Using Machine Learning to Make Computationally Inexpensive Projections of 21st Century Stratospheric Column Ozone Changes in the Tropics

Keeble, James and Yiu, Yu Yeung Scott and Archibald, Alexander T. and O’Connor, Fiona and Sellar, Alistair and Walton, Jeremy and Pyle, John A. (2021) Using Machine Learning to Make Computationally Inexpensive Projections of 21st Century Stratospheric Column Ozone Changes in the Tropics. Frontiers in Earth Science, 8: 592667. ISSN 2296-6463

Abstract

Stratospheric ozone projections in the tropics, modeled using the UKESM1 Earth system model, are explored under different Shared Socioeconomic Pathways (SSPs). Consistent with other studies, it is found that tropical stratospheric column ozone does not return to 1980s values by the end of the 21st century under any SSP scenario as increased ozone mixing ratios in the tropical upper stratosphere are offset by continued ozone decreases in the tropical lower stratosphere. Stratospheric column ozone is projected to be largest under SSP scenarios with the smallest change in radiative forcing, and smallest for SSP scenarios with larger radiative forcing, consistent with a faster Brewer-Dobson circulation at high greenhouse gas loadings. This study explores the use of machine learning (ML) techniques to make accurate, computationally inexpensive projections of tropical stratospheric column ozone. Four ML techniques are investigated: Ridge regression, Lasso regression, Random Forests and Extra Trees. All four techniques investigated here are able to make projections of future tropical stratospheric column ozone which agree well with those made by the UKESM1 Earth system model, often falling within the ensemble spread of UKESM1 simulations for a broad range of SSPs. However, all techniques struggle to make accurate projects for the final decades of the SSP5-8.5 scenario. Accurate projections can only be achieved when the ML methods are trained on sufficient data, including both historical and future simulations. When trained only on historical data, the projections made using models based on ML techniques fail to accurately predict tropical stratospheric ozone changes. Results presented here indicate that, when sufficiently trained, ML models have the potential to make accurate, computationally inexpensive projections of tropical stratospheric column ozone. Further development of these models may reduce the computational burden placed on fully coupled chemistry-climate and Earth system models and enable the exploration of tropical stratospheric column ozone recovery under a much broader range of future emissions scenarios.