Practical Limitations of Machine Learning Approaches for Online Emulation of Simplified Physical Processes in CAM6

Machine learning (ML) approaches have been used to emulate physical parameterizations in climate and weather models in recent years with varying degrees of success. Limitations to these approaches are not well known, particularly with regard to the complexity of the parameterization scheme being emulated. Utilizing a hierarchy of simplified model configurations within NCAR's Community Atmosphere Model (CAM6), we explore the limitations of ML emulators via the offline skill and trained model size, in the context of its relation to online performance. Online performance is probed with respect to both numerical stability and consistency of results when compared to the original numerical methods. Currently tested model configurations begin with a three-dimensional dry setup, primarily for proof-of-concept coupling tasks. We also extend this to a moist version of the dry model, as well as a version of the moist case that is coupled to a simple convection scheme. Each of these cases add additional layers of nonlinearity and complexity to the parameterization schemes. We utilize both random forests and neural networks to emulate the physics tendencies and precipitation rates for each case. In past work, we have shown that as the CAM6 complexity is increased, random forest skill noticeably decreases, regardless of the extensive ML tuning each goes through. In this work, we extend these results to the coupled system. This allows for further exploration and quantification of the limitations associated with ML techniques with regards to varying parameterization complexity in the case of online emulators.