Data-driven discovery and model reduction methods for the atmospheric effects of high altitude emissions

Abstract

Chemistry transport models play a crucial role in the evaluation of the effect of anthropogenic emissions on the atmosphere and climate, but they come with high computational costs and require specialized know-how. This renders them impractical for applications in multidisciplinary optimisation, or regulatory and operational decision-making processes where environmental effects are to be considered. Such applications require computationally efficient surrogate models of the complex chemistry transport models. Here we investigate the use of data-driven discovery and reduced-order modelling methods for this purpose. Specifically, we examine the dynamic mode decomposition (DMD) and proper orthogonal decomposition coupled with the sparse identification of non-linear dynamics (POD-SINDy). We evaluate their ability to reconstruct and forecast changes in the distribution of ozone in response to the introduction of supersonic aircraft as modelled by the GEOS-Chem chemistry transport model. Of the tested methods, we find that optimized DMD and bagging optimized DMD with constrained eigenvalues perform best. These methods can reconstruct and forecast full-atmospheric ozone responses for up to several years without losing stability, at smaller errors than estimates using the spatio-temporal mean of the data. On average, the constrained optimized DMD method reduces the reconstruction error by 63.5 % and that of forecasting by 25.8 % compared to the spatio-temporal mean. For the constrained bagging optimized DMD these reductions are 45.0 % and 23.1 %, respectively. The resulting change in global ozone column, calculated from the reconstructed atmospheres, has an error smaller than 10 %. This is achieved while reducing the computational and storage requirements by several orders of magnitude, which may be a worthwhile tradeoff for some applications.