Themes and readiness scale
Workshop themes
Hybrid ML-physics model development
Where should ML enter the Earth system model — parameterization, full component, correction term, stochastic closure, solver accelerator, data-assimilation layer, or ensemble generator?
Climate emulators and surrogate systems
What is being emulated — a process tendency, a component, a full-state trajectory, a coupled system, a response surface, or an ensemble distribution?
Coupling as the bridge
How do hybrid components and emulator components exchange fluxes and states without hiding physical inconsistencies?
Generalization and climate change
How should emulators be trained and evaluated for forcing levels, mean states, regimes, extremes, and feedbacks that are absent or rare in training data?
Conservation, stability, and long integrations
Which constraints are essential, which can be corrected diagnostically, and which failures are scientifically disqualifying?
Training data, benchmarks, ensembles, and calibration
Which simulations and perturbation ensembles teach an emulator the climate response rather than only the historical attractor? Which common diagnostics compare systems fairly across grids, components, training data, and scientific goals while evaluating uncertainty, physical diversity, and calibration?
Software, reproducibility, and interpretability
What does a sustainable and scientifically auditable hybrid model stack require in coupler APIs, inference serving, data standards, tests, provenance, and community governance? How do we determine whether a model has learned climate physics, a bias correction, or a fragile statistical shortcut?
Readiness scale
Talks, posters, and workshop products will be framed against a simple readiness scale. The scale lets participants be candid about current capability without downplaying progress: it guards against overclaiming while making room for early but scientifically valuable systems.
| Level | Status | Evidence expected |
|---|---|---|
| R0 | Conceptual or offline prototype | Clear target process or component, training data identified, offline diagnostics started, known physical constraints stated. |
| R1 | Offline replay or tendency testing | Skill against held-out simulations, physical-range checks, conservation tendencies, variable and unit conventions, sensitivity to input distribution. |
| R2 | Online component in prescribed-boundary mode | Stable multi-year rollout with prescribed boundary conditions, seasonal cycle, spectra, process budgets, restart tests, and documented failure modes. |
| R3 | Idealized or reduced coupled setting | Demonstrated behavior in lower-order models, SCM/RCE, aquaplanet, slab-ocean, shallow-water/QG, or other controlled coupled tests. |
| R4 | Limited coupled prototype | Interactive coupling with budget monitoring, drift analysis, response to forcing perturbations, variability diagnostics, and sensitivity to coupling frequency. |
| R5 | Intercomparison candidate | Reproducible configuration, documented training data, standard diagnostics, restart files, model card, compute-cost estimate, and willingness to run shared benchmark experiments. |
| R6 | Multi-decadal / centennial coupled response | Stable 100+ year rollouts, conservation closure, unforced natural variability (ENSO, AMOC, and MJO power spectra), and forced-response verification (4×CO2 or 1% CO2 yr−1 sensitivity). |