Studies employing earth system models are challenged by the large number of parameters that control these models, compounded by the high computational cost required for each model evaluation. Previous sensitivity analysis studies of the Energy Exascale Earth System Model (E3SM) land component highlighted subsets of parameters that impact the variability for specific model outputs. These studies were nevertheless challenged by the non-linearity of the input-output maps which limited the accuracy of surrogate models constructed for subsequent uncertainty quantification and propagation.

In this presentation we propose an alternate approach for constructing sparse surrogate models to simulate the input-output dependencies in the E3SM land model. We employ low-rank functional tensor train decomposition to uncover interactions between model components and the parameters that control them. We explore a set of functional representations over the combined stochastic space including both the model parameters and physical space by considering models spanning several adjacent land cells. We compare the efficiency of the approach with low-rank methods spanning the stochastic space only and with sparse regression approximations in a Bayesian setting.