Skip to main content
U.S. flag

An official website of the United States government

Designing observation networks using Earth System Models to reduce uncertainty in regional carbon fluxes

Presentation Date
Wednesday, December 12, 2018 at 8:45am
Walter E Washington Convention Center 150B



Predictive uncertainties in climate models remain high, and confidence in the magnitudes of these uncertainties are low for many variables. A key research focus for Earth System Models (ESMs) is to improve model representations of terrestrial carbon cycling to reduce uncertainties. Uncertainty quantification (UQ) techniques are already being applied to single column versions of land-surface models and simple ecosystem models, but new methods are needed to address uncertainties at regional scales. A large number of relevant land and atmosphere observations exist but have not been incorporated in ESMs for the purpose of estimating predictive uncertainties. Here we present a new framework to address two key science questions related to the application of new UQ techniques:

  1. What is the optimal design of a measurement network to minimize climate model prediction uncertainties?
  2. How much do land model parameter uncertainties contribute to predictive uncertainty at regional scales and which processes are most responsible?

We show an example from the Energy Exascale Earth System Model (E3SM) where we estimate the reduction of model prediction uncertainty in carbon fluxes from an optimal placement of new flux sensors in North America. To do this, we first use a surrogate modeling approach to calibrate model parameters using network land model simulations using flux observations from the AmeriFlux network. Next, we apply statistical methods to upscale these network simulations to continental scale. Then, we optimize the placement of new synthetic observations points for calibration to minimize the posterior error in regional flux predictions. The resulting advance from these methods in understanding about the sensitivity of model outputs to specific processes, and how model uncertainties contribute to prediction uncertainty, will be critically important for future model development and for providing actionable information from models to policymakers.