Extending the CAMS Carbon Monoxide Reanalysis after the Loss of MOPITT: An ML‑Based Approach

Reanalyses of atmospheric composition (AC) combine atmospheric models with satellite retrievals to produce consistent, long‑term, gridded datasets that are widely used to assess trends and variability in atmospheric composition and air quality. However, changes in the availability of the assimilated observations can introduce artificial discontinuities, complicating the interpretation of long‑term trends. Bias correction and careful selection of the assimilated datasets are therefore essential to ensure temporal consistency.

The Copernicus Atmosphere Monitoring Service (CAMS) global AC reanalysis (EAC4) assimilates multiple retrievals of aerosol optical depth, ozone, and nitrogen dioxide, as well as total column carbon monoxide (TCCO) from the MOPITT instrument—the sole CO data source in the system. EAC4 spans 2003 to near‑present and has been extensively used for reporting AC anomalies and trends. The termination of MOPITT operations in January 2025 resulted in a substantial shift in CO fields in the subsequent EAC5 reanalysis, preventing its direct use for diagnosing TCCO anomalies in 2025.

To address this discontinuity, we developed a machine‑learning‑based method to emulate the MOPITT‑driven assimilation impact in EAC4. The ML model predicts monthly mean TCCO fields by learning the relationship between EAC4 and a corresponding control simulation without data assimilation. The control simulation uses the same meteorological fields and the same emissions as EAC4, including the CO wildfire emissions that dominate interannual variability.

We evaluate the agreement between EAC4 TCCO trends and annual anomalies and their ML‑based predictions. We also discuss alternative approaches for deriving TCCO anomalies for 2025, such as the use of the control simulation alone or the TCCO analysis from the operational CAMS forecasting system.

This work represents an initial step toward emulating AC data assimilation using machine learning, with the broader aim of improving the robustness of long‑term AC datasets in the presence of observational gaps.