Applying machine learning parameterization through coarse graining to improve the simulation of multiple climates in a full complexity GCM

Previous work, primarily in an aquaplanet setting, has demonstrated that machine learning can be used in climate models to bring the climate statistics of coarse resolution simulations closer to those of finer resolution runs. The presence of topography and land-surface heterogeneity, however, makes this more challenging in full-complexity models, and requires the development of modified approaches. Here we present the results of some initial work on this problem, in the context of improving 200 km resolution simulations of climate change. To crudely simulate climate change, we run 2-year reference simulations with NOAA's FV3GFS model at 25 km resolution with four sets of prescribed SSTs, one with present-day SSTs, and three others with SSTs perturbed by minus 4 K, plus 4 K, or plus 8 K. To produce training data, we run four 2-year 200 km simulations where we nudge the temperature, specific humidity, horizontal winds, and pressure thickness to the coarsened state of each of the 25 km runs. The nudging tendencies produced can be thought of as state-dependent bias corrections, which can bring the state of the coarse model closer to that of the fine model. As such, we train neural networks (NNs) using data from all climates to predict the nudging tendencies in each vertical column based on the columns atmospheric state. We show that when applied in free-running 200 km simulations in all four climates, the NNs help to improve the simulation of climate statistics such as the time mean spatial patterns of precipitation relative to the 25 km reference runs, compared with 200 km runs without ML corrections, and can also improve the simulation of patterns due to differences between climates. If an additional machine learning model is used to predict the surface radiative fluxes, correcting for biases in the land surface energy budget, we can obtain improvements in the simulation of the time mean spatial pattern of land surface temperature, while maintaining improvements in the simulation of precipitation over land.