Uncertainty quantification of RANS turbulence model parameters with ensemble Kalman methods and machine learning

Turbulence parametrizations in atmospheric boundary layer (ABL) models account for the effects of turbulence that occur in the subgrid-scale. These parameterizations, widely used in numerical weather predictions and wind engineering applications, typically use a pre-determined set of model coefficients that are tuned on limited high-fidelity data from canonical flows. To exploit more high-fidelity data and explore the parameter space more broadly, we pose the calibration of these values as a Bayesian inverse problem in which an optimal set of parameters minimizes the error between model output and high-fidelity data. In this study, we utilize the calibrate-emulate-sample framework for the efficient calibration and uncertainty quantification of turbulence model parameters in a single-column model (SCM). In a perfect-model calibration setting for a conventional neutral ABL case, we learn a prescribed set of truth parameters from noisy approximations of the model output and quantify the uncertainty in parameters. Further, we determine the ABL statistics that minimize the turbulence model parameter uncertainty in the SCM. The aim of this study is two-fold: 1) to quantify uncertainty in ABL turbulence parameterizations, and 2) to develop a Bayesian optimal experimental design framework for measurement acquisition in its application to turbulent ABL wind predictions.