Climate Instability Modeling UQ

Uncertainty Quantification in Climate Models

Model parameter estimation using measurement data, as well as forward propagation of input parameter uncertainty through climate models require model simulations at sampled parameter settings. However, climate models are generally complex and computationally┬ácostly, and therefore only a limited number of model simulations are available. This is a major obstacle for uncertainty quantification methodologies particularly in presence of (a) discontinuous or highly nonlinear model response, or (b) input parameters or output observables that exhibit “fat tails” (i.e., include nonnegligible probability of having extreme values) typically associated with high-consequence climate disruption events.

Our approach is based on a domain decomposition algorithm that first detects the location in the input parameter space where the output changes sharply, and then uses spectral expansions to represent and propagate uncertainties on both sides of the discontinuity. We also rely on modified bases for spectral expansions that sample the input space in specific locations leading to more accurate representations in the tail regions. Our techniques can also lead to optimal experimental design strategies to identify experimental or computational settings that minimize the uncertainty in the resulting predictions. Finally, we investigate model validation and comparison methods with quantified uncertainty assessment for models with varying degree of complexity.