Scribd
Upload
90 views24 pages

Elliptic Grid Generation Based On Laplace Equations and Algebraic Transformations

Elliptic Grid Generation Based on Laplace Equations and Algebraic Transformations

Uploaded by

Anonymous 1rLNlqU
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
90 views24 pages

Elliptic Grid Generation Based On Laplace Equations and Algebraic Transformations

Elliptic Grid Generation Based on Laplace Equations and Algebraic Transformations

Uploaded by

Anonymous 1rLNlqU
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
You are on page 1/ 24
8, 38-61 (1995) Elliptic Grid Generation Based on Laplace Equations and Algebraic Transformations S.P. Sperneuse* atone Aerogpace Ltoratry NLR, PO, Box 90802 1096 BM Amsterdam, The Meterands Received Masch 25, 198; revel October 3, 1938 An olliptie grid generation mathod ix presented to generate ‘boundary conforming gridsin domains in 2D and 3D physical space ‘end on minimal surfaces and porametrized surfaces in 30 physice! space. Tho elipic grid generation method i based on the use ‘fa composite mapping. This composite mapping consists of & honlinear transfinite algebraic wansformation and an elliptic tone iptie transformation is based on the Laplace eae ‘onthe Leplace-Beitram| equations for surtaces. igebraic transformation maps the computational space one- to-one onto @ parameter space. The ollipte transformation maps ‘tho parameter enace one-to-one onto the domains or eurfaces. The ‘composition of thase two mappings is 8 differentiable one-to-one ‘mapping from computational space onto the domaine or surfaces And has a nonvanishing Jacobian, The composite mapping defines the grid point distribution i the interior of the domains ar surfaces. For domains and minimal surfaces, the composite mapping obeys & ‘nonlinear slliptis Poisson system with contol funetions completely ‘defined by the algebraic transformation, The solution ofthe Poisson tysteme is obtained by Picard feration and black-box multignd ‘solvers. For parametrized curved curfaces, it ie nat necessary (0 define and solve a nonlinear elipic Poisson system. Instead ainear elliptic system and an inversion problem is solved to generate the ‘rid inthe interior ofthe surface. © te pessemi hos ts CONTENT 1 tation 2.2 rid genertin. 2.1, Desivaion of the 2D grid generation fxptions 2.2. Dicrezaton ad soon me, 23. Or hoponaliy at Rounds. 24, Ulusetons. 4 Suen sid genera om mini sear 44. Supe grid generat cn puruntied mnfcet, 41. Devon ofthe grid generason equations ‘method. 43, Mutations 42. Djscretization and sation 5. 3D grid generation. $1. Derivation of the 3D pid generation ‘spans 5.2. Dissietizaton and solution method. 3.3 hs 6. Conetshns INTRODUCTION A graphical interactive multi-block grid generator, called ENGRID, has been developed at NLR to construct multi-block, * il: spt nl en s998 Sam PS hy Redbone. ‘Aaah of epedtinn ay fm ret * structured grids for the computation of flows based on the ler and Navier-Stokes equations about complete aircraft Configurations, including propulsion aircraft components [11, 13-15}. Advanced algebraic grid generation techniques arc applied to construct the grids (13, 14]. Extensive use of EN- GRID at NLR and Fokker has demonstrated that the applied techniques are fast and sufficiently robust to create grids for the simulation of lows based on the Euler equations. However, the applied techniques show too often grid folding when Ni vier—Stokes grids are generated in the interior of curved sur= faces and blocks with complex shapes, Therefore, more robust ‘grid generation techniques with a minimum of grid tuning parameters were needed to construct such grids efficiently. For this purpose a new elliptic grid generation method has been developed with a maximum of robustness and a minimum of ‘grid tuning parameters. The method has been incorporated into the ENGRID code and has been applied successfully to generate ‘boundary conforming Navier—Stokes prids in blocks and block- faces with complex shapes. This new elliptic grid generation ‘method is the topic of this paper. Since the pioneering work of Thompson on elliptic grid generation itis known that systems of elliptic second-order patil differentia equations produce the best possible gids in the sense of smoothness and grid point distribution, The systems ‘of ellitic second-ovder partial differential equations are Poise son-lype systems with control functions to be specified. The secret of each “good” elliptic grid is the method to compute the control functions [10] ‘Originally Thompson and Warsi introduced the Poisson sys- tems by considering curvilinear coordinate system which satisfies a system of Laplace equations and which is transformed to another coordinate system [1, 2}. Then this new coordinate system satisfies a system of Poisson equations with control functions completely specified by the transformation between the two coordinate systems. However, Thompson did not use this approach for grid generation. Instead he proposed 10 use the Poisson system with control functions specified directly rather than through a transformation [1] Since then the general approach is to compute the control functions at the boundary nd to interpolate them from the boundaries into the field (1, 3-10) ELLIPTIC GRID GENERATION 39 The main disadvantage of such an approach is that itis then not possible to prove that the system of Poisson equations defines a one-to-one map so that the computed grids may con- tain grid folding. In this paper we will show thet also Thompson's and Warsi's original idea to define the control functions by a transformation, can be used for grid generation. An important advantage of this approach is that the corresponding Poisson system defines 1 one-to-one map if the ttansformation is one-to-one. It will be shown that itis not difficult to construct appropriate one~ {o-one transformations. For this purpose, nonlinear ransfinite algebraic transformations will be used. We will apply this approach to generate boundary conform- ing grids in domains in 2D and 3D physical space and on minimal surfaces and parametrized surfaces in 3D physical space. ‘Thus, the underlying concept of the proposed grid generat ‘method is to use a composite mapping. The idea isto introduce. 1 parameter (coordinate) system in the given domains and sur- faces which only depends on their shape and not om the pre scribed boundary grid point distribution. The parameters are defined as normalized arclength at the boundaries and each parameter obeys the Laplace equation in the interior of a domain, ‘or the Laplace-Beltrami equation in the interior of a surface, For 2D domains and surfaces, the parameter system can be interpreted as a differentiable one-to-one mapping from a unit square, called the parameter space, onto the 2D domain or surface. This mapping is calle the elliptic transformation. The parameter space is a unit cube for a domain in 3D pliysical space. ‘A nonlinear transfinite algebraic transformation is con- structed to control grid point distributions. This transformation ‘maps the computational space onto the parameter space, The computational space is also considered as a unit square for 2D domains and surfaces and as a unit cube for 3D domains. Grids in computational space are always uniform, ‘The algebraic transformation depends on the prescribed bound- ary grid point distribution, The algebraic transformation is con: structed in such a way that the mapping is also differentiable ‘and one-to-one, ‘Thus, the algebraic transformation maps the computational space onto the parameter space and an elliptic transformation maps the parameter space onto the domains or surfaces in physical space. The composition of these two mappings is a differentiable one-to-one mapping from computational space ‘onto the domains or surfaces in physical space and has a nonvan- ishing Jacobian. The composite mapping defines the grid point distribution in the interior of the domains or surfaces. Although the composite mapping is one-to-one, this does root imply that generated grids are always grid folding free, because the discrete equations may not share this robust prop- enty {7}, But it is guaranteed that grid folding will always isappear when the grid i refined. Furthermore itis our experi- ence that grids produced by the composite mapping are hardly ever folded even when Navier—Stokes type of grids are genes ated in domains or surfaces with complex shapes. ‘The elliptic transformation is independent of the prescribed boundary grid point distribution and may thus be considered as a property of the domain or surface itself. The algebraic, twansformation depends on the prescribed grid point distrib tion, As we will see, the interior grid point distribution in parameter space, generated by the algebraic transformation, always a good reflection of the grid point distribution at the boundary of the parameter space, Therefore, the interior grid point distribution in the domains and surfaces is also a good rellection of the prescribed boundary grid point distribution ‘This is not the case for grids solely based on the system of Laplace equations. Then the inherent smoothness ofthe Laplace. ‘operator makes the grids evenly spaced in the interior (for example, a boundary layer will be blown up and completely, disappear). Therefore, grid generators solely based on the sy tem of Laplace equations are unusable in practice. "Thompson [1] and Warsi [2] have shown that the composite mapping obeys an elliptic Poisson system with control fonctions completely defined by the algebraic ransformation. The num- ber of control functions is six for 2D domains and surfaces and 18 for 3D doniains. In our case, the control functions are speci- fied by the algebraic transformation only and it is, therefore, rot needed to compute the control functions at the boundary and ( interpolate them into the interior of the domains or surfaces, asis the case of all well-known elliptic grid generation systems based on Poisson systems [1, 8-10]. ‘Also new and more useful expressions for the control func tions are derived in a short and elegant way which only depend, ‘onthe algebraic transformation itself and not also on the inverse Of this transformation (which occurs in the expressions used by Warsi and Thompson). ‘The computed grids are in general not orthogonal at the boundary. Sometimes, grid orthogonality is very much desired. It is shown that the algebraic transformation can be redefined to obtain a grid which is orthogonal at the boundary. ‘The nonlinear elliptic Poisson equations are solved by Picard iteration. The linearized equations are solved by excel- lent black-box multigrid solvers developed by P. M. de Zeeuw at CW. {16-18} ‘The paper is organized as follows: In Section 2, the Laplace equations, the parameter space and the algebraic transformation, are presented for domains in two-dimensional physical space. ‘The resulting Poisson equations are derived, together with the appropriate expressions of the control functions. The diseretiza- tion and solution of the nonlinear elliptic equations are dis cussed and, also, the orthogonalization of the grid at boundaries, Examples of grids in 2D domains are given, Surface grid generation on minimal surfaces is discussed in Section 3. Itis shown that grid generation on a minimal surface isin fact the same problem 2s grid generation in a domain in, 2D physical space. Illustrations of grids on minimal surfaces, are given 40, S.P. SPEKREUSE Surface grid generation on surfaces with a prescribed shape is teated in Section 4, I is assumed that such surfaces are parametrized and that the parametrization is « differentiable one-to-one mapping from a unit square onto the surface, The generated surface grids are independent of the parametrization. The solution method to genera the grids in the interior of| parametrized surfaces is differen fom that used for minimal surfaces [tis much easierto solve directly the vo linear elliptic partial differential equations defined by the Laplace-Beltramt equations instead of inerchanging the dependent and indepen dent variables, which leads to nonlinear elliptic system of partial differential equations. An inversion problem must then be solved afterwards. Such a simple solution method is only possible for parametrized surfaces. This i due tothe fact that an initial grid folding fee surfece grid on a parametrized surface can be easily generated because the given parametrization is one-to-one. Mlystrations of grids on parametrized surfaces are given. Grid generation in 31) domains is treated in Section S. The elliptic and algebraic transformation is defined. The resulting Poisson equations are derived, together with the appropriate expressions of the conirol functions. The diseretization and solution of the nonlinear ellipic equations is discussed and ‘examples of grids in 3D domains are given. Finally, concluding remarks are made in Section 6 2. 2D GRID GENERATION 2.1. Derivation of the 2D Grid Generation Equations Consider a simply connected bounded domain & in two dimensional space with Cartesian coordinates x = (x,y). Sup pose that is hounded by four edges, E, Fs, By. Ex. et (Ei ) and (Es, Ex) be the two pairs of opposite edges as shown in Fig. Define the computations space‘ asthe unit square in two-dimensional space with Cartesian coordinates & ~ (E, ‘Assume that a mapping x: 3+ 2%) is prescribed which maps the boundary of € one-to-one on the boundary of 9. This mapping defines the boundary grid point distribution. As- ssume that + €= 0 at edge E, and €= 1 at edge Ey, + n= at 0 at edge E; and = | at edge E, We wish to construct a mapping x: H+ % which obeys the boundary conditions and which is a differentiable one-to-one rapping. Furthermore, we requise thatthe interior grid point distribution is a good reflection ofthe prescribed boundary arid point distribution ‘A natural mapping x:"@ ++ 9 exists which obeys these requirements. This mapping will be the composition of an alge- braic transformation and an elliptic transformation based on the Laplace equations. The algebraic transformation is a differ- entiable one-to-one mepping fom computational space @ onto ‘a parameter space ®. The parameter space is also a unit square. Wee will sve below thatthe algebraic transformation will only depend on the prescribed boundary grid point distribution at the four edges of domain ®. The elliptic transformation is a differentcble one-to-one mapping from parameter space ® onto ‘domain &. The elliptic transformation will only depend on the shape of domain 3 and is thus independent of the prescribed boundary grid point distribution. The elliptic transformation ‘may thus be considered as a property of domain 2, The compo- sition of these two mappings defines the interior grid point distribution and is a differentiable one-to-one mapping from computational domain onto domain &. Introduce the parameter space Q as the unit square in a two-dimensional space with Cartesian coordinates $ = (17 Require that the parameters s and t obey: at edge F; and s = 1 at edge F:, + sis the normalized arclength along edges Fy and &, + 1= 0 atedge & and = I at edge Ei, + ris the normalized arclength along edges E, and E>. 1 t 7 o 7 o Pp a ot ® L : ‘Computational space C Parameter space P Domain D FIG. Transfomnatin fom computational (7) space to 2 domain in Cartesian (s ) space ELLIPTIC GRID GENERATION 4 ‘Thus 8:89 ++ 2 is defined by these requirements. In the interior of @ we require that ¢ and rare harmonic funetions of x and y, thus obey the Laplace equations: w @ ‘The two Laplace equations As = 0 and Ar = 0, together with the above specified boundary conditions, define the mapping. 5:21» ®. Note that this mapping only depends on the shape ‘of domain @ and is independent of the prescribed boundary vid point distribution, By interchanging the dependent and independent variables, a nonlinear eliptic partial differential ‘equation can be derived for x: +» @. Thus we have to solve «nonlinear elliptic boundary value problem in ® in order (0 define this mapping. This mapping defines our elliptic transfor- ‘mation. Ic is well known that this mapping is differentiable and ‘one-to-one [4] ‘The algebraic transformation must be a differentiable one- to-one mapping from computational space & onto the parameter space ®. Because x:0°@ > a9 is prescribed and x: a> a@ is defined as described above, it follows that $:4€ +> a0 is also define. From the preceding requirements it follows that 510.) = 0, 0,9) = 560) = 5508, 56,1) @ 56 where the functions s,,.., are monotonically increasing, and 6,0) 10, n) D1 oy my tsa) fem, Where the functions 1. 80 are monotonically increasing ‘Thus the four functions #5 (m),te(m) 56 (+55, (6) ae defined by the boundary grid point distribution. The mapping s: @-> is now defined by two algebraic equa- tions: tel 9 + 56h 6 ta(mX = 5) + tals © Equation (5) implies that a coordinate line €= consti mapped tothe parameter space ® as a straight line: s isa linear function ‘of 1; Eq. (6) implies that a grid line 1 = const is also mapped to as a straight line: ¢ is a linear function of s. For given values of €and 7, the corresponding s and r values are found as the intersection point ofthe two straight lines, For this reason, the system defined by Eqs. (5), (6 is called the “algebraic straight lnc transformation” because ofthe use of straight ines in parameter space #. It canbe easily verified that this system defines «differentiable one-to-one mapping because of the positveness of the Jacobian sey ~ sy > 0. “The system defined by Eqs. (3), (6) can be interpreted as @ transinite interpolation with nonlinear blending functions and resembles the transite interpolation method of Soni (6) “The algebric transformations» and the elite tens- formation x: H @ are differentiable and one-to-one, Ths the composite mapping x:'G > 3 defined as x(£) = xis() is also differemiable and one-to-one. Furthermore, due 10 the ropes ofthe basic mappings, we may indecd expect that the interior grid point distibuton will be a good reflection of the boundary grid point distribution. Inthe remainder of this section, we wll derive the set of nonlinear lpi patil cfer ential equations which the composite mapping x = x(s(€)) has to fui It has already been noted by Warsi and Thompson that the composite mapping will obey an elliptic system of Poisson equations. However, the system of Poisson equations as given iin [1, 2) is not so useful because it contains control functions which depend also on the derivatives of the inverse mapping £:9++4, Itwillbe show below thas its not dificult o obtain expressions for these control functions which only depend on the derivatives of the mapping 82° > ® itselE First, introduce the two covariant base vetors o ‘and define the covariant metric tensor components as the inner product of the covariant base vectors a, = (aa), 1, 2hJ = {1 2 ®) ‘Then the contravariant base vectors a and a! are defined ac- cording to the rules (al.a) = 8). 1.2, j= {1,2 co) ‘with 6} the Kronecker symbol. Define the contravariant metric tensor components a= aia), 1-1 2F= 1, (oS) ala, + ala, a ao so that ay and a2) a S.P, SPEKREUSE Introduce the determinant J? of the covariant metric tensor: P= aydn ~ ai. Now consider an arbitrary function g = 4&9). Then ¢ is also defined in domain & and the Laplacian of dis expressed as Ab = bu t by = Adel by + Jal by), + (ade * SoS (3) which may be found in every textbook on tensor analysis and differential geometry (for example, sce (20, p. 227). Take as special cases respectively = and = 7 Then Eq, (13) yields AEH Flu), + Ua), an= Hua, + Ue) (4) ‘Thus the Laplacian of can also be expressed as AG = ally + 2a!Gyy + by, + AES, + Andy. C15) Substitution of respectively d = s and 6 = £ in this equa tion yields As = alg + 20% 4 a5 4 ME5, +475, (16) ALE alg + 2aleg + ayy AEE Sty. UT) Using these equations and the requirement that s and are hharmonic in domain %, thus, As = 0 and Ar = 0, we find the expressions for the Laplacian of £ and 7, 8 ae «gy where m0 The six coefficients of the vectors Py = (Py, PA)", Pa (Ph, PisJ',and Py = (Ph, Pi)" are so-called control function ent and less useful expressions of these control functions can also be found in 11,21 0) Finally, substitution of = x in Eg, (15) yields Ax = allxy + 20x + a Xy + AER: + AnX,. (21) Substitution of Eg, (18) into this equation and using the fact that Ax = 0 we arrive at the Poisson grid generation system, alg + 2a + a2 (al Py +a Ph + Phx 2) + (GNP 2a"Py + dP Ine = 0. Using Eqs. (8) and (11) we find the following well-known ‘expressions for the contravariant metric tensor components: ene Pat = Oy. Xoh ag = 0%. %)s 23) Fa = ay = (89. ‘Thus the Poisson grid generation system defined by Ea. (22) can be simpliied by multiplication with J. Then we obtain Ang + 2a Rpg + oy + ("Pl + 2a Ph + a PLdxy + (a!Pi, + 20"Phy + a PiL)x, ay and = ye yh a= Uy. Kye = (Ry AD. (25) ‘These equations, together with the expressions for the control functions P given by Eq. (19). form our 2D grid generation system. Grids are computed by solving this quasilinear system of elliptic partial ditferental equations. The discretization of this Poisson system is described in the next section, 22. Discretization and Solution Meshod (Consider a uniform rectangular grid of (N+ 1) x (M+ 1) points in computational space ‘€ defined as y= GSUN, ny = WHIM, 12 0.NF= 00M. 26) ‘Assume that x, is prescribed on the boundary of this grid and consider the computation of x, inthe interior of the computa ‘ional grid based on the solution of the Poisson system defined by Bq. (24) First we will compute the arclength normalized variables s,, and 4, based on the algebraic transformation defined by Fas. (5) and (6). The arclength normalized variables s,, and f are computed atthe boundary of the computational grid as follows, Compute the distance between succceding points at the boundary y= Ny mths Ty = Mixa Kye FM, ELLIPTIC GRID GENERATION limo ita div = [Rw Menah FTW 8) Define the length of edges Fi, Es, Fa, Eu by 09) and the normalized distances as Ay adley day = Aufl, G2 1M, — GO) da" Alley du = Fuullg, F=1uN BD ‘The arclength normalized variables s,;and fal the boundary are then defined by Sy = 0, y= f= OM, 2) wl P= ON, @3) and Sa = Seip tdi Su Sint dan E= aN GAY y= haya Hays Bey taps dye J= eM (3) ‘The arclength normalized variables (s,,1,) inthe interior of the grid are now computed according tothe algebraic straight Tine transformation defined by Eqs. (5) and (6) and are th found by solving simultoneousty the two linear algebraic equa- tions <1) + seals 36) ~ 8) + ts an for each pair (ij) © (VN =I; VM =D, ‘At each grid point (i, j), the six control functions P\,, Ph Ph, Ph, Ph, Ph defined by Eq. (19), are now easily computed using central differences for the discretization of sy, Seas Sms StS MA He Keys En fe by [Next, consider the iterative solution process ofthe nonlinear cliptic Poisson grid generation system defined by Eg. (24) Rewrite this system as, Px t 20K + RX + Sap Tx, = 28) with P= (Ky) O= Bux) R S= PPh + 2QPh + RPL, 69) T= PPI, + 207+ RP a ‘The solution of this system of nonlinear elliptic equations is obtained by Picard iteration, Point 20MIN + ROL, + SEE TM =O, 0) with Pom @yla, Qh= Pi xy), REY = ext eae ay St = PIP S20" PL, + REPL, Tet PLP + 20°1P, + ROP ‘Thus a current approximate solution BAGO NZS OM}, 42) is improved by the following steps: + Compute the coefficients P.O, RMS, TH! by -pplying central differences for the discretization of x}-" and X;'. Note that the six control functions remain unchanged uring the iterative procedure. * Discretize we xy}. 4 by using central differences. The discretization of the mixed derivative x, is done in a way as described in (19} + After the discretization of Xi, xy, Xiy Xf. X5, We arrive at 4 linear system of equations for the unknowns Xi, 1 = 0 Nj = 0... M with Dirichlet boundary conditions, At each intetior gid point (,j) we havea nine-point stencil. This linear system is solved by the excellent black-box multigrid solver MGDSV developed at C.W.L by P.M. de Zeeuw {16}. The multigrid solver MGD9V is called tice to compute the two components xi, and y of xi. The solution of the linear system provides a better approximate solution x ‘The complete process is repeated until a sufficiently accurate solution has been obtained. The initial start solution x” is ob- tained by algebraic grid generation. The final grid is indepen- dent of the initial grid. Moreover, the quality ofthe initial grid is unimportant and severe grid folding of the initial grid is, allowed. {n general, about 10 Picard iterations are enough to ‘obtain a sufficiently accurate solution of the nonlinear elliptic Poisson equations, 23. Orthogonality at Boundaries Grids obvained by the nonlinear elliptic Poisson grid genera- tion system defined by Eq. 24) are grid folding free and have excellent intetior grid point spacing distribution. However, the computed grids are in general not orthogonal atthe boundary and sometimes grids should be orthogonal at the boundary. Especially for Navier-Stokes computations, the orthogonality of the grid in a boundary layer is often desired. “4 S., SPEKREUSE Grid orthogonality at boundaries can be achieved as follows, Suppose that 2 grid has been computed based on the solution ‘of the Poisson grid generation system with control functions specified by the algebraic straight-line transformation. Suppose that it is desired that the grid is orthogonal at all four edges of domain 2. Redefine the elliptic transformation x: 9 +» % by imposing the following new set of boundary conditions forthe harmonic functions » and r: = Oat edge £, and s = 1 at edge E;, + asin = 0 along edges E, and £, where n is the outward normal direction, O at edge F, and = 4 at cdge Ey + aifan = 0 along edges E, and E;, where m is the outward, normal direction ‘These new boundary conditions define a new mapping x P+. Thus s is no longer the normalized arclength along, edges £; and Ey, and £ is no longer the normalized arciength ‘long edges E, and E,. It is not difficult to understand, by applying Gauss’s integral formula for harmonic functions, that ‘the Neumana boundary condition @s/@n = Oimply that sannot hhave a local extremum at edge E and edge E, of domain . ‘Similarly, cannot have a local extremum at edge E, and edge E,. Hence, s i still monotone along edges F, and B., and ris still monotone along edges E, and F:. ‘The Neumann boundary condition as/n = 0 along edges sand E, also imply that a parameter fine s = const is a curve in domain J which is orthogonal at those edges, Similarly, a parameter line t ~ const is a curve in which is omhogonal at edge E, and edge £5 It is possible to compute the new harmonic functions s and 1 directly as functions of the computational coordinates (7) because of the existence of an initial mapping x: €i+ BD, We ‘only have to solve two Laplace equations As = 0 and Ar = 6, together with the above specified boundary conditions, on an existing grid in domain 2. This is an elementary classical problem and the solution ean be obtained rather easily. Some remarks about the discretization of the Laplace equation with Neumann boundary conditions are given at the end of this Write the solution as s = 5(, m) and 1 = i(é, »). For our purposes, the only important information is the solution on the boundary. Redefine the edge functions by (9 -HEO, (Y= HED, 10.7). (43) fem) = HO, ntl) ‘These new edge functions are still monotonically increasing. ‘The algebraic transformation 8: -» 9 is now redefined according to two algebraic equations, 5 = 55 (QH) + 55 OHO, a 1 te (m)He(8) + te (DHS), as) where Hy and H, ate cubic Hermite interpolation functions defined as Hos) = (+25), HYG)= 3-29), OSSS1 46) Note that H(0) = 1, H§(0) = 0, Hy(l) = 0, Hi(1) = O and (0) = 0, Hi(0) = 0, HQ) = 1, Hi) = 0. It follows from Fg. (44) that coordinate Fine €= const is mapped to parameter space & as a cubic curve which is onthogonal at both edge Ey and edge E, in ®. Such a curve in parameter space ® will thus bbe mapped by the new elliptic transformation x: > % as a curve which is orthogonal at both edge Ey and edge E, in 2. Similar observations can be made for coordinate lines 7 const. Thus the grid will be orthogonal at all four edges in domi 2 ‘The composite mapping x:%E + 5 still obeys the Poisson arid generation system defined by Eq. (24). Thus the same system of elliptic equations can be solved to generate an orthog- coal grid at the boundary. The only difference is thot now s Gr @ is defined by Eqs. (#4) and (45) instead of Eqs. (5) and (6). Grid orthogonality at boundaries is obtained in three steps. First compute an initial grid based on the Poisson grid gonera- tion system with control funetions specified according to the algebraic straight line transformation detined by Eqs. (5) and (. Next solve the two Laplace equations As = 0 and At = 0, together with the above specified houndary conditions. on this inital grid to obtain new edge functions t=), =m)» se (2), 5e(4). Then recompute the geid based on the Poisson system but with control functions specified according to the Algebraic transformation defined by Eqs. (44) and (43) Grid orthogonality at boundaries may introduce ged folding, Fortunately, grid folding will not be easily introduced. From Eq, (44) it follows that two different coordinate lines € = &. £= & & # &, are mapped to parameter space ® as wo disjunct cubie curves which are othogonal at both edge Ey and edge Ez in ®, This is due 10 the fat thats () and s,(€) are monotonically increasing functions. The same holds for differ ent coordinate Fines 7) = 7). = Nes m) Ma. For given values of €and 7, the corresponding s and r values are found as the iersection point oftwo cubic curves. However, two'such cubic ‘curves may have more than one intersection point, In that case arid folding will occur. In practice we hardly ever encounter ‘id folding due 1 orthogontlization ‘We have described a method to obtain an orthogonal grid tall four edges of domain 2. ln practice, orthogonality ofthe arid is often only desired at one edge or two or three edges. Suppose, for example, thet itis only desired to have an omthogo- nal arid at edge F In that case, te), fo). and 5e(6) are defined as normalized arciength. Only 556) i8 computed by demanding that s/an = 0 along edge Fin 2. Thus only one ELLIPTIC GRID GENERATION 45 Laplace equation As = 0 has tobe solved to obtain (with Dirichlet boundary conditions at edges E, , Ey arid a Neu- mane boundary condition at edge f). Furthermore itis st cient thatthe algebraic transformation §:% -» 9 is Such that a coordinate line 7 = const is mapped to & as a straight line and that @ coordinate line € = const is mapped 10 as @ parabolic curve which is only onthogonal at edge Fin Finally we will show how the Laplace equation As together with the above specified boundary condition, cretized and solved on an existing gd in . The discretization and solution of Ar ~ O is obtained inthe same way. Consider a uniform rectangular grid of (V+ 1) x GM + 1) point in computational space es defined by Fa. (26). Thus x, is defined forall grid points (jj). From Eq, (13) it follows thats obeys in computational space the lincar second-order elliptic equation (als, + Jas.)g + (als, + Jay) ay div(A grad s) = 0, (48) where the matrix A = ACE, n) is defined as sai (ew ‘At an interior grid point (i,j), the coefficients of matrix A can be directly computed by using central differences for x and x,. Thus Eq, (48) is a linear diffusion problem with given variable coefficients A finite-volume cell-centered approach is used to obtain the discretized equations. Integration of Eq, (48) on a contol vol- ume © C € gives J siv(a grads) aan [jgaemismas 0) ftes.an)de =, where m is the outward unit normal vector and de is a line clement, At an interior grid point, j) in, the discrete equation 1s derived in a straightforward way by applying Eq. (50) for a rectangular control volume , with sizes UN and 17M around (Ei 14)) = GUN, j7M). The result i a nine-point stencil Half control velumes are used for boundary grid points. It is not difficult to show that the Neumann boundary condition, as/an = 0 at a boundary in @ transforms to (grad s, An) = O at the corresponding boundary in . Thus the fux is zero. ‘This observation makes the discretization at boundary grid points also straightforward After discretization, we obtain a fnear system of equations for the unknowns {51,1 = 0. N: j= 0 .. M} with Diieet and Nevmann boundary conditions, At each interior grid point Gi. ) we have a nie-poin stencil. This linear system is solved ty the black-box multigrid solver MGDSY (161. 2.4, Illustrations Examples of grids in 2D domains are shown in Figs. 7-17 (see Section 5.3) All grids are grid-folding free and the interior rid point distribution is a good reflection of the prescribed ‘boundary grid point distribution. An jnitial grid (obtained with algebraic grid generation) is required as the starting sotution, for the nonlinear elliptic Poisson system. The final ellipic grid is independent of the initial grid. Moreover, the quality of the initial grid is unimportant and severe grid folding of the inital _gri is allowed. Figure 7 shows a region about a NACAOOI2 airfoil subdi vided into four domains. The domains have common edges. ‘The total number of edges is 12. The boundary grid poi listibution is preseribed at all 12 edges. Figure 8 shows a complete O-type Euler grid. Grid orthogonality is prescribed atthe interior edges and at the boundary of the airfoil. A close= up near the airfoil of the domains and grid is shown in Fig. 9 and Fig. 10. Figure 11 shows a region about a RAF2822 airfoil, also subdivided into four domains. Again, the boundary grid point distribution is prescribed at all 12 edges and grid orthogonality, is prescribed at the interior cdges and at the boundary of the airfoil. Figure 12 shows a C-type Navier-Stokes grid. A closeup. ‘of the grid ncar the airfoil is shown in Fig. 13. Figure 14 shows an initial grid around a complex artificial houndary With severe grid folding. This initial grid is obtained with an algebraic grid generation method. Figure 15 shows the Navier-Stokes grid around the complex artificial Boundary ‘obtained with the elliptic grid generation method. Grid orthogo- ality is prescribed. This grid illustrates the robustness of the clliptic grid generation method. Figure 16 and Fig. 17 show details ofthe elliptic grid at respectively a convex and aconcave part of the boundary. There is only some slight tendency that {gfid lines are more closely (widely) spaced near convex (con- cave) parts of the boundary. 3, SURFACE GRID GENERATION ON ‘MINIMAL SURFACES. Grid generation on a minimal surface is in facta straightfor- ward extension of grid generation in a domain in 2D physical space. Consider four connected curved edges situated in 3D. physical space. A minimal surface is then defined as a surface bounded by these four edges and with zero mean curvature ‘Thus the shape of the minimal surface is a soap film bounded, by the four curved edges, Again, a parameter system with ‘wo parameters is defined. The two parameters are normalized arclength at the four curved edges, Furthermore, it is required 46 S.P, SPEKREMSE that both parameters obey the Laplace—Beltrami equation for surfaces. These two equations, rogethor withthe requirement that the mean curvaure is identically 210, define a differenti- able one-to-one mapping from parameter space (a unit square) ‘onto the minimal surfece. Thus this mapping is independent of the prescribed boundary gid point distribution at the four edges, ‘The same algebraic transformation as used for domains in 2D physical space is applied to map the computational space (a ‘unit square) onto the parameter space. We will now show that the set of nonlinear elliptic partial differeatia equations which the composite mapping has to Ful- fills the same Poisson s¥stem as defined by Eq, (24) but with x= (x, » 2) instead of x = (3, »)¥. Thus grid generation on 2 minimal suface in 3D physical space isin fact equivalent to arid generation in a domain in 2D physical space. The result that a Poisson system of the form as defined by Ea. (24) ean bbe used to compute a grid on a minimal surface can also be found as a special application of the formulas derived in {3 A in the (wo-dimensional case, consider again four curved edges Ey, Bs, Fs, Ey, but now situated inthe three-dimensional physial space with Cartesian coordinates x = (3 3,2)". Let (E;, 2) and (5, £2) be the to pairs of opposite edges as shown in Fig. 2. Introduce the parameter space @ es the unit square in two mensional space with Cartesian coordinates s = (5,1. Again require hat the parameters s and 1 obey: + sO atedge Ey and s = 1 at edge Es, + sis the normalized arclength along edges Ey and E. at edge Ey andr = | at edge Ea, + ris the normalized arclength along edges Band Ey Furthermore, require that Where A isthe Laplace-Beltamt operator for surfaces and IY is the mean curvature. ‘These three requirements, together with the described bound: ary conditions define a usique mapping x: -» 3°. The shape of the surface defined by this mapping is © minimal surface because of the requirement that the mean curvature H is 2210. ‘The parametrization of the surface is defined by Eqs. (St) and (52) Define the minimal surface $ as F = {als 0 (5, HEF}. 54) Consider a prescribed bourdary grid point distribution at the four edges E.£3. Es, E, ofthe minimal surface ¥. Mathemati- cally, she boundary grid point distribution can be defined as a mapping x:3% + OY, where © is the computational space defined as the unit square in a twowdimensional space with Canesian coordinates & = (Em). Because x05 H> a is prescribed and x: > a is defined as described above, it follows that :9°€ > 29 is also defined In exactly the same way as for the two-dimensional case, the mapping $:"€ > @ is defined by the algebraic straight-line transformation defined by Eqs. (5) and (6). The mapping t+ F is defined by Eqs. (51)=(53). The composite mapping x:%6+ Fis dofnod as x = x(6(6)) and deserbes the interior 6 point distibution on the minimal surface J. Note that this composite mapping will be differentiable and one-to-one. ‘What remains tobe done isto derive the system of nonlinear elliptic partial differential equations which he composite map- ping has to obey. Then the solution of this system defines the {ntcrior grid point disibution on the minimal surface For this purpose, introduce the two covariant base vectors 55) ‘The two covariant base vectors span the tangent plane of Sf at axe m= x, As=0, (51) the corresponding point P. Define the unit surface normal as aro, (52) te H=0, (53) 66) Q t 1 o 1 o f io ol _ ® fl = = y a @ d @ ° ad © "3 x Computational space C Parameter space P Minimal surface S FIG. 2, Transformation rom computation! (6 1) space 10a minal surface $ in Cansian x,y. 2) pace BLLIPTIC GRID GENERATION a7 where /\ is the vector product operator. The contravariant base vectors a! and a are defined according to the rules (a) = 8, F= U2 Mh 67) and a’, aa) (58) “Thus the two contravariant base vectors are also lying in the tangent plane of Fat the corresponding point P. Define the covariant meric tensor components by Eq. (8) and the eoatra- variant metre tensor components by Eq. (10). Then Eqs. (11) and (12) ae sll valid, Again introduce the determinant of the covariant metre tensor J* = ayn = a ‘Now consider an arbitrary function & = ¢(é, 1). Then ¢ is also defined on surface ¥ and the Laplace—Belurami operator of d is expressed as ' 7 (Jal by + Jal", + Ja"*be + Ja*b,),} (59) (ee (20, p. 2271. As in the two-dimensional ease, substitution of @ = End $= 7 into this equation yields Eq. (14). Thus, ‘the Laplace-Beltrami operator of ¢ can also be expressed as defined by Eq, (15). Substitution of respectively 4 = s and & in Eg. (15) and using the requirements expressed by Eas. (51) and (92) yields exactly the same expressions for A€ and, An given in Eqs. (18) and (19), Finally, substitution of 6 = x in Eg, (15) yields Eg. (21). ‘The Laplace—Beltrami operator applied on x obeys 2 famous relation expressed by Ax = 2Hn, (60) where the mean curvature H is defined as HL (allxg + 2a xy + @Xqq, 0) 1) 2 (forexample, see [21, Theorem 1, p.71))- Using the requirement H= 0 yields Ax =0, (62) ‘Thus Eq. (18) and Eq, (21) with Ax = 0 are also valid for minimal surfaces. Following the same derivation as given at the end of Section 2.1, we arrive at exactly the same nonlinear system of elliptic partial differential equations as expressed by Eq, (24). Thus an interior grid point distribution on a minimal surface is found by solving Eq, (24). The only difference com. pared to the two-dimensional ease is that now x = (x, y, 2)? instead of x = (x, 9) ‘The same discretization and solution method as described in Section 2.2 can be used to solve the Poisson grid generation system in order to generate grids on minimal surfaces. The only difference compared to the two-dimensional case is that tee (instead of two) linear systems must be solved during. ‘one Picard iteration, Grid orthogonality at boundaries can be obtained in the same way as described in Section 2.3. (One may ask whether itis useful to implement a method to ‘compute grids on minimal surfaces in a 3D mult-block grid ‘generator code, The answer is yes, Minimal surfaces may be used to define the geometry and grid for a block-face of which only the four face-edges are given. Its also possible to apply ‘minimal surface grid generation when a grid must be generated in a block-face with four face-edges lying in a plane. ‘Then the ‘minimal surface is a plane surface bounded by the four edges. ‘The grids in the 2D domains depicted in Figs. 7-17 were ‘generated in this way and are infact grids on minimal surfaces. ‘An example of a grid on a characteristic minimal surface is shown in Fig. 19. This is a so-called square Scherck surface [21]. The initial algebraic grid is shown in Fig. 18. Figure 20, iMlustrates what happens when the prescribed boundary grid point distribution is changed. This figure clearly shows thatthe shape of the minimal surface is independent of the prescribed boundary grid point distribution, (See Section 5.3 for Figs. 18-20) 4, SURFACE GRID GENERATION ON PARAMETRIZED SURFACES 4.1. Derivation of the Grid Generation Equations In this section we develop a method to generate a grid on a parametized surface which is independent ofthe parametriza- tion. A generated grid only depends on the shape ofthe surface and the prescribed boundary grid point distribution atthe four ceges ofthe surface Consider a bounded surface with a prescribed geometrical shape in three dimensional physical space with Cartesian coor dinates x = (x, y, 2)". Assume that ¥ is parametrized by @ tiffereniable one-to-one mapping xP, (63) where isthe unit square in two-dimensional space with Cartesian coordinates w= (4, 0, Define the four eds E, Ey, By, Exo surface # by + w= Oat edge Ey and u = 1 at edge E, + v= Oat edge Ey and v = | at edge Ey ‘Thus (E), £:) and (E;, E2) ane the two pairs of opposite edges of surface & as shown in Fig. 3. Introduce the parameter space as the unit square in a two-dimensional space with Cartesian 48 S.P, SPEKREUSE 1 t y, 4 o q o : o @ @ o) a = = => dl 9 J @ p @ Et ° me ° 7s ° 7 Computational space G Parameter space Fy Paramster space Ray Parametrized surtace S FIG. 3. Transformation fom computations (fm space to pramewzed surface in Catesan (9 3) space coordinates s = (5, 1)". Again require thatthe parameters s and obey: + 5 Oat edge E, and s = 1 at edge By, + 5 the normalized arclengih along edges F and E, + pe Oat edge £ and 1 = 1 at edge Ey, + Fis the normalized arclength along edges £, and E;. Furthermore, require that As = 0 and Ar = 0, where A is the Laplace-Beltrami operator for surfaces. Hence the parameters sand 1 obey (Jas, + Jats), + las, + Jas) = 0, (64) (lait, + Jat). + (Ja, + JaBh)e= 0, (65) Where a re the eontravariant tensor components and J is defined as the determinant of the covariant metric tensor. The contravariant tensor components a are related to the covariant tensor components a, according to Eg. (11). The covariant metric tensor components are defined by Eq. (8). where the two covariant base vectors are now given by a=, a= (66) ‘Thus the coefficients Ja", Jal, and Jo in Eg. (64) and (65) are functions of w and v and Eqs. (64) and (65) are therefore ‘wo uncoupled second-order linear patil differential equations for the functions $ = s(u v) and t= 1(u, 0) Each boundary point of surface has a unique (s ) parame- ter value at 3%, and a unique (u, v) parameter value at 3, ‘Thus each (u, 0) parameter value at AP, has also a unique (5,1) parameter value at 3. Thus the functions s and 1 are prescribed at the boundary of ®,. Hence, Eq, (64), together with the Dirichlet boundary conditions for s, ean be used to compute 5 = s(u, v) and Eq. (65), together with the Dirichlet boundary conditions for z, can be used to compute t = (u,v). Only two linear partial differential equations have tobe solved to define these mappings. These two mappings are compactly ‘writtenas$: Py i> Py. Note thats: Bi Py isa differentiable one-to-one mapping so that the inverse mapping U: Py > Pay also exists. “Thus the composite mapping x: +> ¥, defined as x = x(u(S), also exists and is differentiable and one-to-one. Note that this mapping x 9, ¥ only depends on the shape of surdace {and is independent of the original parametization X Pane F. The mapping x Fr F may thus be considered asa property of surface ¥ and defines anew unique parametsiza- ion of Consider a prescribed boundary grid point distribution at the four edges E), Fs, Es, Ex. Mathematically, the boundary grid point distribution can be defined as a mapping x:3°G + a, ‘where isthe compatational space defined as the it square in a twowdimensional space with Cartesian coordinates & (Gn Because x: 3° r+ AF is prescribed and x: A> AP is defined as described above, it follows that $:0%G > 3, is also defined Tn exactly the same way as forthe two-dimensional case, the mapping s:‘61-» 9. is now defined by the algebraic straight. line transformation defined by Eqs. (5) and (6). The composition cof the mapping 6:“)-» Dg and the mapping x:0P,1-» F defines x:€ r+ J and describes the interior grid’ point distribution fon surface {%. Note that this composite mapping wil also be differentiable and one-to-one. Although it is possible to derive the system of nonlinear clliptic partial differential equations which the composite map- ping x: € +> £ has to obey, we prefer not to do so because it {s much easier to solve the linear partial differential equations defined by Eqs. (64) and (65) to define the mapping s: Pas > aq, instead of interchanging the dependent and independent variables to obtain the nonlinear partial differential equations for the inverse mapping U:, > Po. Thus the mapping s: G+» Bis computed by solving Bas. (64) and (65) and an inversion problem is solved afterwards to compute the inverse mapping u:F, > Pr ‘This is possible due to the fact that the parametrization x: + F is one-to-one so that an inital grid folding free grid jn surface F can be easily generated, Such an iniial grid is ‘obiained by applying the algebraic straight-line algorithm in parameter space Py. This isa different situation compared to ELLIPTIC GRID GENERATION 49 grid generation in 2D domains or minimal surfaces where it is not possible to generate easily an initial grid folding free grid. Details ofthe solution method are described in the next section, 4.2. Diseretization and Solution Method Consider surface F with a prescribed boundary grid point distribution. Assume that there are M ++ | prescribed boundary grid points on edges F; and F, and N + 1 prescribed boundary arid points on edges &, and Es. A boundary conforming grid in the interior of surface £ is now obtained by the following al- gorithm, Step 1. Compute the corresponding boundary grid points in parameter space Py. A boundary grid point xp of surface is related to a unique boundary grid point wp of parameter space @., by the equation x(t) = Xa, where x: F,1> F isthe given parametrization of surface F. In practic, the corresponding parameter values uy of a boundary grid point xy are often already known, Step 2. Compute an initial grid wf by applying the algebraic straight-line algorithm in parameter space ®,,. Thus wi, is com- puted according to Eqs. (36) and (37) with s, and ty replaced by ul, and v). Compute the corresponding initial grid x}, on surface F by xl, ~ x(u). Step 3. Compute the normalized arclength parameters s and ‘atthe boundary points of surface Sf in a way as deseribed in Section 2.2. Solve the two Laplace~Beltrami equations As = 0, 1 = 0, together wit Dirichlet boundary conditions, on the inital grid x, The same solution procedure as described atthe end of Section 2.3 can be used to discretize and solve the Laplace-Beltrami equations. Note that the Laplace—Beltrarai ‘equation As = 0 isin fact the same equation as defined by Ea, (47); the only difference compared to the two-dimensional case is that now x — (x, ¥, 2), instead of x = (4, y7F. Thus in practice, the Laplace—Beltrami equations As = 0 and Ar = 0 are solved directly in computational space. instead of solving Eqs. (64) and (65) on the nonuniform grid wi, in parameter space ®.,, which would be more complicated. Write the solu- tion of the Laplace-Beltrami equations as fs), = (3, |i = 0...N, j= 0. M}. Step 4. Compute in parameter space ®, the grid {8 = (5j.)) [1 = 0u.Ns j= 0. M) by applying the algebraic slraight live algorithm aecording to Eus. (36) and (37). Step 5, Finally the inversion problem must be solved. Con- sider the parameter space, and consider that Step 3 computed mesh (3) as an embedded nonuniform grid This grid may also be considered asa nonoverlapping subdivision of parame- ter space ®, by NX M patches, where each patch has four comer points For a given interior grid point (7, j), the new position x, on sucface of the inal ged is now obtained as follows. Locate the patch in parameter space to which that Step 4 computed value (5, 1,) belongs. Suppose that (sy, ¢,) belongs to patch {p, 4) a8 shown in Fig. 4 ‘Tae local patch parameters @ and fare now defined by the two bilinear equations, 5G = Shall — ANB) + sheryl —B) oil = a) B + Shogo. ty the(l ~ aX ~ B)+ Fhnygeel — B) + thee =) B + thai eteB. ‘The «wo parameters a and i are solved by Newton iteration. Note that 0S a 1 and 0 = B 1 because (5), f,) belongs {0 patch (p, g)- Compute the corresponding position w= (44. 0y) in parameter space Dy by tu = ubgll ~ aX = B) + abergaell ~ 8) Wheel — B+ ahrrgniB = Ubg(l = a0 = B)+ Bf. — 8) + un gt (1 = ©) B + Upsrg eB, and compute % = x(u,), where x:F,, + VP is the given parametrization, The grid {xj |= 0. j= 0... MP 8 the final surface grid ‘The same algorithm can also be used to obtain an orthogonal grid atthe Boundary of surface ‘F. The only changes that have to be made are at Step 3 and Step 4. At Step 3, the Laplace Beltrami equations must then he solved, together with the New mann boundary condition ds/an = 0 along edges Es and E., and 3t/an = ( along edges E, and E, where n is the outward normal direction. At Step 4, the grid s, must be computed using Eqs. (44) and (45) instead of Eqs. (5) and (6). 4.3. Mlustrations We only consider parametrized surfaces which are defined as interpolated surfaces, constructed from a two-dimensional array of coatrol points and passing through these control points. The surface shape of each patch, spanned between four adjacent control points, is delined by a bicubic polynomial. Hermite interpolation is used to connect the surface shapes othe patches ‘smoothly, For details, see [12]. The parametrization ofa surface, defined by the mapping x: 9, ¥, is constructed such that this ‘mapping is continuously differentiable. The parametrization ‘depends on the position of the control points. ‘As an illustration, consider a surface ¥ which is defined by fan irregular control point mesh in a unit square as shown in Fig. 21. Thus the shape of surface $ is a unit square. Figure 22 shows how a uniform grid in P,, is mapped onto surface 9 by the parametrization x9, J. This figure clearly dem- ‘onstrates thatthe parametrization of depends on the position, Of the control points so S.P. SPEKREUSE Theat . c paet' p.det Septet eee nee ae peta” peta) FIG.4. Patch (pin parameter space 9 ‘The grid in Fig. 22 is also te initial surface grid x as defined. at Step 2 of the grid generation algorithm described above. The corresponding initial grid walso defined at Step 2, isa uniform grid in parameter space ®,,. The new grid w,, defined at Step 5, is shown in Fig. 23, Note that the behaviour of the grid in Fig. 23. is opposite to the behaviour of the grid in Fig, 22. The ccomesponding final surface grid x, also defined at Step 5, is, shown in Fig. 24, As expected, this surface grid is uniform. ‘Thus this example clearly demonstrates thatthe surface grid is independent of the parametrization ‘A more practical example is shown in Fig. 25, where the control point mesh is shown of a surface $F which belongs 10 the lower part of a wing near the intersection of a pylon. Figure 26's closeup of the contral point distribution near the leading, edge of the pylon. Figure 27 shows a detail ofthe initial surface ‘grid xf, as defined at Step 2. This grid is badly distributed and. shows the influence of the control point mesh. Figure 28 shows a detail of the final elliptic surface grid x, as defined at Step 5. The elliptic surface grid shows no dependency of the control point mesh and has an excellent interior gr point distribution. (See Section 5.3 for Figs, 21-28.) 5, 3D GRID GENERATION 5.1, Derivation of the 3D Grid Generation Equations ‘The two-dimensional grid generation method described in Section 2 can be extended t0 three dimensions. Consider @ simply connected bounded domain % in three-dimensional space with Cartesian coordinates x = (x, 2)". Suppose that is bounded by six faces Fi. Fa. Fi, Fas Fo, Fu Let (Fis Fas (Fs, Fi), and (Fs, Fe) be the three pairs of opposite faces. Furthermore, consider the twelve edges {E}, / 12} and assume that these edges are related to the six faces as shown in Fig. 5, Consider the computational space as the unit cube in three- dimensional space with Cartesian coordinates & = (én. DY ‘Assume that a mapping x: € +> 09s prescribed which maps the boundary of ‘ one-to-one on the boundary of 2. This ‘mapping defines the boundary grid point distribution, As sume that + = Oat face F and €= 1 at face Fy +7 = Oat face F, and = 1 at face Fi, + f= Oat face F; and f= 1 at face Fy, We wish to construct a mapping x: @ +> % which obeys the boundary conditions and which is a differentiable one-to-one ‘mapping. Furthermore, we require that the interior grid point distribution is a good reflection of the prescribed boundary grid point distribu As in two dimensions, this mapping will be a composition of an algebraic tranformation and an elliptic transformation based on the Laplace equations. The algebraic transformation is differentiable one-to-one mapping from computational space ELLIPTIC GRID GENERATION 31 Computational space C Parameter space P Domain D FIG. & Transformation from computational (2, £) space to domain in Caresion x, 2) spe ‘onto a parameter space @. The parameter space @ is also a unit cube. The elliptic transformation is a differentiable one- to-one mapping from parameter space ® to domain ©. Introduce the parameter space @ as the unit cube in three- dimensional space with Cartesian coordinates 8 = (5; t, w). Require that the parameters s, 1, and u obey + 5 = Dat face Fy and s = 1 at face Fi, at face Fy and 1 = 1 at face Fy, Oat face Fs and um Lat fice Fe + sis the normalized arclength at edges E\, E3, Es, Bs, *+ ris the normalized arclength at edges Es, Fy, Ess Fay + wis the normalized arclength at edges B, Ey, Eu, Ei From the frst three requirements it follows that + 90. m 2) = O and s(1, m &) + 80,0) = Oand 8 1,0) = + (Em 0) = Oand uC Em, 1 The coordinates (5,1) ae defined st all 12 edges of domain ‘The computational coordinates are defined at the complete boundary of @ and thus also atthe 12 edges. Thus each point at the 12 edges of domain ® has @ unique (& », £) coordinare and a unique (s, 1, u) coordinate, Thus each (& 9, £) value at the 12 edges ofthe unit cube in computational space has also A unique (5, t,u) value, Hence, we may conclude: + (60,0) = se(8. E 1 = 548, 160, = 55,2), (6 1D = 5048. +O. m0) = tanh HCL, 0) = tend 1, = tem. tC, 1 = te, + (0.0, 2) = ual ¥), Wl, 0, 45, (0). WC be = He. gg, WO, 1, 0) * ‘The ewelve edge functions sy, , ts, are monotonically increas ing and are defined by the prescribed boundary point distribu- tion at the 12 edges. The algebraic mapping from computational space to parame ter space, 8:1» Bis now defined as = s,(QKl — NU = 0) + se(Or0 =u) 4 55K = Du + 55,6), 67) 1 tell ~ 51 = a) + tego sCl = 0) + te (MK = su te (adn, (68) = un (KI ~ SKI) + ues = 0) 4 te, (OKL = 8) + we, (st (6%) ‘Note that this mapping only depends on the boundary grid point distribution at the 12 edges of domain Equation (67) implies that a grid plane € to the parameter space ® as a bilinear surface: s isa bilinear function of ¢ and w. Similarly, Eq. (68) and Eq. (69) imply that arid planes 1 = const and £ = const are also mapped to the parameter space ® as bilinear surfaces. For a given computa tional coordinate (&, 7, {) the corresponding (5, fu) value found as the intersection point of three bilinear surfaces. For this reason, the system defined by Eqs, (67)-(69) is called the “algebraic bilinear transformation" because of the use of bilinear surfaces in parameter space ®. The algebraic bilinear ‘transformation isthe three-dimensional equivalent of the two- ‘dimensional algebraic straight-line ranformation. Itcan be eas- ily verified that wo bilincar surfaces corresponding to two different Evelues will never intersect in parameter space ‘The same i true fortwo different sor values. This observation const is mapped 52 S.P, SPEKREDSE ndicates that the algebraic transformation is « differentiable ‘one-to-one mapping. The system defined by Eqs. (67)-(69) can also be interpreted as a transfinite interpolation with nonlinear blending fonctions Because x:8°€+-> 35 is prescribed ands: G++ @ is defined by the algebraic bilinear wansformation, it follows that the (, &.0) coordinates are now defined atthe complete boundary of domain , including the interior ofthe six faces Fis «Fo equi that (5 , u) are harmonic functions inthe interior of Bie, Sut Sy t Se tatty + (70) a # iy + = 0. ‘Thus a linear elliptic boundary value problem defines the map- ping $:% ++ 9 It is still an open theoretical question if this, mapping is one-to-one (5]. A reviewer remarked thatthe proof in [4] is false. In this paper it is assumed that s:% r+ is ‘one-to-one and thus thatthe inverse mapping x: + % exists. ‘This inverse mapping obeys a nonlinear system of elliptic partial differential equations. [Note that the mapping x:%° > @ is not independent of the boundary grid point distribution and may thus not be considered a property of domain %. This is because the (5,1, coordi- rates at the interior of the six boundary faces depend on the ‘boundary grid point distribution. It is possible to define the mapping x:9 \-» @ independently from the boundary grid Point distribution by requiring that the (s, 1, «) coordinates ‘obey the Laplace—Beltrami equations in the interior of the six faces of domain @ but then it is no longer possible to use the simple algebraic bilinear transformation defined by Eas. (67)-@. The algebraic transformations $:€ +> # and the elliptic transformation x: +» @ are assumed to be one-to-one map- pings. Then the composite mapping x:‘€ +> ©, defined as xX = x(6(2)) is differentiable and one-to-one. Furthermore, due to the properties of the basic mappings, we may indeed expect, ‘that the interior grid point distribution will be a good reflection of the boundary point distribution. “The composite mapping x:“€ -» ® obeys an elliptic Poisson system with control functions defined by the algebraic mapping 8: +» ®. This three-dimensional elliptic Poisson system, together with the appropriate expressions of the control func- tions is a straightforward extension of the ewo-dimensional system and will be derived in the remainder of this section Define the three covariant base vectors a= x a land the covariant metric tensor components ay = (a.m) F=U2zII=M2I. — (7D "The three contravariant base vectors al, a, and a? are defined according to the rules f= 4,2,3h/= 41232 (73) “The contravariant metric tensor components a= (8a), §={23,7=425, 4) ful ay a2 dy |{ a" (5) “The three contravariant base vectors can be expressed as ‘ata ate ala + aay + aay, 76) a= ala, +a%e; + a%ay Define Jas the determinant of the covariant mettic tensor. Consider an arbitrary function @ = $(& 1, 2). Then & js also defined in domain © and the Laplacian of ¢ can be expressed as b= Laat de Jay + Jaa + ab + Jab, + Jb m) + (la by + Ja}, + Jao) AAs in the two-dimensional case, substitution of = & = ‘n,and = Cinto this equation yields expressions for AE. A, and Ag. Combining these expressions with Eq. (77) gives NG = al bye + 2a, + 2a hy + by, + 20S (7) $+ abe + ME b+ An dby + AL dy Substitute ¢ = (s, fw)? in Eg, (78) and use the requirement thats, and w are harmonic in domain @, ie, As = 0, 41 0, and du = 0. Then the following expressions forthe Laplacian of & mand £ are ag Aq] =a"P, +20"P, + 2a"P ag, @) + aPPy + 20"Py + oP, ELLIPTIC GRID GENERATION 3 where se Sy T=[% tm & en Ue ty Me ‘The 18 coefficients of the six vectors Py, Pix: Pus, Pas Pas. Pas are so-called control functions. Thus the 18 control functions are completely defined and easily computed for a given alge. rai transformation mapping s = 6(£), Different and less useful, ‘expressions ofthese contro! functions can also be found in {1,2} Finally, substitution of ¢ = x in Eg. (78) and using the Tact that Ax = 0, we arrive at Ne + 2a" ny + Dal xy + agg + 20K @) $day + MEK, + Anny + ALK(= 0. “The final form of the Poisson grid generation system can now be derived from this equation by substitution of Eq, (79) and by multiplication with and expressing the contravarant tensor components in the covariant tensor components according to Eg. (15). The result ean be written as Ay + 2a Rey + Dany + ay + 2a Ky + ay + (@!Phy+20"Phy + 20"P|, + Ph + 2a, + aL yxy + (@'Ph +20"Ph, + 2aPPIy + PP; + 20H, + a"PRI + (@Ph +20%P), + 205Ph, + aPl; + 20°F + a PRR, =o, 3) with tatty — al, «way indy ~ Bids and Ae oR A= OA v= BHR — gs) Ag = (Wy Xk das = Oy XO 9 = Oe X0. Equation (83), together with the expressions for the control functions PS given by Eq. (80), forms our 3D grid generation system, Grids are computed by solving this quasi-linear system of elliptic partial differential equations. The discretization of this Poisson system is described in the next section 5.2. Discretization and Solution Method Consider @ rectangular grid of (V+ 1) x OF + 1) x (L + 1) points in computational space ‘€ defined as Bam S= HM maa nyse b= £2 0..Mj=0.Mik= 0.1. =L, (86) ‘Assume that x; is prescribed on the boundary of this grid and consider the computation of Xx, in the interior of the ‘computational grid based on the solution of the Poisson system defined by Faq, (83). The first task isthe computation of the algebraic transforma- tion. The computation of the arclength normalized values at the 12 edges is straightforward and performed in exactly the same way as deseribed in Seetion 2.2.'The arclength normalized variables (5,3 fs Mie) in the interior of the six boundary faces and in the interior ofthe grid are computed according 10 the algebraic bilinear transformation defined by Eqs. (67)~(69) and are thus found by solving simultancously the thrce bilinear algebraic equations: sua = Saal typi = Md) + Sunahiall Wad 6) saul — oe Sotntne Taya = tayo = $00 = thy) + trys Miya) ” 68) tye = Sydege Fast weal] ~ Sajal fej) + Mas Syl fe (89) + Mayall — Sua + MraeiSisatiae for each pair (i j.®) E (0. N; 0. Mi 0 .. L) Ateach grid point (jj. 9, the 18 control functions defined bby Eq, (80), are now easily computed using central difference representations of the derivatives of s, %, and u. ‘What remains isthe iterative solution process ofthe nonlinear elliptic Poisson grid generation system defined by Eq. (83). ‘The diseretization and the applied Picard iteration process is similar as used to solve the 2D Poiston grid generation systems, and the details are therefore omitted, 54 SP. SPEKRENSE FIG. 6 Srucure of 19poin stent ofthe dscretizad 3D Poison gid generation sem, During a Picard iteration, a linear system of equations must be solved for the unknowns xij4, = O ou Nj =O Mi 0... L. This linear system consists of 19-point stencils with Dirichlet boundary conditions. Figure 6 shaws the structure of the 19-point stencils, Such linear systems are solved by another excellent black-box linear-system solver, also developed at C.WiL. by P. M. de Zeeuw. The black-box linear system solver, called THREED, is based on multigrid and Bi-CGSTAB (17, PIG. 8. Complete Oxy Baler grid 18], The linear system solver THREED is called three times to compute the three components je, Yin ABA Ea OF Xie The complete process is repeated until a sufficiently accurate solution has been obtained. The intial start solution x" is ob- tained by algebraic grid gencration. The fina grid is indepen- an AL { fee FIG. Region about NACAODI2 sii sobevies ito four domains. FIG. 9. Domain boundariss near NACAOO!2 aii ELLIPTIC GRID GENERATION 55 FIG, 10, Ord nour NACAOOI? asl dent ofthe initial grid, Moreover, the quality ofthe initial grid is unimportant and severe grid folding of the initial grid allowed. In general, about 10 Picard iterations are enough to ‘obtain a sufficiently accurate solution of the nonlinear elliptic Poisson equations. L. ¥1G.11.- Region aout RAE2802 airfoil subdivided ino Foor domains. FIG. 12 Complee Caype Navier-Stokes grid 5.3. Mlustrations ‘An example of a multi-block Navier-Stokes grid is shown in Figs. 29-34. The mesh isa four-block CO-type grid around ‘an Onera-M6 wing. The total number of grid cells is 256 FIG, 13. Ged nor RAEDHOD sift, 56 S.. SPEKREUSE, FIG. 14. ntl alyebric grid wih severe gi foling wound x complex FUG. 16, Dei of elipic ri at convex prt of the boundary. sical bowndery. 64 x 48. On the wing, the mesh-width in normal direction of —grid-planes intersecting the wing. Figure 31 is a closeup of the the first grid cell is at the leading edge 10° times the local _wing-tip. Figure 32 is a two-dimensional projection of hori- cchord-length and atthe trailing edge 2 10 times the local id-planes intersecting the wing. Figure 33 and Fig, chord-lengu 34 are (wo-dimensional projections of vertical grid-planes at Figure 29 and Fig, 30 are three-dimensional views of vertical stations halfway in spanwise and chordwise direction, ERY RRQ es FIG, 15, Elipue pd with ombogonality a the boundary. FIG, 17. Dewi of ellipse arid at concave pur of the boundary FLLIPTIC GRID GENERATION 7 FIG. 18, Inia lgcrac gi used for mini surface grid pene, FUG. 19, Minimal surtce grid. Surface is x square Schr suse ‘The grid is prid-folding free and the interior grid point dist bution is « good reflection ofthe prescribed boundary grid point distribution atthe block-faces. 6. CONCLUSIONS ‘An elliptic grid generation method is developed which pro- duce excellent grids in the sense of smoothness, grid point distribution, and regularity. The elliptic grid generation method, is based on the composition of an algebraic and elliptic transfor- ‘mation. The elliptic transformation is based on the Laplace equations for domains and on the Laplace-Beltrami equations for surfaces. The composite mappings obey the familiar grid generation systems of Poisson equations with control functions specified by the algebraic transformation. New expressions for, the control functions are derived which only depend on the algebraic transformation and not also on the inverse of this, transformation. The composite mappings are differentiable and (one-to-one for 2D domains and surfaces and probably also for 3D domains It is described how the proposed clliptic grid generation ‘method can be used to generate boundary conforming grids in, 2D domains, 3D domains, and surfaces. Is showa that surface arid generation on minimal surfaces (soap films) is in fact a straightforward extension of grid generation in 2D domains. It is also shown that grid generation on parametrized surfaces, with a prescribed geometrical shape can be performed very easily by only solving two linear elliptic partial differential equations and an inversion problem, A generated surface grid 0n a parametrized surface is independent of the parametrization < AA SSK FIG. 20. Minimal srfce grid. Shope of sufice is independent ofthe boundary grid poin distribution. 58 ‘SP. SPEKREUSE a FIG. 21. Surface defined by an imegelar contol point mesh fn 3 unit n= FIG. 28, New gai in parameter space ‘eral itself and only depends on the shape of the surface and the The described elliptic grid generation method has been imple- prescribed boundary grid point distribution mented into NLR’s muli-block grid generation code ENGRID For 2D domains and surfaces, itis described how the alge- and is extensively used for the generation of boundary conform- braic wansformation can be redefined to obtain grids which are ing Navier~Stokes grids in blocks and block-faces with com= orthogonal at the boundary. plex shapes, FIG. 22. tidal algtic surface grid obtsined from unite grid im FUG.24. Corespoding elliptic surace rid ri isindependent of pra parameter Spe 9 tation. 60. 5. SPEKREUSE FIG.29, Vertical grid plone itersesting the Over ME wing FIG. 31. Wingip with pars of vertical grt planes FIG. 30, Onera.M6 wig with pas of vertical grid planes FIG. 32. Hrzona gi-planesicerecting the wing ELLIPTIC GRID GENERATION 6 FIG. 38, Grid haiay i spannie dieeten FIG. 34, Giid balay in chordie diction, ACKNOWLEDGMENTS. “The author expresses his grade to P.M. de Zecu, esearch sets at the department of Numerical Mathematics ofthe Cong fr Matiematcs end ‘Computer Ssierce (CWI) in Amsterdam, who developed bs the tvo-dama sional black-boe lings mulligid solver MGDOV and the three

You might also like