You are on page 1of 11
A.6 Computational considerations 589 21000) (G 4 14100/1G 2 O14 10/0 Gp=43 (4.90) oo14il]a 4 ooo012sla 2 The equations can be expanded as 21 4+Q=4 (A91) CH4Q +O = 2 (4.92) C+ 4C3 +4 = 3 (A93) Cr+ 4Crt Cs =4 (A94) Cy+2C5 = 2 (A.95) The first unknown, C; appears only in the first two equations, so it can be eliminated by multiplying (A.92) by 2 and subtracting (A.91) from it, with the result 1Cy + 2C3=0 (A.96) We next eliminate C> by multiplying (A.93) by 7 and subtracting (A.96) from it, with the result 26C4-7C4 = 21. (A97) Continued use of the same technique yields 98C4 + 26Cs = 83 (A.98) and finally 170Cs = 113 (4.99) or Cs = 0.6765. Equation (A.98) then gives 83 — 26 x 0.6765 Cy 98 0.6675 and the remaining unknowns are recovered one by one from equations (A.97, A.96, A.91) respectively as C3 =0.6280, Cp =—0.1794, Cy =2.0897. Anyone who has attempted to solve a system of five simultaneous equations by hand will realize that this procedure, which only works when the matrix is banded, is a great deal quicker than other methods. For a system of V simultaneous equations, it involves only N matrix row operations to find the last unknown, followed by another N operations to recover the full set of unknowns. This is a particularly simple example where the bandwidth of the matrix is only three — i.e. there are only three terms in each of the algebraic equations. In more complex problems, the band will be wider. For example it is of width 6 for the beam. bending formulation of equation (A.74). However, the computational effort involved. in this elimination procedure is still very much less than that required for direct inversion of the global stiffness matrix. 590 A The Finite Element Method. A.6.1 Data storage considerations When the system being analyzed is very complex, the amount of numerical data to be transmitted and stored is very large and this in itself requires special consideration. For example, the stiffness matrix contains a large number of components, but most of these components are zeros and considerable savings can be achieved by identifying the bandwidth of the matrix and storing only the non-zero values. We can also take advantage of the symmetry of the matrix (Kj; =Ki,) to store only half of the off- diagonal components. In the solution procedure for banded matrices used in Example A.6, only a few components of the stiffness matrix are needed at each stage and this permits some added efficiency in data handling. Even more efficiency is obtained by only assem- bling parts of the stiffness matrix as they are needed in the solution procedure. In this way, only a small part of the large stiffness matrix is stored in the computer at any given stage. The resulting algorithm is known as a front solver® and it makes it possible to solve problems involving millions of degrees of freedom, A.7 Use of the finite element method in design The finite element method is typically used ata later stage in the design process when the basic geometry of the system is largely finalized and we need to establish more accurate values for the stresses and displacements.‘ It is extremely widely used in industry and the chances are very high that you will be called upon to use it at some time in your career as an engineer, However, commercial finite element codes are very flexible and user friendly and they cover the vast majority of design needs. It is therefore almost never necessary to write a program yourself, The primary purpose of this appendix is to explain to the reader in a simple context some of the fundamental, processes involved in these codes and to introduce some of the terminology used. ‘The manuals and other supporting materials for the best known codes contain nu- merous examples and it is often possible to modify one of these to cover the problem. under consideration, merely by redefining the number of nodes and their location. ‘The best advice here is to start formulating the problem after the bare minimum of introductory reading, since the output from the program will often teach you how (o correct errors or improve the model. Make sure you can reproduce the example solution on your computer system before you make the modifications. The error messages you will get in early attempts can be obscure and frustrating If you know someone in the company who has used the code before, their advice 3 BM. Irons (1970), A frontal solution program for finite element analyses, International Journal for Numerical Methods in Engineering, Vol.2, pp-5~32. 4 ILis rather paradoxical that we use an avowedly approximate method when we want more accuracy. The reason of course is that by using a finer discretization, the accuracy of the fi- nite element method can be improved as much as we need, By contrast, the ‘exact’ methods discussed elsewhere in this book make an implicit approximation at the beginning when the component is idealized, for example as @ beam or a shell. A8 Summary 591 will be invaluable. Alternatively, if you cut and paste an obscure error message into Google or another search engine, often it will direct you to an explanation of the problem and a way of fixing it. Once the program appears (o be Working satisfactorily, there is a strong temp- tation to breathe a sigh of relief and accept the results as accurate. This is very risky, particularly since most commercial codes have sophisticated postprocessors that can make any results look very impressive and plausible. If the design is impor- tant enough for you to perform a finite element analysis, itis also important enough to make every effort to be sure that there are no errors in your formulation. There are three tests you might perform here:~ (® Compare the finite element results with the analytical solution of the simplest idealized problem you can think of that is somewhat like the real problem. The results should certainly be of the same order of magnitude and there is reason to be suspicious if they differ by more than a factor of 2. (ii) Try to find a system of loads and/or a small modification of the geometry that would cause the complex system to have a simple stress field such as uniform stress or a linearly varying bending stress. Make sure the finite element program confirms this result ii) ‘Test the convergence of the program by making several runs with different de- grees of mesh refinement, particularly in critical areas of high stress. A.8 Summary In this chapter, the basic principles of the finite element method were introduced and applied to problems in one dimension. In mechanics of materials problems, the method is equivalent to the use of the Rayleigh-Ritz method with a piecewise poly- nomial approximation to the displacement. The approximation is defined through a finite set of nodal values that constitute the degrees of freedom for the problem, An alternative mathematical formulation can be obtained from the governing equilibrium equation by choosing the nodal values to minimize the integral of the square of the error over the domain. This leads to the Galerkin approximation in which the same functions are used as approximating (shape) functions and weight functions. The advantage of this method is that it can be applied to any problem governed by a differential equation. The finite element method was applied to problems of axial loading and the bend- ing of beams. The extension of the method to problems in two and three dimensions was also briefly discussed. 5 This argument cuts both ways. If you do not feel high accuracy and reliability are crucial, itis probably not worth doing the finite element analysis, unless of course you simply want to convince your boss of your thoroughness! 592 A The Finite Blement Method Further reading K.J. Bathe (1982), Finite Element Procedures in Engineering Analysis, Prentice: Hall, Englewood Cliffs, NJ Y.K. Cheung and M.F. Yeo (1979), A Practical Introduction to Finite Element Method, Pittman, London. ‘TLR. Hughes (1987), The Finite Element Method, Prentice Hall, Englewood Cliffs, NJ, Chapters 1,2. H. Kardestuncer and D.H. Norrie, eds. (1987), Finite Element Handbook, MeGraw- Hill, New York. Y.W. Kwon and H. Bang (1996), The Finite Element Method using MATLAB, CRC Press, Boca Raton, LA. G. Strang and GJ. Fix (1973), An Analysis of the Finite Element Method, Prentice- Hall, Englewood Cliffs, NJ. OC. Zienkiewicz (1977), The Finite Element Method, McGraw-Hill, New York. Problems Section A.L A.1, Find a piecewise linear approximation to the parabola y=." in the range 0< x< I, using two equal elements and three nodes, 0,0.5, 1. Plot a graph comparing the function and its approximation. A.2. Find a piecewise linear approximation to the function sin(sx/2) in the range 0

You might also like