With SWIFT-AI, we explore machine learning (ML) methods and apply them to stratospheric ozone chemistry. The model will firstly replace the detailed chemistry schemes of Chemistry and Transport Models (CTMs) and secondly be coupled to global climate models - General Circulation Models (GCMs). Modeling the Earth system is a complex task and models typically contain a large number of submodules and parameterizations. This is true, for example, of the atmosphere, hydrosphere, lithosphere, and cryosphere. Atmospheric chemistry is complex and typically involves dozens of chemical species and hundreds of reactions with a wide range of concentrations and chemical lifetimes. SWIFT-AI uses artificial intelligence methods to predict stratospheric ozone trends. In addition to prediction, model uncertainty will be exploited and used during a simulation run. The training data comes from stratospheric chemistry simulations of the Lagrangian CTM ATLAS and consists of the ozone trends and 55 parameters stored at each model point and time step. These data serve as the basis for supervised learning of the highly nonlinear relationship. An earlier version of the SWIFT model used a polynomial approximation approach. In SWIFT-AI, we use neural network capabilities to improve the approximation and develop an uncertainty estimation algorithm to increase the robustness of the model. The overall advantage of this replacement model is essentially the reduced computation time (minutes instead of days) compared to the full chemistry model ATLAS, while still achieving high accuracy.
More information at: https://www.mardata.de/