EC-EarthFlow: a generative model that emulates a European climate model on a day-to-day basis, at a fraction of the computational cost

Trained on EC-Earth3 simulations covering the period 1950–2166, the emulator reproduces daily variability, seasonal cycles and the warming trend in temperatures.

Running a physical climate model over several centuries is expensive, which limits the number of explorable scenarios. Researchers present EC-EarthFlow, a generative model based on flow matching that mimics the outputs of EC-Earth3, one of the climate models developed by a European consortium.

Forecast tomorrow from yesterday

The emulator was trained on transient simulations of EC-Earth3 ranging from 1950 to 2166 in the intermediate scenario SSP2-4.5. It predicts the next day's temperature field from that of the previous days and the annual average temperature, and then chains its own predictions over periods ranging from one month to an extended season.

With this variable alone, the authors find the daily variability, the geographical distribution, the annual cycle and the long-term trend of the physical model, for a much lower computational cost. The model remains stable over long simulations and seems to capture the physical relationships as represented by EC-Earth3.