CMM-2003 Computer Methods in Mechanics June 3-6, 2003, Gliwice, Poland

Problems of computational mechanics related to finite-element analysis

of structural constructions

Anatoly V. Perelmuter and Sergiy Yu. Fialko
*

Software company SCAD Soft
13, Chokolovsky bld., room 508
Kiev, 252680 GSP, Ukraine
e-mail: [email protected]
Abstract

The problems of computational mechanics, concerning with application of finite element analysis to structural constructions, are
discussed. Our attention is addressed to medium-class software for personal computers with which structural constructions are
usually analyzed. The complexity of a system and simplicity of its components, the large-scale of finite element problem, the
heterogeneity of finite elements and its coupling, the estimation of correctness of finite element model, the problems of seismic
analysis, the problem of indeterminacy and so on are the objective of this work.
Keywords: finite element analysis method, structural engineering, seismic analysis, fast solvers

1. Introduction
The contemporary market of industry-oriented software for
structural strength analysis impresses very much by its
versatility and widest functionality. There are real giants on this
market such as ANSYS, ADINA, COSMOS, MSC NASTRAN
and others, not restrained to any particular field of application
but oriented at large-scale problems. A special place in this
sphere is occupied by software intended for analyzing and
designing structural constructions SAP 2000, SCAD,
GTSTRUDL, Robot-Millennium etc. These we will call
medium-class software. Programs like these succeed in
providing features especially appreciated by structural engineers
such as graphical preprocessors and postprocessors,
catalogues of profiles, materials, regional climatic regulations.
They include specific analysis options (construction of
influence lines, seismic analysis etc.) Other special-purpose
software can be mentioned, too, particularly programs oriented
at narrow classes of problems or tutorial purposes.
Our attention will be addressed to medium-class software
for personal computers with which structural constructions are
usually analyzed. Problems related to the analysis of this kind
have their peculiar flair that affects the structure and
functionality of a computer program. There are certain
requirements to analytic methods employed by the software,
too.
Another important circumstance is that the software of this
type is oriented commonly at the level of expertise possessed by
a design engineer rather than a scholar researcher. Therefore the
software should have an intuitive interface and highly
automated functions. These programs should also account for
specifics of the management of structural design activities. In
particular, a typical form of organization used in this industry is
a design team that includes a lead analyst who solves
complicated problems of general nature and a few engineers
who prepare data and solve series of more specific problems.
The latter use simpler satellite programs interfaced with the
main application administered by the analyst.
Our discussion is based on the experience of development
of the SCAD Office [6] and Robot Millennium [16] software
because the authors of this report are members of their
development teams, participate in the support of the software,
and are familiar with both the architecture and functionality of
these programs.
2. Peculiarities of computational analysis in structural
engineering
2.1. Complexity of a system and simplicity of its components
Objects of structural engineering are residential and
public buildings, bridges, tanks, television towers, industrial
buildings, and a great variety of other types of structures (Fig.
1). Civil buildings belong to most widely spread objects of
construction.
All these objects, though much different, have
common peculiarities in their design models:
Bar elements are used extensively in structural models,
unlike most objects of mechanical engineering. Even if the
shape of a structure seems sophisticated, its load-bearing
framework may consist of elements of relatively simple
geometrical configurations. A characteristic example of
this is one of most whimsical buildings, the Gugenheim
Museum in Bilbao (Fig. 2). All the more so, this statement
relates to most of the objects shown in Fig. 1. Design
models of industrial, residential and public buildings, in
their vast majority, consist of sets of rectilinear bars, plates,
and flat shell elements. The latter have rectangular
configurations, as a rule, or contain a number of
rectangular sub-areas. Therefore the civil-engineering-
oriented software systems deal very little with spatial finite
elements which are often used to analyze FEM models of
mechanical engineering objects (an example is a popular
program SolidWorks [2].) These peculiarities of structural
construction models beg for special software to be
developed.
High dimensionality of models required by complicated
geometrical shapes of walls and floors, the use of
automatic mesh generators and an object approach which
treats a structure as a set of story, wall, floor etc. objects.
A noticeable spreading of stiffness properties which causes
an ill conditionality of the respective mathematical model.


2
Joints between elements of different types with different
numbers of degrees of freedom in a node. Often enough, this
circumstance brings the necessity to regularize the models
equation system, and this causes the ill conditionality again.
)

b)
e)

f)

c) d)
g)

h)



Figure 1: Examples of structural
objects: () a residential house; (b) a
bridge; (c) a skyscraper; (d) a
television tower; (e) an industrial
building; (f) an office building; (g) a
tank; (h) an arc of a hangar


)

b)

Figure 2: The Gugenheim museum: (a) its appearance; (b) its load-bearing framework

2.2. Model dimensionality
The structural analysis may involve models and schemes
which are quite typical and by no means break records by
containing 20 to 30 thousand nodes, 30 to 50 thousand
elements of various types (bars, plates, shells, elastic links),
and possessing over one hundred of stiffness property sets. 15
to 30 different loading patterns are usually under
consideration, each one including hundreds of components of
nodal or distributed loads. The dimensionality



of the model grows drastically if one has to analyze load-
bearing constructions of a structure jointly with its soil bed.
An example of this kind is shown in Fig. 3,a where the model
of a structure includes 27,138 equations while the soil
structure model consists of 319,133 equations.




3

Figure 3: A design model of a structure together with its foundation

Usually, this class of problems can hardly be solved by
direct methods because the structure of the adjacency graph
with its root in a pseudo-peripheral node is not extended. In
the example given above, we did not manage to factor the
stiffness matrix using a PC Pentium III (CPU Intel 1000
MHz, RAM 512 MB) because the multi-front method with
nested section ordering required 1292 MB of RAM to store
the maximum front, while the skyline method (RCM
ordering) demanded over 20 GB of disk storage. So, this
problem was solved by iterative methods.

2.3. Heterogeneity of finite elements, problems with
matching those
A plethora of complexities with the creation and
verification of design models are related to a typical
heterogeneity of finite elements often encountered in this
class of computational analyses. It is only a rare case that the
whole structure is represented by elements of the same type
(such as plates). Most often, a single design model includes
bars, plates and other finite elements at the same time.
It is a must for an advanced computational software
system to allow nearly every possible combination of finite
elements of most various types, dimensionalities, sizes and
shapes, different stiffness properties. There are a lot of
dangers here, sometimes revealed and sometimes concealed.
The latter are especially hazardous.
A typical example can be an analysis of a spatial bar
framework together with its slab foundation. This kind of a
design model includes plate finite elements and bars attached
rigidly to the slab. The axes of the column bars should cross
the median surface of the slab in nodes of the finite element
mesh on the slab. If no additional measures are taken, the
design model described above will provide for a perfect
match both between vertical displacements of the slab and the
columns (perpendicular to the plane of the slab) and between
respective slopes in nodes where the plate and bar elements
join one another. Though, bending moments in sections of the
columns near the slab calculated by this model have nothing
in common with the true distribution of internal stresses.
To see this, imagine how the mesh is getting denser and
the user expects the computational results to become more
and more accurate. Though, starting from a certain scale of
the mesh, further densification will be lessening the absolute
values of the bending moments in the bars at points of their
attachment to the slab.
In the limit, as the maximum size of the mesh cell tends
to zero, these bending moments will tend to zero, too. This
means that the design model in question provides a hinged
rather than rigid connection between the framework elements
and the slab. The fact that the user does obtain some formal
nonzero bending moments with a particular finite element
mesh of his choice evidences just an error of discretization
and nothing more. But there is no reason at all to take the
discretization error for an intended credible result!
In the design model presented above, the bars transfer
concentrated bending moments to the slab. As is known, the
solution of this elasticity problem has a logarithmic
singularity in the slopes. Therefore the slope at the
concentrated bending moment application point tends to
infinity as the mesh becomes denser. Consequently, in order
for the work of the concentrated moment at the slope to be
finite, the bending moment itself must be zero. Thus, making
the mesh denser will force the numerical solution to tend to
the hinged column-to-slab attachment case.
Results presented below make a confirmation of that.
Lets consider a square slab clamped along its sides and
having a single column standing in the middle of the slab and
fixed to it rigidly. The free top end of the column is subjected
to an external concentrated force P directed along the global
axis X (Fig. 4,a). The bending moment in the bottom section
of the column will be a constant magnitude not depending on
the size of the finite element mesh because the system is
statically determinate with respect to the column.


4
The calculation of the displacement wn of the columns
free end in the direction of the load using different finite
element meshes (2n by 2n) shows that the deflection of the
column grows almost linearly as n is increasing above n > 32.
This results in an unlimited growth of the slope in the root
section of the column as the finite element mesh becomes
finer (see Fig. 4,b).
Of course, the fact is that it is the user who must be
responsible for correctness of his design model from the
viewpoint of mechanics and adequacy between the finite-
element model and the real structure. Though, there have
been made numerous attempts to prevent the said dangers by
means of software implementation. In particular, flat shell
elements resistant to drilling rotation have been introduced.
This has been claimed to solve the problem with a torsion in
an attached bar. Though, a detailed analysis shows that this
way may lead to serious errors.
A more involved consideration of this problem is
presented in the report [10] at this conference.

) b)
11,6
11,8
12
12,2
12,4
12,6
12,8
0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64

Figure4: A column fixed to a slab: (a) a schematic; (b) a displacement changing under the effect of the force


3. Estimation of correctness of input information
3.1. Trivial checks
It is known that the probability for an error in input data is
much higher in large-scale problems. Engineering psychology
researches state a power dependence of the human error
probability on the volume of information processed by the man.

Figure 5: A spectrum of rigidities

Any contemporary computational analysis program operates
fairly heterogeneous data which describe properties of finite
elements, nodes, loads etc. The heterogeneity of the input
information is of especial complexity for software systems
oriented at performing structural design tasks. It is important for
one to be able to detect and extract deviations from a common
relationship, for example, analyze the spread of rigidity
properties and represent those as a spectrum of rigidities (Fig.
5).
3.2. Validation of kinematical stability in the course of the
matrix decomposition
Checks can be performed in the course of solving a
problem, too. Errors in the model can be detected right during
the solution, particularly such as a kinematical instability and
the lost of positive definiteness of the matrix (in cases there
must be one).
The presence of the kinematical instability is evidenced by a
substantial reduction of the governing element of the matrix
comparing to the respective diagonal element before the
factoring. Though, the detection of a nodes No. and the
respective degree of freedom in the node based on this method
is not always successful.

3.3. Detection of kinematical instability and other errors in a
design model, visualization of kinematical mechanism schemes
based on the natural oscillation analysis
To detect possible mechanism-type motions in a structure,
the SCAD software implements a special analysis mode based
on a Lanczos block method and involving spectral
transformations. The idea of this option is to use a shift
technique which enables one to analyze unconstrained systems
and even mechanisms. In the latter case the systems stiffness
matrix K is singular. Though, the B K K

= matrix is
not singular provided that the shift is chosen appropriately,
and it can be factored. In particular, the matrix B can be
assumed equal to the mass matrix M . Then, the presence of
zero natural frequencies in the model 0
2
=

M K
evidences the kinematical instability, and natural modes that
conform to zero frequencies describe possible movements of the


5
mechanism strictly. included in their original form and will be
reproduced in black and white.
This approach enables us both to determine whether a
particular system is kinematically unstable and to visualize
modes of mechanism-type motion, thus giving the user a hint
where to install additional constraints in order to eliminate the
instability. Fig.6 shows a fragment of a structure that contains
an unconstrained solid body, and one of its possible
mechanism-type modes of motion.
A more detailed presentation of the same material can be
found in the report [13].


Figure 6: Detecting a rigid-body mode of motion

4. A specific type of analysis seismic analysis
4.1. Sum of modal masses and local modes
It is known that the sum of modal masses is a criterion
whether the number of natural modes taken into account is
sufficient. Current seismic regulations require that the sum of
modal masses along each direction be at least 90%. In many
cases this requirement can hardly be met. It often appears that
the lower part of the spectrum includes local oscillation modes
that contribute only a little to the response of the whole system
as it moves during the seismic event. Some kinds of structures
do not even contain oscillation modes which contribute much to
the seismic movement of the system. One finds that in those
structures small pieces of its seismic response are distributed
over a large number of natural modes. In this connection, there
arises an enormous computational problem how to determine
100 to 500 or even more natural frequencies and oscillation
modes.
4.2. Development of specific algorithms Ritz vectors, a
residual mode, a seismic mode etc.
To escape from this situation, the following measures are
usually taken. E. Wilson suggests in his papers that
decompositions by natural modes should be replaced by
decompositions by specially constructed Ritz vectors, more
informative ones for the purpose of seismic analysis [14]. This
method is implemented in the SAP2000 software. A similar
technique has been used in the Robot Millennium software [15].
Another trick is to use a residual mode (a pseudo mode).
Apparently, it should be admitted that the problem has no
easy solution for today. The matter is that the solution of a
dynamics problem in terms of Ritz vectors is strict only if
motion equations are integrated directly. At the same time, most
dynamical problems in the structural engineering practice are
solved by the spectral technique. The latter has some immanent
contradictions when using a basis different from the set of
natural modes, so one may obtain noticeable discrepancies
between the calculated displacements/stresses and those
observed in reality.
4.3. Problem of data amount, filtering
In many cases one can reduce the computational effort
essentially by using only some chosen natural modes that
contribute much to the seismic response of the system. This
kind of filtering is based on an analysis of modal mass values
for each natural mode. This approach is implemented in the
Robot Millennium software [15, 16].
5. Requirements to the softwares response speed
5.1. Sources of requirements
Actually, the reduction of computation time has ceased to
be a critical issue in common structural design activities. The
usual relation of the effort and time is such that the most part of
business time (at least 80%) is spend for preparations and
verifications of input data, and then for reviewing and analyzing
results. Under these circumstances one may think that trying to
reduce the governing equation system solution time from twenty
minutes to two makes no real sense. Developers often express
this point of view, but we believe it is not true.
We deem it useful to indicate a number of practically
important problems in which the speed of solution of
linear/linearized governing equation systems is a really critical
point:
the design in an interactive dialog mode seems
impossible if the software responds too slow; a long and
painful waiting for an answer makes the interactive
design procedure almost fruitless;
in spite of the authors steady belief that an extensively
detailed design model is just a serious mistake of an
analyst, large-scale problems do arise in real design

6
practice for example, as long as a complex
structuresoil system is under consideration where
three-dimensional finite elements are used to simulate
the soil beds behavior;
the solution of nonlinear problems involves multiple
solutions of linearized systems of governing equations
at each step of a step-by-step procedure or at each
iteration;
optimization problems and related multi-variant searches
also lead to the need for multiple solutions;
problems with indeterminate parameters are posed and
solved more and more often recently; for these, one of
most universal techniques is an imitative modeling
(including the dynamical behavior modeling) which
requires the same multiple repetition of linear solution.
These and other related problem formulations can be feasible
only if the software used to perform the task is provided with
efficient algorithms for solving systems of algebraic equations
and eigenvalue calculation.
5.2. Construction of quick solvers
The quick solvers in the FEM analysis available for today
include direct methods for sparse matrices and highly efficient
iterative methods.
The advantage of direct methods is their low sensitivity to
ill-conditioned matrices and to the number of right parts (if
there are not too many), and the possibility to detect the
kinematical instability of the design model. The efficient direct
methods are based on a reduction of filling. The filling means
nonzero elements of the factored matrix standing in positions
where zeros used to be in the original matrix. The less filling is
reached, the higher the efficiency of a direct method. Therefore
the most responsible step for a direct method is the matrix
reordering. Until recently, the most popular reordering
algorithm was an inverse Cuthill McKie method employed to
decrease the profile width [17]. In recent years the most widely
spread methods have become the minimum degree algorithm,
the nested section method, and the multi-section method based
on domain decomposition [18]. Commercial FEM programs
implement direct methods for sparse systems, most often, on the
basis of a multi-front approach [18-22].
Iterative solvers are preferable for large-scale problems in
which the number of equations can be 100,000 to 700,000 or
more. Their drawbacks include a slower convergence in the case
of ill conditionality and a high sensitivity of the computation
time to the big number of right parts. The effective technique to
suppress the ill conditionality effect is a preconditioning.
Suppose a linear static problem b Kx = needs to be solved.
The preconditioning consists of a transition to another problem
b B Kx B
1 1
= where B is a preconditioning operator. If B
is positive definite, the system of equations
k k
r Bz = , where
k k
Kx b r = is a residual vector and k is No. of iteration,
can be solved much faster than the original system, and its
conditionality number ) ( ) (
1
K K B C C <

, then the
preconditioned problem will have a faster convergence than the
original one. In a limit case, when K B = , the preconditioned
problem converges to its exact solution in one iteration. So, the
trick of quick iterative methods is how to construct such
preconditioning which would not require much time and
resources and at the same time provide that 1 ) (
1

K B C .
The theory of iterative methods states that lower modes are
the slowest in convergence. The worse the conditionality of a
problem, the slower their convergence [23]. Hence the idea of
multilevel methods [29, 30]. The key point is to construct a
rough-level model intended for predicting low-mode
components of the solution. The convergence in high-mode
components is ensured by smoothing. The maximum effect is
usually achieved by combining the preconditioned conjugate
gradient method and the multilevel methods idea. This builds
up a family of conjugate gradient methods with multilevel
preconditioning.
Commercial FEM programs employ most often the
conjugate gradient method with preconditioning of an
incomplete Choletsky factorization type, multi-mesh methods,
algebraic multi-level techniques [24, 25], and aggregative multi-
level method [26], methods of space decomposition and
subspace correction [27]. A review of iterative methods is
presented in [28].
Talking about the eigenvalue problem, we should note that
currently most popular methods based on the stiffness matrix
factorization include the block subspace iteration method [14]
and the Lanczos block method [30, 31]. The implementation of
the Lanczos block method with shifts in the SCAD software
will be discussed in [13].
In cases when we cannot reorder the stiffness matrix
efficiently, it is reasonable to use methods not requiring the
matrix to be factored. The most efficient technique is the
conjugate gradient method with preconditioning [28]. Though,
the traditional algorithm of this method suffers from a
convergence suffocation in some cases [33]. This is overcome
by introducing shifts into the preconditioning [34, 35]. A more
detailed discussion of this problem will be given in the report
[36]. The report [37] presents a quick approximate method for
determining natural oscillation frequencies and modes. That one
is a Ritz method which makes use of a gradient procedure with
aggregate multilevel preconditioning to construct an orthogonal
system of basis vectors.
A modern FEM analysis software must implement both
quick direct methods and iterative ones because nobody can say
beforehand what method is going to be most efficient in a
particular case. For example, prominent programs like MSC
NASTRAN, ANSYS, ADINA include both direct sparse matrix
solvers and efficient iterative solver tools at their disposal.
Among civil-engineering-oriented software, we should notice
Robot Millennium which also implements both direct sparse
solvers and an aggregate multilevel iterative solver [15]. The
SCAD software that implements a multi-front solver is worth
mentioning, too.
6. Choosing a most disadvantageous combination of loads
The peculiarity of construction objects is that one has to
deal with a plethora of variations of loads applied to a structure.
This is an essential distinction from common engineering
where this problem is less sharp.
The fact is that even simplest buildings have tens of rooms,
and in each of those the useful load can be present or absent in a
particular moment. It is by no means obvious that the critical
case will be the fully loaded structure. Moreover, we can say for
sure that it is not the case for a good deal of structures. Then
add the necessity to account for a few possible directions of
wind or seismic loads, and numerous possible positions of
movable loads such as those caused by bridge cranes. All this
makes it quite clear how much effort the problem of choosing a

7
design load combination may take. One should note also that a
direct exhaustion of possible variations is difficult even when
there are only twenty or thirty independently acting loads.
In essence, one needs to solve some optimization problem
where one has to find an extremum of the structures response
in the set of possible loaded states of the system. This set may
be of a pretty complex build because some of the applied loads
can be related to one another via logical relationships of the
following types:
incompatible some loads cannot act together for purely
physical reasons, for example, a south wind cannot be
accompanied by a north wind, nor snow can be combined with
the maximum summer heat;
bound certain loads can be treated only as acting
together; this is often the case when different loads are of the
same physical origin and are presented separately only for
convenience;
accompanying one of loads cannot exist without the
presence of another, while the other way round is quite feasible:
for example, the bridge cranes braking force cannot exist
without the pressure of the cranes wheel, while the pressure
can exist without braking;
limited some of jointly acting loads cannot exceed an
established limit in total, for example, loads from bridge cranes
are limited to two cranes on a single pathway or in the same
section.
One of feasible approaches to the solution of the problem is
to represent the logic of interaction between different loads as
an oriented graph [9]. Then the problem can be formulated as a
known problem of searching a network for the biggest flow [8].
Lets give an example of such graph for the situation when
the following elementary loads can be applied:
1 dead weight;
2 snow;
3 wind from the left;
4 wind from the right;
5 maximum pressure of the crane onto the left column;
6 maximum pressure of the crane onto the right column;
7 braking of the cranes to the left transferred to the left
column;
8 braking of the cranes to the right transferred to the left
column;
9 braking of the cranes to the left transferred to the right
column;
10 braking of the cranes to the right transferred to the
right column.
Fig. 7 shows a schematic of the respective graph where
there are arcs 110 conforming to the elementary loads listed
above and four more arcs (dash lines in this figure) conforming
to zero values of the load intensity. These additional arcs enable
one to bypass those loads in the graph which must not
necessarily be included in a design load combination (that is,
which unload the structure).

Figure 7: A graph of the logical structure of loads

7. Problem of indeterminacy
Modern buildings or other structures are, most often,
complex structural multi-element aggregates created to perform
a plenty of different functions. During their lives the structures
go through a long sequence of various working states. The
specifics of structural engineering is such that its final product
(a building, a structure) must combine three features often
contradicting one another: functionality, aesthetics,
designability.
An idealization of the design model and impossibility to
make it a perfect reflection of the real structure create a
situation of some indeterminacy. It is these conditions of
indeterminacy under which design decisions have to be made.
The indeterminacy is caused by either unavailability of
required information (for example, we are not able to know all
future regimes of the structures operation) or its
incompleteness (we can hardly imagine knowing mechanical
constants precisely in any point of the structure). The
unavailability of some types of information and its
incompleteness are key points they cannot be overcome
altogether, and deeply as we could study the problem of our
interest, we may never say we have taken absolutely all into
account in our model.
Though, it is not only the unavailability and incompleteness
of data that causes the indeterminacy to appear. There is also an
ambiguity of the data, that is, a possibility to interpret the same
factors differently. This circumstance requires us to estimate all
possible alternatives. There are known classical approaches to
the indeterminacy which can be classified into the following
decision-making methods:
making use of the probability theory, the decision being
based on the objective earlier experience;
making use of expert estimations, the decision being
based on the subjective experience of an expert (or a
panel of experts);
a minimax estimation, when the best of achievable
solutions is adopted with the assumption of the worst
possible course of events, i.e. the decision is made by the
possible result.
All these options can be used together or separately. They
are intended to estimate the credibility of a design model. There
are other factors, too, which determine how approximate the

8
design model is and what errors, distortions, contradictions may
appear in it.
First, there are design modeling errors (approximation
errors) that appear due to either our knowledges approximate
nature itself or an intentional rough approximation of it. These
errors include using simplified mathematical representations
such as polynomials of low degrees for describing displacement
fields in the finite element analysis, truncation of series in the
Galyorkin method etc. The same category includes errors
caused by discrepancies between scientific theories and
assumptions that are used to simulate different parts of the same
design mode. A typical example is a discrepancy between
concentrated forces as popular models of loads, on one hand,
and plate finite elements, on the other hand. The latter cannot
balance the concentrated actions by finite values of their shear
forces. It is natural that totally mythical values of the shear force
in elements obtained by such calculation result directly from the
said discrepancy between the models.
Second, we should note the approximate character of nearly
all specifiable properties of a model. This is related to
tolerances for sizes, weights and other measurable magnitudes
existing in practice. From the practical viewpoint, both
inaccuracies stated above differ little. Though, in the first case
we deal with a limited accuracy of our simulation (either
intentional or unconscious) while in the second case it is the
original objects properties which cause the limited accuracy.
8. Postprocessing
8.1. Problem with understanding
Results of static or dynamical analysis of a complex system
represented as numbers contain vast arrays of data perceiving
and reviewing which is practically unfeasible. A selective result
printout option available in most programs is of little help, too,
because the analyst does not necessarily know which of the
values he expects to be critical.



)


b)


c)












Figure 8: An example of localized ultimate values of a factor
of interest: (a) isofields in the whole model; (b) displaying
only maximum values of the factor in the
transparent model using color markers; (c) isofields of
maximum values limited to a specific range.
A much better demonstrativeness can be achieved by
using graphical representations of results as curves, color
maps, isofields. These methods compress the data to a great

9
extent, and the information thus becomes more or less
apparent.
Though, even this technique is not always enough to
make a proper analysis, because the graphical information
can be still hardly accessible for the system as a whole (Fig.
8,a). A fragmentation of it will restore the demonstrativeness
but cause another problem how to find a particular
fragment at which specific results of the users interest have
been obtained [12]. Solution of this problem is not trivial at
all for a complex model consisting of tens or hundreds
thousands of nodes and elements. For example, Fig. 8,a
shows a design model with isofields of vertical displacements
drawn on it. Note that this figure does not show an area of
maximum values.
The way out of the situation can be a technique suggested
in the SCAD software. It is based on a control of color
indication and described in detail in the report [11] at this
conference. The point of it is that one uses the color map to
find the factor in question in the transparent model, first,
and then detects the location where the needed values appear
(Fig. 8,b). Next, the color indication is used only for a part of
the isofield that belongs to the range of interest, and all the
other levels are turned off (Fig. 8, c). In this way critical
results of the analysis are localized.
The general depiction given by the graphical
representation of the analysis results accords best with the
well-known statement that the goal of a calculation is an
understanding rather than a raw number. Having analyzed the
general picture, one should always turn to numerical results
that now can be selected from the common data flow
consciously.
When dealing with problems of buckling/stability, one
should be aware of a universal tool for visualizing the stress
and strain distribution in a system. This tool is a picture of the
deformation energy field. If the energy distribution has been
constructed with taking the geometrical stiffness matrix into
account, in this way one acquires the capability of classifying
particular fragments of the system (down to its separate
elements) into one of two following categories: either
restraining or pushing elements (parts) of the system [2]. The
restraining elements facilitate the stable equilibrium of the
system, while the pushing elements play a negative role
because they force (push) the mechanical system to buckle.
The role played by a particular subsystem is checked by
calculating the energy accumulated by this part as it deforms
by a buckling mode. For the system as a whole this energy is
zero. Parts where it is non-positive are pushing ones, while
those with a positive deformation energy can be classified as
restraining subsystems.
Based on numerical values of the energy, pushing
elements of the system can be ordered by the degree of their
blame for the critical state of the system. A contribution of
each of the systems elements to its total energy balance can
serve as a convenient quantitative measure of its
responsibility for the equilibrium stability.
8.2. Meeting requirements of design codes
In the course of the structural design procedure, results of
static and dynamic analyses of constructions are used to
estimate their strength and stability. This process is regulated
by design codes. Unfortunately, design codes are by no
means as strict and non-contradictory as computer mechanics
methods. These documents were initially created when
manual calculations reigned supreme, they absorbed informal
practical experience, and they are based on a great deal of
compromises.



Fig.9. A load-bearing ability area

The design regulations contain numerous empirical
relationships, correction coefficients, and simplifying
assumption which very often can be complicating
assumptions on the way of computer analysis. All these
together create certain problems for software developers,
namely:
how to match the level of accuracy of the analysis with the
accuracy of formulations found in the codes, because a
paradoxical situation often arises a system calculated
with a higher accuracy is inferior to a system calculated
approximately when compared in compliance with the
code requirements;

10
results of static and dynamic analyses need to be made
rougher and interpreted specifically to comply with terms
of the design codes;
concepts operated by the design codes need to be
introduced into finite-element analysis software. For
example, a beam or a column comprising certain groups of
finite elements must possess certain properties;
analysis results need to be specially processed to obtain
some properties of a structural object that cannot be
described in FEM terms, such as the tilt of a building or the
axis of elastic centers.
It is not only software developers who face the said
problems. Computer mechanics experts have to deal with them,
too, because the problems require specific methods to be
developed for their solution. We are presenting only one
example here for the purpose of illustration the issue of a
potential non-convexity of the load-bearing ability area of a
structures elements, in the case this ability must comply with
all design regulations (strength, stability, rigidity). Lets
consider a compressed and bent element of a steel structure and
determine its load-bearing ability area. This area is shown in
Fig. 9 for a steel bar with its design strength Ry = 2050 kg/cm2
and its effective length 600 cm in both principal planes.
The boundary of the load-bearing ability at the segments
AB and AH is defined by the condition of sufficient strength
under the combined effect of the tension and bending, at the BC
and GH segments it is defined by the stability of the plane
bending, and at the CD and GF segments (as well as at DEF) by
the stability out of the moments plane.
By itself, the non-convexity of the area in question may lead
to quite a few unpleasant effects. The most apparent of the
effects is related to the fact that traditional disadvantageous
stress combinations estimated by engineers either do not include
some actions or include them incompletely. In a non-convex
area, though, it is quite possible that the disadvantageous
combination occurs at some intermediate point. For example, if
one variation of loads conform to the C point while another to
the E point (in both cases the load-bearing ability is ensured),
then we can take halves of the limit moment and force and find
ourselves in the K point beyond the admissible area.
There arises a problem of this type find conditions under
which the convex surface of points depicting all possible
stressed states belongs to the load-bearing ability area defined
by structural design code requirements. As far as we know,
there has been found no satisfactory general solution of this
problem fitting for a practical software implementation.
9. Conclusion
The development of finite element software for structural
design creates lot of complex specific problems of
computational mechanics.
The experience of development and practical usage of
software for structural design requires the permanent
replenishment and analysis. The corresponding efforts should be
supporting by scientific community.
References
[1] Vorovich, I.I., Lebedev, L.P. Some questions of continuum
mechanics and mathematical problems in the theory of thin-
wall structures, International Applied Mechanics, Vol. 38,
No 4, 2002. pp. 2-21.
[2] Avedian, A.B., Danilin, A.N. Strength for strength
dummies, CAD and graphics,. No1, pp. 75-83; No2, pp. 63-
68; No3, pp. 39-46, 2000 (In Russian).
[3] Oden, J.T., Belytschko, T., Babuska, I., Hudhes, J.R.
Research Directions in Computational Mechanics,
Computer Methods in Applied Mechanics, Vol. 192, No. 7-
8, pp. 913-922, 2001
[4] Perelmuter, A.V., Slivker, V.I. Numerical Structural
Analysis: Models, Methods and Pitfalls, Springer Verlag,
Berlin-Heidelberg-New York-Hong Kong-London-Milan-
Paris-Tokyo, 2003.
[5] Vasiliev, V.V. To the discussion on the classical plate
theory, Mechanics of solids, N4, pp. 140-149, 1995 (In
Russian).
[6] Karpilovsky, V.S., Kriksunov, E.Z., Perelmuter, A.V. et al.
SCAD for users. Kompas Publishers, Kiev, 2000 (In
Russian).
[7] Grigolyuk, E.I., Shalashilin, V.I. Problems of Nonlinear
Deformation. Dordrecht et al.; Kluwer, 1991.
[8] Ford, L.R., Fulkerson, D.R. Flows in Networks, Princeton
University Press, Princeton, NJ, 1962
[9] Artemenko, V.V., Gordeyev, V.N. A program for
calculating rated stress combinations in conditions of a
complicated logical relationship between loads,
Computational and mechanization facilities in structural
design, No2, pp. 1014, 1967 (In Russian).
[10] Perelmuter, A.V., Slivker, V.I. Problems in matching finite
elements having different dimensionalities, Proceedings
CMM 2003
[11] Kryksuniv, E.Z., Perelmuter, A.V. Slivker, V.I. Techniques
to check properties of complex design models, Proceedings
CMM 2003
[12] Elizarov, S.V., Benin, A.V., Tananaiko, O.D. Modern
methods for analysis of engineering structures used in
railway transport. Finite-element method and the
COSMOS/M software, PGUPS, St.-Petersburg 2002. (In
Russian).
[13] Fialko, S.Yu, Kriksunov, E.Z. and Karpilovsky, V.S. A
block Lanczos method with spectral transformations for
natural vibrations and seismic analysis of large structures in
SCAD software, Proceedings CMM 2003
[14] Wilson, E.L., Three dimensional dynamic analysis of
structures, Computers and Structures, Inc., Berkeley,
California, USA, 1996.
15. Fialko, S.Yu., High-performance iterative and sparse direct
solvers in Robot software for static and dynamic analysis of
large-scale structures, Proceedings of the second European
conference on computational mechanics, Poland, June 26-
29, 2001, 18 p.
[16] Robot Millennium v. 15.0, Users Manual. pp. 244-278.
[17] George, A., Liu, J., Computer solution of large sparse
positive definite systems, Prentice-Hall, Inc., 1981.
[18] Ashcraft, C. and Liu, J.. Robust ordering of sparse matrices
using multisection. Technical Report CS - 96-01 Dept. of
Computer Science, York University. February 1996 to
appear in SIMAX
[19] Duff, I.S., Reid, J.K., The multifrontal solution of
indefinite sparse symmetric linear equations, ACM Trans.
Math. Software, vol. 9, pp. 302-325, 1973.
[20] Gend, P., Oden, J.T., R.A van der Geijn., A parallel
multifrontal algorithm and its implementation, Comput.
Methods Appl. Mech. Engrg., vol. 149, pp. 289-301, 1997.
[21] Fialko, S.Y. Method of nested substructures for analyzing
large-scale finite-element systems, applied to calculation of

11
thin shells with high ribs. Prikladnaya Mekhanika =
Applied Mechanics, vol. 39, No 3, p. 8896. (In Russian}.
[22] Fialko, S.Y. A multi-frontal method for solving large-scale
finite-element problems applied to calculation of thin shells
with massive ribs. Prikladnaya Mekhanika = Applied
Mechanics, vol. 39, No4. (In Russian).
[23] Bakhvalov, S.N., Zhidkov, N.P., Kobelkov, G.M.,
Numerical methods , Nauka, Moscow 1987 (In Russian).
[24] Axelsson, O., Vassilevski, P. Algebraic multilevel precon-
ditioning methods, I, Num.Math., 1989, vol.56, pp.157-177.
[25] Axelsson, O., Vassilevski, P. Algebraic multilevel
preconditioning methods, II, Num.Math., 1990, vol.57,
pp.1569-1590.
[26] Fialko, S. Yu. High-performance aggregation element-by-
element iterative solver for large-scale complex shell
structure problems. Archives of Civil Engineering, XLV, 2,
1999. pp.193-207
[27] Xu J., Iterative methods of space decomposition and
subspace correction, SIAM Review, vol.34: No 4, pp.581-
613, 1992.
[28] Papadrakakis, M, Solving largescale problems in
mechanics, John Wiley & Sons Ltd., 1993.
[29] Brandt, A., Multi-level adaptive solutions to boundary-
value problems, Mathematics of Computations, vol.31, No
138, pp. 333-390, 1977.
[30] Hackbush, W., Trottenberg, U., Multigrid Methods,
Springer-Verlag, Berlin, 1992.
[31] Ericsson, T. Ruhe, A., The spectral transformation Lanczos
method for the numerical solution of large sparse generalize
symmetric eigenvalue problem, Math. Comput., vol.35, pp.
1251-1268, 1980.
[32] Grimes, R.G. Lewis, J.G., Simon, H.D., A shifted block
Lanczos algorithm for solving sparse symmetric generalized
eigenproblems, SIAM J. Matrix Anal. Appl, vol.15, No1: pp.
1-45, 1994.
[33] Young Cho, Yook-Kong Yong, A multi-mesh,
preconditioned conjugate gradient solver for eigenvalue
problems in finite element models, Computers & Structures,
vol. 58, No3, pp. 575-583, 1996.
[34] Feng, Y.T., An integrated Davidson and multigrid solution
approach for very large scale symmetric eigenvalue
problems, Comput. Meths. Appl. Mech. Eng., vol.190, pp.
3543-3563, 1999.
[35] Fialko, S. Aggregation Multilevel Iterative Solver for
Analysis of Large-Scale Finite Element Problems of
Structural Mechanics: Linear Statics and Natural Vibrations.
LNCS vol. 2328, p. 663 ff,
http://link.springer.de/link/service/series/0558/tocs/t2328.htm
[36] Fialko, S. Yu. An aggregation multilevel iterative solver
with shift acceleration for eigenvalue analysis of large-scale
structures, Proceedings CMM 2003
[37] Fialko, S. Yu. High-performance aggregation element-by-
element Ritz-gradient method for structure dynamic
response analysis. CAMES, vol.7, pp. 537-550, 2000