Climate and weather models can only afford solve the partial differential equations governing atmospheric flow at a coarse resolution. They cannot affordably simulate small-scale physical processes like clouds, precipitation, and turbulence, which are typically approximated using hand-tuned “parameterizations”. These parameterizations are typically derived by domain-scientist on pencil and paper based on qualitative features of observations and then hand-tuned to produce a reasonable climate. This development process is somewhat disconnected from data, and despite large uncertainty in the predictions they make, parameterizations have changed little in decades. A recent explosion in observations and high-resolution modeling allows us to bootstrap machine-learned alternatives. In this talk, I will introduce attempts to build machine learning parameterizations for use in increasingly complex simulations of the atmosphere. If successful, these parameterizations would improve the accuracy of precipitation and other forecast over a range of timescales.
Challenges in long-term prediction for climate were discussed.