#27 Parametric sensitivity in ACME-V1 atmosphere model revealed by short Perturbed Parameters Ensemble (PPE) simulations

Abstract

The ACME V1 (NE30_L72) atmosphere model has included many new features in the physics parameterizations. Complex nonlinear interactions between those new features create a big challenge for model tuning. Using the one-at-a-time method, we often encounter cases where the tuning of one parameter leads to an offset of the accomplishment from the tuning of another parameter, or the improvement in one target variable leads to degradation of model fidelity in another target variable. The PPE simulations provide an opportunity to evaluate and optimize model fidelity in a comprehensive and systematic manner. While traditional PPE experiments based on multi-year simulations are hugely expensive, the short ensemble strategy allowed us to finish 12-member ensembles of 3-day simulations for 256 sets of parameter values at a rather modest cost (about 1M CPU hours) and within 5 days of wall clock time.

In this set of simulations, 18 parameters in various physical processes were perturbed simultaneously using the Latin Hypercube sampling method. This poster presents first results from the analysis that aimed at identifying the most influential parameters and quantifying the model response to these parameters for a number of important fidelity metrics. Results from this analysis provide a more complete picture of the model behavior than the one-at-a-time experiments. They also provide information on the tuning potential of different parameters, thus can help guide the tuning activities planned for the next steps.