The Energy Exascale Earth System Model (E3SM) is at the cutting edge of high-resolution climate prediction, but its accuracy is hindered in part by significant uncertainties in land surface processes. Addressing these uncertainties is critical for reducing model biases, prioritizing model developments and improving predictive understanding. Detailed investigation into the impacts of model parameter uncertainties on land surface processes and their outputs is needed.  To enable this capability, we introduce an advanced uncertainty quantification (UQ) framework specifically designed for the E3SM land model (ELM), integrating the offline land model testbed (OLMT) and a neural network-based surrogate modeling approach. The surrogate modeling approach enables otherwise computationally expensive global sensitivity analyses (GSA) and model calibration using Markov Chain Monte Carlo (MCMC) methods, which can significantly reduce uncertainties and improve the predictive power and reliability of ELM and E3SM simulations.

UQ requires model ensemble simulations, which are computationally expensive in a model like ELM. A primary goal of our UQ framework is to capture the relationship between model parameters and output quantities of interest (QoIs) with the smallest possible ensemble size. This can be accomplished using neural network surrogate models trained on ELM simulations that explore a high-dimensional parameter space over a range of gridcells across different biomes and environments. GSA may then be applied to identify the most influential parameters that drive variability in QoIs for calibration or further investigation.

Following GSA, a new model ensemble may be generated with a reduced set of model parameters identified as sensitive, which can then be used to train a new set of surrogate models that are accurate enough for calibration. The UQ framework incorporates model calibration using Markov Chain Monte Carlo (MCMC) methods. MCMC is a powerful statistical technique that can generate posterior distributions of model parameters and QoIs given a set of prior distributions and observational constraints. By performing MCMC on the surrogate models, parameter estimation can be performed efficiently even for complex and computationally demanding models like ELM. This approach ensures that the calibrated model parameters are well-constrained by data.

Here we demonstrate the UQ framework across a range of North American eddy covariance flux towers spanning 3 major plant functional types (PFTs). By leveraging modern surrogate modeling techniques and advanced statistical methods, our UQ framework provides a powerful tool for researchers and policymakers to better understand and mitigate the uncertainties inherent in land surface modeling.