4 INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, VOL. 15, 1813-1822 (1980) A SCHEME FOR THE AUTOMATIC GENERATION OF TRIANGULAR FINITE ELEMENTS EDWARD A. SADEKt Faculty of Engineering, Cairo University, Egypt (A.RE.) SUMMARY ‘This paper describes a method for the automatic triangulation of arbitrary multilateral plane domains. In addition, the method can be used in connection with that suggested by Zienkiewicz and Phillips‘ for the subdivision of curved-boundary domains. The method can be described as general fully automatic and computer oriented. A Fortran computer program has been prepared by the author. Output can be of interactive, graphical or alphanumerical form. The program has been applied to a number of cases. The resulting triangulations are always satisfactory, even when vigorous changes of mesh sizes are encountered. INTRODUCTION One of the main steps of any finite element procedure is the subdivision of the problem under consideration into a finite number of non-separated, non-overlapping elements. This is normally achieved through the following steps: 1. Subdivision of the problem into a number of main subregions. The choice of these subregions may be based on geometrical, metallurgical as well as other considerations. 2. Having defined the different subregions, the next step is to decide either the number of elements along each side of the subregions or the average element size required throughout each subregion. The size of a triangular element is usually defined either as its area or its side length (assuming equilateral form). It is also necessary to decide the appropriate mesh grading to achieve the required accuracy of idealization. ‘These two steps are, to a large extent, a matter of opinion and itis unlikely that even experts will come to the same conclusions in all cases. By the end of this phase, a number of subregions with distributed nodes around their boundaries are chosen. The next step is to subdivide each of these subregions into a number of triangular elements. The method used for this subdivision must satisfy, as much as possible, the following requirements: 1, Since the best form of a triangular finite element is thé equilateral, the resulting elements should not be far from this form (well-conditioned elements). 2. The method should impose no constraints on the liberty of the user in choosing what he thinks is the best distribution of nodes around the boundaries of the different subregions. This, of course, excludes the condition of consistent nodes along common boundaries of adjacent subregions. 3, Minimum cost with respect to both human effort and computation time. + Assistant Professor. Currently: Visiting Research Fellow, Department of Civil end Structural Engineering, UMIST. 0029-5981 /80/1215-1813$01.00 Received 11 January 1979 © 1980 by John Wiley & Sons, Ltd. Revised 16 November 19791814 E. A, SADEK However, the computation time required for subdivision into elements is normally a small fraction of that required to perform the analysis. Thus the last requirement is mainly minimum human effort. Many attempts, using different algorithms, have been made in this direction, such as the methods suggested by Zienkiewicz and Phillips," Yeung and Hsu,® Cavendish” and Bykat." The suggested method, however, is @ new attempt towards satisfactory fulfilment of the above-mentioned requirements. TD Area to be subdivided ZA Area cut out before the next cycle of subdivision * Figures represent node- levels.‘TRIANGULAR FINITE ELEMENTS 1815 BASIS OF THE SCHEME. Given a certain domain with a number of nodes around its boundary, Figure 1(a), we start by determining the most well-conditioned elements which can be formed at each corner of the domain, Figure 1(b). In our discussic, a corner is a boundary point at which the boundary angle 4, Figure 3(a), is not equal to 180°, Having finished this step, the elements formed are considered to be cut out from the original area and the same concept is applied to the remaining area and so on. The nodes around the original boundary of the domain are considered to be of level equals unity. Subdivision at a corner node of level equals unity generates nodes of level equals two and so on. Figures 1(b)-(e) show some steps of subdivision of the domain shown in Figure 1(a) and the levels of the different nodes, The subdivision of a domain is performed in successive stages. In each stage, a continuous boundary layer is cut out. This is shown diagrammatically in Figure 2. Starting with the nodes around the original boundary, triangular elements are generated and a new set of nodes of level © nodes with level = 1 A ox Sx ow 22 3 Contour(t1) Contour If) Figure 2, Node levels equals two are obtained. The first stage is known to be complete when all the nodes bounding the area to be subdivided become of level equals two. The area to be subdivided in the second stage is that bounded by contour (Il), around which nodes of level equals two are distributed. Again, starting with these riodes, a new set of nodes of level equals three are generated and so on. In the case ofa domain with’a number of inside holes, the nodes around the boundary of the domain and those around the boundaries of the holes are all considered of level equals unity. Subdivision is performed once to cut a boundary layer round the outer boundary of the domain and once to cut a boundary layer around the boundary of each hole. “In order to show how a new node is generated, consider the case shown in Figure 3(a). The niimbet of triangular elements that can be gerierated at a node ‘i'.with an internal boundary angle 7is'takén’as the nearest integer of $i/(n/3). The parameter $,/(n/3) is used to ensure1816 EA. SADEK : (a) (d) Generated elements: Figure 3. Generation of a new node ° that a suitable number of triangles are generated so that triangles are as equilateral as possible. Let us assume that the angle os at node 3 is such that two triangular elements can be generated. ‘However, in order to produce two equally well-conditioned elements, the ratio J2/as should be ‘equal to the ratio /a3/13. In other words, the length /23 should be taken as V(Ia.!s). The length 12s taken along a line bisecting the angle $3 determines the position Ay, Figure 3(b). ‘This is the position of the new generated node if only the situation at node 3 is taken into consideration. To allow for the fact that the position of the new generated node will.also affect the shape of theTRIANGULAR FINITE ELEMENTS 1817 elements formed at adjacent nodes, namely nodes 2 and 4, the same procedure is repeated at these nodes. The angle $2 at node 2 is 180° and thus three equal divisions each of 60° can be taken. The condition that makes the three triangles shown in Figure 3(c) equally well condi- tioned is, h/la= ha/ly = ba/ le or fa=N(G.b) and I= 20.8) ‘The length /21 determines another position A>. This position is the position of the new generated node if only the situation at node 2 is taken into consideration. The situation at node 4 indicates a third position A3, Figure 3(4), with the lengths by PVs.) Now three different positions Aj, A> and A; are required for the new generated node A. Position A, best fits the situation at node 3, while Positions Az and A; best fit the situations at nodes 2 and 4, respectively. In order to choose a unique position for the generated node that satisfies as much as possible all these requirements simultaneously, the position of the generated node is taken at the centre of gravity of three equal loads located at the three positions A;, A> and As. ‘This new position will naturally be deviated from either of these three positions. However, in order to allow for the fact that the shape of small elements will be more distorted by this deviation than larger ones, the loads at the positions Ay, A> and As are taken in inverse Proportions to the size of the resulting elements in every case, i.e. (Ble) and as wicct/i3s, wee fl, and wal /Iq as illustrated in Figure 3(¢). In the program used, the values of ws, ws and w, are taken as wi=w/ls, wa=w/ly and: w3=w/lb4 where walis +l +136 ‘The factor ‘n’ can be chosen arbitrarily and fed to the computer as a part of the input data. A Feasonable range of n has been found to be between 05 and 4. Figure 4 illustrates the effect of different values of m ranging between zero and 4. The boundary angle in this example is taken 90° and the length ratio of the elements is 2:1. When n = 0, the distortion of the smaller element is pronounced. With increasing the value of n, the distortion of the smaller element decreases and that of the larger one increases. Opposite to that for n = 0, n = 4 causes more distortion for the larger element than that of the smaller one. This means that by increasing the value of n, ‘more emphasis is given to the shape of the smaller element and vice versa. However, it has been found that a value of m equals unity. fits well with all the problems handled, including the examples given in this paper. ; Figure 5 illustrates the case when two or more corner nodes lie adjacent to each other (nodes 2 and 3 in Figure 5). Since both nodes 2-and 3 are corner nodes, subdivision can be made either according to nodes 1,2 and 3 (node A’), or according to nodes 2,3 and 4 (node A”). In such cases, a reasonable subdivision is obtained by replacing both nodes A’ and A” by one node A located in between A’ and A’,1818 B.A SADEK * Figure 4. Effect of different valies of “1 ‘Also, situations such as that shown in Figure 6 happen quite often before or during the = _ procedure of subdivision. In Figures 6(a) and 6(b), the areas 456 and 234, respectively, could be acceptable element shapes without any further subdivision, Thus check must be made, and ifit js found that the requirements of the user regarding element size and shape are satisfied, these = areas are considered to be cut out from the area to be subdivided before the next cycle of subdivision starts. Since the program in its present form is fully automatic, the user has to decide the limits of size and shape he accepts for the resulting elements. : ‘However, some users think that a semi-automatic scheme is more attractive than a fully automatic one. In this case another version of the program can'be used with the decision left to the user at any stage he likes.“ 3 “The subdivision process is known to be complete when the number of nodes around the boundary of the remaining area reaches three. This means that the remaining area is one single triangular element and hence no more subdivision is required. Before reaching this state, we may reach a state in which the number of nodes is four. In this case, the remaining area isa quadrilateral. It can be divided into two triangles by connecting one of its diagonals. However, =-—_ Generated elements: Figure 5, Average position of a generated node‘TRIANGULAR FINITE ELEMENTS 1819 Domain, Boundary (a) (b) Figure 6. Areas cut out this should not be done randomly and a choice of the diagonal to be connected is necessary. In Reference 4, it has been suggested that the division of 2 quadrilateral into two triangles is to be chosen according to the shorter diagonal for better shape conditioning. This fits quite well in most cases but it does not fit so well in some other cases, such as those shown in Figures 7(a) and 7(b). Another method for performing such division is as follows. To divide a quadrilateral area into two triangular areas, two requirements should be taken into considerati 1. The resulting elements are those of the better shape conditioning. This is achieved by choosing the shorter diagonal division, 2. The areas of the resulting two triangles are the nearest to each other. ‘These requirements are sometimés contradictory. To explain how a subdivision is selected, consider the quadrilateral area shown in Figure 7(c). Figures 7(d) and 7(e) show the two possible subdivisions. For the subdivision shown in Figure 7(d), a factor R given by a, Bilbo b as/az is determined. Similarly, the factor R for the subdivision shown in Figure 7(e) is given by R 7 by ar/ar Raa bibs - . ‘The first term of R gives preference to the subdivision which corresponds to the shorter diagonal. However, the second term gives preference to the subdivision which produces two Nearer areas. Hence the subdivision with larger R is considered to be the one which satisfies one of the requirements and deviates lesser from the second one. The subdivision according to the ‘suggested method is shown by solid lines in Figures 7(a) and 7(b) (longer diagonals) and also that shown in Figure 7(4) (shorter diagonal). It is also possible to put more emphasis on one of these requirements. This is achieved by multiplying one or both terms of R by:arbitrary factors. For example, the subdivision with the shorter diagonal as the only criterion is obtained by mulplying the first term by unity and the second term by zero and so on. * ? s1820 E. A. SADEK wl. ' (a) P tb) PX siageated | Qe Figure 7. Subdivision of a quadrilateral INPUT DATA ‘The input data of a domain may be given in one of the following forms: Form 1 (a) Number of sides of the domain (which is equal to the number of its corners) and co-ordinates of each corner; (b) number of elements along each side of the domain; (c) the gradient along each side, “The gradient is defined as the ratio of length of the last clement on aside of a domain to length of the first element on the same side and'a quadratic variation is assumed. Thus gradient equals unity means constant element size along a side, while gradient larger than unity means increasing element size and so on. ' 1 Form 2 4 ‘The co-ordinates of the nodes all around the boundary of the domain. ‘The direction of rotation around the boundary of a domain must be fixed. In the program used, this direction is always taken clockwise.‘TRIANGULAR FINITE ELEMENTS 1821 EXAMPLES Figures 8 and 9 depict examples of some meshes produced by the method described. The meshes shown in Figure 8 were produced on a Cyber 72 computer with a Textronix 4010 interactive storage-tube graphics device, while those of Figure 9 were produced on an ICL computer. Each of these examples represents one single subregion. In Figure 8 the case of a square plate with a small central hole is depicted. Because of symmetry only one-eighth of the plate was considered. Figure 9 depicts the case of a domain with an inside hole. Figure 8. A square plate with a central hole ‘The method was also:applied to many other examples (not included in this paper). The examples were chosen to cover the wide variety of domain shapes ranging between triangular to multilateral domains, with and without inside holes. The examples also cover the different cases of variable as well as constant element size. Although vigorous changes of element size are not recommended, they have been included to check the capability of the method to handle such situations. CONCLUDING REMARKS ‘The scheme described in this paper is simple, logical and straightforward, It does not include any special algorithm that may fit in some cases and fail in others. By the use of this scheme a general multilateral subregion with arbitrary boundary nodes distribution can be divided into reason- ably well-shaped triangular (and/or quadrilateral) elements. Unlike many of the available methods, the proposed method imposes no constraints on the choice of the different subregions of a domain, It also imposes no constraints on number of elements along different sides of a1822 E, A. SADEK LER LENA ARS ISIN NSS ARSE LESS ERS AROS ANY a Figure 9, A machine element subregion. In other words, there are no constraints on the user in choosing subregion shapes or element distribution over these subregions. By relieving him of these constraints, he will be able to choose the element distribution that best fits the expected stress gradient without the necessity of adding any extra elements. From this point of view, the’ method is believed to be economical in both human effort and computation time. "The method has been applied to a large variety of examples and the elements obtained were always satisfactory even with vigorous changes of element size. The method has been proved to be computationally efficient and to possess a good degree of generality. The method can also be used in connection with the method suggested by Zienkiewicz and Phillips" to handle problems with curved boundaries. Indeed the same concept can be extended to three-dimensional mesh generation using tetrahedral elements. This will be the subject of a future work. ACKNOWLEDGEMENTS ‘This work has been prepared in the Department of Civil and Structural Engineering, UMIST, ‘during the sabbatical leave of the author from Cairo University. The program will be fitted into the Departinental Finite Element Library for future research and development. ‘The author wishes to express his sincere indebtedness to Prof. D. M. Brotton, Head of the Department, for all the facilities offered to him during the preparation of this work. ‘Thanks are also due to Dr. M. A. Millar and Mr. J. L. Kirton for their valuable discussions. ‘The author also wishes to thank Prof. M. Ezzat Mohsin of Cairo University for his interest and continuous encouragement. YE REFERENCES 1. A. Bykat, “Automatic generation of triangular gid: () subdivision of « general polygoa into convex subregions: ( ‘tlangulation of convex polygons’, Int. J. num. Meth. Engng, 10, 1329-1342 (1976) 3 2. JC: Cavendish, “Automatic triangulation of arbitrary planar domains for the finite element method’, Int. J. mur. ‘Meth, Engng, 8, 679-696 (1974). 1 . 3, S-F. YeungandM. B. Hsu, A mesh generation method based on set theory’, Comp. & Struct. 3, 1063-1077 (1973). "G. Zienkiewice and D. V. Phillis, ‘An automatic mosh generation scheme for plane and curved surfaces by srametrie co-ordinates’, Int. J. num. Meth. Engng, 3, 519-528 (1971). au) ( (nb pdspibditeb st