Improving climate models using machine learning
This session of AI and Climate Science focuses on improving climate models using machine learning. The series is hosted by prof. Philip Stier, this session features Chris Bretherton and Noah Brenowitz.
We use machine learning (ML) to correct the physical parameterizations of a real-geography coarse-grid global atmosphere model to make the model evolve more like a reference dataset – either a reanalysis or a fine-grid global storm resolving model simulation. In either case, we run a training simulation in which the temperature, humidity, and optionally the horizontal winds of a coarse model are nudged to the reference on a 3-6 hour time scale. We learn the nudging tendencies as functions of column state and use these in forecasts to correct the physical parameterization tendencies. We demonstrate our approach using the US operational weather forecast model FV3GFS with a coarse 200~km grid. For both reference dataset types, the correction improves the weather forecast skill and the time-mean precipitation patterns by 20% or more.
Climate models integrate many components encoding the physical processes like precipitation and the greenhouse effect that are not represented by large-scale fluid flow. While these components, known as parameterizations are simple individually, their interplay with each other and with the fluid flow creates climate and weather simulations of extraordinary complexity. Replacing existing parameterizations with look-alike machine learning emulators is a popular method, usually motivated by the potential to accelerate these existing schemes and deploy them on GPUs. In this talk, however, I will argue that emulation is a tricky machine-learning problem whose solution will teach us a lot about how to combine machine learning with existing physics-based prediction systems.