Research Interest

At Michigan, my work has focused primarily on the question of feasibility for machine learning (ML) models to emulate the physical parameterizations in global climate models (GCMs). Physical parameterizations are responsible for the sub-grid scale processes occuring in the atmosphere but cannot be explicitly resolved with the fluid flow calculations. Thus, they are parameterized and are responsible for significant bias and uncertainty in GCMs. Working with the Community Atmosphere Model (CAM), the atmospheric component of the Community Earth System Model developed by NCAR, we have been able to train and test ML emulators for various simplified model configurations. We recently showed how offline random forest emulators can be highly skillful for various idealized CAM configurations, but with just minor increases in complexity we do observe non-negligible decreases in skill. An example is depicted in Figure 2 below, showing the effectiveness of our emulators on our parameterized moisture tendencies. The closer the performance measure `R2’ is to one, the better our random forest can be thought of at making the prediction. We see that while both show exceptional skill, there is a noticeable drop off in the slightly more complex case (e). In particular, the region of loss is exactly the region most associated with the additional complexity: a coupled convection scheme for cloud processes which significantly impacts tropical rainfall processes. Further details on the promises and potential limitations of these techniques, as well as comparisons to baseline neural network implementations, can be found in my recently published manuscript [1]. 

References

1. Limon, G. C. and C. Jablonowski (2023). Probing the Skill of Random Forests Emulators for Physical Parameterizations via a Hierarchy of Simple CAM6 Configurations.” Journal of Advances in Modeling Earth Systems, May 25, 2023. https://doi.org/10.1029/2022MS003395.