The Aurora foundation model adapted for atmospheric chemistry: ozone and aerosols simulated 100 times faster, and stable over 50 years

Retrained on the UKCA chemistry-climate model, Aurora reproduces 23 out of 27 fields with an R² greater than 0.85 and learns to correct its ozone biases using MLS satellite data.

Simulating atmospheric chemistry and aerosols is often the most expensive part of a climate model, to the point of being simplified or imposed from the outside. A team on arXiv presents what they describe as the first foundation model capable of representing the interactions between emissions, atmospheric composition and radiation. They refined Aurora, a foundation model of the Earth system, on simulations of the UKCA chemistry-climate model.

Fast, stable and emission-sensitive

The model reproduces weather, composition, aerosols and radiative fluxes on a daily basis, with 23 of the 27 predicted fields reaching an R² greater than 0.85 in both the troposphere and the stratosphere. The average annual biases remain below 4% for ozone, carbon monoxide and temperature, and below 6% for downward radiation. It rotates stably for 50 years of autoregressive simulation, about 100 times faster than UKCA, and responds correctly to modified emission scenarios (tropospheric ozone according to NO and CO releases, surface radiation according to aerosols).

By integrating ozone measurements from the MLS satellite into the training, the authors reduce the average ozone error between 250 and 20 hPa compared to observations by 23%. They specify that the approach is not intended to replace physical modelling, but to offer a low-cost complement capable of learning from observations.