APS/123-QED

On the missing link between pressure drop, viscous dissipation and the turbulent energy spectrum

A. Badillo∗ and O. Matar† Department of Chemical Engineering, Imperial College London, UK. (Dated: April 27, 2021) We present convincing evidence of a direct connection between the pressure drop, viscous dissipa- tion, and the turbulent energy spectrum. We use this finding to explain Nikuradse’s experimental results of pressure drop for turbulent flows in rough pipes, in terms of a modified Kolmogorov length scale that varies with the surface roughness. Furthermore, we use Laufer’s measurements of turbu- lent energy spectra in pipe flow to calculate—to a good approximation—the turbulent component of the pressure drop directly from an averaged turbulent energy spectrum. We also show that the incompressibility assumption, leads to the conclusion that viscous dissipation in fully-developed (laminar and turbulent) pipe flow, cannot increase the temperature of the fluid through viscous heating.

PACS numbers:

↔ For many decades, turbulence has remained stub- τ = µ ∇u + ∇uT
the viscous stress tensor expressed bornly one of the most challenging open problems in in terms of the dynamic viscosity µ and the strain rate. classical physics. Achieving complete understanding of In our analysis, we consider a portion of a pipe, where turbulent flows is hindered by the their multi-scale na- the flow is adiabatic, incompressible and fully developed ture, and the complex mechanisms underlying the dis- (Fig. 1). Hence, all the quantities are statistically invari- tribution of energy to the large number of harmonics, ant along the stream-wise direction, except for the fluid which constitute the fluctuating velocity field. It is well pressure. accepted that vortex stretching is one of the main mecha- nisms responsible for transferring energy from low to high wavenumbers. This mechanism, however, only exists in viscous, three-dimensional turbulence, with no analogue in quantum [1], and two-dimensional turbulence [2]. The pictorial energy cascade proposed by Richardson [3], and then recovered theoretically first by Kolmogorov [4, 5], and then, independently, by Onsager [6], does not reveal much about the energy transfer between consecu- tive harmonics; it only represents an ‘equilibrium’ state of energy distribution arising from a balance between the fluxes of energy going up and down the cascade. Despite the fact that viscous dissipation is irrelevant for quantum turbulence, the scaling laws dominating its energy cas- FIG. 1: Schematic of fully developed pipe flow. The pipe is cade are the same as those in viscous, three-dimensional thermally insulated (adiabatic) and the fluctuations of any turbulence [7–9]. What, then, is the true role of viscous flow quantity are statistically identical at the inlet and outlet dissipation in the formation of the energy cascade, and of the portion of the analyzed pipe. We also present how, what is its connection to macroscopic quantities such as hypothetically, the molecular radial distribution varies along the pipe (wavy lines in the upper part of the figure). The the pressure drop? We address the second question in red circle surrounded by black ones represents, schematically, the present letter. the local arrangement of atoms/molecules in the fluid. At We start our discussion from the conservation of linear ↔ inlet, the pressure is higher than that at the outlet, leading momentum ∂tp + ∇ · (up) = −∇· σ +Σfext, with p = to a slightly higher local degree of order. This higher degree arXiv:2104.12110v1 [physics.flu-dyn] 25 Apr 2021 ρu the linear momentum, ρ the density, u the instanta- of short range order is observed as a larger amplitude of the ↔ second and subsequent peaks in the radial distribution. neous velocity field, σ the stress tensor, and Σfext the summation of all external forces. We further split the ↔ ↔ ↔ By applying a time average to the conservation of mo- stress tensor as σ= P I − τ , with P the pressure and mentum, we arrive to the following expression for the pressure drop 2L ∆P = − ρu2 (1) ∗Electronic address: [email protected] R τ †Electronic address: [email protected] p with R the pipe radius, L the pipe length, uτ = τw/ρ 2

the friction velocity, τw = (µ∂U/∂r)w the wall shear indicate that the local range order (a measure of how stress, and U the stream-wise component of the time- molecules are arranged in the liquid) undergoes impor- averaged velocity field. tant changes as pressure increases. Using the MD data An equivalent expression can be obtained from the from Mahoney and Jorgensen [11], we can estimate the 2 conservation of kinetic energy ek = u /2 (per unit derivative deu/dP . Applying a linear fit to their MD  ↔  ↔ data yields a value of de /dP ∼ −0.826 (see supplemen- mass) ∂ (ρe ) + ∇ · uρe + σ ·u =σ : ∇u + Σf · u. u t k k ext tal material). Since the fluid considered in the MD study By applying a time average to the conservation law for is slightly compressible [11], we would expect a value kinetic energy and invoking incompressible and fully de- slightly different from −1. These molecular studies, show veloped pipe flow conditions, the pressure drop reads the strong degree of correlation between internal energy and the molecular arrangement of the liquid. This would 1 Z D↔ E ∆P = − τ : ∇u dΩ, (2) imply that for strictly incompressible fluids, the energy VA Ω of the flow (pressure + kinetic) is dissipated only into the potential component of the internal energy follow- with h·i a time average, VA = R u · dA the volumetric A ing a reversible path, and not, as commonly believed, flow rate, A the pipe cross section, and V the average into the micro-kinetic component of the internal energy cross section velocity (see supplemental material for a (temperature). For a liquid flowing in a pipe, the higher detailed derivation of conservation laws and the two ex- pressure at the inlet will cause a slightly higher degree of pressions for the pressure drop). Equation (2) was also molecular order compared to that at the outlet (see Fig. presented by Pope [10]. Since the conservation law for 1). kinetic energy is not independent from that of momen- tum, Eqs. (1) and (2) are completely equivalent. Thus, To reveal the connection between Eq. (2) and the tur- we can express the wall shear stress in terms of the total bulent the energy spectrum, we follow the work of Gioia and Chakraborty[12]. They investigated the relationship viscous dissipation. 2 between the friction factor, f = 8τw/ρV and the phe- 1 Z D↔ E nomenological turbulent energy spectrum. Considering τ = − τ : ∇u dΩ, (3) w 2πRLV the Darcy-Weisbach equation, we can express the pres- Ω sure drop as We need to emphasize that Eq. (3) is only valid for fully L ρV 2 developed (laminar or turbulent) and incompressible pipe ∆P = −f , (4) flow. For laminar flow, it is possible to confirm analyti- DH 2 cally the validity of Eqs. (2) and (3). The first interest- ing conclusion from Eq. (2), is that viscous dissipation where DH the hydraulic pipe diameter. The negative cannot increase the temperature of a fluid at the incom- sign is to enforce a decrease in the pressure along the pressibility limit. The conservation of enthalpy per unit pipe. By replacing Eq. (2) into (4) and introducing the ↔ Reynolds decomposition u = U + u0, with U = hui, we mass, h, implies that ρD h + ∇ · q = D P + τ : ∇u (see t t can express the friction factor in terms of two components supplemental material), with D = ∂ +u·∇ the material t t of the viscous dissipation, derivative and q the heat flux. If the pipe is adiabatic and the flow incompressible, the only possibility for the en- thalpy to change, is through viscous heating (dissipation Z Z 0  64 1 0  term). After applying a time average to the conserva- f = dΩ + 0 dΩ (5) Re 8πLV 2 ν ν tion law of enthalpy, the integration of the R.H.S. over Ω Ω the pipe volume, vanishes by virtue of Eq. (2) and, thus, R where ν is the kinematic viscosity, and hρD hi dΩ = 0. Since there is no energy exchange with D↔ E Ω t ρ−1 τ : ∇u =  + 0 is defined as the total time- the environment (adiabatic flow), this requires either a 0 0 T constant fluid temperature or that the local viscous dissi- averaged dissipation, with 0/ν = ∇U: ∇U+∇U : ∇U 0 0 0 0T 0 pation increase and decrease the local fluid temperature and 0/ν = ∇u : ∇u + ∇u : ∇u the mean and in a way that the net change of total enthalpy is null. Be- turbulent components of the viscous dissipation respec- cause the second option would be in violation with the tively. The Reynolds number, Re = VDH /ν, is defined first law of thermodynamics, the fluid temperature must in terms of the hydraulic pipe diameter, DH . be constant at the incompressibility limit. The main difficulty faced by Gioia and Chakraborty For molecular fluids like water, the internal energy can [12], was the determination of the wall shear stress, and suffer important variations by changing the molecular ar- how to relate it to the turbulent energy spectrum and the rangement, keeping an almost constant temperature and surface roughness. They postulated that the wall shear density. A small change in the enthalpy per unit volume, stress can be expressed in terms of a momentum transfer ˜ ˜ h, is given by dh = deu + dP , with eu being the inter- between the bulk flow and the flow near the wall, charac- nal energy per unit volume. If the enthalpy of the fluid terized by a velocity scale us. Expressing the wall stress is constant, then deu/dP = −1. Molecular Dynamics as τw = ρV us, the problem reduces to finding the veloc- (MD) simulations of water at constant temperature [11], ity scale, which might also vary with the amplitude of the 3 surface roughness. Furthermore, to determine the veloc- mogorov spectrum to determine this integral. Hence, we 2/3 −5/3 ity scale they assumed the validity of the phenomeno- choose Rˆ0 = A0˜ q Cd (q), with A0 a dimensionless logical energy spectrum in the proximity of the wall, constant and ˜ a dissipation rate scale (see supplemental which is something arguable. It is well known, however, material for its derivation) that must not be confused 0 R 0 2/3 R ∞ 1/3 that to derive the phenomenological spectrum, the turbu- with 0. Thus, Ω 0/νdΩ ≈ 15ΩA0˜ 0 q Cd (q) dq. lent structures must be rotationally- and translationally- It is now clear that without the correction for the invariant; that is, the flow is isotropic and homogenous, dissipative regime, the total dissipation diverges when which is not the case near the wall. Aware of this lim- using a Kolmogorov spectrum, which would translate itation, Gioia and Chakraborty[12] argue that even for into an infinite friction factor. To recover the proper anisotropic and inhomogeneous flows, the phenomenolog- ical theory still represents a good approximation. Thus, 10 by using directly the energy spectrum E (q) (including corrections for the energetic and dissipative ranges), they 2 R ∞ obtain the velocity scale as us = 1/s E (q)dq. Here, the integration is carried out for wave numbers higher than 1/s = 1/ (r + aη), where a is a dimensionless constant, r the size of the surface roughness, and η the Kolmogorov 1/4 1 length scale. Hence, they are effectively filtering out all f Re eddies larger than s in the calculation of the velocity scale. It is remarkable that for Re → ∞, the friction factor in the work of Gioia and Chakraborty[12] scales as f ∼ (r/R)1/3 despite the absence of a correction for 0.1 the dissipative regime, Cd (q), in the turbulent energy 0.1 1 10 100 1000 10000 spectrum: without this, the total dissipation diverges (λ/R)Re3/4 when using the Kolmogorov spectrum. The authors re- covered the proper scaling for the friction factor, observed in Nikuradse’s experimental results, only by selecting an FIG. 2: Reproduction of Nikuradse’s data over the entire appropriate cutoff wave length (1/s) in the integration of range of Reynolds number. the energy spectrum. To avoid an arbitrary selection of the cutoff wave scaling for the turbulent friction factor, the problem length, we follow a different route. In our analysis, we reduces to choosing an appropriate correction Cd(q). start by investigating the direct connection between For instance, we can recover the correct scaling by the friction factor, defined in Eq. (S28) with the selecting Cd(q) = exp (−βψ [R/η, Re, r/R] ηq), with ψ [R/η, Re, r/R] a dimensionless function, η = ν3/4˜−1/4 turbulent energy spectrum. Considering an spherically 3/4 ˆ  11/12 and volumetric averaged spectrum R0 (see details in the the Kolmogorov’s length scale, β = 30A0I0C supplemental material), along with Taylor’s identities a dimensionless constant, and I and C constants for isotropic turbulence[13], we can approximate the 0  ∞ determined in the supplemental material. By choosing 0 R 2 ˆ −3/4 total viscous dissipation by 0 ≈ 15ν 0 q R0dq. It is  1/3 1/4  3/4  noteworthy to mention that in our analysis, Rˆ0, does ψ = (2R/η) A2 (1 − φ2) + A3φ2 Re (r/R) not correspond to the spectrum of any particular point (see supplemental material) we recover the scaling inside the pipe, but rather an averaged quantity. We −1/4 1/3 fT ∼ Re and fT ∼ (r/R) . The blending function now analyze the asymptotic behavior of Eq. (S28). For −1 R 2 0 h 2i 3/4 Re → 0, Ω 0/νdΩ → 8πLV and 0 → 0. Thus, the φ2 = 1 − 1 + (z/z0) , with z = Re (r/R), appropriate scaling for laminar flow f ∼ Re−1 is recov- bridges the Blasius and Strickler regimes. The constants R ered directly. Conversely, for Re → ∞, Ω 0/νdΩ → 0 z0 = 35.5, A2 = 0.32, A3 = 0.147 were determined from R 0 2 and Ω 0/νdΩ ∼ LV Re, which is a result obtained a fit to Nikuradse’s data. Goldenfeld[14] postulated from imposing a constant friction factor for high that the turbulent friction factor can be expressed Reynolds numbers. The question that remains to be −1/4  3/4  in general as fT = Re G Re (r/R) for the answered now is whether the two scalings, observed −1/4 1/3 Blasius and Strickler regimes, with G an unknown in rough pipe flows, f ∼ Re and f ∼ (r/R) function of the Reynolds number and the relative surface as Re → ∞, can be explained solely in terms of the roughness r/R. By plotting Nikuradse’s data in a fluctuating component of the viscous dissipation. We, plot of f Re1/4 vs (r/R) Re3/4, Goldenfeld was able therefore, center our attention on the second integral T to collapse almost perfectly the experimental points in Eq. (S28). Since we are interested in the dissipative onto a single curve. By utilizing a second function range of the spectrum, we consider only the correction φ = 0.5 (1 + tanh [(Re − Re ) /α ]) to bridge the for this regime C (q). Because Rˆ is isotropic and 1 T 1 d 0 laminar and turbulent regimes, with Re = 2944.27 homogeneous, we will also make use of a modified Kol- T and α1 = 1089.43 (also determined from a fit), we can 4 reproduce Goldenfeld’s findings over the entire range of part of the viscous dissipation. To evaluate the mean Re including the laminar regime. Thus, by expressing viscous dissipation (from the time-averaged velocity), we the friction factor as f = fL (1 − φ1) + fT φ1, with fL use the phenomenological profile derived by Musker[17] the friction factor in the laminar regime, we can closely (see supplemental material). Considering a dimensionless match Nikurade’s results (see Fig. 2). Nikuradse’s pipe length of L/2R = 16, Laufer measured a dimen- 1 2 data contain values from the laminar, transitional and sionless pressure drop of (Pinlet − Pout)/( 2 ρUcenter) = turbulent regimes. To avoid an arbitrary selection of 0.16, where Ucenter is the time–averaged velocity at the the data in the fully turbulent regime, we included the center of the pipe. The mean viscous dissipation ex- full data set in the fit. The value determined for ReT , pressed in the same dimensionless units resulted to be 1 2 coincides well with the Reynolds number for transitional ρ0L/( 2 ρUcenterV ) = 0.058, which represents only the turbulent flows. 36% of the total pressure drop. The results of for the di- ˆ 1 2 We have shown that a modified Kolmogorov spectrum, mensionless pressure drop ∆P = ∆P/( 2 ρUcenter), mean ˆ 1 2 with a variable dissipative length scale (which depends on viscous dissipation 0 = ρ0L/( 2 ρUcenterV ) and turbu- the surface roughness), can reproduce well Nikuradse’s ˆ0 1 2 lent viscous dissipation  0 = ρ0L/( 2 ρUcenterV ) are pre- data. However, we must still provide evidence that an sented in table I. Calculating the pressure drop from isotropic and homogeneous spectrum can be used to de- 0 the addition of ˆ0 + ˆ 0 for all the variants of the aver- termine the turbulent component of the pressure drop. aged spectrum, the error—compared to the experimental Finding a complete data set (experimental or numerical) value—ranges from 6% to 20%. to obtain an averaged spectrum has proven difficult, be- cause we need spectra of velocity fluctuations measured along different directions at several locations inside a TABLE I: Dimensionless component of viscous dissipation pipe. So far, we have found only one public data base obtained from the mean velocity profile and averaged spec- that contains the spectrum for the three components of trum. Arithmetic and geometric (spherical) averages are la- the fluctuating velocity field at various distances from the beled with a. and g. respectively (see supplemental material wall. To calculate an averaged spectrum for fully devel- for details). Calculations considering an inertial correction oped pipe flow, we use the measurements of Laufer[15]. are labeled with “i”. Newer data such as the one produced in the Superpipe at Princeton University [16], only presents the spectrum 0 0 0 0 ∆Pˆ ˆ0 ˆ 0 (a.) ˆ 0 (a.i.) ˆ 0 (g.) ˆ 0 (g.i.) for the streamwise component of the fluctuating veloc- 0.16 0.058 0.092 0.085 0.074 0.070 ity field. Looking at the individual spectra presented by Laufer (see supplemental material), we can see that at low wave numbers, the streamwise fluctuations, u0u0, contain a much higher energy level than the other two 0 0 0 0 In conclusion, we have shown that at the incompress- spanwise velocity components v v and w w (at a given ibility limit, viscous dissipation cannot increase the tem- distance from the wall). This means that most of the perature of the fluid and the flow’s energy can only dissi- turbulence kinetic energy is contained in the streamwise 0 0 pated into the micro-potential component of the internal velocity component u u . However, the situation is quite energy. We also showed that when introducing a correc- different at high wave numbers, where it is the spanwise tion to the Kolmogorov scale, as a function of the surface fluctuations that contain most of the energy. Since vis- roughness, we can reproduce accurately Nikuradse’s [18] cous dissipation is relevant only at high wave numbers, data for pressure drops in rough pipes. This leads, effec- this means that the spanwise fluctuations are the main tively, to a new correcting term for the energy spectrum responsible for dissipating the turbulence kinetic energy. in the dissipative regime. The fact that Nikurase’s re- An interesting picture appears then, where turbulence sults can be closely reproduced by modifying only the kinetic energy in created mostly on the streamwise fluc- dissipative part of the energy spectrum, is a good indica- tuating component of the velocity field, but dissipated tion that small scales in turbulence, play a much impor- mostly by the spanwise fluctuations. At those high wave tant role than previously thought. To validate the use numbers, the energy content of the spanwise fluctuations of an isotropic and homogeneous spectrum, in the cal- is about five times larger than that of the streamwise culation of the pressure drop, we used the experimental component near the pipe wall, ruling out the possibil- data from Laufer [15]. Our calculations show that the ity of a bias in the measurements. This observation in viscous dissipation component associated to the mean Laufer’s measurements, indicate that to properly evalu- velocity profile, only accounts for a modest fraction of ate the viscous dissipation, we must have the spectra of the total pressure drop. By using Laufer’s experimental the spanwise velocity fluctuations, because they are the spectra, we obtained and averaged isotropic and homo- mean responsible for viscous dissipation. Using only the geneous spectrum, which leads to a good agreement be- spectrum along the streamwise direction, leads to a much tween the measured and calculated pressure drop. These lower value for the total turbulent viscous dissipation. results constitute a strong evidence of a direct connection We use Laufer’s data set to obtain an averaged spectrum between the pressure drop, viscous dissipation, and the (see supplemental material), and calculate the turbulent turbulent energy spectrum. 5

ACKNOWLEDGEMENTS paper), p = ρu the linear momentum, fext the external ↔ ↔ ↔ forces, σ= P I − τ the stress tensor written in terms ↔ We acknowledge financial support from the Engi- of the pressure P and the viscous stress tensor τ , ρe the neering and Physical Sciences Research Council, UK, total energy per unit volume (internal + kinetic) and q through the Programme Grant MEMPHIS (grant num- the heat flux. The term corresponding to the chemical ber EP/K003976/1). potential is added for generality purposes, where varia- tion of the chemical potential produce work (e.g. osmotic pressure). For single phase mixtures assumed to be under Appendix A: Conservation laws chemical equilibrium, the chemical potential and density are constant and, therefore, the contribution of this term We start our derivation by the most general statement to the energy and momentum equations is null. However, of conservation of a (n)−rank tensor Φ(n) inside a control for two phase flows, ∇ (µρ) leads to the surface tension a dynamic volume Ω (t) force inside of the phase boundary [19]. An equation for kinetic energy can be obtained by mul- Z ! Z Z d (n) (n+1) (n) tiplying (inner product) Eq. (S4) by the velocity and Φ dΩ + jΦ · dΓ = SΦ dΩ applying mass conservation: dt Ω(t) dΓ(t) Ω(t) (S1)   ↔  ∂ (ρek) ↔ (n+1) + ∇ · [u − uΓ] ρek + σ −µρ I · u = where jΦ represents the (n + 1) −rank tensor flux of ∂t (n) Φ through the surface of the control volume Γ (t), and 2 (S6) ↔ ↔ u (n) (n) S is any source of Φ . Applying the divergence and σ −µρ I : ∇u + Σfext · u + ∇ · jd Φ 2 Reynolds transport theorem, we arrive at the differential form of the conservation law The last term in the RHS of Eq. (S6) corresponds to (n) a change in the kinetic energy due to a variation of the ∂Φ  (n+1) (n) (n) + ∇ · jΦ − uΓΦ = SΦ (S2) density with the local concentration. By subtracting Eq. ∂t (S6) from Eq. (S5), we obtain an equation for the internal with uΓ the boundary velocity of the control volume. energy: All conservation laws can be written in this way, pro- (n) ∂ (ρeu) vided the proper expressions for total flux jΦ . In this + ∇ · ([u − uΓ] ρeu) = work, we are concerned with mass, momentum and en- ∂t 2 (S7) ↔ ↔ u ergy conservation, thus, the fluxes for these quantities −∇ · q − σ −µρ I : ∇u − ∇ · j (1) (2) 2 d are given by: mass jρ = uρ + jd, momentum jp = ↔  ↔ ↔ (1) ↔ up+ σ −µρ I , energy je = pe + q + σ −µρ I · u. The last term in Eq. (S7) modifies the internal en- ergy due to diffusional fluxes changing the concentration. ↔ ↔ The term σ −µρ I corresponds to the momentum flux Writing Eq. (S7) in a non-conservatice manner ↔ ↔ r due to stresses and chemical species, thus, σ −µρ I ·u D e ↔ ↔ ρ u = −∇ · q − σ −µρ I : ∇u represents the work rate done by the fluid. By further Dt (S8) considering external forces as a source of momentum and  u2  + e − ∇ · j energy (if these forces exert work in the system), we ar- u 2 d rive at the final set of conservation laws, r ∂ρ where D /Dt = ∂/∂t + ur · ∇ is the material derivative, + ∇ · ([u − u ] ρ) = −∇ · j (S3) based on the relative velocity between the fluid and the ∂t Γ d control volume boundary ur = u − uΓ. Applying the first law of thermodynamics, the change in the internal ∂p ↔ ↔ energy is given dEu = δQ − δW , where heat and work + ∇ · ([u − uΓ] p) = −∇ · σ −µρ I + Σfext (S4) satisfy δQ ≤ T dS and δW = P dV −Σµ dN respectively. ∂t i i At the incompressibility limit, we have δQ = T dS. By dividing the variation of the internal energy by the mass ∂ (ρe) ↔ ↔  of the system and applying the material derivative leads +∇·([u − u ] ρe) = −∇·q−∇· σ −µρ I · u +Σf ·u ∂t Γ extto (S5) r r r −1 r D eu D s D ρ D (Ni/m) Here, ρ is the density, µ = Σ (µ ρ φ ) /ρ the averaged ≤ T − P + Σµi (S9) i i i Dt Dt Dt Dt chemical potential per unit mass, φ the volume fraction of each phase or specie, u the velocity of the fluid, jd a diffu- the last term can be re-expressed in terms of an effec- sional mass flux (whose form is irrelevant for the current tive chemical potential per unit mass as (µ/ρ) Drρ/Dt. 6

Hence, the change in the internal energy is with u = hui and P = hP i time-averaged quantities. To arrive to this equation, we use the Reynolds decomposi- Dre Drs 1 Drρ 0 u ≤ T + (P + µρ) (S10) tion for velocity u (x, t) = u (x) + u (x, t) and pressure Dt Dt ρ2 Dt P (x, t) = P (x) + P 0 (x, t). For fully developed flow, ρ∇ · (uu) = 0 and, hence, ∇P = −∇ · (ρ hu0u0i − µ∇u − and substituting Eq. (S10) into Eq. (S8) leads to the µ∇uT ). Because the Reynolds stresses ρ hu0u0i are the inequality for the entropy same at the inlet and outlet due to the fully developed 0 0 r flow assumption, the surface integral of ∇·(ρ hu u i) van- D s ↔ ↔ ρT ≥ −∇ · q − σ −µρ I : ∇u ishes, from which we arrive to Dt (S11) r  2  2L 2 1 D ρ u ∆P = − ρuτ (S18) − (P + µρ) + eu − ∇ · j R ρ2 Dt 2 d p with uτ = τw/ρ the friction velocity, τw the constant By replacing the thermodynamic definition of the en- wall shear stress, L the pipe length, and R the pipe ra- −1 thalpy per unit mass h = eu + P ρ into Eq. (S7) we dius. For the second expression of the pressure drop, we arrive at integrate over the surface the time-averaged conservation ∂ (ρh) ∂P equation of kinetic energy + ∇ · ([u − uΓ] ρh + q) = + ∇ · ([u − uΓ] P ) ∂t ∂t Z D ↔ E Z D↔ E 2 uρek + uP − τ ·u · dA = − τ : ∇u dΩ ↔ ↔ u − σ −µρ I : ∇u − ∇ · j A Ω 2 d (S19) (S12) Since for fully developed turbulent pipe flow fluctuat- ing quantities are translation invariant, we have that D↔ E Replacing the definition of the stress tensor R R A huρeki · dA = 0 and A τ ·u · dA = 0. This means ↔ ↔ ↔ σ= P I − τ and considering a pure single phase that the average flow rate of kinetic energy and the work incompressible system with a fixed control volume done by viscous stresses at inlet and outlet are equal. By (uΓ = 0), the equations for kinetic energy, internal applying the Reynolds decomposition to the pressure and energy, entropy and enthalpy respectively reduce to velocity, we arrive to

Z Z D↔ E ∂ (ρek)  ↔  ↔ hP 0u0i + P u
· dA = − τ : ∇u dΩ (S20) + ∇ · uρek + uP − τ ·u = − τ : ∇u + Σfext · u ∂t A Ω (S13) Since the velocity is zero at walls, the surface integrals reduce only to the inlet and outlet. Invoking translational

∂ (ρeu) ↔ invariance along the stream wise direction, the integra- + ∇ · (uρeu) = −∇ · q+ τ : ∇u (S14) D E ∂t tion of Pbub over the surface vanishes and, hence

1 Z D↔ E ∆P = − τ : ∇u dΩ (S21) Ds ↔ VA ρT = −∇ · q+ τ : ∇u (S15) Ω Dt R where VA = A u · dA is the volumetric flow rate. For fully developed laminar pipe flow the velocity field has ∂ (ρh) only one component in the axial direction u = (0, 0, uz), + ∇ · (uρh) = −∇ · q+ then the tensor velocity gradient in cylindrical coordi- ∂t (S16) nates is given by ∂P ↔ + ∇ · (uP ) + τ : ∇u  ∂uz  ∂t 0 0 ∂r ∇u =  0 0 0  (S22) It is important to remark that Eq. (S15), is only valid 0 0 0 at the incompressibility limit. For compressible (real) ↔ 2 flows, this equation changes to an inequality, following ∂uz
Thus, the local dissipation is τ : ∇u = µ ∂r . the second law of thermodynamics. Replacing the parabolic velocity profile  2 uz (r) = 2V 1 − (r/R) into Eq. (S21) leads to Appendix B: Pressure drop 8πµV L ∆P = − (S23) We obtain the first expression for the pressure drop, A from the time-averaged momentum equation The same result is obtained when using the parabolic velocity profile to calculate the wall shear stress in Eq. ρ∇ · (uu) = −∇P − ∇ · (ρ hu0u0i − µ∇u − µ∇uT ) (S17) (S18) 7

Appendix C: Fitting Molecular Dynamics data near the wall, we will see that the contradiction is only apparent. Pope [10] presents in his book separate equa- Molecular Dynamics simulations, are the only tools tions (5.131 and 5.132) for conservation of the mean and that can allows us to quantify simultaneously the effects turbulence kinetic energy. Using our nomenclature, these of pressure on the internal energy and the local range or- equations read der of a liquid. It is well known that changes in the molec- ∂ek ular arrangement of a liquid has an associated change in + u · ∇ek + ∇ · T = −P − E (S24) ∂t the internal energy (due to changes in the energy of each molecular bond). In the case of water at high pressures, it is possible to modify the angle of the hydrogen bonds, ∂e TKE + u · ∇e + ∇ · T’ = P − E0 (S25) while keeping and almost unchanged density. This could ∂t TKE allows us to modify greatly the internal energy with pres- sure, while keeping a constant density. As mentioned in with T = u · hu0u0i + uP /ρ − 2νhu · ∇ui, T0 = the main body of the article, for a strictly incompress- 0.5hu0 · u0u0i + u0P 0/ρ − 2νhu0u0i, the production ible fluid, the derivative of the internal energy respect to of kinetic energy P = −hu0u0i : ∇u, mean dissipation ↔ T
pressure is equal to deu/dP = −1. We used the the MD E = hτ i : ∇u + ∇u /ρ and turbulent dissipation data from Mahoney and Jorgensen [11] to estimate this D↔0 E E0 = τ : ∇u0 + ∇u0T
/ρ . A detailed analysis of con- derivative. Figure (3) shows the MD data along with a linear fit, whose slope resulted to be −0.826. This in- servation of kinetic energy, Eq. (S6), shows that the local ↔ dicates a slightly compressible behavior of the MD fluid instantanous viscous dissipation is given by E =τ : ∇u/ρ. simulated by these authors. Having a slope close to −1 Because of the symmetry of the stress tensor, the dissi- means that a large fraction of the internal energy that pation can also be expressed in terms of the strain rate ↔ is borrowed, when pressure is increased, is paid back fol- tensor as E =τ : ∇u + ∇uT
/2/ρ. In our article, we lowing a reversible path. Flow compressibility leads to derived the equation for the pressure drop assuming con- irreversible changes of the internal energy, with the con- servation of (total) kinetic energy in a pipe under fully sequent generation of sensible heat. developed flow conditions. Applying this flow regime to Eqs. (S24 and S25) and integrating over the pipe volume, 0 leads to Z Z Z -100 slope = -0.826 uP · dA = − PdΩ − EdΩ (S26)

) A Ω Ω 3 -200 (J/m -6 Z Z

10 -300 0

× 0 = PdΩ − E dΩ (S27) ) Ω Ω ref -400

(U-U Conservation of kinetic energy requires the fulfillment of

-500 both equations simultaneously. This means that under the current flow conditions, the global production of tur- -600 bulence kinetic energy balances exactly the global turbu- 0 100 200 300 400 500 600 lent component of the viscous dissipation, in agreement × -6 3 (P-Pref) 10 (J/m ) with Townsend’s hypothesis. However, by replacing Eq. (S27) into (S26), we arrive to Eq. (S21). Equation (S26) FIG. 3: MD results for the variation of internal energy with tells us that the pressure gradient is in charge of creating pressure. The data were obtained from Ref. [11]. Reference turbulence kinetic energy and overcome the mean vis- values for pressure and internal energy were taken at 1 atmo- cous dissipation, but once kinetic energy is created, it is sphere. To express the internal energy in units of energy per dissipated into internal energy requiring more energy–in unit volume, we used a molar mass for water of 18.015 g/mol; the form of pressure–to sustain the steady state and fully the density was taken from the MD data. developed flow conditions.

Appendix E: Determining the dissipation scale Appendix D: Derivation of the second equation for the pressure drop from conventional time averaged equations. To determine the dissipation scale ˜, we can make use of the friction factor defined in the main body of the article In spite that Eq. (S21) appears to be in contrast with Z Z Townsend’s hypothesis [20], where the production of tur- 2DH 1 2DH 1 0 f = 3 0dΩ + 3 0dΩ (S28) bulence kinetic energy balances the viscous dissipation V Ω Ω V Ω Ω 8

where Γ is the gamma function. Minimizing I0 respect to As discussed in the text, the component associated to α is equivalent to minimize the dissipation rate. Figure 0 vanishes for high Reynolds numbers. Thus, to have (4) presents the variation of I0 as a function of α, with a a constant friction factor as Re → ∞, the second term minimum value located at α = 2.8883. Making use of the must be constant. Defining the dissipation scale as Z 1 1 0 ˜ = 0dΩ (S29) 0.95 Ω Ω 0.9 we arrive at once to 0.85 V 3

0 0.8 ˜ = C (S30) I DH 0.75 with C a dimensionless constant. This dissipation scale is consistent with the upper bound derived by Doering 0.7 and Constantin[21] in shear driven turbulence. It is inter- 0.65 esting to note that this dissipation scale is only valid for 0.6 the Strickler regime (Re → ∞). The fact that the fric- 0 5 10 15 20 tion factor changes from the Blausius regime (dominated α by a power law with an exponent equal to −1/4), to the Strickler regime where the friction factor is constant and FIG. 4: Variation of the integral I0 respect to α (the power independent from the Reynolds number, indicates that exponent in correcting term in the dissipation controlled there is a flow transition. Just as the transition from regime). laminar to fully turbulent occurs in a Reynolds number interval, this second transition also happens in a certain dissipation scale Eq. (S30), we can express the turbulent interval, which in this case, depends on the surface rough- component of the friction factor as ness. Because our main hypothesis is that the friction factor depends exclusively on the viscous dissipation, the  C 2/3 1  g−4/3  f = 30D2 A I  (S35) different plateaus observed by Nikuradse[18] for the fric- T H 0 0 1/4 3/4 DH Re Re tion factor at high Reynolds number, indicates that the flow transition should happen only at high wave num- Comparing this expression to the one provided by Gold- bers. In other words, we should observe modifications to elfeld [14] f = G/Re1/4, we arrive at the energy spectrum only in the dissipative zone. T    C 2/3  g−4/3  G Re3/4 (λ/R) = 30D2 A I  H 0 0 3/4 Appendix F: Determining the scaling function DH Re ψ (R/η, Re, λ/R) (S36) Introducing Kolmogorov scale, η = ν3/4˜−1/4, we obtain Using the phenomenological spectrum, the turbulent component of the fraction factor is written as   G Re3/4 (λ/R) = γηg−4/3 (S37) 64 15A ˜2/3Ω Z ∞ f = 0 q1/3 exp (− [g (η, λ/D ) q]α) dq T Re 8πV 2 H 0  2/3 (S31) C 1/4 with γ = 30DH A0I0 C expressed in units of with g being a function having units of length. Following DH length1/3. We now propose the following form for G the variable change y = g (η, λ/DH ) q, the integral is expressed as    λ 1/3 Z ∞ G Re3/4 (λ/R) = A (1 − φ (z))+A φ (z)Re1/4 −4/3 1/3 α 2 2 3 2 I = g (η, λ/DH ) y exp(−y )dy (S32) R 0 (S38)

Evaluating the integral leads to 3/4 where z = Re (λ/R). The blending function φ2(z) I = g (η, λ/D )−4/3 I (S33) provides a smooth transition between the Blasius and H 0 Strickles regimes. Following Bobby and Joseph [22], we with have chosen 1  4  1 I0 = Γ (S34) φ2(z) = 1 − 2 (S39) α 3α 1 + (z/z0) 9

By equating Eqs. (S37,S38), we finally arrive at orientation-independent correlation plus corrections, we

↔ ↔ ↔ ↔   have Rijk=R0 + ∂x Rijk ∗∆x+ ∂δx Rijk ∗∆δxk+··· , γ3/4 0 0 g = η   ↔  h i3/4  where (∗) indicates convolution (see section J). Here, R0 1/4 1/4 1/3 η A2 (1 − φ2) + A3φ2Re (λ/R) is the homogeneous tensor correlation under rotational (S40) invariance. Since γ3/4 has units of length1/4, the large parenthesis We assume now that the Taylor expansion is con- is dimensionless. Thus, by choosing the dimensionless vergent and that first order corrections are the lead-  3/4 constant β = 30A I C11/12 we identify ing terms. The first and second terms represent the 0 0  departure from homogeneity and isotropy, respectively. ↔ (2R/η)1/4 Thus, the homogeneous isotropic tensor R0 can be in- ψ = (S41) terpreted as a spherical average followed by space av- h i3/4 ↔ 1/4 1/3 A2 (1 − φ2) + A3φ2Re (λ/R) erage of Rijk. With this, the energy spectrum can be finally written as Rˆ ijk (x, q ) = Rˆ0 (q) + δRˆ t (x, q) ∆x + Equation S36 implies that the turbulent component of k δRˆ (δx, δq ) ∆δx + ··· . This enables us to calcu- the friction factor is given by r k late each component of the viscous dissipation eas- D E 1/3 0
2 R ∞ 2 ˆ  λ  ily, that is, ∂iuj = qi Rjji (x, δxi = 0) dqi f = A (1 − φ (z)) Re−1/4 + A φ (z) (S42) −∞ T 2 2 3 2 R 0 0 R ∞ 2 ˆ
and ∂iuj∂jui = −∞ q(ij)Rji(ij) x, δx(ij) = 0 dq(ij). These expressions are exact and the subscripts (ij) in- This equation arises from the direct blending of the Bla- dicate a spectrum obtained along δx = δx + δx . sius and Strickles regimes (through φ ), with no further k i j 2 Hence, the connection between the energy spectrum theoretical support other than experimental observation. and the pressure drop is directly realized through the Nonetheless, and despite its phenomenological character, Darcy-Weisbach equation (defined in the main body) this form of the turbulent friction factor provide us with and Eq. (S28). Following Taylor’s [13] identities for the needed information to explain Nikuradse’s results in isotropic turbulence, we have that the fluctuating com- terms of an averaged Kolmogorov length scale that varies ponent of the dissipation rate can be approximated by with the surface roughness. By including the laminar ∞ 0 ≈ 15ν R q2Rˆ dq + O (∆x + ∆δx ). contribution, the total friction factor reads 0 0 0 k Our hypothesis is that errors introduced in the calcu-  1/3 lation of the total turbulent viscous dissipations, are of 64 −1/4 λ f = (1 − φ1)+A2 (1 − φ2) φ1Re +A3φ2φ1 second order when using an averaged spectrum. This was Re R confirmed by using the experimental data from Laufer (S43) [15]. This is the function used when fitting Nikuradse’s re- sults.

Appendix H: Spherically and spatially averaged spectrum Appendix G: The idea behind the averaged spectrum In this letter, we propose that the pressure drop can To investigate the connection between the fric- be determined from an spherically and spatially averaged tion factor, with the turbulent energy spectrum, we spectrum. We calculate the spherical average by using ↔ an arithmetic and geometric average, and compare its introduce the tensor correlation Rijk (x, δxk) = influence on the calculation fo the pressure drop. These 0 0 ui (x, t) uj (x + δxk, t) , where i and j represent the two spherical averages are given by components of the fluctuating velocity field and k the direction in which the correlation is being calculated. 1 ↔ Es (q, r) = (Eu0u0 (q, r) + Ev0v0 (q, r) + Ew0w0 (q, r)) Since δx can be chosen independently from x, R 3 k ijk (S44) is a 3-rank tensor, with eighteen independent com- ponents. The position and orientation-dependent en- ergy spectrum is directly obtained by applying the 1 E (q, r) = (E 0 0 (q, r) E 0 0 (q, r) E 0 0 (q, r)) 3 (S45) Fourier transform to the tensor correlation, this is, s u u v v w w ∞ ↔ ˆ R −2πiqk·δxk Rijk (x, qk) = −∞ e Rijk (x, δxk) dδxk. Since The geometric average is the only one that allows us to no assumption has been made in the definition of express the spherically averaged correction in the dissi- the tensor correlation, this definition of the energy pative regime Cd with the same functional form as the spectrum is exact. Using a Taylor expansion to ex- individual spectra. Thus, for the geometrically aver- press the tensor correlation in terms of a position and aged spectrum, the dissipation correction is expressed as 10

α −Bq Bu0u0 +Bv0v0 +Bw0w0 e , with B = 3 . Since the constant Since the experimental results from Laufer are presented D in the fitted spectrum is different for each velocity com- in logarithmic scale, we apply a logarithmic fit −C  D  ponent, the correction to the inertial regime 1 + α q   5 ∗ D cannot be expressed in terms of an averaged exponent ln {E (q)} = A∗ − q∗ −Beαq −C ln 1 + (S48) C. However, and as seen in the calculation of the total 3 eαq∗ dissipation rate (see results in the main text), the effects of the inertial regime are negligible. The spatial aver- where A∗ = ln (A), q∗ = ln (q) and α = 2.8883 (de- age was carried out by using a linear interpolation of the termined from the minimization of the dissipation rate, spherically averaged spectra between the measured radial neglecting the inertial correction). The results of the fit positions are presented in table 1 for each individual spectrum. For the u0 component of the fluctuating velocity field, Laufer 1 Z R [15] measured the spectrum at five different radial posi- hE (q)i = 2 Es (q, r) 2πrdr πR 0 tions. However, for the other two velocity components, n (S46) 2 X Z Ri+1 only three radial locations were reported. In order to ≈ E (q, r) rdr R2 s have the three spectra at each of the five radial loca- max i=1 Ri tions, an interpolation scheme was used to obtain the fitting parameters at the two missing radial positions. where Rmax is the closest (measured) position to the The interpolation scheme was based on the anisotropy of wall. each fitting parameter in the v0 and w0 directions, respect to u0. The anisotropy at each radial location was calcu- lated as the ratio of the fitting parameter in the v0 or w0 Appendix I: Fitting Individual spectra over the value obtained for u0. For example, we obtain B at the three pipe locations for u0u0, v0v0 and w0w0, and The function for fitting each individual spectrum is then we calculate the ratio Bv0v0 /Bu0u0 . Then, an inter- given by polating curve was fitted for this ratio (anisotropy) as a function of the radial position (individual curves for each fitting coefficient), which was then used to calculate the  −C coefficients for the missing spectra. Figures 5, 6 and 7 α D −5/3 −Bq show the measured (circles) and fitted (continuous lines) E (q) = Aq e 1 + α (S47) q spectra at three radial locations. with the parenthesis and exponential term constituting the corrections to the inertial and dissipation regimes re- 106 u’u’ spectively. Hence, the departure from the 5/3 law might 5 v’v’ be interpreted as the combined effect introduced by the 10 w’w’ inertial and dissipation regimes. 104

3

) 10 2 /s TABLE II: Fitted coefficients for individual spectra. (∗) In- 3 102 terpolated value (experimental data is not available). E (cm 101

100 (R-r)/R = 1 Stress r/R A∗ B C D η (cm) -1 u0u0 0 7.147 2.866E-4 0.4896 0.01512 0.0368 10 0 0 u u 0.309 7.207 1.272E-4 0.4017 0.01198 0.0330 10-2 u0u0 0.72 7.532 1.473E-5 0.2669 0.03245 0.0237 0.01 0.1 1 10 100 1000 -1 u0u0 0.926 8.291 2.433E-5 0.3170 0.09888 0.0208 q (cm ) u0u0 0.9918 9.481 2.144E-6 0.2530 100 .0 0.0117 0 0 v v 0 7.621 7.562E-5 0.6333 0.07428 0.0280 FIG. 5: Fitted spectra at the pipe center v0v0 0.309 7.833 4.831E-5 0.6255 0.07369 0.0252 v0v0(∗) 0.72 8.411 7.093E-6 0.4176 2.23390 − − − v0v0 0.926 9.382 1.303E-5 0.4506 20.4591 0.0148 v0v0(∗) 0.9918 10.771 1.201E-6 0.3429 22033.4 − − − w0w0 0 7.684 8.130E-5 0.6434 0.07824 0.0278 w0w0 0.309 8.030 9.208E-5 0.6483 0.06212 0.0259 w0w0(∗) 0.72 8.610 1.726E-5 0.4030 0.97417 − − − Appendix J: Taylor expansion w0w0 0.926 9.600 3.415E-5 0.3940 6.5343 0.0157 0 0 w w (∗) 0.9918 11.056 3.209E-6 0.2852 6695.1 − − − To explain the meaning of the Taylor expansion used in the main body of the article, we expand a function 11

6 10 Since xi can be chosen arbitrarily, we can average all u’u’ 5 v’v’ these equations to obtain 10 w’w’

104

3 N

) 10 2 1 X 0 /s φ (x) = (φ (x ) + φ (x )(x − x ) + 3 i i i 102 N i=0 (S50) E (cm 101  1 00 2 φ (xi)(x − xi) + ... 100 (R-r)/R = 0.691 2

10-1

10-2 0.01 0.1 1 10 100 1000 Clearly, each term of this expansion represents an aver- q (cm-1) aged quantity. If we now go to the continuum limit, we arrive to

FIG. 6: Fitted spectra at (R − r) /R = 0.691 1 D E φ (x) = hφ (y)i+hφ0 (y)(x − y)i+ φ00 (y)(x − y)2 +... 106 2 u’u’ (S51) 5 v’v’ 10 w’w’ with 104

3 ) 10 T/2 2 Z

/s 1 3 102 hφ (y)i = lim φ (y) dy T T →∞ −T/2

E (cm 1 10 Z T/2 0 1 0 100 (R-r)/R = 0.074 hφ (y)(x − y)i = lim φ (y)(x − y) dy T T →∞ −T/2 -1 10 Z T/2 D 00 2E 1 00 2 10-2 φ (y)(x − y) = lim φ (y)(x − y) dy 0.01 0.1 1 10 100 1000 T T →∞ −T/2 q (cm-1) (S52)

FIG. 7: Fitted spectra at (R − r) /R = 0.074 The integrals of the higher order terms represent, evi- dently, a convolution (denoted by ∗). Thus, we arrive to

φ (x) around a point x0

1 0 1 00 2 φ (x) = φ (x ) + φ0 (x )(x − x ) + φ00 (x )(x − x )2 + ... φ (x) = hφ (y)i + φ (y) ∗ (x − y) + φ (y) ∗ (x − y) + ... 0 0 0 2 0 0 2 (S53) or This peculiar form of the Taylor expansion allows us the 1 φ (x) = φ (x ) + φ0 (x )(x − x ) + φ00 (x )(x − x )2 + ... express any function in terms of its expected value, plus 1 1 1 2 1 1 averaged corrections. or 1 φ (x) = φ (x ) + φ0 (x )(x − x ) + φ00 (x )(x − x )2 + ... n n n 2 n n (S49) [1] C. Barenghi, L. Skrbek, and K. Sreenivasan, Proc. Natl. (1998). Acad. Sci. USA 111, 4647 (2014). [8] M. Maltrud and G. Vallis, J. Fluid Mech. 228, 321 [2] T. Tran, P. Chakraborty, N. Guttenberg, A. Prescott, (1991). H. Kellay, W. Goldburg, N. Goldenfeld, and G. Gioia, [9] L. Smith and V. Yakhot, Phys. Rev. Lett. 71, 352 (1993). Nature Phys. 6, 438 (2010). [10] S. Pope, Turbulent flows (Cambridge University Press, [3] L. F. Richardson, Weather Prediction by Numerical Pro- 2000). cess (Cambridge University Press, 1922). [11] M. Mahoney and W. Jorgensen, Journal of Chemical [4] A. Kolmogorov, Dokl. Akad. Nauk. SSSR 30, 299 (1941). Physics 112, 8910 (2000). [5] A. Kolmogorov, Dokl. Akad. Nauk. SSSR 32, 19 (1941). [12] G. Gioia and P. Chakraborty, Phys. Rev. Lett. 96, [6] G. Eyink and K. Sreenivasan, Review of Modern Physics 044502 (2006). 78, 87 (2006). [13] G. Taylor, Proc. Roy. Soc. A 151, 421 (1935). [7] J. Maurer and P. Tabeling, Europhysics Letters 43, 29 [14] N. Goldenfeld, Phys. Rev. Lett. 96, 044503 (2006). 12

[15] J. Laufer, National Advisory Committee for Aeronautics [19] A. Badillo, Phys. Rev. E 86, 041603 (2012). (1954). [20] A. Townsend, The structure of turbulent shear flows [16] B. J. Rosenberg, M. Hultmark, M. Vallikivi, S. C. C. (Cambridge University Press, 1976). Bailey, and A. J. Smits, J. Fluid Mech. 731, 46 (2013). [21] C. Doering and P. Constantin, Phys. Rev. Lett. 69, 1648 [17] A. Musker, Journal of The American Institute of Aero- (1992). nautics and Astronautics 17, 655 (1979). [22] Y. Bobby and D. Joseph, Physica D 239, 1318 (2010). [18] J. Nikuradse, National Advisory Committee for Aeronau- tics (1965).