CHAPTER INTRODUCTION This chapter outlines the finite clement method—what it is, to what problems it may be applied, and how it should be used. Details of these matters and appropriate theory occupy the remainder of the book. 1.1 FINITE ELEMENT ANALYSIS Finite element analysis (FEA), also called the finite element method (FEM), is a method for numerical solution of field problems. A field problem requires that we determine the spatial distribution of one or more dependent variables. Thus we may seek the distribution of temperature in the piston of an engine, or we may seek the distribution of displacements and stresses in a paving slab. Mathematically, a field problem is described by differential equations or by an integral expression. Either description may be used to formulate finite clements, Finite element (FE) formulations, in ready-to-use form, are contained in general- purpose FEA programs. It is possible to use FEA programs while having little knowledge of the analysis method or the problem to which it is applied, inviting consequences that may range from embarrassing to disastrous. Individual finite elements can be visualized as small pieces of a structure. The word “finite” distinguishes these pieces from infinitesimal elements used in calculus. In each finite element a field quantity is allowed to have only a simple spatial variation, perhaps described by polynomial terms up to x7, xy, and y*. The actual variation in the region spanned by an element is almost certainly more complicated, so FEA provides an approx- imate solution. Elements ate connected at points called nodes (Fig, 1.1-1). The assem- blage of elements is called a finite element structure, the word “structure” being used in a general sense to mean a defined body or region. The particular arrangement of elements is called a mesh. Numerically, an FE mesh is represented by a system of algebraic equations to be solved for unknowns at nodes. Nodal unknowns are values of the field quantity and, depending on element type, perhaps also its first derivatives. The solution for nodal quan- tities, when combined with the assumed field in any given clement, completely determines the spatial variation of the field in that element. Thus the field quantity over the entire structure is approximated element by element, in piecewise fashion. Although an FEA solution is not exact (unless the problem is so simple that FEA is probably inappropriate), the solution can be improved by using more elements to represent the structure. FEA has advantages over most other numerical analysis methods, including versatility and physical appeal. + FEA is applicable to any field problem: heat transfer, stress analysis, magnetic fields, and so on. + There is no geometric restriction. The body or region analyzed may have any shape.2 INTRODUCTION Figure 1.1-1. A two-dimensional model of a gear tooth. All nodes and elements lie in the plane of the figure. Supports are not shown. Boundary conditions and loading are not restricted. For example, in stress analysis, any portion of a body may be supported, while distributed or concentrated forces may be applied to any other portion. Material properties are not restricted to isotropy and may change from one element to another or even within an element. Components that have different behaviors, and different mathematical descriptions, can be combined. Thus a single FE model might contain bar, beam, plate, cable, and friction elements. An FE structure closely resembles the actual body or region to be analyzed. The approximation is easily improved by grading the mesh so that more elements appear where field gradients are high and more resolution is required. Other numerical methods have arisen since FEA appeared, but at present only FEA can confidently claim all these attributes. Overview of the Book. The rest of this chapter elaborates on the foregoing remarks in an introductory way. The most easily understood finite elements are those for structures com- posed of axial elements and beams. They are discussed in Chapter 2, along with some FEA procedures used with all element types. Simple triangular and rectangular elements for plane problems are discussed in Chapter 3. General formulation methods for finite ele- ments are treated in Chapters 4 and 5. Chapters 6 and 7 discuss arbitrarily shaped quadri- lateral elements and elements that may have curved sides. Chapters 8 through 10 discuss modeling procedures, sources of error, and how a sequence of analyses with successively1.2. Problem Classification, Modeling, and Discretization 3 refined meshes may be deemed to be adequately converged. Chapters 11 through 13 deat with the application of FEA to dynamics and vibrations, thermal problems, and a few problems that involve fluids and incompressible media. Chapters 14 through 18 are devoted to topics in structural mechanics, including material nonlinearity. Appendices deal with matrix manipulations, equation solving, and eigenvalue extraction. Over many years, element formulations and analysis procedures have been modified, extended, and fine-tuned to improve performance. Throughout the book we emphasize the more basic concepts, elements, and procedures. Refinements are often beyond the scope of this book but are included in commercial software and are described in references cited. 1.2 PROBLEM CLASSIFICATION, MODELING, AND DISCRETIZATION Classification. ‘The first step in solving a problem is to identify it. What are the more important physical phenomena involved? Is the problem time-independent or time-dependent? (In stress analysis terminology, we ask whether the problem is static or dynamic.) Is nonlinearity involved, so that iterative solution is necessary? What results are sought from analysis? What accuracy is required? Answers to such questions influence how much information must be gathered to carry out an analysis, how the problem is modeled, and what method of solution is adopted. ‘complicated problem may not lie entirely in one category. An example is a fluid-structure interaction problem, such as earthquake excitation of a storage tank that contains liquid. Motion of the liquid makes a thin-walled tank deflect, and deflection modifies the liquid motion. Therefore, structural displacement and fluid motion fields cannot be considered sepa rately; calculations must take their interaction into account. This example involves what may be called direct or mutual coupling, in which each field influences the other. There is also what may be called indirect or sequential coupling, in which only one field influences the other. An example is ordinary analysis for thermal stresses, where temperature influences stresses but stresses have negligible influence on temperature. Modeling. An analytical method is applied to a model problem rather than to an actual physical problem. Even laboratory experiments use models unless the actual physical struc- ture is tested. A model for analysis can be devised after the physical nature of the problem has been understood. In modeling, the analyst seeks to exclude superfluous detail but include all essential features, so that analysis of the model is not unnecessarily complicated yet provides results that describe the actual problem with sufficient accuracy. A geometric model becomes a mathematical model when its behavior is described, or approximated, by selected differential equations and boundary conditions. The equations, depending on their particular forms, may incorporate restrictions such as homogeneity, isotropy, constancy of material properties, and smallness of strains and rotations. It is important to recognize that FEA is simulation, not reality. FEA is applied to the mathematical model. Even very accurate FEA may be at odds with physical reality if the mathematical model is inappropriate or inadequate. ‘A mathematical model is an idealization, in which geometry, material properties, loads, and/or boundary conditions are simplified based on the analyst’s understanding of what4 INTRODUCTION features are important or unimportant in obtaining the results required. As examples in stress analysis, material may be regarded as homogeneous, isotropic, and linearly elastic (although common materials are otherwise); a load distributed over a small area may be regarded as concentrated at a point (which is not physically possible); a support maybe designated as fixed (although no support is completely rigid); a reentrant comer may be introduced but high stresses there ignored (if stresses elsewhere are sought); and an almost-flat structure may be modeled as two-dimensional (if stress variation in the thickness direction is considered to be practically zero, or regarded as linear as it often is for bending). Behavior of an axisymmetric pressure vessel might be described by equations of axisymmetric elasticity or by equations of axisymmetric shells, depending on whether the wall thickness is judged to be thick or relatively thin. Modeling decisions such as these precede FEA. Discretization. A mathematical model is discretized by dividing it into a mesh of finite elements. Thus a fully continuous field is represented by a piecewise continuous field defined by a finite number of nodal quantities and simple interpolation within each ele- ment, Clearly, discretization introduces another approximation. Relative to reality, two sources of error have now been introduced: modeling error and discretization error. Mod- ling error can be reduced by improving the model; discretization error can be reduced by using more elements. Even if discretization error could be reduced to zero, reality is not perfectly represented because modeling error remains. Also, as a computer does arith- metic, it introduces numerical error by using numbers of finite precision to represent data and the results of manipulation. Numerical error is usually small but can be made large by some physical situations and by poor discretization. ‘As a very simple example of modeling and discretization, consider a tapered support post, as in Fig. 1.2-1. Its cross-sectional area varies from Ay, at the bottom to A, at the top. In modeling, we elect to show the ground as a rigid support. Once we have omitted defor- ation of the ground from the model, deflection at the top is due entirely to shortening of ctual post, We may presume that the state of sires is uniaxial at_every cross section, ane is an acceptable approximation if taper is slight. In the mathematical model of upiaxialstress, coordinate is the only independent le. This representation is consistent with uniform stress across WHE bottom. Similarly, in this model the manner of load distribution on top does not matter; only its magnitude P is important. The distributed load of the post’s own weight usually may be neglected. We may also assume that the model material is homogeneous and linearly elastic. Thus, if the material is concrete, we ignore its cemented-particle structure and consequent local stress variations, and ignore For uniaxial stress and linear elasticity, we can obtain a valid discretized model by rep- resenting the tapered model by a stack of uniform members, each of elastic modulus F but different cross-sectional area A, as shown in Fig. 1-2-T- Discretization error can be reduced by increasing the number of members, This manner of discretizing a tapered structure is not at all new, but it can be regarded as a simple instance of FEA, as explained in the next section. After completing an analysis, it is important to check the results. In Fig. 1.2-1 it is obvi- ous that, prior to discretization, axial stresses throughout the model have magnitudes between P/Ay and P/A,. Even simple checks such as this may detect a large error, due perhaps to a blunder in data input.1.3. Interpolation. Elements, Nodes, and D.O.F. Model Discretized models P 0, YY A, P A, tr =P 1 Ay 2 Ay 3 As Ay Rigid Physical Finite element support representation representation 1.3 INTERPOLATION. ELEMENTS, NODES, AND D.O.F. Figure 1.2-1. Steps in modeling and FE analysis of a tapered support post. The essence of FEA is approximation by piecewise interpolation of a field quantity, Usu- ally, polynomial interpolation is used. Here we illustrate the method by applying it to the tapered bar shown in Fig. 1.3-1, modeled as a problem of uniaxial stress. We will briefly describe a bar element and the nature of results it provides. The results have features in common with results produced by most other FE analyses, however complicated the phys- ical problem may be. Meet Foe don ot 7 ee | po ree —r == Exact x 0. 0 Ly ° Figure 1.3-1. A tapered bar, discretized by three uniform two-node elements.6 INTRODUCTION, Bar Element. Figure 1.3-1 shows a discretization of a tapered bar into three elements, each uniform, linearly elastic, and of length L. The field quantity is axial displacement u. Along a typical element, say the middle one, u is stated in terms of nodal displacements uz and u3 by the equation : us (1-2) + dus (13-1) where s is an axial coordinate along the element. Clearly, Eq. 1.3-1 expresses a linear vari- ation of u with s that has values u = up ats = Oandu = u3 ats = L. The same form applies to the rightmost element, but with nodal displacements uz and ug. Similarly the form applies to the leftmost element, but with u, = 0 because the left end of the structure is fixed. Linear displacement u = u(s) dictates that axial strain ¢ is constant over an ele- ment. From the stress-strain relation « = Ee and the elementary definition of strain as change in length divided by original length, and with u, = 0, we obtain the following expressions for axial stress in the respective elements of Fig. 1.3-1. u 2 o2=EF mg = EAS O44 = (1.3-2) The problem of Fig. 1.3-1 is simple enough that the FE solution can be obtained without matrix formulations and systematic manipulation procedures. Instead we solve for nodal displacements using methods of elementary mechanics of materials, as follows. The three bar elements are each uniform, of respective cross-sectional areas 6A, 44, and 2A, which are cross-sectional areas of the tapered bar at element midpoints. Nodal dis- placements can be obtained from the elementary expression for elongation of a bar under axial load. Thus PL PL PL m=O m= Se ma mt Te ms tae (13:3) ~ These displacements can be expressed in terms of overall length Lr by the substitution L = Ly/3. Next, element stresses can be obtained from Eqs. 1.3-2. To check results, we can simply divide load P by element cross-sectional areas, because this simple problem is statically determinate. Results are plotted in Fig. 1.3-1. The displacement plot is reminiscent of using straight lines for numerical interpolation between points on a continuous curve, but here points do not lie on the correct curve. In other words, nodal values of field quantities are not exact. The cause is discretization error. Only for certain very simple problems are nodal values exact. “Exact” means full agreement with behavior of the mathematical model, not neces- sarily agreement with physical reality. The stairstep axial stress plot shows that stresses in this example are accurate at element centers. Elsewhere stresses are represented less accu- rately than displacements, as should be expected in FEA results because most types of finite elements are based on displacement fields, and stresses are usually computed from displacement gradients. An alternative form of Eq. 1.3-1 is 4 = a+ ays (13-4)1.3 Interpolation. Elements, Nodes, and D.O.F. a) where a; and az are constants that can be expressed in terms of wy and 3 by requiring that u = uyat 5 = Oandu = uz ats = L, Thea; are known as generalized coordinates or generalized degrees of freedom. The term “degrees of freedom” is abbreviated “d.o.f.” and is explained at the end of this section. . Other Elements. From Figs. 1.1-1 and 1.3-1 one may surmise (incorrectly) that an FE discretization is obtained by conceptually sawing an actual structure into small fragments and then reconnecting them at convenient points. If applied to the gear tooth of Fig. 1.1-1, this process would result in triangular fragments connected at comer and midside nodes. Such a patchwork would be weaker than the actual structure, with strain concentrations near nodes, sliding of some fragments on one another, and gaps between others. These defects do not arise in FEA because elements are based on simple fields, which do not contain terms capable of representing a strain concentration, and which provide interele- ment compatibility. Fields used for some simple two-dimensional elements are as fol- lows. In subsequent chapters we explain how these fields are used in FEA. Figure 1.3-2a shows a three-node triangular element that can represent a two-dimensional field 6 = (xy). As examples, might represent temperature, voltage, hydraulic head in seepage flow, or lateral deflection of an inflated membrane. In the form of Eq. 1.3-4, the ele- ment field is $= ataxtay (1.35) ‘The three a; can be expressed in terms of values of ¢ at the three element nodes, as will be shown in Chapter 3. Figure 1.3-2b shows that a mesh of these elements approximates a smooth function @ = (xy) by a surface of triangular facets. The four-node rectangular element shown in Fig. 1.3-2b has the field = a +axx + ayy tagry (13-6) and the six-node triangular element with midside nodes has the field @ = ay $agx + ayy tag? tasty tag” (13-7) ‘The four-node rectangular element displays ¢ = (uy) over the element as a surface that may be flat or warped. The six-node triangular element can display a parabolic ¢ surface. In all these elements, the variation of @ along an element edge is completely determined by values of ¢ at nodes on that edge. Therefore, adjacent elements that share nodes along a common edge automatically display the same function ¢ along the entire shared edge, and no incompatibilities such as gaps appear between elements. Summing up, we may say that FEA is an analysis method in which a field variable is approximated by connecting simple interpolation functions, each defined over a small region. The region is called a finite element. The interpolation function (such as Eq. 1.3-5) is adapted to the number of nodes in the element type, and amplitudes of the a; are deter- mined by numerical values of the field quantity at specific points called nodes. Elements ‘Some elements, discussed subsequently, are incompatible, Along element edges, but not at nodes, gaps or overlaps can appear between adjacent elements. Such elements are formulated in away that enhances coarse-mesh accuracy and causes incompatibilities to tend toward zero asa mesh is refined,8 INTRODUCTION ¢ | e 3-node 6-node triangular triangular element element 4s 3 ay > a ~ 4-node rectangular element @ ® Figure 1.3-2. (a) A three-node triangular element. (b) A smooth function $ = (x,y) can be approximated by various element types. as are connected at nodes, where they share values of the field quantity (and may also share one or more of derivatives of the field quantity, depending on element type). Nodes are also locations where loads are applied and boundary conditions are imposed, Degrees of Freedom (d.o.f.). Degrees of freedom are independent quantities that gover the spatial variation of a field. For example, Eq. 1.3-7 defines a field @ = (x,y) that has six d.o.f,, namely the six q;. Six nodal d.o.f. ; can be used instead to define the same field. Each element in Fig. 1.1-1 has 12 d..f., namely six nodal displacements u; that govern the x-direction displacement field u = u(x,y) and six nodal displacements v; that govern the y-direction displacement field v = v(x,y). The u; and v; are displacements of spe- cific points; in general the a; are not. 1.4 EXAMPLE APPLICATIONS. HISTORY OF FEA Applications. Figure 1.4-1 shows an application of FEA that dates from 1965 [1.1]. The structure is an axisymmetric solid, whose axis of revolution lies above the créss section shown, Each finite element is a toroidal ring of triangular cross section. Each element has a node (or in this case a nodal circle) at each vertex. Field quantities at each node are tem- perature for heat conduction analysis, and radial and axial displacements for stress analy- sis. The same discretization can be used for both analyses. Computed nodal temperatures are transferred to the stress analysis model and used to determine thermal stresses. In Fig. 1.4-2, FEA is applied to an induction motor. Only part of the motor is shown; symmetry is exploited by modeling only a repetitive portion. The mesh of triangular ele- ments spans spaces between pole pieces as well as the pole pieces themselves, For magne- tostatic analysis, nodal unknowns are values of the magnetic potential. Examples of FEA could be given from many other areas of application. The concept of piecewise interpolation is common to all of them. However, familiarity with concepts of FEA does not confer competence in all applications. For example, the problem of Fig. 1.4-11.4 Example Applications. History of FEA 21.37 diameter Figure 1.4-1. Cross section of a multi-material rocket nozzle, showing construction (left portion) and possible finite element mesh (right portion), from [1.1]. poses no great challenge to today’s stress analyst, but if presented with the magnetics prob- lem of Fig. 1.4-2, a stress analyst may not know what kind of result to seek or what input data is required. It is important to understand the physics of the problem. FEA was accepted by industry soon after its introduction, for reasons suggested by the foregoing two applications. Finite elements can represent structures of arbitrarily complex geometry, A discretized model resembles the actual body or region. Each element can be Figure 1.4-2, Part of an induction motor. Computed magnetic flux contours for zero rotor speed are shown by the right-hand figure. (Courtesy of A. O. Smith Corp., Data Systems Division, Milwaukee, Wisconsin.)10 INTRODUCTION regarded as a piece of the actual structure (but having idealized behavior). Systematic pro- cedures of FEA allow calculations to be almost completely automated. Unfortunately, automation makes it possible to do FEA with so little understanding that results may be worthless. Some critics say that most FEA results are worthless. It is important to under- stand how finite elements behave, and important to check for errors. History. In 1851, to derive the differential equation of the surface of minimum area bounded by a given closed curve in space, Schellbach discretized a surface into right trian- gles and wrote a finite difference expression for the total discretized area [1.2]. He pro- posed no other application or generalization of the idea. FEA is now regarded as a way to avoid differential equations by replacing them with an approximating set of algebraic equations. Starting in 1906, researchers noted that a framework having many bars in a regular pat- tem behaves much like an isotropic clastic body [1.3,1.4]. Application to problems of plane elasticity and plate bending was reported in 1941 [1.5]. This work exploits well- known methods for analysis of framed structures but cannot be applied to bodies of arbi- trary shape. Also, rather than discretization of a continuum into smaller pieces, structural members of a different type are substituted. The framework method may be regarded as a precursor to FEA rather than an early form of it. The FE method as we know it today seems to have originated with Courant in his 1943 paper, which is the written version of a 1941 lecture to the American Mathematical Soci- ety [1.6]. Courant determined the torsional rigidity of a hollow shaft by dividing the cross section into triangles and interpolating a stress function ¢ linearly over each triangle from values of ¢ at net-points (or nodes, as we now call them). He does not mention Schell- bach’s work. Courant notes that the method “suggests a wide generalization which pro- vides great flexibility and seems to have considerable practical value.” Practical applications did not appear until aeronautical engineers developed the method, apparently without knowing of Courant’s work. Engineers in the aeronautical industry made remarkable progress in the early to mid- 1950s, although some of the work was not published until much later due to company poli- cies. Early in this period, equations from conventional analysis methods were solved on the smail computers then available, In the United States, conventional methods proved inade- quate for wings of low aspect ratio, so Tumer devised a three-node triangular element to model the wing skin [1.7]. In England, Taig did similar work [1.8]. In Germany, Argyris included FEA concepts in a set of influential papers about matrix procedures [1.9]. Details may be found in the references cited; see also [1.10-1.13]. The name “finite element” was coined by Clough in 1960. Many new elements for stress analysis were soon developed, largely by intuition and physical argument. In 1963, FEA acquired respectability in academia when it was recognized as a form of the Rayleigh-Ritz method, a classical approximation technique. Thus FEA was seen not just asa special trick for stress analysis but as a widely applicable method having a sound mathematical basis. Papers about heat conduction and seepage flow using FEA appeared in 1965. General-purpose computer programs for FEA emerged in the late 1960s and early 1970s. Since the late 1970s, computer graphics of increasing power have been attached to FE software, making FEA attractive enough to be used in actual design. Pre- viously, FEA was so tedious that it was used mainly to verify a design already com- pleted or to study a structure that had failed.1.5 Solving a Problem by FEA u Computational demands of practical FEA are so extensive that computer implementa- tion is mandatory. Analyses that involve more than 100,000 d.o.f. are not uncommon. It is no accident that developments in computers and programming languages were contempo- raneous with early developments in FEA. The first textbook about FEA appeared in 1967 [1.14]. By 1995, Mackerle {1.15} estimated that about 3800 papers about FEA were being published annually, and that the cumulative total of FEA publications amounted to some 380 books, 400 conference proceedings, and 56,000 papers (excluding papers on fluid mechanics). Mackerle also counted 310 general- purpose FE computer programs. 1,5 SOLVING A PROBLEM BY FEA Solving a practical problem by FEA involves leaming about the problem, preparing a mathematical model, discretizing it, having the computer do calculations, and checking results. Most often, more than one cycle through these steps is required. Time spent by the computer is a small fraction of time spent by the analyst, but the analyst must have an understanding of what the computer is doing. Material of the present section is discussed in detail in Chapter 10. Problem Classification. As summarized in Section 1.2, the analyst must understand the nature of the problem. Without this step a proper model cannot be devised, nor can FEA software be told what to do. At present, software does not automatically decide that non- linear analysis is to be undertaken if stresses are high enough to produce yielding, that buckling is to be considered if thin sections carry compressive load, and so on. Although the trend is for software to be given more decision-making capability, the analyst should not abdicate control. Software has limitations and almost certainly contains errors, yet the engineer, not the software provider, is legally responsible for results obtained. Mathematical Model. Before undertaking FE discretization and a numerical solution, we devise a model problem for analysis. This step involves deciding what features are important to the purpose at hand, so that unnecessary detail can be omitted, and deciding what theory or mathematical formulation describes behavior. Thus we may ignore geo- metric irregularities, regard some loads as concentrated, and say that some supports-are fixed. Material may be idealized as linear and isotropic. Depending on the dimensions, loading, and boundary conditions of this idealization, we may decide that behavior is described by beam theory, by plate-bending theory, by equations of plane elasticity, or by some other analysis theory. The simplified problem, with the analysis theory to be applied in solving it, constitutes the mathematical model. Because subsequent FEA is approximate and pertains only to the mathematical model, FEA is two or three steps removed from reality. Modeling decisions are influenced by what information is sought, what accuracy is required, the anticipated expense of FEA, and its capabilities and limitations. Also, initial modeling decisions are provisional. It is likely that results of the first FEA will suggest refinements, in geometry (perhaps by restoring geometric irregularities previously omitted), in applicable theory (perhaps by adding in-plane stretch- ing terms to plate-bending theory), and so on.12 INTRODUCTION ‘As an example of modeling, consider the very simple problem depicted in Fig. 1.5-1a. The ring thickness (measured normal to the figure) is uniform and is considerably less than the ring diameter. The material is considered linearly elastic, homogeneous, and iso- tropic. We ask for stresses and deflections due to the ring’s own weight as it rests on the ground. It is easy to arrive at the plane model in Fig. 1.5-1b, in which symmetry about the vertical centerline has been exploited. Details of pressure applied by the ground have been discarded, replaced by a point support. If mean radius R is pethaps St or more, the largest stresses in the actual problem are circumferential flexural stresses. Then the theoretically infinite stresses at D associated with a point support are not important (and could only be calculated as high stresses by conventional finite elements). A two-dimensional model is adequate. Instead, if the physical structure is not a ring but a long, thin-walled pipe, should the model be three-dimensional? Probably not. Stress analysts recognize that deflections and stresses are essentially constant along the length, that they vary only near ends, and that the variation has only a small effect on the largest magnitudes of deflection and stress. However, in a long pipe the situation is more nearly plane strain than plane stress. Thus the model is changed, and appropriate data must be supplied to software and appropriate anal- ysis options chosen. The foregoing conceptual models become complete mathematical models when we decide on the appropriate analysis theory. For a slender ring, it can be beam theory. For a not-so-slender ring, plane elasticity theory is appropriate. For a thin-walled pipe in which end effects are to be represented, thin-shell theory is appropriate. Elements based on the respective theories would be used for FEA of the respective mathematical models. Why not use a three-dimensional model? After all, reality is always three dimensional, and elements for three-dimensional FEA are available. The reason is cost. Demands on the analyst’s time and computer resources are likely to increase by a factor of 10 or more in going from two dimensions to three. t i Ti (Gravity RUPLI \ Figure 1.5-1. (a) Ring in the vertical plane that rests on the ground, loaded by its own weight. (b) Mathematical © o model,15. Solving a Problem by FEA 13 A gear tooth poses a more complicated problem than the ring. In Fig. 1.1-1, supports are not shown. Is it satisfactory to impose full fixity along ABCD, where elastic support is actually provided by the remainder of the gear? An actual gear is not of uniform thickness; is a two-dimensional model satisfactory? Is load P uniformly distributed in the z-direction across the tooth? If not, stresses will be considerably higher, and the situation cannot be considered two-dimensional. Preliminary Analysis. Before going from a mathematical model to FEA, at least one pre- liminary solution should be obtained, using whatever means are conveniently available— simple analytical calculations, handbook formulas, trusted previous solutions, or experiment. Some of this effort may lead to a better mathematical model. Subsequently it will be used to check computed results. If we do this work before FEA rather than after, we reduce a natural tendency to find answers that support whatever FEA results have already been obtained, especially if it took considerable effort to get them, It is easy to make mistakes in supplying data to software, and even a crude preliminary solution may detect a result that errs greatly due to a mistake in data input. Preliminary analysis for the ring problem of Fig. 1.5-1 is easy if R is considerably greater than f. Formulas for deflection and stress in a slender ring are available in hand- books [1.16]. If R is comparable to ¢, these formulas are approximate but still useful for checking. Finite Element Analysis. Use of general-purpose FEA software involves the following steps. * Preprocessing: Input data describes geometry, material properties, loads, and bound- ary conditions. Software can automatically prepare much of the FE mesh, but must be given direction as to the type of element and the mesh density desired. That is, the analyst must choose one or more element formulations that suit the mathematical model, and state how large or how small elements should be in selected portions of the FE model. All data should be reviewed for correctness before proceeding. Numerical analysis: Software automatically generates matrices that describe the behavior of each element, combines these matrices into a large matrix equation that represents the FE structure, and solves this equation to determine values of field quantities at nodes. Substantial additional calculations are performed if behavior is nonlinear or time-dependent. Postprocessing: The FEA solution and quantities derived from it are listed or graphi- cally displayed. This step is also automatic, except that the analyst must tell the soft- ware what lists or displays to prepare. In stress analysis, typical displays include the deformed shape, with deformations exaggerated and probably animated, and stresses of various types on various planes. Check the Results. First, we examine results qualitatively and ask if they “look right”— that is, are there obvious errors? Have we solved the problem we intended to solve, or some other problem? Boundary conditions are often misrepresented; does the deformed FE strue- ture show displacements where there should not be any? Are expected symmetries present in the results? If answers to such questions are satisfactory, FEA results are compared with4 INTRODUCTION Unaveraged stress bands Averaged stress bands ® ® Figure 1.5.2. Stress bands in aportion of a mesh of rectangular elements. (a) Without nodal averaging, interelement discontinuity suggests how good or bad results are. (b) After nodal averaging, continuity prevails, but important information is lost. =|) solutions from preliminary analysis, and with any other useful information that may be available. For example, let us qualitatively examine the problem of Fig. 1.5-1. Points along CD and AF should move downward but not left or right. Points along BE should move down- ward and rightward. Vertical stresses should be compressive near B and D. Horizontal stresses should be tensile near A and C, compressive near D and F. Stresses normal to boundaries ABC and DEF should be zero, but will not be exactly so because the solution is approximate. Similarly, due to symmetry, shear stress should be zero along CD and AF, and stress contours of flexural stress should be normal to CD and AF, but will not be exactly so. ‘One way to judge the adequacy of a discretization is to look at plots of stress (or plots of heat flux in thermal analysis). Software can plot either stress contours or “stress bands,” which are zones of color. Different colors are used for different levels of stress. Stress is related to gradients of the field quantity, and gradients in a given element depend on field quantities at nodes attached to that element only. Therefore, as will be shown subse- quently, stress bands are discontinuous across interelement boundaries. Strong discontinu- ities indicate too coarse a discretization, whereas practically continuous bands suggest unnecessarily fine discretization [1.17]. In Fig. 1.5-2a, bands are discontinuous but not badly so, and the discretization may be adequate for the purpose intended. Software can be instructed to display bands computed from nodal average values of stress (or of flux). Thus interelement discontinuities are removed. The resulting picture is visually more pleasing, but information useful in judging the quality of computed results is lost. Bands plotted from nodal averages, Fig. 1.5-2b, may suggest that results are of higher quality than is actually the case. Expect to Revise. Rarely is the first FE analysis satisfactory. Obvious blunders must be corrected. Uncomfortably large discrepancies between what is expected and what is com- puted demand explanation. Either physical understanding or the FE model, or both, may be at fault. Disagreements must be satisfactorily resolved by repair of the mathematical model and/or the FE model. After another analysis cycle, the discretization may be judged1.6 Learning and Using FEA 15 START ‘Consider the ‘Obtain approximate Plan a finite element - physical actions results for subsequent discretization of the = involved. Devise ‘comparison with mathematical model. Preprocess: (or improve) a FEA results. Bild the finite atheatical element mode} model im the computer. Revise the ‘Are error ina estimates cece is at faule discretization. Generate and. i ! 1 ' ' ' { ' Physical understanding small? Does ' or FEA modeling? nen 1 ' ' ' ' I t ' \ \ ‘solve equations of the finite element model. do little to alter FBA results? Postprocess: Display computed results for examination, Are FEA results free of obvious errors, such as disagreement with boundary conditions intended? Are FEA results physically reasonable? Do FEA results agree well enough with predictions and approximations obtained by other means? Figure 1.5-3. Outline of a finite element analysis project. inadequate, perhaps being too coarse in some places. Then mesh revision is required, fol- lowed by another analysis. In analyzing a new problem, it is almost always appropriate to begin with a simple FE model, to which detail is added as the analyst leams more. Each revision is an expected step on the way to an adequate solution, not a penalty for failure in the preceding attempt. The flow of an analysis project by FEA is outlined in Fig. 1.5-3. 1.6 LEARNING AND USING FEA, Why study the theory of FEA? Satisfactory elements and versatile analysis procedures are already available in widely used software, and software has become so accommodating that even an inept user can obtain a result. Seasoned practitioners stress that reliable results are obtained only when the analyst understands the problem, how to model it, behavior of finite elements, assumptions and limitations built into the software, input data formats, and when the analyst checks for errors at all stages. It is not realistic to demand that analysts understand details of all elements and procedures, but misuse of FEA can be avoided only by those who understand fundamentals. For example, it is important to real- ie that each individual element has very limited ability to represent spatial variation of the field quantity, and to understand how this ability differs from one element type to another. Older engineers sometimes complain that younger engineers have naive faith in com- puter programs, value computer skills over analytical skills, and lack the ability to produce “ballpark” answers. Such deficits can be overcome while learning FEA. A student can use FEA to analyze problems for which results are already available and known to be reliable, discovering and fixing the inevitable mistakes in modeling, data input, and software options until computed results agree with established results. Problems for which results are not available can be solved analytically—crudely if necessary—and then solved by FEA, with the process repeated until results are reconciled. This exercise will improve16 INTRODUCTION analytical skills as well as FE skills. Initial failure to achieve agreement may be discourag- ing, but it is more instructive than success. ‘A study of computer misuse in engineering [1.18] considers cases in which incorrect results caused damage in the form of expensive delay, a need to redesign, poor perfor- mance, or even collapse. Of 52 cases cited, 7 were due to hardware error, 13 to software error, 30 to user error, and 2 to other causes. User error was usually associated with poor modeling, and sometimes with poor understanding of software limitations and input data formats. Most errors could have been caught early had users been careful to check results. Often, after damage was done, the cause of the trouble was found by consultants who used hand calculation to check computer output. ‘A cautionary example is depicted in Fig. 1.6-1. Here a straight beam with hinge sup- ports is loaded by a pressure pulse that causes yielding of the material and vibration of the beam. Analysis seeks to track lateral displacement at the midpoint as a function of time. Resulis plotted come from 10 reputable analysis codes operated by users regarded as expert (1.19]. Yet if any of the curves is correct, we cannot tell which one it is. Admittedly, the problem is difficult: results indicate “strong sensitivities of both physical and computa- tional nature” [1.19]. This example reminds us that analysis software is based on theory and approximation, and that a user may push the software beyond its range of validity [1.20], k—————— 200 mm ———-——>} — 420 mm Ke lems pressure pulse ne Displacement, mm Time, ms Figure 1.6-1. Lateral midpoint displacement versus time for a beam loaded by a pressure pulse [1.19], reproduced courtesy of ASME. The material is elastic-perfectly plastic. Plots were generated by various analysts and various software packages.Analytical Problems Wy ANALYTICAL PROBLEMS 1.3-1 (a) For the problem of Fig. 1.3-1, determine both exact and finite element values of axial displacement u at x = L7/3, x = 2L7/3, and x = Ly. Hence, verify the plots of u versus x. (b) In similar fashion, verify the plots of o versus xin Fig. 1.3-1. 1.3-2 Strain 6, is given by the expression ¢, = ou/dx. What expression for e, is obtained when w in a four-node plane element is given by the right-hand side of Eq. 1.3-6? For a mesh of such elements, what can you say about interelement continuity of ,? 13-3 (a) Ina three-node triangle, field quantity $ can be written as @ = ay +.ayx + ayy, where the a; are generalized d.of. For the particular shape of triangle shown, express ¢ in the form = fi +fo¢y +fab3, where the fj are functions of x, y, a, and b. Suggestion: Obtain three equations for the a; from the conditions = $j at = y = 0,6 = gyatx = aandy = 0,andd = gyatx = Oandy = (b, ¢, d) In similar fashion, obtain expressions = fy; + foby + fads for the trian- gles shown. TT By y each \ | z b b : 3 | | x x k—a—l LN ke ape a asa ® % © @ Problem 1.3-3 13-4 For the plane quadrilateral element shown, imagine that field quantity ¢ has the form @ = a; + ayx + aay + ayxy, where the a; are generalized d.o.f. How does $ vary with x or y along each side? Do you think this element will be compatible with neighboring elements that may be attached to it? y non L\ 1 2 Problem 1.3-4 Problem 1.4-1 AV TY 14-1 The sketch shows a propped cantilever beam under uniformly distributed load, as it might be sketched in a book about mechanics of materials. What idealizations of reality may have been introduced in arriving at this model? 14-2 A cylindrical pipe, shown in cross section, has nominal temperatures T, on the inside and T) on the outside. The standard analytical solution for temperature T at arbitrary radius r in the pipe is18 INTRODUCTION In(r/r,) T=T+(- May However, actual circumstances may differ sufficiently from the ideal that this equa- tion is not accurate enough, and temperature distribution must instead be deter- mined from FEA. What are some of these circumstances? Problem 1.4-2 Problem 1.4-3 1.43 The sketch shows the cross section of a concrete gravity dam. The V symbols indi- cate water surfaces. Imagine that stress analysis for loading due to hydraulic pres- sure is required. Has anything of importance been omitted from the sketch? What considerations influence the mathematical model devised? What additional infor- mation will be needed before undertaking numerical analysis? 1.4-4 In the gravity dam of Problem 1.4-3, imagine that the rock is slightly porous and that analysis for seepage flow under the dam is required. Answer the questions posed in Problem 1.4-3. 14-5 Two steel plates are connected by a single rivet to form a lap joint, as shown, Axial load is applied. If accurate and detailed stress analysis is required, what aspects of material properties, geometry, and loading must be considered in planning an analysis? rh ToT rae Problem 1.4-5CHAPTER ONE-DIMENSIONAL ELEMENTS AND COMPUTATIONAL PROCEDURES ‘We consider straight elements that have a node at each end, and apply physical arguments to obtain matrices that represent element behavior. Then we use these comparatively simple elements and matrices to explain computational procedures that are generally applicable in FEA, regardless of element type. Thus in this chapter we survey the entire computational process of linear static FEA: formulation of element matrices, their assembly into a struc- tural matrix, application of loads and boundary conditions, solution of structural equations, and extraction of gradients (element strains and stresses, in this chapter) 2.1 INTRODUCTION One-dimensional elements include a straight bar loaded axially, a straight beam loaded lat- erally, a bar that conducts heat or electricity, and so on. In structural terminology, a bar can resist only axial load, whereas a beam, in its most general sense, can resist axial, Lateral, and twisting loads. In time-independent analysis, a truss of n members can be modeled by nbar elements, and a frame having n straight members usually requires n beam elements. ‘A beam continuous over two or more supports can usually be modeled using one beam element per span between supports. Thus, when one-dimensional elements are used for static analysis, the discretization phase of modeling becomes trivial, and for stress analysis the name “matrix methods of structural mechanics” may be used in preference to “FEA.” However, bar and beam elements are provided in FEA software and are much used, both as stand-alone elements and in combination with finite elements of other types, For exam- ple, beam elements can be attached to plate elements to model stiffened plates. In this chapter we restrict our attention to linear problems, which means that material properties are essentially unchanged by loading (by force or moment, by temperature, by voltage, and so on). In mechanical problems, linearity also requires that deformations be small enough that equilibrium equations can be written using original geometry rather than deformed geometry. That is, we exclude nonlinear behavior such as yielding of steel, crum- bling of concrete, opening or closing of gaps, and lateral deflection large enough to generate membrane-stretching action. Also, we consider only steady-state problems, which are called static or (more properly) quasistatic in structural mechanics. As an approximation, if structure is loaded by a cyclic force whose frequency is less than about one-quarter the structure’s lowest natural frequency of vibration, the loading can be regarded as quasistatic, and analysis of the type described in the present chapter is acceptable. A finite element has a characteristic matrix, which is a stiffness matrix for load-deformation analysis, a conductivity matrix for heat conduction analysis, and so on. One-dimensional ele- ments are simple enough that the characteristic matrix can usually be formulated by the 19