Can an AI-based coupled ocean-atmosphere emulator of a global climate model correctly simulate its natural climate variability and climate response to radiative forcing?

Ai2 and M2LiNES have partnered to develop the SAM-ACE coupled emulator, trained on simulations with a reference climate model, GFDL's AM4. SAM-ACE has 8 terrain-following layers spanning the atmosphere (including the stratosphere), 19 constant-depth layers spanning the ocean, a simple AI sea-ice scheme, and a 1 degree lat/lon horizontal grid. The atmosphere has a 6-hour rollout step, and the ocean has a 5 day rollout step. The atmosphere model is based on SFNO; the ocean model has a U-Net architecture. When trained on 160 years of outputs from a reference CM4 pre-industrial simulation, SAM-ACE reproduces its time-mean climate (including its seasonal cycle), and the dominant period and amplitude of ENSO. SAM-ACE requires 1000x less computation per simulated year than does the AM4 simulation that it is trained on. We are currently training SAM-ACE on model outputs from changing climates to see if it can be used to efficiently create accurate large ensembles of simulations for typical cases used in the Coupled Model Intercomparison Project (CMIP), which provides most the climate projections that are used in IPCC climate assessments done roughly every seven years.

Bio

Chris Bretherton directs a climate modeling team at Ai2 in Seattle which uses AI trained on global climate and global storm-resolving model output and observational data to improve climate model simulations. He is an Emeritus Professor of the Atmospheric Science and Applied Mathematics Departments at the University of Washington, where for 35 years he studied cloud formation and turbulence and improved their simulation in atmospheric models. He is an American Meteorological Society Charney Award winner, IPCC author, AMS and AGU Fellow, and a member of the National Academy of Sciences.