Aurora AI Foundation Model Matches UKCA Chemistry Outputs at 100x Speed with 23/27 Fields at R²>0.85

Researchers fine-tuned the Aurora Earth-system foundation model on paired UKCA chemistry-climate simulations that include interactive emissions, composition and radiation. The resulting emulator captures full daily meteorology, aerosols and radiative fluxes without the expensive chemical solver, achieving R² above 0.85 on 23 of 27 predicted fields in both troposphere and stratosphere. It also learns causal responses to emission perturbations, reproducing ozone changes from altered NO and CO and surface radiation changes from aerosols.

Incorporating MLS satellite ozone retrievals during training reduced zonal-mean ozone error by 23% between 250 and 20 hPa relative to the original UKCA fields. This hybrid approach therefore corrects process-model biases at modest additional cost while preserving physical consistency. The work builds on prior Aurora weather results and UKCA evaluations but is the first to close the loop between emissions, composition and radiation in a single foundation-model architecture.

Policy-relevant air-quality and climate applications stand to benefit because the model runs roughly 100 times faster than UKCA, enabling large ensembles of future emission scenarios that remain computationally prohibitive with conventional chemistry-climate models. Limitations include reliance on UKCA as training truth and the absence of full stratospheric heterogeneous chemistry. Strengthening evidence would require out-of-sample validation against independent satellite and aircraft campaigns plus explicit comparison to other emulators such as those derived from GEOS-Chem or CAM-chem.