Uncertainty Quantification for the OCO-2 Mission: A Monte Carlo Framework Using a Surrogate Model

Abstract:

The OCO-2 retrieval algorithm is based on the optimal estimation (OE) framework of Rodgers (2000): it
applies Bayes' Theorem to the problem of inferring geophysical states from soundings observed by the
instrument. The algorithm ingests a sounding and outputs the putative posterior mean vector and
covariance matrix that, under the assumption of Gaussianity, give a probabilistic description of the
underlying state vector. The algorithm also uses a host of other inputs that are treated as fixed and
known when at least some of them are not. Thus, some additional uncertainty beyond that captured by the
posterior covariance is imparted to the algorithm output. The OCO-2 Team is implementing a framework to
quantify this additional uncertainty by simulating a representative set of synthetic state vectors,
generating corresponding synthetic soundings, performing retrievals on them, and comparing the retrievals
to the original synthetic truth. A simplified forward model is used both to generate the soundings and in
the retrieval in order to generate a large ensemble of results in a short period of time. The empirical
joint distribution of the synthetic true and retrieved state vectors can be interrogated to provide
estimates of the impact of uncertain inputs on the retrieval. In this talk, we describe this framework,
its rationale, and how it will ultimately provide users with adjustments that more fully account for uncertainty.