Exploring Compressive Sensing for Atmosphere Model Emulation

Atmospheric models developed for climate research and Earth system modeling are computationally expensive. For high-resolution models (e.g., those with 25 km or smaller grid spacing) and during intensive model development phases, it is impractical to repeatedly conduct hundreds or more simulations for sensitivity analysis and parameter estimation. Our study assesses the utility of compressive sensing for characterizing parametric sensitivities in the atmosphere component of the US Department of Energy's Earth system model ACME. Surrogate models are constructed and their errors evaluated for a wide range of model output fields. Our results suggest that compressive sensing is a promising technique for establishing objective and efficient strategies of sensitivity analysis and parameter estimation. We will present detailed results from our analysis and discuss the plans for future investigations.