Advances in Machine Learning for Climate and Weather Simulation
September 05, 2024
Mike Pritchard
Hosted by Kristen Rasmussen
Abstract
In the past two years, the use of machine learning (ML) in atmospheric simulation has rapidly advanced, leading to significant breakthroughs. This talk will present highlights from collaborative research involving NVIDIA, UC Irvine, UC Berkeley, LBNL, and the LEAP NSF Science & Technology Center at Columbia. I will cover:* Hybrid Physics-AI Climate Simulation: Advances in ML-based parameterization of explicit embedded convection, and takeaways from involving ML researchers through scalable benchmarks.
* Global Autoregressive Synoptic Weather Prediction: Discoveries of emergent physics, strategies for ensemble calibration, and pioneering efforts in using “huge ensemble” counterfactual hindcasts to assess extreme event risk.
* Ocean-Coupled Global Subseasonal Prediction: Multi-component autoregressive Earth System modeling and first evidence of learnt El Niño dynamics.
* Generative AI for km-Scale Super-Resolution & Score-Based Data Assimilation: Using diffusion models from image and video domains for statistical downscaling and simplifying data assimilation.
* Regional Autoregressive km-Scale Weather Prediction: A related breakthrough in modeling convective dynamics with generative AI relevant to fully emulating cloud-resolving simulations.