Precipitation Microphysics Retrieval from Polarimetric Weather Radar Data using Physically Constrained Machine Learning Techniques

Polarimetric radar data (PRD) provide detailed insights into the microphysical properties of hydrometeors. However, the microphysics retrieval for multiple species of hydrometeors—rain, snow, graupel, hail, and their mixtures—using only PRD presents a significant challenge due to the limited independent information available. Traditional microphysics retrieval methods, including physics-based inversions and empirical formulas, often suffer from significant biases and inaccuracies. Statistical approaches, such as variational methods, become increasingly complex and computationally intensive as the number of features and output parameters increases. Recently, machine learning techniques, like deep neural networks (DNNs) and convolutional neural networks (CNNs) such as U-Net, have been introduced to overcome these limitations. DNNs exploit correlations among radar variables to improve the accuracy of the retrieval for rain microphysics parameters, such as mass- or volume-weighted mean diameter (D_m) and liquid water content (W). However, despite these advances, DNNs alone cannot fully resolve the complexity of the problem and require the incorporation of physical constraints to improve feasibility and accuracy. To further improve the retrieval, advanced machine learning techniques (e.g., U-Net) are adapted to incorporate additional spatial information in this study. Enhanced hydrometeor classification information is also incorporated to reduce problem complexity and improve retrieval accuracy. Initial training and testing of the U-Net model are conducted using simulated PRD from the Weather Research and Forecasting model, combined with newly developed forward polarimetric radar operators and a melting model. The retrieved results are compared between statistical, and machine learning-based methods, both with and without physical constraints. Preliminary results suggest that U-Net, when combined with physical constraints, provides accurate and reasonable retrievals, with the spatial distribution of both D_m and W for all seven hydrometeor mixtures agreeing well with expectations.