Surrogate-assisted Bayesian Uncertainty Quantification of a Great Lakes Coupled Regional Climate Model
Abstract
As part of the U.S. Department of Energy's Office of Science's efforts to further our predictive understanding of coastal systems, a coupled atmosphere-land-lake modeling system has been developed for regional climate assessments of the Great Lakes region under the COMPASS (Coastal Observations, Mechanisms, and Predictions Across Systems and Scales) project. A critical component of this project is to better understand the relative contributions of internal variability, model parameters, and physics parametrizations to uncertainties in the coupled model. In this study we investigate coupled model parameterizations of the lake mixing, atmospheric convection and radiation schemes, and the atmosphere-lake surface flux. A neural network (NN) based surrogate approximation is trained using model results to predict time series of quantities including lake surface temperature and surface air temperature, humidity, and wind velocities. The surrogate model is employed first for a sampling-based global sensitivity analysis to determine relative uncertainty contributions of each parameter toward output quantities of interest. Further, the NN surrogate is employed in a Bayesian inference framework, sampled via Markov chain Monte Carlo, to calibrate model parameters given observed in-situ data. Calibrated model parameters, when propagated through the model, exhibit significantly better match to the observational data compared to the unconstrained, nominal parameter settings.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.A46F..08P