The Equilibrium Response of Atmospheric Machine-Learning Models to Uniform Sea Surface Temperature Warming

TL;DR

While the state-of-the-art ML models reproduce key aspects of the physical model response, particularly the response of precipitation, some exhibit notable departures from robust physical responses, including radiative responses and land region warming.

Abstract

Machine learning models for the global atmosphere that are capable of producing stable, multi-year simulations of Earth's climate have recently been developed. However, the ability of these ML models to generalize beyond the training distribution remains an open question. In this study, we evaluate the climate response of several state-of-the-art ML models (ACE2-ERA5, NeuralGCM, and cBottle) to a uniform sea surface temperature warming, a widely used benchmark for evaluating climate change. We assess each ML model's performance relative to a physics-based general circulation model (GFDL's AM4) across key diagnostics, including surface air temperature, precipitation, temperature and wind profiles, and top-of-the-atmosphere radiation. While the ML models reproduce key aspects of the physical model response, particularly the response of precipitation, some exhibit notable departures from robust physical responses, including radiative responses and land region warming. Our results highlight the promise and current limitations of ML models for climate change applications and suggest that further improvements are needed for robust out-of-sample generalization.

Figures (11)

• Figure 1: Annual mean surface air temperature for (left) the climatology and (right) the response to a uniform +2 K SST perturbation for cBottle, ACE2, AM4, and NeuralGCM from top to bottom. ERA5 is included in panel (e) for reference of the mean state. Note that NeuralGCM does not directly output surface air temperature. Here we use temperature at 1000 hPa to illustrate the temperature pattern near the surface.
• Figure 2: Annual mean precipitation for (left) the climatology and (right) the response to a uniform +2 K SST perturbation for cBottle, ACE2, NeuralGCM and AM4 from top to bottom. GPCP is included in panel (e) for reference.
• Figure 3: The zonal mean change of precipitation minus evaporation for ACE2, NeuralGCM and AM4 from top to bottom. The dashed lines in each panel is the thermodynamic component,approximated as $\alpha \Delta T(\bar{P}-\bar{E})$ with $\Delta T=2K$.
• Figure 4: Zonal mean 99.9th percentile of daily precipitation for (a) cBottle, (b) ACE2, (c) NeuralGCM, and (d) AM4. Results are shown for the control simulation (black) and the +2K simulation (red).
• Figure 5: Daily mean precipitation vs binned column water vapor in the tropics (30°S–30°N) for (a) cBottle, (b) ACE2, (c) NeuralGCM, and (d) AM4 for the control simulation (blue) and the +2K simulation (orange).