High-Fidelity Velocimetry of Cloud Cavitation using Deep Learning Optical Flow

Cloud cavitation is characterized by the violent growth and collapse of vaporous structures, presenting significant challenges for accurate velocimetry. High-speed imaging is an essential tool, and this work focuses on optimizing velocity field estimation from such data. We analyze high-speed videos of cloud cavitation captured at 130,000 frames per second, comparing the performance of three distinct techniques: traditional Particle Image Velocimetry (PIV) using PIVLab, the dense Farneback optical flow algorithm, and RAFT (Recurrent All-Pairs Field Transforms), a modern deep learning optical flow architecture.

Our comparative analysis reveals that RAFT provides markedly superior velocity estimations. It successfully captures fine-scale flow structures and transient events with significantly higher fidelity than both Farneback and PIV, which tend to spatially average these features. We show that the high temporal and spatial resolution of our dataset is critical for RAFT's superior performance. By leveraging these high-quality velocity fields, we demonstrate an enhanced ability to study key cavitation physics, including the interaction of bubbles with shock waves and the evolution of turbulence within the collapsing cloud. Our findings highlight the potential of deep learning methods to transform quantitative analysis in high-speed fluid dynamics.