Stable machine-learning parameterization of subgrid processes for climate modeling at a range of resolutions

Abstract:

The representation of sub-grid processes contributes to uncertainty in climate predictions by earth system models. Specifically, the parameterization of convection and clouds is a major contributor to the uncertainty in changes in temperature, rainfall distribution, and severe storm frequency. An increasing number of studies show that machine learning can be used to build data-driven parameterizations directly from high-resolution model output. However, such parameterizations have been prone to issues of instability and climate drift when implemented in a coarse-resolution model, and parameterizations learned from three-dimensional model output have not yet been successful for simulations of a given climate state. Here we learn a parameterization of subgrid processes from coarse-grained output of a three-dimensional high-resolution atmospheric model in an idealized aquaplanet configuration. We show that implementing the parameterization in a coarse-resolution model leads to stable simulations that replicate the climate of the high-resolution simulation. Both random forests and neural networks can be used to achieve a stable and accurate parameterization that obeys energy conservation. Furthermore, we show that the stability of the simulations is independent of the specific choice of architecture or hyperparameters. This implies that the accurate calculation of the instantaneous subgrid fluxes and tendencies from the atmospheric state plays an important role in achieving stable simulations. Retraining for different coarse-graining factors shows the parameterization performs best at smaller horizontal grid spacings, which suggests that machine-learning parameterization can be successful at grid spacings that are normally considered to be in the "grey zone". Our results yield insights into parameterization performance across length scales, and they also demonstrate the potential for learning parameterizations from global convection-permitting simulations that are now emerging.