-day forecast on Texas Advanced Computing Center's NVIDIA H100. This consistent and high-resolution hindcast archive is a critical resource for evaluating the evolution of weather prediction models, assessing changes in forecast accuracy over time, and understanding the impacts of various meteorological phenomena.

Key findings include the identification of model strengths and weaknesses in capturing vital atmospheric processes, particularly in forecasting extreme weather events such as hurricanes, heatwaves, and cold spells. Additionally, we explore the implications of these hindcasts for climate research and equity, reanalysis comparisons, and future improvements in AI-based weather prediction models.

This work underscores the potential of AI-driven models in advancing meteorological science and provides a robust framework for future research in hindcasting and predictive analytics.