Exploring Ways to Reduce Biases in a Hybrid Global Climate Model With Machine‐Learned Moist Physics

Abstract In a previous study (Han et al., 2023, https://doi.org/10.1029/2022ms003508 ), we implemented a deep convolutional residual neural network for moist physics into the 3‐D real‐geography CAM5 and carried out a stable multi‐year integration successfully. However, the simulation has large temperature and moisture biases in high latitude troposphere and dry bias in precipitation over tropical land. In this study, we explore several ways to reduce these biases. First, we train two shallow neural networks to correct the biases, one with relative humidity included in the input and one without. Second, instead of attempting to correct the biases with an additional neural network, we expand the input and output variables of the neural net by including radiation‐related physical quantities as well as land fraction to enhance the land‐atmosphere coupling and to account for the land‐ocean contrast. Both approaches can alleviate the high‐latitude temperature and moisture biases in the atmosphere as well as land surface temperature biases. However, the corrector approach fails to improve tropical land precipitation, and including extra surface radiative fluxes as neural net input and output only leads to limited improvement in tropical land precipitation. Both approaches made the tropical oceanic precipitation biases worse. These results suggest that something else may be responsible for tropical land precipitation biases. Our findings offer valuable insights that future research can build upon. Subsequent efforts should involve more capable deep learning architectures, stronger physical constraints, and possibly continued use of correction strategies to address persistent errors in hybrid GCM simulations.