E3SM AI for Model Emulation – A Pilot Study
Artificial intelligence (AI) and machine learning (ML) are transforming weather and Earth system modeling by enabling faster, more accurate, and data-driven predictions that complement traditional physics-based models. To harness these advances, the E3SM Project plans to release AI-enabled emulators alongside each physics-based model, including both standalone component models and a fully coupled system tailored to E3SM’s scientific needs and software requirements.
To advance this effort, E3SM has launched a pilot study to develop the workflow for releasing these AI-based emulators. The study aims to pave the way for an AI-enabled version of E3SMv4, backed by clear evidence of its performance and readiness for use by the E3SM and DOE users, just like current physics-based models.
A new E3SM AI group will lead this pilot. The group brings together staff from eight E3SM teams with expertise spanning infrastructure, component and coupled model development, and scientific and statistical evaluation. Their first priority is to build a robust software framework to integrate AI-enabled emulators into the E3SM code base, followed by iterative testing and improvement. The goal: a seamless, end-to-end AI workflow that meets E3SM’s software and science requirements. Ongoing AI efforts within other groups will continue under their existing leadership, with the new AI group coordinating and tracking progress across the entire project.
This work builds on E3SM’s collaboration with the Allen Institute for AI (AI2) and its ACE emulator, which models the global atmosphere using large autoregressive neural networks with a Spherical Fourier Neural Operator (SFNO) backbone. ACE incorporates physical constraints directly into its architecture, improving conservation and stability during long simulations. It will be coupled with Samudra, a global ocean emulator, forming the ACE–Samudra system — the starting point for this new AI initiative, which will be further customized to meet E3SM’s specific science goals.
The initial members of the E3SM AI group are Oscar Diaz-Ibarra, Olawale Ikuyajolu, Noel Keen, Wuyin Lin, Naser Mahfouz, Andrew Nolan, Finn Rebassoo, Khachik Sargsyan, Claudia Tebaldi, and Jon Wolfe. While the group’s initial staffing is limited, E3SM’s task-based structure and “help where it is needed” ethos ensure that additional expertise can be brought in as needed. The AI group will work closely with relevant E3SM staff to achieve its goals and will document progress in an accessible and transparent manner. It will also collaborate with existing E3SM groups and engage with related efforts across the broader DOE community and beyond.
