NVIDIA Earth-2: 2024 AI Research Highlights Spanning Emulation, Forecasting, Huge Ensembles, Downscaling & Data Assimilation.

NVIDIA Earth-2 continues to develop and drive AI technologies that transform how weather and climate are simulated and accelerate the extraction of actionable information from vast quantities of data. We summarize open-source research accomplishments highlighting breakthroughs in successful probabilistic huge ensemble calibration of our deterministic global AI weather prediction system FourCastNet, a new ocean-coupled autoregressive seasonal forecasting system prototype, our latest advances in applying generative AI for kilometer-scale data assimilation, statistical downscaling, and emulation of cloud-resolving simulations, and step-shifts in hybrid multi-scale climate simulation. We stitch together these breakthroughs laying out a vision for end-to-end km-scale AI-based forecasting in the context of weather. We also demonstrate the tools and technologies that NVIDIA is building to accelerate scientific breakthroughs and enable easy access and use of these technologies for stakeholders. To conclude, we emphasize the potential of generative AI, in particular diffusion models, in tackling increasingly important challenges in weather and climate.