Seminar: Gaussian Process Emulation of Computationally Expensive Physics Models

Speaker: Dr. Otto Lamminpaa, NASA JPL

Abstract

Recent advances in atmospheric remote sensing have made global high accuracy measurements of CO2 possible using satellite instruments. Inferring CO2 concentrations from measurements (retrieval) is an inverse problem that’s solved using iterative methods. The gigantic data volume and need for rigorous uncertainty quantification for the final product poses enormous computational challenges due to expensive physics models used. To tackle this problem, we propose a Gaussian Process emulator for the physics model of NASA’s OCO-2 satellite. The implementation leverages novel Kernel Flows method for parameter learning. We finally demonstrate the effectiveness of our emulator by performing CO2 retrievals.

See Dr. Lamminpaa's bio here. 

*This event is for NCAR/UCAR/UCP staff only. For employees, add this event to your calendar.