Paula Harder

• Accurate modeling of weather and climate is critical for taking effective action to combat climate change. Predicted and observed quantities such as precipitation, clouds, aerosols, wind speed, and temperature impact decisions in sectors such as agriculture, energy, health, and transportation. While these quantities are often required at a fine geographical and temporal scale to ensure informed decision-making, most climate and weather models are extremely computationally expensive to run, resulting in coarse-resolution predictions. Recent advances in deep learning (DL) make it an attractive tool for speeding up simulations. The two main ways to decrease computational efforts with DL are downscaling, the increase of the resolution directly on the predicted climate variables, and emulation, the replacement of model parts to achieve faster runs initially. This thesis leverages DL models for accelerated climate forecasting while making sure the methods are feasible for physical modeling. Standard DL approaches often violate simple physical constraints such as positivity or conservation properties. We develop novel methodologies to incorporate those constraints into the training and into architectures of DL. First, we look into so-called soft-constraining methods that introduce an additional regularisation term. Then several hard-constraining methods that change the neural networks (NNs) architectures, by adding final constraining layers, are discussed. We consider two application cases to test our constraining methodology and evaluate the potential of DL for speeding up climate modeling. The first test case is downscaling. We not only show how our hard-constraining layers guarantee the constraints to be satisfied, but also increase the overall predictive performance. In the second employment of our constrained DL models, the aerosol microphysics module in the global ICON climate model is replaced by a NN. We both investigate offline performance, as well as implement the NN in Fortran to run it online within ICON, achieving a stable an accurate coupled simulation. We discuss challenges and choices for a successful deployment of DL in climate and weather simulations.