Statistical Emulator For Fast Uncertainty Quantification

Otto Lamminpaa,  Jet Propulsion Laboratory,  otto.m.lamminpaa@jpl.nasa.gov (Presenter)
Jonathan Hobbs,  Jet Propulsion Laboratory,  jonathan.m.hobbs@jpl.nasa.gov
Amy Braverman,  Jet Propulsion Laboratory,  amy.braverman@jpl.nasa.gov
Pulong Ma,  Duke University,  pulong.ma@duke.edu
Anirban Mondal,  Case Western Reserve University,  anirban.mondal@case.edu

In recent years, satellite-based observations of atmospheric carbon dioxide (CO2) concentrations have emerged as a means of providing data with global coverage and high spatial resolution. Retrieval of CO2 concentrations from measured radiances usually requires iterative solvers and thus repeated evaluations of a computationally expensive atmospheric radiative transfer physics model. This makes it prohibitively expensive to perform rigorous Uncertainty Quantification (UQ) for the retrieval, which is crucial for further science applications of the CO2 data. In addition to Markov Chain Monte Carlo (MCMC) methods for full posterior uncertainty analysis, these UQ efforts include large scale Observing System Uncertainty Experiments (OSUEs). To remedy this computational problem, we propose and implement a Gaussian Process based statistical emulator for the forward model used in NASA’s Orbiting Carbon Observatory 2 (OCO-2) satellite’s CO2 retrieval algorithm. Our approach leverages Functional Principal Component Analysis (FPCA) to find a low dimensional basis for the radiance data. The corresponding FPCA scores are then learned from inputs consisting of synthetic atmospheric state vectors after employing a gradient-based Active Subspace dimension reduction scheme.

Poster: Poster_Lamminpaa__109_25.pdf 

Presentation Type: Poster

Session: 2.5c Uncertainty quantification and bias correction techniques

Session Date: Tuesday (6/15) 12:00 PM