The newest member of the Ai2 Climate Emulator (ACE) family, and the first ACE model filed here: a stochastic spherical Fourier neural operator on a 1° grid with a 6-hour step and an energy-conservation constraint, trained to emulate NOAA's SHiELD atmosphere model about 68 times faster than the physical model. The accompanying paper addresses a failure of earlier ACE versions, which were accurate only where sea surface temperature and CO2 were correlated in training and behaved unphysically under a +4 K SST perturbation or an abrupt CO2 quadrupling; the "+" training separates the two forcings so the emulator responds correctly to each. Weights released 10 September 2026 under Apache 2.0; the parameter count is not stated on the card.

On October 12 Ai2 followed with HiRO-ACEv2 (Apache 2.0; arXiv 2610.17693). It pairs a sibling emulator, ACE2S-X-SHiELD, which emulates the 3 km X-SHiELD storm-resolving model coarsened to 100 km, with a separately trained diffusion downscaler that regenerates 3 km precipitation, 10 m winds and sea-level pressure for any region between 66°S and 70°N. The authors report matching X-SHiELD's distributions out to the 99.9999th percentile at more than 300 times lower inference cost than the numerical model. The card notes that the emulator saw only nine years of X-SHiELD output and is not expected to generalize beyond those forcings.

Model Details

License Apache 2.0

Paper

scienceopen-weight