PIT: $0009-2509(97}00090-0 ‘Chrcat Engineering Sect, No.2, No. 24 pp 4681-4698, 1997 1997 Elser Sonne Ld" A night served Pete n Gest Bastin nos 56597 $100 000 Dimensional analysis, scaling, and orders of magnitude Gianni Astarita’ Dipartimento di Ingegneria dei Materiali e della Produzione, Universita’ di Napoli Federico I, Piazzale Tecchio, 80125 Napoli, Italy Abstract—The modern viewpoint which regards dimensional analysis as the basis for a theory of scaling is discussed. Several specific examples are illustrated to show how an ingenuous application of dimensional analysis and scaling may yield some very significant insight on the solution of the problem considered. Conclusions of a hopefully somewhat general nature are extracted from the examples considered. © 1997 Elsevier Science Ltd Keywords: Dimensional analysis; scaling; orders of magnitude INTRODUCTION Professor Sharma's contributions to chemical engin- ering science have been mostly in what might be called ‘classical’ chemical engineering, or in what is nowadays described as the core areas’. The core areas have (in my opinion, unfortunately) rather come out ‘of fashion in the last decade, but it seems appropriate that a paper dedicated to the 60th birthday of Prof. Sharma should be one in classical chemical engineer ing, and 1 am convinced that the real essence of chemical engineering is the ability to reduce an inex- tricably complex real life problem to something which ccan be (at least in the best possible case) reduced to aback of the envelope’ calculation. Such calculations are invariably based on dimensional analysis, appro- priate scaling, and order of magnitude estimates. The concept of ‘dimensions’ goes back to Greek science; Euclid (ca. 300 BC; see Heiberg and Menge, 1883-1916, Heath, 1921) considered restrictions im- posed upon geometrical operations based on require ments of homogeneity; Ptolemy (ca.120 A.D; see Heiberg, 1887; Schnabel, 1938) wrote a book entitled “On Dimension’. The restrictions were in fact overly strict ones, since for instance multiplication of four quantities all having dimensions of a length was re- garded as forbidden (Heath, 1921) (dimensions and directions in space were confused with each other), and, @ fortiori, multiplication (or division) of two quantities having different dimensions was also for- bidden. Indeed, one has to go all the way to Newton (1678) to find the idea that one could divide (or multi- ply) two quantities with different dimensions—that what nowadays are called ‘derived’ or ‘secondary’ quantities in dimensional analysis are legitimate ones. Euler’s(1763) significant contributions to the develop- ment of dimensional analysis went largely unnoticed until the second half of the 19th century. ‘The modern viewpoint on dimensional analysis wa probably originated by Rayleigh: in the index to his “Theory of Sound’ (1877-1878) there is the entry “Method of Dimensions’. One may discern three suc cessive phases of development of dimensional analy- sis: in its original formulation it was called “the principle of similitude’ (Rayleigh, 1915), and it had to do essentially with the identification of the dimension- less groups relevant to any given problem, and thus of some results (which only in hindsight may be regarded as elementary) ofthe type: the period of a pendulum is, independent of its mass’. The best discussion ever of classical dimensional analysis (which is also quite relevant to the more recent viewpoint) is more than 70 years old (Bridgman, 1922). We urge the reader to study that book—it's not a long one, and it is easy to read, and once one has reached its end lots of things cone had learned by heart but did not really under- stand have become crystal clear. [Earlier versions of the r-theorem usually attributed to Buckingham (1914) are due to Rayleigh (1879), Vashy (1892), Riabouchinsky (1911), and in fact the theorem can be traced back to Fourier (1822); the early history of dimensional analysis is discussed by Macagno (1971)]. Applications of dimensional analysis to en- xineering problems was by no means unheard of in the pioneers’ time, as witnessed by the works of such giants as Einstein (1911), Reynolds (1883) and Lord Rayleigh (1892, 1904, 1915). [In his 1915 paper, Rayleigh wrote a sentence which is perhaps the stron- gest tribute ever given to the power of dimensional analysis: “It happens not infrequently that results in the form of ‘laws’ are put forward as novelties on the basis of elaborate experiments, which might have been predicted a priori after a few minutes’ consideration’. Perhaps, nowadays it takes more than a few minutes, 46814682 but the concept is still valid; 80 years after Rayleigh, Aris (1995) has written: ‘I would suggest that the engineer's use of dimensionless parameters is a notion as profound and far-reaching as any in his reper- toite’]. Bertrand (1878), Carvalho (1891), Clavenad (1892), and Vashy (1892) were the French counterpart of the English work carried out mostly by Rayleigh. The second phase of dimensional analysis was the one where it was regarded as the basis for scaleup (and, had this period actively survived to the present, also as 2 basis for the scaledown needed when minia- turization has become so important). The scaleup concept is based on the fact that, provided the values of all but one of the dimensionless groups is the same in the (small scale) model as they will be in the (large scale) prototype, so will the value of the last one. The idea of using small models goes back to Vitruvius (56 B.C.) and Leonardo da Vinci (ca. 1500) who writes “Vitruvius says that small models are of no avail for ascertaining the behaviour of large ones; and I here propose to prove that this conclusion is a false one’ Leonardo's approach works in the simplest cases, but it does not take much of a complication to become hopeless, vindicating Vitruvius’ viewpoint (Astarita, 1985). This approach reached its peak in the 1950s and 1960s, when interminable lists of dimensionless groups were compiled and recorded (Boucher and Alves, 1959; Catchpole and Fulford, 1966; Fulford and Caichpole, 1968; Klinkenberg and Mooy, 1948; for an understanding of the prevailing viewpoint at the times, see Boelter et al, 1946; Johnstone and Thring, 1957; Langhaar, 1951; Murphy, 1950; Sedov, 1959; Zierip, 1971). This period also witnessed the tentative applications of dimensional analysis to s ology (Dodd, 1963), psychology (Lehman and Craig, 1963), and economy (De Jong, 1967) The third, contemporary phase is the one where dimensional analysis is regarded as giving the basis for appropriate scaling. The importance of scaling cannot be overemphasized; suffice it to say that the development of scaling laws for polymers by De Gennes (1979) has made macroscopic (as distin- guished from high-energy, subatomic particle) physics again deserving of the Nobel prize. What, very simply stated, I mean by scaling is as follows. Consider problem the mathematical description of which con- tains some (dimensional) variable y—for instance, distance from the centerline in a cylindrical tube. 1s, a value of y= 1m large, small, or about average? yhas a maximum value equal to the radius of the tube Y: hence the new variable y’ = y/¥ is not only dimen- sionless, itis known to be of order unity: dividing by Y is the appropriate scaling of the y variable. In this example the scale Y is obvious, in many cases what the appropriate scale may be has to come out from a rather ingenuousos analysis of the problem at hand—which also means one needs to know what the problem at hand is, which is by far not a trivial requirement (Astarita 1990). Aris (1995) gives a de- lightful poetic interpretation of appropriate scaling by comparing Keats’ dystich: ‘St Agnes’ Eve—ah, bitter G. Astarita chill it was/the ow for all his feathers was a-cold’ with a possible prosaic alternative: “St Agnes’ Eve—ah, bitter chill it was/the temp. was ten above, right on the line’ the ows feathers provide the appropriate scale for temperature. The discussion about scaling in the preceeding paragraph has already established the connection with the concept of order of magnitude, This is possibly the most difficult concept to grasp for an engineering student; and the more computers, pocket calculators and the like become available, the more difficult it becomes to grasp it. In the old times of the ide rule, the student soon learned the hard way that trying to keep track of digits she could not read on the slide rule was hopeless. Today one gets cight digits as, an absolute minimum, and an instructor who tries to point out that there really are only three important numbers, 0, 1 and ob, is looked upon as one who is trying to be funny. And yet, given two phenomena, if wwe learn to compare them in the appropriate way, what possibilities do we foresee? Either the first one can be dismissed as irrelevant (their relative import- ance is of order zero), or both are relevant (of order 1), ‘or the second one can be neglected (of order a0). Stokes law for the friction on a sphere is quite good up to Re ~ 08, after all, and yet it is based on assuming Re = 0.So what difference does it make whether Re is, say, 0.02 or 0.2? And yet today’s engineering student (assuming he is a good student who has some feeling for orders of magnitude estimates) is likely 10 come up with the statement that he can use Stokes’ law because Re = 1.5467032E-2 which is less than unity (it is indeed) The three techniques of dimensional analysis, ap- propriate scaling, and intelligent estimation of orders of magnitude, if used jointly, constitute an extremely powerful tool for the chemical engineer. Very often more than nine-tenths of what one can ever hope to know about a problem can be obtained from this tool, without actually solving the problem; the remaining one-tenth requires painstaking algebra and/or lots of computer time, it adds very little to our understanding of the problem, and if we have not done the first part right, all that the algebra and the computer will produce will be a lot of nonsense. Of course when nonsense comes out of a computer people have a lot of respect for it, and that is exactly the problem. Some sort of Murphy’s law of course applies here, as it always does. The tool is a very powerful one, and for that very reason it is difficult to use, and it cannot be taught as an algorithm. [Algorithms are never very powerful; they only solve problems within the class for which the algorithm has been developed, and one does not develop an algorithm unless the correspond- ing class of problems is s0 well understood as to be conceptually trivial}. Rather than trying to hide this, ‘unpleasant fact under a coat which might superficially look like an algorithm, this paper is organized as follows. The next section discusses a number of exam- ples where cither an exact solution, or a much more detailed analysis, is available, so that the approximateDimensional analysis estimates can be compared with what can be found in standard textbooks. The main point here is that while the analyses available in the literature are on occasion quite complex, the essence of the results can be ob- tained by very simple arguments— by a back of the envelope calculation. In the following section some points of a hopefully general nature are discussed this is as close to developing an algorithm as I was able to do. Following that, there is a section dedicated to the quasi-steady state approximation, a procedure of analysis which is particularly illuminating as to how much one can get from an approach based on dimensional analysis, scaling, and order of magnitude estimates. All the examples discussed only involve sealar quantities; for a discussion of dimensional anal- ysis as applied to vectors and tensors, see Dorgelo and Schouten (1946) 2. EXAMPLES OF SCALING (a) Flat plate with suction (based on Schlichting 1979) Consider a flat plate at zero incidence (located at y=0, x>0) which is invested by a fluid stream which, at x =— 0, has only one nonzero velocity component, vy = U. Suction is exerted through the plate, so that for y = 0 the y component of velocity is bound to be vy = — V, with V < U. Itis easy to see that the solution of the Euler equations is such that, at y=0, they result in vyo=U, vy =—V. and Po/dx = 0 (The subscript 0 is used to indicate values obtained from the solution of the Euler equations). Should there be no suction, the boundary layer thickness would be proportional to ./x, but in the presence of suction the thickness 6 may eventually approach some constant value 6... What is the value of 5? The value of U is unlikely to play any role: after all, should V be zero, there would be no value of 6. There is no characteristic length—the plate is infinite ly long in the x direction, So there are only three parameters: viscosity 2, density ©, and the suction velocity V. There is only one way to combine these to have the dimensions of a length, say 4/®V. Thus one concludes that 5. & WoVv wo This is tentative, so one proceeds by defining a di- mensionless distance from the plate, Y, as Y (So = OV in Q which is nothing else but a Reynolds number based on distance from the plate and suction velocity. Now if indeed 5 is constant along x, 50 is vg say dv,/0x = 0. But the continuity equation now implies that Avy/éy=0, and the boundary conditions (BCs) (vy = — Vat both ¥ = Oand ¥ = 00) can be satisfied with oy = — V everywhere. Let v, = f(). The equa- tion of motion in the x direction becomes: (— Vagjay) = ped? fidy? 8 4683 subject to f(0) the ¥ variable: affay +a? fa¥? =0 “) 0, feo) = 1. Transforming this to subject to the same BCs. There are no more para- meters, everything is of order unity, so the steady boundary layer thickness in eq. (1) must be right. In this case eq. (4) can be integrated to eq, (5) below, but the step is really not needed to reach the conclusion that eg. (1) holds true: Sat —exp(— ¥). 6 ‘The next question is: How far from x = 0 does one need to go before 6 is approximately 6,” First look at dimensional analysis. This time U is sure to play a role—so the required length Lis L = 5..f(U/V). At x= 0 the momentum flux for ¥ between 0 and 1 is ©U%5,,. At x = L, the momentum flux is significantly less than that, so the difference is of order ®U3,. This is balanced by the average shear stress at y = 0,1. Therefore, Lx OU?5,/t 6 ‘One only needs an estimate oft. At sufficiently large oc the shear stress must be of order 4U/3.., and if one uses this as an estimate one gets: Lx OUS%in = AU/OV? =5,(U/V). (1) This estimate has the form predicted by dimen- sional analysis, and itis indeed much larger than 3. in the pragmatically relevant case where U > V. I thisis not sufficiently convincing, one may reason as follows. Near x = 0, suction cannot have much of an effect, so 3 will grow as if there were no suction: 6? = ux/OU, Substituting 6, for # and L for x, one gets the same answer. Indeed, assuming that the appropriate length scale in the y direction is y/®V, and the appropriate one in the x direction is UV", define ¥ as before, and X as x/L. The velocity scales are obvious, so one defines u= »,/U, and w= vy/V. The equations of continuity and motion become: ujeX + Ow/oY =O @) uoulOX + wowjo¥ = 2u/2¥? } to within the usual boundary layer theory approxima- tions. The BCs are X=0, “1 (10) Y=0, =-1 ay Y=o w=-l (1) ‘There are no parameters in the differential equa- tions, the BC’s are all in terms of 1’s and zero's, so the assumed sealing must be right. Lessons 1o be learned: 1. In the estimation of orders of magnitude, the cruder the better, Factors of say 5 can safely be neglected; numerical coefficients should not clutter the algebra 2. Always try to define dimensionless variables in such a way that there are as few parameters in the4684 differential equations as possible, and boundary conditions present only 0's and 1°s. If this is obtain- able, one knows for sure that all scaling factors have been chosen appropriately. (b) Free convection [based on Acrivos (1960), Acrivos et al. (1960), Ng and Hartnett (1988), Minale and Astarita (1993)] In the classical theory of heat transfer in a flowing fluid the standard hypothesis is that the flow field is the same as would be observed in the absence of an imposed temperature difference. This is usually called the ‘isothermal’ approximation. Its consequence is that the energy balance equation becomes linear (if the effect of frictional heating is neglected), and hence that the heat transfer coefficient fis independent of the applied temperature difference 67. Indeed, one ‘obtains the general result that Nu = f(Re, Pr), where 4T does not appear on the right hand side. [Pr is of course the Prandtl number, a misnomer if there is one. ‘The Prandtl number is neatly introduced and dis- cussed in Rayleigh's 1915 paper]. The ‘isothermal’ assumption is obviously untenable in the free convec- tion case, where the applied temperature difference is the very reason why the fluid is set in motion. It follows that, in analyzing free convection, one should in principle solve the motion and energy equations simultaneously, an almost entirely hopeless task First the physics. There are three forces (per unit volume) acting: buoyancy forces G, viscous forces V, and inertia forces J. From the classical analysis of the Navier Stokes equations it is known that V = U/L? and I = U/L, where U is the appropriate scaling factor for velocity, and L the one for geometrical distances. The scaling factor U is not known in ad- vance. The order of magnitude of the buoyant forces is ‘@gB5T, where g is gravity and f is the coefficient of cubic expansion. The physics of the problem are as follows: the system will develop a flow field such that the combined effect of viscous and inertia forces bal- ances the buoyant forces. Should the velocity scale be U=/@L, the Reynolds number would be unity, and both the vis- cous and inertia forces would be 42/L*. At larger values of U inertia forces predominate over viscous ones, and vice versa. It follows that, if the buoyant forces are much larger than 12/®L?, they need to be balanced essentially by inertia forces. The condition is OBST > 2/OL*: Gr = O*L*G6T|e > 1 (13) The Grashof number Gr has emerged naturally from the analysis as the ratio of the imposed buoy- ancy force and the force arising when viscous and inertia forces are of the same order of magnitude. Since 13 implies that buoyancy forces are balanced by inertia forces, Og ST = @U*/L, say the characteristic velocity is U > (BST L). The Reynolds number is thus seen to be the square root of the Grashof num- ber: Gr>t, Rex J/Gr. (ad) G, Astarita In the converse situation where Gr < 1, buoyant forces are balanced by viscous forces, Ogh6T ~ HU/L?, and hence Ux@gBdT L7/y, and the Reynolds number is equal to the Grashof number: Gr <1, Rex Gr. as) Since the Grashof number determines the value of the Reynolds number, Nu = /(Gr, Pr). Now one may proceed to the scaling part of the analysis, which should yield the form of the function f(,) ‘The scaling of temperature does not present any problem: = (T — T9)/ST is of order unity. Let star- red quantities be dimensional; x* is a coordinate moving upwards from the lowest point of the body considered, and y* is orthogonal to the body surface; the corresponding velocity components are w* and 2% D is the diffusivity of momentum, D = y/®, and ais the diffusivity of heat, x = k/@c, so that Pr = D/a. The dimensional equations are: Momentum in the x* direction: wrout ox + vOUm/Oy* = BB dT 0p + Do? uk/ay*? 16) where @ is a geometrical parameter describing the local orientation of the body surface; @ is expected to be of order unity. Mass: aut /ox* + Ov): a7 Energy: we B0/2x* + v*aO/ay* = a08/6y*?. (18) The scale along x* is obviously the vertical dimen- sion of the body, L, say x = x*/L. One therefore needs to establish three scaling factors: U, V and Y, appropriate to the scaling of u* =Uu, o* = Ve, = Yy5so that u, ¢ and y are of order unity. The continuity equation immediately sets a first constraint between the scales, eq. (19) below, so that the dimensionless continuity equation becomes eq. (20) below: UY=VL (a9) Gu/ax + dvjay = 0. (20) Treg, (19) is satisfied, the energy equation is (U/L(ua9/ex + vé0jey) = (x/¥7)0*O/2y?, (21) Equation (21) could conceivably be satisfied with the two terms on the LHS of equal order of magnitude and opposite sign, and then the RHS would have a much smaller order of magnitude. This would, how- ever, imply that conduction is negligible, and this does not make sense, Hence the RHS and at least one of the terms on the LHS must have the same order of magni- tude, which gives one more constraint: = aLjU 2)Dimensional analysis which reduces the dimensionless energy equation to the following parameter free form: Uab/éx + ve0/oy = eAjey2 (23) If egs (19) and (22) are satisfied the balance of momentum reduces to (U?/L(udujox + véwjoy) = gh6T#O + (U2Pr/L\Ou/ay? 24) Viscous forces must play a role near the solid sur- face, and one therefore multiplies the whole equation by L/U?Pr so as to make unity the coefficient of the viscous forces (LgpsT/U2Pr}¢9 5) (1/Prtueujex + vou/ey) + Puléy? Buoyant forces have to be significant, and in fact near the solid surface they have to be of the same order of magnitude as the viscous forces; it follows that the coefficient of 0 must be of order unity, which sets the value of the scaling factor U: U? = Lops Pr 26) Al scaling factors have been determined, say: V2 = Lop aT Pr? Gr-¥ 27) Y= LPr Gr 28) No parameters appear in any of the equations or the boundary conditions, except for the 1/Pr coetfic- ient of the inertia forces in the equation of motion, First, consider gases: the Prandt! number is of order unity, and so no term drops out of the equations. What that means, physically, is that the thickness of the ‘thermal boundary layer’, ¥ (the distance from the solid surface over which @ is appreciably different from zero), is also the distance over which inertia terms become significant-—it is equal to the thickness of the momentum boundary layer. Notice that in this case 28 yields ¥/L = Gr-"*, and since Gr = Re? one recovers the classical result that ¥/L = 1/,J/Re. In the case of ordinary liquids Pr > 1 and the inertia terms drop out—the thermal boundary layer is much thin- ner than the momentum boundary layer, and its thickness ¥ is small enough that inertia forces are negligible when 6 is significantly different from unity. For both cases considered above, Vu would be calculated as — Léd/éyf In the case where Pr = 1 (gases), this yields Nu = KGr'f(Pr), The value of J( Jis unity when Pr = 1 and it can’t be much differ- cnt from unity in any gas; furthermore, the constant K is expected to be of order unity. In the case where Pr> 1, the inertia terms drop out, and one obtains Nu = K(Gr PA), with again K expected to be of order unity. Notice that the exponent 1/4 comes out simply from the scaling of ¥. Now suppose one had started on this problem by straightforward dimensional analysis, say by listing the parameters on which h may depend h=h® 1, g, B 8T, Lyk, 0. 029) 4685 Identification of the appropriate list of parameters is by no means a trivial exercise: in hindsight one knows that thelist in eg. 29) is correct, but if this were a problem one had never heard about, would one hhave come to the same list? Be that as it may, there are 9 parameters in eq. (29), and 4 fundamental units mass, length, time, and temperature—so that one would wrongly conclude there ate $ dimensionless groups. Can one add heat to the list of fundamental dimensions? Heat is energy, and hence it has dimen- sions ML?/T?, However, frictional heating has been neglected, so that there is no conversion of mechanical to thermal energy, and thus indeed one can use heat as a fourth fundamental unit. Had the Griffith (or Brink- man) number «U7/K6T not been negligibly small we could not have used calories as an independent dimension. Now there are 9 parameters and 5 dimensions, but the right answer is 3, not 4 dimensionless groups. One should now argue that /f enters the problem only insofar as it determines the 5 corresponding to the imposed 57, and 6® in turn is relevant only because there is gravity-hence # and g can only enter through their product, and one is finally home. Well, would the reader ever have thought of that without knowing the answer beforehand? Lessons to be learned: 1. The fundamental dimen- sions to be used are not a sacred cow. They can be chosen in different ways depending on what the prob- fem under consideration is, and which approxima- tions ones in principle willing to make (in the case at hhand, neglect of frictional heating) 2, Its best to find dimensionless groups by looking at the order of magnitude of different terms in the balance equations which are relevant to the problem, and then taking ratios of the latter. The resulting dimensionless groups will then be identified with ra- tios of physically significant quantities. This step can often be performed without actually writing down the appropriate differential equations, (That heat can be used as an independent dimen- sion had been already recognized by Rayleigh (1915). Riabouchinski brought up, in this regard, what seemed a paradox—heat is energy and it has no inde- pendent dimensions like mass, length and time. Rayleigh, in his response, stated that there would indeed be a paradox if “further knowledge of the nature of heat afforded by molecular theory put us in a worse position than before in dealing with a particu- lar problem”. The paradox is neatly solved in Bridg- ‘man's (1922) and Sedov’s (1959) books.] (6) Transport between two fluid phases Consider two fluid phases in contact with each other, the phase of interest being much more viscous than the other one. Let» be the velocity component in the interface plane, and let x be the normal to the interface pointing into the phase of interest. Since the tangential stress at the interface must be equal in4686 value and opposite in sign in the two phases, the values of 6o/2x at the interface in the two phases are in inverse proportion to their viscosities, and hence @o/0x in the more viscous phase can be set approxim- ately to zero. It follows that, near the interface, the more viscous phase moves as a rigid body: heat trans- fer orthogonal to the interface in the more viscous phase takes place essentially by unsteady conduction (except in very special cases, there is no velocity com- ponent in the x direction). The differential equation for transport is aT fat = ad? T/ax? G0) subject to T (x, 0) = To, T(0, #) = Ty. The solution of this problem is very well known, but should one not know it, one would proceed to dimensional analysis. Temperature is normalized by defining (T ~T¢)(Ty — To) so that the equation becomes: e0/at = ad 2; A(x, 0) = 0, (0,t)=0. (31) Since there is no characteristic length, there must be a similarity variable. The only one possible is $= x/y(a), and indeed substituting @ = 6(6) into eq. (31) it does reduces it to an ordinary differential equation for 6(). The BC at x = 0 transforms to one at @ = 0, which however corresponds also to t = 20, which makes sense, From the definitions one sees that the instantaneous Nusselt number at some time t is = 8 (W)L// (at), ic. to within a numerical constant it is L//(at), where L is the characteristic length used in the definition of the Nusselt number. One is, however, interested in the average value of Nu over some inter- val of time ¢*, which measures how long an clement of the viscous phase stays at the interface. Since the average of the inverse square root is again the inverse square root, one concludes that Nu = L/./(at*. The viscosity of the phase of interest cannot play any role, since it has essentially been assumed to be infinitely large. Thus only the product of the Reynolds and Prandtl numbers can appear. Nu is proportional to the square root of the specific heat c, since a=kc, and hence Nux J/(RePr). All that is needed is the proportionality factor, and that can be determined by one single experiment. It is important to notice that the result has been obtained without ever trying to integrate any differential equation, and that the similarity variable has been deduced by straightforward application of dimensional analysis. From that point on, it has simply been a matter of applying the definitions. Now focus attention on the other phase, where av/dx-is most certainly not zero; let us say it has some value rat the interface. For the viscous phase one was willing to conclude that it moved as a rigid body because Ge/dx was 2ero—one was willing to use for va Taylor series expansion truncated at the second term. By the same token, one should now be willing to approximate the velocity tangential to the interface with rx, with x pointing into the less viscous phase G. Astarita (velocity at the interface has been normalized to zer0). Let y be the direction of the velocity v. The differential equation is x0 0/Ay = (xfs) O70 /0x2 (2) where conduction in the y direction has been neglect- ed since convection is presumably predominant. The BCs are A(x, 0) = 0, 4(0, y) = 1, and again the lack of ‘an external length scale leads us to think that there ‘must be a similarity variable. There are again two independent variables (x and y) and one parameter (a/2), but this time there are infinitely many different ways of constructing a dimensionless combination of those. Jn the previous case, the two independent vari- ables (x and 1) had different dimensions; in this case x and y have the same dimensions, that is the source of the problem. If only x and y had different dimen- sions! But what is wrong with thinking as if they did have different dimensions? Let us say that y has di- mensions L, and x has dimensions L’ It follows that a/t has dimensions L/L. Now there is only one possible dimensionless combination, 6 = x(ay/s)"*, and this is thus the similarity variable—@ = 0(¢). Sub- stitution into eq. (32) indeed reduces it to an ordinary differential equation. The local value of Nu corresponding to some value of y is obtained from the definitions as L(c/ay*)!'®, and the same is true of the average Nu over some finite distance y* which is L(c/2y*)" to within a nu- merical factor. This means that Nu is proportional to the cubic root of Pr—as we know to be the case. One cannot deduce how Nu depends on Re, though. In fact, x equals the characteristic velocity divided by the thickness of the momentum boundary layer 4, and hence one gets a D/é ratio in the final expression, and one does not know how this ratio may depend on Re. Lessons to be learned: 1. An even more extreme example of the freedom of choice of the fundamental dimensions has been presented: two different dimen- sions for length. Of course, the freedom of choice has to be used with quite a bit of care. 2. The existence of a similarity variable, or lack thereof, can be found out by judicious dimensional analysis. Of course a similarity solution may exist even when there is no similarity variable. [Using different length scales indifferent directions was considered as early as 1892 by Williams; many applications of the method are discussed by Huntley (1953). The method should be used with great care; ee the discussion by Panton (1984),] (d) The tibrating sphere [based on Batchelor (1967) Consider a sphere of diameter D which is vibrating with frequency © and amplitude 6, with 6 < D. The interest is in such questions as what the force F with which the sphere acts on the fluid may be, and how much mechanical power is needed to keep the sphere vibrating.Dimensional analysis Straightforward dimensional analysis starts from the following list of parameters: (14, ©, 9, D, 5,1) 63) since of course the force will vibrate itself and there- fore it will depend on time t. The force F is going to appear in a drag coefficient, which, with DQ the velo- city scale, is F/D*@Q?, Counting parameters and di- mensions one has 4 dimensionless groups, and two of those are easily identified, say Or and 5/D—the frst one a dimensionless time along the cycle of vibration, the second one a geometrical form factor. The last group is a Reynolds number D?Q0/x, and thus F/D*Q? = f(Qt, 6/D, D°QYy), G4) The force F is going to vibrate itself, and therefore one is really interested in two separate things: the amplitude of the force vibration, F*, and its phase angle with the sphere vibration, ¢, with @ = 0 when the force F is always in phase with the sphere velocity (in which case the instantaneous rate of work done by the sphere on the fluid will be always positive). All the work done by the sphere is dissipative when = 0, Conversely, if § = #/2, the total work done by the sphere over a cycle is zero (no dissipation). Both F* and ¢ do not depend on time, and, since one is seeking a solution for 4/D ~0, one looks for something like: FA /D'O0? = f(D?Q0/y) 63) with @ given by a similar equation The physics of the problem suggests that at very small frequencies, Stokes’ law is going to apply, and therefore ¢ is going to be zero; while at high frequen- cies, itis like a swing—one needs very little power to keep it going, just a tiny little push at each swing, so that ¢ will approach x/2. However, how small a fre- quency is small enough? The problem is one of scal- ing. There are 4 terms to deal with: the Eulerian inertia 0e/01, the convected inertia v-grad v, the pressure gradient, and the viscous force term. The pressure gradient term cannot be scaled—it will ad- just itself to whatever the other terms demand. The force will be of the same order of magnitude as the form drag, hence | grad p| = F/D°, Time is obviously to be scaled as 1, but the length scale may, at this stage, be either 5 or D, so call it L, remembering that Lis at most D and at least 6. The following orders of magnitude are estimated: Eulerian inertia: @50? Convected inertia: ®5207/L Viscous forces: 430Q/L? Pressure gradient: F*/D* Now suppose that © < D2, In this case, no matter what L may be, the viscous force is the pre- dominant one, and Stokes law will apply, FP = uO 1. 160 < w/@D?, then P* = 4005, and = 0. The force is almost entitely dissipative. This estimate holds up to =O) = w@D?, where F* = p26/OD = FE 4687 Now suppose 1/5? <0. No matter what L may be, the viscous forces are negligible, and the convected inertia is either of the same order of the Eulerian one or negligible with respect to it. Hence the Navier-Stokes equations reduce approximately to on/ot = — grad p, and since F* = D3|grad p| one gets an estimate of F* x O402D". Notice that this time grad p has the opposite sign—for half a eycle the fluid pushes the sphere, and for the other half the sphere pushes the fluid, This is not surprising, because if viscous forces are negligible there is very little dissi- pation, and hence F is almost entirely conservative: 2. HOD w/O5?, then F* x @SQD, and $ ~ x/2. “The force is almost exactly conservative. This estimate holds down to Q=Q, = y/5?, where F* = D078? = FE Now FE/F ~ (Dj3)*, while 94/0, ~ (D/6) It fol- lows that there is a very smooth transition between the slope | line which holds for © < , and the slope 2 line holding at © >. The two lines cross each other at © = 0, which suggests that the true critical frequency is 2, and hence that the appropriate length seale for the viscous forces is D. Lessons to be learned: 1. Dimensional analysis, scaling, and order of magnitude analysis are inter- twined with each other. 2 Some thermodynamic thinking about dissipa- tion is very often useful. {€) Too much of a good thing In examples (a}-{d), one has tried to reduce the parameters to their minimum possible number, to have the differential equations and boundary condi- tions all in terms of zeros and 1's, and so on. Well, sometimes there is too much of a good thing, and in this subsection I give an example of that. Consider a steady point source of some obnoxious chemical, ofstrength $ (moles/s). There is no wind, itis all steady diffusion in an infinite body of fluid, and the steady concentration distribution is c= Si4xDr. 36) At the other extreme of a very strong steady wind, with velocity w in the z direction, it is better to use cylindrical coordinates, with @ = 0 the downwind di- rection, and y the distance from the downwind axis. Since the wind is strong one neglects diffusion in the downwind direction z. The solution (valid only at + >-0) is [Williamson (1973) the equation given there is in error by a factor of 2] (S/4nDz) exp( — uy*/4Dz) G7) ‘Now consider the case of a wind which is not quite so strong, One would expect to find a dimensionless group, say Q, containing the wind speed w in the numerator, When Q is large, one expects eq. (37) to hold, when it is small, one expects eq. (36) to hold.4688 Going back again to spherical coordinates, there is only one length scale available, D/u, and so R = ru/D. Concentration is best scaled as t = 4xD?c/Su, so that eq, (36) reduces to x = 1/R. The equation is Laple = cos @0r/2R —sin 662/68 (38) and there is no dimensionless group Q! This is 100 much of a good thing: one wanted a dimensionless group to come out of the equations, and it has not Furthermore, what are the boundary conditions? + goes to zero as R goes to a0; @r/60 is zero at 8 = 0 and at 6 = 7/2; and the fourth one is that, given any closed surface encompassing the source, whatever ‘comes out oft per unit time must be S, which isa very nasty BC indeed, ‘Now engineering sense suggests that one only needs to worry about the EPA recommendations—there will be some concentration value co which cannot be exceeded. co establishes a concentration scale, and this will produce a dimensionless group of the @ type. It might be objected that the solution of the problem cannot depend on what value of co the EPA decides to choose, but it is always dangerous to regard good engineering sense as silly, so one defines T as c/co, one switches again to cylindrical coordinates to obtain Lapl T= of @9) where Q = Su/4xD*co. @ may not have physical meaning, but eq, (39) does tell us that, when Q ap- proaches zero, the equation is Lapl+ = 0, and that has eq. (36) as the solution. Ifo is large ¢q, (36) holds—or, in other words, it must hold near the origin: the nasty boundary condition is now substituted by—in spheri- cal coordinates: R0, r+ /R (49) [This problem admits a surprisingly simple exact solution (Seinfeld, 1985): 1/R)expLR(cos 6-1)/2) ay Exactly downwind (cos@ = 1), the wind has no ef- fect whatsoever on the concentration level!] {3 GENERALIZATIONS (a) Peculiarity of the power law Suppose one has decided to write down a constitut- ive equation for some variable y, and one has decided that its value is uniquely determined by the value of some other variable x, sa: y=s00. @2) The case of interest is the one where both y and xx are dimensional, and their dimensions are different from each other. For instance, for a chemical reaction ¥y Would be the rate of reaction, and x would be the ‘concentration of one particular species; for a purely viscous non-Newtonian fluid y would be the stress tensor and x the rate of deformation tensor. The following question is of importance: ifeq, (42)is one of G. Astarita the equations which enter any problem we may wish to consider, how many dimensional parameters does it contribute to the starting list of classical dimen- sional analysis? What one means by saying that some physical quantity y is dimensional is that the number which ‘measures its value depends on the set of units one has chosen. One can extend this concept: an operator f( ) is dimensional if its mathematical form depends on the units chosen to measure both its argument x and its value y. Obviously, ifat least one of either x or y is dimensional, f() is dimensional. One can always transform eq, (42) in such a way that a dimensionless operator F( ) appears init by introducing two dimen- sional parameters Y and X, having the same dimen- sions of y and x, respectively, and then write: IY = F(X), (43) One has of course unlimited latitude of choice for the parameters X and Y, but the result to be stressed at this point is the following one: in order to have a dimensionless operator, one needs two dimensional parameters, Are there exceptions? The answer is af- firmative, as shown by the example of a power law: ya Kx (4) Equation (44) produces, when itis cast in the form of eq. (43): ys (KY/X9)x7 a9) which contains only one dimensional parameter, the group KY/X", Notice that the linear case (n = 1) is a special case of the power law form. ‘Now it takes a bit of algebra—not in any significant way different from the one leading to the classical formulation of Buckingham’s theorem—to convince oneself that the power law form of constitutive equa- tion is the only one for which a single dimensional parameter is all one needs, The result is related to the fact that the only algebraic operations which are legit- imate with dimensions are multiplication and raising toa power. Be that as it may, the fact remains that the power law is dimensionally peculiar. In order to con- vince oneself of how peculiar it is, consider the deceiv- ingly simple case where eq. (42) can be expressed as (46) There are two dimensional parameters, aand b, and no amount of juggling of algebra can collapse those two parameters into one. ‘One might here reason that the linear case has some special status—in the sense of a Taylor series expan- sion truncated at the first order in x, after appropriate normalization to guarantee that /(0) = 0—but the power law has not, and that therefore one really has only two cases: either eq, (42) is linear, and there is only one parameter, or it is not, and then there are two. But in the case of chemical kinetics, are not norder ones reasonable? Consider the isothermal ef- fectiveness factor in a porous catalyst having a flat yratbx.Dimensional analysis plate geometry with half thickness L. For norder kinetics, the differential equation is: Déd?cjdx? = be 4) where, for reasons which will soon become clear, 1 use the somewhat unusual symbol b for the 'n-order kin- tic constant’. Equation (47) is subject to c= c, at ), defdx = O at x = L, and hence it is natural to scale distances and concentrations as y= x/L, 1 = cj This reduces the differential equation to er dy? ee (48) subject to r=1 at y=0, de/dy =0 at y= 1. The Thiele modulus ¢, is defined by $2 = bL?/De}~", ic. itis recognized as the usual Thiele modulus for Hinear kinetics, with the pseudo-first-order kinetic constant being evaluated at the surface conditions, keq = bet", Now the effectiveness factor E is given by a) [aa ), {eae = ($2), 9. 4) Since the solution of eq. (48) is obviously of the form +r =1(y, 6), €4. (49) implies that the effectiveness fac- tor is a unique function of the Thiele modulus cal- culated at the surface conditions, E = E(g.). ‘Now suppose that the actual kinetics are of the Langmuir isotherm type, say oma (50) dx? (1 + Key This time its natural to define «= Ke. The equiva let kinetic constant tthe surice conditions snow keq = kill + Ke), and the second part of eq. (49) still holds. The differential equation is ar ay (+ el +2 (si) ‘The solution of eq, (51) is of the type t = 1(0, bs #0. and so one cannot conclude that E is a unique func- tion of the Thiele modulus caluculated at the surface conditions, say in this case E = E(f,, 1). The result is, not surprising: there is one more dimensional para- meter, and hence one more dimensionless group than in the former case. The discussion above shows that the dimensional peculiarity of the power law form is not a trivial one, since it leads to conclusions which would not hold for any other nonlinear form. There are other examples in the literature. For a power law non-Newtonian fluid, ‘one can define a modified Reynolds number (analog- ous to @, in the example above), and conclude, for instance, that the ratio of the entry length in a pipe to its diameter is & unique function of this modified Reynolds number; the same conclusion is not reached for any other form of the constitutive equation for the stress. Also, for any fluid for which the stress is a unique, though not necessarily linear, function of the rate of deformation, a variational principle for ereep- 4689 ing flow of the type of the Helmholtz-Korteweg one can be constructed (Fredrickson, 1964; Astarita, 197), In the special case of power law fluids (includ- ing Newtonian ones) the quantity to be minimized is proportional to the rate of entropy production, so that the variational principle can be (wrongly) inter- preted as a principle of minimum dissipation; for any other nonlinear form of the constitutive equation, the quantity to be minimized is not the total rate of energy dissipation. This peculiarity of the power law cquation is related to its dimensional peculiarity (Astarita, 1983), Now one comes to the question of why power law equations have this habit of popping up here, there, and everywhere. One looks at where they come from, and one reaches a somewhat unexpected conclusion: data are plotted on logarithmic paper, as engineers are wont to do for a variety of reasons, and straight lines are fit to them—again, engineers like straight lines. OF course the equation of a straight line in logarithmic paper is a power law, and that is where the origin lies, in most of the cases. Now those same data could have been plotted on some other type of paper (semilogarithmic, hyperbolic, whatever) and perhaps a straight line would have fit them equally well. But in this case one would have obtained a non- linear constitutive equation of type other than power law to represent those very same data. One thus concludes that one should be very careful about con- clusions drawn from dimensional analysis when power law constitutive equations are used (unless n= 1}: such conclusions are peculiar, and different conclusions would be reached with any other type of datacfitting constitutive equation (0) Time and length scales In many problems in chemical engineering, a very powerful tool is the comparison of time scales of different phenomena. For instance, all approxima~ tions of the ‘quasi-steady-state’ type (to be discussed in the next section) are obviously based on the consid eration that the time scale of unsteadinesss is in some sense very large as compared to the time scale of the phenomenon one is considering. In this subsection, 1 discuss some conceptual issues associated with time scales, and I also show that often a comparison of time scales is equivalent to a comparison of length scales. First consider classical transport phenomena, One always has a diffusivity D, and the appropriate time scale is L?/D, where L is some external length scale. Now consider any problem where a transport phe- nomenon takes place simultaneously with some other phenomenon which is characterized by an intrinsic time scale 9, such as a first order reaction where the kinetic constant k is the inverse of the intrinsic time scale of the reaction, # = 1/k. The ratio of the two time scales is L*/D8, or, in the case of a first order reaction, it is the square of the Thiele modulus ¢. [It is of interest that what is usually called the Thiele modulus4690 was in fact introduced and analyzed as early as 1909 by Juttner.] Now of course ¢ is usually regarded as, a dimensionless measure of the catalysts linear di- mension L, and in this sense one concludes that the intrinsic length scale one is comparing L with is (DO. Given any phenomenon where a diffusivity D and an intrinsic time scale 6 play simultaneously a role, the corresponding intrinsic length scale is Vi00. Intrinsic time scales may arise from two different types of constitutive equations. The first type is that of Kinetic phenomena, ic. phenomena the rate of which is governed by their own intrinsic kinetics. Chemical reactions are of course the example which immediate- ly comes to mind, but they do not exhaust the field Whenever the thermodynamic state includes some internal state variable, the rate of change of the latter is given by a kinetic equation, and hence an intrinsic time scale can—with some attention—be identified. The second type of intrinsic time scale is the one arising when relaxation phenomena are considered; any relaxation phenomenon is or is not relevant to the analysis of any given problem according to whether the intrinsic time scale, i. the relaxation time, is or is, not comparable with other time scales of the process. Intrinsic time scales arise from constitutive equa- tions, but as seen in the previous section some care must be exercised in extracting the time scale when the constitutive equation is not linear: one cannot simply carry over results from the linear case to the nonlinear one by using time scales evaluated at some particular point. There is, in fact, a degree of arbitrari- ness: one may choose any definition one may wish for the intrinsic time scale, but one should then be careful about order of magnitude analysis. Consider a rather general case of kinetic equation, say one where the rate of change of some internal state variable w depends on the value of w itself, say Sow). (52) Since w is an internal state variable, it will have some equilibrium value w*, such that at w= w* aw/ét = 0. It is therefore useful to rewrite eq. (52) in the following form: ewyee a) ow (33) so that the function f() is now constrained by F(0) = 0, sgnf= sen(w* —w). The operator f¢ } is still dimensional, even if w has been normalized to be dimensionless (as can always be easily done). One can easily rewrite eg. (53) in a form which contains a di- mensionless operator F( ) subject to the same con- straints; Owjée = Fiw* — wy (54) where @ is an intrinsic time scale. However, the con- straints F(Q)=0, sgnF =sgn(w*—w) do not uniquely determine the value of @, and there is there- fore still freedom of choice. G. Astarita A possibility which is always present is as follows. If one is considering this kinetic problem, it must be because w is either initially, or at some external boundary, different from w*—otherwise there would be no kinetics to consider. In other words, one always has some physically significant driving force w= wh —w which is different from zero. In the example discussed in the previous section, c, played the role of the driving force. One can now require F (dw) to be unity, which of course uniquely deter- mines the value of 0. Furthermore, one now has a physical interpretation of @: it is the intrinsic time scale corresponding to the imposed driving force dw: Another possibility is as follows. Since F(0) = 0, sgn F = sgn(w* —w), one can (perhaps) expand F( ) in a Taylor series near equilibrium and write FOv* — w) = (w* — w)F (0) + O[Ov* — WP]. (55) Requiring that F'(0) = 1, the value of @ is uniquely determined. In this case. 0 is recognized as the intrin- sic time scale near equilibrium—in the Langmuir iso- therm case it would be the constant 1/k (since equilibrium corresponds to c= 0). Quite obviously the two choices are both legitimate, but one should be very careful about the implications of dimensional analysis based on either one of them. While the first choice, F(6w) = 1, is always pos- sible—after all, f(6w) must have some finite vvalue—the second choice is possible only ifthe kinetic constitutive equation can be expanded in a truncated Taylor series at the equilibrium point, ic. if F'(O) is finite. Now this may seem a trivial requirement, but it deserves some discussion. Consider the case of an irreversible dimerization reaction, for which Ge/ét = — he. In this case the equilibrium point is c= 0, and the kinetic function ccannot be expressed as a Taylor series at the equilib- rium point truncated at the second order, since F (0) =0 and one cannot require it to be unity. This, does seem a rather convincing counterexample. How- ever, no reaction is truly irreversible, and therefore the correct kinetic equation for a dimerization reaction is bic? — ex/K) (56) eejee where > is the dimer concentration and K is the concentration based equilibrium constant. Given 4 dimer concentration c, one may define an equilib- rium concentration of monomer as c* = /(c2/K}:* is the monomer concentration which would equili brate the local instantaneous dimer concentration. ‘The RHS of eq. (56) would thus be written as bc? ~c¥2), Now suppose that c= c*(I +0}, with <1, ie. the actual concentration is close to the equilibrium one. One obtains: Gclt = 2bec®? + O(e?) hele —c%) (57) i.e, the rate of reaction is linear in the displacement from equilibrium. It follows that F(0) is nonzero, and 4 (near equilibrium) is identified with 1 /(2bc%)}—which of course diverges only in the case of a truly irrevers- ible reaction,Dimensional analysis The other point to be made is that the assumption that F'(0) is finite is equivalent to the assumption that F() is invertible at equilibrium—an assumption which is needed in order to get the usual result that the affinity iszero at equilibrium (Astarita and Ocone, 1988; Astarita, 1989). Since the latter result is inva- riably regarded as the correct one in the theory of equilibrium, the assumption that F'(0)is finite is noth- ing else but the assumption one makes, albeit in most cases implicitly, o obtain the classical theory of equi- librium, Now go back to the case where one wishes to identify 0 from the requirement that F(5w) = 1, as, done in the previous section. While this sets the value of @, the value one gets does depend on the imposed driving force dv. This may seem to be a consequence of the nonlinearity of the constitutive equation one is considering, since for a first order reaction 0 = I/k whatever the driving force may be. However, this conclusion is not true in general, and the following example shows that one may have a perfectly linear problem where the ratio of two relevant time scales depends on the imposed driving force. The example is that of the classical Stefan problem of the freezing of water (Stefan, 1891). The time scale ofa straightforward transport problem is L?/D. If, on a sample with halfthickness L initially at some tem- perature Ty, the faces are brought to temperature To + 67 at time zero and kept there afterwards, the amount of heat to be removed per unit exposed sur- face is cL 6T, and under a driving force 57 it will take a time of order L2/D to do that: the amount to be removed and the rate of removal are both propor- tional to the imposed driving force, and hence the time needed for removal is independent of ST. How- ever, consider the case where one has to remove an amount of heat governed by the latent heat I, say an amount per unit surface which is TL—this is not proportional to dT. The driving force is the same as before, and thus also the rate of removal is of the same order of magnitude, so that now the time required is, (L4DYE/e6T), So the Stefan number St = T/e5T is seen to be the ratio of the time needed to remove the latent heat to that needed to remove the sensible heat under the same driving force—and that ratio does depend on the imposed driving force ST in spite of the fact that all equations governing the phenomenon are linear. Time and length scales are not always inter-change- able. A good example can be extracted from an el- egant analysis by Amon and Denson (1984) of the growth of gas bubbles in a polymeric foam. The gas bubbles contain only the volatile component; in the bulk of the polymeric phase, the volatile concentra- tion is such that the equilibrium vapor pressure p* is, larger than the local pressure p; this provides a driving force for mass transfer to the bubbles and hence for their growth. The problem is to estimate the growth rate of the bubbles. Two asymptotic cases can be envisaged. First, there is no hydrodynamic resistance to the growth of the bubbles. In this case (neglecting 4691 surface tension effects) the pressure inside the bubbles is equal to the external one p, and p* —p is the riving force for mass transfer which takes place by diffusion. The length scale is obviously the inverse of the cubic root of the number of bubbles per unit volume, n, and hence the time scale is the time scale of diffusion, 6 = n-?9/D. However, there are hydrodynamic resistances to the bubbles’ growth: a hydrodynamic hoop stress develops due to the extensional flow field in the thin shell surrounding every bubble, and hence the pres- sure inside the bubble must be larger than p to bal- ance the hoop stress. Now the pressure inside the bubbles cannot exceed p* (it will in fact be p* if there is no diffusional resistance to mass transfer), and therefore p* —p is the maximum difference which may balance the hoop stress. One does not need to solve the (surprisingly simple) problem of the flow in the outside shell: the viscosity must play a role, and the ratio (p* ~ p)is has units of time, Hence in this cease the time scale is 8, = (p* — piu The density of the fluid does not play any role because polymers have viscosities so high that the Reynolds number is practically zero in every conceiv- able situation. The length scale cannot contribute to the determination of 6, as classical dimensional anal- ysis shows directly. However, it is physically more illuminating to consider that the extensional flow shell is very thin (the continuity equation tells us that the radial velocity decays as the inverse of the square of the distance from the center of the bubble), and that therefore the viscosity-dominated growth problem is, one without an external length scale. In other words, in this ease we may consider the problem of an iso- lated growing bubble for determining 6, The ratio of the two time scales has been called the Denson number Ds = 6/0, =n" ?°(p* = p/Du (Maffettone and Astarita, 1993). Dsis the ratio of a D- controlled time scale to a y-controlled one, and it may appear strange that D and 1 both appear in the denominator. However, the larger is D, the less likely itis that the phenomenon is D-controlled; in contrast with this, the larger is a, the more likely it is y- controlled. The interest of this example is that it seems to be one where consideration of time scales is forced conus. Could this same problem be expressed in terms of length scales? A length scale exists, and if Ds is to be interpreted as (the square ofa length scale ratio, one would need to regard [Dp(p* — p)}'? as the other length scale. This in no way is an obvious length scale, and in fact the very fact that both D and jt appear in it seems to imply that in order to find it one would need to solve the complete problem—which is defeating the purpose of finding scales. So one would conclude that this is indeed a problem where time scales are ‘natural’. (0) Nusselt and Biot type numbers ‘What is the physical meaning of dimensionless groups of the type of the Nusselt number? Let us first see what one means by the classical Nusselt number4692 arising in the theory of heat transfer, Nu = hL/K, where h is the heat transfer coefficient, L is some characteristic length, and k is the thermal conductiv- ity. The heat transfer coefficient is in turn defined from = h8T, where q is the heat flux orthogonal to the bounding walls, and 67 is the driving force. The obvious analog in the theory of mass transfer is the Sherwood number kL/D. where k is the mass transfer coelficient, and D is diffusivity; kis again defined from N=kéc, with N the mass flux and éc the driving force. Now consider the case of momentum transfer, where things are not quite so trivial. For the sake of simplicity, consider first turbulent flow in a tube. The equivalent of q and N would be the momentum flux at the wall, say the wall shear stress t. The imposed driving force would be the average velocity V. Thus the equivalent of and k would be 7/¥, and since t= LP/A, where L is the tube diameter and P is the pressure drop per unit length, one would have a trans- fer coefficient given by LP/4V. Now P isexpressible in terms of the friction factor f as P= fDV?/L, and therefore the transfer coefficient is—to within numer- ical factors—fOV. The equivalent of the Nusselt num- ber would thus be fDVL/p = fRe. The result can be extended to the case of flow around a submerged object, i cis now interpreted as the average flux of the momentum component in the direction of relative motion, and one would have the drag force divided by the cube of the linear dimension as the equivalent of P. The Nusselt number equivalent would again come out to be the product of the drag coefficient and the Reynolds number. ‘Now go back to the simple heat transfer problem, Let x be @ coordinate orthogonal to the wall, with x = (identifying the bounding wall itself, Should one know the space distribution of temperature T, one would calculate q as k(@T/éx)o, and hence Nu = (6T/éx)oL/6T. Variations of T are of order 57, so that one estimates (€T/éx)g to be of order ST/l, where | is the distance over which the significant variation of T takes place (I is often called, somewhat improperly, the thermal boundary layer thickness). ‘One therefore identities Nu with Lil. Had one chosen the scale properly, one would have chosen {instead of L,and Nu would have been of order unity by defini tion. One is unlikely to choose the scaling factor for length smaller than the actual one—for instance, in a tube how could the proper scaling factor for dis- tance be significantly more than the diameter, which one always chooses as the scaling factor?—and thus the Nusselt number type of dimensionless group is expected to be either of order unity or much larger than that. [It will be much larger than unity if one has overestimated the sealing factor for length by taking, say, the tube diameter while in actual fact the whole variation of T takes place in a thin layer of thickness Inear the tube wall] The situation is entirely different in the case of the Biot number. Bi is defined as hL/k, where L is the characteristic dimension of the phase where heat con- duction is being considered, k is the conductivity of G. Astarita that same phase, but h is the heat transfer coefficient in another phase. It follows that the Biot number may have any value whatsoever, ranging from zero (the exposed surface is essentially adiabatic) to infinity (here is no external heat transfer resistance). It is instructive to consider the deceivingly simple problem of a slab of half thickness L, initially at temperature To. which at time zero is brought in contact with a fluid phase at temperature T); his the heat transfer coefficient in the fluid phase. In dimensional form, the differential equation and boundary conditions are GOT let = KET IE (58) Thx, 0) = To (59) 0, -kOT/Ox=h(T,-T) (60) L, aT jx =0. (61) Define # as (7; — T)(F, — To), s0 that the system becomes homogeneous. The appropriate length scale is L,y = x/L. The scaling of time is mote delicate, One possible scale is @el.?/k, which is the heat diffusion time inside the slab; this would lead to defining += ki/®cL?, which is known to eliminate all para- eters from the differential equation. However, Pel. jh also has dimensions of time, and it could be used as the time scale, The question which arises is of course which one is the realistic time scale; in other words, how long does it take for the slab to approach the final condition T = T,? The answer may be @cL*/k (the internal diffusion time), or @eL/h (the external resistance time), and at this stage the best one can do is to suspect that it will be the one which happens to be larger. Now notice that the ratio of the internal to the external time scale is Bi. Choosing (somewhat arbitrarily) the internal time seale, and hence the di- mensionless time ¢, the dimensionless formulation be- ‘comes: cofer = 0/69 (62) t=0, 6=1 (63) 0, c0/0y = Bid (64) yal, eéy=0. (65) First consider the case of a negligible external res- istance to heat transfer, say 1/Bi—+0. Equation (64) becomes (0, r) = 0, and thus the problem reduces to the classical one for which the time scale is indeed £1. However, in the case of a very high external resistance Bi—+0 one expects temperature within the slab to be uniform at all times, T = T(t) in which case one simply needs to integrate ®eLAT dt = h(T, — T) subject to eq. (59). Now the alternate time scale @cL,/h emerges quite naturally: 6 = exp( — Bir), So now one concludes that the time scale for final equilibration is + IiBi, which is very significantly larger than unity Could one have obtained the same result directly from eqs (62)-65)? After all, 0 = exp( — Biz) must be the zero order approximation to the solution of eqs 62-65,Dimensional analysis If one integrates eq, (62) over 0 1, 6 = exp( — Bit) + Bi(y — y2/2) is the first order solution, [The complete solution of this problem is of course in Carlsaw and Jaeger (1959). Bi +0 corresponds, in Carlslaw and Jaeger (pp. 114-115), to hl approaching zero, With this, the first root a: = \/(hl) < 1, and since the second root must be > mone has az > 2, and hence the second term in the sum in eq, (8) decays very rapidly indeed as com- pared to the first one, The first term is of course the zero order solution] 4 QUASI-STEADY.STATE APPROXIMATION A very common, and very powerful approximation which chemical engincers often use is the quasi- steady-state approximation (QSSA), which is useful whenever there are widely different time scales in- volved in a given process. [The QSSA can, rather surprisingly, be traced back to a 1913 paper by Bodenstein, see p. 343 of his paper]. Two typical examples are discussed in the following. The general message which I try to convey is the following one powerful as the QSSA is, it must be used with care. One first needs to thoroughly understand the physics 4693 of the problem at hand; next, one has to decide what kind of QSSA might be possible, and do a preliminary analysis ofthe implications of it; finally, one has to do the actual order of magnitude analysis of the problem. Hopefully, the two examples below will show how surprisingly simple final results can be obtained by following such a procedure. (0) Ziegler-Natta polymerization In Ziegler-Natta polymerization, a solid catalyst is introduced into the system in the form of very small particles, with radius Ro. The catalyst is sparged into a liquid phase of monomer, and polymerization oc- curs until rather large particles of polymer are pro- duced, Let us say one is willing to consume one part of catalyst per 1,000,000 parts of polymer; the final vol- ume ratio needs to be 10°, ie. the final polymer particle must have a radius 100 times larger than Ro. First the physics. A growing spherical shell of poly- mer forms on the catalyst; let R(@) be its radius. The monomer surrounding the growing polymer particle must (a) dissolve into the polymer. (b) diffuse towards the catalyst surface. (c) react on the catalyst surface. Step (a) is described by the requirement that, at the outer limit of the polymer layer, the concentration of ‘monomer equals its solubility co: ra RO, m Step (b) is modeled by writing the diffusion equa- tion in the polymer shell: D[Gefér* + Gc/Erhr] = ecjet. (72) Step (c) may be modeled by assuming that the reaction rate (per unit catalyst surface) is first order With respect to the monomer concentration at the catalyst surface. One thus writes, with k a kinetic constant per unit surface area having units of a velo- city Roy (73) However, R(t) is not known, and one needs a mass balance argument: whatever reacts on the catalyst surface contributes to the growth of R(®). If cy is the density of the polymer expressed as moles of mono- ‘mer units per unit volume, the equation should be may (75) Doclor = ke. ke(Ro, OR} = Rie, AR/dt ROO) = Ro. ‘The QSSA consists in regarding the diffusion phe- nomenon as steady, and hence the time scale of the reaction, which is the time scale of unsteadiness, must be large as compared to the diffusion time scale. The only combination of parameters having units of a ve- locity is D/Re: in addition we have the dimensionless group ¢,/co, and so the condition for the QSSA is expected 10 be k-< f(i/co)D/Ro. The form of function4604 J() can be guessed by considering that c, appears only in eg. (74) and there what matters is really the ratio kjey, so presumably the condition is kco/ey < D/Ro. This completes the preliminary analysis. The appropriate dimensionless variables and para- meters are of course b = c/eq: 2 = 1/Ro; = R/Ros 1 = tD/R3; A = keoRo/Dey, and the (expected) QSSA condition is 4 < 1. The equations reduce to S3+2(2) z= objet a) a z=) b=1 77 z=, dbjte=Ab 78) Abt, = 6? d6ae ”) 6-1 (80) The frst step of the QSSA consists in setting the RHS of eq. (76) to zero to obtain b= AG—AdINO+AS-—A BL) 2) ‘The QSSA holds provided 26/0 is negligibly small as compared to any one of the terms on the LHS of 4. (76), say: dolar = Ad + Ad— AY; (0) =0. Hd (63) é <2 Consider what happens at the outer layer of the polymer shell, z = (2). Let Db/Dr be the ‘substantial’ derivative of b along g(t), which is zero because of eq. (77) Db_ ab, (cb ag oF e+) ab ‘do ab 2 --(3)% i) Hence cq (8) is satisfied provided that (86) 2a, Equation (82) indicates that dg/dr decreases more than linearly with increasing @, and hence that eq. (86) is hardest to satisfy the smaller ¢ is. Since the smallest value of @ is unity, one requires eq. (86) to be satisfied at $= 1 to obtain the QSSA condition in the ex- pected form: A@2 (87) Equation (83) is not satisfied at z 0, where ebjex = — A, 22bjéz = 24: there isa neighborhood of time zero where the QSSA is not valid near the surface of the catalyst the QSSA is almost always the outer solution of a singular perturbation problem (Nayieh, 1973). At times other than zero and at 7 = 1 ee) (88) 1/261 + AG — A), G. Astarita Since the final value of @ of interest is 100 o more, eq. (88) in fact guarantees that, at later times, the QSSA js justified near the catalyst surface whatever the value of A may be. Having convinced ourselves that ifeg, (87) is satisfied we can indeed use the QSSA, one integrates eq, (82): Ara (A+ (68 = 1/3 — AWG? 12. (89) However, since eq. (82) holds only if eq. (87) is satisfied, inner consistency requires 10 only trust eq, (89) to within O(4), so that one gets the very simple final result: Pel + 3dr (90) This is an important point of a general nature. If a simplified model is used, one needs to establish the order of approximation; in the case at hand, the ap- proximation is valid to within terms of order A. It follows that the solution of the equations describing the simplified model is valid only to within the same order, and hence that terms of order A? and higher have to be neglected: it does not make sense to keep terms of an order which has been considered as n ligibly small when setting up the model. This may seem as a trivial conclusion, but in fact there is an abundance of results in the literature where terms are retained in the solution which are of the same order as terms which have implicitly been neglected. {b) Chemical vapor deposition T now move to another example which provides a somewhat unusual type of QSSA. Consider a CVD process where a gaseous stream containing an or- ganometallic compound C is produced by exposing a metal surface M to a low pressure gas containing a finite concentration of free organic radicals R (cg. an ethane gas at low pressure and high temperature will have a finite concentration of CHs radicals). The following chemical reaction may take place: R4+M=C 01) ‘The intrinsic rate ofthis reaction is extremely high, so that it may be regarded as instantaneous in a diffu- sion-reaction model. Let pg be the partial pressure of R in the gas phase; if one assumes that the solid surface is exposed metal M, the partial pressure of radicals at the surface will be zero, and thus the flux of radicals to the metal surface is given by M kepe (92) where ky is the pressure-based mass transfer coeffic- ient of the radical in the gas phase. The flux of radicals to the surface equals the num- ber of moles of C formed per unit surface, and if the ‘metal has to stay exposed an equal flux of C from the surface to the gas phase must establish itself. The transport equation which determines the flux of C away from the interface is ke(Pcr — Peo) with pre: the partial pressure of C at the interface, and Peo its value in the bulk of the gas. Since the flux ofDimensional analysis € must balance that of R, pcr will adjust itself to the appropriate value. However, the partial pressure of the organometallic cannot exceed its vapor pressure p*, so that, ifthe value of pcr needed to balance the radical flux exceeds p*, a layer of liquid organometal- lic will form, and the radical flux will be slowed down by the need to diffuse through the liquid layer in order to reach the metal surface. The highest possible value of pcr is p*, and hence the largest possible flux of C is Ne,ax = kep* 3) so that if Np as given by eq. (92)is larger than Ne, axe a liquid layer will indeed form, and the partial pres- sure of radicals at the surface will be Jarger than zero. In the limiting case where diffusion through the liquid layer is the rate-determining phenomenon, the inter- face partial pressure will in fact be equal to the bulk ‘gas partial pressure pp. Let c(x, ) be the concentration of radicals in the liquid organometallic layer, with x being the distance from the original position of the exposed surface, and 1 being time elapsed since the surface was exposed to the gas. The distribution (x,t) is determined by the ‘one-dimensional unsteady diffusion equation: Dé efOx? = dejar. (94) The liquid-metal surface is located at some position X= s(@). At that surface, the chemical reaction takes place at an essentially infinite rate, and therefore x=s(), c=0. (95) Now let dy and gc be the molar densities of the metal and of the organometallic, respectively, and define V = Neide and = du/dc (notice that B is likely to be close to unity}, A simple mass balance yields the result that the thickness f(¢) of the liquid layer increases at a rate given by dhjat = pdsjdt —V 08) It follows that the position of the exposed surface 4) moves downwards with a velocity V+ (1 Bids/dte: dq/dt = V + (1 ldsidt; 4(0)=0. 97) In addition to eq, (95), the other boundary condi- tion imposed on eg. (94) is, if co is the solubility of the radical corresponding to px =al, =o (98) In order to formulate a well posed free boundary problem, one needs an additional boundary condi- tion, since the position of the liquid-metal interface (0) is not known in advance [there are two moving boundaries in the problem, q(t) and s(), but once s(t)is, determined, q(t) is given by eq. (97)]. This is obtained from a mass balance at that surface. The flux of radicals reaching the surface is — Déc/dx, and this, represents the number of metal moles transformed to 4695 organometallic per unit surface area and time. Thus one can write SQ, — Dac/ox = Oy dside; s(0) = 0. (99) ‘The equations above constitute a well posed if somewhat unusual Stefan-like problem. The problem as formulated contains 4 dimensional parameters: D, V, dy, and co. An intrinsic length scale D/V and an intrinsic time scale D/V? suggest themselves naturally. Thus the dimensionless distance y, time +, exposed surface position 0, liquid-metal surface position 1. concentration 4, and Stefan number St, are defined as y= xv/D (100) rep (01) He) = g(e)¥/D (102) Ha) = st0)V/D (103) (104) (105) so that (106) d8jde = 1 + (1 Adulde (107) y= 00), (108) y= ue =0 (109) y= nl), — adjoy = Stduide; (0) =0 (110) The problem represented by eqs. (106)-(110) cannot of course have a completely steady solution, ic. an asymptotic solution where é¢/ér = 0. However, a QSSA asymptote may exist where the thickness of the liquid layer is constant, so that the exposed surface and the metal-liquid surface travel at equal and con- stant speed. Suppose that this is the case, with K the (dimensionless) constant thickness of the liquid layer: us) — 0) = K Equations (41) and (45) are satisfied if @ = 1/f and = K + 1/B. The concentration in the liquid layer is not constant in time at any value of y, since y is the distance from the original position of the gas-metal surface. However, concentration may be constant in time at any value of z= y —1/f, since z measures distance from the instantaneous position of the liquid-gas surface. This suggests that z is the sim- ilarity variable. Substitution of @ = f(2) gives indeed: ay S +s ip=0 (112) 4) = 1, [(K)=0 (113) —S(K) = Seip. (14) [Of course (0) = 0 is not satisfied since the QSSA only holds at sufficiently long times—again, it is the4696 outer solution of a singular perturbation]. The solution is. f= [exp( — 2/f) — exp( — K/)I/E1 — exp( — Kip] (115) K =piin(l + 0, (116) tis now of interest to estimate the time required in order to approach the quasi-steady state asymptote. The initial rate of growth of the liquid layer thickness is infinitely large, so one cannot base the estimate on that. At small times, |d0/dz| is significantly less than dude. In fet, at time zero the conditions are identical with those of the classical Stefan problem, and thus dujde is proportional to 1/,/t. Equation (107) now shows that the dominant term for d#/de at small times is the second one, and since f has a value close to unity, one concludes that [dix] is much less than dujdz. [Notice that if f > 1, d@/dr will be negative at small values of t.] Thus initially the process is domin- problem equations: in (118) For the problem at hand, St is necessarily a large number. It follows that eqs (116) and (118) can be expanded in powers of 1/St, and with the expansions ‘truncated at the first erm one obtains K = fist (119) A Stal (120) Requiring p(x) as given by eq. (117) to be equal to K provides the required estimate of the time needed to approach the quasi-steady-state solution, 1 hs p/m/ASt (2p This is guaranteed to be a very small number, and hence the time needed to reach the quasi-steady-state solution is significantly less than D/V? CONCLUSIONS Having known Prof. Sharma for a longer time than either he or I care to acknowledge, { am sure he enjoys the intellectual beauty, as well as the pragmatical usefulness, of back of the envelope’ calculations of the type which, I hope, I have shown can be done by Judicious application of dimensional analysis, scaling ‘arguments, and order of magnitude estimates. How- ever, it must be acknowledged that the exponentially increasing computing power available nowadays has relegated such calculations, at least in the mind of computer enthusiasts, into the realm of things that only old-timers find useful to do (and having quoted quite a number of papers which are older than 70 years I have certainly qualified myself as an old- timer), while modern scientists, computer literate as they are, scorn them as useless: after all, the argument seems to go, we kniow what the equations are, and the G. Astarita problem is simply to solve them, and who can do that better than a computer? (Do we really know what the equations are? Physicists at the end of the 19th cen- tury thought they did—and then came relativity) Even an appeal to the irreplaceable strength of the human brain's creativity is easily dismissed by com- puter enthusiasts: creativity, the argument goes, is nothing else than an ability to synapse out of previous knowledge seemingly new ideas, and therefore the computer can be creative as well as, in fact better than the human brain, since storing, immediate retrieval, and cross-referencing of existing knowledge is done best by computers. After all, there is such a thing as Artificial Intelligence! (But I have always felt that the adjective ‘artificial’ conveys the message that it is not quite as good as the real thing). ‘The argument is a metaphysical one, and there is no scientific way of establishing beyond doubt who is right and who is wrong, Hopefully this paper has, provided some ammunition to one of the two camps. Personally, I am convinced that the ability to ascer- tain what are the essential features of any given prob- lem will always be needed before the problem's solution is delegated to a computer; and I am also convinced that Prof. Sharma shares this viewpoint of mine. [tis therefore with great pleasure that I acknow- ledge the honor of being able to dedicate this paper to the 60th birthday of Prof. Sharma. NOTATION So many different examples have been discussed, that [ have been forced to use the same symbol with different meanings in different subsections. Hence no notation section is given. In a paper on dimensional analysis, requiring the reader to keep track, without help from a notation section, of symbols as they are introduced, and of the dimensions the corresponding quantities have, seems particularly appropriate, REFERENCES. Acrivos, A. (1960) A theoretical analysis of laminar natural convection heat transfer to non-Newtonian fluids, A..ChE.J. 6, 584-590, Acrivos, A, Shah, M. J. and Pete, E. E. (1960) Mo- mentum and heat transfer in laminar boundary layer flow of non-Newtonian fluids past external surfaces. A.1.ChE.J. 6, 312-317 Amon, M. and Denson, C. D. (1984) A study of the dynamics of foam growth: analysis of the growth of closely spaced spherical bubbles. Polym. Eng. Sci 24, 1026. Aris, R. (ed), (199) Aesthetic considerations in science and engineering, Department of Chemical Engineering and Materials Science, Univ. of Min- nesota. Astarita, G, (1977) Variational principles and entropy production in creeping flow of non-Newtonian fluids, J. Non-Newt. Fluid Mech. 2, 343-351. Astarita, G. (1983) On the relationship between a di- mensional and a thermodynamic peculiarity of the power-law constitutive equation. J. non Fluid Mech, 13, 223-224.Dimensional analysis Astarita, G. (1985) Scaleup: overview, closing re- ‘marks, and cautions, Scaleup of Chemical Processes, eds, A. Bisio and E. Kabel, pp. 677-690, Wiley, New York. Astarita, G. (1989) Thermodynamics. An Advanced Texthook for Chemical Engineers. Plenum Press, New York. Astarita, G. (1990) The engineering reality of the yield stress. J. Rheol. 34, 275-277. Astarita, G. and Ocone, R. (1988) Lumping non-linear Kinetics. A.J.ChEJ. 34, 1299-1308. Batchelor, G. K. (1967) An Introduction to Fluid Dy- namics. Cambridge University Press, Cambridge. Bertrand, J. (1878) Sur 'Homgénéité dans les For- mules de Physique, C. R, Acad. des Sei. 86, 915-920, Bodenstein, M. (1913) Eine Theorie der Photochemis- chen Reaktionsgeschwindgkeiten. Z. Physik, Chem: 85, 329-348, Boelter, L. M. K., Cherry, V. H., Johnson, H. A. and Martinelli, R. C. (1946) Heat Transfer Notes. University of California Press, Berkeley and Los Angeles. Boucher, D. F. and Alves, G. E. (1959) Dimensionless ‘Numbers. Chem. Eng. Progr., $519), 55-64; 75-83. Bridgman, P., (1922) Dimensional Analysis. Yale Uni- versity Press, New Haven, Buckingham, E. (1914) On physically similar systems: illustrations of the use of dimensional equations. Phys, Rev. Ser. 2, 4, 345-356. Carlslaw, H. S. and Jaeger, J. C, (1959) Conduction of Heat in Solids. 2nd edn. Oxford University Press, Oxford Carvalho, E. (1891) Sur une Similitude des Fonctions des Machines. La Lumiére Electrique 42, 506-507. chpole, J. P. and Fulford, G. (1966) Dimensionless Groups. Ind. Engng Chem. $83), 46-60. Clavenad, M. (1892) Sur une Rélation entre la Vitesse de Propagation, le Coefficient de selfInduc- tion, et Ja Capacité. La Lumiére Electrique 46, 215-218 De Gennes, P. G. (1979) Sealing Concepts in Polymer Physics. Cornell University Press, Ithaca De Jong, F. J. (1967) Dimensional Analysis for Eco nomists. Amsterdam, Dodd, S. C. (1963) The Probable Acts of Man. lowa City Dorgelo, H. B. and Schouten, J, A. (1946) On unities and dimensions. In: Proc. Kon, Net. Akad. Wet. 48, 124, Einstein, A. (1911) Elementare Betrachtungen ueber die Termische Molekular Bewegungen in Festen Korpern. Ann, Phys. 38, 679-694. Euclid, Stoichéia (Elements). (ca 300 B.C) In J. L. Heiberg and H. Menge, Leiprig 1883-1916, 8 vols. translation in T. L. Heath, The Thirteen Books of Euclid’s Elements Translated from the Text of Heiberg, with Introduction and Commentary, 2 vols, 2nd edn Cambridge, 1926 Euler, L, (1765) Theoria Motus Corporum Solidorum seu Rigidorum, Leonhardi Euleri Opera Omnia Swiss Society of Natural Sciences, Zurich (1954-1973). Fourier, J.B. (1822) Théorie Analytique de la Chaleur. Paris. Fulford, G. D. and Catchpole, J. P.(1968) Dimension less groups. Ind. Engng Chem. 603), 71-78. c 4697 Heath, L. T., (1921) A History of Greek Mathematics. Oxford University Press, Oxford. Huntley, H. BE. (1953) Dimensional MeDonald, London. Johnstone, R. E. and Thring, M. W. (1957) Pilot Plants, Models and Scaleup Methods in Chemical Engineering. McGraw-Hill, New York Juttner, F. (1909) Reaktionskinetik und diffusion Z. Phys. Chem, 68, 595-623. Klinkenberg, A. and Mooy, H. H. (1948) Dimension- less groups in fiuid friction, heat, and material transfer. Chem. Engng Progr. 44(1), 17-36. Langhaar, H. L. (1951) Dimensional Analysis and the Theory of Models. Wiley, New York. Lehman, P, J. and Craig, E. A. (1963) Dimensional analysis in applied psychological research. J. Ps)~ chology 56, 223-226 Leonardo da Vinci, Ms. F, back cover. from Johnstone and Thring, (1957) Macagno, F. 0. (1971) Historico-critical review of dimensional analysis. J. Franklin Inst. 292, 391-402, Maffettone, P. L. and Astarita, G. (1993) Devol- atilization of polymers, Adv. Transp. Proc. 9, 419-444, Minale, M. and Astarita, G. (1993) Heat transfer anal- ysis of the basal melting of Antarctic Ice shelves, ACh E.J. 39, 2019-2026, Murphy. G. (1950) Similieude in Engineering. Ronald Press, New York. Nayfeh, A. (1973) Perturbation Methods, Wiley, New York Newton, 1 (1687) Philosophiae Naturalis Principia ‘Mathematica, Book 2, Lemma 2. Streeter, London. Ng, M. L. and Hartnett, J. P. (1988) Free convection heat transfer from horizontal wires to pscudo- Analysis, Pranslation plastic fluids. Int. J. Heat Mass Transfer 31, 441—a49 Panton, R. L. (1984) Incompressile Flow, Ch. & Wiley, New York. Ptolemy, C, (ca 140 A.D.) On dimensions. In Claudii Prolemoi Opera quae extant Omnia, ed. 3. L. Heiberg, Vol. 2. Leipzig 1887; and P. Schnabel, Text und Karten des Ptolemaeus. Leipzig, 1938. Rayleigh, Lord, (1877-1878) The Theory of Sound. London. Rayleigh, Lord, (1879) On the capillary phenomena of Jets, Proc. Roy. Soc. 29, 71-97, Rayleigh, Lord, (1892) On the question of the stability of the flow of fluids. Phil. Mag. 34, 89-70. Rayleigh, Lord, (1904) Fluid friction on even surfaces. Phil. Mag. 8, 66-67. Rayleigh, Lord, (1915) The principle of similitude. Nature, 95 (2368), 66-68, Reynolds, O. (1883) An experimental investigation of the circumstances which determine whether the motion of water shalll be direct or sinuous, and of the law of resistance in parallel channels. Phil Trans. Roy, Soc. 174, 935. Riabouchinsky, D., (1911) Méthode des Variables de Dimensions Zéro et son Application en Aérody- namique, L’Aérophile, pp. 407-408. Schlichting, H. (1979) Boundary Layer Theory, 7th edn. McGraw-Hill, New York. Sedov, L. 1. (1959) In Similarity and Dimensional Methods in Mechanics, ed. M. Holt, Academic Press, New York.4698 G. Astarita Seinfeld, J. H. (1985) Atmospheric Chemistry and Phys- ics of Air Pollution, Wiley, New York. Stefan, J. (1891) Ueber die Theorie der Fisbildung insbesondere ueber die Eisbildung in Polarmeere. Ann. Phys. Chem. N.F. 42, 261-269. Vashy, A. (1892) Sur les Lois de Similitude en Phy- sique. Annales Telégraphiques 19, 25-28. Vitruvius, (35 B.C.) De Architectura, Williams, W. (1892) On the relation of the dimensions of physical quantities to directions in space. Phil. Mag. 34-38, 234-271, Williamson, S. J.(1973) Fundamentals of Air Pollution, Addison-Wesley, New York. Zierip, J. (1971) Similarity Laws and Modeling, Gas Dynamics Series, Vol 2. Marcel Dekker, New York.