3 Model Families
cBottle, ACE2-ERA5, and NeuralGCM
Model Development and Evaluation
Climate Emulators
Cross-architecture emulator evaluation shows that mean-state agreement alone is insufficient; process-level and distributional fidelity control predictive reliability under warming.
Generative + Autoregressive + Hybrid
Uniform SST Warming
Out-of-Sample Diagnostics
10-Year Runs
Control and warming trajectories for emulator stress tests
ArXiv (2025)
Equilibrium response to uniform sea-surface-temperature warming
Related Publication
Zhang, B., and T. M. Merlis, 2025: The Equilibrium Response of Atmospheric Machine-Learning Models to Uniform Sea Surface Temperature Warming. arXiv, 2510.02415.
Scientific Logic
- Question: How reliably do climate emulators reproduce baseline climate structure and warming responses relevant to impacts?
- Method: Cross-model evaluation of generative, autoregressive, and hybrid emulator configurations against reference model fields in control and warming-like states.
- Mechanism: Architecture-dependent skill in preserving spatial covariance, variability, and extremes determines whether mean-state agreement translates to trustworthy anomaly responses.
- Main Findings: Some emulator families match global means but diverge regionally and in distribution tails, motivating process-aware constraints before impact attribution.
Scientific Focus
How reliably do next-generation climate emulators reproduce mean-state and warming-response behavior, particularly when evaluated outside their training distribution?
Model Set
- cBottle: non-autoregressive diffusion framework generating hourly fields from SST forcing.
- ACE2-ERA5: autoregressive SFNO architecture at 1 degree / 6-hour resolution.
- NeuralGCM: hybrid dynamical-core plus neural-physics model enabling stable decadal runs.
Key Findings
- Different emulator families show distinct strengths in spatial pattern, variability, and tail-behavior reproduction.
- Skill in control climate does not guarantee robust response under warming perturbations.
- Process-aware diagnostics are required before using emulators for attribution-sensitive applications.
Figure
Comparison (Extracted)
