Foundation Models as New Backbone for Medium-range Forecasting and Data Assimilation

Explore how foundation models are revolutionizing medium-range weather forecasting and data assimilation in this 44-minute conference talk by NASA's Katherine Breen. Discover the competitive performance of machine learning models trained on large reanalysis datasets compared to traditional numerical weather prediction systems through rigorous quantitative evaluation using standard forecast skill metrics across multiple atmospheric variables, pressure levels, and regions. Learn about NASA's pioneering research efforts to develop machine learning-based data assimilation methods that combine learned representations with scientific transparency, physical consistency, and reproducibility within operational forecast systems. Examine both the strengths and current limitations of state-of-the-art foundation models in weather prediction, and engage with critical discussions about the growing influence of privately developed forecasting models versus the need for open, independently evaluated, and scientifically grounded model development. Gain insights into the future intersection of artificial intelligence and atmospheric science, including considerations about proprietary versus scientific objectives in weather and climate research applications.