[Paper Review] ACE2: Accurately learning subseasonal to decadal atmospheric variability and forced responses

Oliver Watt‐Meyer, Brian Henn+7 more|arXiv (Cornell University)|Nov 18, 2024

Underwater Acoustics ResearchEarth and Planetary Sciences3 citations

TL;DR

ACE2 is a 450M-parameter, autoregressive machine learning emulator that accurately simulates subseasonal to decadal atmospheric variability and forced responses over 1940–2020 at 1° resolution and 6-hour temporal frequency. It conserves global dry air mass and moisture, enables stable long-term integration (1500 simulated years/day), and reproduces key phenomena like the Madden–Julian Oscillation, tropical cyclones, and the atmospheric response to El Niño and long-term warming trends.

ABSTRACT

Existing machine learning models of weather variability are not formulated to enable assessment of their response to varying external boundary conditions such as sea surface temperature and greenhouse gases. Here we present ACE2 (Ai2 Climate Emulator version 2) and its application to reproducing atmospheric variability over the past 80 years on timescales from days to decades. ACE2 is a 450M-parameter autoregressive machine learning emulator, operating with 6-hour temporal resolution, 1° horizontal resolution and eight vertical layers. It exactly conserves global dry air mass and moisture and can be stepped forward stably for arbitrarily many steps with a throughput of about 1500 simulated years per wall clock day. ACE2 generates emergent phenomena such as tropical cyclones, the Madden Julian Oscillation, and sudden stratospheric warmings. Furthermore, it accurately reproduces the atmospheric response to El Niño variability and global trends of temperature over the past 80 years. However, its sensitivities to separately changing sea surface temperature and carbon dioxide are not entirely realistic.

Motivation & Objective

• To develop a machine learning climate emulator capable of accurately simulating atmospheric variability and forced responses over subseasonal to decadal timescales.
• To enable stable, long-term integration of the emulator with exact conservation of dry air mass and atmospheric moisture.
• To train the model on both reanalysis (ERA5) and physics-based model output (SHiELD) to ensure robustness and generalization.
• To evaluate the emulator’s ability to reproduce observed climate trends, ENSO responses, and emergent weather phenomena.
• To demonstrate that the model can be used for efficient climate simulation and analysis, including ensemble exploration and scenario interpolation.

Proposed method

• ACE2 is a 450M-parameter autoregressive transformer-based model trained to predict two 6-hour steps ahead from atmospheric state variables.
• The model operates at 1° horizontal resolution with eight vertical layers and uses 6-hourly time steps with user-specified sea surface temperature (SST) and CO2 boundary conditions.
• It enforces exact conservation of global dry air mass and atmospheric moisture through architectural and loss function design.
• The model is trained on two datasets: ERA5 reanalysis (1940–1995, 2011–2019) and an AMIP-style SHiELD simulation (1940–2020).
• Checkpoint selection is based on climate skill, using global RMSE and R² metrics for time-mean and annual-mean variables.
• Evaluation includes regression of atmospheric variables against the Niño 3.4 index to assess ENSO response accuracy.

Figure 1: Global- and annual-mean series for a) 2-meter air temperature and c) total water path over 81-year evaluations of ACE2-ERA5 and ACE2-SHiELD. For each ACE2 evaluation, a three-member initial condition (IC) ensemble of the model (each initialized one day apart) is shown in solid lines, and t

Experimental results

Research questions

• RQ1 Can a machine learning emulator accurately simulate atmospheric variability across timescales from days to decades?
• RQ2 Does the model reproduce the observed atmospheric response to El Niño-Southern Oscillation (ENSO) variability?
• RQ3 Can the emulator capture long-term global temperature trends and total water path changes over the past 80 years?
• RQ4 How well does the model conserve mass and moisture over long integration periods?
• RQ5 What is the model’s performance in generating emergent phenomena such as tropical cyclones and sudden stratospheric warmings?

Key findings

• ACE2 achieves a throughput of approximately 1500 simulated years per wall clock day, enabling efficient long-term climate simulation.
• The model accurately reproduces the atmospheric response to El Niño variability, with global area-weighted RMS difference in response maps below 0.5 K for key variables.
• It captures long-term trends in global mean temperature and total water path with R² values exceeding 0.95 against reference datasets.
• ACE2 generates emergent phenomena including tropical cyclones, the Madden–Julian Oscillation, and sudden stratospheric warmings.
• The model conserves global dry air mass and atmospheric moisture exactly over long integrations, enabling stable forward simulation for arbitrarily many steps.
• Compared to physics-based SHiELD, ACE2 is ~100 times faster and 700 times less energy-intensive at 1° resolution, and ~50 times faster and 25 times less energy-intensive than a coarser C24 SHiELD simulation.

Figure 2: a) - c): Zonal- and time-mean for ACE2 (solid) and its reference datasets (dashed) over test period spanning 2001-01-01 to 2010-12-31, for selected variables. d) - f): ACE2-ERA5 time-mean biases over this time period. g) - i): ACE2-SHiELD time-mean biases over this time period. Results for