Presenters: Benjamin Wagman, Andrew Yarger, Lyndsay Shand and Andy Salinger
Title: Autotuning E3SM using a surrogate model for climatological spatial fields
Abstract:
Tuning an Earth system model is a time-consuming effort traditionally led by subject matter experts. We are developing an automated tuning method that uses a perturbed parameter ensemble (PPE) and machine learning (ML)/uncertainty quantification (UQ)/optimization tools to select parameter values. We present the workflow and the results of our optimization for E3SMv2. The automated tuning method reduces the RMSE relative to the E3SMv2 release by 3.9% on average for a set of eleven spatial targets over four seasons. We also study the effects of PPE sample size, PPE climatology length (number of simulated years), and surrogate spatial resolution, to determine the most efficient methods for our next effort, the automated tuning of E3SMv3.