A machine learning approach to representing heterogeneous land-atmosphere interactions in an Earth system model

The spatial and temporal patterns over the land surface span over a wide range of scales smaller than those resolved by current Earth system models (ESMs). Accurate representation of this complex subgrid-scale heterogeneity in land-atmosphere interactions remains a significant challenge in Earth system modeling, largely due to limited direct observations and incomplete process-level understanding. In this study, we develop a deep neural network (DNN) emulator that effectively captures these interactions within the framework of the Department of Energy’s Energy Exascale Earth System Model (E3SM). The emulator is trained on a one-year simulation using a Regionally Refined Model (RRM) over the continental United States, which provides high-resolution data that captures the subgrid variability at the land-atmosphere interface. We assess the impact of the emulated surface heterogeneity on the climate system. Preliminary results indicate that the DNN emulator significantly impacts the model’s ability to simulate regional climate patterns, clouds, and precipitation, particularly in regions with complex land surface characteristics. The underlying physical mechanism will be discussed. This work demonstrates the successful application of a machine learning approach to represent subgrid variability in an ESM that traditional parameterizations cannot achieve.