Sensor-based Temporal Super-Resolution for Turbulent Separated Flows

The high Reynolds number turbulent separated flow over a Gaussian speed-bump benchmark geometry has presented challenges for predicting flow separation. Moreover, the lack of time-resolved (TR) experimental data on the Bump has limited progress on understanding the link between the unsteady dynamics and energy transfer mechanisms, which would advance turbulence models. The above challenges motivate the present work: to provide TR estimates of the velocity field from undersampled particle image velocimetry (PIV) data. We propose a data-driven estimation technique that uses oversampled surface-mounted pressure sensors and long short-term memory (LSTM) neural networks to predict transient dynamics that are inherently aliased from the undersampled PIV time-series. The method leverages the TR pressure dynamics to estimate a low-dimensional representation of the velocity field. Spectral analysis of the flowfields after up-sampling the 15 Hz PIV data to 3000 Hz reveals (i) a low-frequency breathing mode that is associated with the contraction and expansion of the separation bubble, and (ii) a medium-frequency mode describing the flapping of the shear layer. The modal dynamics are then used to explore physical mechanisms that govern the turbulence transport within the flow.