[Paper Review] ACE: A fast, skillful learned global atmospheric model for climate prediction

Oliver Watt‐Meyer, Brian Henn+4 more|arXiv (Cornell University)|Oct 3, 2023

Meteorological Phenomena and Simulations29 citations

TL;DR

ACE is a 200M-parameter SFNO-based emulator of a 100-km global atmospheric model, achieving 100x faster run-time and near-conservation of moisture, with stable 100-year forecasts and strong climate fidelity.

ABSTRACT

Dataset for Ai2 Climate Emulator This dataset contains a minimal example set of files and configuration to use for inference with the Ai2 Climate Emulator (ACE). Please see https://github.com/ai2cm/ace to install the necessary software. The included checkpoint is the same ace checkpoint as referenced in (https://zenodo.org/records/10791087). See README.md for a description of the included files.v1.1: The initial condition zarr stores were erroneously missing all values in the first upload. These files have been fixed in this update.

Motivation & Objective

• Demonstrate that a neural emulator can provide long-term, physically consistent climate predictions.
• Show stability and near-conservation of moisture in long simulations (up to 100 years).
• Assess ACE’s ability to reproduce the reference physics-based model’s climate under standard forcing.
• Evaluate generalization to unseen sea surface temperature forcing and compare to a coarser baseline.

Proposed method

• Use the Spherical Fourier Neural Operator (SFNO) to predict 6-hour ahead atmospheric states on a 100 km global grid.
• Train on an 11-member FV3GFS reference ensemble with 10 years of data, using annually repeating SST and fixed greenhouse/gas conditions.
• Classify variables into prognostic inputs/outputs, forcing inputs, and diagnostic outputs to enable physical constraints like moisture conservation.
• Apply residual scaling to normalize variables so prediction errors contribute equally to the loss across fields.
• Evaluate using time-dependent global means, RMSE, and biases against a 2x-coarser FV3GFS baseline and against the reference emulator.
• Explore zero-shot generalization to historical SST forcing and assess stability over 100-year forecasts.

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.

Experimental results

Research questions

• RQ1 Can ACE reproduce the reference physics-based atmospheric model’s climate across multiple variables at 100 km resolution?
• RQ2 Does ACE maintain long-term stability and conserve column moisture with minimal explicit constraints?
• RQ3 How does ACE perform relative to a 2x coarser FV3GFS baseline in key climate metrics?
• RQ4 Is ACE robust to unseen sea surface temperature forcing outside its training distribution?
• RQ5 What are the computational and energy efficiencies of ACE compared to the reference model?

Key findings

• ACE achieves stable simulations for at least 100 years and reproduces seasonal cycles in near-surface temperature and total water path.
• ACE has a global RMSE reduction for 41 of 44 output variables compared to the 2x coarser baseline.
• ACE nearly conserves column moisture; global moisture budgets show small violations relative to typical precipitation/evaporation scales.
• ACE runs about 100x faster and uses about 100x less energy than the reference FV3GFS model under typical resources.
• Zero-shot forcing with historic SST data demonstrates stability, with some biases (e.g., cold land temperatures) but overall modest RMSE and bias."
• Most outputs show lower time-mean RMSE for ACE versus the baseline (Figure 7).

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 6 and 7 ).