OPTIMIZED MICROPHYSICS RETRIEVALS FROM POLARIMETRIC RADAR DATA USING PHYSICAL CONSTRAINTS AND MACHINE LEARNING

Abstract

Polarimetric weather radar observations have been widely used over the past two decades to infer precipitation structure and hydrometeor microphysics. However, their full potential remains limited by insufficient treatment of measurement and model errors, the lack of independent information in radar measurements, and the difficulty of linking polarimetric radar variables to model microphysical parameters. This dissertation aims to improve the retrieval of precipitation microphysics from polarimetric radar data (PRD), by addressing three key components: (1) development of efficient forward operators that connect model microphysical parameters to polarimetric radar variables, (2) estimation of measurement errors in polarimetric radar variables particularly polarimetric phased array radar (PPAR) measurements, and (3) development of physics-constrained retrieval algorithms that more effectively integrate or deal better with observations, forward operators, and their associated errors.Each of these components is addressed in the following projects. First, multi-frequency polarimetric forward operators are developed to connect polarimetric variables with model state parameters, based on various mixing formulas and T-matrix scattering calculations. Machine learning (ML) parameterizations are introduced to improve computational efficiency and representativeness. Second, measurement errors in PPAR systems are analyzed using a modified error calculations that leverages their high temporal resolution to access PRD quality. The performance of planar and cylindrical PPAR configurations is evaluated, with emphasis on off-broadside observartions. Third, a physics-constrained ML approach (i.g., physically-informed neural networks, PINN) retrieval algorithm is developed to retrieve the microphysical properties of mixed hydrometeors by combining forward operators, radar observations, and improved treatment of uncertanties in both physical constraints and observations. The developed forward operators and retrieval algorithms are evaluated through retrieval experiments using Weather Research and Forecasting (WRF) simulations, which demonstrate their capability for multispecies hydrometeor retrieval. These methods are further applied to operational NEXRAD observations for rain and hail cases. Overall, this work aims to better utilize estimated measurement errors, forward operators, and advanced retrieval algorithms to advance quantitative precipitation retrieval.