You are on page 1of 11

Acta Mechanica 163, 39–49 (2003) DOI 10.

1007/s00707-003-1008-3

Acta Mechanica
Printed in Austria

Constructal multi-scale structure for maximal heat transfer density
A. Bejan, Durham, North Carolina, and Y. Fautrelle, Grenoble, France
Received January 3, 2003 Published online: June 12, 2003 Ó Springer-Verlag 2003

Summary. This paper presents a new concept for generating the multi-scale structure of a finite-size flow system that has maximum heat transfer density–maximum heat transfer rate installed in a fixed volume. Laminar forced convection and parallel isothermal blades fill the volume. The spacings between adjacent blades of progressively smaller scales are optimized based on constructal theory: the goal is maximum heat transfer density. The smaller blades are installed in the fresh-fluid regions that sandwich the tips of the boundary layers of longer blades. The overall pressure difference is constrained. As the number of length scales increases, the flow rate decreases and the volume averaged heat transfer density increases. There exists a smallest (cutoff) length scale below which heat transfer surfaces are no longer lined by distinct (slender) boundary layers. Multi-scale flow structures for maximum heat transfer rate density can be developed in an analogous fashion for natural convection. The constructal multi-scale algorithms are deduced from principles, unlike in fractal geometry where algorithms are assumed.

1 Geometry
A key result of constructal theory is the prediction of optimal spacings for the internal flow structure of volumes that must transfer heat and mass to the maximum. This body of work comprises both forced and natural convection, and is reviewed in [1], [2]. Optimal spacings have been determined for several configurations, depending on the shape of the heat transfer surface that is distributed through the volume: stacks of parallel plates, bundles of cylinders in crossflow, and arrays of staggered plates. In each configuration, the reported optimal spacing is a single value, i.e., a single length scale that is distributed uniformly through the available volume. Figure 1 shows qualitatively why an optimal spacing exists. The heat-generating blade shown at the top of the figure is one of many parallel blades that fill a much larger package. Isolated in Fig. 1 is the volume allocated to a single blade. The size of this volume is fixed, and is represented by the rectangular space V shown in Fig. 1a–c. The shape of V is not fixed. The fluid sweeps the blade and convects heat. The boundary layers are regions that work (are active) in a heat transfer sense. When V is considerably wider than the boundary layers (Fig. 1a), most of V is occupied by coolant that does not help the heat transfer enterprise. When V is much narrower than the boundary layers (Fig. 1c), most of V is occupied by thermally fully developed flow, i.e., by coolant that warms up in the downstream direction. Such a fluid is overworked, and becomes poorer as a coolant.

[3]). if the structure is to be uniform. the geometries of single-spacing structures vary periodically. At the most. 2.  À1=2 Ux : ð1Þ dðxÞ ffi 5x m In this expression U and m are the free stream velocity and the kinematic viscosity. 1. D1 . The optimal volume shape for the flow associated with one blade The best configuration is in-between. Fautrelle heat-generating blade V unused volume a V b overworked fluid V c Fig.. This observation is the same as taking the argument of Fig. Three such blades are shown in Fig. The technique consists of placing more heat transfer in regions of the volume HL0 where the boundary layers are thinner. assume laminar forced convection driven by the imposed pressure difference DP. For concreteness. D2 . Is the stack of Fig.g. 2 the best way to pack heat transfer into a fixed volume? It is. In Fig. 1. 1b. and have negligible thickness. The boundary layers merge in the plane of the outlet.40 cold fluid convective body A. . or eliminated. 2 can be improved if more length scales (D0 . .) are available. Those regions are situated immediately downstream of the entrance plane. 2 is uniform. Here the volume is just wide enough to house the flow regions that work. where L0D0 has replaced the two-dimensional volume V of Fig. . The key observation is that the structure of Fig. Furthermore. but only when a single length scale is to be used. x ¼ 0. the wedges of fluid contained between the tips of opposing . 1 to a finer level: regions that do not work in a heat transfer sense must either be put to work. as in the case of arrays of cylinders and staggered plates. Bejan and Y. 2. the fluid Prandtl number is of order 1. so that the velocity and thermal boundary-layer thicknesses are both represented by the Blasius thickness (e. Assume also that the blades are isothermal. because it does not change from x ¼ 0 to x ¼ L0. Tw . The structure of Fig. This transversal dimension of V represents the optimal spacing between the parallel blades in the stack. that is. The spacing D0 is the single length scale that is distributed uniformly through the twodimensional flow structure of length L0 and width H. Fig.

The approximation is due to the assumption that the presence of the L1 boundary layers does not affect significantly the downstream development (x > L0/4) of the L0 boundary layers. They can be involved if heat-generating blades of shorter length (L1) are installed on their planes of symmetry. the image generated by the algorithm (3) is not a fractal. We are consistent. 1. but the shape of the boundary layer region is the same for all the blades. L2).Multi-scale structure for maximal heat transfer density 41 TW U. The order-of-magnitude correctness of this assumption is clear. because of this. Because d increases as x1/2. Each new L1 blade is coated by boundary layers described by Eq. 4 1 Di ¼ DiÀ1 2 ði ¼ 1. because the blades are all swept by the same flow (U). [5]. every structure with merging boundary layers will be optimal. 2. no matter how complicated. and the smallest size (Dm. . . (1). ∆P d (x) D0 H 0 x L0 Fig. we invoke one more time the optimal packing principle of Fig. This assumption is made for the sake of simplicity. Two such blades are shown in the upper-left corner of Fig. In other words. This new design is shown in Fig. L2 ¼ L1/4. as we will see in Eq. The length scales become smaller (L0. 3. The sequence of decreasing length scales is finite. It is a Euclidean image [4]. . and it comes from geometry: the edges of the L1 and L0 boundary layers intersect when 1 L1 ¼ L0 : 4 ð2Þ Note that by choosing L1 such that the boundary layers that coat the L1 blade merge with surrounding boundary layers at the downstream end of the L1 blade. ð3Þ where we shall see that m is finite. as in all the constructal tree structures [1]. mÞ. Lm) is known. 3. The wedges of isothermal fluid (T0) remaining between adjacent L0 and L1 blades can be populated with a new generation of even shorter blades. . the boundary layers of the L1 blade merge with the boundary layers of the L0 blades at a downstream position that is approximately equal to L0/4. The merging and expiring boundary layers are arranged according to the algorithm 1 L i ¼ L i À1 . Optimal package of parallel plates with one spacing boundary layers are not involved in transferring heat. . 2. L1. not infinite. and. T0. (21).

because there are as many L1 blades as there are D0 spacings. At scales smaller than L1. . 3. One is attractive: the total surface of temperature Tw installed in the HL0 volume increases. n0 ¼ H . 2. Bejan and Y.42 A. the number of blades of one size doubles with every step. Let n0 be the number of L0 blades in the uniform structure of Fig. ni ¼ 2niÀ1 . 3. D0 ð4Þ where  1=2 mL 0 D0 ffi 2dðL0 Þ ffi 10 : U ð5Þ The number of L1 blades is n1 ¼ n0 . 3. The other is detrimental: the flow resistance increases. i ¼ 2. . m: ð6Þ Two conflicting effects emerge as the structure grows in the sequence started in Fig. The important question is how the volume is being used: what happens to the heat transfer rate density as complexity increases? 2 Heat transfer The total heat transfer rate from the Tw surfaces to the T0 fluid can be estimated by summing up the contributions made by the individual blades. the flow rate driven by the fixed DP decreases. . the Pohlhausen solution for Prandtl numbers of order 1 [3]. . and so does the heat transfer rate associated with a single boundary layer. The heat transfer rate through one side of the L0 blade is equal (in an order of magnitude sense) to the heat transfer rate associated with a laminar boundary layer. we note that the number of blades of a certain size increases as the blade size decreases. cf. Fautrelle L2 D2 D0 D1 L1 0 x L0 Fig. . Optimal multi-scale package of parallel plates To complete the description of the sequential construction of the multi-scale flow structure.

and it is related to all the friction forces felt by the blades.   ULi 1=2 q0i ffi 1:328kDT ni : ð9Þ m The heat transfer rate of all the blades is the sum   m X UL0 1=2 0 0 q ¼ qi ffi 1:328kDT n0 S. 3 is   UL0 1=2 0 00 0 L0 ffi 1:328kDT n0 q0 ¼ 2n 0 q : ð8Þ m The same calculation can be performed for any group of blades of one size. [3]). We rely on the same approximation as in the case of heat transfer. (10) is not specified. and their combined contribution to the total heat transfer rate of the package of Fig. and k is the fluid thermal conductivity. DT ¼ Tw À T0 . S¼1þ m : 2 ð12Þ ð10Þ ð11Þ This analysis confirms the trend noted at the end of the preceding section: the total heat transfer rate increases monotonically as the complexity of the structure (m) increases. (8). in view of Eqs. 3 Fluid friction It is necessary to evaluate the flow resistance of the multi-scale structure. The total force felt by the blades of size Li is Fi ¼ 2ni si Li ffi 1:328qðmLi Þ1=2 ni U 3=2 : The total force for the multi-scale package is F¼ m X i ¼0 ð14Þ Fi ffi 1:328qðmL0 Þ1=2 n0 U 3=2 S: ð15Þ . respectively. There are 2n0 such boundary layers.Multi-scale structure for maximal heat transfer density 43   00 q UL0 1=2 0 L0 ffi 0:664 : DT k m ð7Þ 2 00 Here q 0 [W/m ] is the L0-averaged heat flux.. Their total heat transfer rate q0i [W/m] is given by a formula similar to Eq.g. and estimate the friction force along one face of one blade by using the solution for the laminar boundary layer (e. because the velocity U that appears in Eq. Li. The pressure difference DP is specified. 2 Cfi ¼ 1:328 ðUL0 =mÞ1=2 : ð13Þ Here si and Cfi are the averaged shear stress and skin friction coefficient. (3) and (6). m i ¼0 where S is the dimensionless geometric parameter       n1 L1 1=2 n2 L2 1=2 nm Lm 1=2 þ þÁÁÁ þ S¼1þ n0 L0 n0 L0 n0 L 0 or. 1 si ffi Cfi qU 2 . in which n0 and L0 are replaced by ni and Li .

the right side of Eq. The shortest blade length Lm below which the boundary layer heat transfer mechanism breaks down is Lm  Dm : In view of Eqs. Fautrelle This force is balanced by the longitudinal force imposed on the control volume. (5) and (17) we find the smallest scale.. DT . (3). q0 H ffi 0:36 Be1=2 S1=2 . (10) and (17). i. The alternative to using the global conductance is the heat transfer rate density. Equation (22) establishes m as a slowly varying monotonic function of Be1/4. Both quantities increase with the applied pressure difference (Be) and the complexity of the flow structure (S). . 2 and 3 are valid when boundary layers exist. we obtain the order of magnitude of the average velocity of the fluid that bathes the structure:  1=2 DP : ð17Þ U ffi 2:7 qS This result confirms the second trend anticipated at the end of Sect. This function can be substituted in Eq. by combining Eqs. Be ¼ DPL2 0 : la ð19Þ In this expression l and a are the fluid viscosity and thermal diffusivity. (15). (10) and (17) yields the dimensionless global thermal conductance. H . L0 Þ. L0 kDT ð18Þ where Be is the dimensionless pressure drop that Bhattacharjee and Grosshandler [6] and Petrescu [7] termed the Bejan number. (20). by using Eqs. 4 Heat transfer rate density: the smallest scale Putting together the results of the heat transfer and fluid flow analyses. Figure 3 makes it clear that boundary layers are less slender when their longitudinal scales (Li ) are shorter. To be distinct. q000 ¼ q0 =HL0 . 1: the flow slows down as the complexity of the structure (S. the effect of increasing S is beneficial from the point of view of packing more heat transfer in a given volume.e. which occurs at the level m given by  m1=4 2m 1 þ  0:17Be1=4 : ð22Þ 2 In view of the order-of-magnitude character of the analysis based on Eq. (18) to see the complete effect of Be on the global heat transfer performance. this means that L0  2m D0 : ð21Þ ð20Þ Finally. in spite of the conflicting effects of S in Eqs. 3 and the analyses of Sects. Bejan and Y. Optimized complexity is the route to maximal global performance in a morphing flow system [1]. boundary layers must be slender. (16) with n0 ¼ H =D0 and the D0 formula (5). In conclusion. How large can the factor S be? The answer follows from the observation that the geometry of Fig. when they are distinct. or m) increases. (22) is essentially (Be/103)1/4.44 A. when its constraints are specified ðDP. we find how the structure performs globally. Eliminating U between Eqs. DPH ¼ F : ð16Þ Finally.

4a. The monotonic effect of m is such that each new length scale (m) contributes to global performance less than the preceding length scale (m À 1). Eq. then the resulting image is a fractal and m and the heat transfer density (23) tend to infinity. If the number of L0 blades in the package is n0 . because the fluid must be able to flow through the structure. a b Fig. more like the drawing shown in Fig. (1). 5. the required complexity (m) increases monotonically with the imposed pressure difference (Be). then the number of L1 blades is n1 ¼ 2n0 . This rule follows from the nonlinear shape of each boundary layer region. new branches form on both sides of the tip of an existing blade. 4b. Is this type of growth more advantageous from the point of view of packing maximal heat transfer into a given volume? The symmetric-growth alternative to Fig. and smaller smallest scales. 4a). The structure becomes not only more complex but also finer. 3 is arbitrarily continued ad infinitum. the new (smaller) branches grow on only one side of the tips of the existing branches. If the construction started in Fig. and that in this case the tree is asymmetric. because it shows that the population of Li blades can be viewed as the branches of a tree structure. In one D0 spacing we placed two L1 blades separated by the distance D1 ¼ D0 =3. 3 is shown in Fig. It is a useful way to think. Asymmetric (a) versus symmetric (b) trees built with blades of decreasing lengths and increasing numbers . In Fig.Multi-scale structure for maximal heat transfer density 45  1=2 q0 H 1 ffi 0:36 Be1=2 1 þ m : L0 2 kDT ð23Þ In conclusion. 5 Symmetric growth Imagine that the blades of the multi-scale structure of Fig. The new boundary layers merge with neighboring boundary layers when L1 ¼ L0 =9. 3 are connected solidly at their downstream end – connected to the neighboring blade of immediately larger scale (see Fig. Nature strikes us with images of tree-shaped structures endowed with symmetry. Here. This is certainly not a flow design recommendation. More flow means more length scales. 4.

. After using Eq. 3 ð26Þ where ms is the number of gap filling steps used in the symmetric construction. 6. We see that the complexity (ms ) increases as the flow becomes stronger (i. (15). (11). 3 ði ¼ 1. by comparing Eqs. Fautrelle D0 D1 L1 L0 Fig. Furthermore. The smallest blade length Lms cannot be smaller than Lms  Dms . with the difference that Eq. by placing two new blades in the near-entrance region of every spacing. (17) for U. (23). (20). Fig. The optimal symmetric structure is simpler because it is based on fewer construction steps than the asymmetric structure. 2. . (17) and (18). namely  1=2 q0s H 1=2 2 ffi 0:36 Be 1 þ ms .e. as Be increases). we arrive at  1=4 2  ðBe=103 Þ1=4 : ð28Þ 3ms 1 þ ms 3 Equation (28) is plotted in Fig.46 A. Bejan and Y. with Ss in place of S. 5. (28). The relation between ms and m is roughly ms =m  ln 2= ln 3 ¼ 0:63. The number ms can be determined based on the same argument as in Eq. . and is a reminder that the number of construction steps (ms ) is not the same as in the asymmetric construction (m). 3. . 9 ni ¼ 3iÀ1 2n0 1 Di ¼ DiÀ1 . (10). The dimensions of the growing structure follow the algorithm: 1 L i ¼ L i À1 . ms Þ: ð24Þ ð25Þ The heat transfer and fluid mechanics analyses lead again to Eqs. Optimal multi-scale package of parallel plates with two smaller blades around each larger blade The construction can be continued toward smaller scales. Eq. ð27Þ where Lms ¼ 9Àms L0 and Dms ¼ 3Àms D0 . we conclude that ms < m. This trend is the same as in the asymmetric structure. (22). The subscript s refers to the symmetric construction of Fig. 5. (28) and (22) when Be is specified.. and is only weakly dependent on Be. (5) for D0 and Eq. ð29Þ L0 3 kDT for which the function ms(Be1/4) is furnished by Eq. The performance of the symmetric structure is given by the same relation as Eq. The performance relative to the asymmetric case of Eq. (11) now yields 2 S ¼ 1 þ ms . (23) is measured by the ratio .

The number of scales of the multi-scale flow structure (m) increases slowly as the flow becomes stronger. Boundary layers become thinner as Be increases. The starting regions of the boundary layers are thinner.3) 0. Larger surfaces are surrounded by thicker working volumes.5) q′s(Fig. ms) as functions of the imposed pressure difference (Be). but smaller than 1. They are surrounded by fresh flow that can be put to good use: heat transfer blades with smaller and smaller lengths can be inserted in the fresh fluid that enters the smaller and smaller channels formed between existing blades. 5 is only marginally less effective than the asymmetric packing of Fig. because m and ms are functions of Be. At the same time.Multi-scale structure for maximal heat transfer density 6 5 m 4 ms 3 ms 47 m Fig. the total heat transfer surface . we described a new concept for generating a multi-scale flow structure that maximizes the heat transfer density installed in a fixed volume. The bottom graph shows the relative heat transfer density of the symmetric vs. and the flow rate decreases.5 2 1 0 1 10 100 (Be / 103)1/4 1 q′s(Fig. The symmetric structure of Fig. 3.95 0.3 Fig. asymmetric multi-scale structures q0s ffi q0  1 þ 2ms =3 1 þ m=2 1=2 ð30Þ which is close to 1.9 1 10 (Be / 103)1/4 100 Fig. The flow strength is accounted for by the pressure difference maintained across the structure (Be). The number of length scales (m. 6 Conclusions In this paper. the working volume has a thickness that scales with the square root of the length of the streamwise heat transfer surface. The structure becomes less permeable. Two trends compete as the number of length scales increases. In laminar forced convection. and this means that more small-scale heat transfer blades can be inserted in the interstitial spaces of the entrance region of the complex flow structure. Figure 6 shows that the ratio q0s =q0 is a function of Be. 6. The method consists of exploiting every available flow volume element for the purpose of transferring heat.

. at least qualitatively. then the role of the overall pressure difference DP is played by the difference between two hydrostatic pressure heads. Bejan and Y. Finally. (19) we find that the dimensionless group that replaces Be in natural convection is the Rayleigh number Ra ¼ gbDTL3 0 : am ð32Þ Other than the Be ! Ra transformation. as a flow mechanism on which to build the multi-scale structure. fractal geometry is descriptive. i. [8]. Forced convection was used in this paper only as a working example. the constructal theory used in this paper generated the construction algorithms (Eqs. Constructal theory delivers the algorithm as an optimization result. Eqs. and were distributed throughout the volume. In brief. the algorithms are selected such that the resulting structures resemble flow structures observed in nature. (24) and (25)). the missing link has been the origin of the algorithm [10]. The algorithms were generated by the constructal principle of optimization of global performance subject to global constraints [1]. 2 is rotated by 90° counterclockwise. This increase occurs at a decreasing rate. . The net and most important result is that the heat transfer density increases as the number of length scales increases. the features described for forced convection in this paper. Contrary to the fractal approach. all the features due to the generation of multi-scale blade structure for natural convection should mirror. By substituting the DP expression (31) into Eq. not predictive [1]. This principle was invoked every time the optimal spacing between two blades was used. and g is the gravitational acceleration aligned vertically downward (against x in Fig. A completely analogous multi-scale structure can be deduced for laminar natural convection. if the structure of Fig. With regard to fractal geometry and why some fractal structures happen to resemble natural flow structures. We showed that there exists a characteristic length scale below which heat transfer surfaces are no longer lined by boundary layers. For this reason. the flow architecture constructed in this paper is a theoretical comment on fractal geometry. from the constructal principle [1]. (6). the effective DP due to buoyancy is DP ¼ qgbDT L0 . [9]. The complete analogy that exists between optimal spacings in forced convection and natural convection was described by Petrescu [7]. b is the coefficient of volumetric thermal expansion. Fautrelle increases. Optimal spacings were assigned to all the length scales. including the smallest scale cutoffs. 2). (22) and (28). The heat transfer augmentation effect due to the insertion of a single plate in the entrance region of the parallel-plates channel was noted and documented numerically by Aihara et al. ð31Þ where DT is ðTw À T0 Þ. If the Boussinesq approximation applies. Fractal structures are generated by assuming (postulating) certain algorithms. (3). and if the flow is driven upward by the buoyancy effect. and the other for the L0 fluid column of temperature Tw . meaning that each new (smaller) length scale contributes less to the global enterprise than the preceding length scale. one for the fluid column of height L0 and temperature T0 . This smallest scale serves as cutoff for the algorithm that generates the multi-scale structure. It is not a theory.48 A.e. In much of the current fractal literature.

O. Department of Mechanical Engineering and Materials Science. [10] Nottale.: Shape and structure. from engineering to nature. Akaku. UK: Cambridge University Press 2000.: The formation of a wall jet near a high temperature wall under microgravity environment. Italy.: Fractal space-time and microphysics. Boca Raton: CRC Press 1996. [9] Bradshaw. 983 (2001). [5] Avnir. 1996.. P. [2] Kim. A. O. W. S. Gori. W. Durham. AIAA J. 765. Lidar. A. Biham. U..S. J. L.: Shape and structure. 95. Box 90300. Int. Rome. Grosshandler.P. Authors’ addresses: A. T. L. Lee. Singapore: World Scientific 1993. 711–716 (1988). [3] Bejan. [4] Bejan.: Advanced engineering thermodynamics.: Is the geometry of nature fractal? Science 279. S. J. Duke University. Y. A. p. New York: Wiley 1997. 39. 38402 Saint Martin d’Heres Cedex. 2nd ed. D. Malcai.. Bejan. from engineering to nature. F. Cambridge..: Comments on the optimal spacing of parallel plates cooled by forced convection.: Augmentation of free-convection heat transfer between vertical parallel plates by inserting an auxiliary plate. 2nd European ThermalSciences and 14th UIT National Heat Transfer Conference... M.. Sasaco.Multi-scale structure for maximal heat transfer density 49 References [1] Bejan. 2nd ed. S. ASME HTD 96. [7] Petrescu. EPM-MADYLAM Laboratory ENSHMG. A. France . S. D. Institut National Polytechnique de Grenoble. 1283 (1994).A. B. North Carolina..: Convection heat transfer. [8] Aihara. 39–40 (1998). [6] Bhattacharjee. Ohara. Heat Mass Transfer 37.. Fautrelle..: Air cooling technology for electronic equipment. New York: Wiley 1995. T. NC 27708-0300.