Correcting a 200 km Resolution Climate Model in Multiple Climates by Machine Learning From 25 km Resolution Simulations

Abstract

Bretherton et al. (2022, https://doi.org/10.1029/2021MS002794) demonstrated a successful approach for using machine learning (ML) to help a coarse-resolution global atmosphere model with real geography (a ∼200 km version of NOAA's FV3GFS) evolve more like a fine-resolution model, at the scales resolved by both. This study extends that work for application in multiple climates and multi-year ML-corrected simulations. Here four fine-resolution (∼25 km) 2 year reference simulations are run using FV3GFS with climatological sea surface temperatures perturbed uniformly by −4, 0, +4, and +8 K. A data set of state-dependent corrective tendencies is then derived through nudging the ∼200 km model to the coarsened state of the fine-resolution simulations in each climate. Along with the surface radiative fluxes, the corrective tendencies of temperature and specific humidity are machine-learned as functions of the column state. ML predictions for the fluxes and corrective tendencies are applied in 5.25 years ∼200 km resolution simulations in each climate, and improve the spatial pattern errors of land precipitation by 8%–28% and land surface temperature by 19%–25% across the four climates. The ML has a neutral impact on the pattern error of oceanic precipitation. Plain Language Summary Previous work demonstrated how to use machine learning to help a computationally efficient coarse-grid climate model behave like a more realistic, but expensive, fine-grid reference simulation that we could only afford to run for 40 days. The machine learning was interpreted as correcting errors in the representation of uncertain small-scale cloud, precipitation, and turbulence processes on the model simulations. By using a fine-grid model with a grid spacing eight times as large as our previous reference that runs tens of times faster, we extend that approach to multi-year coarse-grid simulations of a range of climates, both warmer and colder than the present day. Different random starting guesses (“seeds”) lead to slightly different machine learning corrections even with exactly the same training protocol. When applied interactively in 1 year coarse-grid simulations, the machine learning corrections consistently improve the time-mean pattern of rainfall and surface temperature over land versus fine-grid reference simulations in each of the climates we trained against. These machine learning models can be used successfully to enhance the accuracy of 5 year simulations in all climates.