Abstract
Quantifying sensitives of model results to uncertain parameters and tuning the parameter values to improve a model’s simulation/prediction skills are important tasks in the development of Earth System Models like E3SM. A systematic exploration of the parameter space typically requires a large number of simulations and hence are computationally expensive or even prohibitive. Manual tuning of parameter values based on experts’ educated guesses is tedious and the results cannot be guaranteed to be reproducible. This project addresses these challenges by applying cutting-edge methods from the applied math field. A signal processing technique called compressive sensing is used to construct emulators (i.e., statistical surrogate models) that describe the relationships between simulated climate features and the values of uncertain parameters in the E3SMv1 atmosphere model. The emulators are then used for sensitivity analysis and automatic parameter tuning. The first phase of the project focuses on large-scale features of the simulated climate. The presentation will demonstrate that the compressive-sensing-based emulation method requires substantially fewer E3SM simulations and provides higher emulation accuracy. Variance analysis based on those surrogate models gives robust results in the identification of impactful model parameters. A multi-objective optimization algorithm manages to identify parameter sets that lead to the desired model features within the user-specified tolerance. These results provide a solid basis for future steps, where we plan to develop emulation methods for spatial and temporal variations in the simulated climate, and provide our algorithms to the model developers in the form of user-friendly software packages.