Quantifying Uncertainty in E3SM via Functional Tensor Network Approximations

Abstract

The predictive performance of the Energy Exascale Earth System Model (E3SM) is challenged by the modeling choices for a large ensemble of physical processes. This results in a large number of uncertain parameters and computationally expensive numerical simulations which makes both forward and inverse uncertainty quantification studies difficult. To overcome these challenges, we will focus on constructing surrogate models that exploit the model structure via low-rank functional tensor networks approximations. We will approximate the components of the large scale model with functional forms tailored to the model behavior. We will then cast the training of the functional tensor network model in a Bayesian framework and use a Stein variational inference approach to construct a probabilistic model that approximates the discrepancy between the surrogate and the original model predictions. We will focus on the land model component of E3SM and present results pertaining to global sensitivity analysis and model calibration at a regional scale.