Efficient Adaptive Beamforming for Weather Observations with the NSSL Advanced Technology Demonstrator using Deep Neural Networks

Phased-array radars are under consideration as future replacements for the Weather Surveillance 1988 – Doppler (WSR-88D) radars, which are expected to remain operational until the 2035-2040 timeframe through ongoing maintenance and engineering support. The Advanced Technology Demonstrator (ATD) at the National Severe Storms Laboratory has been developed and deployed to demonstrate the capability of phased-array radars to collect calibrated, dual-polarization weather data with rapid update rates. Featuring an overlapped-subarray architecture, the ATD is capable of adaptive digital beamforming on receive, allowing for antenna pattern adjustments that enhance the accuracy of radar variable estimates, particularly in areas with significant reflectivity gradients. A "Constrained Capon" method has recently been developed, introducing additional constraints on the mainlobe shape to the traditional Capon technique, resulting in effective minimization of sidelobes and proper calibration. However, the complexity of this new technique requires significant computation time, necessitating increased efficiency for eventual real-time operations. This study focuses on developing a deep neural network (DNN) approach to drastically increase the speed of utilizing the Constrained Capon method. We present approaches for case selection and data curation steps, including how data are processed with a subarray in-phase/quadrature (IQ) data simulator, methods for gradient detection and range gate selection, and techniques for addressing data imbalance challenges. Results using a medium-sized training dataset are shown, and comparisons between Fourier, Capon, Constrained Capon, and the DNN model for Constrained Capon are presented. Velocity, spectrum width, and the polarimetric variable estimates are provided to demonstrate viability beyond just reflectivity factor. The path forward for development and testing is discussed, including scaling to much larger datasets, appropriate validation metrics, and eventual application to ATD data collected with spoiled beams.