A Stochastic Neural-Network Parameterization for Coarse-grid Climate Models

Abstract

Coarse-grid climate models require parameterizations to include the effect of unresolved sub-grid processes. Recently, machine learning approaches have shown promise in producing more accurate parameterizations than existing physical-based approaches. However, these machine learning approaches to parameterization are deterministic so they fail to capture the probability distribution of possible sub-grid processes, which can cause mean state bias in simulations. A stochastic parameterization that is built on top of a machine learning parameterization is introduced. The stochastic approach consists of a Markov chain model that switches between states that represent residual ranges of the machine learning parameterization and a regression model that produces a stochastic adjustments to the machine learning parameterization given a Markov state. The approach produces the correct probability distribution of outputs when predicting on a training a dataset obtained from coarsening a high resolution simulation over 160-km by 160-km grid cells, but fails to correct the mean state bias issues in a climate model simulation. However, the climate model used for the simulation has systematic issues that make it difficult to effectively evaluate parameterizations: a hyper-diffusion scheme wipes out stochastic effects, and it is optimized for high resolution, not coarse resolution, simulations. Further evaluation is necessary to determine whether the stochastic method improves a deterministic machine learning scheme in a full climate model simulation.