The Ai2 Climate Emulator (ACE2): Learning atmospheric variability and forced response on seasonal and decadal timescales

Deep learning models have demonstrated state of the art performance for medium range global
weather prediction. However, their abilities to accurately represent atmospheric variability and
forced responses on time scales spanning from days to decades have not been explored. Here,
we present the Ai2 Climate Emulator version 2 (ACE2), versions of which are trained on a
global atmosphere-land model (AGCM) and on an observational reanalysis (ERA5).
The training datasets consist of prognostic (temperature, humidity, wind, and surface pressure)
and diagnostic (e.g., surface precipitation and radiative fluxes) variables at 6-hourly frequency,
coarsened to 1° horizontal resolution and 8 terrain-following finite volume vertical layers. ACE2
is based on NVIDIA’s Spherical Fourier Neural Operator architecture and is trained to step the
atmospheric state forward by 6 hours. Forcing variables include sea surface temperature (SST)
and insolation, while diagnostic variables include the surface and top of atmosphere energy and
moisture fluxes necessary for interpretability and budget calculations. In an update since ACE
version 1, new explicit constraints enforce exact conservation of global dry air mass and
atmospheric moisture.
Using either ERA5 or the AGCM (GFDL’s SHiELD version of NOAA’s operational global weather
forecast model) as training and evaluation data, ACE2 makes stable and accurate 80-year
simulations initialized on 1 January 1940—over 100,000 autoregressive forward steps simulated
in 1.5 hours on a single H100 GPU. Annual-mean near-surface air temperature and total water
path track the expected response to the warming sea surface temperatures. Evaluating on ten
held-out test years, time-mean biases are small. For example the global RMSEs of time-mean
surface precipitation rate are 0.3-0.4 mm/day for both ERA5 and the AGCM, which is smaller
than the difference between them over this period. Similar results were found when training
ACE2 on data from a different AGCM, the DOE EAMv2.
On seasonal timescales, the circulation and precipitation response of the atmosphere to El Niño
SST variability is accurately emulated. In the tropics, ACE2 faithfully reproduces equatorial
waves across the frequency and wavenumber spectrum seen in ERA5, and has variability with
similar characteristics to the Madden Julian Oscillation. Finally, unlike other existing deep
learning atmospheric emulators, we are able to evaluate stratospheric variability since ACE
includes a vertical layer representing the mid- to upper-stratosphere. The stratospheric polar
vortex exhibits the expected seasonal cycle and hemispheric asymmetries in terms of strength
and variability, but the quasi-biennial oscillation is intermittent and does not reliably have proper
amplitude or period.
We conclude that ACE2 reproduces, with reasonable fidelity, the atmospheric response to
forcing from SSTs on seasonal and annual timescales. This represents a necessary step for
using deep learning models to forecast longer than medium-range weather, as well as for
training such models to simulate the response to increasing greenhouse gas concentrations in a
coupled ocean-atmosphere system.