Regime identification for stratified wakes from limited measurements: a library-based sparse regression formulation

TL;DR

This paper presents a library-based sparse regression method to identify flow regimes in stratified wakes from limited measurements, using DMD modes and a calibration algorithm to estimate Reynolds and Froude numbers.

Contribution

It introduces a novel sparse regression framework leveraging DMD modes for regime identification from limited data in stratified wake flows.

Findings

Successfully identified flow regimes from limited velocity data.

Demonstrated effectiveness on both numerical and experimental datasets.

Provided a confidence metric for regime classification accuracy.

Abstract

Bluff body wakes in stratified fluids are known to exhibit a rich range of dynamic behavior that can be categorized into different regimes based on Reynolds number ($Re$) and Froude number ($Fr$). Topological differences in wake structure across these different regimes have been clarified recently through the use of Dynamic Mode Decomposition (DMD) on Direct Numerical Simulation (DNS) and laboratory data for a sphere in a stratified fluid for $Re\in [200,1000]$ and $Fr\in[0.5,16]$. In this work, we attempt to identify the dynamic regime from limited measurement data in a stratified wake with (nominally) unknown $Re$ and $Fr$. A large database of candidate basis functions is compiled by pooling the DMD modes obtained in prior DNS. A sparse model is built using the Forward Regression with Orthogonal Least Squares (FROLS) algorithm, which sequentially identifies DMD modes that best represent the data and calibrates their amplitude and phase. After calibration, the velocity field can be reconstructed using a weighted combination of the dominant DMD modes. The dynamic regime for the measurements is estimated via a projection-weighted average of $Re$ and $Fr$ corresponding to the identified modes. Regime identification is carried out from a limited number of 2D velocity snapshots from numerical and experimental datasets, as well as 3 point measurements in the wake of the body. A metric to assess confidence is introduced based on the observed predictive capability. This approach holds promise for the implementation of data-driven fluid pattern classifiers.