ACE: A fast, skillful learned global atmospheric model for climate prediction

TL;DR

ACE demonstrates that a skillful, physically mindful climate emulator can be trained to reproduce a high-fidelity atmospheric model with long-term stability, achieving around 100x speedup and comparable climate fidelity. By using the Spherical Fourier Neural Operator and a carefully designed prognostic-forcing-diagnostic setup, ACE preserves mass/moisture budgets and produces realistic surface fluxes while enabling zero-shot generalization to unseen SST forcing. The work shows strong long-term stability, favorable climate biases relative to a baseline, and substantial computational efficiency, highlighting the potential for democratizing climate modeling. Remaining challenges include generalization to broader forcing regimes, explicit conservation constraints, and coupling with ocean/land components to form a complete climate system emulator.

Abstract

Existing ML-based atmospheric models are not suitable for climate prediction, which requires long-term stability and physical consistency. We present ACE (AI2 Climate Emulator), a 200M-parameter, autoregressive machine learning emulator of an existing comprehensive 100-km resolution global atmospheric model. The formulation of ACE allows evaluation of physical laws such as the conservation of mass and moisture. The emulator is stable for 100 years, nearly conserves column moisture without explicit constraints and faithfully reproduces the reference model's climate, outperforming a challenging baseline on over 90% of tracked variables. ACE requires nearly 100x less wall clock time and is 100x more energy efficient than the reference model using typically available resources. Without fine-tuning, ACE can stably generalize to a previously unseen historical sea surface temperature dataset.

Figures (15)

• Figure 1: Global mean timeseries of (top) near-surface air temperature $T_7$ and (bottom) total water path computed as $\mathrm{TWP}=\frac{1}{g}\sum_k q_k^T \, dp_k$. For clarity, the daily average is plotted.
• Figure 2: 10-year mean bias in surface precipitation rate. Titles show global and time-mean RMSE and bias in units of mm/day (Equations \ref{['eq:globaltimermse']} and \ref{['eq:globaltimebias']}).
• Figure 3: Snapshot of the terms in the column moisture budget (Equation \ref{['eq:twp']}) one year into simulation for (top) reference data and (bottom) ACE simulation. Given chaotic nature of atmosphere, we do not expect details to match between the reference and ACE simulations. If column-integrated moisture is exactly conserved, the rightmost column should equal zero, as it is for the reference data.
• Figure 4: ACE forced with a realistic SST dataset spanning 1990-2020. Showing (left) timeseries of annual mean and global mean $T_7$, with individual lines corresonding to different initial conditions and (right) time-mean bias of $T_7$ for a single initial condition with time-mean RMSE and global bias (Equations \ref{['eq:globaltimermse']} and \ref{['eq:globaltimebias']}) reported in the title. Here, the "reference" is the physics based model (FV3GFS) run for the same period and forced by the same SST pattern.
• Figure 5: Diagram summarizing the flow of input and output variables. Prognostic variables are fed back into the model autoregressively. Forcing variables are read from an external dataset and appended to the prognostic variables at each step. The network outputs diagnostic variables, which contribute to the loss but are not passed back as inputs for the next step.