Dept. of Geosciences Colloquium: Stable and accurate machine-learning parameterization of subgrid processes for climate modelling at a range of resolutions
Dr. Janni Yuval, MIT
Global climate models represent small-scale processes, such as clouds and convection, using subgrid models known as parameterizations. Traditional parameterizations are usually based on simplified physical models and inaccuracies in these parameterizations are a main cause for the large uncertainty in temperature, precipitation and wind projections. One alternative to traditional parameterizations is to use machine learning to learn new parameterizations which are data driven. However, such parameterizations have been prone to issues of instability. I will show how machine learning algorithms (namely, neural networks and random forest) can be used to learn new parameterizations from the output of a three-dimensional high-resolution atmospheric model, while obeying physical constraints such as energy conservation. Implementing these parameterizations in the atmospheric model at coarse resolution leads to stable simulations that replicate the climate of the high-resolution simulation, and capture important statistics such as precipitation extremes. Retraining parameterizations for different coarse-graining factors shows that the parameterizations perform best at smaller horizontal grid spacings, which suggests that machine-learning parameterization can be successful at small grid spacings where traditional parameterizations fail. Reliable and accurate parameterizations are urgently needed for decreasing the uncertainty in climate projections, and the results I present demonstrate the potential for learning such parameterizations from global convection-permitting simulations that are now emerging.
Event Organizers: Dr. Roy Barkan and Dr. Asaf Inbal