Abstract:

I will review how generative AI is revolutionizing tasks across the Earth System Modeling stack, focusing on important developments from the past year from NVIDIA Research. First, I will share tricks for performance optimization to minimize the inference cost when using corrective diffusion (CorrDiff) in AI for multivariate super-resolution and channel synthesis. Next, I will introduce how conditional diffusion models can be used in place of autoregressive emulators of GCMs for controllable climate informatics to efficiently generate samples with desired extreme events at locations of interest. This includes how to reconfigure noise schedules inherited from computer vision architectures and how to scale these techniques to the 12.5M pixel ambition of global cloud resolving climate models — a recipe called “Climate in a Bottle”. Finally I will introduce three AI weather breakthroughs from January 2026: (i) in medium-range forecasting with skill surpassing that of GenCast using multi-scale diffusion transformers, (ii) in convection-resolving nowcasting under direct observation space with skill surpassing the US HRRR, and (iii) in AI-based global data assimilation, demonstrating the promise and still unfulfilled potential of this research frontier for end-to-end AI forecasting.