"Recent advances in machine learning have enabled data driven models for weather and climate prediction that are approaching the skill of traditional numerical weather prediction systems. These approaches, ranging from hybrid physics and machine learning methods to full model emulation, offer dramatic reductions in computational cost and open the door to new experiments, large ensembles, and interactive workflows.
The speaker presents recent work within NCAR’s CREDIT, Community Research Earth Digital Twin, framework, focusing on stable decadal scale autoregressive prediction. They present results with CAMulator, a data driven emulator of the NSF NCAR Community Atmosphere Model, CAM, and examine its architecture, training strategy, and long horizon behavior. Results show that the model reproduces key atmospheric dynamics while maintaining stability over extended forecasts.

They then explore a pathway toward coupled data driven Earth system modeling by coupling CAMulator with a process based ocean model. This required building a robust interface between legacy Fortran infrastructure and modern Python based machine learning systems. They discuss both the technical challenges and the solutions that enabled this hybrid coupling, and present early results demonstrating the potential of such systems to accelerate Earth system prediction and experimentation.