AI for weather & climate physics applications: Advances from planetary to km-scales

AI has rapidly changed the paradigm for simulating atmospheric dynamic from planetary to storm-resolving scales. In the first part of this talk I will review how a deterministic global AI weather model developed at NVIDIA has been re-calibrated for probabilistic prediction to generate huge ensemble counterfactuals of historical heat waves, relevant for climate risk calibration to low likelihood / high impact event exposure. Next, I will review emerging NVIDIA technologies based on generative AI – for (i) spatial downscaling and new channel synthesis, (ii) its extension to ambitious domain sizes, (iii) dynamical downscaling and mesoscale forecasting and (iv) generative data fusion. I will conclude with some remarks on the exciting potential for end-to-end AI forecasting systems that portend a more interactive and computational efficient paradigm for simulating atmospheric and climate physics.