Escalating adventures in using ML to parameterize explicit convection in climate models and new perspectives from weather forecasting in industry.
* Register (or log in) to the AI4G Neural Network to add this session to your agenda or watch the replay
Mike Pritchard, Director for Climate Simulation Research at Nvidia and Professor of Earth System Sciences, University of California, Irvine, will begin by discussing the problem that drew him to machine learning (ML), in academia – outsourcing the full physics package of computationally expensive superparameterized climate simulations to ML, focusing on latest attempts to discover what it takes to achieve reliable, reproducible prognostic stability and skill. This part will conclude with an outlook on the current potential of multi-scale climate models to generate useful training datasets for ML parameterization, emphasizing advances in simulation algorithms that are beginning to achieve realistic stratocumulus clouds at high interior grid resolution, powered by GPU supercomputers. Mike Pritchard will then switch themes to discuss his new outlook from escalating adventures in industry, with Nvidia, focusing on the intriguing potential of fully data-driven transformer-based ML methods to entirely subsume global atmospheric prediction models. Highlights will include the impressive rate of increases of ML weather prediction skill over the past year, both at Nvidia and by other industrial players, the associated outlook for massive ensemble forecasting, and potential for such methods to solve compression and latency problems in making climate model intercomparison data archives more useful to stakeholders, as part of Nvidia’s “Earth-2” initiative.
This live event includes a 30-minute networking event hosted on the AI for Good Neural Network. This is your opportunity to ask questions, interact with the panelists and participants and build connections with the AI for Good community.