Understanding Sensitivity of Climate with Perturbed Parameter Ensembles

Efficient and accurate model parameter estimation is a challenge in climate modeling. There has been renewed interest in parameter estimation methods, in part due to a confluence of recent scientific advancements. Large observational datasets are becoming more readily, increased computational capacity has enabled the generation of ensembles of model simulations with strategic perturbations to model parameters and machine learning algorithms are emerging as valuable tools for high-dimensional data-driven tasks such as parameter estimation. Given these advancements, many groups are exploring and developing methods for more accurately and efficiently estimating model parameters and using this for scientific exploration. For example, to evaluate parameterizations in models, simulations have often been performed with only one change to one parameterization at the time. This is highly inefficient and can be both time-consuming and computationally costly. I will present results from a Perturbed Parameter Ensemble (PPE) using CESM-Community Atmosphere Model (CAM) 6, utilizing the Latin hypercube sampling technique to reduce the needed number of ensembles, while making sure the entire parameter spaces are used. In our case 45 parameters in the microphysics, convective, turbulence and aerosol schemes were perturbed over 262 simulations. Machine learning can then be used to fill out the density of parameter combinations.