Climate change is dramatically affecting our planet, resulting in extreme events such as floods, droughts, wildfires, heatwaves, sea level rise, and many other processes. These events annually incur billions of dollars in damages and lead to hundreds of fatalities worldwide. Earth System Models (ESMs) offer an opportunity to help us understand the main drivers of these events, and improve Earth system predictability across different spatial and temporal scales. Modeling land-atmosphere interactions in ESMs has been shown to improve climate predictions. However, Land Surface Models (LSMs) are subject to potential bias or errors due to various sources of uncertainties. To quantify, characterize and reduce these uncertainties, one approach is to generate model simulations within a Bayesian framework. This is most often performed through Bayesian inference and Data Assimilation (DA). The value of Earth system modeling relies on the degree to which the uncertainties in the ESM and its components are quantified and accounted for. 

This project aims to demonstrate the application of the latest advances in the data assimilation theory to enhance the predictability of the Energy Exascale Earth System Model (E3SM) while addressing different sources of uncertainties (i.e., model parameters, meteorological forcing data, and initial condition), which is very useful for hindcast experiments, as well as enhancing the capability for future prediction. This project will further improve E3SM’s Land Model (ELM) using a state-of-the-art Machine Learning-enhanced DA technique. This development will be conducted through a comprehensive uncertainty analysis, wherein the team will study how the impact of different uncertainty sources on ELM and coupled land-atmosphere simulations influences predictions of mountain snowpack and streamflow in the western United States and drought in the Great Plains.