Hybrid & Emulated ESMs 2027

About the workshop

Dates: 12–14 May 2027
Venue: NSF NCAR Mesa Laboratory, Boulder, Colorado
Size: approximately 50–80 participants
Format: plenary and contributed talks, a poster session and reception, walk-and-talks, and working groups

Scientific motivation

Machine learning now produces climate-relevant emulators that are fast, skillful, and increasingly stable in long rollouts; hybrid model components that augment process-based Earth system models; and the first fully differentiable physics-based climate and weather models. Coupled climate modeling, however, sets a stricter standard than matching observed metrics or forecast skill. A coupled system must exchange fluxes and state variables across components, preserve essential budgets, generate credible internal variability, and remain stable over many simulated years. Climate modeling also requires extrapolation beyond the observational record: the model must represent feedbacks and respond physically to forcing.

Current systems demonstrate important pieces of this problem — examples include CAMulator, the ACE2 family, ArchesClimate, LUCIE, Samudra, SamudrACE-style coupled atmosphere–ocean emulation, ICON-ML2, CESM3-MLe, the CliMA project, NeuralGCM, the JAX Circulation Model, and ClimSim/ClimSim-Online parameterization workflows. These are early state-of-the-art demonstrations rather than evidence that coupled hybrid modeling or coupled climate emulation is mature. Most proposed use cases remain ahead of present capability. The workshop therefore aims to define a ladder of evidence, from low-order models through idealized test cases and component emulators to coupled prototypes.

The workshop builds on, and will disseminate, work from the ClimateBench2 project and the meeting held at Ai2 in fall 2026.

Who should attend

We seek a deliberately balanced mix of climate dynamicists, Earth system model developers, machine learning scientists, research software engineers, and early-career researchers. The program gives visibility to both threads — hybrid ML-physics modeling and climate emulation — without splitting the meeting, and to the shared work at their interface: coupling, diagnostics, software, and test cases.

Format

Planned outputs

  1. Community white paper outline. An outline for a community white paper on hybrid ML-physics climate modeling and climate emulation, with named drafting leads.
  2. Benchmark and diagnostics specification. For online coupled hybrid models and emulators — conservation, drift, feedbacks, climate sensitivity, internal variability, extremes, and out-of-distribution forcing.
  3. Coupling interface checklist. For atmosphere, ocean, sea ice, land, chemistry, radiation, and coupler time-stepping.
  4. Lower-order and idealized test-case ladder. Spanning toy systems, single-column models, aquaplanets, slab oceans, AMIP-style tests, and coupled prototypes.
  5. Coupled model intercomparison pilot plan. Covering 3–5 candidate systems for coupled emulation and coupled runs, with 6–12 month milestones.
  6. Shared failure-mode catalog. For emulator coupling, including unstable rollouts, muted variability, energy leaks, nonphysical feedbacks, and regime-locking.

Background reading