Recent Advances in Computational Mechanics – Łodygowski, Rakowski & Litewka (Eds)

© 2014 Taylor & Francis Group, London, ISBN 978-1-138-02482-3

Multi-load truss topology optimization using the adaptive

ground structure approach

T. Sokół
Warsaw University of Technology, Warsaw, Poland

ABSTRACT: The paper presents a new method to determine optimal topologies of multi-load trusses which
can be regarded as generalized Michell structures subjected to multiple load conditions. The method is based on
the adaptive ground structure approach and stress-based formulation of the optimization problem. It leads to a
huge but linear programming problem. Thus, from definition, the proposed method assures finding the global
optimum for the given discretization of the design domain, represented here by the fully connected ground
structure. The method makes use of both active set and interior point methods and allows solving a large-
scale optimization problem even with more than billion of design variables. The efficiency and robustness of the
proposed method have been confirmed in several benchmark tests. Moreover, an important result of this research
is finding a new exact-analytical solution of a specific symmetrical two-load case problem. This solution was
predicted numerically using the proposed method, and then adjusted to obtain the exact formulae describing the
layout of the optimal structure using the concept of component loads.

1 INTRODUCTION shape and cross-section areas. Moreover, the main dif-

ficulty of the classical approach has recently been
Michell trusses are the most efficient structures capa- overcome by new linear programming methods based
ble of transmitting the given load to the given supports on the adaptive ground structure approach, introduced
using the smallest amount of structural material. In fact by Gilbert and Tyas (2003) and then modified by Sokół
they are not real trusses because they are composed (2011a, b). The latter method has been tested using
of an infinite number of infinitesimal bars. Neverthe- ground structures with more than a billion potential
less, they define the shape and topology of the optimal bars giving the excellent agreement with known exact
structure, thus they play a significant role in structural solutions of Michell problems; see Sokół et al. (2010,
topology optimization. The exact analytical solutions 2011, 2012), Lewiński & Sokół (2013). The main goal
of Michell trusses are very hard to obtain but can be of the present paper is to show that this method can be
accurately approximated numerically by trusses com- enhanced and applied to multi-load case problems.
posed of large but finite number of bars. It requires, The problem of multiple load cases is not new and
however, solving large-scale numerical optimization was investigated in many papers (see for example:
problems which can be handled only by appropriate Achtziger (1993, 1998), Bendsøe et al. (1994), Hemp
optimization methods. This problem is more complex (1973), Kirsch (1989), Nagtegaal & Prager (1973),
for the case of multiple loads since then one has to deal Pedersen (1972), Rozvany et al. (1978, 1992, 1993,
with independent member forces (design variables) for 2013)) and some basic properties of the optimal multi-
each load condition. Hence the size of the optimization load trusses are known. For instance and contrary to
problem increases several times. classical Michell solutions, the bars in the optimal
The classical ground structure approach was intro- multi-load truss have not to be orthogonal nor fully
duced by Dorn, Gomory and Greenberg (1964), who stressed and the optimal trusses are often statically
showed that truss topology optimization problem can indeterminate structures. In the present paper some
be formulated on the basis of plastic design principles new properties of multi-load trusses will be discussed.
and solved using the linear programming methods. The
above paper received little attention in recent decades.
This approach leads to a large optimization problem 2 PRIMAL AND DUAL FORMULATIONS OF
with many design variables and corresponding con- MULTI-LOAD STRESS-BASED
straints, and therefore might be regarded as the “brutal OPTIMIZATION OF TRUSSES
force” or inefficient method. But contrary to many
other approaches, it assures finding a possibly light- Let us start the discussion by reviewing various formu-
est truss in a global sense. The solution obtained by lations of the stress-based multi-load truss optimiza-
this method determines at once the optimal topology, tion problem and then derive the formulae needed to

9
 determine which new candidate bars should be acti- prefer a standard form of the optimization problem in
vated in a subsequent iteration. These formulae define which all design variables are greater than or equal to
adjoint strains (or elongations) in the optimal truss zero. In addition, the form (1) cannot be applied in
and can be treated as generalized Michell (1904) opti- the adaptive ground structure method, discussed in the
mality criteria. They allow also defining a proper stop next section. The reason of this can be understood by
condition for the method developed, see Section 3. examining the dual form of the problem (1) which can
Any optimization problem can be written in dif- be written as
ferent forms which are mathematically equivalent but
can lead to significantly different calculation times
using the given optimization method (see Achtziger
(1993, 1998), Bendsøe et al. (1994)). In other words,
the formulation of the optimization problem should be
matched to the method applied.
According to well known duality principles, the
plastic design optimization problem can be written as
well in primal as dual form. Both of them play an
Downloaded by [Warsaw University of Technology] at 09:21 13 December 2015

important role in the proposed method and should be

considered together. in which vectors u(l) , e+ −
(l) and e(l) denote appropriate
The primal (lower bound) form of the truss topol- Lagrange multipliers. Note that they are independent
ogy optimization problem for different stress limits for every load condition. The physical interpretation of
for tension σ T and compression σ C , and for multiple these multipliers is as follow: u(l) denote the adjoint
load conditions can be defined as follows (c.f. Bendsøe nodal displacements; e+ −
(l) and e(l) are the vectors of
et al. (1994)): adjoint elongations of members, both considered in
positive manner for tension and compression. The
objective function W is the sum of works of the forces
P(l) over u(l) so one can loosely say that we maxi-
mize the ‘total compliance’ over admissible adjoint
displacements and member elongations.
In the problem (3) all dual variables are coupled
through equality constraints and must be perfectly
matched what makes trouble with the convergence of
The goal of this problem is to find the truss of min- the optimization process. Moreover this formulation is
imal volume of structural material and subject to not applicable in the adaptive ground structure method
equilibrium equations (defining the sets of statically (discussed later). Thus it is recommended to explicitly
admissible member forces) and to inequality con- specify that the cross-section areas A are non-negative
straints arising from the allowable stresses in tension
and compression. Thus in this natural form, the objec-
tive function V is the total volume of bars of the truss;
M denotes the number of potential bars; L is the vector
of bar lengths; A is the vector of member cross-section
areas (the main design variables); B is the geometric
matrix including directional cosines of bars; P(l) are
the vectors of nodal loads, independent for every load
case l = 1, 2, …, K, where K denotes the number of
load conditions; and S(l) are the vectors of member
forces corresponding to loads P(l) . Formally, the constraints (4)4 are not necessary
Note that problem (1) belongs to worst-case formu- because they follow directly from (1)3 . Nevertheless
lations which define ‘true designs’ for multiple and it is worth adding them explicitly to obtain a more
independent load conditions because the constraints relaxed dual form
(1)3 simply define the smallest required cross-section
area of every bar:

It is to be remarked that the more popular (easier to use)

weighted-average formulations cannot strictly satisfy
(2).
The concise form (1) is not convenient for direct Here, instead of equations (3)2 we have inequal-
use in linear programming methods because the design ity constraints (5)2 letting dual variables to be eas-
variables can be positive or negative. These methods ier adjusted. Moreover, formulation (5) enables to

10
 derive the optimality criteria suitable for the adaptive The dual (upper bound) form of the multi-load case
ground structure approach (they are however discussed problem (6) can be written as
later).
The formulations (4) and (5) were applied in the first
version of the program developed. It enabled finding
accurate solutions of some relatively simple multi-
load case problems for which the ground structure did
not have to be very dense. Unfortunately, the larger
problems quickly revealed that the convergence of the
interior point method based on the form (4) is rather
slow. Therefore, in the next stage of the study author
decided to convert (4) to a more applicable (standard)
form. It can be done by separating member forces S
into tension and compression forces: T and C respec-
It is worth noting that using the primal-dual interior
tively (cf. Sokól (2011)). This leads to the following
point method (proposed by Karmarkar (1984) and later
primal form
Downloaded by [Warsaw University of Technology] at 09:21 13 December 2015

modified by Mehrotra (1992) and Wright (1997)) it is

enough to solve the problem (6) because we get the
dual variables of the problem (9) for free. They are
necessary to derive a suitable recipe of activating new
bars in the ground structure.
Let us consider for the moment only one, i-th bar
of the truss. Its elongation for load case l is equal to

where Bi denotes i-th row of matrix B. Now one can

easily deduce that in an optimal solution one of the
constraints (9)3 or (9)4 has to be active and three
which is almost two times greater than (4) but here all scenarios are possible:
design variables are non-negative. This significantly
improves the convergence of the interior point method
and reduces the computation time.
After arranging the total vector of design variables
in the order: {A, T(1) , C(1) , …, T(K) , C(K) }, the coef-
ficient matrix of the problem (6) can be written in the
following block-matrix form: which can be reduced to a more concise form

After dividing the elongations e+ −

(l),i and e(l) by the
length Li one can easily replace (10) and (11) by similar
formulas based on adjoint member strains

where The constraints (9)2,3,4 determine the domain of

feasible adjoint member elongations and allow to for-
mulate the following theorem of generalized Michell
optimality criteria for multi-load trusses.

Theorem:
In the stress-based multi-load truss optimization prob-
lem the optimal solution has to satisfy the following
conditions:
Note that all matrices defined in (7) and (8) are very
sparse and this sparsity has to be utilized to achieve 1) for every bar of the truss the adjoint multi-load
a good efficiency of computations. In addition, the strains are restricted by
main matrix H has a regular-repetitive form which also
should be taken into account.

11
 where

and

2) moreover, the non-zero cross-section area Ai is

needed only for ‘fully strained’ bar:

The term “fully strained” corresponds to the total Figure 1. The two-load case problem.
normalized adjoint strain defined in (14). The con-
ditions (14) determine the domain of feasible adjoint

(σT ε+ −
(iter−1)
strain fields and can be utilized in the adaptive ground iteration: u(l) ⇒ εi = (l),i + σC ε(l),i )
structure method discussed in the next section. (see Eqs (9), (11), (12)),
Downloaded by [Warsaw University of Technology] at 09:21 13 December 2015

– if εi ≥ 1 − tol, then activate (add) i-th bar,

– otherwise, if εi < 0.3 and d < dmax then deacti-
3 THE METHOD OF ADAPTIVE GROUND vate (remove) bar,
STRUCTURES WITH SELECTIVE SUBSETS – if d < dmax and the number of added bars is too
OF ACTIVE BARS small then go to step 4.
In the proposed method the solution is obtained and 6. Check the stopping criterion:
improved iteratively. Each iteration begins with a prop- – if d = dmax and there are no new bars added
erly adjusted set of active bars using the information then finish (we approach the optimum solution
from the previous solution. Inactive bars are assumed because for all potential bars i = 1:M the con-
to have zero cross-section areas and zero member straints (13) are satisfied and the solution can
forces and are eliminated from the system. Then the not be further improved).
significantly reduced problem (6) for m active bars
(m << M ) is solved using the primal-dual version of 7. Solve primal problem (6) and get dual variables
(iter)
the interior point method. After solution the dual vari- u(l) .
ables of the problem (9) are extracted without any 8. Repeat from step 3.
additional cost. The optimality criteria discussed in
the previous section serve as a hint for activating or The program implementing the above algorithm has
deactivating new bars and also determine the stop con- been written in Mathematica (Wolfram 2003).
dition (after obtaining the globally optimal solution;
c.f. Rozvany & Sokół (2013)).
The step by step procedure for the proposed method 4 EXAMPLES
can be described as follows:
Let us consider a two-load case problem of two sym-
First iteration: metrically distributed but independent loads P(1) and
1. Set iter = 1, d = 1 and generate the initial ground P(2) , see Fig. 1. In this example we assume that:
structure Nx × Ny :1 × 1 with bars connecting the a) σT = σC = σ0 ; b) the magnitudes of the loads are
neighbouring nodes (here only horizontal, vertical equal: |P(1) | = |P(2) | = P; c) the distances of applied
and 45◦ slope bars are included, see Sokół (2011b). loads √ measured√from the supports are defined by
These bars will always be active (even for zero cross d/L = 2/(1 + 2) ≈ 0.585786. The optimal volumes
sectional area). They play a role of groundwork to given in subsequent figures are scaled by a reference
obtain nodal displacements even in the voids. volume defined as V0 = P L/σ0 .
2. Solve the problem (6) for this initial ground struc- In Figs 2a and b the solutions of related one-load
(1)
ture and get the dual variables u(l) . case problems are presented just to show that they do
not have much in common with the two-load case solu-
Next iterations: tion of Fig. 3a. They are rather special cases of the
three forces problem, discussed in Sokół & Lewiński
3. Increment the number of iteration ++iter.
(2010).
4. Increment the distance of connections:
Numerical solution of Fig. 3a suggests that the opti-
d : = max(dmax , d+1),
mal truss is composed of two almost independent
dx : = max(dxmax , dx + 1),
trusses (the common nodes are located only at the
dy : = max(dymax , dy + 1).
supports and at the points of application of forces).
5. Select the new set of active bars in the ground
This rather surprising solution can be better under-
structure Nx × Ny :dx × dy :
stood if we refer to the concept of component loads,
– for every new bar compute normalized strain see Rozvany & Hill (1978) and Rozvany et al. (2013).
using the displacement fields from the previous It was checked numerically that the layout and volume

12
Downloaded by [Warsaw University of Technology] at 09:21 13 December 2015

Figure 2. Numerical solutions of one-load case problems

relevant to the problem of Fig. 1.

Figure 4. Exact layouts and volumes of the problems of

Fig. 3.

problem which required about 75 million design vari-

ables (it is 2.5 times greater than in the one-load case
problems). Nevertheless, it was also noted that the con-
vergence of the interior point method implemented in
Mathematica is slower for two-load case. This topic
is currently being investigated and will be reported in
future publications.
The numerical solutions of Fig. 3 can be refined
to obtain the exact analytical solutions by superposi-
tion (Rozvany & Hill (1978)), see Fig. 4. The solution
of Fig. 4a can be derived after Sokół & Lewiński
(2010, 2011). Using the notation from this paper
and after solving the transcendental system (see eqns
5.10 in Sokól & Lewinski (2010)) one can obtain
the three angles defining the topology: γ2 = 24.1351◦ ;
θ2 = 27.5555◦ ; θ1 = γ2 + θ2 = 51.6906◦ . The height of
this structure is equal to h = 0.630069L and the total
volume is equal to
Figure 3. Numerical solutions of: a) two-load case problem,
b) symmetric and c) asymmetric component loads.

The solution of Fig. 4b is easy to obtain due to circu-

shown in Fig. 3a can be obtained by superposition lar fans of angles ¾ π. The member forces in the exter-
of solutions of Figs 3b and c which were obtained nal chords are equal to vertical reactions (horizontal
using the older, one-load case version of the program reactions vanish due to antisymmetric load). The√ mag-
developed by the author (2011). nitude of the reactions is equal to R = P/[2(1 + 2)].
The solutions of Figs 2 and 3 were performed using After simple algebraic calculations we obtain
the ground structures of the same densities 140 × 49
with about 15 million potential bars. The one-load
case solutions of Figs 2ab and 3bc were obtained
after 10 iterations with the execution time less than Finally, the exact solution presented in Fig. 4c is
two minutes. The two-load case solution of Fig. 3a just the superposition of Figs 4a and b with the total
was obtained after 12 iterations and took about 25 volume equal to
minutes. This significantly longer time of computa-
tion results from the bigger size of the two-load case

13
Downloaded by [Warsaw University of Technology] at 09:21 13 December 2015

Figure 5. Numerical solutions of a ‘bridge problem’ for top half plane domain and two pin supports, a) one-load case with
uniformly distributed load q = P/L, b) multi-load case: |P(l) | = P, l = 1, 2, …, 9.

The numerical volume of Fig. 3a is only 0.14% is similar to πV0 (this is suspected value of the exact
worse. Obviously these volumes (Figs 3a and 4c) solution not proved yet).
are less than the sum of the volumes of structures The optimal multi-load trusses of Fig. 5b and 6b
presented in Fig. 2. are quite complicated but indicate that the generalized
The program presented in the previous section was Michel multi-load structures resemble multilayered
written for a general multi-load case problem with any laminates. This is very important conclusion, perhaps
number of load cases. The more complex examples of new in the literature.
‘modified bridge problems’ with nine-load cases are
presented in Figs 5 and 6. They differ only in supports:
pin-pin (Fig. 5) and pin-roller (Fig. 6). To broaden
the discussion we present also in Figs 5a and 6a the 5 CONCLUDING REMARKS
numerical solutions of one-load case problems for con-
tinuous and uniformly distributed load q = P/L, where The method proposed in the present paper results
L denotes the half span length. The solution in Fig. 5a in a significant reduction of the size of the prob-
has also been derived by Pichugin et al. (2012). These lem because most of the unnecessary zero bars are
two solutions were obtained using the ground struc- eliminated a priori from the ground structure.The solu-
tures 105 × 100 with almost 35 million bars and using tion is obtained iteratively using suitably chosen small
the symmetry of the problem (right part of the struc- subsets of active bars, and instead of one large opti-
ture was analyzed and then symmetrically reflected mization problem a few much smaller problems are
to the left). It is worth nothing that the problem of solved. Moreover, the convergence of the proposed
Fig. 5a was also calculated for much denser ground method is very good because usually it is enough to
structure 400 × 280 with over 3.5 billion potential bars perform 10 to 15 iterations. The program written by
giving the optimal volume equal to 3.15257V0 which the author is still being developed and improved but

14
Downloaded by [Warsaw University of Technology] at 09:21 13 December 2015

Figure 6. Numerical solutions of a bridge problem for top half plane domain and pin and roller supports, a) one-load case
with uniformly distributed load q = P/L, b) multi-load case: |P(l) | = P, l = 1, 2, …, 9.

the preliminary results obtained in this paper indi- of Science and Higher Education, entitled: Topology
cate its high reliability. These results clearly indicate Optimization of Engineering Structures. Simultaneous
also that the optimal stress-based multi-load truss can shaping and local material properties determination.
resemble the multilayered laminate in which every I would also like to express my sincere thanks to
layer is composed of the orthotropic material resulting Prof. George I.N. Rozvany for inspiring discussions
from the classical one-load Michell solution obtained concerning stress-based multi-load truss topology
using the concept of component loads (Rozvany and optimization.
Hill (1978)). In addition, the generalized optimal-
ity criteria for multi-load trusses were derived in a
concise form. REFERENCES
Achtziger, W. 1993. Minimax compliance truss topology sub-
ject to multiple loadings. In: M.P. Bendsøe & C.A. Mota
ACKNOWLEDGEMENTS Soares (eds.), Topology optimization of structures: 43–54,
Dordrecht, Kluwer,
The paper was prepared within the Research Grant Achtziger, W. 1998. Multiple load truss topology and sizing
no N506 071338, financed by the Polish Ministry optimization: Some properties of minimax compliance,

15
 Journal of Optimization Theory and Applications 98(2): Rozvany, G.I.N. & Sokół, T. 2013. Validation of numerical
255–280. methods by analytical benchmarks, and verification of
Bendsøe, M.P., Ben-Tal, A. & Zowe, J. 1994. Optimiza- exact solutions by numerical methods. In G.I.N. Rozvany
tion methods for truss geometry and topology design, & T. Lewiński (eds.), Topology Optimization in Structural
Structural Optimization 7(3): 141–159. and Continuum Mechanics. CISM 549: 53–69, Udine,
Dorn, W.S., Gomory R.E. & Greenberg, H.J. 1964. Automatic Springer.
design of optimal structures, Jurnal de Mecanique 3(1): Rozvany, G.I.N., Sokół, T. & Pomezanski, V. 2013. Exten-
25–52. sion of Michell’s theory to exact stress-based multi-load
Gilbert, M. & Tyas, A. 2003. Layout optimization of truss optimization, 10th World Congress on Structural
large-scale pin-jointed frames, Engineering Computa- and Multidisciplinary Optimization, Orlando, 19–24 May
tions 20(8): 1044–1064. 2013.
Hemp, W.S. 1973. Optimum structures, Clarendon, Oxford. Rozvany, G.I.N., Zhou, M. & Birker, T. 1993. Why multi-load
Karmarkar, N. 1984. A new polynomial time algorithm for topology designs based on orthogonal microstructures
linear programming, Combinatorica 4(4): 373–395. are in general non-optimal. Structural optimization 6(3):
Kirsch, U. 1989. Optimal topologies of structures. Applied 200–204.
Mechanics Reviews 42(8): 223–239. Sokół, T. 2011a. A 99 line code for discretized Michell
Lewiński, T. & Sokół, T. 2013. On basic properties of truss optimization written in Mathematica, Structural and
Michell’s structures. In G.I.N. Rozvany & T. Lewiński Multidisciplinary Optimization 43(2): 181–190.
Downloaded by [Warsaw University of Technology] at 09:21 13 December 2015

(eds.), Topology Optimization in Structural and Contin- Sokół, T. 2011b. Topology optimization of large-scale trusses
uum Mechanics. CISM 549: 87–128, Udine, Springer. using ground structure approach with selective sub-
Mehrotra, S. 1992. On the implementation of a primal-dual sets of active bars. In A. Borkowski, T. Lewiński, G.
interior point method, SIAM Journal on Optimization 2(4): Dzierżanowski (eds.), 19th International Conference on
575–601, Computer Methods in Mechanics, Warsaw, 9–12 May
Michell, A.G.M. 1904. The limits of economy of mate- 2011.
rial in frame structures. Philosophical Magazine 8(47): Sokół, T. & Lewiński, T. 2010. On the solution of the three
589–597. forces problem and its application in optimally designing
Nagtegaal, J.C. & Prager, W. 1973. Optimal layout of a truss symmetric plane frameworks of least weight, Structural
for alternative loads. International Journal of Mechanical and Multidisciplinary Optimization 42(6): 835–853.
Sciences 15: 583–592. Sokół, T. & Lewiński, T. 2011. On the three forces problem
Pedersen, P. 1972. On the optimal layout of multi-purpose in truss topology optimization. Analytical and numerical
trusses. Computers & Structures 2(5–6): 695–712. solutions. In H. Yamakawa (ed.), 9th World Congress on
Pichugin, A.V., Tyas, A. & Gilbert, M. 2012. On the optimal- Structural and Multidisciplinary Optimization, Shizuoka,
ity of Hemp’s arch with vertical hangers, Structural and 13–17 June 2011.
Multidisciplinary Optimization, 46(1): 17–25. Sokół, T. & Rozvany, G.I.N. 2012. New analytical bench-
Rozvany, G.I.N. 1992. Optimal layout theory: analytical marks for topology optimization and their implications.
solutions for elastic structures with several deflection Part I: bi-symmetric trusses with two point loads between
constraints and load conditions. Structural Optimization supports, Structural and Multidisciplinary Optimization
4(3–4): 247–249. 46(4): 477–486.
Rozvany, G.I.N. & Hill, R.H. 1978. Optimal plastic design: Wolfram, S. 2003. The Mathematica book, 5th ed.
superposition principles and bounds on the minimum cost, Champaign: Wolfram Media.
Computer Methods in Applied Mechanics and Engineer- Wright, S. 1997. Primal-Dual Interior-Point Methods,
ing 13(2): 151–173. Philadelphia, PA, SIAM.

16