On Friday, November 10th, we will have an OCP Seminar given by Dr. Andre Perkins, a Resarch Engineer at the Allen Institute for Artifical Intelligence.
This will be a hybrid in-person/zoom seminar taking place in the Geochemistry Seminar Room on the first floor of Comer. Please email the event contact for the zoom information. The title and abstract are provided down below.
Title: Improving Climate Simulations Using Machine Learning
Climate simulation is a complex and expensive task, and current climate models still have significant biases in their simulations of important climate variables such as precipitation, clouds, and convection. Observational-based simulation and fine-grid simulations provide data sources with reduced biases, but they are limited in their utility for long-term simulation. Machine learning (ML) provides a mechanism to bridge these gaps and provide relevant information for climate timescales.
In this talk, I will present an overview of the AI2 climate modeling team's research in partnership with GFDL and NVIDIA on using ML to improve climate simulations. I will discuss our work on two promising approaches: (1) hybrid modeling, which combines a physics-based model with ML corrections based on observations or a high-fidelity reference, and (2) full emulation, which trains an ML model to directly predict the output of a climate model.
Our hybrid modeling approach can improve a climate model's precipitation biases by approximately 30% and works across multiple climates. Our emulation approach has resulted in a climate emulator that is stable for at least 10 years and retains realistic weather variability and an accurate representation of the climate. These successes are a first step towards providing efficient, accurate, and trustworthy improvements to climate models, but many challenges still remain.