Disentangling the Effects of Sea Surface Temperature and CO 2 in Global Machine Learned Weather‐Climate Emulators

While previous versions of the Ai2 Climate Emulator (ACE) have been trained with CO 2 as a forcing, they are only accurate within a narrow range of scenarios, for example, climate over the last 80 years forced by observed sea surface temperature (SST), sea ice, and CO 2 (AMIP), or equilibrium or near‐equilibrium climates with CO 2 concentrations ranging from 1× to 4× that of the present day. Attempting to simulate climate forced by AMIP SST perturbed by +4 K, or the response to an abrupt quadrupling of CO 2 , results in unphysical behavior. We attribute this to these models being trained on data sets where the SST and CO 2 are correlated, limiting their ability to accurately learn their separate effects. In this study we introduce a new class of “random‐CO 2 ” reference simulations where the SST and CO 2 are prescribed to vary independently. We train ACE on a balance of AMIP, equilibrium‐climate, and random‐CO 2 data, with a total energy conservation constraint for improved interpretability. It is data‐efficient and not only accurately emulates its reference model in scenarios in which previous models excelled, but also scenarios like AMIP +4 K and slab‐ocean‐coupled abrupt 4 × CO 2 where they did not. However, it is not yet a full‐fledged AI climate model; it has simplified or prescribed representations of other Earth system components like the ocean, land, and sea ice, and it does not expose other known climate drivers as forcings. As configured here, it also relies solely on physics‐based model output for training data, inheriting that model's observational biases.