Abstract/Summary

Weather prediction hinges on mathematical models implemented into software to predict the future state of the atmosphere. Despite remarkable progress, computational constraints and user demands are choking this progress. In this PhD thesis, new machine learning (ML) methods are presented to improve the parametrization of two common schemes used within numerical weather prediction (NWP): radiation and urban land surface. First, a fast and accurate ML emulator for simulating three-dimensional cloud effects as a correction term to a fast parametrization scheme is developed rather than replacing the entire radiation scheme. Second, as ML emulators' training data can be scarce or expensive, a cheap method based on statistical copulas is implemented to generate data like the original across variables and dimensions. Third, the urban land surface model Town Energy Balance (TEB) is coupled to the Weather Research and Forecasting (WRF) model through a modular implementation, verified by an integration test, to evaluate a newly devised urban neural network (UNN) emulator. By training the UNN on the mean output from several urban land surface schemes, the UNN is more accurate, cheaper to run, and simpler to set up than TEB. Furthermore, when coupled to WRF, the UNN is numerically stable with lower errors than the reference WRF-TEB implementation.