Estimation of a Time-Resolved Velocity Field Over an Open Cavity via Recurrent Neural Network.

This study presents a data-driven approach to reconstructing time-resolved cavity flow dynamics from sparse wall pressure measurements using snapshot Proper Orthogonal Decomposition (POD) and long short-term memory (LSTM) neural networks. Velocity and surface pressure data were experimentally measured in a Mach 1.4 open cavity flow regime. The velocity fields were decomposed into a truncated POD representation and paired with synchronized surface pressure measurements to train individual LSTM models for each POD mode. The trained networks were then utilized to predict time-resolved POD expansion coefficients which enabled reconstruction of a time-resolved flow field based on the surface pressure measurements as the input. The reconstructed flow fields exhibit key features of cavity flow, including shear layer flapping and alternating regions of velocity fluctuations. Spectral analysis shows strong agreement between the predicted expansion coefficients and pressure measurements, indicating that the LSTM models were able to capture global cavity oscillations. Analysis of the resulting time-resolved flow fields demonstrate the capability of LSTM models to reconstruct high-energy structures with high fidelity and shows the potential of pressure-driven models for estimating the dominant flow dynamics of cavity flow.