US 5,754,181 AGrant
Computer Aided Design System
Issue Date:1998-05-19
•7 Claims
•2 Drawing Sheets
Abstract
A computer-aided design system in which a mechanical quantity of a structure can be determined from stored geometrical representation of the structure, the material properties thereof and the loads imposed thereon, comprising: generating data describing a structure as a mesh; generating by application of the finite element method to the mesh, the elements of a stiffness matrix, a loading vector, and a vector including an associated degree of freedom index for each row and column of the stiffness matrix, generating from the stiffness matrix and the loading vector elements of matrix A and a right hand-side vector f, where A and f are related through the relation Ax=f, the vector x representing a mechanical quantity at a plurality of points of the mesh; and a linear solver of the preconditioned conjugate gradient type for generating, for subsequent processing or display, the elements of the vector x from A and f using a preconditioning matrix, the system being characterised in that the linear solver comprises generating a principal submatrix from the stiffness matrix by setting to zero those elements therein in positions such that the degree of freedom index associated with their row is not equal to the degree of freedom index associated with their column and generating the preconditioning matrix from the principal submatrix.
Metadata
Assignee
- International Business Machines Corporation
Inventors
- Vardy Amdursky
- Ilan Efrat
- Shlomo Shlafman
Application Information
Application Number:US 6906583
Filing Date:1996-07-31
Priority Date:1993-02-08
Art Unit:242
Classifications
IPC:
G06T 1700
Field of Search:
395364119;120754;505;508;572
Patent Drawings (2 sheets)
Description
Summary of the Invention
Disclosed herein is an extension of the ordinary preconditioned conjugate gradient method which enables solutions for a plurality of loadings and which avoids the necessity of "starting from scratch" each time. The invention, in one aspect, is a technique for converting a stored geometrical representation of the structure, the material properties thereof and the loads imposed thereon into a visualization of a mechanical quantity of the structure. Initially, a mesh is generated from the stored geometrical representation of the structure. The technique then includes generating, by application of the finite element method to the mesh, the elements of the stiffness matrix and a loading vector; generating from the stiffness matrix and the loading vector a matrix A and a right hand-side vector f, where A and f are related through the relation Ax=f, the vector x representing the mechanical quantity at a plurality of points of the mesh; generating a preconditioner to accelerate the solution to Ax=f; approximating solutions of a plurality of problems Ax.sub.k =f.sub.k, the plurality of problems having a common matrix A and a plurality of different loadings/right hand-sides f.sub.k, including recursively constructing for each k an A-orthogonal basis W(k) using W(1), W(2) . . . W(k-1), and obtaining in approximation to x.sub.k using W(k); and providing a user the solutions x.sub.k.
In another aspect, the invention relates to a technique for generating a numerically stable A-projection y.sub.q of x upon an A-orthogonal basis sequence W, which includes: partitioning the A-orthogonal basis sequence W into a plurality of subsequences; calculating an A-projection of x upon a given subsequence of the plurality of subsequences using a residual value r.sub.i-1, including calculating a subsequent residual vector r.sub.i, for i=1; accumulating the A-projection into a cumulative A-projection y.sub.i of x, for i=1; repeating the calculation of the A-projection for i=2 . . . q upon respective subsequent subsequences of the plurality of subsequences; and repeating the accumulating such that the numerically stable cumulative a-projection y.sub.q of x upon W results having associated therewith a numerically stable residual vector r.sub.q.
By using the above techniques, a fast solution is provided when switching from one loading to another, while the stiffness matrix remains unaltered. This procedure eliminates the need to "start from scratch" each time the loading is replaced by a new loading, which is typical of many of the iterative solution techniques currently available. The disclosed technique therefore considerably shortens the processing time required for the entire design phase.
Brief Description of the Drawings
An embodiment of the invention will now be described, by way of example only, with reference to the drawings, wherein:
FIG. 1 is a schematic diagram showing a computer aided design system according to the invention;
FIG. 2 is a schematic diagram showing the FEM analysis module;
FIG. 3 is a schematic diagram showing the solver;
FIG. 4 is a schematic diagram showing the linear solver; and
FIG. 5 is a flow chart showing the generation of the preconditioning matrix .
Detailed Description of the Invention
FIG. 1 is a schematic diagram of a computer aided design system according to the invention. The geometry of a structure, finite element mesh information, elements information, material properties, Boundary conditions and loading information are stored in storage device 10 under the control of Data Base Management System (DBMS) 20. This data is input via a front-end graphical interface 30. A display system 40, which includes a visual display unit and a printer, is used for displaying images of the structure and the stress & strain tensor, as well as the displacements (deformation), at prescribed portions of the structure using appropriate data visualization techniques. The display system also includes software for relevant report generation for display or printing.
The information stored in the database is used to control numerical control module 50 which generates numerical control code for controlling machine tools used in manufacturing the structure being designed. Tool design module 60 handles the design of the tools required to make the structure. Inspection and testing module 70 generates data for use in inspection and testing of the manufactured structure. The general operation of these modules is well understood in the art and will not be further described herein.
The CAD system also includes a finite element method (FEM) module 80 which enables a stored structure to be tested and various of its properties to be determined using finite element analysis.
FIG. 2 is a schematic diagram showing the Finite element analysis module 80 and the interface with DBMS 20. FE module 20 comprises a graphical mesh generator 90 to generate the division of the structure into elements, equation generator 100 for generating the system of algebraic equations which describe the behavior of the structure, and solver 110 for solving the generated equations to calculate the desired property. Discretization of the problem by finite element method (FEM) leads to a linear system of equations, generated by the equation generator, Ax=f, where A=(A.sub.ij) is a stiffness matrix, f=(f.sub.i)--right hand side (r.h.s) vector and x=(x.sub.i)--vector of unknowns i, j=1,2, . . . ,n.
In structural analysis, it is generally necessary to solve the following types of elastic problems, where the displacements of parts of the structure are unknown parameters:
a) Linear Equilibrium problems;
b) Non-Linear Equilibrium problems;
c) Dynamical problems which include forces of inertia; and
d) Eigenvalue/eigenvector problems--Normal mode dynamics.
Various theories of elasticity can be used to formulate these problems, all of them resulting in a system of elliptic partial differential equations. For dynamical problems, the space derivative part is elliptic. Various types of finite elements can be used, including Serendipic and P-elements.
When the finite element method is applied to the relevant system of partial differential equations which corresponds to a linear equilibrium problem, one obtains, a system of linear algebraic equations, in which a stiffness matrix multiplies the vector of unknown displacements, while a certain right-hand side vector prescribes the loadings information.
In non-linear equilibrium problems as well as in various dynamical problems, the FEM discretization leads to a system of algebraic equations, and the application of various iterative techniques devised for their solution results, again, within each iteration, in a similar system of linear algebraic equations.
There exist a maximum of 6 degrees of freedom at each point of an elastic body:
u.sub.1, u.sub.2, u.sub.3 --translations; and
u.sub.4, u.sub.5, u.sub.6 --rotations.
Each entry x.sub.i of the solution vector x corresponds to a node j and a degree of freedom m .epsilon. (1,2, . . . , 6). A vector L of degrees of freedom is generated which has, as i-th component, that value of m which corresponds to the component x.sub.i.
Solver module 110 is shown schematically in FIG. 3 and comprises two parts--an iteration mechanism 120, for handling non-linear and dynamic problems, and a linear solver 130 for solving both linear equilibrium problems as well as the linear systems arising within each iteration in solving non-linear and dynamic problems.
Linear solver 130 is the critical element in the operation of the whole CAD system. It must be capable of handling Single & Multi-Point Constraints, be efficient and sufficiently accurate even handling large condition numbers and maintain its efficiency in solving for cases with variable loadings, or arbitrarily specified multi-loading systems, by avoiding, as far as possible, the need to start from scratch whenever a new loading-system is being treated.
FIG. 4 is a schematic block diagram illustrating the linear solver. It takes as input the stiffness matrix and the right hand side vector generated by equation generator 100. From these a preconditioning matrix is generated and, using this preconditioning matrix, the preconditioned conjugate gradients method is applied using a suitable computer subroutine to solve the equations generated by the equation generator. The preconditioned conjugate gradients method was originated by R. Hestenes, E. Steifel, `Methods of Conjugate Gradients for Solving Linear Equations,` Journal of Research of the National Bureau of Standards, 49, 409-435 (1952). The present form of PCG is described in the book by Gene H. Golub, Charles F. Van Loan, `Matrix Computations,` John Hopkins University Press (1984).
In order to solve the equation Ax=f, the PCG method with the preconditioner K proceeds as follows. A trial solution x.sub.0 is selected. Vectors r.sub.0 and P.sub.0 are defined as r.sub.0 =f-Ax.sub.0 and P.sub.0 =K.sup.-1 r.sub.0.
The following steps are then performed iteratively until convergence is achieved:
where the expression in the form of (Vector.sub.-- x, Vector.sub.-- y) represents the scalar product of the vectors Vector.sub.-- x and Vector.sub.-- y.
A suitable efficient matrix-vector multiplication, which can be adapted for the particular computer being used, is used for the multiplication operations.
FIG. 5 shows the steps performed in the generation of the preconditioning matrix. A `principal` submatrix B=(b.sub.ij) is extracted from the stiffness matrix A=(a.sub.ij) by setting to zero those elements in the stiffness matrix for which the degree of freedom index associated with their row is not equal to the degree of freedom index associated with their column, using a method described in the following pseudo-code:
do i=1, n
do j=1, n
if l.sub.j =l.sub.i then b.sub.ij =a.sub.ij
else b.sub.ij =0
enddo
enddo
Where l.sub.i are components of vector L. The resulting matrix is a matrix in which the coupling between certain degrees of freedom is eliminated.
Following both a theoretical and experimental verification, the inventors have found that by using this process either the matrix B defined above or the matrix B' can be used as preconditioning matrices for the stiffness matrix, where B' is the result of applying the incomplete LU factorization to the matrix B.
The incomplete L-U factorization is obtained as follows. If it is desired to solve the matrix equation Ax=f, where A is a large sparse nxn matrix, choosing a Gaussian elimination would lead us to a factorization A=LU. However this method has the problem that the number of non-zero elements is much greater than the number of non-zero elements in the sparse matrix A, and since only the non-zero elements are stored this results in a high cost in terms of storage and complexity.
To avoid these difficulties an incomplete factorization technique is used to obtain a splitting A=LU--R, where the fill-in has been controlled by zeroing previously determined coefficients in L and U during the elimination. Such a technique is described in D. S. Kershaw, `The Incomplete Cholesky-Conjugate Gradient Method for the Iterative Solution of Systems of Linear Equations,` Journal of Computational Physics 26, 43-65 (1978).
In the case of a boundary value problem for an elastic body with rigid inclusions, the problem is usually formulated as a boundary value problem for an elastic body with holes instead of rigid inclusions, but with multipoint constraints generated by the rigid inclusions. It is known that a rigid element has only 6 independent degrees of freedom in 3D problems (3 in 2D). For example it may consist of the degrees of freedom of a single node.
Other degrees of freedom for 3D rigid elements can be found by the kinematic relations,
where u(u1,u2,u3) is the vector of displacements and .omega. (u4,u5,u6) is the vector of rotations, j is any node with dependent degrees of freedom and r.sub.j0 is a pointer vector from point 0 to point j.
If the total number of degrees of freedom g consists of n independent degrees of freedom and m dependent degrees of freedom, then the relation between the dependent degrees of freedom and the independent ones can be written in matrix form as u.sub.m =G.sub.mn u.sub.n, where G.sub.mn is an m.times.n rectangular matrix.
The saddle point formulation and finite element method are used for discretization of the problem, giving rise to the following algebraic equation, ##EQU1## where K.sub.gg denotes the FEM matrix for the elastic body with holes replacing the rigid inclusions. This will be referred to as the `hole stiffness matrix.` G.sub.mg is the matrix of multipoint constraints, P.sub.g is the loading vector and .lambda. the vector of Lagrange multipliers. I.sub.mm is the m.times.m unit matrix, and G.sup.t.sub.mg denotes the transpose of G.sub.mg.
In this case a principal submatrix K.sup.d.sub.gg is extracted from the hole stiffness matrix K.sub.gg using the method described above. It has been found by the inventors that a reduced resolving matrix
where ##EQU2## or the result of applying the incomplete factorization to R.sub.nn as above can be used as a preconditioner for the resolving matrix K.sub.nn in the equation,
This equation can then be solved in the same way using the preconditioned conjugate gradients method.
In practice, many applications of structural analysis require a system of algebraic equations to be solved for several loading vectors/right-hand sides. For example, this situation arises in linear problems with several loadings, in the inverse power method for normal mode dynamics and in iterative methods for non-linear problems.
One way of achieving this is by treating each system separately, for each i, i=1 . . . k, via PCG, using the above preconditioner. However, this would be expensive in terms of computer time.
These situations are handled in this embodiment of the invention in the following manner, which is both efficient & numerically reliable.
If the system to be solved is Ax.sub.k =f.sub.k, for k=1, 2, 3 . . . N, the linear problem for the first right hand side vector f.sub.1 is solved using the method described above. The sequence of coordinate vectors to the base sequence W=P(1)={p.sup.1.sub.1, p.sup.1.sub.2 . . . , p.sup.1.sub.m1 } obtained during the PCG procedure described above are stored. For the subsequent right hand side vectors f.sub.k the solutions x.sub.k are found in two steps. In the first step the A--orthogonal projection y.sub.k of x.sub.k on the space generated by the basic sequence W, i.e. the span W, as well as the corresponding residual r.sub.k are calculated. In the second step, the linear problem Az.sub.k =r.sub.k is solved using the PCGM algorithm described below. The sequence of coordinate vectors P(k)={p.sub.1.sup.k, p.sub.2.sup.k, . . . , p.sub.mk.sup.k } is now added to the basic sequence W to obtain an updated extended basic sequence. The approximate solution for x.sub.k is given by x.sub.k =y.sub.k +z.sub.k.
A straightforward calculation of the A-orthogonal projection y.sub.k of x.sub.k on the current span W, i.e., ##EQU3## suffers from loss of precision, since small inner products of relatively large vectors are hard to resolve accurately. Thus in the preferred embodiment a cycling scheme is used for computing the projection y of x on the W, as follows.
A parameter s is chosen between 1 and m. A quantity q is defined as q=m/s!. Variables y, y and y are initialized to zero. r.sub.0 is set to f.sub.k. For i=1, 2, . . . , q, the following steps are performed.
Then set y=Y.sub.q and r=r.sub.q.
If the parameter s is increased, the accuracy improves but more calculations are required. A typical good order of magnitude for s is 10.
The purpose of PCGM is to obtain--in an efficient & reliable way--the solutions x(1), . . . , x(k) to the systems of equations Ax(i)=f(i), (i=1 . . . k), corresponding to the loading vectors f(1), . . . f(k).
The PCGM consists of k subsequent STAGES, each depending on the results of the previous stages.
Each stage (i) produces the following:
1) A set of vectors P(i), which is added to a previously-constructed set W(i-1), so as to form the new set of vectors W(i). The way by which the vectors in P(i) are constructed is provided below, but it is important to emphasize that as a result of this construction, ALL the vectors in W(i) come out to be MUTUALLY A-ORTHOGONAL. By this is meant that if u,v are any two vectors in W(i), then u.sup.t Av=0, where here "t" stands for the "transpose" operation for vectors.
2) An approximate solution x(i) to the system Ax(i)=f(i).
The details of this construction are as follows: The set of vectors W(1) and x(1) are constructed via PCG (Preconditioned Conjugate Gradients) applied to the matrix A and f(1), and using the above-described preconditioner.
Having completed the i-th stage, the i+1 stage starts by construction of a vector y(i+1)--which is the numerically stable A-projection of x(i+1) on SPAN W(i)--and is done via the "PROJECTION" cyclic method, as described above.
Then, a corresponding numerically stable "residual" r(i+1) is formed, which is in fact provided by:
The various vectors in P(i+1) are now constructed in the same way this is done in ordinary PCG, using the above described preconditioner,--with the following exception:
Let v(1), . . . , v(i) each denote the LAST vector in P(1), . . . P(i) respectively. Then, once a vector Q' is constructed as one of the `conjugate-directions` obtained during the PCG process--it is REPLACED by the vector ##EQU5## (In order to orthogonalyze Q' on W(i))
This new vector Q is now added to P(i+1), all the vectors of which are stored. It is THIS operation which assures the mutual A-orthogonality of all the vectors in W(i+1).
The construction of an approximate solution x(i+1) then follows the procedure used in PCG. All the vectors in P(i+1) are now added to W(i), so as to form W(i+1), which is stored.
The difference from the repeated, 'from scratch'approach is that the "conjugate directions" constructed at each stage serve also in obtaining the solutions for the next loading/stage. Furthermore, within each stage, the formation of "A-PROJECTIONS" and "CONJUGATE DIRECTIONS" are modified considerably, but these are necessary in order to obtain numerical stability and efficiency in computational speed.
The PCGM algorithm ensures that the basic sequence W is orthogonalised by imposing that in the conjugate gradients procedure all the vectors of P(i)={p.sup.i.sub.1, p.sup.i.sub.2 . . . p.sup.i.sub.mi } are orthogonal only to the last vectors of P(k), k=1, 2, . . . , i-1. The procedure is as follows.
x.sup.0 is set to y, r.sup.0 is set to r and p.sup.0 is set to K.sup.-1 r.sup.0, where K is the preconditioning matrix. An index j is initialized to 1.
For k=i the following steps are performed.
(1) p.sup.j-1 is set as Q above
(2) A value a.sup.j is set to (K.sup.-1 r.sup.j-1 r.sup.j-1)/(p.sup.j-1, Ap.sup.j-1), and x.sup.j is set to x.sup.j +a.sup.j p.sup.j-1 and r.sup.j to r.sup.j-1 -a.sup.j Ap.sup.j-1.
(3) A value .beta..sup.j is set to (K.sup.-1 r.sup.j, r.sup.j)/(K.sup.-1 r.sup.j-1, r.sup.j-1) and p.sup.j is set to Kr.sup.j +.beta..sup.j p.sup.j-1 and then to ##EQU6##
(4) If a stopping criterion is satisfied, then m.sub.k is set to j, and P(k) is set to , and, {p.sub.1.sup.k, p.sub.2.sup.k, . . . , P.sub.mk.sup.k } ##EQU7## otherwise j is set to j+1 and the process repeated.
The invention can be implemented as a set of tools for a general digital computer, to serve in structural-analysis packages, which yield fast & robust iterative solutions, especially for medium and large 3-D problems, and enable the Multi-Point Constraints and the several loadings cases to be handled efficiently when the number of Loading-Systems being solved does not exceed certain limits, which, with present-day machines seems to be around several tenths. These tools enable performance advantage to be taken of low-accuracy specs, and their storage requirements are very modest.
The inner structure of most structural analysis packages makes them ideal candidates for incorporating a solver which operates according to the invention. The packages are large, and are constructed from several modules that communicate information through a common interface. More so, because typical problems take considerable machine resources, most packages would have restart facilities, and additional tools for enhanced end-user input.
Even though the Solver module is only one among the several modules that makes the structural analysis package, it is the limiting one in terms of computer resources like memory, disk space and machine time. Incorporating a solver in accordance with this invention into the package will enable the user to execute larger problems using considerably less resources than those required by Direct Solvers.
Claims
What is claimed is:
1. A computer-aided design system for converting a stored geometrical representation of a structure, the material properties thereof and the loads imposed thereon into a visualization of a mechanical quantity of the structure, said system comprising: means to generate a mesh from said stored geometrical representation of the structure, the material properties thereof and the loads imposed thereon, said mesh describing the structure; means to generate, by application of the finite element method to the mesh, the elements of a stiffness matrix, and a loading vector; means to generate from the stiffness matrix and the loading vector a matrix A and a right hand-side vector f, and where A and f are related through the relation Ax=f, the vector x representing the mechanical quantity at a plurality of points of the mesh; means to generate a preconditioner to accelerate the solution to Ax=f; means for approximating solutions of a plurality of problems Ax.sub.k =f.sub.k, said plurality of problems having a common matrix A and a plurality of different loadings/right hand sides f.sub.k, including: means for recursively constructing for each k an A-orthogonal basis W(k) using W(1), W(2) . . . W(k-1), and obtaining an approximation x.sub.k to x.sub.k using W(k); and means for providing a user of said system said solutions x.sub.k.
2. The computer aided design system of claim 1, wherein the means for recursively constructing for each k includes: means for constructing a preconditioning matrix K; preconditioned conjugate gradient means for obtaining said W(1), W(2) . . . W(k-1); means for calculating a numerically stable A-projection y.sub.i of solution x.sub.i upon basis W(i-1); means for approximating a solution z.sub.i of Az.sub.i =r.sub.i including: preconditioned conjugate gradient means for constructing an A-orthogonal basis P(i) wherein each of the vectors of P(i) is A-orthogonal to each of the last vectors in W(1), W(2) . . . W(i-1) respectively, and means for forming the approximation z.sub.i to z.sub.i using P(i).
3. The computer aided design system of claim 2, wherein the means for recursively constructing for each k further includes: means for adding y.sub.i to z.sub.i to obtain x.sub.i ; and wherein the system further includes: means for obtaining W(i) by joining all the vectors in P(i) to W(i-1).
4. The computer aided design system of claim 3, wherein the means for calculating a numerically stable A-projection y.sub.i of solution x.sub.i upon said basis W(i-1) includes: means for partitioning the A-orthogonal basis sequence W(i-1) into a plurality of subsequences; means for calculating an A-projection of x upon a given subsequence of the plurality of subsequences using a previous residual value r.sub.j-1, including means for calculating a subsequent residual vector r.sub.j, for j=1; means for accumulating said A-projection into a cumulative A-projection y.sub.j of x, for j=1; means for repeating said calculating an A-projection for j=2 . . . q upon respective subsequent subsequences of the plurality of subsequences; and means for repeating said accumulating such that a numerically stable cumulative A-projection y.sub.q of x upon W results having associated therewith a numerically stable residual vector r.sub.q, where y.sub.i is set to y.sub.q.
5. A computer-aided design system for converting a stored geometrical representation of a structure, the material properties thereof and the loads imposed thereon into a visualization of a mechanical quantity of the structure, said system comprising: means to generate a mesh from said stored geometrical representation of the structure, the material properties thereof and the loads imposed thereon, said mesh describing the structure; means to generate, by application of the finite element method to the mesh, the elements of a stiffness matrix, and a loading vector; means to generate from the stiffness matrix and the loading vector a matrix A and a right hand-side vector f, and where A and f are related through the relation Ax=f, the vector x representing the mechanical quantity at a plurality of points of the mesh; means to generate a numerically stable A-projection y.sub.q of x upon an A-orthogonal basis sequence W, including: means for partitioning the A-orthogonal basis sequence W into a plurality of subsequences; means for calculating an A-projection of x upon a given subsequence of the plurality of subsequences using a previous residual value r.sub.i-1, including means for calculating a subsequent residual vector r.sub.i, for i=1; means for accumulating said A-projection into a cumulative A-projection y.sub.i of x, for i=1; means for repeating said calculating an A-projection for i=2 . . . q upon respective subsequent subsequences of the plurality of subsequences; and means for repeating said accumulating such that the numerically stable cumulative A-projection y.sub.q of x upon W results having associated therewith a numerically stable residual vector r.sub.q.
6. The computer aided design system of claim 5, wherein the basis sequence W comprises {u.sub.1, U.sub.2, . . . , u.sub.m }, and wherein the means for partitioning the basis sequence W into a plurality of subsequences includes means for partitioning the basis sequence W into q subsequences, each of the subsequences comprising {u.sub.j1 . . . u.sub.j2 }, where: and
7. The computer aided design system of claim 6, wherein said means for calculating, said means for accumulating, said means for repeating said calculating, and said means for repeating said accumulating are combined in means for performing the following operations for each index i from 1 to q: ##EQU8##
Patent Citations (1)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US5136538 | 1992-08-01 | Karmarker et al. |
Non-Patent Literature (3)
- Ralston, A. (Ed.), Encyclopedia of Computer Science and Engineering, Van Nostrand Reinhold Company Inc., pp. 631-632, 927-932, 1983.
- McCallum, P., "Getting Better Results from Linear FEA", Machine Design, v. 64, n. 2, pp. 123-126, Feb. 20, 1992.
- Saad, Y., "On the Lanczos Method for Solving Symmetric Linear Systems with Several Right-Hand Sides", Mathematics of Computation, v. 48, n. 178, pp. 651-662, Apr., 1987.