Machine-Learning-Based and Process-Oriented Model Calibration on Turbulence and Clouds

The emerging global storm resolving models are approaching kilometer scale resolution, however, sub-grid scale parameterization is still needed for turbulence and shallow convective clouds. Traditional tuning methods to deal with model’s parametric uncertainties are labor intensive and hardly feasible for such computational-expensive new generation of global models. In addition, such practice often involves with regional biases in process analysis due to compensating errors during the pursue of an optimal global performance on average. To better understand uncertainties related to parameterized processes such as turbulence moments and transports, we establish a case library of convective regimes based on long-term ground-based observations from Atmospheric Radiation Measurements (ARM) fixed sites and mobile facilities supported by US Department of Energy (DOE). We then use this case library in model validation and improvement including discoveries of structural errors and fine-tunes of parametric uncertainties. In this study, we use machine learning tools to create surrogate models to emulate simulations of convective boundary layer and surface-forced continental shallow cumulus clouds by doubly periodic Simple Cloud Resolving E3SM Atmospheric Model (DP-SCREAM). DP-SCREAM is a stand-alone mode with the same physics packages as in DOE’s global storm resolving model, SCREAM. By doing so, a full spectrum of sensitivities to uncertainties in model parameter values could be learned, as well as the sensitivities to the choices of cost functions in constraining posterior joint distribution of parameters for optimal performance against observations (or large eddy simulations of the same cases). In this study, we specifically focus on fine-tune model performance in simulating boundary layer turbulence and clouds and aim to dissect model uncertainties for each convective regime and complement the view from bottom to up, e.g., from processes upscale to larger scales. We expect such process-oriented tuning may help better understand biases related to individual convective regimes, thus, to reduce compensating errors and lead to optimal performance at regional and global scales.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. IM Release number: LLNL-ABS-868419