Forward Model Emulation and Computationally Efficient Atmospheric Retrievals for NASA’s Microwave Limb Sounder

• Activity Number: 431 - Contributed Poster Presentations: Section on Statistics and the Environment
• Type: Contributed
• Date/Time: Wednesday, August 10, 2022 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Maggie Johnson* and Joaquim Teixeira and Nathaniel Livesey and William Read and Michael Schwartz and Paul Wagner and Amy Braverman
• Companies: Jet Propulsion Laboratory, California Institute of Technology and Jet Propulsion Laboratory, California Institute of Technology and Jet Propulsion Laboratory, California Institute of Technology and Jet Propulsion Laboratory, California Institute of Technology and Jet Propulsion Laboratory, California Institute of Technology and Jet Propulsion Laboratory, California Institute of Technology and Jet Propulsion Laboratory, California Institute of Technology
• Keywords: Gaussian process emulation; atmospheric retrievals; MAP estimation

Abstract

NASA's Microwave Limb Sounder (MLS) has been collecting data on the chemistry and dynamics of the upper troposphere, stratosphere, and mesosphere since its launch aboard EOS-Aura in July 2004. MLS measures radiance in the forward limb, and a retrieval algorithm relying on a computationally expensive forward model is used to infer vertical profiles of atmospheric constituents, such as water vapor, temperature, etc., from radiance measurements. In this talk, we present methodology to provide computationally efficient MLS retrievals by approximating the MLS forward model with a Gaussian process emulator. The emulator methodology incorporates dimension reduction in the radiance space through multivariate functional principal component analysis and kernel-based sufficient dimension reduction in the state space. A likelihood approximation is used to estimate independent Gaussian processes in the reduced space using a large sample of historical MLS retrievals. Emulator-based retrievals provide orders of magnitude speed up over the physics-based retrievals with potential to enable near-real-time data processing and large-scale uncertainty experiments.