Hostname: page-component-7b7567c8c-tnkh4 Total loading time: 0 Render date: 2026-11-16T13:13:49.549Z Has data issue: false hasContentIssue false

Metric for attractor overlap

Published online by Cambridge University Press:  12 July 2019

Rishabh Ishar
Affiliation:
Department of Mechanical Engineering, Punjab Engineering College, Chandigarh 160012, India
Eurika Kaiser
Affiliation:
University of Washington, Department of Mechanical Engineering, Stevens Way, Box 352600, Seattle, WA 98195, USA
Marek Morzyński
Affiliation:
Poznań University of Technology, Chair of Virtual Engineering, Jana Pawla II 24, PL 60-965 Poznań, Poland
Daniel Fernex
Affiliation:
Institut für Strömungsmechanik, Technische Universität Braunschweig, Hermann-Blenk-Str. 37, 38108 Braunschweig, Germany
Richard Semaan
Affiliation:
Institut für Strömungsmechanik, Technische Universität Braunschweig, Hermann-Blenk-Str. 37, 38108 Braunschweig, Germany
Marian Albers
Affiliation:
Institute of Aerodynamics, RWTH Aachen University, Wüllnerstr. 5a, 52062 Aachen, Germany
Pascal S. Meysonnat
Affiliation:
Institute of Aerodynamics, RWTH Aachen University, Wüllnerstr. 5a, 52062 Aachen, Germany
Wolfgang Schröder
Affiliation:
Institute of Aerodynamics, RWTH Aachen University, Wüllnerstr. 5a, 52062 Aachen, GermanyForschungszentrum Jülich, JARA-High-Performance Computing, 52425 Jülich, Germany
Bernd R. Noack
Affiliation:
Institut für Strömungsmechanik, Technische Universität Braunschweig, Hermann-Blenk-Str. 37, 38108 Braunschweig, GermanyLIMSI, CNRS, Université Paris-Saclay, Bât 507, rue du Belvédère, Campus Universitaire, F-91403 Orsay, FranceInstitut für Strömungsmechanik und Technische Akustik (ISTA), Technische Universität Berlin, Müller-Breslau-Straße 8, 10623 Berlin, GermanyInstitute for Turbulence-Noise-Vibration Interaction and Control, Harbin Institute of Technology, Shenzhen Campus, 518055 Shenzhen, China

Abstract

We present the first general metric for attractor overlap (MAO) facilitating an unsupervised comparison of flow data sets. The starting point is two or more attractors, i.e. ensembles of states representing different operating conditions. The proposed metric generalizes the standard Hilbert-space distance between two snapshot-to-snapshot ensembles of two attractors. A reduced-order analysis for big data and many attractors is enabled by coarse graining the snapshots into representative clusters with corresponding centroids and population probabilities. For a large number of attractors, MAO is augmented by proximity maps for the snapshots, the centroids and the attractors, giving scientifically interpretable visual access to the closeness of the states. The coherent structures belonging to the overlap and disjoint states between these attractors are distilled by a few representative centroids. We employ MAO for two quite different actuated flow configurations: a two-dimensional wake with vortices in a narrow frequency range and three-dimensional wall turbulence with a broadband spectrum. In the first application, seven control laws are applied to the fluidic pinball, i.e. the two-dimensional flow around three circular cylinders whose centres form an equilateral triangle pointing in the upstream direction. These seven operating conditions comprise unforced shedding, boat tailing, base bleed, high- and low-frequency forcing as well as two opposing Magnus effects. In the second example, MAO is applied to three-dimensional simulation data from an open-loop drag reduction study of a turbulent boundary layer. The actuation mechanisms of 38 spanwise travelling transversal surface waves are investigated. MAO compares and classifies these actuated flows in agreement with physical intuition. For instance, the first feature coordinate of the attractor proximity map correlates with drag for the fluidic pinball and for the turbulent boundary layer. MAO has a large spectrum of potential applications ranging from a quantitative comparison between numerical simulations and experimental particle-image velocimetry data to the analysis of simulations representing a myriad of different operating conditions.

Information

Type
JFM Papers
Copyright
© 2019 Cambridge University Press 
Figure 0

Figure 1.Schematic of the pipeline for the metric of attractor overlap and associated proximity maps. Here, $P_{1}$, $P_{2}$ and $P_{3}$ refer to data sets outlined in § 3, i.e. sets of snapshots which are statistically representative of an attractor under a given operating condition. Of course, the number of data sets is typically much larger. $D$ is the metric of attractor overlap (§ 2.1). Its two arguments $P$, $Q$ represent data sets, like $P_{1}$, $P_{2}$ and $P_{3}$. The proximity map (top right) is explained in § 2.3. The coarse graining by clusters (centre row) is discussed in § 2.4 and the identification of overlapping and disjoint attractor regions (bottom row) is outlined in § 2.5.

Figure 1

Figure 2.Schematic of the metric of attractor overlap for continuous data.

Figure 2

Figure 3.Configuration of the fluidic pinball. The Cartesian coordinate system $(x,y)$ is depicted at the centre of the cylinders.

Figure 3

Figure 4.Grid of the fluidic pinball simulation.

Figure 4

Figure 5.Operating conditions of the fluidic pinball simulation. (c) Time series of the actuation commands for each cylinder over all phases together. (a) The flow of each phase is illustrated with a vorticity snapshot corresponding to maximum lift or minimum drag of the converged phase. (b) The time series of the total drag $F_{x}$ and total lift $F_{y}$ are displayed.

Figure 5

Table 1.Summary of the different control phases of the fluidic pinball simulation and the associated actuation mechanism.

Figure 6

Figure 6.Overview of the physical domain of the actuated turbulent boundary layer flow, where $L_{x}$, $L_{y}$ and $L_{z}$ are the dimensions of the domain in the Cartesian directions, $\unicode[STIX]{x1D706}$ is the wavelength of the spanwise travelling wave, $x_{0}$ marks the onset of the actuation and $x_{\unicode[STIX]{x1D70F}_{w},start}$ and $x_{\unicode[STIX]{x1D70F}_{w},end}$ denote the interval of the integration of the wall-shear stress $\unicode[STIX]{x1D70F}_{w}$.

Figure 7

Table 2.Flow and grid parameters of the non-actuated reference case and the 37 actuated cases.

Figure 8

Figure 7.Illustration of the turbulent boundary layer flow: (a,d) non-actuated reference case $N_{1}$, (b,e) actuated case with highest drag reduction $N_{24}$ and (c,f) actuated case with lowest drag reduction $N_{2}$; (a–c) contour plots of the instantaneous Cartesian velocity components $u$, $v$ and $w$ in a $y{-}z$ plane at $x/\unicode[STIX]{x1D703}\approx 65$; (d–f) time evolution of the instantaneous relative drag reduction $\unicode[STIX]{x0394}c_{d}$.

Figure 9

Figure 8.Contour plot of the $\unicode[STIX]{x1D706}_{2}$-criterion (Jeong & Hussain 1995), coloured by the instantaneous streamwise velocity, for three turbulent boundary layer flows; non-actuated reference case $N_{1}$, actuated highest drag reduction case $N_{24}$, actuated lowest drag reduction case $N_{2}$.

Figure 10

Table 3.Actuation parameters of the turbulent boundary layer simulations, where each set-up is denoted by a case number $N$. The quantity $\unicode[STIX]{x1D706}^{+}$ is the spanwise wavelength of the travelling wave, $T^{+}$ is the period and $A^{+}$ is the amplitude, all given in inner units, i.e. non-dimensionalized with the kinematic viscosity $\unicode[STIX]{x1D708}$ and the friction velocity $u_{\unicode[STIX]{x1D70F}}$. Each block includes set-ups with varying period and amplitude for a constant wavelength. The list includes the values of the averaged relative drag reduction $\unicode[STIX]{x0394}c_{d}$, and the averaged relative skin-friction reduction $\unicode[STIX]{x0394}c_{f}$.

Figure 11

Figure 9.Cluster centroids of the fluidic pinball simulation. The centroids are sorted in order of appearance neglecting the first 50 time units (transients) of each phase. The visualization depicts the vorticity distribution: green, red and blue represent vanishing, positive and negative values, respectively.

Figure 12

Figure 10.Cluster affiliation $k$ of the fluidic pinball simulation as a function of time $t$. The applicable control laws are described in table 1 and the corresponding centroid $\boldsymbol{c}_{k}$ is depicted in figure 9. By construction, new cluster indices are a monotonically increasing function of time, neglecting the initial actuation transients. Grey curve sections correspond to transients while red sections mark the post-transient data taken for the analysis.

Figure 13

Figure 11.Population probabilities of the clusters for the seven fluidic pinball phases (table 1). The abscissa denotes the phase $l$ and the ordinate the cluster index $k$. Each box represents a non-vanishing probability. The yellow to red tones indicate the population probability $P_{k}^{l}$ from (2.17). Probabilities below 1 % are kept white. Note that steady boat tailing ($l=2$) is represented by a single centroid.

Figure 14

Figure 12.Metric of attractor overlap (2.18) for the seven control laws of the fluidic pinball (table 1). The value of $D_{geom}(\boldsymbol{P}^{l},\boldsymbol{P}^{n})$ is shown in the $l$th row and $n$th column as colour code from white to red and by the size of the circle as illustrated by the caption on the right.

Figure 15

Figure 13.Proximity map of the fluidic pinball showing the centroids (large circles) together with snapshots (small circles). Snapshots from transients are displayed as grey dots, while the data for the analysis are colour coded by cluster affiliation.

Figure 16

Figure 14.Attractor proximity map of the fluidic pinball: (a) based on the snapshot-based MAO (2.5) and (b) on a cluster-based MAO (2.18) with $K=50$ clusters. The points $P_{1}\ldots P_{7}$ correspond to the seven control laws in table 1. Note that the snapshot-based MAO represents a cluster-based MAO with maximum number of clusters.

Figure 17

Figure 15.Population probabilities of the clusters for the 38 turbulent boundary layer simulations (see table 3). The abscissa denotes the data set $l$ and the ordinate the cluster index $k$. Each box represents a non-vanishing probability. The yellow to red tones indicate the population probability or relative frequency $P_{k}^{l}$ from (2.17). Probabilities below 1 % are kept white. The actuation wavelength is indicated by the three vertical line separating simulations with no actuation, $\unicode[STIX]{x1D706}^{+}=200$, $\unicode[STIX]{x1D706}^{+}=500$ and $\unicode[STIX]{x1D706}^{+}=1000$ from left to right. The actuation amplitude $A^{+}$ and time $T^{+}$ are shown in the top caption.

Figure 18

Figure 16.Metric of attractor overlap (2.18) of the turbulent boundary layer with 38 control laws (see table 3). The visualization is analogous to figure 12.

Figure 19

Figure 17.Proximity map of the actuated turbulent boundary layer. The figure displays the centroids (large circles) together with snapshots (small circles). The data for the analysis are colour coded by cluster affiliation.

Figure 20

Figure 18.Attractor proximity map of the turbulent boundary layer simulations (see table 3): (a) colour coded with relative drag reduction $\unicode[STIX]{x0394}c_{D}$; (b) colour coded with wavelength $\unicode[STIX]{x1D706}^{+}$. The actuation parameters of each symbol is indicated in (a) by a number corresponding to the index from table 3.

Figure 21

Figure 19.The POD modes $\boldsymbol{u}_{i}(\boldsymbol{x})$, $i=1,\ldots ,10$. Visualization as in figure 9.

Figure 22

Figure 20.POD amplitudes $a_{i}(t)$, $i=1,\ldots ,10$. Snapshot sequences after the decay of the transients are employed for the analysis (coloured in red).

Figure 23

Figure 21.Distances between attractors: (a) Kullback–Leibler divergence and (b) Jensen–Shannon divergence based on the cluster probability distribution for each of the seven control phases. The row and column of $D_{KL}(P\Vert Q)$ correspond to the first ($\boldsymbol{P}$) and second arguments ($\boldsymbol{Q}$). The value of the divergence is colour coded and indicated by the size of the circle (see the corresponding caption).