Building smarter climate models with machine learning

Climate models can not affordably simulate small-scale physical processes so they must approximate the effect of important phenomena like clouds, precipitation, and turbulence with parameterizations. Unfortunately, uncertainty in these sub-grid-scale parameterizations limits our faith in our coarse-resolution climate models and contributes to well-known climate model biases. Traditional parameterizations designed by human experts have changed little in decades despite a recent explosion in our ability to observe the Earth-system and run explicitly-resolved simulations. The brute-force solution to this problem is to decrease the grid-spacing of atmospheric models until they can resolve individual thunderstorms, but this is not yet feasible for climate-scale simulations. Machine learning (ML) presents another path forward by replacing traditional schemes with nonlinear regression models trained from shorter high-resolution simulations or observations. In this talk, I will introduce a recent body of work building machine learning parameterizations for use in increasingly complex simulations of the atmosphere. I will particularly focus on how we ensure that ML parameterizations play nicely with fluid-mechanics simulations and on interpreting their behavior. While these efforts are targeted on improving atmospheric models, similar techniques could be applied to sub-grid-scale problems throughout the Earth sciences.