ACE2: Accurately learning subseasonal to decadal atmospheric variability and forced responses

TL;DR

ACE2 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.

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.

Figures (23)

• 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 the reference dataset is shown in dashed lines (e.g., ACE2-ERA5 vs. ERA5 itself). The validation and test periods are shaded in dark gray and light gray, respectively. As a baseline, the ACE-climSST model WattMeyer2023 forced with the historical SST is also shown for total water path (2-meter air temperature was not predicted by this model). The "forced SST" in a) is the prescribed SST averaged over 45$^{\circ}$S to 45$^{\circ}$N in the SHiELD simulation (ERA5 SSTs are similar though not identical). The R$^2$ of the 81-year series are shown in b) and d). For ACE2-SHiELD, the skill metrics for each of four trained models are shown. Error bars indicate the range over three IC ensemble members for each model. SHiELD reference variability is the R$^2$ computed between the two SHiELD ensemble members.
• 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 a single initialization of each ACE2 model are shown.
• Figure 3: Global RMSE between the time-mean of ACE2 and its reference dataset (ERA5 or SHiELD). Error bars indicate the 95% confidence interval based on the IC ensemble. Also included are NeuralGCM error against ERA5, SHiELD reference variability, the error of ACE-climSST evaluated against the SHiELD dataset, and the error of the SHiELD simulations against ERA5. ACE-climSST did not predict 2-meter temperature or 500hPa height. NeuralGCM Kochkov2024 results are only available for total water path for a single year (2020), and so we also show 2020-only results of ACE2-ERA5.
• Figure 4: Maps of regression coefficients of predicted and reference dataset surface precipitation against the Niño 3.4 index over the 10-year test period. Single model initializations are shown. a) ACE2-ERA5, b) ACE2-SHiELD, c) ACE-climSST evaluated on SHiELD, d) ERA5 reference, e) SHiELD reference, all for the 10-year test period. Titles of panels a-c indicate the RMSE of the predicted map against its reference map; the numbers in parenthesis are for the two other initializations that are not shown. For e), the SHiELD reference variability is calculated as the RMSE between the regression coefficient maps of the two ensemble members.
• Figure 5: Tracks of tropical cyclone-like features over the 2001-2010 period for a) the IBTrACS dataset, b) ERA5, c) the C96 SHiELD model and d) ACE2-ERA5. The tracks for b)-d) are determined based on minima in sea-level pressure along with maxima in upper troposphere temperature. See main text for details. The average number of tropical cyclones across the globe per year is shown in the title of each panel, although the IBTrACS dataset is not directly comparable to the detections in other panels which use a tracking algorithm applied to 1° resolution data.