A GPU-Accelerated Polarimetric Radar Time-Series Emulator

Polarimetric radars have observed regions of negative differential reflectivity near tornadoes, but the responsible scattering mechanism has yet to be fully understood and characterized. It has been suggested that the signature could be due to common debris alignment and/or dominant scattering from debris particles that have extremely large radar cross section (RCS). While the conjecture is plausible, physical in-situ validation and quantification are challenging and almost prohibitive due to the associated danger in the vicinity of tornadoes. However, numerical simulation with rigorous dynamics can be devised to emulate realistic trajectories of debris within a tornado. Combined with realistic polarimetric RCS modeling of each debris particle, the ensemble sum of all the contributions from all the particles could realistically reflect what a polarimetric Doppler radar would measure in practice. This serves as our primary motivation to develop a physically-based radar emulator. There are two key novel aspects of this work: The realistic derivation of debris trajectory based on an air-drag model and the representative diversity of RCS contributions from each debris particle. The debris air-drag and angular-moment parameters are developed through wind tunnel measurements while the polarimetric RCS parameters are obtained through realistic modeling and anechoic-chamber measurements, as described in a separate paper. The goal of the radar emulator is to produce in-phase and quadrature-phase time-series data. A Monte Carlo method is utilized to coherently integrate contributions from all particles within the simulation domain; individual contributions are weighted and phase-shifted based on the positions of each particle and radar operational parameters such as antenna and range-weight patterns. Due to the independent nature of each particle, there is an inherent potential to massively parallelize the workload using the Graphic Processing Unit (GPU) through the OpenCL (Open Compute Language) framework, and that was utilized in this effort. Compared to a traditional implementation, the GPU-accelerated implementation has allowed us to achieve an 80-fold increase in computational speed, which roughly translates a 1-year simulation time to approximately 18 days. A detailed description of the radar emulator and example data will be presented.