Toward the Application of Nonlinear Ensemble Data Assimilation Methods to Convective-Scale Parameter Estimation

Recent work has demonstrated that convective-scale model parameters, such as those related to cloud microphysical schemes, are nonlinearly related to dynamic/thermodynamic variables in forecasts and observations. This leads to errors when data assimilation (DA) schemes based on linear-Gaussian assumptions are used to estimate the uncertain model parameters. Nonlinear modifications to the standard ensemble Kalman filter (EnKF) have been shown to perform better for systems governed by convective dynamics, and recent algorithms leveraging advances in AI/ML appear to be especially promising.

 

In this talk, we will present results from previous experiments that demonstrate how and why linear EnKF methods fall short for the challenging task of nonlinear parameter estimation. We will discuss the potential improvements that may result from a new class of ensemble DA algorithms leveraging the powerful framework of latent Gaussian models. In particular, two generalizations of the classical EnKF will be described – one which exploits the special mathematical properties of invertible neural networks (ECTF) and another one based on ideas from measure transport in the context of two-step ensemble filtering (TGA-EnKF). The advantages of these new methods will be illustrated through idealized DA experiments, which will then motivate further discussion on their applicability to convective-scale DA problems.