Machine learning parameterizations for ozone: climate model transferability

Many climate modeling studies have demonstrated the importance of two-way interactions between ozone and atmospheric dynamics. However, atmospheric
chemistry models needed for calculating changes in ozone are computationally expensive. Nowack et al. [1] highlighted the potential of machine learning-based ozone parameterizations in constant climate forcing simulations,
with ozone being predicted as a function of the atmospheric temperature state. Here we investigate the role of additional time-lagged temperature information under preindustrial forcing conditions. In particular, we test if the use of Long Short-Term Memory (LSTM) neural networks can significantly improve the predictive skill of the parameterization. We then introduce a novel workflow
to transfer the regression model to the new UK Earth System Model (UKESM). For this, we show for the first time how machine learning parameterizations could be transferred between climate models, a pivotal step to making any such parameterization widely applicable in climate science. Our results imply that ozone parameterizations could have much-extended scope as they are
not bound to individual climate models but, once trained,could be used in a number of different models. We hope to stimulate similar transferability tests regarding machine learning parameterizations developed for other Earth
system model components such as ocean eddy modeling, convection, clouds, or carbon cycle schemes.