Using braids to quantify interface growth and coherence in a rotor-oscillator flow

Margaux Filippi; Marko Budisic; Michael R. Allshouse; Severine Atis; Jean-Luc Thiffeault; Thomas Peacock

doi:10.1103/PhysRevFluids.5.054504

Using braids to quantify interface growth and coherence in a rotor-oscillator flow

Filippi, Margaux;Budisic, Marko;Allshouse, Michael R.;Atis, Severine;Thiffeault, Jean-Luc;Peacock, Thomas 2019-10-28 00:00:00 Using braids to quantify interface growth and coherence in a rotor-oscillator ﬂow Margaux Filippi Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A. and Woods Hole Oceanographic Institution, Woods Hole, MA 02543, U.S.A. Marko Budiˇ si´ c Department of Mathematics, Clarkson University, Potsdam, NY 13699, U.S.A. Michael R. Allshouse Department of Mechanical and Industrial Engineering, Northeastern University, Boston, MA 02115, U.S.A. S´ everine Atis Department of Physics, University of Chicago, Chicago, IL 60637, U.S.A. Jean-Luc Thiﬀeault Department of Mathematics, University of Wisconsin – Madison, Madison, WI 53706, U.S.A. Thomas Peacock Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A. (Dated: May 26, 2020) The growth rate of material interfaces is an important proxy for mixing and reaction rates in ﬂuid dynamics, and can also be used to identify regions of coherence. Estimating such growth rates can be diﬃcult, since they depend on detailed properties of the velocity ﬁeld, such as its derivatives, that are hard to measure directly. When an experiment gives only sparse trajectory data, it is natural to encode planar trajectories as mathematical braids, which are topological objects that contain information on the mixing characteristics of the ﬂow, in particular through their action on topological loops. We test such braid methods on an experimental system, the rotor-oscillator ﬂow, which is well-described by a theoretical model. We conduct a series of laboratory experiments to collect particle tracking and particle image velocimetry data, and use the particle tracks to identify regions of coherence within the ﬂow that match the results obtained from the model velocity ﬁeld. We then use the data to estimate growth rates of material interface, using both the braid approach and numerical simulations. The interface growth rates follow similar qualitative trends in both the experiment and model, but have signiﬁcant quantitative diﬀerences, suggesting that the two are not as similar as ﬁrst seems. Our results shows that there are challenges in using the braid approach to analyze data, in particular the need for long trajectories, but that these are not insurmountable. I. INTRODUCTION and arise in engineered systems and technological pro- cesses [19–23]. Characterization of material mixing in ﬂuid ﬂows is an Quantiﬁcation of mixing typically focuses on charac- active research area at the intersection of mathematics, terizing the amount of stirring by chaotic advection [2], engineering and physics. Two common types of anal- which in turn relates to the role that stretch-and-fold ysis are to (1) quantify the overall amount of stirring and stretch-and-stack mechanisms play in mixing of ma- and mixing that occurs in the system and (2) to iden- terial [1, 5, 6, 24, 25]. The strength of these mecha- tify structures that enhance or reduce material mixing. nisms is computed by estimating local rates of exponen- The earliest examples of a systematic mathematical ap- tial stretching of the material [26, 27] over a timescale as- proach to these problems uses techniques from nonlinear sociated with folding or stacking. Alternatively, chaotic dynamics that rely on having the ﬂuid velocity ﬁeld as advection can be quantiﬁed by studying norms of scalar a continuous, often diﬀerentiable, function of space and ﬁelds advected in the ﬂow [28–30]. A review of associated time [1–4]. (See for example [1, 5–11] for comprehen- topics with additional references can be found in [11]. sive reviews.) By now, these techniques have been suﬃ- The search for structures that enhance or prevent mix- ciently adapted and reﬁned to be applicable to data from ing typically focuses on features of the ﬂow that remain real ﬂuid ﬂows that, for example, arise in and around liv- coherent over relevant timescales and organize transport ing organisms [12–15], govern ecological processes [16–18] by, for example, attracting, repelling, or containing the material. The most well-known of such structures are the Lagrangian Coherent Structures (LCS) [31], which are margaux@mit.edu distinguished low-dimensional barriers to transport, and arXiv:1910.12779v2 [physics.flu-dyn] 23 May 2020 2 Almost-Coherent Sets [32], which are regions in which uate how well suited braid theory is for the characteri- the transport is contained. In all cases, the true challenge zation of experimental ﬂuid ﬂows. The ﬂow studied here remains to deﬁne the objects of interest so that they en- is the rotor-oscillator Stokes ﬂow, a canonical example compass all relevant physical phenomena, and to propose of a ﬂow ﬁeld that possesses both chaotic and coherent and implement an algorithm that identiﬁes them, espe- regions, described by Hackborn et al. [77] and Weldon cially in transient or aperiodic ﬂows [31, 33–37]. The de- et al. [19]. In contrast to ﬂows such as the double-gyre, tection of coherent structures oﬀers especially-signiﬁcant Bickley jet, or the Duﬃng oscillator, which have been applications to geophysical ﬂows, where these techniques more commonly used to analyze material transport, the have been used to understand climate change and plan re- rotor-oscillator ﬂow has been analyzed both as an ana- sponses to ecological catastrophes [17, 38–41]. The most lytical model [77] and as an experimental ﬂow [19]. The commonly used tools to detect coherent structures are braid-based analysis will be twofold: ﬁrst, it will pro- based on material deformation [31, 42–44] and on prob- vide an identiﬁcation of regions of the ﬂow surrounded abilistic [32, 45] properties of the ﬂow. A comparison by minimally growing material lines (typically, coherent of the wide range of approaches for detecting coherent material sets); second, an estimation of maximum growth structures can be found in [36]. rate of material interfaces (typically found in the chaotic region). Both calculations will be applied to the analyt- Input data in the form of continuous velocity ﬁeld ical model and to the experimental data, in an eﬀort to stands in contrast to observations in oceans and atmo- expose and address challenges that come with analyzing sphere that are recorded by deploying sparse sensors that non-idealized ﬂows. measure physical properties (temperature, salinity, etc.) The paper is organized as follows. Section II presents of the ﬂuid ﬂow as they are advected [46–48]. The rela- tive sparsity of sensors prevents a reliable estimation of a the model and experimental ﬂows used in this study. Sec- tion III gives a short summary of the braid representation continuous velocity ﬁeld; to analyze such sparse datasets, a number of methods have been developed that require of the ﬂow kinematics. Next, the paper demonstrates how two particular analysis methods can be applied to only a ﬁnite set of discretized trajectories [49–54]. model and experimental ﬂows: Section IV explains the Sparse-data methods include the braid dynamics meth- detection of coherent structures, comparing the analysis ods employed in this paper, which require only a set of of numerically-advected and experimental particle tra- discrete trajectories instead of the full velocity ﬁeld. Fol- jectories, while Section V explains estimation of material lowing Boyland et al. [55], the study of the topological growth, comparing the analysis of model and experimen- properties of ﬂuid stirring has developed into an active tal velocity ﬁelds. Section VI summarizes what has been research area. The topological approach is particularly learned by employing braid dynamics to study model and well-suited to the study of mixing by rods, vortices, or experimental rotor-oscillator ﬂows. otherwise distinguished Lagrangian trajectories [56–64]. In particular Thiﬀeault [56, 65] and Allshouse and Thif- feault [49], Budiˇ si´ c and Thiﬀeault [66] used braid theory to characterize mixing and coherent structures from pla- II. THE ROTOR-OSCILLATOR FLOW nar ﬂows solely from particle trajectories, forming the basis for the approaches used in this paper. The rotor-oscillator ﬂow is a planar, non-autonomous, In planar ﬂows, the rate of exponential growth of ma- incompressible ﬂow where the motion of the ﬂuid is in- terial interfaces corresponds to the topological entropy duced by a rotor that simultaneously rotates about its of the ﬂow [55, 56, 65, 67–70], which quantiﬁes the com- axis and oscillates along a channel. Experimentally, plexity of trajectories evolving in a dynamical system. particle motion is conﬁned to a two-dimensional ﬂuid Techniques for estimating topological entropy (without layer in a rectangular tank, with advection induced by estimating growth rate ﬁrst) typically use precise veloc- a fast-spinning cylinder that also oscillates in the longi- ity ﬁelds to compute unstable periodic orbits [71–74], or tudinal direction. This section summarizes the analytic intricate partitions of the ﬂow [75, 76]. Budiˇ si´ c and Thif- model and the experimental realization; further details feault [66] developed a method for calculating the Finite- are found in [19, 77, 78]. Time Braiding Exponent (FTBE), which is an approxi- mation for the topological entropy that is more applicable to ﬁnite-time, sparse datasets. The FTBE provides a ro- A. Analytic model bust measure of mixing that approaches the topological entropy as the number of trajectories is increased. More recently, Roberts et al. [64] developed a braid-free ap- Hackborn et al. [77] studied the rotor-oscillator ﬂow, proach that also estimates topological entropy based on an asymmetric ﬂow between parallel plates driven by the the relationship with growth of material lines, without rotation of a cylinder (rotor ) and the longitudinal oscilla- detailed knowledge of the velocity ﬁeld. tion of one of the walls of the ﬂow tank (oscillator ). In a Since the methods of Budiˇ si´ c and Thiﬀeault [66] and suitable parameter regime, the ﬂow exhibits a main vor- Allshouse and Thiﬀeault [49] were developed and applied tex around the rotor, with two secondary recirculating to model ﬂows, the principal aim of this paper is to eval- vortices on each side of the rotor. Each of those vortices 3 −1 −0.5 0.5 −3 −2 −1 0 1 2 3 −0.5 −0.25 FIG. 1. Geometry for the rotor-oscillator ﬂow. The walls are at x = ±h, with the rotor oscillating in y at constant x = c. 0.25 0.5 −1.5 −1 −0.5 0 0.5 1 1.5 creates its own, progressively-weaker, array of vortices, which will not play a role in this work. We take coordinates (x, y) ∈ [−h, h]×R, with the walls FIG. 2. Poincar´ e map for the model rotor-oscillator ﬂow of the channel at x = ±h. The cylindrical rotor starts at [19, 77] with c = 0.54, = 0.125, λ = 2π/5 simulated for y = 0 at t = 0 and moves along the line x = c. The rotor 300 forcing cycles. The blue and red colors depend on the av- has radius a and angular velocity ω. A diagram of the erage value of the y coordinate along simulated trajectories. geometry is shown in Figure 1. (The peculiar interchange Bottom panel shows a detailed view of the dashed frame. The of x and y axes is for consistency with earlier work [77].) Poincar´ e map is sampled at the rotor’s oscillation period τ . Without longitudinal (y) oscillation, the rotating cylin- with non-dimensional parameters der induces, in the Stokes limit, a steady velocity ﬁeld. In non-dimensional coordinates, the span-wise width is 2 2 2 = V h/2a ω, λ = h Ω/a ω, (4) x ∈ [−1, 1] with the position of the rotor c = 0.54 set to match [77]. We absorb the rotation frequency ω into the that combine the channel width h, magnitude of oscil- non-dimensional time t, so the resulting stream function lation V and angular frequencies of rotation ω and of ψ(x, y) [77, 78] is oscillation Ω. The use of a translated steady solution is valid as long as Ω is not so large as to invalidate the ψ(x, y) = log[f (x, y)] + g(x, k) cos(ky)dk (1) Stokes approximation. | {z } | {z } Passive particles are advected via point vortex boundary eﬀects x ˙ (t) = −∂ ψ(x, y − sin λt), where (5) y ˙ (t) = ∂ ψ(x, y − sin λt). cosh(πy/2)− cos(π(c− x)/2) f (x, y) = ; Fixing c = 0.54, Hackborn et al. [77] varied and λ to cosh(πy/2) + cos(π(c + x)/2) identify regimes with chaotic dynamics and islands of co- 2[tanh k cosh kx− x sinh kx] cosh kc herence using Poincar´ e maps. Figure 2 shows a Poincar´ e g(x, k) = sinh 2k + 2k map generated with = 0.125 and λ = 2π/5. 2[coth k sinh kx− x cosh kx] sinh kc Timescales that govern the evolution of a single La- + . sinh 2k − 2k grangian trajectory vary widely depending on its loca- tion. Trajectories that encounter the rotor will circle Following [77], we assume the rotor to be a point vortex around it rapidly; trajectories that remain on the outside with a vanishing diameter (a → 0) while keeping a ω edge of both the rotor and recirculating vortices evolve on constant. The time non-dimensionalization is kept as timescales separated by two or more orders of magnitude follows: compared to the rotor rotation. ta ω → t. (2) B. Experimental ﬂow The longitudinal oscillation of frequency Ω is mod- The experimental apparatus was inspired by Weldon eled by the time-periodic translation of the steady stream et al. [19], and is depicted in Figure 3. The main diﬀer- function ψ(x, y, t): ence compared to the analytical model of Hackborn et al. Ψ(x, y, t) = ψ(x, y − sin λt), (3) [77] is in apparent change of the reference frame: they ﬁx 4 the rotor and moved the walls, while Weldon et al. [19] ﬁx the walls and move the rotor. The rotor-oscillator system was recreated in a 90 mm× 402 mm acrylic ﬂow tank. For the rotor, an alu- minum rod of diameter a = 3.165 mm was attached to a stepper motor through a plastic sleeve for thermal insu- lation. To longitudinally oscillate the rotor, the stepper motor was mounted to a longitudinal traverse controlled by a second stepper motor. The tank and the traverse were mounted onto alu- minum frames and aligned horizontally. To visualize a horizontal cross-section of the system, a front-faced mir- (a) ror was placed underneath the tank at a 45 angle. A camera was mounted facing the mirror with the rod at the center of the ﬁeld of view. The experimental images were acquired through the mirror reﬂection at 45 and recorded with a LaVision Imager Pro X 4M CCD cam- vegetable oil era of resolution 2042× 2042 pixels with a 28 mm lens. Image acquisition, calibration and cross-correlation was performed by LaVision’s DaVis imaging software. To mitigate three-dimensional eﬀects due to evaporation or interaction with the bottom wall, the ﬂuid system was trapped between a top layer of vegetable oil and a bot- tom layer of FC-40 coolant. Two separate experiments were used to record proper- (b) ties of the ﬂow: a particle tracking experiment (PT, Sec- tion II B 1), to record Lagrangian trajectories, and a par- FIG. 3. Schematics of the experimental set-up for the ticle image velocimetry experiment (PIV, Section II B 2), Hackborn–Weldon rotor-oscillator ﬂow [19, 77], with still to record instantaneous velocity ﬁelds. Table I summa- walls and an oscillating rotor. (a) Experimental apparatus set up. (b) Fluid system to set up a two-dimensional ﬂow rizes the experimental parameters for the two experi- through density layers for the PT experiment. In the PIV ments. For all experiments, the Reynolds number was experiment, the tracer particles were mixed to a single layer estimated to be O(1). of glycerol and a laser sheet was projected horizontally. 1. Particle tracking (PT) experiment There were 77 918 frames (time steps) recorded, for a total run duration of 7791.8 s, corresponding to about Tracer particles were custom-made from cellulose ac- 82.3 periods. The particle trajectories were obtained etate polymer spheres by Cospheric, with diameter through a MATLAB package, the Tracking Code Repos- 1.78 mm and density 1.285 g/cm . The particles were itory [79]. For the particle tracking experiment, the non- painted ﬂuorescent to create a high light contrast and dimensional parameters were = 0.1274 and λ = 1.2806. thereby enhance image acquisition. The coated par- The physical period of oscillation was τ ˜ = 94.7163 s. ticles were then ﬁltered by density to ensure they re- mained on a virtually two-dimensional plane, the sur- face between the lower layer of salty glycerol, of den- 2. Particle image velocimetry (PIV) experiment sity 1.297 g/cm , and the upper layer of pure glycerol, of density 1.261 g/cm . The denser glycerol solution was mixed with salt and water to match the viscosity of Tracers for the PIV experiment were hollow glass −3 2 pure glycerol, measured to be ν = 1.1197× 10 m /s. spheres of mean diameter 8−12 μm mixed with glycerol. The ﬂuorescent particles were illuminated with an ul- The Reynolds number of the glass spheres at the rotor’s −4 traviolet light in a dark room. With the bead diame- boundary was Re ≈ 1.9977× 10 and are assumed sphere −3 ter D = 1.78× 10 m, the viscosity ν, a rod of radius to move as passive tracers. A pulsed Nd:YAG laser was a = 3.165 mm and angular velocity of ω = 10.4720 rad/s, powered to illuminate the glycerol–glass spheres solution the Reynold number of the beads at the rotor’s bound- in a horizontal plane normal to the tank front wall. To re- (ωa)D −2 ary was Re = ≈ 5.2689× 10 . The velocity duce artifacts and noise in the velocity ﬁelds, the domain bead being maximized at the rotor’s boundary, Re is an was partitioned into two for image processing in DaVis; bead upper bound for the ﬂow ﬁeld, so the neutrally buoyant this helped compensate for the wide range of velocities beads are assumed to move as passive tracers. between the rod’s rotation and the slower motion on the The images were acquired at a frequency f = 10 Hz. outer domain. First, the region around the rod was pro- 5 TABLE I. Experimental parameters. symbol description value (PIV) value (PT) h half-width of the channel 0.045 m a rotor radius 0.003 165 m c distance of rotor from the wall 0.023 76 m τ ˜ translational period of oscillation 96.5201 s 94.7163 s −4 −4 V translational velocity magnitude 5.8275× 10 m 5.9385× 10 m Ω translational angular velocity 0.0651 rad/s 0.0663 rad/s ω rotational angular velocity 10.4720 rad/s Re Reynolds number 2.6640 f sampling frequency 1 Hz 10 Hz non-dim. oscillation magnitude (V h/2a ω) 0.1250 0.1274 2 2 λ non-dim. oscillation ang. frequency (h Ω/ωa ) 2π/5 ≈ 1.2566 1.2806 In calculating the Reynolds number for the experiments, we use ωa for the characteristic linear velocity and 2h as the characteristic length, resulting in Re = 2ωah/ν. Hackborn 2 −2 et al. [77] use a as the characteristic length, resulting in Re = a ω/ν ≈ 9.3682× 10 here. Eﬀective frequency of sampling. See Section II B 2 for details. cessed at f = 10 Hz and bin-averaged in 10-image blocks; on to the trajectory integrator. Trajectories were inte- second, the outer domain was processed at f = 1 Hz and grated using the variable-step, variable-order MATLAB matched to the bin centers of the ﬁrst part. The re- ode15s [81] algorithm to manage the stiﬀness of the dif- sulting eﬀective sampling rate was f = 1 Hz. There were ferential equations, which is a consequence of the large 10 730 frames recorded, for a total run duration of 1073 s, diﬀerence in timescales of primary and secondary rotors. corresponding to about 11.1 periods. For the PIV exper- To simulate times longer than the recorded number of iment, the non-dimensional parameters were = 0.1250 oscillation periods, the PIV velocity ﬁeld was periodized and λ = 1.2566. The physical period of rotor oscillation in time. was τ ˜ = 96.5201 s, which corresponds to 5 dimensionless time units. III. BRAID DYNAMICS The velocity ﬁelds were post-processed using the ap- proach of Kelley and Ouellette [80] to impose incom- pressibility for the PIV-obtained velocity ﬁeld; this was Braid theory is an algebraic way of characterizing and necessary because of the multiple repeated periods over classifying continuous maps based on their topological which the PIV data was used for our investigations. Fig- properties. In our application, the continuous maps are ure 4 demonstrates that instantaneous streamlines of the ﬂow maps generated by the two-dimensional (un- the model and experimental velocity ﬁelds qualitatively steady) ﬂuid velocity ﬁeld over a particular time interval, match. Due to experimental constraints, however, the as studied by [55–57, 65, 67]. The “input data” for the streamlines obtained from the PIV experiments deviate braid theory characterization of ﬂuid ﬂows is a set of N from the streamlines obtained from the model velocity continuous particle trajectories evolved concurrently over ﬁelds. Some of these discrepancies come from the fun- a ﬁnite time interval; in particular, the analysis does not damental diﬀerences between the model and experimen- require access to the velocity ﬁeld or its gradients. tal ﬂow. First, the model assumes the rotor to be a Braids are constructed from a set of N particle tra- point vortex, whereas a physical rod had to be present 2 jectories p (t) ∈ R , i = 1, 2, . . . , N , with 0 ≤ t ≤ T . to stir the ﬂuid. In spite of the small rod diameter of A physical braid is the embedding of trajectories in the a = 3.165 mm, the ﬂows deviate around the rotor. Sec- three-dimensional space-time (Fig. 5(a)), where individ- ond, the model assumes a Stokes ﬂow with zero Reynolds ual trajectories (strands ) weave around each other. A number: with a 6= 0 as the characteristic length, the ex- topological braid (Fig. 5(c)) is a reduced representation −2 perimental Reynolds number is Re ≈ 9.3682× 10 , as of the physical braid that retains only the sequence of ex- described in Table I. Additionally, the deviations are es- changes of the strand order with respect to a ﬁxed space- pecially prominent near the lateral boundaries, due to the time plane onto which the strands are projected. This laser beam’s reﬂection at the walls of the tank, which in- plane can be chosen largely arbitrarily for the purposes troduced artifacts in the velocity ﬁelds. The PIV record of this article [56, 65, 66, 82]. is not available for the full width of the channel, but Strand exchanges can be represented by a sequence of rather in the approximate band −0.85 < x < 0.85 in the symbols σ called Artin braid generators. The generator nondimensional spanwise coordinate. σ represents the crossing of the ith strand in front of the To create synthetic trajectories in the experimental (i + 1)th strand in Fig. 5(c), where the index i indicates velocity ﬁeld, the velocities were linearly interpolated the order of strands from left to right just before the −1 between spatial nodes and time samples before passed crossing occurs; the inverse generator σ represents the i 6 −1 −0.5 0.5 t t −−2 2 −1 0 1 2 2 (a) −1 −0.5 x (a) Physical braid (b) Projection onto (c) Topological x-coordinate braid 0.5 FIG. 5. A physical braid and a corresponding topologi- cal braid generated from ﬁve trajectories. In all diagrams the time ﬂows from bottom to top. The sequence of gen- −−2 2 −1 0 1 2 2 erators (ordered from left to right in increasing time) is −1 −1 −1 −1 B = σ σ σ σ σ σ σ σ σ σ σ . 4 1 2 3 2 4 3 2 2 4 4 (b) FIG. 4. Snapshot of instantaneous streamlines of (a) the model velocity ﬁeld [19, 77] with c = 0.54, = 0.125, λ = 2π/5 (a) and (b) the PIV-recorded experimental velocity ﬁeld at the same time instance. (b) crossing of ith strand behind the (i + 1)th strand. A topological braid constructed from a given set of FIG. 6. Two topological loops before (left column) and af- trajectories captures topological information about the ter (right column) action of the braid from Figure 5. The ﬂow. The topological analog of material advection in strands of the braid are shown in black as cross-sections. The ﬂuid ﬂows is the action of a braid on topological loops. A Dynnikov vectors before and after the braid action in (a) topological loop is a collection of closed, non-intersecting B B 0 −2 3 0 −1 0 are [ 1 1 1 0 0 0 ] −→ [ ] and in (b) [ 0 0 0 0 0 1 ] −→ curves that enclose two or more (but not all) strands in [ 0 −1 1 0 −1 0 ]. the braid (Figure 6). The curves in a loop are “pulled tight,” i.e., they can be pictured as rubber bands tightly wrapped around strands of the physical braid. The ac- tures and between the punctures, respectively. The Dyn- tion of a braid B on a loop `, denoted by B`, is visualized nikov coordinate vector is calculated from intersection by sliding the rubber bands along the physical braid in numbers as the direction of time. As strands exchange places, the loop is forced to stretch since it cannot pass through the a a ··· a b b ··· b , where 1 2 N−2 1 2 N−2 (6) strands (Figure 6). a = (α − α )/2, b = (β − β )/2. n 2n 2n−1 n n+1 n Computationally, any loop acted on by the N -strand braid can be represented by a vector of 2N − 4 signed The action of each braid generator σ is then repre- 2N−4 2N−4 integers, and vice-versa, through the Dynnikov coor- sented by a piecewise-linear map σ : Z → Z ; dinatization of loops [65, 83, 84]. Figure 7 demon- explicit expressions for σ can be found in [65, 84]. strates how to calculate the Dynnokov coordinates of All braid computations and visualizations in this pa- a loop around N = 5 strands. The intersection num- per have been produced using the MATLAB toolbox bers α , α ,··· , α count the number of intersections braidlab [82]. 1 3 2N−3 between the tightened loop and axes above the punc- To measure the amount of stretching, we compute the tures, while α , α ,··· , α and β ,··· , β count “length” of a topological loop|`| as the number of times it 2 4 2N−4 1 N−1 the number of intersections with axes below the punc- intersects the horizontal axis that would pass through all 7 α α α β 1 β 3 β 5 β 1 2 3 4 work improves the algorithm to more eﬃciently detect co- herent structures for larger trajectory sets. The method is then applied to a synthetic set of trajectories based on the model velocity ﬁeld presented in II A and on experi- 1 2 3 4 5 mentally measured trajectories, allowing for comparison between theory and experiments. A. Theory α α α 2 4 6 There are several diﬀerent deﬁnitions of coherent struc- FIG. 7. Coordinate system based on intersection numbers tures used for characterizing patches of passive tracer α , β used to calculate Dynnikov coordinates (6). In this i i that do not disperse under Lagrangian transport [31, 44, example of N = 5 strands, α = 0, α = 2, β = 2, 1,3,5 2,4,6 1,2,3,4 87]. In the context of braid theory, Allshouse and Thif- 1 1 1 0 0 0 resulting in the Dynnikov vector [ ]. Precise location of strands in the x, y space is not important, only their order feault [49] deﬁne coherent sets as a set of trajectories along the projection coordinate. surrounded by an initially “simple” topological loop that does not grow, or grows sub-exponentially, over the du- ration of the braid. Sub-exponential growth can occur points in Figure 6. Although the quantity |`| on its own as particles rotate as a patch, so we do not consider this does not correspond to spatial length of a material curve, type of growth as indicative of long term mixing. We it is useful for characterizing the relative loop growth now summarize the relevant techniques from [49] with factor modiﬁcations to make the presentation more compact. Simple topological loops are those whose Dynnikov co- |B`|/|`|, (7) ordinates (see Section III) contain only {±1, 0}, i.e., their 2N−4 2N−4 coordinate vectors are {±1, 0} ⊂ Z . As the measuring the ratio of loop length before and after appli- magnitude of elements in the coordinate vectors relate cation of the braid B. The loop growth factor will play a to the number of folds of a loop around particles, these prominent role in the following sections as a criterion for loops will have one or no folds as they surround parti- searching for maximally and minimally growing loops `, cles (see ﬁgure Figure 8 for examples). As mentioned given a speciﬁc braid B. in [49], enumerating all simple loops and checking how The topological braid can be interpreted as a dramat- fast they grow under the action of the braid is com- ically reduced representation of the full ﬂow using only putationally intractable, even for modestly-size braids: 2N−4 2N−4 N a ﬁnite number of trajectories, with the physical trajec- #({±1, 0} ) = 3 ∼ 9 , a number that is pro- tories represented by topological operations, and mate- hibitively large for even N ≈ 20 strands in a braid. rial advection represented by stretching of “rubber-band” Therefore, we look to reduce the number of simple loops loops. Despite their simplicity, braid-based calculations considered, while still retaining a way to determine which provide bounds on the rate of material stretching in the advected particles are in coherent structures. full ﬂow, and approximate boundaries of Lagrangian co- To reduce the number of loops considered, Allshouse herent structures, as explained in Sections IV, V. Addi- and Thiﬀeault [49] focuses on pair-loops that enclose tionally, the reduction also yields signiﬁcant speed up as only two out of the N particles. Although there are compared to material line advection, which can rely on = N (N− 1)/2 particle pairs to choose, there are still expensive interpolation methods [85]. many simple loops that enclose any given pair. In con- trast to [49] where loops were completely above or below all particles between the particle pair, here we consider IV. DETECTION OF COHERENT a single pair-loop for each pair, namely the one that can STRUCTURES be “pinched” to a straight line in the initial conditions plane; we term these straight pair-loops. This creates a set of loops to consider, which is half as many The braid theory approach to detect coherent struc- tures oﬀers a unique mechanism for partitioning diﬀer- as in [49], while also yielding loops that have a simpler spatial shapes. Figure 8(a) contains the lines (red and ent regions of the ﬂow based on the entanglement of tra- green) connecting pairs of particles, and the correspond- jectories, as opposed to their relative positions [50] and material deformation [86]. A coherent structure identi- ing straight pair-loops in standard form are presented in Figure 8(b). ﬁed using braids is deﬁned as a topological loop that does not grow under the action of the braid. The loop provides The identiﬁcation of coherent structures as loops that an approximate physical boundary that delineates an in- do not grow under the action of the braid is broken into ternal ﬂuid subdomain that does not mix with exterior two parts: identiﬁcation of sets of particles enclosed by a ﬂuid. We present an overview of the theory developed by coherent structure, and creation of a loop around those Allshouse and Thiﬀeault [49], and show how the current particles that does not grow. The starting point for iden- 8 value, the number of coherent structures initially grows as more pair-loops become admissible. Eventually, ad- ditional pair-loops that are added will act as connec- tions between pre-existing connected components, merg- ing pairs of coherent structures. This will result in a de- crease in the number of coherent structures, and growth in size of individual coherent structures, until there is only one large structure containing all particles within the system. Once a set of particles has been identiﬁed as being con- tained by a coherent structure, the corresponding struc- ture is the composite of the straight pair-loops that con- nect the set of particles. In the example presented in Figure 8, three particles (colored in blue) are connected by straight pair-loops, a green loop and a red loop. The ﬁnal step in the coherent structure identiﬁcation is to FIG. 8. (a) Particles (dots) and loops in physical space. Parti- merge the non-growing straight pair-loops into a single cles in the identiﬁed structure (blue) are connected by straight non-straight loop, which we consider to be the bound- lines (solid) and surrounded by the corresponding straight ary of the coherent structure. In Figure 8, the coherent pair-loop (dashed and dotted lines). The coherent structure structure surrounds both straight pair-loops without en- loop (blue line) links the two straight pair-loops. (b) The veloping any of the particles outside the coherent struc- particles and loops in standard form, with points aligned on ture. 0 1 0 −1 0 −1 0 0 0 1 the real axis. The Dynnikov vectors are [ ] 0 1 1 −1 0 −1 0 0 0 1 for the blue curve, [ ] for the red curve and [ 0 1 0 0 0 −1 0 1 0 0 ] for the green curve. B. Coherent structures in the model system tifying sets of particles that are inside a coherent struc- While the braid method for detecting coherent struc- ture is the forming of straight pair-loops ` connecting ij tures was validated by Allshouse and Thiﬀeault [49], the particles p , p ∈ R in the initial condition plane. We i j use of only straight pair-loops is a modiﬁcation that re- represent the collection of N trajectories by vertices in quires similar veriﬁcation. To do so, we analyze a set (a family of ) undirected graphs G; the set of particles of trajectories generated using the model velocity ﬁeld enclosed by a coherent structures will correspond to con- presented in section II A. Because only a single pair- nected components in these graphs. The growth of the loop per pair of trajectories is analyzed, it is possible straight pair loops under the action of the braid will de- to process a larger number of trajectories. Increasing termine which edges e are included or excluded in G. ij the number of trajectories while keeping the domain size Based on the loop growth rate (7), compute L as ij constant has the potential to more accurately locate the the exponential growth rate of the straight pair-loop ` ij boundary of the coherent set with the braid method. As under the action of the braid demonstrated by Allshouse and Thiﬀeault [49], the braid method requires multiple particles to be located within 1 |B` | ij L = log . (8) ij a coherent structure in order for it to be identiﬁed; by T |` | ij increasing the number of trajectories analyzed, it is thus more likely that smaller structures will be identiﬁed. For The graph of vertices G is fully-connected and undi- rected as there is a pair-loop between each pair of ver- the numerically calculated trajectories, 300 trajectories are analyzed and the resulting coherent structures can tices, with edge weight L . Fix a threshold Λ ≥ 0 and ij construct a family of subgraphs G ⊂ G, retaining only be compared with the Poincar´ e map in Figure 2. edges that satisfy L ≤ Λ. G can be interpreted as a The numerically calculated trajectories are randomly ij Λ subgraph of G where connections are made only by sim- distributed throughout the entire domain and are ad- ple straight pair-loops whose exponential growth rate is vected for 50 periods to match the PT experiment. A less than Λ. sampling rate of 100 measurements per period is suﬃ- For any Λ there may be vertices in G that remain dis- cient to properly represent the braid. The 300-trajectory connected from any other vertex. In particular, vertices braid is represented by approximately 2.15 million gener- that are disconnected at Λ = 0 are deemed incoherent, ators. The application of the braid to all of the straight as they are not enclosed by any slow-growing pair-loops. pair-loops is the bottleneck of the algorithm and takes On the other hand, coherent structures will correspond to three hours on a laptop. We use a growth rate thresh- vertices that are mutually connected (connected compo- old of L = 1/T to identify the straight pair-loops that ij nents of the graph). Particles are grouped into a coherent do not grow rapidly over the time interval. Using this structure by greedy agglomeration of the connected ver- growth rate, we identify loops that grow by a factor of tices. We expect that, as Λ is increased from its lowest e over the time interval (see (8)), which is less than 1% 9 FIG. 9. (a) The initial positions of N = 300 analytic model trajectories (dots), colored based on their coherent structure assignment with unassigned particles in black. The lines FIG. 10. (a) The initial positions of N = 33 experimen- connecting particles indicate straight pair-loops that have a tal trajectories (dots), and identiﬁed coherent loops (black growth ratio less than the threshold, with less opaque lines lines). Three coherent structures are identiﬁed (red, blue and representing loops that grow more. (b) and (c) are zoomed in green). The straight pair-loops that have a growth rate of less regions corresponding to the gray dashed boxes in (a). The than the threshold are represented as straight lines connect- coherent structures corresponding to slowly growing loops for ing the particles. The less opaque the line, the greater the each set of points is drawn in the corresponding color. growth rate. (b) The ﬁnal position of the trajectories and the deformed coherent loops. of all the pair-loops. The network of particles connected by a slow-growing, that are a physical distance of 0.01 cm apart. Because of straight pair-loop is presented as lines in Figure 9(a). The their close proximity, these particles do not separate over particles that are connected by a network of these pair- the 50 periods of advection, despite being in the chaotic loops form a coherent set, where each set has a diﬀerent sea. coloring of both the particles and the network connec- tions. Finally, the coherent structure, which is repre- sented by the slow-growing loop that surrounds all the C. Coherent structures in the experimental system particles of a particular coherent set, can be identiﬁed. These resulting slow-growing loops are presented in Fig- ure 9(b) and (c). Having identiﬁed the main coherent structures in the model system, we next apply the method to experimen- Four diﬀerent types of coherent structures were iden- tally observed trajectories. The particle trajectories gen- tiﬁed by the method. The two structures in red and erated by the experimental system (see section II B 1 and blue represent each of the main recirculating vortices. table I) can be applied to the coherent structure method Over the time interval, the topological length of the left to identify the structures in the ﬂow ﬁeld. The ﬁrst 50 and the right recirculating vortex loops grow by a fac- periods of advection are considered, throughout which 33 tor of 1.24 and 1.48, respectively. Near the left vortex trajectories remained in the imaging domain for the dura- (blue), there are four additional structures identiﬁed at tion of the time interval. These 33 trajectories produce a its perimeter: these structures are plotted in magenta, braid made up of 9,592 generators. Using the same length purple, yellow and dark purple. Each of these features ratio threshold as the analytic model trajectory analysis, is comprised of two or three particles and in all cases, we identiﬁed three coherent sets of particles. Each of the the structures closely orbit the recirculating vortex with- sets is surrounded by a non-growing loop. Applying the out penetrating the boundary. Another type of coherent method to this dataset takes less than 10 seconds. structure identiﬁed is presented in light and dark green in Figure 9(a). These two structures correspond to par- The initial and ﬁnal position of the particles and the ticles that orbit around the chaotic sea and the recir- corresponding coherent structures are presented in Fig- culating vortices. The dark green structure grows by a ure 10. The two main recirculating vortices are identiﬁed. factor of 1.47 while the light green structure shrinks to The set of particles orbiting around the outside of the 0.6 times its initial length due to the particles becoming domain is also identiﬁed. For the experiment, particles closer in the topological projection. Finally, the fourth that pass through the chaotic sea only during the ﬁnal type of structure is identiﬁed in gray in Figure 9(a) near periods also belong to this outer structure because not (1.89,0.13) cm. This structure is made up of two particles enough mixing with the structure has occurred for some 10 of these particles to be excluded from the orbit structure. trast between particles and background ﬂow that is high For all three cases, the braided length of the topological enough for particles to be detected at each time step. The loop |B`| either shrinks or remains the same. The left tracking issue was mostly due to the wide range of veloc- and right recirculating vortex structures correspond to ities between particles located close to the rotor and par- ratios of 1 and 0.86, respectively, while the orbiting struc- ticles in the outer region, which made the construction of ture ratios 0.92, demonstrating that perceivably-coherent trajectories from their respective particles diﬃcult. The structures result in roughly-constant lengths of topolog- other main challenge to obtaining a suﬃcient spatiotem- ical loops. Having demonstrated the coherent structures poral coverage is that as the particle density increases approach to the PT trajectories, the analysis was reit- the particle-particle interactions become more important erated on datasets of synthetic PIV trajectories. Each and can cause the ﬂow to signiﬁcantly deviate from the recirculating vortex was detected with resolution depen- analytic model. Some of the features of interest present dent on the trajectory initialization. The Supplemental in the ﬂow being small-sized, it was therefore diﬃcult to Material Figure [88] presents the coherent structures de- seed the spatial domain to discern all of them. Nonethe- tected from two sample sets of synthetic trajectories, one less, the remaining trajectories were already suﬃcient to with 33 trajectories and one with 100 trajectories. succeed at detecting the largest coherent structures using the braid theory approach. D. Comparison V. ESTIMATING THE GROWTH RATE OF MATERIAL INTERFACES These results demonstrate the ability of the braid the- ory approach to detected coherent structures accurately Newhouse and Yomdin [68–70] established that, in pla- and robustly in an experimental, periodically forced ﬂow. nar ﬂows, the exponential growth rate of material inter- The method found similar qualitative results using the faces is equal to the topological entropy of the ﬂow. In experimental and analytic model trajectories. Despite turn, topological entropy of the ﬂow is bounded below the lower particle density in the experimental system, the by braid entropy, a measure of complexity of braids of coherent structures representing the vortex edges were any periodic trajectories in the ﬂow [55, 58]. Adding tra- detected, even with the structures containing as few as jectories to a braid can only increase the braid entropy, three particles, which shows the reliability of the method therefore a reasonable approximation strategy involves even for sparse datasets. The small-scale features de- computing braid entropies of braids comprising more and tected in the model ﬂow were not identiﬁed from the more trajectories in an eﬀort to narrow the gap between experimental data because too few particles were seeded the braid and topological entropy. Applying this strat- in those regions. egy to experimentally-recorded or simulated trajectories When comparing the braid-based coherent structures is diﬃcult to do directly, as most trajectories are not pe- to the Poincar´ e map for the model ﬂow there are simi- riodic. Nevertheless, Budiˇ si´ c and Thiﬀeault [66] demon- larities and diﬀerences. First, the main recirculating vor- strated that computing a so-called Finite-Time Braiding tices, represented in red and blue for both the model and Exponent (FTBE) provides a reasonable data-driven es- experimental examples, correspond to the islands identi- timate of braid entropy and consequently an estimate ﬁed in Figure 2(b). The second feature identiﬁed is an of the rate of exponential growth of material interfaces. outer structure corresponding to a set of particles that In this section, we apply this procedure to evaluate the are advected at very low velocities around the main vor- degree of correspondence between the experimental and tex system. This outer structure encompassed a much model rotor-oscillator ﬂows. higher proportion of particles for the experimental ﬂow than for the theoretical ﬂow. This may be due to the sparsity of the dataset: if fewer trajectories are present, A. Theory then they may be less likely to entangle with the chaotic sea. The most straightforward way to quantify the com- To improve the resolution of the braid-based coher- plexity of a braid is to count the number of crossings in ent structures, more particle trajectories are needed to the braid, while avoiding counting trivial crossings. For improve the spatiotemporal coverage. In laboratory ex- example, two braids with two strands each: periments, however, there are two main limitations to ac- complishing this. Due to imagery and tracking issues, the trajectories of several experimental particles could not be fully reconstructed for all periods. This is why only 33 out of 80 trajectories were retained for the analysis. The braid approach requires all trajectories to span the both have 3 crossings, since the two central crossings in same time interval and it needs a continuous dataset, i.e., the top braid can be disentangled even with the end- no missing datapoints within trajectories. The imagery points pinned. More rigorously, before counting crossings issue was mostly due to the challenge of having a con- the braid is reduced by employing braid group rules that 11 cancel out such intersections. Deﬁne braid length #B to be the number of intersections after the braid is reduced. Given a topological braid B that corresponds to N N,T trajectories collected over a time interval of duration T , the average braid growth rate of B is N,T #B N,T G(B ) = . (9) N,T FIG. 11. Top: Loop ` used to compute the Finite- The number of intersections alone is a poor indicator Time Braiding Exponent, with Dynnikov vector ` = of complexity. For example, the two 3-stranded braids 0 0 0 0 −1 −1 −1 −1 [ ]. Bottom: The loop after action of the braid from Figure 5, with Dynnikov vector B` = −1 0 8 1 0 −6 −3 7 [ ]. Black dots: strands belonging to the braid; right-most red dot: the auxiliary “anchor” strand. The length of a loop is the number of intersections with the dashed axis (|` | = 4,|B` | = 36). 0 0 have the same length, but the strands in the bottom braid are intertwined in a more intricate manner. Compared to the connection between topological en- The topological entropy of a braid, or simply braid en- tropy and the rate of material growth, the connection tropy, measures the complexity of a braid by quantifying between FTBEs and the rate of material growth is more how fast loops grow as they are slid along the braid (as heuristic. Nevertheless, numerical investigations in a in Figure 6). Braid entropy can be computed by ﬁnding chaotic ﬂow [66] shows that FTBEs can indeed be used the largest rate of growth of a topological loop under a as a reasonable proxy to the topological entropy of a repeated action of the braid [56, 89]. When the physi- braid when in the absence of periodic Lagrangian tra- cal braid comprises periodic trajectories, the entropy of jectories from the ﬂow. Consequently, FTBEs charac- the associated topological braid is a lower bound for the terizes the rate of growth of material lines in this work. topological entropy of the ﬂow. This property was suc- Braid growth G will be used as the independent variable cessfully exploited to create stirring protocols for rod- in comparing FTBE of braids with diﬀerent number of based stirrers [23, 55, 57, 90], to design spatially pe- strands. riodic mixers [91], and to analyze stirring of a passive tracer by compact coherent structures (so-called ghost rods ) [57, 58, 63, 92]. B. Comparison of the model and experimental When trajectories in the physical braid are not peri- ﬂows odic, the theory is less developed compared to the peri- odic case, and published works rely on numerical simu- lations [56, 65]. Nevertheless, when trajectory segments We compare the complexity of numerical trajectories are long enough, it is possible to deﬁne the Finite-Time in the model and experimental velocity ﬁelds. The pa- Braiding Exponent (FTBE) [66], which measures the ex- rameters of the PIV experiment in Table I correspond ponential growth rate of a particular topological loop to a nondimensional oscillation period τ = 5 and non- during a single application of a braid. (By compari- dimensional parameters = 0.1250 and λ = 1.2566. son, topological entropy measures asymptotic exponen- We initialized trajectories in the chaotic region where tial growth with respect to repeated application of the primary and secondary vortices meet (see Figure 4), as braid.) this is where exponential growth of material lines is ex- The Finite-Time Braiding Exponent is given by pected, and simulated them for 30 oscillation periods (T = 30τ = 150). As the PIV velocity ﬁeld was recorded 1 |B ` | N,T 0 FTBE(B ) = log , (10) over 11 oscillation periods (see II B 2), to achieve T = 150 N,T T |` | we repeated three times the ﬁrst 10 oscillation periods. where ` is the topological loop in Figure 11 (top). The Integrating PIV velocity ﬁeld for long times com- particular loop ` is chosen so that any nontrivial braid pounds measurement errors in the velocities, especially results in a nonzero deformation [66]. To ensure this, the in the slow regions of the ﬂow, resulting in unphysical components of the loop are “anchored” on the boundary trajectories, e.g., trajectories that exit the domain. We of the domain, which is topologically equivalent to adding removed such trajectories from the PIV dataset and re- an additional strand (depicted as the hollow circle) that placed them with additional simulated trajectories until does not participate in dynamics. Justiﬁcations for the we obtained a full set of N = 600 well-behaved trajecto- choice of ` is given at length in [93, 94]. ries. We collected the same number of trajectories from Compared to (8) in section IV, the FTBE is calculated the model velocity ﬁeld. Even though the model and PIV for the speciﬁc loop ` instead of a set of simple pair- velocity ﬁelds are qualitatively similar (see Figure 4), the loops. Further information on FTBE calculations can be precise location of the saddle-like point sitting between found in [66]. the primary and secondary vortices is sligthly diﬀerent. 12 −1 Model 0.15 PIV −0.5 0.1 0.15 0.2 0.25 0.1 0.5 0.05 −3 −2 −1 0 1 2 3 −6 −4 −2 0 2 4 6 (a) Samples of analytic model velocity ﬁeld trajectories # Std. dev. from mean (a) FTBE −1 −0.5 0.15 Model PIV 1.2 1.3 1.4 1.5 0.1 0.5 ×10 0.05 −3 −2 −1 0 1 2 3 (b) Samples of PIV velocity ﬁeld trajectories −6 −4 −2 0 2 4 6 FIG. 12. Poincar´ e plots of trajectories in braids used to com- # Std. dev. from mean pare FTBE and braid growth rate between analytic model and (b) Growth rate PIV velocity ﬁelds. Trajectories were initialized uniformly in- side the highlighted rectangles (a) (x, y) ∈ [−0.25,−0.5] × [−0.5, 0.5] for model velocities; (b) (x, y) ∈ [−0.1,−0.5] × FIG. 13. Standardized distributions of FTBEs and growth [−0.5, 0.5] for PIV velocities. The Poincar´ e map is taken at rates with S = 500 samples of n = 300-stranded sub- the rotor’s oscillation period τ . braids (out of N = 600 trajectories, T = 30τ = 150); non- standardized histograms are inset into each graph. To ensure that trajectories are seeded in the chaotic zone in both cases, we used a diﬀerent initialization rectangle shapes of distributions are similar, but with quantita- for the two velocity ﬁelds. Consequently, Poincar´ e plots tive diﬀerences. The model velocity ﬁeld shows a greater in Figure 12 show that both datasets have similar quali- variance in the growth rate, while FTBE distributions tative features in the region of interest. Finally, we con- are non-overlapping, indicating a signiﬁcant separation verted sets of trajectories into braids using the MATLAB of recorded FTBE values. The model velocity ﬁeld may allow initialization of rare-event trajectories more easily, toolbox braidlab [82]. e.g., in intricate periodic islands that would be washed The braids of trajectories are characterized by their out in experimental system, potentially explaining the braid growth rates (9) and FTBE values (10), which re- variance in growth. The distance between FTBE distri- spectively quantify how many non-trivial strand crossings butions appears to be more signiﬁcant, indicating that are generated, and the amount of stretching that these even though the number of strand intersections is com- crossings impose on topological material lines. We ana- parable, the intersections for the model ﬂow generate no- lyze ensembles of S = 500 braids, each with n strands, ticeably more complexity than the experiment. in order to estimate distributions of FTBE and braid growth. By randomly choosing n trajectories out of the Presently, we do not know of a trajectory mechanism full set of N = 600 simulated trajectories, for each num- that would explain this diﬀerence. We do not think that ber of strands n = 3, . . . , N/2 = 300, we create S realiza- the diﬀerence in initialization window (see Figure 12) tions of braids B . By reusing trajectories in several could cause this deviation, as trajectories in both cases n,T braids, fewer trajectories need to be simulated, avoiding occupy the chaotic region surrounding the vortices de- the costly initialization and resimulation of trajectories spite the slight numerical diﬀerences in locations of those regions. for individual realizations. Further discussion of the re- sampling step is in [66]. Additionally, artiﬁcially degrading the quality of the Figure 13 shows histograms of FTBEs and braid model does not change the outcome. We have sampled growth rates for both sets of braids with S = 500 realiza- the model velocity ﬁeld on the same grid as the PIV tions of the largest number of strands n = 300, subsam- velocity ﬁeld, and performed the same space-time inter- pled from the full simulated set of N . From a qualitative polation that was used for the PIV velocity ﬁeld. The comparison with Poincar´ e plots, one would expect simi- resulting interpolated model velocity ﬁeld had the same lar results for both model and PIV velocity ﬁelds. The FTBE values as the continuous model velocity ﬁeld, sug- Proportion Proportion 13 gesting that discretization and interpolation themselves FTBE and braid growth should increase with n. Fig- are not the cause for the lower FTBE values in the PIV ure 15 shows the observed dependence. The growth rate ﬂow. These results are not presented in ﬁgures, as they for both ﬂows increases quadratically with the number of do not diﬀer from the model ﬂow results. strands, which can be explained by noting that n strands Ultimately, the quantitative diﬀerence between FTBE can form = O(n ) pairs. In chaotic regions, statis- values may speak to the sensitivity of the braid dynamics tical quantities tend to mimic those in stochastic pro- techniques, in particular FTBE, to diﬀerences between cesses with independent increments, so over large peri- ﬂows that are otherwise challenging to determine. ods of time any of those possible crossings should In order to exhibit the dependence of distributions of be equally likely. As more pairs are conﬁned in a ﬁnite FTBEs and growth rates on duration T and number of space, we observe a quadratic increase in the number of strands n for braids B , we compute summary statis- crossings for a ﬁxed time interval. n,T tics and graph them against T and n. FTBE distribu- FTBE increases with n in both velocity ﬁelds, follow- tions are represented by their maximum values, as braid ing the shape previously observed in [66] for which there entropy of any braid is a lower bound for the topological is no good mathematical model, as far as we are aware. entropy of the ﬂow. Growth rate distributions are sim- Topological entropy of the ﬂow acts as the upper bound ply represented by mean values, as there are no special for the curve, but it is not clear if the bound is tight. In considerations for this quantity. planar ﬂows, topological entropy corresponds to the max- The maximum FTBE in a sample should provide imal rate of exponential growth of material lines. Since the best-available bound to topological entropy for the FTBE is non-zero, we infer that both ﬂows exhibit suﬃciently-long duration T [66]. Figure 14 demonstrates exponential growth of material interfaces, although the that the values of maximum FTBE and mean growth sta- FTBE estimates of these rates diﬀer quantitatively. We bilize well ahead the end of simulations, which leads us do not know if the diﬀerence is dynamically signiﬁcant. to conclude that studying trajectories of length T = 30τ Future work could contrast FTBE results to those ob- is representative of values in the long-time limit. tained by the recently-developed eTEC technique [64] as an independent comparison. Model (sample max., 300 str.) 0.4 Model (single braid, 600 str.) PIV (sample max., 300 str.) PIV (single braid, 600 str.) 0.2 0.2 0.1 Model PIV 50 100 150 Trajectory duration T 0 200 400 600 Number of strands n (a) Max. FTBE (a) FTBE of sub-braids (T = 150) 1,500 Model PIV 1,000 Model PIV −1 50 100 150 1 2 3 10 10 10 Trajectory duration T Number of strands n (b) Mean growth rate (b) Growth rate of sub-braids (T = 150 ) FIG. 14. Dependence of the maximum FTBE and mean FIG. 15. Eﬀect of increase in braid size (number of strands) growth rate on trajectory duration T . The comparison is be- on FTBE and growth rate of braids of trajectories from model tween the maximum FTBE over a sample of n = 300-stranded and PIV ﬂows. Filled-in marks correspond to maximum of a braids and a single braid of n = 600 strands, all from analytic sample of n-stranded braids. Hollow marks correspond to the model and PIV ﬂows. single 600-stranded braid. Intuitively, the more strands participate in a braid, the higher the potential for complexity. Consequently Finally, we point out some challenges encountered in chaotic regions, where we initialized trajectories, both here. An eﬀective comparison of complexity based on Mean G(B ) n,T max FTBE(B ) n,T G (B ) n N,T max FTBE(B ) N,T 14 braid dynamics requires Lagrangian trajectories that • At least a few particles should be seeded in each cover many periods of oscillations, which is experimen- structure to be detected; tally challenging for several reasons. As explained in Sec- • Quantitative comparison of FTBE values is unreli- tion II B 2, the analytical model assumes that the rotor able in regions without signiﬁcant mixing; is a point vortex and that the Reynolds number is zero. • Reliable calculation of FTBEs requires either long Additionally, for the FTBE to be computed, the motors Lagrangian trajectories, or measured velocity ﬁelds of the rotor and the oscillator must run for many hours, that can be used to generate such trajectories. which inevitably introduces some amount of thermal en- As explained in Sections IV D and V B, we have had ergy into the ﬂow and slightly changes its viscosity over to overcome several diﬃculties in order to obtain a set of time. For these reasons, we only used 10 periods of exper- experimental particle tracks that were both suﬃciently imental data to reconstruct the ﬂow, which required arti- dense and gapless. Other experimental setups may face ﬁcially repeating the velocity ﬁeld to obtain Lagrangian similar challenges, which currently limits the range of ap- trajectories over 30 periods. The chosen periods were plication of methods based on braid dynamics. Nonethe- the ﬁrst 10 full periods after the ﬁrst peak of oscillation, less, in experimental conditions that do satisfy the re- discarding the transient dynamics after the initial accel- quirements, braid dynamics techniques can be competi- eration. tive with alternative techniques for identifying coherent structures or quantifying the amount of mixing in a two- dimensional ﬂow. We do point out that the recently- VI. DISCUSSION developed eTEC technique [64] shows promise in over- coming at least some of the obstacles mentioned here. We highlight three directions for future work to in- We demonstrated that methods based on braid dynam- crease the usefulness of braid dynamics techniques for ics can identify qualitative features of material transport experimental data. First, many more existing (gappy) in model and experimental rotor-oscillator ﬂows. In par- datasets could be processed if there was a technique ticular, Section IV shows that braid dynamics can distin- for ﬁlling-in missing segments in Lagrangian trajecto- guish between nearby transport structures—secondary ries that do not introduce signiﬁcant additional entropy. vortices and a recirculating region—even when only a few Second, fewer strands would be needed if there existed Lagrangian trajectories are seeded in them, with struc- a theoretical model for growth of FTBEs with number tures being detected from as few as two trajectories. For of strands. Finally, the eﬃciency of search for coherent quantities measuring material deformation (FTBE and structures could be improved by considering the geom- growth rate), we demonstrated in Section V that our two etry of the space of topological loops. Some of these qualitatively-similar ﬂows result in similar distributions, directions may be amenable to modern machine learning with parallel trends in their time-evolution. techniques. Based on these results, we expect that the braid dy- namics methods will be useful in situations where only a small number of Lagrangian trajectories from a two- ACKNOWLEDGMENTS dimensional ﬂow is available, either due to experimental challenges, or because only a rough estimate of ﬂow prop- erties is desired. We thank Gustaaf (Guus) Jacobs for his help running During the comparison of model and experimental CFD models for preliminary analysis of the experimental data, it became clear that processed data needs to sat- setup, Nicholas Ouellette and Douglas Kelley for their isfy, at a minimum, the following requirements in order assistance with processing PIV data and Greg Voth for to be successfully processed using computational braid a helpful discussion about the comparison between the dynamics methods: PIV data and the model. This research was supported • There cannot be signiﬁcant gaps in the discrete by the U.S. National Science Foundation (NSF), under time series of particle tracks; Grants No. CMMI-1233935 and AGS-1520825. [1] Julio M. Ottino, The Kinematics of Mixing: Stretch- ical study of transport, mixing and chaos in an unsteady ing, Chaos, and Transport, Cambridge Texts in Applied vortical ﬂow,” Journal of Fluid Mechanics 214, 347–394 Mathematics (Cambridge University Press, Cambridge, (1990). 1989). [5] Rob Sturman, Julio M. Ottino, and Stephen Wiggins, [2] Hassan Aref, “Stirring by chaotic advection,” Journal of The Mathematical Foundations of Mixing, Cambridge Fluid Mechanics 143, 1–21 (1984). Monographs on Applied and Computational Mathemat- [3] R S MacKay, J D Meiss, and I C Percival, “Transport in ics, Vol. 22 (Cambridge University Press, Cambridge, Hamiltonian systems,” Physica D. Nonlinear Phenomena 2006). 13, 55–81 (1984). [6] S. Childress and A.D. Gilbert, Stretch, Twist, Fold: The [4] V. Rom-Kedar, A. Leonard, and S. Wiggins, “An analyt- Fast Dynamo, Lecture Notes in Physics Monographs 15 (Springer Berlin Heidelberg, 1995). tion of rod-stirring devices,” SIAM Review 53, 723–743 [7] J. D. Meiss, “Thirty years of turnstiles and transport,” (2011). Chaos: An Interdisciplinary Journal of Nonlinear Science [24] S. Wiggins and J. M. Ottino, “Foundations of chaotic 25, 097602 (2015). mixing,” Philosophical Transactions of the Royal Society [8] Sanjeeva Balasuriya, Barriers and Transport in Unsteady of London Series a-Mathematical Physical and Engineer- Flows: A Melnikov Approach (SIAM, 2016). ing Sciences 362, 937–970 (2004). [9] Erik M. Bollt and Naratip Santitissadeekorn, Applied [25] Emmanuel Villermaux, “Mixing Versus Stirring,” Annual and Computational Measurable Dynamics, Mathemati- Review of Fluid Mechanics 51, 245–273 (2019). cal Modeling and Computation, Vol. 18 (Society for In- [26] Jean-Luc Thiﬀeault and Allen H. Boozer, “Geometrical dustrial and Applied Mathematics (SIAM), Philadelphia, constraints on ﬁnite-time Lyapunov exponents in two and PA, 2013). three dimensions,” Chaos: An Interdisciplinary Journal [10] A. Leonard, V. Rom-Kedar, and S. Wiggins, “Fluid mix- of Nonlinear Science 11, 16–28 (2001). ing and dynamical systems,” Nuclear Physics B - Pro- [27] Jean-Luc Thiﬀeault, “Stretching and curvature of mate- ceedings Supplements 2, 179–190 (1987). rial lines in chaotic ﬂows,” Physica D. Nonlinear Phe- [11] Hassan Aref, John R. Blake, Marko Budiˇ si´ c, Silvana S.S. nomena 198, 169–181 (2004). Cardoso, Julyan H.E. Cartwright, Herman J.H. Clercx, [28] George Mathew, Igor Mezi´ c, and Linda Petzold, “A mul- Kamal El Omari, Ulrike Feudel, Ramin Golestanian, Em- tiscale measure for mixing,” Physica D. Nonlinear Phe- manuelle Gouillart, GertJan F. van Heijst, Tatyana S. nomena 211, 23–46 (2005). Krasnopolskaya, Yves Le Guer, Robert S. MacKay, [29] J.-L. Thiﬀeault, “Scalar decay in chaotic mixing,” Trans- Vyacheslav V. Meleshko, Guy Metcalfe, Igor Mezi´ c, port and Mixing in Geophysical Flows Lecture Notes in Alessandro P.S. de Moura, Oreste Piro, Michel F. M. Physics, 3–36 (2008). Speetjens, Rob Sturman, Jean-Luc Thiﬀeault, and Idan [30] J.-L. Thiﬀeault, “Using multiscale norms to quantify Tuval, “Frontiers of chaotic advection,” Reviews of Mod- mixing and transport,” Nonlinearity 25, R1–R44 (2012), ern Physics 89, 025007 (2017). arXiv:1105.1101. [12] Melissa A. Green, Clarence W. Rowley, and Alexander J. [31] George Haller, “Lagrangian Coherent Structures,” An- Smits, “Using hyperbolic Lagrangian coherent structures nual Review of Fluid Mechanics 47, 137–162 (2015). to investigate vortices in bioinspired ﬂuid ﬂows,” Chaos: [32] Gary Froyland and Kathrin Padberg, “Almost-invariant An Interdisciplinary Journal of Nonlinear Science 20, sets and invariant manifolds—connecting probabilistic 017510 (2010). and geometric descriptions of coherent structures in [13] Shawn C. Shadden, John O. Dabiri, and Jerrold E. Mars- ﬂows,” Physica D. Nonlinear Phenomena 238, 1507–1523 den, “Lagrangian analysis of ﬂuid transport in empirical (2009). vortex ring ﬂows,” Physics of Fluids 18, 047105 (2006). [33] Thomas Peacock and George Haller, “Lagrangian co- [14] Kakani Katija and John O. Dabiri, “A viscosity-enhanced herent structures: The hidden skeleton of ﬂuid ﬂows,” mechanism for biogenic ocean mixing,” Nature 460, 624– Physics Today 66, 41–47 (2013). 626 (2009). [34] Roger M. Samelson, “Lagrangian Motion, Coherent [15] Amanda J. Tan, Eric Roberts, Kevin A. Mitchell, and Structures, and Lines of Persistent Material Strain,” An- Linda S. Hirst, “Investigating Quality of Mixing of a Bi- nual Review of Marine Science 5, 137–163 (2013). ological Active Nematic,” Biophysical Journal 114, 650a [35] Michael R. Allshouse and Thomas Peacock, “Lagrangian (2018). based methods for coherent structure detection,” Chaos: [16] Jean-Luc Thiﬀeault, “Chaos in the Gulf,” Science Mag- An Interdisciplinary Journal of Nonlinear Science 25, azine (2010). 097617 (2015). [17] Igor Mezi´ c, Sophie Loire, Vladimir A Fonoberov, and [36] Alireza Hadjighasem, Mohammad Farazmand, Daniel Patrick J Hogan, “A New Mixing Diagnostic and Gulf Oil Blazevski, Gary Froyland, and George Haller, “A critical Spill Movement,” Science Magazine 330, 486–489 (2010). comparison of Lagrangian methods for coherent structure [18] Mar´ ıa J. Olascoaga and George Haller, “Forecasting sud- detection,” Chaos: An Interdisciplinary Journal of Non- den changes in environmental pollution patterns,” Pro- linear Science 27, 053104 (2017). ceedings of the National Academy of Sciences 109, 4738– [37] Sanjeeva Balasuriya, Nicholas T. Ouellette, and Irina I. 4743 (2012). Rypina, “Generalized Lagrangian coherent structures,” [19] M. Weldon, T. Peacock, G. B. Jacobs, M. Helu, and Physica D: Nonlinear Phenomena 372, 31–51 (2018). G. Haller, “Experimental and numerical investigation of [38] Michael R. Allshouse, Gregory N. Ivey, Ryan J. Lowe, the kinematic theory of unsteady separation,” Journal of Nicole L. Jones, C. J. Beegle-Krause, Jiangtao Xu, and Fluid Mechanics 611, 1–11 (2008). Thomas Peacock, “Impact of windage on ocean surface [20] Yong Wang, George Haller, Andrzej Banaszuk, and Lagrangian coherent structures,” Environmental Fluid Gilead Tadmor, “Closed-loop Lagrangian separation con- Mechanics 17, 473–483 (2017). trol in a bluﬀ body shear ﬂow model,” Physics of Fluids [39] Francisco J. Beron-Vera, Mar´ ıa J. Olascoaga, George 15, 2251–2266 (2003). Haller, Mohammad Farazmand, Joaqu´ ın Trinane ˜ s, and [21] B Noack, Igor Mezi´ c, G Tadmor, and Andrzej Banaszuk, Yan Wang, “Dissipative inertial transport patterns near “Optimal mixing in recirculation zones,” Physics of Flu- coherent Lagrangian eddies in the ocean,” Chaos: An In- ids (2004). terdisciplinary Journal of Nonlinear Science 25, 087412 [22] George Mathew, Igor Mezi´ c, Symeon Grivopoulos, (2015). Umesh Vaidya, and Linda Petzold, “Optimal control of [40] P. Tallapragada, S. D. Ross, and D. G. Schmale, “La- mixing in Stokes ﬂuid ﬂows,” Journal of Fluid Mechanics grangian coherent structures are associated with ﬂuctua- (2007). tions in airborne microbial populations,” Chaos: An In- [23] M. D. Finn and J.-L. Thiﬀeault, “Topological optimiza- terdisciplinary Journal of Nonlinear Science 21, 033122 16 (2011). “Topological ﬂuid mechanics of stirring,” Journal of Fluid [41] Amir E. BozorgMagham, Shane D. Ross, and David G. Mechanics 403, 277–304 (2000). Schmale III, “Real-time prediction of atmospheric La- [56] Jean-Luc Thiﬀeault, “Measuring Topological Chaos,” grangian coherent structures based on forecast data: An Physical Review Letters 94, 084502 (2005). application and error analysis,” Physica D: Nonlinear [57] Jean-Luc Thiﬀeault and Matthew D. Finn, “Topology, Phenomena 258, 47–60 (2013). braids and mixing in ﬂuids,” Philosophical Transactions [42] George Haller and Themistoklis Sapsis, “Lagrangian co- of the Royal Society A: Mathematical, Physical and En- herent structures and the smallest ﬁnite-time Lyapunov gineering Sciences 364, 3251–3266 (2006). exponent,” Chaos: An Interdisciplinary Journal of Non- [58] Emmanuelle Gouillart, Jean-Luc Thiﬀeault, and linear Science 21, 023115 (2011). Matthew D. Finn, “Topological mixing with ghost rods,” [43] Marko Budiˇ si´ c, Stefan Siegmund, Doan Thai Son, and Physical Review E. Statistical, Nonlinear, and Soft Mat- Igor Mezi´ c, “Mesochronic classiﬁcation of trajectories in ter Physics 73, 036311, 8 (2006). incompressible 3D vector ﬁelds over ﬁnite times,” Dis- [59] Matthew D Finn and Jean-Luc Thiﬀeault, “Topological crete and Continuous Dynamical Systems - Series S 9, entropy of braids on the torus,” SIAM Journal on Applied 923–958 (2016). Dynamical Systems 6, 79–98 (electronic) (2007). [44] Tian Ma and Erik M. Bollt, “Shape Coherence and [60] Nicolas Francois, Hua Xia, Horst Punzmann, Benjamin Finite-Time Curvature Evolution,” International Journal Faber, and Michael Shats, “Braid Entropy of Two- of Bifurcation and Chaos 25, 1550076 (2015). Dimensional Turbulence,” Scientiﬁc Reports 5, 18564 [45] Gary Froyland and Kathrin Padberg-Gehle, “Finite-time (2015). entropy: A probabilistic approach for measuring nonlin- [61] S. Candelaresi, D. I. Pontin, and G. Hornig, “Quantify- ear stretching,” Physica D: Nonlinear Phenomena 241, ing the tangling of trajectories using the topological en- 1612–1628 (2012). tropy,” Chaos: An Interdisciplinary Journal of Nonlinear [46] Dean Roemmich, Gregory C. Johnson, Stephen Riser, Science 27, 093102 (2017). Russ Davis, John Gilson, W. Brechner Owens, Silvia L. [62] C. K. Taylor and Stefan G. Llewellyn Smith, “Dy- Garzoli, Claudia Schmid, and Mark Ignaszewski, “The namics and transport properties of three surface quasi- Argo Program: Observing the Global Ocean with Proﬁl- geostrophic point vortices,” Chaos: An Interdisciplinary ing Floats,” Oceanography 22, 34–43 (2009). Journal of Nonlinear Science 26, 113117 (2016). [47] Ma¨ elle Nodet, “Variational assimilation of Lagrangian [63] Piyush Grover, Shane D Ross, Mark A Stremler, and data in oceanography,” Inverse Problems 22, 245–263 Pankaj Kumar, “Topological chaos, braiding and bi- (2006). furcation of almost-cyclic sets,” Chaos: An Interdisci- [48] Emanuel F. Coelho, P. Hogan, G. Jacobs, P. Thoppil, plinary Journal of Nonlinear Science 22, 043135–043135– H. S. Huntley, B. K. Haus, B. L. Lipphardt, A. D. Kir- 16 (2012). wan, E. H. Ryan, J. Olascoaga, F. Beron-Vera, A. C. [64] Eric Roberts, Suzanne Sindi, Spencer A. Smith, and Poje, A. Griﬀa, T. M. Ozg¨ okmen, A. J. Mariano, G. Nov- Kevin A. Mitchell, “Ensemble-based topological entropy elli, A. C. Haza, D. Bogucki, S. S. Chen, M. Curcic, calculation (E-tec),” Chaos: An Interdisciplinary Journal M. Iskandarani, F. Judt, N. Laxague, A. J. H. M. Re- of Nonlinear Science 29, 013124 (2019). niers, A. Valle-Levinson, and M. Wei, “Ocean current [65] Jean-Luc Thiﬀeault, “Braids of entangled particle trajec- estimation using a Multi-Model Ensemble Kalman Fil- tories,” Chaos: An Interdisciplinary Journal of Nonlinear ter during the Grand Lagrangian Deployment experiment Science 20, 017516–017514 (2010). (GLAD),” Ocean Modelling 87, 86–106 (2015). [66] Marko Budiˇ si´ c and Jean-Luc Thiﬀeault, “Finite-time [49] Michael Allshouse and Jean-Luc Thiﬀeault, “Detecting braiding exponents,” Chaos: An Interdisciplinary Jour- coherent structures using braids,” Physica D. Nonlinear nal of Nonlinear Science 25, 087407 (2015). Phenomena , 95–105 (2012). [67] Philip Boyland, “Topological methods in surface dynam- [50] Gary Froyland and Kathrin Padberg-Gehle, “A rough- ics,” Topology and its Applications 58, 223–298 (1994). and-ready cluster-based approach for extracting ﬁnite- [68] Sheldon Newhouse and Thea Pignataro, “On the estima- time coherent sets from sparse and incomplete trajectory tion of topological entropy,” Journal of Statistical Physics data,” Chaos: An Interdisciplinary Journal of Nonlinear 72, 1331–1351 (1993). Science 25, 087406 (2015). [69] Sheldon E. Newhouse, “Entropy and volume,” Ergodic [51] Alireza Hadjighasem, Daniel Karrasch, Hiroshi Ter- Theory and Dynamical Systems 8, 283–299 (1988). amoto, and George Haller, “Spectral-clustering approach [70] Y. Yomdin, “Volume growth and entropy,” Israel Journal to Lagrangian vortex detection,” Physical Review E 93, of Mathematics 57, 285–300 (1987). 063107 (2016). [71] Ditza Auerbach, Predrag Cvitanovi´ c, Jean-Pierre Eck- [52] K. L. Schlueter-Kuck and J. O. Dabiri, “Coherent struc- mann, Gemunu Gunaratne, and Itamar Procaccia, “Ex- ture colouring: identiﬁcation of coherent structures from ploring chaotic motion through periodic orbits,” Physical sparse data using graph theory,” Journal of Fluid Me- Review Letters 58, 2387–2389 (1987). chanics 811, 468–486 (2017). [72] Ruslan L. Davidchack, Ying-Cheng Lai, Aaron Kle- [53] I. I. Rypina and L. J. Pratt, “Trajectory encounter vol- banoﬀ, and Erik M. Bollt, “Towards complete detection ume as a diagnostic of mixing potential in ﬂuid ﬂows,” of unstable periodic orbits in chaotic systems,” Physics Nonlin. Processes Geophys. 24, 189–202 (2017). Letters A 287, 99–104 (2001). [54] Matthew O. Williams, Irina I. Rypina, and Clarence W. [73] Ruslan L. Davidchack, Ying-Cheng Lai, Erik M. Bollt, Rowley, “Identifying ﬁnite-time coherent sets from lim- and Mukeshwar Dhamala, “Estimating generating par- ited quantities of Lagrangian data,” Chaos: An Interdis- titions of chaotic systems by unstable periodic orbits,” ciplinary Journal of Nonlinear Science 25, 087408 (2015). Physical Review E 61, 1353–1356 (2000). [55] Philip L Boyland, Hassan Aref, and Mark A Stremler, [74] Ruslan L. Davidchack and Ying-Cheng Lai, “Eﬃcient al- 17 gorithm for detecting unstable periodic orbits in chaotic 114101 (2011). systems,” Physical Review E 60, 6172–6175 (1999). [93] Ivan Dynnikov and Bert Wiest, “On the complexity of [75] Gary Froyland, Oliver Junge, and Gunter Ochs, “A braids,” Journal of the European Mathematical Society rough-and-ready cluster-based approach for extracting (JEMS) 9, 801–840 (2007). ﬁnit-time coherent sets from sparse and incomplete tra- [94] Patrick Dehornoy, “Eﬃcient solutions to the braid iso- jectory data,” Chaos 25, 68–84 (2001). topy problem,” Discrete Applied Mathematics Appli- [76] Erik M. Bollt, Theodore Stanford, Ying-Cheng Lai, and cations of Algebra to Cryptography, 156, 3091–3112 (2008). Karol Zyczkowski, “What symbolic dynamics do we get with a misplaced partition?: On the validity of thresh- old crossings analysis of chaotic time-series,” Physica D: Nonlinear Phenomena 154, 259–286 (2001). [77] W. W. Hackborn, M. E. Ulucakli, and T. Yuster, “A theoretical and experimental study of hyperbolic and de- generate mixing regions in a chaotic Stokes ﬂow,” Journal of Fluid Mechanics 346, 23–48 (1997). [78] W. W. Hackborn, “Asymmetric Stokes ﬂow between par- allel planes due to a rotlet,” Journal of Fluid Mechanics 218, 531–546 (1990). [79] Daniel Blair and Eric Dufresne, “The matlab parti- cle tracking code repository,” http://site.physics. georgetown.edu/matlab/ (1999). [80] Douglas H. Kelley and Nicholas T. Ouellette, “Onset of three-dimensionality in electromagnetically driven thin- layer ﬂows,” Physics of Fluids 23, 045103 (2011). [81] Lawrence F. Shampine, Mark W. Reichelt, and Jacek A. Kierzenka, “Solving index-i daes in matlab and simulink,” SIAM Review 41, 538–552 (1999). [82] Jean-Luc Thiﬀeault and Marko Budiˇ si´ c, “Braidlab: A Software Package for Braids and Loops,” arXiv:1410.0849 [math] (2014), version 3.2.2, arXiv:1410.0849 [math]. [83] I. A. Dynnikov, “On a Yang-Baxter map and the De- hornoy ordering,” Russian Mathematical Surveys 57, 592 (2002). [84] Toby Hall and S. Oyku ¨ Yurtta¸ s, “On the topological en- tropy of families of braids,” Topology and its Applica- tions 156, 1554–1564 (2009). [85] Z.P. You, E. J. Kostelich, and J. A. Yorke, “Calculating stable and unstable manifolds,” International Journal of Bifurcation and Chaos 1, 605–623 (1991). [86] George Haller, “Distinguished material surfaces and co- herent structures in three-dimensional ﬂuid ﬂows,” Phys- ica D: Nonlinear Phenomena 149, 248–277 (2001). [87] Gary Froyland, Simon Lloyd, and Naratip Santitis- sadeekorn, “Coherent sets for nonautonomous dynamical systems,” Physica D: Nonlinear Phenomena 239, 1527– 1541 (2010). [88] See Supplemental Material at [URL will be inserted by publisher] for the Figure presenting the coherent struc- tures detected from two sample sets of synthetic trajec- tories. [89] Jacques-Olivier Moussaﬁr, “On computing the entropy of braids,” Functional Analysis and Other Mathematics 1, 37–46 (2006). [90] Jean-Luc Thiﬀeault, “The mathematics of taﬀy pullers,” Mathematical Intelligencer 40, 26–35 (2018), https://arxiv.org/abs/1608.00152. [91] Matthew D. Finn, Jean-Luc Thiﬀeault, and Emmanuelle Gouillart, “Topological chaos in spatially periodic mix- ers,” Physica D: Nonlinear Phenomena 221, 92–100 (2006). [92] Mark A Stremler, Shane D Ross, Piyush Grover, and Pankaj Kumar, “Topological Chaos and Periodic Braid- ing of Almost-Cyclic Sets,” Physical Review Letters 106, Supplementary material to “Using braids to quantify interface growth and coherence in a rotor-oscillator ﬂow” Margaux Filippi, Marko Budiˇ si´ c, Michael R. Allshouse, S´ everine Atis, Jean-Luc Thiﬀeault, and Thomas Peacock (Dated: May 26, 2020) The particle tracking experiment directly provided the set of trajectories analyzed to ﬁnd coherent structures. It is also possible to detect coherent structures from Par- ticle Image Velocimetry (PIV) by generating a synthetic set of trajectories from the measured velocity ﬁeld. Using the same 10 periods of data used to create the synthetic trajectories for the FTBE analysis, two sets of 33 and 100 trajectories are generated. To prepare the initial conidtions, particles are uniformly seeded in −1.5 < x < 1.5, −0.5 < y < 0.2, then advected for 5 initial periods to minimize the inﬂuence of the initial seeding. After such initialization, the trajectories are recorded for 50 periods of wall-oscillation and converted to braids, which are then used in detection of coherent structures as explained in Sec. IV of the paper. The resulting coherent structures and the network of non- growing pair loops are presented in Figure 1. The smaller data set produces a similar result to the one found using the particle tracking trajectories where each vortex contains only two particles. The larger dataset include many particles in each of the recircu- lating vortex. The perceived increase in the vortex foot- FIG. 1. Coherent structures analysis on datasets of synthetic print demonstrates the importance of highly sampling PIV trajectories. Each recirculating vortex was detected with the coherent structure to get a better approximation for resolution dependent on the trajectory initialization. (a) The the feature. analysis of 33 trajectories detected a blue coherent structure within the left recirculating vortex and a red coherent struc- ture within the right recirculating vortex, each structure con- taining 2 trajectories. (b) The analysis of 100 trajectories detected a blue coherent structure on the left with 13 trajec- tories, a red coherent structure on the right with 5 trajectories and a green coherent structure within the right vortex with 2 trajectories. margaux@mit.edu arXiv:1910.12779v2 [physics.flu-dyn] 23 May 2020 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Nonlinear Sciences arXiv (Cornell University) http://www.deepdyve.com/lp/arxiv-cornell-university/using-braids-to-quantify-interface-growth-and-coherence-in-a-rotor-hHqpuKSN0B

Loading next page...

References (129)

A. BozorgMagham, S. Ross, D. Schmale (2013)
Real-time prediction of atmospheric Lagrangian coherent structures based on forecast data: An application and error analysis
Physica D: Nonlinear Phenomena, 258
朱克勤 (2006)
Journal of Fluid Mechanics创刊50周年
G. Haller, T. Sapsis (2011)
Lagrangian coherent structures and the smallest finite-time Lyapunov exponent.
Chaos, 21 2
G. Haller (2001)
Distinguished material surfaces and coherent structures in three-dimensional fluid flows
Physica D: Nonlinear Phenomena, 149
Eric Roberts, Suzanne Sindi, Spencer Smith, K. Mitchell (2018)
Ensemble-based topological entropy calculation (E-tec).
Chaos, 29 1
H. Aref (1984)
Stirring by chaotic advection
Journal of Fluid Mechanics, 143
Michael Allshouse, Jean-Luc Thiffeault (2011)
Detecting coherent structures using braids
Physica D: Nonlinear Phenomena, 241
Marko Budivsi'c, S. Siegmund, Doan Son, Igor Mezi'c (2015)
Mesochronic classification of trajectories in incompressible 3D vector fields over finite times
arXiv: Chaotic Dynamics
G. Haller (2015)
Lagrangian Coherent Structures
Annual Review of Fluid Mechanics, 47
M. Finn, Jean-Luc Thiffeault (2010)
Topological Optimization of Rod-Stirring Devices
SIAM Rev., 53
(2015)
Ocean Modelling 87
(2002)
Russian Mathematical Surveys 57
D. Roemmich, G. Johnson, S. Riser, R. Davis, J. Gilson, W. Owens, S. Garzoli, C. Schmid, M. Ignaszewski (2009)
The Argo Program : observing the global ocean with profiling floats
Oceanography, 22
P. Grover, S. Ross, M. Stremler, Pankaj Kumar (2012)
Topological chaos, braiding and bifurcation of almost-cyclic sets.
Chaos, 22 4
Ditza Auerbach, P. Cvitanovic, J. Eckmann, G. Gunaratne, I. Procaccia (1987)
Exploring chaotic motion through periodic orbits.
Physical review letters, 58 23
G. Froyland, O. Junge, G. Ochs (2001)
Rigorous computation of topological entropy with respect to a finite partition
Physica D: Nonlinear Phenomena, 154
Hannaneh Najdataei, Y. Nikolakopoulos, M. Papatriantafilou, P. Tsigas, Vincenzo Gulisano (2019)
STRETCH
Proceedings of the 13th ACM International Conference on Distributed and Event-based Systems
(1989)
The Kinematics of Mixing: Stretching
E. Villermaux (2019)
Mixing Versus Stirring
Annual Review of Fluid Mechanics
(2015)
International Journal of Bifurcation and Chaos 25
G. Froyland, Kathrin Padberg-Gehle (2015)
A rough-and-ready cluster-based approach for extracting finite-time coherent sets from sparse and incomplete trajectory data.
Chaos, 25 8
R. Davidchack, Y. Lai, Aaron Klebanoff, E. Bollt (2001)
Towards complete detection of unstable periodic orbits in chaotic systems
Physics Letters A, 287
S. Candelaresi, D. Pontin, Gunnar Hornig (2017)
Quantifying the tangling of trajectories using the topological entropy.
Chaos, 27 9
M. Budišić, Jean-Luc Thiffeault (2015)
Finite-time braiding exponents.
Chaos, 25 8
Alexander Rovira, Nicolas Müller, Weiwen Deng, Chudi Ndubaku, Richmond Sarpong (2019)
Bio-inspired synthesis of xishacorenes A, B, and C, and a new congener from fuscol† †Electronic supplementary information (ESI) available. See DOI: 10.1039/c9sc02572c
Chemical Science, 10
F. Beron-Vera, M. Olascoaga, G. Haller, M. Farazmand, J. Triñanes, Yan Wang (2014)
Dissipative inertial transport patterns near coherent Lagrangian eddies in the ocean.
Chaos, 25 8
G. Mathew, I. Mezić, L. Petzold (2005)
A multiscale measure for mixing
Physica D: Nonlinear Phenomena, 211
G. Froyland, Simon Lloyd, N. Santitissadeekorn (2009)
Coherent sets for nonautonomous dynamical systems
Physica D: Nonlinear Phenomena, 239
J. Moussafir (2006)
On computing the entropy of braids
Functional Analysis and Other Mathematics, 1
S. Ana, P.Elav arasi (2018)
Topology
International Journal of Mathematics Trends and Technology
R. Samelson (2013)
Lagrangian motion, coherent structures, and lines of persistent material strain.
Annual review of marine science, 5
Melissa Green, C. Rowley, A. Smits (2010)
Using hyperbolic Lagrangian coherent structures to investigate vortices in bioinspired fluid flows.
Chaos, 20 1
E. Coelho, P. Hogan, G. Jacobs, P. Thoppil, H. Huntley, B. Haus, B. Lipphardt, A. Kirwan, E. Ryan, J. Olascoaga, F. Beron-Vera, A. Poje, A. Griffa, T. Özgökmen, A. Mariano, G. Novelli, A. Haza, D. Bogucki, Shuyi Chen, M. Curcic, M. Iskandarani, F. Judt, N. Laxague, A. Reniers, A. Valle‐Levinson, M. Wei (2015)
Ocean Current Estimation Using a Multi-Model Ensemble Kalman Filter During the Grand Lagrangian Deployment Experiment (GLAD)
Ocean Modelling, 87
I. Dynnikov, B. Wiest (2004)
On the complexity of braids
Journal of the European Mathematical Society, 9
S. Barnett, Mohamed Babiker, Miles Padgett (2017)
Philosophical Transactions of the Royal Society A : Mathematical
(2018)
Biophysical Journal 114
Phanindra Tallapragada, S. Ross, D. Schmale (2011)
Lagrangian coherent structures are associated with fluctuations in airborne microbial populations.
Chaos, 21 3
Kristy Schlueter-Kuck, J. Dabiri (2016)
Coherent structure colouring: identification of coherent structures from sparse data using graph theory
Journal of Fluid Mechanics, 811
V. I, B. Siegel, L. Rosen, C. Giumarro
PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES.
Science, 44 1129
(2018)
Mathematical Intelligencer 40
Jean-Luc Thiffeault (2009)
Braids of entangled particle trajectories.
Chaos, 20 1
Alireza Hadjighasem, D. Karrasch, H. Teramoto, G. Haller (2015)
Spectral-clustering approach to Lagrangian vortex detection.
Physical review. E, 93 6
V. Rom-Kedar, A. Leonard, S. Wiggins (1990)
An analytical study of transport, mixing and chaos in an unsteady vortical flow
Journal of Fluid Mechanics, 214
J. Meiss (2015)
Thirty years of turnstiles and transport.
Chaos, 25 9
M. Finn, Jean-Luc Thiffeault, E. Gouillart (2005)
Topological chaos in spatially periodic mixers
Physica D: Nonlinear Phenomena, 221
for the figure presenting the coherent structures detected from two sample sets of synthetic trajectories
(2016)
Discrete and Continuous Dynamical Systems - Series S 9
(1995)
Physical Review E
Y. Yomdin (1987)
Volume growth and entropy
Israel Journal of Mathematics, 57
Jean-Luc Thiffeault (2005)
Scalar Decay in Chaotic Mixing
, 744
(2015)
Chaos: An Interdisciplinary Journal of Nonlinear Science 25
Jean-Luc Thiffeault (2010)
Chaos in the Gulf
Science, 330
Michael Allshouse, G. Ivey, Ryan Lowe, N. Jones, C. Beegle-Krause, Jiangtao Xu, T. Peacock (2017)
Impact of windage on ocean surface Lagrangian coherent structures
Environmental Fluid Mechanics, 17
Jean-Luc Thiffeault (2002)
Stretching and curvature of material lines in chaotic flows
Physica D: Nonlinear Phenomena, 198
Zhiping You, E. Kostelich, J. Yorke (1991)
CALCULATING STABLE AND UNSTABLE MANIFOLDS
International Journal of Bifurcation and Chaos, 01
(2008)
Discrete Applied Mathematics Applications of Algebra to Cryptography
(2015). Multifractal analysis of Birkhoff averages for countable Markov maps. Ergodic Theory and Dynamical Systems (8),
G. Froyland, K. Padberg (2009)
Almost-invariant sets and invariant manifolds — Connecting probabilistic and geometric descriptions of coherent structures in flows
Physica D: Nonlinear Phenomena, 238
K. Katija, J. Dabiri (2009)
A viscosity-enhanced mechanism for biogenic ocean mixing
Nature, 460
L. Rey-Bellet, L. Young (2017)
Ergodic Theory and Dynamical Systems :
(1999)
The matlab particle tracking code repository,
T. Peacock, G. Haller (2013)
Lagrangian coherent structures: The hidden skeleton of fluid flows
Physics Today, 66
(2013)
Physics Today 66
E. D’Aniello (2013)
ON TOPOLOGICAL CHAOS
(2017)
Nonlin
M. Nodet (2007)
Variational assimilation of Lagrangian data in oceanography
Inverse Problems, 22
G. Mathew, I. Mezić, S. Grivopoulos, U. Vaidya, L. Petzold (2006)
Optimal control of mixing in Stokes fluid flows
Journal of Fluid Mechanics, 580
E. Bollt, N. Santitissadeekorn (2013)
Applied and Computational Measurable Dynamics
, 18
Sanjeeva Balasuriya (2016)
Barriers and transport in unsteady flows - a Melnikov approach
, 21
Jean-Luc Thiffeault (2016)
The Mathematics of Taffy Pullers
The Mathematical Intelligencer, 40
(2007)
Journal of the European Mathematical Society (JEMS) 9
Patrick Dehornoy (2007)
Efficient solutions to the braid isotopy problem
Discret. Appl. Math., 156
(2001)
Physics Letters A 287
Jean-Luc Thiffeault (2011)
Using multiscale norms to quantify mixing and transport
Nonlinearity, 25
W. Hackborn, M. Ulucakli, T. Yuster (1997)
A theoretical and experimental study of hyperbolic and degenerate mixing regions in a chaotic Stokes flow
Journal of Fluid Mechanics, 346
(2018)
Physica D: Nonlinear Phenomena 372
M. Stremler, S. Ross, P. Grover, Pankaj Kumar (2011)
Topological chaos and periodic braiding of almost-cyclic sets.
Physical review letters, 106 11
Michael Allshouse, T. Peacock (2015)
Lagrangian based methods for coherent structure detection.
Chaos, 25 9
B. Noack, I. Mezić, G. Tadmor, A. Banaszuk (2004)
Optimal mixing in recirculation zones
Physics of Fluids, 16
Stephen Wiggins, J. Ottino (2004)
Foundations of chaotic mixing
Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 362
Sanjeeva Balasuriya, N. Ouellette, I. Rypina (2018)
Generalized Lagrangian coherent structures
Physica D: Nonlinear Phenomena
E. Bollt, T. Stanford, K. Życzkowski, Y. Lai (2001)
What symbolic dynamics do we get with a misplaced partition? On the validity of threshold crossings analysis of chaotic time-series
Physica D: Nonlinear Phenomena, 154
Toby Hall, S. Yurttaş (2008)
On the topological entropy of families of braids
arXiv: Dynamical Systems
(1987)
Israel Journal of Mathematics 57
Alireza Hadjighasem, M. Farazmand, D. Blazevski, G. Froyland, G. Haller (2017)
A critical comparison of Lagrangian methods for coherent structure detection.
Chaos, 27 5
(2003)
Physics of Fluids 15
S. Childress, A. Gilbert, E. Ott (1995)
Stretch, Twist, Fold: The Fast Dynamo
Yong Wang, G. Haller, A. Banaszuk, G. Tadmor (2003)
Closed-loop Lagrangian separation control in a bluff body shear flow model
Physics of Fluids, 15
Jean-Luc Thiffeault, A. Boozer (2000)
Geometrical constraints on finite-time Lyapunov exponents in two and three dimensions.
Chaos, 11 1
Jean-Luc Thiffeault, M. Budišić (2014)
Braidlab: A Software Package for Braids and Loops
arXiv: Geometric Topology
M. Finn, Jean-Luc Thiffeault (2006)
Topological Entropy of Braids on the Torus
SIAM J. Appl. Dyn. Syst., 6
Jean-Luc Thiffeault (2004)
Measuring topological chaos.
Physical review letters, 94 8
D. Kelley, N. Ouellette (2011)
Onset of three-dimensionality in electromagnetically driven thin-layer flows
Physics of Fluids, 23
I. Rypina, L. Pratt (2017)
Trajectory encounter volume as a diagnostic of mixing potential in fluid flows
Nonlinear Processes in Geophysics, 24
Jean-Luc Thiffeault, M. Finn (2006)
Topology, braids and mixing in fluids
Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 364
(1987)
Nuclear Physics B - Proceedings Supplements 2
C. Taylor, Stefan Smith (2016)
Dynamics and transport properties of three surface quasigeostrophic point vortices.
Chaos, 26 11
G. Froyland, Kathrin Padberg-Gehle (2012)
Finite-time entropy: A probabilistic approach for measuring nonlinear stretching
Physica D: Nonlinear Phenomena, 241
(2006)
Inverse Problems 22
Amanda Tan, Eric Roberts, K. Mitchell, L. Hirst (2018)
Investigating Quality of Mixing of a Biological Active Nematic
Biophysical Journal, 114
M. Olascoaga, G. Haller (2012)
Forecasting sudden changes in environmental pollution patterns
Proceedings of the National Academy of Sciences, 109
S. Shadden, J. Dabiri, J. Marsden (2006)
Lagrangian analysis of fluid transport in empirical vortex ring flows
Physics of Fluids, 18
I. Mezić, Sophie Loire, V. Fonoberov, P. Hogan (2010)
A New Mixing Diagnostic and Gulf Oil Spill Movement
Science, 330
E. Gouillart, Jean-Luc Thiffeault, M. Finn (2005)
Topological mixing with ghost rods.
Physical review. E, Statistical, nonlinear, and soft matter physics, 73 3 Pt 2
A. Leonard, V. Rom-Kedar, S. Wiggins (1987)
Fluid Mixing and Dynamical Systems
, 2
(2009)
Oceanography 22
I. Dynnikov (2002)
On a Yang-Baxter map and the Dehornoy ordering
Russian Mathematical Surveys, 57
S. Newhouse (1988)
Entropy and volume
Ergodic Theory and Dynamical Systems, 8
H. Aref, J. Blake, Marko Budivsi'c, S. Cardoso, J. Cartwright, H. Clercx, K. Omari, U. Feudel, R. Golestanian, E. Gouillart, G. Heijst, T. Krasnopolskaya, Y. Guer, R. MacKay, V. Meleshko, G. Metcalfe, Igor Mezi'c, A. Moura, O. Piro, M. Speetjens, R. Sturman, Jean-Luc Thiffeault, Idan Tuval (2014)
Frontiers of chaotic advection
Reviews of Modern Physics, 89
J. Weiss, A. Provenzale (2008)
Transport and mixing in geophysical flows
P. Boyland, H. Aref, M. Stremler (2000)
Topological fluid mechanics of stirring
Journal of Fluid Mechanics, 403
N. Francois, H. Xia, H. Punzmann, B. Faber, M. Shats (2015)
Braid Entropy of Two-Dimensional Turbulence
Scientific Reports, 5
R. MacKay, J. Meiss, I. Percival (1984)
Transport in Hamiltonian Systems
Hamiltonian Dynamical Systems
R. Sturman, J. Ottino, Stephen Wiggins (2006)
The Mathematical Foundations of Mixing by Rob Sturman
W. Hackborn (1990)
Asymmetric Stokes flow between parallel planes due to a rotlet
Journal of Fluid Mechanics, 218
P. Boyland (1994)
Topological methods in surface dynamics
Topology and its Applications, 58
S. Newhouse, T. Pignataro (1993)
On the estimation of topological entropy
Journal of Statistical Physics, 72
(1989)
Physical Review Letters 63
(2009)
Nature 460
Matthew Williams, I. Rypina, C. Rowley (2014)
Identifying finite-time coherent sets from limited quantities of Lagrangian data.
Chaos, 25 8
(2006)
Functional Analysis and Other Mathematics 1
(2015)
Scientific Reports 5
L. Shampine, M. Reichelt, J. Kierzenka (1999)
Solving Index-1 DAEs in MATLAB and Simulink
SIAM Rev., 41
R. Davidchack, Ying-Cheng Lai, E. Bollt, Mukeshwar Dhamala (2000)
Estimating generating partitions of chaotic systems by unstable periodic orbits
Physical review. E, Statistical physics, plasmas, fluids, and related interdisciplinary topics, 61 2
R. Davidchack, Ying-Cheng Lai (1999)
Efficient algorithm for detecting unstable periodic orbits in chaotic systems.
Physical review. E, Statistical physics, plasmas, fluids, and related interdisciplinary topics, 60 5 Pt B
(2013)
Annual Review of Marine Science 5
T. Ma, E. Bollt (2014)
Shape Coherence and Finite-Time Curvature Evolution
Int. J. Bifurc. Chaos, 25
(2010)
Science Magazine 330
M. Weldon, T. Peacock, G. Jacobs, M. Helu, G. Haller (2008)
Experimental and numerical investigation of the kinematic theory of unsteady separation
Journal of Fluid Mechanics, 611

ISSN: 2469-990X
eISSN: ARCH-3338
DOI: 10.1103/PhysRevFluids.5.054504
Publisher site: See Article on Publisher Site

Abstract

Using braids to quantify interface growth and coherence in a rotor-oscillator ﬂow Margaux Filippi Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A. and Woods Hole Oceanographic Institution, Woods Hole, MA 02543, U.S.A. Marko Budiˇ si´ c Department of Mathematics, Clarkson University, Potsdam, NY 13699, U.S.A. Michael R. Allshouse Department of Mechanical and Industrial Engineering, Northeastern University, Boston, MA 02115, U.S.A. S´ everine Atis Department of Physics, University of Chicago, Chicago, IL 60637, U.S.A. Jean-Luc Thiﬀeault Department of Mathematics, University of Wisconsin – Madison, Madison, WI 53706, U.S.A. Thomas Peacock Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A. (Dated: May 26, 2020) The growth rate of material interfaces is an important proxy for mixing and reaction rates in ﬂuid dynamics, and can also be used to identify regions of coherence. Estimating such growth rates can be diﬃcult, since they depend on detailed properties of the velocity ﬁeld, such as its derivatives, that are hard to measure directly. When an experiment gives only sparse trajectory data, it is natural to encode planar trajectories as mathematical braids, which are topological objects that contain information on the mixing characteristics of the ﬂow, in particular through their action on topological loops. We test such braid methods on an experimental system, the rotor-oscillator ﬂow, which is well-described by a theoretical model. We conduct a series of laboratory experiments to collect particle tracking and particle image velocimetry data, and use the particle tracks to identify regions of coherence within the ﬂow that match the results obtained from the model velocity ﬁeld. We then use the data to estimate growth rates of material interface, using both the braid approach and numerical simulations. The interface growth rates follow similar qualitative trends in both the experiment and model, but have signiﬁcant quantitative diﬀerences, suggesting that the two are not as similar as ﬁrst seems. Our results shows that there are challenges in using the braid approach to analyze data, in particular the need for long trajectories, but that these are not insurmountable. I. INTRODUCTION and arise in engineered systems and technological pro- cesses [19–23]. Characterization of material mixing in ﬂuid ﬂows is an Quantiﬁcation of mixing typically focuses on charac- active research area at the intersection of mathematics, terizing the amount of stirring by chaotic advection [2], engineering and physics. Two common types of anal- which in turn relates to the role that stretch-and-fold ysis are to (1) quantify the overall amount of stirring and stretch-and-stack mechanisms play in mixing of ma- and mixing that occurs in the system and (2) to iden- terial [1, 5, 6, 24, 25]. The strength of these mecha- tify structures that enhance or reduce material mixing. nisms is computed by estimating local rates of exponen- The earliest examples of a systematic mathematical ap- tial stretching of the material [26, 27] over a timescale as- proach to these problems uses techniques from nonlinear sociated with folding or stacking. Alternatively, chaotic dynamics that rely on having the ﬂuid velocity ﬁeld as advection can be quantiﬁed by studying norms of scalar a continuous, often diﬀerentiable, function of space and ﬁelds advected in the ﬂow [28–30]. A review of associated time [1–4]. (See for example [1, 5–11] for comprehen- topics with additional references can be found in [11]. sive reviews.) By now, these techniques have been suﬃ- The search for structures that enhance or prevent mix- ciently adapted and reﬁned to be applicable to data from ing typically focuses on features of the ﬂow that remain real ﬂuid ﬂows that, for example, arise in and around liv- coherent over relevant timescales and organize transport ing organisms [12–15], govern ecological processes [16–18] by, for example, attracting, repelling, or containing the material. The most well-known of such structures are the Lagrangian Coherent Structures (LCS) [31], which are margaux@mit.edu distinguished low-dimensional barriers to transport, and arXiv:1910.12779v2 [physics.flu-dyn] 23 May 2020 2 Almost-Coherent Sets [32], which are regions in which uate how well suited braid theory is for the characteri- the transport is contained. In all cases, the true challenge zation of experimental ﬂuid ﬂows. The ﬂow studied here remains to deﬁne the objects of interest so that they en- is the rotor-oscillator Stokes ﬂow, a canonical example compass all relevant physical phenomena, and to propose of a ﬂow ﬁeld that possesses both chaotic and coherent and implement an algorithm that identiﬁes them, espe- regions, described by Hackborn et al. [77] and Weldon cially in transient or aperiodic ﬂows [31, 33–37]. The de- et al. [19]. In contrast to ﬂows such as the double-gyre, tection of coherent structures oﬀers especially-signiﬁcant Bickley jet, or the Duﬃng oscillator, which have been applications to geophysical ﬂows, where these techniques more commonly used to analyze material transport, the have been used to understand climate change and plan re- rotor-oscillator ﬂow has been analyzed both as an ana- sponses to ecological catastrophes [17, 38–41]. The most lytical model [77] and as an experimental ﬂow [19]. The commonly used tools to detect coherent structures are braid-based analysis will be twofold: ﬁrst, it will pro- based on material deformation [31, 42–44] and on prob- vide an identiﬁcation of regions of the ﬂow surrounded abilistic [32, 45] properties of the ﬂow. A comparison by minimally growing material lines (typically, coherent of the wide range of approaches for detecting coherent material sets); second, an estimation of maximum growth structures can be found in [36]. rate of material interfaces (typically found in the chaotic region). Both calculations will be applied to the analyt- Input data in the form of continuous velocity ﬁeld ical model and to the experimental data, in an eﬀort to stands in contrast to observations in oceans and atmo- expose and address challenges that come with analyzing sphere that are recorded by deploying sparse sensors that non-idealized ﬂows. measure physical properties (temperature, salinity, etc.) The paper is organized as follows. Section II presents of the ﬂuid ﬂow as they are advected [46–48]. The rela- tive sparsity of sensors prevents a reliable estimation of a the model and experimental ﬂows used in this study. Sec- tion III gives a short summary of the braid representation continuous velocity ﬁeld; to analyze such sparse datasets, a number of methods have been developed that require of the ﬂow kinematics. Next, the paper demonstrates how two particular analysis methods can be applied to only a ﬁnite set of discretized trajectories [49–54]. model and experimental ﬂows: Section IV explains the Sparse-data methods include the braid dynamics meth- detection of coherent structures, comparing the analysis ods employed in this paper, which require only a set of of numerically-advected and experimental particle tra- discrete trajectories instead of the full velocity ﬁeld. Fol- jectories, while Section V explains estimation of material lowing Boyland et al. [55], the study of the topological growth, comparing the analysis of model and experimen- properties of ﬂuid stirring has developed into an active tal velocity ﬁelds. Section VI summarizes what has been research area. The topological approach is particularly learned by employing braid dynamics to study model and well-suited to the study of mixing by rods, vortices, or experimental rotor-oscillator ﬂows. otherwise distinguished Lagrangian trajectories [56–64]. In particular Thiﬀeault [56, 65] and Allshouse and Thif- feault [49], Budiˇ si´ c and Thiﬀeault [66] used braid theory to characterize mixing and coherent structures from pla- II. THE ROTOR-OSCILLATOR FLOW nar ﬂows solely from particle trajectories, forming the basis for the approaches used in this paper. The rotor-oscillator ﬂow is a planar, non-autonomous, In planar ﬂows, the rate of exponential growth of ma- incompressible ﬂow where the motion of the ﬂuid is in- terial interfaces corresponds to the topological entropy duced by a rotor that simultaneously rotates about its of the ﬂow [55, 56, 65, 67–70], which quantiﬁes the com- axis and oscillates along a channel. Experimentally, plexity of trajectories evolving in a dynamical system. particle motion is conﬁned to a two-dimensional ﬂuid Techniques for estimating topological entropy (without layer in a rectangular tank, with advection induced by estimating growth rate ﬁrst) typically use precise veloc- a fast-spinning cylinder that also oscillates in the longi- ity ﬁelds to compute unstable periodic orbits [71–74], or tudinal direction. This section summarizes the analytic intricate partitions of the ﬂow [75, 76]. Budiˇ si´ c and Thif- model and the experimental realization; further details feault [66] developed a method for calculating the Finite- are found in [19, 77, 78]. Time Braiding Exponent (FTBE), which is an approxi- mation for the topological entropy that is more applicable to ﬁnite-time, sparse datasets. The FTBE provides a ro- A. Analytic model bust measure of mixing that approaches the topological entropy as the number of trajectories is increased. More recently, Roberts et al. [64] developed a braid-free ap- Hackborn et al. [77] studied the rotor-oscillator ﬂow, proach that also estimates topological entropy based on an asymmetric ﬂow between parallel plates driven by the the relationship with growth of material lines, without rotation of a cylinder (rotor ) and the longitudinal oscilla- detailed knowledge of the velocity ﬁeld. tion of one of the walls of the ﬂow tank (oscillator ). In a Since the methods of Budiˇ si´ c and Thiﬀeault [66] and suitable parameter regime, the ﬂow exhibits a main vor- Allshouse and Thiﬀeault [49] were developed and applied tex around the rotor, with two secondary recirculating to model ﬂows, the principal aim of this paper is to eval- vortices on each side of the rotor. Each of those vortices 3 −1 −0.5 0.5 −3 −2 −1 0 1 2 3 −0.5 −0.25 FIG. 1. Geometry for the rotor-oscillator ﬂow. The walls are at x = ±h, with the rotor oscillating in y at constant x = c. 0.25 0.5 −1.5 −1 −0.5 0 0.5 1 1.5 creates its own, progressively-weaker, array of vortices, which will not play a role in this work. We take coordinates (x, y) ∈ [−h, h]×R, with the walls FIG. 2. Poincar´ e map for the model rotor-oscillator ﬂow of the channel at x = ±h. The cylindrical rotor starts at [19, 77] with c = 0.54, = 0.125, λ = 2π/5 simulated for y = 0 at t = 0 and moves along the line x = c. The rotor 300 forcing cycles. The blue and red colors depend on the av- has radius a and angular velocity ω. A diagram of the erage value of the y coordinate along simulated trajectories. geometry is shown in Figure 1. (The peculiar interchange Bottom panel shows a detailed view of the dashed frame. The of x and y axes is for consistency with earlier work [77].) Poincar´ e map is sampled at the rotor’s oscillation period τ . Without longitudinal (y) oscillation, the rotating cylin- with non-dimensional parameters der induces, in the Stokes limit, a steady velocity ﬁeld. In non-dimensional coordinates, the span-wise width is 2 2 2 = V h/2a ω, λ = h Ω/a ω, (4) x ∈ [−1, 1] with the position of the rotor c = 0.54 set to match [77]. We absorb the rotation frequency ω into the that combine the channel width h, magnitude of oscil- non-dimensional time t, so the resulting stream function lation V and angular frequencies of rotation ω and of ψ(x, y) [77, 78] is oscillation Ω. The use of a translated steady solution is valid as long as Ω is not so large as to invalidate the ψ(x, y) = log[f (x, y)] + g(x, k) cos(ky)dk (1) Stokes approximation. | {z } | {z } Passive particles are advected via point vortex boundary eﬀects x ˙ (t) = −∂ ψ(x, y − sin λt), where (5) y ˙ (t) = ∂ ψ(x, y − sin λt). cosh(πy/2)− cos(π(c− x)/2) f (x, y) = ; Fixing c = 0.54, Hackborn et al. [77] varied and λ to cosh(πy/2) + cos(π(c + x)/2) identify regimes with chaotic dynamics and islands of co- 2[tanh k cosh kx− x sinh kx] cosh kc herence using Poincar´ e maps. Figure 2 shows a Poincar´ e g(x, k) = sinh 2k + 2k map generated with = 0.125 and λ = 2π/5. 2[coth k sinh kx− x cosh kx] sinh kc Timescales that govern the evolution of a single La- + . sinh 2k − 2k grangian trajectory vary widely depending on its loca- tion. Trajectories that encounter the rotor will circle Following [77], we assume the rotor to be a point vortex around it rapidly; trajectories that remain on the outside with a vanishing diameter (a → 0) while keeping a ω edge of both the rotor and recirculating vortices evolve on constant. The time non-dimensionalization is kept as timescales separated by two or more orders of magnitude follows: compared to the rotor rotation. ta ω → t. (2) B. Experimental ﬂow The longitudinal oscillation of frequency Ω is mod- The experimental apparatus was inspired by Weldon eled by the time-periodic translation of the steady stream et al. [19], and is depicted in Figure 3. The main diﬀer- function ψ(x, y, t): ence compared to the analytical model of Hackborn et al. Ψ(x, y, t) = ψ(x, y − sin λt), (3) [77] is in apparent change of the reference frame: they ﬁx 4 the rotor and moved the walls, while Weldon et al. [19] ﬁx the walls and move the rotor. The rotor-oscillator system was recreated in a 90 mm× 402 mm acrylic ﬂow tank. For the rotor, an alu- minum rod of diameter a = 3.165 mm was attached to a stepper motor through a plastic sleeve for thermal insu- lation. To longitudinally oscillate the rotor, the stepper motor was mounted to a longitudinal traverse controlled by a second stepper motor. The tank and the traverse were mounted onto alu- minum frames and aligned horizontally. To visualize a horizontal cross-section of the system, a front-faced mir- (a) ror was placed underneath the tank at a 45 angle. A camera was mounted facing the mirror with the rod at the center of the ﬁeld of view. The experimental images were acquired through the mirror reﬂection at 45 and recorded with a LaVision Imager Pro X 4M CCD cam- vegetable oil era of resolution 2042× 2042 pixels with a 28 mm lens. Image acquisition, calibration and cross-correlation was performed by LaVision’s DaVis imaging software. To mitigate three-dimensional eﬀects due to evaporation or interaction with the bottom wall, the ﬂuid system was trapped between a top layer of vegetable oil and a bot- tom layer of FC-40 coolant. Two separate experiments were used to record proper- (b) ties of the ﬂow: a particle tracking experiment (PT, Sec- tion II B 1), to record Lagrangian trajectories, and a par- FIG. 3. Schematics of the experimental set-up for the ticle image velocimetry experiment (PIV, Section II B 2), Hackborn–Weldon rotor-oscillator ﬂow [19, 77], with still to record instantaneous velocity ﬁelds. Table I summa- walls and an oscillating rotor. (a) Experimental apparatus set up. (b) Fluid system to set up a two-dimensional ﬂow rizes the experimental parameters for the two experi- through density layers for the PT experiment. In the PIV ments. For all experiments, the Reynolds number was experiment, the tracer particles were mixed to a single layer estimated to be O(1). of glycerol and a laser sheet was projected horizontally. 1. Particle tracking (PT) experiment There were 77 918 frames (time steps) recorded, for a total run duration of 7791.8 s, corresponding to about Tracer particles were custom-made from cellulose ac- 82.3 periods. The particle trajectories were obtained etate polymer spheres by Cospheric, with diameter through a MATLAB package, the Tracking Code Repos- 1.78 mm and density 1.285 g/cm . The particles were itory [79]. For the particle tracking experiment, the non- painted ﬂuorescent to create a high light contrast and dimensional parameters were = 0.1274 and λ = 1.2806. thereby enhance image acquisition. The coated par- The physical period of oscillation was τ ˜ = 94.7163 s. ticles were then ﬁltered by density to ensure they re- mained on a virtually two-dimensional plane, the sur- face between the lower layer of salty glycerol, of den- 2. Particle image velocimetry (PIV) experiment sity 1.297 g/cm , and the upper layer of pure glycerol, of density 1.261 g/cm . The denser glycerol solution was mixed with salt and water to match the viscosity of Tracers for the PIV experiment were hollow glass −3 2 pure glycerol, measured to be ν = 1.1197× 10 m /s. spheres of mean diameter 8−12 μm mixed with glycerol. The ﬂuorescent particles were illuminated with an ul- The Reynolds number of the glass spheres at the rotor’s −4 traviolet light in a dark room. With the bead diame- boundary was Re ≈ 1.9977× 10 and are assumed sphere −3 ter D = 1.78× 10 m, the viscosity ν, a rod of radius to move as passive tracers. A pulsed Nd:YAG laser was a = 3.165 mm and angular velocity of ω = 10.4720 rad/s, powered to illuminate the glycerol–glass spheres solution the Reynold number of the beads at the rotor’s bound- in a horizontal plane normal to the tank front wall. To re- (ωa)D −2 ary was Re = ≈ 5.2689× 10 . The velocity duce artifacts and noise in the velocity ﬁelds, the domain bead being maximized at the rotor’s boundary, Re is an was partitioned into two for image processing in DaVis; bead upper bound for the ﬂow ﬁeld, so the neutrally buoyant this helped compensate for the wide range of velocities beads are assumed to move as passive tracers. between the rod’s rotation and the slower motion on the The images were acquired at a frequency f = 10 Hz. outer domain. First, the region around the rod was pro- 5 TABLE I. Experimental parameters. symbol description value (PIV) value (PT) h half-width of the channel 0.045 m a rotor radius 0.003 165 m c distance of rotor from the wall 0.023 76 m τ ˜ translational period of oscillation 96.5201 s 94.7163 s −4 −4 V translational velocity magnitude 5.8275× 10 m 5.9385× 10 m Ω translational angular velocity 0.0651 rad/s 0.0663 rad/s ω rotational angular velocity 10.4720 rad/s Re Reynolds number 2.6640 f sampling frequency 1 Hz 10 Hz non-dim. oscillation magnitude (V h/2a ω) 0.1250 0.1274 2 2 λ non-dim. oscillation ang. frequency (h Ω/ωa ) 2π/5 ≈ 1.2566 1.2806 In calculating the Reynolds number for the experiments, we use ωa for the characteristic linear velocity and 2h as the characteristic length, resulting in Re = 2ωah/ν. Hackborn 2 −2 et al. [77] use a as the characteristic length, resulting in Re = a ω/ν ≈ 9.3682× 10 here. Eﬀective frequency of sampling. See Section II B 2 for details. cessed at f = 10 Hz and bin-averaged in 10-image blocks; on to the trajectory integrator. Trajectories were inte- second, the outer domain was processed at f = 1 Hz and grated using the variable-step, variable-order MATLAB matched to the bin centers of the ﬁrst part. The re- ode15s [81] algorithm to manage the stiﬀness of the dif- sulting eﬀective sampling rate was f = 1 Hz. There were ferential equations, which is a consequence of the large 10 730 frames recorded, for a total run duration of 1073 s, diﬀerence in timescales of primary and secondary rotors. corresponding to about 11.1 periods. For the PIV exper- To simulate times longer than the recorded number of iment, the non-dimensional parameters were = 0.1250 oscillation periods, the PIV velocity ﬁeld was periodized and λ = 1.2566. The physical period of rotor oscillation in time. was τ ˜ = 96.5201 s, which corresponds to 5 dimensionless time units. III. BRAID DYNAMICS The velocity ﬁelds were post-processed using the ap- proach of Kelley and Ouellette [80] to impose incom- pressibility for the PIV-obtained velocity ﬁeld; this was Braid theory is an algebraic way of characterizing and necessary because of the multiple repeated periods over classifying continuous maps based on their topological which the PIV data was used for our investigations. Fig- properties. In our application, the continuous maps are ure 4 demonstrates that instantaneous streamlines of the ﬂow maps generated by the two-dimensional (un- the model and experimental velocity ﬁelds qualitatively steady) ﬂuid velocity ﬁeld over a particular time interval, match. Due to experimental constraints, however, the as studied by [55–57, 65, 67]. The “input data” for the streamlines obtained from the PIV experiments deviate braid theory characterization of ﬂuid ﬂows is a set of N from the streamlines obtained from the model velocity continuous particle trajectories evolved concurrently over ﬁelds. Some of these discrepancies come from the fun- a ﬁnite time interval; in particular, the analysis does not damental diﬀerences between the model and experimen- require access to the velocity ﬁeld or its gradients. tal ﬂow. First, the model assumes the rotor to be a Braids are constructed from a set of N particle tra- point vortex, whereas a physical rod had to be present 2 jectories p (t) ∈ R , i = 1, 2, . . . , N , with 0 ≤ t ≤ T . to stir the ﬂuid. In spite of the small rod diameter of A physical braid is the embedding of trajectories in the a = 3.165 mm, the ﬂows deviate around the rotor. Sec- three-dimensional space-time (Fig. 5(a)), where individ- ond, the model assumes a Stokes ﬂow with zero Reynolds ual trajectories (strands ) weave around each other. A number: with a 6= 0 as the characteristic length, the ex- topological braid (Fig. 5(c)) is a reduced representation −2 perimental Reynolds number is Re ≈ 9.3682× 10 , as of the physical braid that retains only the sequence of ex- described in Table I. Additionally, the deviations are es- changes of the strand order with respect to a ﬁxed space- pecially prominent near the lateral boundaries, due to the time plane onto which the strands are projected. This laser beam’s reﬂection at the walls of the tank, which in- plane can be chosen largely arbitrarily for the purposes troduced artifacts in the velocity ﬁelds. The PIV record of this article [56, 65, 66, 82]. is not available for the full width of the channel, but Strand exchanges can be represented by a sequence of rather in the approximate band −0.85 < x < 0.85 in the symbols σ called Artin braid generators. The generator nondimensional spanwise coordinate. σ represents the crossing of the ith strand in front of the To create synthetic trajectories in the experimental (i + 1)th strand in Fig. 5(c), where the index i indicates velocity ﬁeld, the velocities were linearly interpolated the order of strands from left to right just before the −1 between spatial nodes and time samples before passed crossing occurs; the inverse generator σ represents the i 6 −1 −0.5 0.5 t t −−2 2 −1 0 1 2 2 (a) −1 −0.5 x (a) Physical braid (b) Projection onto (c) Topological x-coordinate braid 0.5 FIG. 5. A physical braid and a corresponding topologi- cal braid generated from ﬁve trajectories. In all diagrams the time ﬂows from bottom to top. The sequence of gen- −−2 2 −1 0 1 2 2 erators (ordered from left to right in increasing time) is −1 −1 −1 −1 B = σ σ σ σ σ σ σ σ σ σ σ . 4 1 2 3 2 4 3 2 2 4 4 (b) FIG. 4. Snapshot of instantaneous streamlines of (a) the model velocity ﬁeld [19, 77] with c = 0.54, = 0.125, λ = 2π/5 (a) and (b) the PIV-recorded experimental velocity ﬁeld at the same time instance. (b) crossing of ith strand behind the (i + 1)th strand. A topological braid constructed from a given set of FIG. 6. Two topological loops before (left column) and af- trajectories captures topological information about the ter (right column) action of the braid from Figure 5. The ﬂow. The topological analog of material advection in strands of the braid are shown in black as cross-sections. The ﬂuid ﬂows is the action of a braid on topological loops. A Dynnikov vectors before and after the braid action in (a) topological loop is a collection of closed, non-intersecting B B 0 −2 3 0 −1 0 are [ 1 1 1 0 0 0 ] −→ [ ] and in (b) [ 0 0 0 0 0 1 ] −→ curves that enclose two or more (but not all) strands in [ 0 −1 1 0 −1 0 ]. the braid (Figure 6). The curves in a loop are “pulled tight,” i.e., they can be pictured as rubber bands tightly wrapped around strands of the physical braid. The ac- tures and between the punctures, respectively. The Dyn- tion of a braid B on a loop `, denoted by B`, is visualized nikov coordinate vector is calculated from intersection by sliding the rubber bands along the physical braid in numbers as the direction of time. As strands exchange places, the loop is forced to stretch since it cannot pass through the a a ··· a b b ··· b , where 1 2 N−2 1 2 N−2 (6) strands (Figure 6). a = (α − α )/2, b = (β − β )/2. n 2n 2n−1 n n+1 n Computationally, any loop acted on by the N -strand braid can be represented by a vector of 2N − 4 signed The action of each braid generator σ is then repre- 2N−4 2N−4 integers, and vice-versa, through the Dynnikov coor- sented by a piecewise-linear map σ : Z → Z ; dinatization of loops [65, 83, 84]. Figure 7 demon- explicit expressions for σ can be found in [65, 84]. strates how to calculate the Dynnokov coordinates of All braid computations and visualizations in this pa- a loop around N = 5 strands. The intersection num- per have been produced using the MATLAB toolbox bers α , α ,··· , α count the number of intersections braidlab [82]. 1 3 2N−3 between the tightened loop and axes above the punc- To measure the amount of stretching, we compute the tures, while α , α ,··· , α and β ,··· , β count “length” of a topological loop|`| as the number of times it 2 4 2N−4 1 N−1 the number of intersections with axes below the punc- intersects the horizontal axis that would pass through all 7 α α α β 1 β 3 β 5 β 1 2 3 4 work improves the algorithm to more eﬃciently detect co- herent structures for larger trajectory sets. The method is then applied to a synthetic set of trajectories based on the model velocity ﬁeld presented in II A and on experi- 1 2 3 4 5 mentally measured trajectories, allowing for comparison between theory and experiments. A. Theory α α α 2 4 6 There are several diﬀerent deﬁnitions of coherent struc- FIG. 7. Coordinate system based on intersection numbers tures used for characterizing patches of passive tracer α , β used to calculate Dynnikov coordinates (6). In this i i that do not disperse under Lagrangian transport [31, 44, example of N = 5 strands, α = 0, α = 2, β = 2, 1,3,5 2,4,6 1,2,3,4 87]. In the context of braid theory, Allshouse and Thif- 1 1 1 0 0 0 resulting in the Dynnikov vector [ ]. Precise location of strands in the x, y space is not important, only their order feault [49] deﬁne coherent sets as a set of trajectories along the projection coordinate. surrounded by an initially “simple” topological loop that does not grow, or grows sub-exponentially, over the du- ration of the braid. Sub-exponential growth can occur points in Figure 6. Although the quantity |`| on its own as particles rotate as a patch, so we do not consider this does not correspond to spatial length of a material curve, type of growth as indicative of long term mixing. We it is useful for characterizing the relative loop growth now summarize the relevant techniques from [49] with factor modiﬁcations to make the presentation more compact. Simple topological loops are those whose Dynnikov co- |B`|/|`|, (7) ordinates (see Section III) contain only {±1, 0}, i.e., their 2N−4 2N−4 coordinate vectors are {±1, 0} ⊂ Z . As the measuring the ratio of loop length before and after appli- magnitude of elements in the coordinate vectors relate cation of the braid B. The loop growth factor will play a to the number of folds of a loop around particles, these prominent role in the following sections as a criterion for loops will have one or no folds as they surround parti- searching for maximally and minimally growing loops `, cles (see ﬁgure Figure 8 for examples). As mentioned given a speciﬁc braid B. in [49], enumerating all simple loops and checking how The topological braid can be interpreted as a dramat- fast they grow under the action of the braid is com- ically reduced representation of the full ﬂow using only putationally intractable, even for modestly-size braids: 2N−4 2N−4 N a ﬁnite number of trajectories, with the physical trajec- #({±1, 0} ) = 3 ∼ 9 , a number that is pro- tories represented by topological operations, and mate- hibitively large for even N ≈ 20 strands in a braid. rial advection represented by stretching of “rubber-band” Therefore, we look to reduce the number of simple loops loops. Despite their simplicity, braid-based calculations considered, while still retaining a way to determine which provide bounds on the rate of material stretching in the advected particles are in coherent structures. full ﬂow, and approximate boundaries of Lagrangian co- To reduce the number of loops considered, Allshouse herent structures, as explained in Sections IV, V. Addi- and Thiﬀeault [49] focuses on pair-loops that enclose tionally, the reduction also yields signiﬁcant speed up as only two out of the N particles. Although there are compared to material line advection, which can rely on = N (N− 1)/2 particle pairs to choose, there are still expensive interpolation methods [85]. many simple loops that enclose any given pair. In con- trast to [49] where loops were completely above or below all particles between the particle pair, here we consider IV. DETECTION OF COHERENT a single pair-loop for each pair, namely the one that can STRUCTURES be “pinched” to a straight line in the initial conditions plane; we term these straight pair-loops. This creates a set of loops to consider, which is half as many The braid theory approach to detect coherent struc- tures oﬀers a unique mechanism for partitioning diﬀer- as in [49], while also yielding loops that have a simpler spatial shapes. Figure 8(a) contains the lines (red and ent regions of the ﬂow based on the entanglement of tra- green) connecting pairs of particles, and the correspond- jectories, as opposed to their relative positions [50] and material deformation [86]. A coherent structure identi- ing straight pair-loops in standard form are presented in Figure 8(b). ﬁed using braids is deﬁned as a topological loop that does not grow under the action of the braid. The loop provides The identiﬁcation of coherent structures as loops that an approximate physical boundary that delineates an in- do not grow under the action of the braid is broken into ternal ﬂuid subdomain that does not mix with exterior two parts: identiﬁcation of sets of particles enclosed by a ﬂuid. We present an overview of the theory developed by coherent structure, and creation of a loop around those Allshouse and Thiﬀeault [49], and show how the current particles that does not grow. The starting point for iden- 8 value, the number of coherent structures initially grows as more pair-loops become admissible. Eventually, ad- ditional pair-loops that are added will act as connec- tions between pre-existing connected components, merg- ing pairs of coherent structures. This will result in a de- crease in the number of coherent structures, and growth in size of individual coherent structures, until there is only one large structure containing all particles within the system. Once a set of particles has been identiﬁed as being con- tained by a coherent structure, the corresponding struc- ture is the composite of the straight pair-loops that con- nect the set of particles. In the example presented in Figure 8, three particles (colored in blue) are connected by straight pair-loops, a green loop and a red loop. The ﬁnal step in the coherent structure identiﬁcation is to FIG. 8. (a) Particles (dots) and loops in physical space. Parti- merge the non-growing straight pair-loops into a single cles in the identiﬁed structure (blue) are connected by straight non-straight loop, which we consider to be the bound- lines (solid) and surrounded by the corresponding straight ary of the coherent structure. In Figure 8, the coherent pair-loop (dashed and dotted lines). The coherent structure structure surrounds both straight pair-loops without en- loop (blue line) links the two straight pair-loops. (b) The veloping any of the particles outside the coherent struc- particles and loops in standard form, with points aligned on ture. 0 1 0 −1 0 −1 0 0 0 1 the real axis. The Dynnikov vectors are [ ] 0 1 1 −1 0 −1 0 0 0 1 for the blue curve, [ ] for the red curve and [ 0 1 0 0 0 −1 0 1 0 0 ] for the green curve. B. Coherent structures in the model system tifying sets of particles that are inside a coherent struc- While the braid method for detecting coherent struc- ture is the forming of straight pair-loops ` connecting ij tures was validated by Allshouse and Thiﬀeault [49], the particles p , p ∈ R in the initial condition plane. We i j use of only straight pair-loops is a modiﬁcation that re- represent the collection of N trajectories by vertices in quires similar veriﬁcation. To do so, we analyze a set (a family of ) undirected graphs G; the set of particles of trajectories generated using the model velocity ﬁeld enclosed by a coherent structures will correspond to con- presented in section II A. Because only a single pair- nected components in these graphs. The growth of the loop per pair of trajectories is analyzed, it is possible straight pair loops under the action of the braid will de- to process a larger number of trajectories. Increasing termine which edges e are included or excluded in G. ij the number of trajectories while keeping the domain size Based on the loop growth rate (7), compute L as ij constant has the potential to more accurately locate the the exponential growth rate of the straight pair-loop ` ij boundary of the coherent set with the braid method. As under the action of the braid demonstrated by Allshouse and Thiﬀeault [49], the braid method requires multiple particles to be located within 1 |B` | ij L = log . (8) ij a coherent structure in order for it to be identiﬁed; by T |` | ij increasing the number of trajectories analyzed, it is thus more likely that smaller structures will be identiﬁed. For The graph of vertices G is fully-connected and undi- rected as there is a pair-loop between each pair of ver- the numerically calculated trajectories, 300 trajectories are analyzed and the resulting coherent structures can tices, with edge weight L . Fix a threshold Λ ≥ 0 and ij construct a family of subgraphs G ⊂ G, retaining only be compared with the Poincar´ e map in Figure 2. edges that satisfy L ≤ Λ. G can be interpreted as a The numerically calculated trajectories are randomly ij Λ subgraph of G where connections are made only by sim- distributed throughout the entire domain and are ad- ple straight pair-loops whose exponential growth rate is vected for 50 periods to match the PT experiment. A less than Λ. sampling rate of 100 measurements per period is suﬃ- For any Λ there may be vertices in G that remain dis- cient to properly represent the braid. The 300-trajectory connected from any other vertex. In particular, vertices braid is represented by approximately 2.15 million gener- that are disconnected at Λ = 0 are deemed incoherent, ators. The application of the braid to all of the straight as they are not enclosed by any slow-growing pair-loops. pair-loops is the bottleneck of the algorithm and takes On the other hand, coherent structures will correspond to three hours on a laptop. We use a growth rate thresh- vertices that are mutually connected (connected compo- old of L = 1/T to identify the straight pair-loops that ij nents of the graph). Particles are grouped into a coherent do not grow rapidly over the time interval. Using this structure by greedy agglomeration of the connected ver- growth rate, we identify loops that grow by a factor of tices. We expect that, as Λ is increased from its lowest e over the time interval (see (8)), which is less than 1% 9 FIG. 9. (a) The initial positions of N = 300 analytic model trajectories (dots), colored based on their coherent structure assignment with unassigned particles in black. The lines FIG. 10. (a) The initial positions of N = 33 experimen- connecting particles indicate straight pair-loops that have a tal trajectories (dots), and identiﬁed coherent loops (black growth ratio less than the threshold, with less opaque lines lines). Three coherent structures are identiﬁed (red, blue and representing loops that grow more. (b) and (c) are zoomed in green). The straight pair-loops that have a growth rate of less regions corresponding to the gray dashed boxes in (a). The than the threshold are represented as straight lines connect- coherent structures corresponding to slowly growing loops for ing the particles. The less opaque the line, the greater the each set of points is drawn in the corresponding color. growth rate. (b) The ﬁnal position of the trajectories and the deformed coherent loops. of all the pair-loops. The network of particles connected by a slow-growing, that are a physical distance of 0.01 cm apart. Because of straight pair-loop is presented as lines in Figure 9(a). The their close proximity, these particles do not separate over particles that are connected by a network of these pair- the 50 periods of advection, despite being in the chaotic loops form a coherent set, where each set has a diﬀerent sea. coloring of both the particles and the network connec- tions. Finally, the coherent structure, which is repre- sented by the slow-growing loop that surrounds all the C. Coherent structures in the experimental system particles of a particular coherent set, can be identiﬁed. These resulting slow-growing loops are presented in Fig- ure 9(b) and (c). Having identiﬁed the main coherent structures in the model system, we next apply the method to experimen- Four diﬀerent types of coherent structures were iden- tally observed trajectories. The particle trajectories gen- tiﬁed by the method. The two structures in red and erated by the experimental system (see section II B 1 and blue represent each of the main recirculating vortices. table I) can be applied to the coherent structure method Over the time interval, the topological length of the left to identify the structures in the ﬂow ﬁeld. The ﬁrst 50 and the right recirculating vortex loops grow by a fac- periods of advection are considered, throughout which 33 tor of 1.24 and 1.48, respectively. Near the left vortex trajectories remained in the imaging domain for the dura- (blue), there are four additional structures identiﬁed at tion of the time interval. These 33 trajectories produce a its perimeter: these structures are plotted in magenta, braid made up of 9,592 generators. Using the same length purple, yellow and dark purple. Each of these features ratio threshold as the analytic model trajectory analysis, is comprised of two or three particles and in all cases, we identiﬁed three coherent sets of particles. Each of the the structures closely orbit the recirculating vortex with- sets is surrounded by a non-growing loop. Applying the out penetrating the boundary. Another type of coherent method to this dataset takes less than 10 seconds. structure identiﬁed is presented in light and dark green in Figure 9(a). These two structures correspond to par- The initial and ﬁnal position of the particles and the ticles that orbit around the chaotic sea and the recir- corresponding coherent structures are presented in Fig- culating vortices. The dark green structure grows by a ure 10. The two main recirculating vortices are identiﬁed. factor of 1.47 while the light green structure shrinks to The set of particles orbiting around the outside of the 0.6 times its initial length due to the particles becoming domain is also identiﬁed. For the experiment, particles closer in the topological projection. Finally, the fourth that pass through the chaotic sea only during the ﬁnal type of structure is identiﬁed in gray in Figure 9(a) near periods also belong to this outer structure because not (1.89,0.13) cm. This structure is made up of two particles enough mixing with the structure has occurred for some 10 of these particles to be excluded from the orbit structure. trast between particles and background ﬂow that is high For all three cases, the braided length of the topological enough for particles to be detected at each time step. The loop |B`| either shrinks or remains the same. The left tracking issue was mostly due to the wide range of veloc- and right recirculating vortex structures correspond to ities between particles located close to the rotor and par- ratios of 1 and 0.86, respectively, while the orbiting struc- ticles in the outer region, which made the construction of ture ratios 0.92, demonstrating that perceivably-coherent trajectories from their respective particles diﬃcult. The structures result in roughly-constant lengths of topolog- other main challenge to obtaining a suﬃcient spatiotem- ical loops. Having demonstrated the coherent structures poral coverage is that as the particle density increases approach to the PT trajectories, the analysis was reit- the particle-particle interactions become more important erated on datasets of synthetic PIV trajectories. Each and can cause the ﬂow to signiﬁcantly deviate from the recirculating vortex was detected with resolution depen- analytic model. Some of the features of interest present dent on the trajectory initialization. The Supplemental in the ﬂow being small-sized, it was therefore diﬃcult to Material Figure [88] presents the coherent structures de- seed the spatial domain to discern all of them. Nonethe- tected from two sample sets of synthetic trajectories, one less, the remaining trajectories were already suﬃcient to with 33 trajectories and one with 100 trajectories. succeed at detecting the largest coherent structures using the braid theory approach. D. Comparison V. ESTIMATING THE GROWTH RATE OF MATERIAL INTERFACES These results demonstrate the ability of the braid the- ory approach to detected coherent structures accurately Newhouse and Yomdin [68–70] established that, in pla- and robustly in an experimental, periodically forced ﬂow. nar ﬂows, the exponential growth rate of material inter- The method found similar qualitative results using the faces is equal to the topological entropy of the ﬂow. In experimental and analytic model trajectories. Despite turn, topological entropy of the ﬂow is bounded below the lower particle density in the experimental system, the by braid entropy, a measure of complexity of braids of coherent structures representing the vortex edges were any periodic trajectories in the ﬂow [55, 58]. Adding tra- detected, even with the structures containing as few as jectories to a braid can only increase the braid entropy, three particles, which shows the reliability of the method therefore a reasonable approximation strategy involves even for sparse datasets. The small-scale features de- computing braid entropies of braids comprising more and tected in the model ﬂow were not identiﬁed from the more trajectories in an eﬀort to narrow the gap between experimental data because too few particles were seeded the braid and topological entropy. Applying this strat- in those regions. egy to experimentally-recorded or simulated trajectories When comparing the braid-based coherent structures is diﬃcult to do directly, as most trajectories are not pe- to the Poincar´ e map for the model ﬂow there are simi- riodic. Nevertheless, Budiˇ si´ c and Thiﬀeault [66] demon- larities and diﬀerences. First, the main recirculating vor- strated that computing a so-called Finite-Time Braiding tices, represented in red and blue for both the model and Exponent (FTBE) provides a reasonable data-driven es- experimental examples, correspond to the islands identi- timate of braid entropy and consequently an estimate ﬁed in Figure 2(b). The second feature identiﬁed is an of the rate of exponential growth of material interfaces. outer structure corresponding to a set of particles that In this section, we apply this procedure to evaluate the are advected at very low velocities around the main vor- degree of correspondence between the experimental and tex system. This outer structure encompassed a much model rotor-oscillator ﬂows. higher proportion of particles for the experimental ﬂow than for the theoretical ﬂow. This may be due to the sparsity of the dataset: if fewer trajectories are present, A. Theory then they may be less likely to entangle with the chaotic sea. The most straightforward way to quantify the com- To improve the resolution of the braid-based coher- plexity of a braid is to count the number of crossings in ent structures, more particle trajectories are needed to the braid, while avoiding counting trivial crossings. For improve the spatiotemporal coverage. In laboratory ex- example, two braids with two strands each: periments, however, there are two main limitations to ac- complishing this. Due to imagery and tracking issues, the trajectories of several experimental particles could not be fully reconstructed for all periods. This is why only 33 out of 80 trajectories were retained for the analysis. The braid approach requires all trajectories to span the both have 3 crossings, since the two central crossings in same time interval and it needs a continuous dataset, i.e., the top braid can be disentangled even with the end- no missing datapoints within trajectories. The imagery points pinned. More rigorously, before counting crossings issue was mostly due to the challenge of having a con- the braid is reduced by employing braid group rules that 11 cancel out such intersections. Deﬁne braid length #B to be the number of intersections after the braid is reduced. Given a topological braid B that corresponds to N N,T trajectories collected over a time interval of duration T , the average braid growth rate of B is N,T #B N,T G(B ) = . (9) N,T FIG. 11. Top: Loop ` used to compute the Finite- The number of intersections alone is a poor indicator Time Braiding Exponent, with Dynnikov vector ` = of complexity. For example, the two 3-stranded braids 0 0 0 0 −1 −1 −1 −1 [ ]. Bottom: The loop after action of the braid from Figure 5, with Dynnikov vector B` = −1 0 8 1 0 −6 −3 7 [ ]. Black dots: strands belonging to the braid; right-most red dot: the auxiliary “anchor” strand. The length of a loop is the number of intersections with the dashed axis (|` | = 4,|B` | = 36). 0 0 have the same length, but the strands in the bottom braid are intertwined in a more intricate manner. Compared to the connection between topological en- The topological entropy of a braid, or simply braid en- tropy and the rate of material growth, the connection tropy, measures the complexity of a braid by quantifying between FTBEs and the rate of material growth is more how fast loops grow as they are slid along the braid (as heuristic. Nevertheless, numerical investigations in a in Figure 6). Braid entropy can be computed by ﬁnding chaotic ﬂow [66] shows that FTBEs can indeed be used the largest rate of growth of a topological loop under a as a reasonable proxy to the topological entropy of a repeated action of the braid [56, 89]. When the physi- braid when in the absence of periodic Lagrangian tra- cal braid comprises periodic trajectories, the entropy of jectories from the ﬂow. Consequently, FTBEs charac- the associated topological braid is a lower bound for the terizes the rate of growth of material lines in this work. topological entropy of the ﬂow. This property was suc- Braid growth G will be used as the independent variable cessfully exploited to create stirring protocols for rod- in comparing FTBE of braids with diﬀerent number of based stirrers [23, 55, 57, 90], to design spatially pe- strands. riodic mixers [91], and to analyze stirring of a passive tracer by compact coherent structures (so-called ghost rods ) [57, 58, 63, 92]. B. Comparison of the model and experimental When trajectories in the physical braid are not peri- ﬂows odic, the theory is less developed compared to the peri- odic case, and published works rely on numerical simu- lations [56, 65]. Nevertheless, when trajectory segments We compare the complexity of numerical trajectories are long enough, it is possible to deﬁne the Finite-Time in the model and experimental velocity ﬁelds. The pa- Braiding Exponent (FTBE) [66], which measures the ex- rameters of the PIV experiment in Table I correspond ponential growth rate of a particular topological loop to a nondimensional oscillation period τ = 5 and non- during a single application of a braid. (By compari- dimensional parameters = 0.1250 and λ = 1.2566. son, topological entropy measures asymptotic exponen- We initialized trajectories in the chaotic region where tial growth with respect to repeated application of the primary and secondary vortices meet (see Figure 4), as braid.) this is where exponential growth of material lines is ex- The Finite-Time Braiding Exponent is given by pected, and simulated them for 30 oscillation periods (T = 30τ = 150). As the PIV velocity ﬁeld was recorded 1 |B ` | N,T 0 FTBE(B ) = log , (10) over 11 oscillation periods (see II B 2), to achieve T = 150 N,T T |` | we repeated three times the ﬁrst 10 oscillation periods. where ` is the topological loop in Figure 11 (top). The Integrating PIV velocity ﬁeld for long times com- particular loop ` is chosen so that any nontrivial braid pounds measurement errors in the velocities, especially results in a nonzero deformation [66]. To ensure this, the in the slow regions of the ﬂow, resulting in unphysical components of the loop are “anchored” on the boundary trajectories, e.g., trajectories that exit the domain. We of the domain, which is topologically equivalent to adding removed such trajectories from the PIV dataset and re- an additional strand (depicted as the hollow circle) that placed them with additional simulated trajectories until does not participate in dynamics. Justiﬁcations for the we obtained a full set of N = 600 well-behaved trajecto- choice of ` is given at length in [93, 94]. ries. We collected the same number of trajectories from Compared to (8) in section IV, the FTBE is calculated the model velocity ﬁeld. Even though the model and PIV for the speciﬁc loop ` instead of a set of simple pair- velocity ﬁelds are qualitatively similar (see Figure 4), the loops. Further information on FTBE calculations can be precise location of the saddle-like point sitting between found in [66]. the primary and secondary vortices is sligthly diﬀerent. 12 −1 Model 0.15 PIV −0.5 0.1 0.15 0.2 0.25 0.1 0.5 0.05 −3 −2 −1 0 1 2 3 −6 −4 −2 0 2 4 6 (a) Samples of analytic model velocity ﬁeld trajectories # Std. dev. from mean (a) FTBE −1 −0.5 0.15 Model PIV 1.2 1.3 1.4 1.5 0.1 0.5 ×10 0.05 −3 −2 −1 0 1 2 3 (b) Samples of PIV velocity ﬁeld trajectories −6 −4 −2 0 2 4 6 FIG. 12. Poincar´ e plots of trajectories in braids used to com- # Std. dev. from mean pare FTBE and braid growth rate between analytic model and (b) Growth rate PIV velocity ﬁelds. Trajectories were initialized uniformly in- side the highlighted rectangles (a) (x, y) ∈ [−0.25,−0.5] × [−0.5, 0.5] for model velocities; (b) (x, y) ∈ [−0.1,−0.5] × FIG. 13. Standardized distributions of FTBEs and growth [−0.5, 0.5] for PIV velocities. The Poincar´ e map is taken at rates with S = 500 samples of n = 300-stranded sub- the rotor’s oscillation period τ . braids (out of N = 600 trajectories, T = 30τ = 150); non- standardized histograms are inset into each graph. To ensure that trajectories are seeded in the chaotic zone in both cases, we used a diﬀerent initialization rectangle shapes of distributions are similar, but with quantita- for the two velocity ﬁelds. Consequently, Poincar´ e plots tive diﬀerences. The model velocity ﬁeld shows a greater in Figure 12 show that both datasets have similar quali- variance in the growth rate, while FTBE distributions tative features in the region of interest. Finally, we con- are non-overlapping, indicating a signiﬁcant separation verted sets of trajectories into braids using the MATLAB of recorded FTBE values. The model velocity ﬁeld may allow initialization of rare-event trajectories more easily, toolbox braidlab [82]. e.g., in intricate periodic islands that would be washed The braids of trajectories are characterized by their out in experimental system, potentially explaining the braid growth rates (9) and FTBE values (10), which re- variance in growth. The distance between FTBE distri- spectively quantify how many non-trivial strand crossings butions appears to be more signiﬁcant, indicating that are generated, and the amount of stretching that these even though the number of strand intersections is com- crossings impose on topological material lines. We ana- parable, the intersections for the model ﬂow generate no- lyze ensembles of S = 500 braids, each with n strands, ticeably more complexity than the experiment. in order to estimate distributions of FTBE and braid growth. By randomly choosing n trajectories out of the Presently, we do not know of a trajectory mechanism full set of N = 600 simulated trajectories, for each num- that would explain this diﬀerence. We do not think that ber of strands n = 3, . . . , N/2 = 300, we create S realiza- the diﬀerence in initialization window (see Figure 12) tions of braids B . By reusing trajectories in several could cause this deviation, as trajectories in both cases n,T braids, fewer trajectories need to be simulated, avoiding occupy the chaotic region surrounding the vortices de- the costly initialization and resimulation of trajectories spite the slight numerical diﬀerences in locations of those regions. for individual realizations. Further discussion of the re- sampling step is in [66]. Additionally, artiﬁcially degrading the quality of the Figure 13 shows histograms of FTBEs and braid model does not change the outcome. We have sampled growth rates for both sets of braids with S = 500 realiza- the model velocity ﬁeld on the same grid as the PIV tions of the largest number of strands n = 300, subsam- velocity ﬁeld, and performed the same space-time inter- pled from the full simulated set of N . From a qualitative polation that was used for the PIV velocity ﬁeld. The comparison with Poincar´ e plots, one would expect simi- resulting interpolated model velocity ﬁeld had the same lar results for both model and PIV velocity ﬁelds. The FTBE values as the continuous model velocity ﬁeld, sug- Proportion Proportion 13 gesting that discretization and interpolation themselves FTBE and braid growth should increase with n. Fig- are not the cause for the lower FTBE values in the PIV ure 15 shows the observed dependence. The growth rate ﬂow. These results are not presented in ﬁgures, as they for both ﬂows increases quadratically with the number of do not diﬀer from the model ﬂow results. strands, which can be explained by noting that n strands Ultimately, the quantitative diﬀerence between FTBE can form = O(n ) pairs. In chaotic regions, statis- values may speak to the sensitivity of the braid dynamics tical quantities tend to mimic those in stochastic pro- techniques, in particular FTBE, to diﬀerences between cesses with independent increments, so over large peri- ﬂows that are otherwise challenging to determine. ods of time any of those possible crossings should In order to exhibit the dependence of distributions of be equally likely. As more pairs are conﬁned in a ﬁnite FTBEs and growth rates on duration T and number of space, we observe a quadratic increase in the number of strands n for braids B , we compute summary statis- crossings for a ﬁxed time interval. n,T tics and graph them against T and n. FTBE distribu- FTBE increases with n in both velocity ﬁelds, follow- tions are represented by their maximum values, as braid ing the shape previously observed in [66] for which there entropy of any braid is a lower bound for the topological is no good mathematical model, as far as we are aware. entropy of the ﬂow. Growth rate distributions are sim- Topological entropy of the ﬂow acts as the upper bound ply represented by mean values, as there are no special for the curve, but it is not clear if the bound is tight. In considerations for this quantity. planar ﬂows, topological entropy corresponds to the max- The maximum FTBE in a sample should provide imal rate of exponential growth of material lines. Since the best-available bound to topological entropy for the FTBE is non-zero, we infer that both ﬂows exhibit suﬃciently-long duration T [66]. Figure 14 demonstrates exponential growth of material interfaces, although the that the values of maximum FTBE and mean growth sta- FTBE estimates of these rates diﬀer quantitatively. We bilize well ahead the end of simulations, which leads us do not know if the diﬀerence is dynamically signiﬁcant. to conclude that studying trajectories of length T = 30τ Future work could contrast FTBE results to those ob- is representative of values in the long-time limit. tained by the recently-developed eTEC technique [64] as an independent comparison. Model (sample max., 300 str.) 0.4 Model (single braid, 600 str.) PIV (sample max., 300 str.) PIV (single braid, 600 str.) 0.2 0.2 0.1 Model PIV 50 100 150 Trajectory duration T 0 200 400 600 Number of strands n (a) Max. FTBE (a) FTBE of sub-braids (T = 150) 1,500 Model PIV 1,000 Model PIV −1 50 100 150 1 2 3 10 10 10 Trajectory duration T Number of strands n (b) Mean growth rate (b) Growth rate of sub-braids (T = 150 ) FIG. 14. Dependence of the maximum FTBE and mean FIG. 15. Eﬀect of increase in braid size (number of strands) growth rate on trajectory duration T . The comparison is be- on FTBE and growth rate of braids of trajectories from model tween the maximum FTBE over a sample of n = 300-stranded and PIV ﬂows. Filled-in marks correspond to maximum of a braids and a single braid of n = 600 strands, all from analytic sample of n-stranded braids. Hollow marks correspond to the model and PIV ﬂows. single 600-stranded braid. Intuitively, the more strands participate in a braid, the higher the potential for complexity. Consequently Finally, we point out some challenges encountered in chaotic regions, where we initialized trajectories, both here. An eﬀective comparison of complexity based on Mean G(B ) n,T max FTBE(B ) n,T G (B ) n N,T max FTBE(B ) N,T 14 braid dynamics requires Lagrangian trajectories that • At least a few particles should be seeded in each cover many periods of oscillations, which is experimen- structure to be detected; tally challenging for several reasons. As explained in Sec- • Quantitative comparison of FTBE values is unreli- tion II B 2, the analytical model assumes that the rotor able in regions without signiﬁcant mixing; is a point vortex and that the Reynolds number is zero. • Reliable calculation of FTBEs requires either long Additionally, for the FTBE to be computed, the motors Lagrangian trajectories, or measured velocity ﬁelds of the rotor and the oscillator must run for many hours, that can be used to generate such trajectories. which inevitably introduces some amount of thermal en- As explained in Sections IV D and V B, we have had ergy into the ﬂow and slightly changes its viscosity over to overcome several diﬃculties in order to obtain a set of time. For these reasons, we only used 10 periods of exper- experimental particle tracks that were both suﬃciently imental data to reconstruct the ﬂow, which required arti- dense and gapless. Other experimental setups may face ﬁcially repeating the velocity ﬁeld to obtain Lagrangian similar challenges, which currently limits the range of ap- trajectories over 30 periods. The chosen periods were plication of methods based on braid dynamics. Nonethe- the ﬁrst 10 full periods after the ﬁrst peak of oscillation, less, in experimental conditions that do satisfy the re- discarding the transient dynamics after the initial accel- quirements, braid dynamics techniques can be competi- eration. tive with alternative techniques for identifying coherent structures or quantifying the amount of mixing in a two- dimensional ﬂow. We do point out that the recently- VI. DISCUSSION developed eTEC technique [64] shows promise in over- coming at least some of the obstacles mentioned here. We highlight three directions for future work to in- We demonstrated that methods based on braid dynam- crease the usefulness of braid dynamics techniques for ics can identify qualitative features of material transport experimental data. First, many more existing (gappy) in model and experimental rotor-oscillator ﬂows. In par- datasets could be processed if there was a technique ticular, Section IV shows that braid dynamics can distin- for ﬁlling-in missing segments in Lagrangian trajecto- guish between nearby transport structures—secondary ries that do not introduce signiﬁcant additional entropy. vortices and a recirculating region—even when only a few Second, fewer strands would be needed if there existed Lagrangian trajectories are seeded in them, with struc- a theoretical model for growth of FTBEs with number tures being detected from as few as two trajectories. For of strands. Finally, the eﬃciency of search for coherent quantities measuring material deformation (FTBE and structures could be improved by considering the geom- growth rate), we demonstrated in Section V that our two etry of the space of topological loops. Some of these qualitatively-similar ﬂows result in similar distributions, directions may be amenable to modern machine learning with parallel trends in their time-evolution. techniques. Based on these results, we expect that the braid dy- namics methods will be useful in situations where only a small number of Lagrangian trajectories from a two- ACKNOWLEDGMENTS dimensional ﬂow is available, either due to experimental challenges, or because only a rough estimate of ﬂow prop- erties is desired. We thank Gustaaf (Guus) Jacobs for his help running During the comparison of model and experimental CFD models for preliminary analysis of the experimental data, it became clear that processed data needs to sat- setup, Nicholas Ouellette and Douglas Kelley for their isfy, at a minimum, the following requirements in order assistance with processing PIV data and Greg Voth for to be successfully processed using computational braid a helpful discussion about the comparison between the dynamics methods: PIV data and the model. This research was supported • There cannot be signiﬁcant gaps in the discrete by the U.S. National Science Foundation (NSF), under time series of particle tracks; Grants No. CMMI-1233935 and AGS-1520825. [1] Julio M. Ottino, The Kinematics of Mixing: Stretch- ical study of transport, mixing and chaos in an unsteady ing, Chaos, and Transport, Cambridge Texts in Applied vortical ﬂow,” Journal of Fluid Mechanics 214, 347–394 Mathematics (Cambridge University Press, Cambridge, (1990). 1989). [5] Rob Sturman, Julio M. Ottino, and Stephen Wiggins, [2] Hassan Aref, “Stirring by chaotic advection,” Journal of The Mathematical Foundations of Mixing, Cambridge Fluid Mechanics 143, 1–21 (1984). Monographs on Applied and Computational Mathemat- [3] R S MacKay, J D Meiss, and I C Percival, “Transport in ics, Vol. 22 (Cambridge University Press, Cambridge, Hamiltonian systems,” Physica D. Nonlinear Phenomena 2006). 13, 55–81 (1984). [6] S. Childress and A.D. Gilbert, Stretch, Twist, Fold: The [4] V. Rom-Kedar, A. Leonard, and S. Wiggins, “An analyt- Fast Dynamo, Lecture Notes in Physics Monographs 15 (Springer Berlin Heidelberg, 1995). tion of rod-stirring devices,” SIAM Review 53, 723–743 [7] J. D. Meiss, “Thirty years of turnstiles and transport,” (2011). Chaos: An Interdisciplinary Journal of Nonlinear Science [24] S. Wiggins and J. M. Ottino, “Foundations of chaotic 25, 097602 (2015). mixing,” Philosophical Transactions of the Royal Society [8] Sanjeeva Balasuriya, Barriers and Transport in Unsteady of London Series a-Mathematical Physical and Engineer- Flows: A Melnikov Approach (SIAM, 2016). ing Sciences 362, 937–970 (2004). [9] Erik M. Bollt and Naratip Santitissadeekorn, Applied [25] Emmanuel Villermaux, “Mixing Versus Stirring,” Annual and Computational Measurable Dynamics, Mathemati- Review of Fluid Mechanics 51, 245–273 (2019). cal Modeling and Computation, Vol. 18 (Society for In- [26] Jean-Luc Thiﬀeault and Allen H. Boozer, “Geometrical dustrial and Applied Mathematics (SIAM), Philadelphia, constraints on ﬁnite-time Lyapunov exponents in two and PA, 2013). three dimensions,” Chaos: An Interdisciplinary Journal [10] A. Leonard, V. Rom-Kedar, and S. Wiggins, “Fluid mix- of Nonlinear Science 11, 16–28 (2001). ing and dynamical systems,” Nuclear Physics B - Pro- [27] Jean-Luc Thiﬀeault, “Stretching and curvature of mate- ceedings Supplements 2, 179–190 (1987). rial lines in chaotic ﬂows,” Physica D. Nonlinear Phe- [11] Hassan Aref, John R. Blake, Marko Budiˇ si´ c, Silvana S.S. nomena 198, 169–181 (2004). Cardoso, Julyan H.E. Cartwright, Herman J.H. Clercx, [28] George Mathew, Igor Mezi´ c, and Linda Petzold, “A mul- Kamal El Omari, Ulrike Feudel, Ramin Golestanian, Em- tiscale measure for mixing,” Physica D. Nonlinear Phe- manuelle Gouillart, GertJan F. van Heijst, Tatyana S. nomena 211, 23–46 (2005). Krasnopolskaya, Yves Le Guer, Robert S. MacKay, [29] J.-L. Thiﬀeault, “Scalar decay in chaotic mixing,” Trans- Vyacheslav V. Meleshko, Guy Metcalfe, Igor Mezi´ c, port and Mixing in Geophysical Flows Lecture Notes in Alessandro P.S. de Moura, Oreste Piro, Michel F. M. Physics, 3–36 (2008). Speetjens, Rob Sturman, Jean-Luc Thiﬀeault, and Idan [30] J.-L. Thiﬀeault, “Using multiscale norms to quantify Tuval, “Frontiers of chaotic advection,” Reviews of Mod- mixing and transport,” Nonlinearity 25, R1–R44 (2012), ern Physics 89, 025007 (2017). arXiv:1105.1101. [12] Melissa A. Green, Clarence W. Rowley, and Alexander J. [31] George Haller, “Lagrangian Coherent Structures,” An- Smits, “Using hyperbolic Lagrangian coherent structures nual Review of Fluid Mechanics 47, 137–162 (2015). to investigate vortices in bioinspired ﬂuid ﬂows,” Chaos: [32] Gary Froyland and Kathrin Padberg, “Almost-invariant An Interdisciplinary Journal of Nonlinear Science 20, sets and invariant manifolds—connecting probabilistic 017510 (2010). and geometric descriptions of coherent structures in [13] Shawn C. Shadden, John O. Dabiri, and Jerrold E. Mars- ﬂows,” Physica D. Nonlinear Phenomena 238, 1507–1523 den, “Lagrangian analysis of ﬂuid transport in empirical (2009). vortex ring ﬂows,” Physics of Fluids 18, 047105 (2006). [33] Thomas Peacock and George Haller, “Lagrangian co- [14] Kakani Katija and John O. Dabiri, “A viscosity-enhanced herent structures: The hidden skeleton of ﬂuid ﬂows,” mechanism for biogenic ocean mixing,” Nature 460, 624– Physics Today 66, 41–47 (2013). 626 (2009). [34] Roger M. Samelson, “Lagrangian Motion, Coherent [15] Amanda J. Tan, Eric Roberts, Kevin A. Mitchell, and Structures, and Lines of Persistent Material Strain,” An- Linda S. Hirst, “Investigating Quality of Mixing of a Bi- nual Review of Marine Science 5, 137–163 (2013). ological Active Nematic,” Biophysical Journal 114, 650a [35] Michael R. Allshouse and Thomas Peacock, “Lagrangian (2018). based methods for coherent structure detection,” Chaos: [16] Jean-Luc Thiﬀeault, “Chaos in the Gulf,” Science Mag- An Interdisciplinary Journal of Nonlinear Science 25, azine (2010). 097617 (2015). [17] Igor Mezi´ c, Sophie Loire, Vladimir A Fonoberov, and [36] Alireza Hadjighasem, Mohammad Farazmand, Daniel Patrick J Hogan, “A New Mixing Diagnostic and Gulf Oil Blazevski, Gary Froyland, and George Haller, “A critical Spill Movement,” Science Magazine 330, 486–489 (2010). comparison of Lagrangian methods for coherent structure [18] Mar´ ıa J. Olascoaga and George Haller, “Forecasting sud- detection,” Chaos: An Interdisciplinary Journal of Non- den changes in environmental pollution patterns,” Pro- linear Science 27, 053104 (2017). ceedings of the National Academy of Sciences 109, 4738– [37] Sanjeeva Balasuriya, Nicholas T. Ouellette, and Irina I. 4743 (2012). Rypina, “Generalized Lagrangian coherent structures,” [19] M. Weldon, T. Peacock, G. B. Jacobs, M. Helu, and Physica D: Nonlinear Phenomena 372, 31–51 (2018). G. Haller, “Experimental and numerical investigation of [38] Michael R. Allshouse, Gregory N. Ivey, Ryan J. Lowe, the kinematic theory of unsteady separation,” Journal of Nicole L. Jones, C. J. Beegle-Krause, Jiangtao Xu, and Fluid Mechanics 611, 1–11 (2008). Thomas Peacock, “Impact of windage on ocean surface [20] Yong Wang, George Haller, Andrzej Banaszuk, and Lagrangian coherent structures,” Environmental Fluid Gilead Tadmor, “Closed-loop Lagrangian separation con- Mechanics 17, 473–483 (2017). trol in a bluﬀ body shear ﬂow model,” Physics of Fluids [39] Francisco J. Beron-Vera, Mar´ ıa J. Olascoaga, George 15, 2251–2266 (2003). Haller, Mohammad Farazmand, Joaqu´ ın Trinane ˜ s, and [21] B Noack, Igor Mezi´ c, G Tadmor, and Andrzej Banaszuk, Yan Wang, “Dissipative inertial transport patterns near “Optimal mixing in recirculation zones,” Physics of Flu- coherent Lagrangian eddies in the ocean,” Chaos: An In- ids (2004). terdisciplinary Journal of Nonlinear Science 25, 087412 [22] George Mathew, Igor Mezi´ c, Symeon Grivopoulos, (2015). Umesh Vaidya, and Linda Petzold, “Optimal control of [40] P. Tallapragada, S. D. Ross, and D. G. Schmale, “La- mixing in Stokes ﬂuid ﬂows,” Journal of Fluid Mechanics grangian coherent structures are associated with ﬂuctua- (2007). tions in airborne microbial populations,” Chaos: An In- [23] M. D. Finn and J.-L. Thiﬀeault, “Topological optimiza- terdisciplinary Journal of Nonlinear Science 21, 033122 16 (2011). “Topological ﬂuid mechanics of stirring,” Journal of Fluid [41] Amir E. BozorgMagham, Shane D. Ross, and David G. Mechanics 403, 277–304 (2000). Schmale III, “Real-time prediction of atmospheric La- [56] Jean-Luc Thiﬀeault, “Measuring Topological Chaos,” grangian coherent structures based on forecast data: An Physical Review Letters 94, 084502 (2005). application and error analysis,” Physica D: Nonlinear [57] Jean-Luc Thiﬀeault and Matthew D. Finn, “Topology, Phenomena 258, 47–60 (2013). braids and mixing in ﬂuids,” Philosophical Transactions [42] George Haller and Themistoklis Sapsis, “Lagrangian co- of the Royal Society A: Mathematical, Physical and En- herent structures and the smallest ﬁnite-time Lyapunov gineering Sciences 364, 3251–3266 (2006). exponent,” Chaos: An Interdisciplinary Journal of Non- [58] Emmanuelle Gouillart, Jean-Luc Thiﬀeault, and linear Science 21, 023115 (2011). Matthew D. Finn, “Topological mixing with ghost rods,” [43] Marko Budiˇ si´ c, Stefan Siegmund, Doan Thai Son, and Physical Review E. Statistical, Nonlinear, and Soft Mat- Igor Mezi´ c, “Mesochronic classiﬁcation of trajectories in ter Physics 73, 036311, 8 (2006). incompressible 3D vector ﬁelds over ﬁnite times,” Dis- [59] Matthew D Finn and Jean-Luc Thiﬀeault, “Topological crete and Continuous Dynamical Systems - Series S 9, entropy of braids on the torus,” SIAM Journal on Applied 923–958 (2016). Dynamical Systems 6, 79–98 (electronic) (2007). [44] Tian Ma and Erik M. Bollt, “Shape Coherence and [60] Nicolas Francois, Hua Xia, Horst Punzmann, Benjamin Finite-Time Curvature Evolution,” International Journal Faber, and Michael Shats, “Braid Entropy of Two- of Bifurcation and Chaos 25, 1550076 (2015). Dimensional Turbulence,” Scientiﬁc Reports 5, 18564 [45] Gary Froyland and Kathrin Padberg-Gehle, “Finite-time (2015). entropy: A probabilistic approach for measuring nonlin- [61] S. Candelaresi, D. I. Pontin, and G. Hornig, “Quantify- ear stretching,” Physica D: Nonlinear Phenomena 241, ing the tangling of trajectories using the topological en- 1612–1628 (2012). tropy,” Chaos: An Interdisciplinary Journal of Nonlinear [46] Dean Roemmich, Gregory C. Johnson, Stephen Riser, Science 27, 093102 (2017). Russ Davis, John Gilson, W. Brechner Owens, Silvia L. [62] C. K. Taylor and Stefan G. Llewellyn Smith, “Dy- Garzoli, Claudia Schmid, and Mark Ignaszewski, “The namics and transport properties of three surface quasi- Argo Program: Observing the Global Ocean with Proﬁl- geostrophic point vortices,” Chaos: An Interdisciplinary ing Floats,” Oceanography 22, 34–43 (2009). Journal of Nonlinear Science 26, 113117 (2016). [47] Ma¨ elle Nodet, “Variational assimilation of Lagrangian [63] Piyush Grover, Shane D Ross, Mark A Stremler, and data in oceanography,” Inverse Problems 22, 245–263 Pankaj Kumar, “Topological chaos, braiding and bi- (2006). furcation of almost-cyclic sets,” Chaos: An Interdisci- [48] Emanuel F. Coelho, P. Hogan, G. Jacobs, P. Thoppil, plinary Journal of Nonlinear Science 22, 043135–043135– H. S. Huntley, B. K. Haus, B. L. Lipphardt, A. D. Kir- 16 (2012). wan, E. H. Ryan, J. Olascoaga, F. Beron-Vera, A. C. [64] Eric Roberts, Suzanne Sindi, Spencer A. Smith, and Poje, A. Griﬀa, T. M. Ozg¨ okmen, A. J. Mariano, G. Nov- Kevin A. Mitchell, “Ensemble-based topological entropy elli, A. C. Haza, D. Bogucki, S. S. Chen, M. Curcic, calculation (E-tec),” Chaos: An Interdisciplinary Journal M. Iskandarani, F. Judt, N. Laxague, A. J. H. M. Re- of Nonlinear Science 29, 013124 (2019). niers, A. Valle-Levinson, and M. Wei, “Ocean current [65] Jean-Luc Thiﬀeault, “Braids of entangled particle trajec- estimation using a Multi-Model Ensemble Kalman Fil- tories,” Chaos: An Interdisciplinary Journal of Nonlinear ter during the Grand Lagrangian Deployment experiment Science 20, 017516–017514 (2010). (GLAD),” Ocean Modelling 87, 86–106 (2015). [66] Marko Budiˇ si´ c and Jean-Luc Thiﬀeault, “Finite-time [49] Michael Allshouse and Jean-Luc Thiﬀeault, “Detecting braiding exponents,” Chaos: An Interdisciplinary Jour- coherent structures using braids,” Physica D. Nonlinear nal of Nonlinear Science 25, 087407 (2015). Phenomena , 95–105 (2012). [67] Philip Boyland, “Topological methods in surface dynam- [50] Gary Froyland and Kathrin Padberg-Gehle, “A rough- ics,” Topology and its Applications 58, 223–298 (1994). and-ready cluster-based approach for extracting ﬁnite- [68] Sheldon Newhouse and Thea Pignataro, “On the estima- time coherent sets from sparse and incomplete trajectory tion of topological entropy,” Journal of Statistical Physics data,” Chaos: An Interdisciplinary Journal of Nonlinear 72, 1331–1351 (1993). Science 25, 087406 (2015). [69] Sheldon E. Newhouse, “Entropy and volume,” Ergodic [51] Alireza Hadjighasem, Daniel Karrasch, Hiroshi Ter- Theory and Dynamical Systems 8, 283–299 (1988). amoto, and George Haller, “Spectral-clustering approach [70] Y. Yomdin, “Volume growth and entropy,” Israel Journal to Lagrangian vortex detection,” Physical Review E 93, of Mathematics 57, 285–300 (1987). 063107 (2016). [71] Ditza Auerbach, Predrag Cvitanovi´ c, Jean-Pierre Eck- [52] K. L. Schlueter-Kuck and J. O. Dabiri, “Coherent struc- mann, Gemunu Gunaratne, and Itamar Procaccia, “Ex- ture colouring: identiﬁcation of coherent structures from ploring chaotic motion through periodic orbits,” Physical sparse data using graph theory,” Journal of Fluid Me- Review Letters 58, 2387–2389 (1987). chanics 811, 468–486 (2017). [72] Ruslan L. Davidchack, Ying-Cheng Lai, Aaron Kle- [53] I. I. Rypina and L. J. Pratt, “Trajectory encounter vol- banoﬀ, and Erik M. Bollt, “Towards complete detection ume as a diagnostic of mixing potential in ﬂuid ﬂows,” of unstable periodic orbits in chaotic systems,” Physics Nonlin. Processes Geophys. 24, 189–202 (2017). Letters A 287, 99–104 (2001). [54] Matthew O. Williams, Irina I. Rypina, and Clarence W. [73] Ruslan L. Davidchack, Ying-Cheng Lai, Erik M. Bollt, Rowley, “Identifying ﬁnite-time coherent sets from lim- and Mukeshwar Dhamala, “Estimating generating par- ited quantities of Lagrangian data,” Chaos: An Interdis- titions of chaotic systems by unstable periodic orbits,” ciplinary Journal of Nonlinear Science 25, 087408 (2015). Physical Review E 61, 1353–1356 (2000). [55] Philip L Boyland, Hassan Aref, and Mark A Stremler, [74] Ruslan L. Davidchack and Ying-Cheng Lai, “Eﬃcient al- 17 gorithm for detecting unstable periodic orbits in chaotic 114101 (2011). systems,” Physical Review E 60, 6172–6175 (1999). [93] Ivan Dynnikov and Bert Wiest, “On the complexity of [75] Gary Froyland, Oliver Junge, and Gunter Ochs, “A braids,” Journal of the European Mathematical Society rough-and-ready cluster-based approach for extracting (JEMS) 9, 801–840 (2007). ﬁnit-time coherent sets from sparse and incomplete tra- [94] Patrick Dehornoy, “Eﬃcient solutions to the braid iso- jectory data,” Chaos 25, 68–84 (2001). topy problem,” Discrete Applied Mathematics Appli- [76] Erik M. Bollt, Theodore Stanford, Ying-Cheng Lai, and cations of Algebra to Cryptography, 156, 3091–3112 (2008). Karol Zyczkowski, “What symbolic dynamics do we get with a misplaced partition?: On the validity of thresh- old crossings analysis of chaotic time-series,” Physica D: Nonlinear Phenomena 154, 259–286 (2001). [77] W. W. Hackborn, M. E. Ulucakli, and T. Yuster, “A theoretical and experimental study of hyperbolic and de- generate mixing regions in a chaotic Stokes ﬂow,” Journal of Fluid Mechanics 346, 23–48 (1997). [78] W. W. Hackborn, “Asymmetric Stokes ﬂow between par- allel planes due to a rotlet,” Journal of Fluid Mechanics 218, 531–546 (1990). [79] Daniel Blair and Eric Dufresne, “The matlab parti- cle tracking code repository,” http://site.physics. georgetown.edu/matlab/ (1999). [80] Douglas H. Kelley and Nicholas T. Ouellette, “Onset of three-dimensionality in electromagnetically driven thin- layer ﬂows,” Physics of Fluids 23, 045103 (2011). [81] Lawrence F. Shampine, Mark W. Reichelt, and Jacek A. Kierzenka, “Solving index-i daes in matlab and simulink,” SIAM Review 41, 538–552 (1999). [82] Jean-Luc Thiﬀeault and Marko Budiˇ si´ c, “Braidlab: A Software Package for Braids and Loops,” arXiv:1410.0849 [math] (2014), version 3.2.2, arXiv:1410.0849 [math]. [83] I. A. Dynnikov, “On a Yang-Baxter map and the De- hornoy ordering,” Russian Mathematical Surveys 57, 592 (2002). [84] Toby Hall and S. Oyku ¨ Yurtta¸ s, “On the topological en- tropy of families of braids,” Topology and its Applica- tions 156, 1554–1564 (2009). [85] Z.P. You, E. J. Kostelich, and J. A. Yorke, “Calculating stable and unstable manifolds,” International Journal of Bifurcation and Chaos 1, 605–623 (1991). [86] George Haller, “Distinguished material surfaces and co- herent structures in three-dimensional ﬂuid ﬂows,” Phys- ica D: Nonlinear Phenomena 149, 248–277 (2001). [87] Gary Froyland, Simon Lloyd, and Naratip Santitis- sadeekorn, “Coherent sets for nonautonomous dynamical systems,” Physica D: Nonlinear Phenomena 239, 1527– 1541 (2010). [88] See Supplemental Material at [URL will be inserted by publisher] for the Figure presenting the coherent struc- tures detected from two sample sets of synthetic trajec- tories. [89] Jacques-Olivier Moussaﬁr, “On computing the entropy of braids,” Functional Analysis and Other Mathematics 1, 37–46 (2006). [90] Jean-Luc Thiﬀeault, “The mathematics of taﬀy pullers,” Mathematical Intelligencer 40, 26–35 (2018), https://arxiv.org/abs/1608.00152. [91] Matthew D. Finn, Jean-Luc Thiﬀeault, and Emmanuelle Gouillart, “Topological chaos in spatially periodic mix- ers,” Physica D: Nonlinear Phenomena 221, 92–100 (2006). [92] Mark A Stremler, Shane D Ross, Piyush Grover, and Pankaj Kumar, “Topological Chaos and Periodic Braid- ing of Almost-Cyclic Sets,” Physical Review Letters 106, Supplementary material to “Using braids to quantify interface growth and coherence in a rotor-oscillator ﬂow” Margaux Filippi, Marko Budiˇ si´ c, Michael R. Allshouse, S´ everine Atis, Jean-Luc Thiﬀeault, and Thomas Peacock (Dated: May 26, 2020) The particle tracking experiment directly provided the set of trajectories analyzed to ﬁnd coherent structures. It is also possible to detect coherent structures from Par- ticle Image Velocimetry (PIV) by generating a synthetic set of trajectories from the measured velocity ﬁeld. Using the same 10 periods of data used to create the synthetic trajectories for the FTBE analysis, two sets of 33 and 100 trajectories are generated. To prepare the initial conidtions, particles are uniformly seeded in −1.5 < x < 1.5, −0.5 < y < 0.2, then advected for 5 initial periods to minimize the inﬂuence of the initial seeding. After such initialization, the trajectories are recorded for 50 periods of wall-oscillation and converted to braids, which are then used in detection of coherent structures as explained in Sec. IV of the paper. The resulting coherent structures and the network of non- growing pair loops are presented in Figure 1. The smaller data set produces a similar result to the one found using the particle tracking trajectories where each vortex contains only two particles. The larger dataset include many particles in each of the recircu- lating vortex. The perceived increase in the vortex foot- FIG. 1. Coherent structures analysis on datasets of synthetic print demonstrates the importance of highly sampling PIV trajectories. Each recirculating vortex was detected with the coherent structure to get a better approximation for resolution dependent on the trajectory initialization. (a) The the feature. analysis of 33 trajectories detected a blue coherent structure within the left recirculating vortex and a red coherent struc- ture within the right recirculating vortex, each structure con- taining 2 trajectories. (b) The analysis of 100 trajectories detected a blue coherent structure on the left with 13 trajec- tories, a red coherent structure on the right with 5 trajectories and a green coherent structure within the right vortex with 2 trajectories. margaux@mit.edu arXiv:1910.12779v2 [physics.flu-dyn] 23 May 2020

Journal

Nonlinear Sciences – arXiv (Cornell University)

Published: Oct 28, 2019

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Using braids to quantify interface growth and coherence in a rotor-oscillator flow

Using braids to quantify interface growth and coherence in a rotor-oscillator flow

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Using braids to quantify interface growth and coherence in a rotor-oscillator flow

Using braids to quantify interface growth and coherence in a rotor-oscillator flow

References (129)

Abstract

Journal

Recommended Articles

There are no references for this article.

Our policy towards the use of cookies