{JOURNAL oF COMPUTATIONAL PHYSICS 106, 125-138 (1993) Efficient Unstructured Mesh Generation by Means of Delaunay Triangulation and Bowyer-Watson Algorithm S, Renay Dipartimento Igegreria Aerospaciaie,Poleenico di Milan, Vie Golg 40, 20133 Milano, lay Received July 22.1991 This work i dovoted to the description of an efficent unstructured ‘mesh gensttion metned entry bases onthe Delaunay trangultion, ‘The distinctive characteristic of the proposed method is thet point positions and connections are computed simultaneously. This result {is achieved by teking adventage of the sequential way in which the Bowyer-Watson algorithm computes the Delaunay teangulation. “Two mathods aye proposed which have great geometical faxbiity, in that thoy allow us to toat domsine of arbitrary shepe and topology fand 10 genarete arbivatiy roruniform meshes. The mathods are ‘computationally efiient and ate applicable both in two ard tee mensions. © 199 Aenen Ps ne 1. INTRODUCTION When looking for the numerical solution of fluid dynamic ‘equations defined in the continuum such as the full poten- lial, Euler, or Navier-Stokes equations, itis necessary to select a finite number of points in the domain and to connect them so as to define a grid, on which a discrete version of the original partial dilferential problem is constructed. There are two main classes of grids, which differ in the way in which the mesh points are connected to each other. Ifthe internal points are connected to their neighbours in a way independent of their position, the mesh is called struc- tured. When the pattern of the connections varies from point to point, the mesh is called unstructured. In the struc- tured ease, the connectivity of the grid is implicitly taken into account by storing the point data into the elements of, ‘a matrix. On the contrary, the connectivity of unstructured grids must be explicitly described by an appropriate data steucture, thus making the solution algorithms on unsteuc- tured grids more expensive than those on structuzed grids. However, the greater geometrical flexibility offered by Uunstructured grids can be crucial when dealing with domains of complex geometry or when the mesh has to be adapted to complicated features of the flow field. The inereasing attention that such problems have received in recent years has therefore made unstructured grids of common use in computational fluid dynamics. A great effort has consequently been directed toward the development of unstructured mesh generation methods. Among the available techniques, those based on the Delsunay 1 angulation are particularly suited to an adaptive solution strategy since the Delaunay construction allows the addi- ‘ion of new points to an existing triangulation without the need for remeshing the whole domain, In earlier applications of Delaunay triangulation to unstructured grid generation, the Delaunay construction, hhas been used to connect a preassigned distribution of points, so as to obtain a triangulation satisfying certain ‘geometrical properties. In the most common mesh genera- tion methods of this class, in fact, the spatial distribution of the mesh points is determined by means of some appropriate technique in a step preceding that of estab- lishing their connections (see, eg, Baker [1], Jameson, Baker, and Weatherill [41, and Weatherill (10, In this work we describe a completely different approach, which takes full advantage of the sequential way in which the Delaunay triangulation is constructed. The distinctive ‘characteristic of the new method is that point positions and connections are computed simultaneously. Two different mesh generation algorithms will be considered in the present work, the only difference between them being the way in which points are positioned in the domain. Both algorithms are of the greatest logical simplicity allowing ‘the treatment of domains of arbitrary shape and a complete control of the size of the criangles over the computational domain, Moreover, the algorithms are computationaily, very efficent and are applicable to two and three- dimensional problems. The structure of the paper is organized as follows. In Section 2 we recall the well-known Delaunay triangulation and the associated Dirichlet tessellation. Section 3 describes. how Bowyer-Watson algorithm, which is widely employed to construct the Delaunay triangulation, can be con- veniently used in the context of mesh generation. In Section we introduce an initial triangulation of the 1o21.99199$590, \Conyiht © 1989 by Aces Pro ne, Alig of epodacin a form eed126 boundary points which represents the starting point of the ‘two mesh generation methods to be described in detail in Sections § and 6, respectively. In Section 7 the potentialities offered by the proposed algorithms are investigated by applying them to some representative test cases. Section 8 is devoted to the conclusions 2. DIRICHLET TESSELLATION AND DELAUNAY TRIANGULATION A systematic way for generating a triangulation of aset of points can be obtained by considering a geometrical con- struction first introduced by Dirichlet in 1850. Consider an arbitrary set of points P,, i= 1,....N. For any point P, we define a convex polygon ¥; characterized by the property that every point of ¥; is nearer to P, than to any other P ‘These convex polygons are called Voronoi regions and theit union is called Dirichlet tesseltaron, The boundary between two Voronoi regions ¥7 and ¥ facing each other lies midway points P, and P, and is there- fore a segment of the axis of the line which joins P; with P,, By connecting only those points whose Voronoi regions have a common edge, a triangulation known as Delaunay triangulation is obtained. An example of a Dirichlet tessellation with the associated Delaunay triangulation of the convex hull of a small set of points is shown in Fig. 1 ‘This geometrical construction can be casily extended to more dimensions. Inthe three-dimensional space, for exam- ple, the boundary between two facing Voronai regions is a plane polygon, and, by conneeting only the points whose FIG. 1 (continuous) of the convex hull small et of points Dirichlet teseltion (dashed) and Delaunay triangulation S. REBAY Voronai polyhedra have a polygon in common, a sct of tetrahedra is obtained. Tt should be noted that Delaunay triangulations and Dirichlet tessellations can be considered the geometical dual of each other, in the sense that for every triangle J there exists a vertex C; of the tessellation, and conversely, for every polygon ¥; there exists a vertex P, of the triangula tion (see Fig. 1).In addition, for every edge ofthe triangula- tion there exists a corresponding segment of the Dirichlet tessellation, In the following, vertices and segments of the Dirichlet tesseliation will be also called Voronoi vertices and Voronoi segments, respectively. 3. BOWYER-WATSON ALGORITHM AND MESH GENERATION ‘The Delaunay triangulation recalled in the preceding section is a construction characterized by certain specific geometrical properties, Some of these properties are of a local nature and allow us to devise algorithms which com- pote the Delaunay triangulation of an arbitrary set of points only by means of a sequence of purely local operations, Among the methods belonging to this class, the Bowyer [2] andjor Watson [9] algorithm is very convenient in a ‘mesh generation procedure. The method is based on the so- called circumcircle property which guarantees that no point of a Delaunay triangulation can lie within the circle cit- cumscribed to any triangle. The Bowyer-Watson algorithm is essentially a “reconnection” method, since it computes how an existing Delaunay triangulation is to be modified because of the insertion of a new point. A simplc example showing the sequence of operations neoded to insert a new point into an existing triangulation and to generate a new triangulation is given in Fig. 2. The algorithm removes from the existing grid all the triangles which violate the circum: circle property because of the insertion of the new point. It can be shown that (1) all these triangles are always con- ‘iguous, thus forming a connected cavity surrounding the newly inserted point, and that (2) by joining the vertices of this cavity with the internal new point, a Delaunay tri- angulation is always obtained [2, 9]. As a consequence, the Delaunay triangulation of an arbitrary set of points can be constructed in a purely sequential manner starting from FIG. 2. Reconnection of an existing grid around 2 newly inserted point by means of the Bowyer-Watson algorithmUNSTRUCTURED MESH GENERATION very simple initial Delaunay triangulation enclosing. all points to be triangulated (for example, that formed by ‘wo adjacent triangles) and adding one point after another until all points have been considered. The algorithm effiency depends on how quickly the search for the triangles to be deleted at each point insertion is performed and ths is made much easier by the knowiedge of the neighbouring triangles to each triangle. Infact, since all the triangles to be deleted are always contiguous, a tree search among neighbouring triangles can be used to find all the other triangles to be deleted after the fist one. In the typical case, the number of triangles to be deleted at cach point insertion does not depend on the numberof al existing triangles. As a consequence, if the information pertaining to the neighbouring triangles is available and an O(log, N) multidimensional search for the firs triangle to be deleted is employed, the algorithm can compute the Delaunay triangulation of a set of 1Y points in O(N log.) operations. In special cases, however, the number of triangles to be deleted at each point insertion can be very large. In the worst possible situation, when all existing triangles have to be deleted at each point insertion, the operation count ofthe Bowyer-Watson algorithm degrades t0 ON"). It has, however, Been shown that a proper order- ing ofthe input points can cure this problem. Unstructured mesh generation techniques which use the Bowyer-Watson algorithm to compute the connections ofa preassigned distribution of points have been sucossfully employed for both two and three-dimensional problems [1 4, 10) ‘The sequential nature of the Bowyer-Watson algorithm can be further exploited to compute mesh point positions and connections simultaneously if, starting from an available triangulation with points, the position of the (N=I)th point is chosen according to some suitable ‘geometrical criterion which depends on the existing tri angulation. Within such an approach, itis very important t0 choose carefully the position where £0 insert the new point in the existing mesh, to avoid a nonsmooth point distribu- tion. Even if the Delaunay construction guarantees in advance that the points will be connected so as to produce triangles as equilateral as possible, a poor point distribution ‘can eventually lead to an unsatisfactory triangulation. In this respec, the vertices and the segments ofthe Dirichlet tessellation are promising locations for placing a new point since they represent a geometrical locus which falls, by construction, midway between the triangulation points. In both cases, the new point is inserted in a position which avoids the formation of edges that are too short so tha it is unlikely for the new triangles 0 have a very elongated shape. Two different mesh generation methods in which the new point is inserted at a Voronoi vertex or on a segment of| the Dirichlet tessellation will be considered in Sections $ and 6, respectively 127 In the first _method, which follows closely an idea originally introduced inthe pioneering work of Holmes and Snyder [5], the new point is introduced in the existing triangulation at the Voronoi vertex (centre of the ci- ccumseribed circle) corresponding to the “worst” triangle, defined as the one having the most diferent circumscribed circle radius in relation to that required forthe final grid. In this way, since the worst triangle is obviously eliminated from the triangulation, we improve the quality ofthe grid at every new point insertion. The mesh generation process is terminated when all “bad” triangles have been eliminated, leaving a grid formed only by suitable triangles Tn the second method here described, the new point is instead inserted in an attempt to generate one or possibly several new triangles having from the very beginning the size presribed for the final grid. This result is aevomplished by placing the new point along a segment of the Dirichlet {essellation, in a position carefully chosen to guarantee that the newly generaied triangles have the sought for size. As ‘willbe shown in the examples reported in Section 7, it turns ‘out that the grids generated by means of the second method are characterized by triangles of a generally better shape in comparison with those obtained by means of the fist method. This is especially true for triangles located near the boundaries ofthe domain, a region where @ regular spatial discretization is of crucial importance to perform accurate computations. Both mesh generation algorithms offer complete control of the triangle sizes and allow an implementation with ‘operation count O(N log N). This result is accomplished by means of a proper ordering of the triangle data Sophisticated multidimensional efficient search routines are therefore not required by the proposed methods 4 INITIAL TRIANGULATION OF ‘THE BOUNDARY POINTS Both mesh generation methods described in the paper choose the position ofthe (N+ 1)th mesh point on the basis of the grid connecting the already existing N points. It is therefore necessary to provide an inital grid (o start the resh generation process. To this aim, all the boundary points are triangulated by means of the Bowyer-Watson algorithm. The initial boundary triangulation must be body conform ing: ie, all boundary edges must be included in the tric angulation. Since the Delaunay triangulation of a given set of points is a unique construction, there is no guarantee that, for an arbitrary distribution of boundary points, the resulting triangulation will be boundary conforming However, by performing a check on the initial triangul tion, all possible missing boundary edges can be detected. It-can be proved that, by repeated insertions of new mesh128 points at the midpoint of the missing boundary edges, a boundary conforming triangulation is always obtained; see, for instance, [11]. This intial check is a relatively cheap procedure, since the number of boundary points inthe mesh is generally small compared to the total umber of mesh points to be generated—asymptotically O/N) in two dimensions. Two different types of triangles can be identified at this stage—the triangles internal to the domain to be meshed and the triangles external to it. Its important to keep track ‘of both internal and external triangles. Internal triangles must be considered in view of the fact that the new mesh points must be generated inside the domain, Moreover, to prevent the elimination of boundary edges during mesh generation, itis suficient to check that no external triangle is ever deleted from the triangulation. In this way, the boundary conforming property possessed by the initial triangulation remains invariant during the entire mesh generation procedure. Different strategies can be adopted to deal with a point that would cause an extemal triangle 0 be deleted. For ‘example, an attempt could be made to find a different loca- tion for the new point, Alternatively, the boundary could be locally refined. In this work, a very simple strategy turned ‘out to be effective—points which would cause an external triangle to be deleted are simply discarded. In principle, this, procedure could lead to a grid which locally does not satisfy the length scale criterion, Numerical investigation seems, however, to indicate that this simple boundary treatment, when used in conjunction with the point generation methods described in the following sections, leads in practice to a faily regular point distribution near to the boundaries In the following sections we will describe in detail the wo ‘aforementioned strategies for defining the position of the new mesh point inside the domain on the basis of an existing set of internal triangles, ‘5. VORONOI-VERTEX POINT INSERTION METHOD As outlined in Section 3, a promising location to place a new point is the centre of the circle circumscribed to a triangle. With this choice, the new point is in fact inserted. in a position where, by the circumcircle property, the grid, is coarsest. ‘The choice of the “insertion triangle,” namely the triangle where to insert the new point, can be done according to dif ferent criteria. For example, in the grid generation method, ‘of Holmes and Snyder, the new points are introduced to eliminate in turn triangles with large areas and triangles having bad “aspect ratios,” the fatter being defined as the ratio between of the radii the inscribed and circumscribed 5. REBAY circles. In this way, all triangles that are large or far from equilateral are eliminated from the triangulation, which therefore (ends to cause clustering according to the geometrical discretization of the boundary points. A disad- vantage of (he method is that the coarseness of the tri- angulation over the entire domain is contrelled only by the graduation of the boundary points, so that the dimensions of the internal triangles far from the boundaries may turn ‘out to be inappropriate for the considered problem. ‘A simple insertion criterion is suggested here which allows control of the grid coarseness over the entire domain. We begin by defining a function f(x) which prescribes the value of a characteristic dimension of the triangles, say the radius of the circumscribed circle, as a function of the triangle position, We define for each triangle % with circumscribed circle radius p, the ratio fade TF. where, indicates the position of the centre of the circle cir- ‘cumscribed to the triangle J,. By inserting the new mesh point at the centre of the citeumscribed circle having the largest value of a, it is possible to reach eventually a mesh n which max, a, I return to step 6, else stop. 6 VORONOLSEGMENT POINT INSERTION METHOD. In this section, the second mesh generation method relying on the Bowyer-Watson slgorithm is described. The basic difference between this method and that presented in the previous section is that the new mesh point is taken on 2 segment of the Dirichlet tessllation, instead of at a Voronoi vertex. Furthermore and more important, instead of eliminating the worst triangle as in the frst method, the new point is here inserted in an attempt to generate immediately one or possibly more new triangles having the required size feom the very beginning. It turns out that the quality of the meshes obtained by means of the second method appears to be generally superior to that of the rmeshes provided by the previous one. 6.1, Point Insertion Strategy ‘The starting point of the mesh generation is an initial boundary triangulation obtained by connecting all bound- ary points as described in Section 4. Since the now point is introduced on a segment of the Dirichlet tesetlation and External ue FIG. 3. Hievarchialsructure of wiange types Accepted (Done) Active Nonaccepted Waiting 9 there is a one-to-one correspondence between the segments of the tessellation and the edges of the triangulation, the position of the new point will be established in a way wach depends on suitable couples of ncighbouring triangles and their common edge. To this aim, i is convenient (0 sub- divide al tiangles present in the initial boundary triangula- tion into several types as explained in the following In the first place, it is necessary to distinguish between triangles external and triangles internal to the domain, in the same way as described in Section 4, Then, the internal triangles must be further subdivided into two types— triangles already having and triangles not yet having the size requested by their position in the domain. These two types of internal triangles will be referred to as aecepted and rronaccepted triangles, respectively. In most common cases, the internal triangles of the intial boundary triangulation are all nonaccepted triangles. Since points are inserted by considering nonaccepted triangies with a neighbour already having the required size, it is necessary to introduce @ further subdivision of the nonaccepted triangles—namely, active and waiting triangles. An active triangle is defined as a nonaccepted triangle having at least one accepted or one external triangle among its neighbours. All the nonaccepted triangles not active are considered waiting. A waiting triangle is therefore a triangle surrounded only by active or waiting triangles. For the sake of clarity, the hierarchical structure describing all the types of triangles involved in the mesh generation process is depicted in Fig. 3. The new point is always inserted by considering the Voronoi segment associated with « triangulation edge shared by the active triangle with largest circumscribed circle radius and one of its accepted and/or external neighbours. In this way, by virtue of the circumeircle property of the Delaunay triangulation, new points are automatically positioned in regions of the domain, where the mesh is coarsest. If more than one accepied and/or external neighbour exists, the segment associated with the shortest triangulation edge is chosen. This choice is purely empirical, and other possibilities might give good results as ‘well. Al test cases reported in Section 7 have, however, been obtained according to the above eriterion, FIG. 4. Accepted or externa eight), active angle (igh) and the new point130 6.2. Positioning of the New Point The position of the new point along the Voronoi segment is chosen in an attempt to generate a new triangle having the size prescribed at its location, With reference to Fig. 4, let * be the segment of the Dirichlet (essellation associated with the triangulation edge PQ in common with the two aforementioned neighbouring triangles, the left being the accepted or external triangle and the right being the active one. The new point Xs placed on on the side of the active triangle in a position chosen so that the triangle formed by connecting with P and Q has the size prescribed for the final grid, The position of X is computed as follows. Let Cy be the centre of the circle circumscribed to the active triangle (the right endpoint of segment in Fig. 4}, and jet -M be the point in common to edge PQ and segment 7 (midpoint of PQ). Furthermore, let Pa Lu) be the value of the circumscribed radius prescribed for the final (ciangulation at point x, (obtained, as usual, by means of a background grid). Ideally, we would like 10 locate the new point ¥ on segment / atthe intersection of F with the circle passing through points P and @ and hhaving circumscribed radius equal to py. However, it might hhappen that the prescribed value py is not appropriate, since any citcle through P and Q has a radius p > 5|PQ! Furthermore, a real intersection point ¥ exists only for circles having @ radius smaller than that ofthe circle passing through P, Q, and C,; ie rag) 2 an where p=$PQ| and g=|C4M| (see Fig. 4), For these reasons, a limited value of py is defined as ie e+e ‘The position of the new point X is therefore defined by the relationship Xyoxutde, where the distance d=|XM) of X from M and the unit veetor e are d= put Ju? Ker Xcue 5. REBAY ‘Now, itis important to note that itis not assured that, by inserting X/as explained above and then retriangulating, the triangle XPQ will be actually produced by the Bowyer ‘Watson algorithm. However, point ¥ is positioned by construction inside the cirele circumscribed to the active triangle, which is therefore always deleted from the t angulation. As a consequence, we are guaranteed that the new triangulation will contain two edges PX and QX, having by construction a length consistent with that required for the final triangulation, As a matter of fact, itis inessential whether the triangle XPQ is generated or not, since We are guaranteed that @ Delaunay triangulation always connects points $0 as to produce triangles as equi- lateral as possible. Asa final remark, note also that, when 7 is too short ‘with respect to |PQI, the new point will be inserted at the ‘Voronoi vertex of the active triangle. In this special situa- tion, the point location is therefore coincident with that generated by the method described in Section 5. 63. Assignment of the New Triangles As anticipated in Section 3, at each point insertion the Bowyer-Watson algorithm generates a new triangulation ‘which differs from the previous one only locally around the newly inserted point, For this reason, and by virtue of the definitions employed to classify the triangles, a new grid in which all triangles are classified consistently with the defini tions given in Section 6.1 can be constructed by considering ‘only (1) triangles having the mew point as one of their ver tices (new triangles) and (2) triangles adjacent to the new ‘ones (see Fig. 5). At each point insertion, these triangles FIG.S ‘The nen triangles generated by the Bowyer-Watson sigorthon because of insertion of point X (continuous) and. ther seighbours (dashed)UNSTRUCTURED MESH GENERATION have to be reassigned to the different types described above. according to their size and relative position. As a conse- quence, new accepted triangles will replace old nonaccepted cones starting from the boundary and moving to the interior of the domain, in a fashion similar to that characterizing the triangle propagation occurring in methods in which new points are gencrated along a propagation front, such as that originally proposed in [6] for two-dimensional applica- tions and subsequently extended to three dimensions in (7. There is no need to subdivide the new triangles into exter- nal and internal, since, by coastruction, no external triangle is ever deleted by the Bowyer-Watson algorithm and all the new triangles are therefore internal, However, it is necessary to subdivide the new triangles into accepted or nonaccepted and (o further subdivide the new nonaccepted triangles and their nonaccepted neighbours into active or waiting. The new triangles J; are classified as accepted or non- accepted depending on their circumeiscle radii p,. For every new triangle, consider the midpoint M, of the edge opposite {0 the new point (see Fig. 5) and evaluate m= Sex; the ratio the triangle J; is considered accepted, otherwise it i con- sidered nonaccepted. The value of 3 is purely empirical Good results have been obtained choosing 5 = 1. The new nonaccepted triangles are then subdivided into active and waiting in a way that is analogous to that described in Section 6.1 Since no external triangle is ever deleted, there is no need to subdivide triangles adjacent to the new ones into external and internal. In fact they will simply retain the external or internal attribute they had prior to the new point insertion, Moreover, since the insertion of the new point does not affect the geometry of the triangles adjacent to the new ones, itis not necessary to subdivide these triangles into accepted or nonaccepted. As a consequence, also the accepted or nonaccepted attribute will be retained. It is, however, necessary to perform a subdivision of the nonaccepted triangles adjacent to the new ones into active or waiting, according to the same rules used for the new triangles. The imesh generation method stops when there are no more active triangles left. ‘An example showing the triangle progression during the ‘mesh generation process is given in Figs. 6 and 7 for the simple case of an elliptical domain, where a uniform triangle distribution has been required, ‘As explained above, the new point is inserted on a BL Voronoi segment relative to the active triangle having the largest circumeircle radius, and this implies that computa- tional efficiency can be achieved by sorting active triangles, according to the values of their radii. The second mesh generation algorithm can therefore be summarized as follows: 1, Triangulate ali the boundary points by means of the Bowyer-Watson algorithm. 2. Check if the triangulation is body conforming and recover all missing boundary edges. 3. Divide all triangles into internal and external. Divide internal triangles into accepted and nonaccepted on the a sis i: ~~ MS 6 Two sceesive steps of the mesh generation of an elliptical “p" indicates acepted (done) tangles, “a” active triangles snd “1 walking tangles The frst step shows the intial boundary ‘angulation,