1: %\documentclass[aps,pre,preprint,groupedaddress]{revtex4}
2: \documentclass[pre,twocolumn,showpacs,amsmath,amssymb,floatfix]{revtex4}
3: %\documentclass[pre,preprint,showpacs,amsmath,amssymb,floatfix]{revtex4}
4:
5: \usepackage{graphicx}% Include figure files
6: \usepackage{dcolumn} % Align table columns on decimal point
7: \usepackage{bm} % Bold math
8: \usepackage{color}
9:
10: \def\AD#1{{\textcolor{magenta}{#1}}}
11: % COLOR %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
12: \def\NOTE#1{{\textcolor{red}{\bf [#1]}}} % note
13: \def\DEL#1{{\textcolor{green}{#1}}} % suggested deletions
14: \def\ADD#1{{\textcolor{blue}{#1}}} % addition
15: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
16:
17: %\topmargin -6pt
18: \begin{document}
19:
20: \title{Non-local interactions in hydrodynamic turbulence at high Reynolds
21: numbers: \\ the slow emergence of scaling laws}
22: \author{P.D. Mininni$^{1,2}$, A. Alexakis$^3$, and A. Pouquet$^2$}
23: \affiliation{$^1$ Departamento de F\'\i sica, Facultad de Ciencias Exactas y
24: Naturales, Universidad de Buenos Aires, Ciudad Universitaria, 1428
25: Buenos Aires, Argentina. \\
26: $^2$ NCAR, P.O. Box 3000, Boulder, Colorado 80307-3000, U.S.A.\\
27: $^3$ Laboratoire Cassiop\'ee, Observatoire de la C\^ote d'Azur,
28: BP 4229, Nice Cedex 04, France.}
29: \date{\today}
30:
31: \begin{abstract}
32: We analyze the data stemming from a forced incompressible
33: hydrodynamic simulation on a grid of $2048^3$ regularly spaced
34: points, with a Taylor Reynolds number of $R_{\lambda}\sim 1300$. The
35: forcing is given by the Taylor-Green flow, which shares similarities
36: with the flow in several laboratory experiments, and the computation
37: is run for ten turnover times in the turbulent steady state. At this
38: Reynolds number the anisotropic large scale flow pattern, the
39: inertial range, the bottleneck, and the dissipative range are
40: clearly visible, thus providing a good test case for the study of
41: turbulence as it appears in nature. Triadic interactions, the
42: locality of energy fluxes, and structure functions of the velocity
43: increments are computed. A comparison with runs at lower Reynolds
44: numbers is performed, and shows the emergence of scaling laws for
45: the relative amplitude of local and non-local interactions in
46: spectral space. The scalings of the Kolmogorov constant, and of
47: skewness and flatness of velocity increments, performed as well and
48: are consistent with previous experimental results. Furthermore,
49: the accumulation of energy in the small-scales associated with the
50: bottleneck seems to occur on a span of wavenumbers that is
51: independent of the Reynolds number, possibly ruling out an inertial
52: range explanation for it. Finally, intermittency exponents seem to
53: depart from standard models at high $R_{\lambda}$, leaving the
54: interpretation of intermittency an open problem.
55: \end{abstract}
56: \pacs{47.27.ek; 47.27.Ak; 47.27.Jv; 47.27.Gs}
57: \maketitle
58:
59: \section{Introduction}
60: Turbulence prevails in the universe, and its multi-scale properties
61: affect the global dynamics of geophysical and astrophysical flows at
62: large scale, e.g. through a non-zero energy dissipation even at
63: very high Reynolds number $R_e$. Furthermore, small-scale strong
64: intermittent events, such as the emergence of tornadoes and hurricanes
65: in atmospheric flows, may be very disruptive to the global dynamics
66: and to the structure of turbulent flows. Typically energy is supplied
67: to the flows in the large scales, e.g., by a large scale instability.
68: The flow at these scales is inhomogeneous and anisotropic. In
69: the standard picture of turbulence, the energy cascades to smaller
70: scales due to the stretching of vortices by interactions with similar
71: size eddies. It is then believed that at sufficiently small scales
72: the statistics of the flow are independent of the exact forcing
73: mechanism, and as a result, its properties are universal. For this
74: reason, typical investigations of turbulence consider flows that are
75: forced in the large scales by a random statistically isotropic and
76: homogeneous body force \cite{Haugen06,Kaneda03}. However, how fast
77: (and for which measured quantities) is isotropy, homogeneity, and
78: universality obtained is still an open question.
79:
80: The return to isotropy has been investigated thoroughly in the past,
81: by analysis of data from experiments and direct numerical simulations
82: (DNS) \cite{Sreenivasan97,Shen00,Pope,Kurien00,Biferale01a,Biferale02}.
83: However, lack of computational power limited the numerical investigations
84: of anisotropic forced flows to moderate Reynolds numbers, for which a
85: clear distinction of the inertial range from the bottleneck, and
86: from the dissipative range, cannot be made. Only recently the fast
87: increase of computational power permitted DNS to resolve
88: sufficiently small scales, such that a flow due to an inhomogeneous
89: and anisotropic forcing develops a clear inertial range with constant
90: energy flux. As a result, this kind of questions can be addressed
91: anew. To give an estimate of the size of the desired grid,
92: we mention that in recent simulations \cite{Mininni06} an
93: incipient inertial range was achieved for a resolution of $1024^3$
94: grid points, while for a $512^3$ run the range of scales between the
95: large scale forcing and the bottleneck was much less than an order
96: of magnitude. In all cases, the flow was resolved since
97: $k_{max}/k_\eta \gtrsim 1$, with $k_{max}$ the maximum wavenumber in the
98: simulation and $k_\eta$ the dissipation wavenumber built on the Kolmogorov
99: phenomenology.
100:
101: Of particular interest in the study of turbulent flows is the issue
102: of universality. It is now known that two-dimensional turbulence
103: possesses classes of universality \cite{Bernard06}, and at least for
104: linear systems such as the advection of a passive tracer, there is
105: evidence of universality of the scaling exponents of the
106: fluctuations \cite{Biferale04}. However, recent numerical
107: simulations of three dimensional turbulence \cite{Mininni06} showed
108: that scaling exponents of two different flows (one non-helical, the
109: other fully helical) were measurably different at similar Reynolds
110: number. It is yet unclear whether this is an effect of anisotropies
111: in the flow, or of a finite Reynolds numbers. If this is a finite
112: Reynolds number effect, one then needs to ask how fast its
113: convergence to the universal value is obtained. If the convergence
114: rate is sufficiently slow then finite Reynolds effects should be
115: considered when studying turbulent flows that appear in nature, at
116: very large but finite Reynolds numbers. Thus, the question of the
117: universal properties of turbulent flows at high Reynolds numbers
118: remains somewhat open.
119:
120: The recovery of isotropy, the differences observed in the scaling
121: exponents, and the slow emergence of scaling laws have been recently
122: considered in the context of the influence of the large scales on
123: the properties of turbulent fluctuations
124: \cite{Laval01,Alexakis05b,Mininni06}. The study of nonlocal interactions
125: between large and small scales has been carried in experiments and in
126: simulations
127: \cite{Domaradzki88,Domaradzki90,Kerr90,Yeung91,Okhitani92,Zhou93,Zhou93b,Brasseur94,Yeung95,Zhou96,Kishida99,Carlier01,Verma05,Alexakis05b,Mininni06,Poulain06}
128: at small and moderate Reynolds numbers. In simulations with $1024^3$
129: grid points \cite{Alexakis05b}, it was found that although most of the
130: flux is due to local interactions, non-local interactions with the large
131: scale flow are responsible for $\approx 20 \%$ of the total flux. It is
132: however unclear how the amplitude of these interactions scale with the
133: Reynolds number.
134:
135: In this context, we solve numerically the equations for an incompressible
136: fluid with constant mass density. The Navier-Stokes equation reads
137: \begin{equation}
138: \partial_t {\bf u} + {\bf u}\cdot \nabla {\bf u} = - \nabla {\cal P}
139: + \nu \nabla^2 {\bf u} +{\bf F} ,
140: \label{eq:momentum} \end{equation}
141: with $\nabla \cdot {\bf u} =0 $,
142: where ${\bf u}$ is the velocity field, ${\cal P}$ is the pressure divided
143: by the mass density, and $\nu$ is the kinematic viscosity. Here, ${\bf F}$
144: is an external force that drives the turbulence. The mode with the largest
145: wavevector in the Fourier transform of ${\bf F}$ is defined as $k_F$, with
146: the forcing scale given by $2 \pi / k_F$. We also define the viscous
147: dissipation wavenumber as $k_\eta=(\epsilon/\nu^3)^{1/4}$, where
148: $\epsilon$ is the energy injection rate (as a result, the Kolmogorov scale
149: is $\eta = 2\pi/k_\eta$).
150:
151: The results in the following sections stem from the analysis of a series
152: of DNS of Eq. (\ref{eq:momentum}) using a parallel pseudospectral
153: code in a three dimensional box of size $2\pi$ with periodic boundary
154: conditions, up to a resolution of $2048^3$ grid points. The
155: equations are evolved in time using a second order Runge-Kutta method,
156: and the code uses the $2/3$-rule for dealiasing. As a result, the
157: maximum wavenumber is $k_{max} = N/3$ where $N$ is the number of grid
158: points in each direction.
159:
160: With $L$ and $\lambda$ defined as
161: \begin{equation}
162: L = 2\pi \frac{\int{E(k) k^{-1} dk}}{\int{E(k) dk}}, \ \ \
163: \lambda = 2\pi \left(\frac{\int{E(k) dk}}{\int{E(k) k^2 dk}}\right)^{1/2},
164: \label{eq:integral} \end{equation}
165: the integral scale and Taylor scale respectively, the Reynolds
166: number is $R_e = UL/\nu$ and the Taylor based Reynolds number is
167: $R_\lambda = U\lambda/\nu$. Here, $U=\left< {\bf u}^2 \right>^{1/2}$
168: is the r.m.s. velocity and $E(k)$ the energy spectrum. The large scale
169: turnover time is $T=U/L$. Note that, with these definitions, $R_e$ and
170: $R_\lambda$ used in this paper are larger than the ones stemming
171: from the definitions used by the experimental community (see e.g.,
172: \cite{Frisch}) by a factor of $2 \pi (3/5)^{1/2}\approx 4.87$.
173:
174: \begin{table}
175: \caption{\label{table:runs}Parameters used in the simulations. $N$ is
176: the linear grid resolution, $\nu$ the kinematic viscosity, $R_e$
177: the Reynolds number, and $R_\lambda$ the Taylor based Reynolds
178: number.}
179: \begin{ruledtabular}
180: \begin{tabular}{ccccc}
181: Run & $N$ & $\nu$ & $R_e$ & $R_\lambda$ \\
182: \hline
183: I & 256 &$2\times 10^{-3}$ & 675 & 300 \\
184: II & 512 &$1.5\times 10^{-3}$& 875 & 350 \\
185: III & 1024 &$3\times 10^{-4}$ & 3950 & 800 \\
186: IV & 2048 &$1.2\times 10^{-4}$& 9970 & 1300
187: \end{tabular}
188: \end{ruledtabular}
189: \end{table}
190:
191: Simulations were done with the same external forcing (see Table
192: \ref{table:runs} for the parameters of all the runs), with $U\approx 1$
193: in all steady states. The forcing ${\bf F}$ corresponds to a Taylor-Green
194: (TG) flow \cite{Taylor37}
195: {\setlength\arraycolsep{2pt}
196: \begin{eqnarray}
197: {\bf F} &=& f_0 \left[ \sin(k_F x) \cos(k_F y)
198: \cos(k_F z) \hat{x} - \right. {} \nonumber \\
199: && {} \left. - \cos(k_F x) \sin(k_F y)
200: \cos(k_F z) \hat{y} \right] ,
201: \label{eq:TG}
202: \end{eqnarray}}
203: \noindent
204: where $f_0$ is the forcing amplitude, and $k_F=2$. The
205: turbulent flow that results has no net helicity, although local
206: regions with strong positive and negative helicity develop.
207:
208: \section{\label{sec:2048}The slow emergence of a Kolmogorov-like scaling}
209:
210: \begin{figure}
211: \includegraphics[width=8cm]{fig1}
212: \caption{(a) Energy spectrum in run IV compensated by $k^{-5/3}$. The
213: inset shows the energy flux. (b) Energy spectrum in runs III (dotted)
214: and IV (solid) compensated by $k^{-4/3}$. Wavenumbers are normalized
215: by the dissipation wavenumber $k_\eta$.}
216: \label{fig:spectrum} \end{figure}
217:
218: We first concentrate on the global dynamics of the $2048^3$ run (run IV).
219: Figure \ref{fig:spectrum}(a) shows the compensated energy spectrum in this
220: run, as well as the corresponding energy flux $\Pi(k)$, both taken in the
221: turbulent steady state after the initial transient. The energy flux is
222: constant in a wide range of scales, as expected in a Kolmogorov cascade,
223: but the compensated spectrum has a more complex structure in that same
224: range of scales. The salient features of this spectrum are well-known from
225: previous studies. Small scales before the dissipative range show the
226: so-called bottleneck effect with a slope shallower than $k^{-5/3}$. On
227: the other hand, larger scales have a spectrum with a slope slightly steeper
228: than $k^{-5/3}$, an effect that is even clearer in the simulation performed
229: at larger spatial resolution \cite{Kaneda03} on a grid of $4096^3$ points;
230: this small discrepancy with a Kolmogorov spectrum is attributed to
231: intermittency, i.e. to the spatial scarcity of strong events leading to
232: non-Gaussian wings in the probability distribution functions of velocity
233: gradients.
234:
235: The bottleneck effect is not fully understood but clearly corresponds to
236: an accumulation of energy at the onset of the dissipation range. It
237: has been attributed to the quenching of local interactions close
238: to the dissipative scales \cite{Herring82,Falkovich94,Lohse95,Martinez97},
239: or to a cascade of helicity \cite{Kurien04} whose energy spectrum
240: would follow a $k^{-4/3}$ power law. The quenching of local interactions
241: in the bottleneck was measured directly in simulations in \cite{Mininni06},
242: and will be also shown here for run IV (see below, Figs.
243: \ref{fig:T3}-\ref{fig:scaling}). The $k^{-4/3}$ spectrum is also compatible
244: with the present data, as shown in Fig. \ref{fig:spectrum}(b) giving
245: %, as well as with that at lower resolution \cite{Mininni06}. This can be
246: %seen in Fig. \ref{fig:spectrum}(b), which shows
247: the energy spectra in runs III and IV compensated by $k^{-4/3}$. However,
248: we observe that the width of the bottleneck appears to be independent of
249: the Reynolds number; this indicates that the origin of the bottleneck is
250: more likely a dissipative viscous effect than an inertial range effect.
251: If helicity plays a role in the formation of the bottleneck, it has to be
252: connected to the local generation of helicity at small scales due to the
253: viscous term in the Navier-Stokes equation. Purely helical structures are
254: exact solutions of the Navier-Stokes equation, and as a result an increase
255: of helicity in the small scales could quench local interactions and the
256: cascade rate (as assumed in Ref. \cite{Kurien04}).
257:
258:
259: \begin{figure}
260: \includegraphics[width=8cm]{fig2}
261: \caption{(a) Amplitude of the triadic interactions $T_3(K,P,Q)$ for $Q=40$
262: as a function of $K$ and $P$ in run IV. (b) Shell-to-shell energy
263: transfer $T_3(K,Q)$ in the same run; several values of $Q$ are
264: superimposed.}
265: \label{fig:T3} \end{figure}
266:
267: The relative strength of local versus non-local interactions between Fourier
268: modes in the shell-to-shell transfer, and in the energy flux can be measured
269: in numerical simulations with the help of a variety of transfer functions
270: \cite{Kraichnan71,Lesieur,Verma04,Alexakis05,Alexakis05b}. Specifically, the
271: amplitude of the basic triadic interactions between the modes in shells $K$,
272: $P$ and $Q$ is defined as:
273: %\cite{Kraichnan71,Lesieur,Verma04,Alexakis05,Alexakis05b}
274: \begin{equation}
275: T_3(K,P,Q) = -\int {\bf u}_K \cdot ({\bf u}_P \cdot \nabla) {\bf u}_Q
276: d{\bf x}^3 ,
277: \label{trans_eq3} \end{equation}
278: where the notation ${\bf u}_K$ denotes the velocity field filtered
279: to preserve only the modes in Fourier space with wavenumbers in the interval
280: $[K,K+1)$. Picking a wavenumber in the inertial range (here $Q=40$), we show
281: in Fig. \ref{fig:T3}(a) its amplitude as a function of $P$ and $K-Q$ for
282: run IV. Specific values of two levels are indicated as a reference (the
283: maximum, indicated by the arrow, corresponds to $P=k_F$). As a comparison,
284: in run III, $\max\{T_3(K,P,Q=40)\} \approx 1.4 \times 10^{-3}$ indicating that
285: a decrease of the relative amplitude of the non-local triadic interactions
286: with the large scale flow ($P=k_F$) occurs as the Reynolds number increases.
287: However, the non-local coupling of the modes with $P \approx k_F$ is still
288: dominant in run IV.
289:
290: The relevance of these interactions in the transfer of energy
291: between scales can be quantified by studying the shell-to-shell transfer
292: and the net and partial fluxes. The energy transfer from the shell $Q$
293: to the shell $K$, integrating over the intermediate wavenumber, is defined as:
294: \begin{equation}
295: T_2(K,Q) = \sum_P T_3(K,P,Q) = -\int {\bf u}_K \cdot ( {\bf u} \cdot
296: \nabla) {\bf u}_Q d{\bf x}^3 \ .
297: \label{trans_eq2} \end{equation}
298: It has the same qualitative behavior as in runs at lower
299: Reynolds number [see Fig. \ref{fig:T3}(b)]. The minimum of $T_2$ for
300: $K-Q \approx -k_F$ for all values of $Q$, and the maximum for
301: $K-Q \approx k_F$, both denote that the energy is transfered from the nearby
302: shell $K-k_f$ to the $Q$ shell, and transfered from this shell to the nearby
303: shell $K+k_f$. As a result, as we increase the Reynolds number, the
304: shell-to-shell energy transfer is still local but not self-similar,
305: mediated by strong non-local triadic interactions with the large scale
306: flow at $k_F$ \cite{Domaradzki90,Ohkitani92,Zhou93,Yeung95,Verma04,Alexakis05b,Mininni06,Poulain06}.
307:
308: \begin{figure}
309: \includegraphics[width=8cm]{fig3}
310: \caption{(a) Ratio of large-scale to total energy flux
311: $\Pi_{\textrm LS}(k)/\Pi(k)$ as a function of wavenumber for
312: runs I (dash-dot), II (dash), III (dot), and IV (solid).
313: (b) Ratio of nonlocal to total energy flux $\Pi_{\textrm NL}(k)/\Pi(k)$
314: for the same runs; wavenumbers are in units of $k_\eta$. The fluxes are
315: defined in Eqs. (\ref{eq:flux}), (\ref{eq:fluxls}), and
316: (\ref{eq:fluxnl}).}
317: \label{fig:ratio} \end{figure}
318:
319: It has been observed that although the individual non-local triadic
320: interactions are strong, as modes are summed to obtain the energy
321: flux, non-local effects become less relevant. To quantify further the net
322: contribution of the local and non-local effects to the energy flux,
323: we introduce the total flux
324: \begin{equation}
325: \Pi(k) = -\sum_{K=0}^k \sum_P \sum_Q T_3(K,P,Q) ,
326: \label{eq:flux}
327: \end{equation}
328: the energy flux due to the non-local interactions with {\it only} the
329: large scale flow
330: \begin{equation}
331: \Pi_{\textrm LS}(k) = -\sum_{K=0}^k \sum_{P=0}^{6} \sum_Q T_3(K,P,Q)
332: , \label{eq:fluxls}
333: \end{equation}
334: and the energy flux due to all the interactions outside the octave
335: around wavenumber $k$ (i.e., all non-local interactions)
336: \begin{equation}
337: \Pi_{\textrm NL}(k) = -\sum_{K=0}^k \sum_{P=0}^{k/2} \sum_Q T_3(K,P,Q) .
338: \label{eq:fluxnl}
339: \end{equation}
340:
341: Figure \ref{fig:T3}(a) shows the ratio $\Pi_{\textrm LS}(k)/\Pi(k)$ as
342: a function of wavenumber for run IV. The same ratio for the lower
343: resolution runs in Table \ref{table:runs} are also shown here as a
344: reference. If the cascade is due to local interactions, this ratio
345: should decrease as smaller scales are reached. We observe however that,
346: at small scales, a plateau obtains within which this ratio remains
347: relatively constant. This is observed in runs III and IV, the two
348: runs at the highest Reynolds numbers. Note also that the plateau
349: lengthens as $R_{\lambda}$ increases: the length of the plateau
350: corresponds roughly to the length of the inertial range (including the
351: bottleneck) at those Reynolds numbers. Finally, the amplitude of the
352: plateau decreases as the Reynolds number is increased, indicating a
353: smaller contribution of the interactions with the large scale flow,
354: relative to the total flux. A detailed study of its dependence with
355: Reynolds number is discussed in the next section.
356:
357: As previously mentioned, the ratio $\Pi_{\textrm LS}/\Pi$ does not
358: increase in the range of wavenumbers that spans the bottleneck. It is the
359: contribution of all non-local interactions (interactions with all the modes
360: outside the octave around a given wavenumber $k$) that becomes dominant in
361: this range. Figure \ref{fig:T3}(b) shows the ratio
362: $\Pi_{\textrm NL}(k)/\Pi(k)$ for the runs in Table \ref{table:runs} (note
363: the wavenumbers are plotted in units of $k_\eta$). As scales closer to the
364: dissipative range are considered, the contribution of all the non-local
365: interactions increases, in agreement with the findings in Ref.
366: \cite{Herring82}. Moreover, the amplitude of $\Pi_{\textrm NL}(k)/\Pi(k)$
367: when $k$ is in units of $k_\eta$ is roughly independent of the Reynolds,
368: in agreement with a width of the bottleneck independent of $R_e$ and
369: controlled by the growth of $\Pi_{\textrm NL}(k)$ as $k$ gets closer to
370: the dissipation scale.
371:
372: It is worth noting that even at the highest Reynolds number examined here,
373: there is still a significant contribution of nonlocal interactions
374: ($\Pi_{\textrm LS}$ and $\Pi_{\textrm NL}$) to the total energy flux in
375: the inertial range. The comparison with runs at smaller resolution shows
376: a qualitative agreement and the persistence of the described non-local
377: effects. What the new computation at $R_{\lambda}\sim 1300$ allows,
378: though, is to determine the scaling of the relative importance of nonlocal
379: effects in Navier-Stokes turbulence when the Reynolds number is increased,
380: as we discuss next.
381:
382: \section{Scaling laws in turbulent flows}
383:
384: \begin{figure}
385: \includegraphics[width=8cm]{fig4}
386: \caption{Scaling of (a) the flux ratio $\Pi_{\textrm LS}/\Pi$ and (b) the
387: non-local flux ratio $\Pi_{\textrm NL}/\Pi$ as a function of Reynolds
388: number. Both ratios are evaluated at the Taylor scale, and several
389: slopes are indicated as references.}
390: \label{fig:scaling} \end{figure}
391:
392: Numerical simulations do not excel in the determination of scaling
393: laws in turbulent flows. The resolutions allowed by present day computers
394: barely allow for the existence of a well-defined inertial range. Indeed,
395: the observation of Fig. \ref{fig:spectrum} shows that, at this Taylor
396: Reynolds number, the Kolmogorov inertial range covers less than one order
397: of magnitude in scale (although, as noted before, the flux is constant in
398: a larger range of scales). This could be an indicator that solutions more
399: complex than simple power laws hold in the inertial range \cite{Tsuji04}.
400: The pioneering computations of the Japanese group on the Earth Simulator
401: using random forcing has allowed, however, for some scaling laws to emerge,
402: although, as these authors observed, not all physical quantities of
403: interest converge to asymptotic values at the same rate
404: \cite{Ishihara05,Kaneda06}. We here
405: %attempt to obtain
406: display such scaling laws for the particular flow studied, namely the
407: Taylor-Green flow, relevant to several laboratory experiments. In
408: particular, we are interested in the scaling of the relative amplitude
409: of local and non-local interactions, as well as other quantities often
410: studied in the context of turbulent flows, whose scaling will be used
411: as a criteria to classify the runs \cite{Vanatta80,Ishihara05,Kaneda06}.
412: It is worth mentioning in this context that, with four runs, we can only
413: show the results are consistent (or at least, not inconsistent) with
414: a particular scaling.
415:
416: \begin{figure}
417: \includegraphics[width=8cm]{fig5}
418: \caption{Kolmogorov constant $C$ as a function of Reynolds number; the
419: solid line gives the best fit $C = 4.60 R_e^{-0.16} + 0.64$.}
420: \label{fig:constant} \end{figure}
421:
422: \begin{figure}
423: \includegraphics[width=8cm]{fig6}
424: \caption{(a) Kurtosis and (b) skewness of the velocity increments as a
425: function of the Taylor based Reynolds number. Results are given for
426: two different increments: the Taylor scale ($+$), and the
427: dissipation scale ($*$). The slopes indicated as a reference are
428: from experimental results.}
429: \label{fig:skewness} \end{figure}
430:
431: Figure \ref{fig:scaling}(a) gives the scaling of the flux ratio
432: $\Pi_{\textrm LS}(k)/\Pi(k)$ with the Reynolds number. To this end,
433: we take the Taylor scale $\lambda$ as a reference scale in the
434: inertial range, and we evaluate $\Pi_{\textrm LS}(k)/\Pi(k)$ for
435: each run at the Taylor wavenumber $k_\lambda = 2 \pi/\lambda$. The
436: best fit to all the runs gives $\Pi_{\textrm LS}/\Pi \sim R_e^{-0.7}$,
437: although the dependence of the ratio $\Pi_{\textrm LS}/\Pi$ with $R_e$
438: seems to change slightly for Run IV. A best fit of the last two points
439: (runs III and IV) gives a dependence $\sim R_e^{-0.6}$ (as was
440: discussed in Sect. \ref{sec:2048}, these two runs show a developed
441: inertial range).
442:
443: We also evaluate the ratio $\Pi_{\textrm NL}/\Pi$ at the Taylor
444: wavenumber; its dependence with the Reynolds number is shown in Fig.
445: \ref{fig:scaling}(b). Here, the ratio in runs III and IV is
446: compatible with a slower decay $\sim R_e^{-0.4}$. The anomalous
447: behavior of runs I and II in Fig. \ref{fig:scaling}(b) is due to the
448: fact that in these runs at lower resolution, the sum over $P$ from
449: the smallest wavenumber $k_{min}=1$ to $k_\lambda/2$ in Eq.
450: (\ref{eq:fluxnl}) defines bands that are too narrow in Fourier
451: space. In other words, it is linked to the lack of a well-defined
452: inertial range in the simulations at lower Reynolds numbers, and we
453: can only expect scaling to obtain in the limit of large $R_e$.
454:
455: Figure \ref{fig:scaling} indicates that as the Reynolds number is
456: increased, the contribution of the non-local interactions with the
457: large scale flow to the total flux decreases (as well as the
458: contribution of all non-local interactions, albeit at a slower
459: rate). On dimensional grounds $\Pi_{\textrm LS}(k_\lambda) \sim
460: U_L u_\lambda^2/L$. Here, $U_L$ is a characteristic velocity at the
461: large scale $L$, and $u_\lambda$ is a characteristic velocity at the
462: Taylor scale (note that this relation does not take into account
463: that structures are in fact multiscale \cite{Alexakis05b}). On the
464: other hand, for $\Pi_{\textrm LS}/\Pi \ll 1$, we have
465: $\Pi(k_\lambda) \sim u_\lambda^3/\lambda$. As a result, we may
466: expect $\Pi_{\textrm LS}(k_{\lambda})/\Pi(k_\lambda) \sim
467: R_e^{-1/2}$. The condition $\Pi_{\textrm LS}/\Pi \ll 1$ is not
468: satisfied in the simulations at lower resolution, and it is unclear
469: whether the departure of Run IV in Fig. \ref{fig:scaling} represents
470: a convergence to a different scaling than $\sim R_e^{-0.7}$ at very
471: large Reynolds numbers. This will require further studies at higher
472: numerical resolutions, a feat reachable with petascale computing.
473:
474: Other scaling laws can be observed in this series of runs; in
475: particular, it is worth comparing the scaling of quantities for
476: which data exist from laboratory experiments or from previous
477: simulations. Figure \ref{fig:constant} shows the Kolmogorov constant
478: $C_K$ as defined by the inertial range spectrum $E(k)=C_K
479: \epsilon^{2/3} k^{-5/3}$. As a reference, we computed a best fit of
480: the form $C_K=a R_e^b + c$, as suggested e.g. in \cite{Tsuji04}, and
481: obtained $a=4.60$, $b=-0.16$, and $c=0.64$. The value of $c$
482: (that represents the asymptotic value of the Kolmogorov
483: constant for infinite $R_e$) obtained from this fit is in good
484: agreement with experimental results and atmospheric observations
485: \cite{Mydlarski96,Tsuji04}, although the values of $a$ and $b$
486: differ. We also note that the measured value of the Kolmogorov
487: constant for the $2048^3$ runs is more than double the value of the
488: expected asymptotic limit $c$, indicating that we are still far away
489: from an asymptotic behavior for large $R_e$.
490:
491: Figure \ref{fig:skewness} shows the skewness
492: \begin{equation}
493: S=\left< \delta u_L(r)^3 \right> / \left< \delta u_L(r)^2 \right>^{3/2},
494: \end{equation}
495: and kurtosis
496: \begin{equation}
497: K=\left< \delta u_L(r)^4 \right> / \left< \delta u_L(r)^2 \right>^{2},
498: \end{equation}
499: of the longitudinal velocity increment $u_L= {\bf u} \cdot {\bf r}/r$
500: \begin{equation}
501: \delta u_L(r) = u_L({\bf x+r}) - u_L({\bf x}),
502: \end{equation}
503: i.e. the component of the velocity
504: in the direction of the increment. The skewness and kurtosis were
505: evaluated at two scales, $r=\lambda$, the Taylor scale, and $r=\eta$,
506: the dissipation scale. In the latter case, only the results from runs III
507: and IV show a dependence with $R_\lambda$ which is consistent with
508: experimental results \cite{Vanatta80}. The behavior of these two runs
509: further confirms that high Reynolds numbers are needed to observe scaling of
510: turbulent quantities.
511:
512: \section{Intermittency and structures}
513:
514: The Taylor-Green flows computed here correspond to an experimental
515: configuration of two counter-rotating cylinders, studied in the
516: laboratory for fluid turbulence as well as in the context of the
517: generation of magnetic fields in liquid metals. These flows
518: present both inhomogeneities and anisotropies in the large scales,
519: a resolved inertial range followed by a bottleneck, and a dissipative
520: range. One may study the rate at which the symmetries of the Navier-Stokes
521: equations are recovered in the small scales, and whether the statistical
522: properties of the small scales are universal. In this section we address
523: the specific question of the properties of the small scales through the
524: evaluation of the anomalous exponents $\zeta_p$ of the longitudinal
525: structure functions $S_p$ of the velocity field, defined as:
526: \begin{equation}
527: S_p= \langle \delta u_L(r)^p \rangle \sim r^{\zeta_p},
528: \end{equation}
529: assuming homogeneity and isotropy. In order to obtain better scaling
530: laws, we use the Extended Self-Similarity hypothesis (ESS)
531: \cite{Benzi93,Benzi93b} in the particular context of plotting $S_p$
532: as a function of $S_3$.
533:
534: \begin{figure}
535: \includegraphics[width=8cm]{fig7}
536: \caption{Scaling exponents using the ESS hypothesis in the $1024^3$ and
537: $2048^3$ runs. The scaling predicted by Kolmogorov and by the
538: She-L\'ev\^eque model are also given as a reference.}
539: \label{fig:exponents} \end{figure}
540:
541: Figure \ref{fig:exponents} shows the scaling exponents $\zeta_p$ in the
542: $1024^3$ and $2048^3$ runs, computed using the ESS hypothesis. Similar
543: results are obtained without ESS and doing the fit only in the inertial
544: range, defined as the range of scales where the so-called 4/5th law of
545: Kolmogorov %\cite{k41-45}
546: is satisfied, namely $S_3(r) \sim r$. If we
547: define stronger intermittency as stronger departure from the Kolmogorov
548: scaling $\zeta_p=p/3$, we note that as we increase the Reynolds number,
549: the intermittency increases as well, albeit slowly. Furthermore, for
550: higher $R_{\lambda}$ (run IV), the departure from the She-L\'ev\^eque
551: model \cite{She94} increases (compared with run III), even for fixed
552: values of $p$. The differences between $\zeta_p$ for runs III and
553: IV, albeit small, are at least one order of magnitude larger than the
554: errors in the fit using ESS. As an example, in run III
555: $\zeta_6=1.746 \pm 0.003$ and $\zeta_8=2.136 \pm 0.007$, while in run IV
556: $\zeta_6=1.7284 \pm 0.0004$ and $\zeta_8=2.0968 \pm 0.0007$.
557:
558: Here it is worth separating the discussion in two parts. On the one
559: hand, the increase of the departure from the She-L\'ev\^eque model
560: as the Reynolds number and spatial resolution are increased
561: indicates that the departure is not the result of lack of
562: statistics. This change in the exponents for simulations with the
563: same forcing at different Reynolds numbers shows that huge Reynolds
564: are required to obtain convergence of high order statistics. In fact,
565: the larger the moment $p$ examined, the larger the relative difference
566: between the $\zeta_p$ exponents measured in the two runs.
567: On the other hand, it was shown in Ref. \cite{Mininni06} that
568: differences in the scaling exponents were measurable when
569: considering two different forcings at similar Reynolds numbers.
570: These differences could be due to anisotropies in the flow, and in
571: that case an SO(3) decomposition could be used to study whether the
572: scaling exponents of the isotropic component of the flow are
573: universal. However, if there is a significant return to isotropy in
574: the small scales, we then also expect the isotropic component to
575: dominate when the Reynolds number is large enough.
576: %\NOTE{K41-45: kolmogorov 1941 4/5th law}
577:
578: \begin{figure}
579: \vskip0.3truein \includegraphics[width=8.7cm]{fig8}
580: \caption{(Color online) Left: rendering of vorticity intensity in a small
581: region of run IV. Only regions with
582: $|\boldsymbol{\omega}| \ge \max\{|\boldsymbol{\omega}|\}/6.5$ are shown
583: ($\boldsymbol{\omega}=\nabla \times {\bf v}$). Note the clustering of
584: filaments into larger vorticity structures. The bars on
585: the bottom indicate respectively the integral, Taylor, and dissipation
586: scales. Right: rendering of relative helicity in the same region (red
587: is $-1$ and blue is 1). Only regions with absolute value larger than
588: $0.92$ are shown.}
589: \label{fig:structure} \end{figure}
590:
591: The intermittency of the flow is linked to the presence of strong
592: spatially separated structures in the form of vortex filaments. The high
593: $R_{\lambda}$ computation (run IV) displays the same large-scale
594: structure of bands as the run presented in \cite{Mininni06}. Conditional
595: statistics analysis as the ones performed in \cite{Mininni06} keep
596: showing a correlation between large scale shear and small scale gradients
597: and enhanced intermittency. It has been noted by several authors that
598: filaments tend to cluster into larger filamentary structures; this is
599: observed e.g. for supersonic turbulence \cite{Porter98} and in the
600: interstellar medium,
601: %\cite{falga}
602: and it has been analyzed quantitatively in \cite{Moisy04}. When
603: individual structures are studied in real space, filament-like clusters
604: formed by smaller vortex filaments are observed here again (see Fig.
605: \ref{fig:structure}), something that was not seen in simulations of
606: the TG flow at lower resolution. This could be interpreted as a
607: manifestation of self-similarity, and a more quantitative analysis will
608: be presented elsewhere. In particular, it would be of interest to compute
609: the inter-cluster distance, and the intra-cluster inter-filament distance,
610: to see whether the space-filling factor of such flows diminish with
611: increasing Reynolds number. Note that the vortex cluster reaches a
612: global length comparable to the integral scale of the flow (indicated
613: in Fig. \ref{fig:structure}); as such, they may be a real-space
614: manifestation of the trace of non-local interactions between
615: small-scales (dominated by vortices) and large scales (dominated by the
616: forcing), giving a coherence length to the flow.
617:
618: Figure \ref{fig:structure} also shows the density of relative
619: helicity ${\bf v} \cdot \boldsymbol{\omega} (|{\bf
620: v}||\boldsymbol{\omega}|)^{-1}$ ($\boldsymbol{\omega}=\nabla \times
621: {\bf v}$). Regions in blue and red correspond respectively to
622: regions of maximum alignment or anti-alignment between the two
623: fields (only regions with absolute relative helicity larger than
624: $0.92$ are shown). Note that regions with large relative helicity
625: correspond to small vortex tubes, but the filament-like clusters
626: have no coherent helicity. Regions with strong alignment fill a
627: substantial portion of the subvolume, even though the global
628: (relative) helicity of the flow is close to zero.
629:
630: \section{Discussion and Conclusion}
631:
632: The data presented in this paper has allowed for a refined analysis of
633: the behavior and structure of turbulent flows as the Reynolds number is
634: increased. We have in particular showed that: (i) the bottleneck appears
635: to have a constant width for the two higher $R_e$ runs;
636: %in the $1024^3$ and $2048^3$ runs;
637: hence, it is probably linked to the dissipation range, and to the
638: depletion of nonlinearities as we approach this range; (ii) the
639: scaling with $R_e$ of the non-local energy fluxes, which indicates a
640: weakening of non-local interactions as $R_e$ increases. These first
641: two results taken together point out to the fact that the bottleneck
642: may not disappear in the limit of very high Reynolds number, since
643: it has been argued that its existence is linked to the relative
644: scarcity of non-local interactions in Navier-Stokes turbulence, by
645: opposition to, e.g., the magnetohydrodynamic (MHD) case. Indeed,
646: when coupling the velocity to a magnetic field in the
647: %magnetohydrodynamic
648: MHD limit, it was shown that the transfer of
649: energy itself was non-local, and that the bottleneck was absent in
650: numerical simulations of such flows; this can be understood in the
651: following manner: as one approaches the dissipation range, few
652: triadic interactions are available but in a flow for which the
653: nonlinear transfer is nonlocal, the energy near the dissipative
654: range can still be transfered efficiently to smaller scales since
655: small-scale fluctuations are transfered by the large scales
656: \cite{Alexakis05}. Finally, the departure of the anomalous exponents
657: of velocity structure functions from standard models of
658: intermittency such as the She-L\'ev\^eque model seems to increase as
659: the Reynolds number is increased.
660:
661: As noted before in \cite{Kaneda06}, convergence to the asymptotic
662: turbulence regime appears to be very slow: even though the nonlocal
663: interactions do diminish with Reynolds number, they are still measurable
664: at these resolutions. In run IV on a $2048^3$ grid at
665: $R_{\lambda}\sim 1300$, of the order of $10\%$ of the energy flux is
666: due to non-local interactions with the large scale flow, and the
667: dependence of the energy flux ratio $\Pi_{\textrm LS}/\Pi$ with $R_e$ for
668: very large $R_e$ is still unclear.
669: %Concerning high order structure functions, the exponent for the sixth
670: %order function already shows a significant (larger than error bars)
671: %change from the lower resolution run.
672: This not only raises the question of the determination of higher
673: order quantities at moderate Reynolds numbers in simulations and
674: experiments, but it also opens the door for a non-universal behavior
675: of turbulent flows which may have to be studied in more detail than
676: was previously hoped for.
677:
678: \begin{acknowledgments}
679: Computer time was provided by NCAR and by the National Science
680: Foundation Terascale Computing System at the Pittsburgh Supercomputing
681: Center. PDM and AP acknowledge invaluable support from Raghu Reddy at
682: PSC. PDM acknowledges discussions with D.O. G\'omez. PDM is a member
683: of the Carrera del Investigador Cient\'{\i}fico of CONICET. AA
684: acknowledges support from Observatoire de la C\^ote d'Azur and Rotary
685: Club's district 1730. The NSF grant CMG-0327888 at NCAR supported this
686: work in part. Three-dimensional visualizations of the flows were done
687: using VAPOR, a software for interactive visualization and analysis of
688: terascale datasets \cite{Clyne07}.
689: \end{acknowledgments}
690:
691: \bibliography{ms}
692:
693: \end{document}
694: