NSF NCAR Workshop · 12–14 May 2027
Emulated and Hybrid ML-Physics Earth System Models
What sequence of component demonstrations and coupled prototypes would make hybrid models and climate emulators trustworthy enough for coupled climate science?
This three-day workshop at the NSF NCAR Mesa Laboratory brings together two overlapping communities: hybrid ML-physics model developers, who place statistically learned estimators inside or alongside process-based Earth system models, and climate emulator developers, who use learned models to represent components, coupled systems, or climate responses. Both communities combine physical modeling and statistical learning but distribute the predictive burden between them differently. The workshop asks what each approach can contribute to the other.
The scope covers hybrid ML-physics Earth system models, differentiable Earth system models, and Earth system emulators. It is distinct from weather emulation and short-range forecasting: the emphasis is on coupled climate, long integrations, conservation, internal variability, and the response to forcing.
Talks are concentrated in plenary and contributed sessions, leaving substantial time for structured discussion, technical design, walk-and-talks, and small-group drafting. Participants will leave with a community roadmap: a coupling benchmark specification, an agreed diagnostic suite for stability and generalization, a hierarchy of lower-order and idealized test cases, and a practical plan for staged intercomparison across hybrid models and emulators.
Read more about the workshop → · Themes and readiness scale → · Agenda →