Assessing the Offline and Online Performance of Physical Parameterization Emulators via Simplified Climate Model Configurations

In recent years, machine learning (ML) methodologies have been applied to emulate physical parameterizations in climate and weather models, presenting variable levels of success. However, the constraints of these approaches, particularly concerning the complexity of the parameterization schemes, remain inadequately understood. This study investigates the limitations and potential tradeoffs of ML emulators using simplified model configurations within NCAR's Community Atmosphere Model (CAM6). We evaluate the relationship between offline skill and trained model size to better understand the impact on the online performance of the coupled model. Our analysis focuses primarily on numerical stability and consistency of the resulting simulations against original numerical methods. The initial model configurations tested are a moist Held-Suarez model along with a modified configuration that incorporated a simple convection scheme, which introduces an additional layer of nonlinearity and complexity. We employ random forests and neural networks to emulate physics tendencies across the two cases. Previous work has indicated that increasing CAM6 complexity in this way leads to quantifiable decreases in offline random forest performance, despite extensive ML tuning. This research extends these findings to coupled systems, providing deeper insights into the potential trade offs and limitations of ML emulation of physical parameterizations.