Constraining the representation of aerosols in GCMs: explainable AI and Lagrangian approaches to identify key drivers of the natural aerosol lifecycle.

Aerosols remain one of the largest sources of uncertainty in climate modelling, with the representation of the natural baseline significantly hampering constraint on aerosol induced radiative forcing. This thesis begins with an introduction to aerosols and their role in the climate system, as well as discussing current sources of uncertainty. The ability of two global climate models (GCMs) to replicate the aerosol lifecycle of an effusive volcanic eruption is evaluated using a Lagrangian framework. The eruption resulted in an increase in the modal diameter of the accumulation mode across the three surface sites utilised in the study compared to the baseline. The UK Earth System Model was not able to replicate this change without the inclusion of an organic mediated boundary layer nucleation scheme. A major factor limiting the GCMs ability to represent the perturbation of aerosol associated with the eruption was the representation of the background aerosol conditions, with the difference in the baseline dwarfing the impact of the eruption. To improve understanding of the processes governing aerosol properties in environments dominated by natural aerosol sources, an explainable artificial intelligence framework utilising airmass history was developed. The developed framework leverages airmass history to predict aerosol concentrations, as the properties of the aerosol size distribution act as a fingerprint for the processes that have acted on the aerosol population during transport. The models built utilising the framework were able to accurately replicate the seasonal cycle and potential source regions for the Antarctic case study. From model interrogation, it was found that the Antarctic aerosol properties are dominated by seasonal processes, with changes in surface solar radiation and the melt of sea ice being key factors. The contribution from the free troposphere was found to be particularly important for the Aitken mode, whereas during the summer boundary layer transport contributed more to the accumulation mode. The framework was then successfully applied to three additional sites across very different environments: boreal forest, marine continental and Arctic. The models were able to accurately replicate the seasonal cycle for each site and the spatial distribution of potential source regions. Model interrogation revealed distinct dominating processes in each environment, however significant contributions from anthropogenic emissions was found for all sites. Finally, the framework was successfully applied to two GCMs for a case study in the Antarctic, to identify differences in the representation of the dominant processes leading to low GCM biases in this region. This holistic approach to investigate sources of GCM bias can then inform targeted improvements to parameterisations. The framework developed in this thesis allows for the first time comprehensive analysis of aerosol processes and sources of structural uncertainty in GCMs using the air-mass history. This will enable constraint on structural errors in current GCM representation of natural aerosols, therefore improving our ability to predict the impact of anthropogenic aerosol forcing, and elucidating one of the most pressing issues in climate modelling.<p></p>

Author dc:creator

• Ellie Duncan (21049856)

• aerosols
• model evaluation
• Lagrangian
• machine learning

Statement dc:rights

• All rights reserved
• Open Access after 2026-11-18

Identifier

10779/exe.30643628.v1

OAI identifier oai:identifier

oai:figshare.com:article/30643628