Cement and Concrete Research 34 (2004) 409 – 420

Genetic algorithm in mix proportioning of high-performance concrete

Chul-Hyun Lim, Young-Soo Yoon *, Joong-Hoon Kim

Department of Civil and Environmental Engineering, Korea University, Seoul, Republic of Korea
Received 27 March 2003; accepted 19 August 2003

Abstract

High-performance concrete is defined as concrete that meets special combinations of performance and uniformity requirements that
cannot always be achieved routinely using conventional constituents and normal mixing, placing, and curing practices. Ever since the term
high-performance concrete was introduced into the industry, it had widely used in large-scale concrete construction that demands high
strength, high flowability, and high durability. To obtain such performances that cannot be obtained from conventional concrete and by the
current method, a large number of trial mixes are required to select the desired combination of materials that meets special performance.
Therefore, in this paper, using genetic algorithm that is a global optimization technique modeled on biological evolutionary process—natural
selection and natural genetics—and can be used to find a near optimal solution to a problem that may have many solutions, the new design
method for high-performance concrete mixtures is suggested to reduce the number of trial mixtures with desired properties in the field test.
Experimental and analytic investigations were carried out to develop the design method for high-performance concrete mixtures and to verify
the proposed mix design.
D 2003 Elsevier Ltd. All rights reserved.

Keywords: High-performance concrete; Genetic algorithm; Mixture proportioning; Compressive strength; Workability

1. Introduction constructors all over the world are finding that using high-
performance concrete allows them to build more serviceable
1.1. High-performance concrete structures at comparable cost. High-performance concrete is
being used for structures in aggressive environments: ma-
The parts of the world in which large-scale concrete rine structures, highway bridges and pavements, nuclear
construction takes place have extended enormously. Due to structures, tunnels, precast units, etc. [11,12].
the recent trends in construction industries (i.e., increased The major difference between conventional concrete and
number of heavily reinforced concrete structures), construc- high-performance concrete is essentially the use of chemical
tion of large and taller structures, and developments of and mineral admixtures. The use of chemical admixtures
construction techniques (i.e., efficient concrete pumping reduces the water content, thereby at the same time reduces
techniques), the industries and companies in general strive the porosity within the hydrated cement paste. The reduction
to cast massive volume of concrete. When this large volume in the water content to a very low value with high dosage of
of concrete is used for construction, the safety and durability chemical admixtures is undesirable, and the effectiveness of
of cast concrete become fundamental issues. To ensure these chemical admixtures such as superplasticizer principally
issues, much effort has been focused on the developments of depends on the ambient temperature, cement chemistry,
high-performance concrete [1– 10]. and fineness. Mineral admixtures, also called as cement
High-performance concrete is designed to give optimized replacement materials, act as pozzolanic materials as well as
performance characteristics for a given set of materials, fine fillers; thereby, the microstructure of hardened cement
usage, and exposure conditions, consistent with strength, matrix becomes denser and stronger. At ambient tempera-
workability, service life, and durability. Engineers and ture, their chemical reaction with calcium hydroxide is
generally slow. However, the finer and more vitreous the
* Corresponding author. Tel.: +82-2-3290-3320; fax: +82-2-928-7656. pozzolan is, the faster will be this reaction. If durability is of
E-mail address: [email protected] (Y.-S. Yoon). primary interest, then the slow rate of setting and hardening

0008-8846/$ – see front matter D 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cemconres.2003.08.018
410 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420

associated with the incorporation of fly ash or slag in

concrete is advantageous. Also, the mineral admixtures are
generally industrial by-products and their use can provide a
major economic benefit. Thus, the combined use of super-
plasticizer and cement replacement materials can lead to
economical high-performance concrete with enhanced
strength, workability, and durability. It is also reported that
the concrete containing cement replacement materials typ-
ically provides lower permeability, reduced heat of hydra-
tion, reduced alkali –aggregate reaction, higher strength at
later ages, and increased resistance to attack from sulfates. Fig. 2. Two-point crossover.
However, the effect of cement replacement materials on the
performance of concrete varies markedly with their proper- ability to keep existing parts of solution, which have a
ties [13]. positive effect on the outcome, and proceed with optimizing
To obtain the special combinations of performance and the nonoptimal part. The approach is global, as possible as
uniformity requirements, a near-optimum mix proportion of many samples from different parts of the solution space are
high-performance concrete is very important. However, examined simultaneously. Then the transition rules, which
there have not been any guide specifications on the mix combine and change those samples in order to improve the
proportion of high-performance concrete. Required mix solutions already found, are probabilistic and not determin-
proportions are therefore obtained by trial and error method istic. This enables genetic algorithm to reach a global
that is based on existing data and by the method that is optimum without being fixed in local optima. Genetic
based on specifications for conventional concrete mixtures. algorithm, known as a very efficient heuristic algorithm
Such methods for mix proportioning require large number that has been widely used in the various fields of engineer-
of trial mixes to select the desired combination of materials, ing [23 – 27], gives therefore more accurate results than
whereas a good mix proportioning procedure has to mini- other algorithms in the mix proportioning problem having
mize the number of trial mixes and achieve an economical many local solutions.
and satisfactory mixture with desired properties [14,15].
This paper presents a new design method for high-perfor-
mance concrete mixtures using genetic algorithm to mini- 2. Genetic algorithm
mize the number of trial mixes and provide appropriate mix
proportion. The general form of genetic algorithm, as described in
Genetic algorithm, which is first formalized as an opti- the book written by Goldberg, is composed of three major
mization method by Holland, is a global optimization processes, i.e., selection, crossover, and mutation [28].
technique for high dimensional, nonlinear, and noisy prob-
lem and a stochastic search technique based on the mech- 2.1. Selection
anism of natural selection and natural genetics [22]. Genetic
algorithm, differing from conventional search techniques, The fundamental principle of genetic algorithm starts
starts with an initial set of random solutions called popula- from essentially Darwinian natural selection. The initial
tion. Each individual in the population is called a chromo- individuals were selected in this process. Selection provides
some, representing a solution to the problem at hand. The the driving force in genetic algorithm, and selection pressure
evolution operation simulates the process of Darwinian is critical in it. Typically, low selection pressure is indicated
evolution to create population from generation to genera- at the start of the genetic algorithm search in favor of a wide
tion. The success of genetic algorithm is founded in its exploration of the search space, while high selection pres-
sure is recommended at the end in order to exploit the most

Fig. 1. Single-point crossover. Fig. 3. Uniform crossover.

 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420 411

population k, fk is the size of population k, and fi is the size

of whole population.
Therefore, a wheel can be made according to these
probabilities. The selection process is based on spinning
the roulette wheel pop  size times. Each time, a single
Fig. 4. Mutation.
chromosome is selected for the new population as described
in the last selection.
promising regions in the search space. The selection directs
genetic algorithm search toward promising regions in the 2.1.2. Ranking selection
search space. This theory introduced to overcome the scaling problems
of the direct fitness-based approach. The idea sorts the
2.1.1. Roulette wheel selection (proportional selection) population from the best to the worst and assigns the
Roulette wheel selection is the best-known selecting selection probability of each chromosome according to the
method due to its simplicity. It is usually presented in every ranking. This method ranks the individuals according to
introduction to genetic algorithm. The basic idea is to their normalized fitness. Two methods are in common use:
determine selection probability; also called survival proba- linear ranking and exponential ranking.
bility for each chromosome proportional to the fitness value.
For chromosome with fitness, its selection probability is 2.1.3. Tournament selection
calculated as follows: This approach contains both random and deterministic
features simultaneously. This method randomly chooses a
fk
Pk ¼ ð1Þ set of chromosomes and picks out the best one from the set
X
popsize
for reproduction. The number of chromosomes in the set is
fi
called tournament size. A common tournament size is 2.
i¼1
This is called binary tournament. Selection probabilities are
where Pk is the selection probability of a chromosome from normally calculated and successive pairs of chromosomes
are drawn using roulette wheel selection. After drawing a
pair of the chromosomes with higher fitness, these are

Fig. 5. Genetic algorithm process. Fig. 6. Research procedure.

412 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420

Table 1 Table 3
Properties of cement Properties of fly ash
Properties Cement Properties Fly ash
Specific gravity 3.15 Specific gravity 2.13
Fineness (m2/kg) 310 Fineness (m2/kg) 360
SiO2 (%) 21.3 SiO2 (%) 63.5
Al2O3 (%) 4.7 Al2O3 (%) 28.5
Fe2O3 (%) 3.1 Fe2O3 (%) 4.2
CaO (%) 63.1 CaO (%) 1.2
MgO (%) 3 MgO (%) 0.8
Loss on ignition (%) 0.8 Moisture content (%) 0.2
Loss on ignition (%) 3.6

inserted in the new population. The process continues until

the population is full. 2.2.2. Two-point crossover
Two-point crossover is shown in Fig. 2. It differs from
2.2. Crossover single-point crossover merely in the point that two random
cuts are made; so three pieces have to be put together in
Crossover simulates the sexual generation of a child or order to produce an offspring.
offspring from two parents. This is performed by taking
parts of the bit string of one of the parents and the other 2.2.3. Uniform crossover
parts from the other parent and by combining both in the Crossover can be implemented with uniform crossover
child. The crossover rate is defined as the ratio of the operator, which has been shown to be superior to traditional
number of offspring produced in each generation to the crossover strategies for combinatorial problem. Uniform
population size. This ratio controls the expected number crossover first generates a random crossover mask and then
(Crossover Rate  Population Size) of chromosomes to exchanges relative genes between parents according to the
undergo the crossover operation. A higher crossover rate mask. A crossover mask is simply a binary string with the
allows exploration in the large solution space and reduces same size of chromosome. The parity of each bit in the mask
the chances of settling for a false optimum; but if this rate determines, for each corresponding bit in an offspring, which
is too high, it results in the wastage of a lot of computation parent it will receive that bit from. It is illustrated in Fig. 3.
time in exploring unpromising regions of the solution
space. There are three kinds of crossover: single-point, 2.3. Mutation
two-point, and uniform crossover.
The crossover operator is able to generate all possible
2.2.1. Single-point crossover values of genotypes even if the population contains only
This kind of crossover operator is analogous to that of the copy of a specific allele value. But as the genetic algorithm
binary implementation. The basic one is one-cut-point proceeds to generate new genotypes, it is always possible to
crossover. Let two parents be u=[u1, u2, u3. . .un] and lose the last copy of an allele value. The second operator
v = [v1, v2, v3. . .vn] if they are crossed after the kth position, prevents the population of genotypes from loosing a specific
and the resulting offsprings are shown as in Eqs. (2) and (3) value of an allele. Mutation is performed with a given
and Fig. 1 shows a single-point crossover. probability. The mutation rate is defined as the percentages
of new genes to the total number of genes in the population
u ¼ ½u1 ; u2 ; u3 ; . . . . . . ; uk ; ukþ1 ; ukþ2 ; . . . . . . ; un  ð2Þ and controls the rate at which new genes are introduced into
the population for trial. If it is too slow, many genes that
would have been useful are never tried out. But if it is too
v ¼ ½v1 ; v2 ; v3 ; . . . . . . ; vk ; vkþ1 ; vkþ2 ; . . . . . . ; vn  ð3Þ high, there will be much random perturbation, the offspring

Table 4
Properties of silica fume
Table 2 Properties Silica fume
Properties of aggregates
Specific gravity 2.1
Properties Coarse aggregate Fine aggregate
Fineness (m2/g) 2200
Fineness modulus 7.2 2.94 SiO2 (%) 92
Specific gravity 2.7 2.61 Al2O3 (%) 1.3
Water absorption (%) 0.6 0.8 Fe2O3 (%) 2.4
Unit weight (kg/cm3) 1480 1590 MgO (%) 0.4
Abrasion (%) 18.6 – Moisture content (%) 0.1
 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420 413

Table 5 Table 7
Properties of superplasticizer Mix proportion (40 – 80 MPa)
Properties Superplasticizer Gmax W/B (%) W (kg/m3) Gv (%) FA (%)
Color tone Deep brown 19 mm 30 160 32 0
State Liquid 35 170 34 10
Specific gravity 1.22 40 180 36 20
Solid content (%) 40.9 45
pH 7.5 Gv: volume of coarse aggregate.

will start losing their resemblance to the parents, and the

algorithm will lose the ability to learn from the history of the used was crushed granite with a specific gravity of 2.7 and a
search. Fig. 4 illustrates mutation. fineness modulus of 7.2, its maximum particle size is 19
Boundary mutation changes one of the parameters of the mm. The fine aggregate was quartz sand with a specific
parent and changes it randomly either to its upper or lower gravity of 2.61 and a fineness modulus of 2.94. A naphtha-
bound, multi-nonuniform mutation changes all of the lene superplasticizer was used to keep the water to binder
parameters of the parent based on a nonuniform probability (W/B) ratio of concrete at a very low level. An air-entraining
distribution, nonuniform mutation changes one of the agent was used. A class F fly ash and a silica fume produced
parameters of the parent based on a nonuniform probability by Elkem, Norway, were used. The detailed properties of
distribution, and uniform mutation changes one of the these materials are shown in Tables 1 –6.
parameters of the parent based on a uniform probability
distribution. 3.3. Mix proportions
After creating initial population composed of strings
substituting for mix proportions, fitness increases through The mix proportions necessary to obtain a compressive
the repeating process of selection, crossover, and mutation. strength between 40 and 80 MPa are presented in Table 7.
When fitness is satisfied, the repeating process is terminated There are nonsilica fume mixes. The W/B varies between
and optimal solution is approached. This process is illus- 0.30 and 0.45, and the amount of fly ash used varies from
trated in Fig. 5. 0% to 20% of the total binder. The range of Gv (the volume
of coarse aggregate) is 32 –36%. The range of water content
(W) is 160– 180 kg/m3, and the content of superplasticizer
3. Experimental program and air-entraining agent are 0 – 2% and 0.010 – 0.013%,
respectively, when expressed as a percentage of dry solids
3.1. Procedure to the binder content.
The mix proportions necessary to obtain a compressive
The 189 sets of mixtures were used for this experiment strength between 80 and 120 MPa are presented in Table 8.
program. Based on the results of mixing test, using Statistical The silica fume dosage varies from 8% to 25%, the W/B
Packages for Social Science (SPSS) version 10, the fitness ratio between 0.18 and 0.27. The range of the ratio of the
functions were developed from the 181 sets of mixtures weight of fine aggregate to the weight of all aggregate (s/a)
[18,19]. After that, the fitness functions were applied in is 35 –39%. The range of W is 140 – 165 kg/m3, and the
genetic algorithm, using MATLAB version 5.1 [20,21], and content of superplasticizer is 2 –5% of binder content.
the results from genetic algorithm were compared to the eight Based on Tables 7 and 8, the 189 batches were mixed in
sets of test mixtures for the verification and validity of the this experiment. The 181 mixtures are listed in Tables 9 and
algorithm. This procedure is illustrated in Fig. 6. 10 according to compressive strength. The eight mixtures
for the verification and validity of the algorithm are listed in
3.2. Material properties Tables 15 and 16.

All materials except silica fume used in this experiment 3.4. Mixing
were produced in South Korea. Portland cement in accor-
dance with ASTM type I was used. The coarse aggregate In many countries, most concrete today is batched and
mixed in ready-mixed concrete plants where the batching is

Table 6
Properties of air-entraining agent Table 8
Properties Air-entraining agent Mix proportion (80 – 120 MPa)

Color tone Colorless Gmax W/B (%) W (kg/m3) SF (%)

State Liquid 19 mm 23, 25, 27 155, 160, 165 5, 10, 15
Specific gravity 1.05 20, 23, 25 145, 150, 155 10, 15, 20
Solid content (%) 6 18, 20, 22 140, 145, 150 15, 20, 25
414 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420

Table 9 Table 9 (continued)

Tested mixtures (40 – 80 MPa)
Number fcV Slump W/B W s/a FA AE SP
Number fcV Slump W/B W s/a FA AE SP (MPa) (mm) (%) (kg/m3) (%) (%) (kg/m3) (kg/m3)
(MPa) (mm) (%) (kg/m3) (%) (%) (kg/m3) (kg/m3)
63 46 175 40 170 50 20 0.043 2.13
1 74 215 30 160 48 10 0.069 8.00 64 47 190 40 170 48 0 0.043 2.13
2 74 245 30 160 48 20 0.069 8.00 65 47 195 40 170 47 10 0.043 2.13
3 71 200 30 160 46 0 0.069 8.00 66 46 195 40 170 47 20 0.043 2.13
4 72 210 30 160 45 10 0.069 8.00 67 47 170 40 170 45 0 0.043 2.13
5 69 205 30 160 44 20 0.069 8.00 68 46 200 40 170 44 10 0.043 2.13
6 69 240 30 160 42 0 0.069 8.00 69 44 180 40 170 44 20 0.043 2.13
7 68 210 30 160 42 10 0.069 8.00 70 45 210 40 180 49 0 0.045 2.25
8 65 225 30 160 41 20 0.069 8.00 71 44 205 40 180 49 10 0.045 2.25
9 66 210 30 170 47 0 0.074 8.50 72 43 205 40 180 48 20 0.045 2.25
10 66 260 30 170 46 20 0.074 8.50 73 45 210 40 180 46 0 0.045 2.25
11 65 225 30 170 44 0 0.074 8.50 74 44 200 40 180 46 10 0.045 2.25
12 65 205 30 170 43 10 0.074 8.50 75 44 210 40 180 45 20 0.045 2.25
13 63 200 30 170 42 20 0.074 8.50 76 44 220 40 180 43 0 0.045 2.25
14 64 245 30 170 41 0 0.074 8.50 77 42 195 40 180 42 10 0.045 2.25
15 63 225 30 170 40 10 0.074 8.50 78 43 220 40 180 42 20 0.045 2.25
16 63 260 30 170 39 20 0.074 8.50 79 47 180 45 160 53 0 0.036 3.56
17 61 220 30 180 45 0 0.078 7.50 80 46 140 45 160 53 10 0.036 3.56
18 62 195 30 180 44 10 0.078 7.50 81 45 130 45 160 52 20 0.036 3.56
19 62 250 30 180 44 20 0.078 7.50 82 45 160 45 160 50 0 0.036 3.56
20 62 210 30 180 42 0 0.078 7.50 83 43 160 45 160 50 10 0.036 3.56
21 61 210 30 180 41 10 0.078 7.50 84 45 170 45 160 49 20 0.036 3.56
22 58 200 30 180 40 20 0.078 7.50 85 44 120 45 160 47 0 0.036 3.56
23 61 225 30 180 38 0 0.078 7.50 86 43 160 45 160 47 10 0.036 3.56
24 61 210 30 180 38 10 0.078 7.50 87 44 200 45 160 46 20 0.036 3.56
25 61 240 30 180 37 20 0.078 7.50 88 46 175 45 170 52 0 0.038 1.89
26 63 145 35 160 51 0 0.059 5.71 89 42 130 45 170 51 10 0.038 1.89
27 63 250 35 160 50 10 0.059 5.71 90 42 100 45 170 51 20 0.038 1.89
28 62 240 35 160 50 20 0.059 5.71 91 43 190 45 170 49 0 0.038 1.89
29 63 175 35 160 48 0 0.059 5.71 92 42 165 45 170 48 10 0.038 1.89
30 63 195 35 160 47 10 0.059 5.71 93 42 190 45 170 48 20 0.038 1.89
31 59 245 35 160 47 20 0.059 5.71 94 43 200 45 170 46 0 0.038 1.89
32 63 185 35 160 45 0 0.059 5.71 95 42 185 45 170 45 10 0.038 1.89
33 62 230 35 160 44 10 0.059 5.71 96 42 180 45 170 45 20 0.038 1.89
34 59 240 35 160 43 20 0.059 5.71 97 42 230 45 180 51 0 0.040 2.00
35 60 195 35 170 49 0 0.063 4.86 98 42 210 45 180 50 10 0.040 2.00
36 58 225 35 170 49 10 0.063 4.86 99 41 175 45 180 50 20 0.040 2.00
37 56 200 35 170 48 20 0.063 4.86 100 42 170 45 180 47 0 0.040 2.00
38 59 195 35 170 46 0 0.063 4.86 101 41 185 45 180 47 20 0.040 2.00
39 58 240 35 170 45 10 0.063 4.86 102 43 175 45 180 44 0 0.040 2.00
40 58 225 35 170 45 20 0.063 4.86 103 40 220 45 180 44 10 0.040 2.00
41 57 220 35 170 43 0 0.063 4.86 104 38 170 45 180 43 20 0.040 2.00
42 55 225 35 170 42 20 0.063 4.86
43 55 195 35 180 48 0 0.067 3.86
44 54 195 35 180 47 10 0.067 3.86
generally automatic or semiautomatic. For instance, in the
45 52 200 35 180 46 20 0.067 3.86
46 56 150 35 180 44 0 0.067 3.86 United States during the period 1966– 1980, the proportion
47 51 190 35 180 44 10 0.067 3.86 of manually batched concrete decreased from 54% to less
48 48 170 35 180 43 20 0.067 3.86 than 25% [16]. Ready-mixed concrete is defined as concrete
49 53 190 35 180 41 0 0.067 3.86 that is manufactured for delivery to a purchaser in a plastic
50 46 220 35 180 40 10 0.067 5.14
and unhardened state. In consideration of the current trend
51 48 210 35 180 40 20 0.067 5.14
52 51 170 40 160 52 0 0.040 4.00 of using ready-mixed concrete, concrete for the test was
53 49 95 40 160 52 10 0.040 2.57 mixed according to the relevant flowchart shown in Fig. 7.
54 49 220 40 160 51 20 0.040 4.00
55 50 210 40 160 49 0 0.040 4.00 3.5. Compressive strength test
56 49 205 40 160 49 10 0.040 4.00
57 49 220 40 160 48 20 0.040 4.00
58 50 230 40 160 46 0 0.040 4.00 Specimens for compressive strength test were made in
59 49 195 40 160 46 10 0.040 4.00 100  200-mm cylinder molds. The test was conducted in
60 47 210 40 160 45 20 0.040 4.00 accordance with ASTM C 684-95. The specimens were
61 49 205 40 170 51 0 0.043 2.13 demolded at 24 h and cured in nonlime water at 20 F 3 jC
62 48 195 40 170 50 10 0.043 2.13
for 28 days then tested.
 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420 415

Table 10 Table 10 (continued)

Tested Mixtures (80 – 120 MPa)
Number fcV Slump W/B W s/a SF SP
Number fcV Slump W/B W s/a SF SP (MPa) (mm) (%) (kg/m3) (%) (%) (kg/m3)
(MPa) (mm) (%) (kg/m3) (%) (%) (kg/m3)
63 99 200 25 155 37 15 11.70
1 115 185 18 140 35 15 31.10 64 93 200 25 155 37 20 14.20
2 122 200 18 140 35 20 32.10 65 89 230 25 160 39 5 16.80
3 123 190 18 140 35 25 36.50 66 91 215 25 160 39 10 13.40
4 113 215 18 145 35 15 27.70 67 86 230 25 165 39 5 16.80
5 116 220 18 145 35 20 28.70 68 98 230 25 165 39 10 15.70
6 119 210 18 145 35 25 28.70 69 92 225 25 165 39 15 16.60
7 109 220 18 150 35 15 28.20 70 88 180 27 155 39 5 10.90
8 115 200 18 150 35 20 33.90 71 93 190 27 155 39 10 12.90
9 117 200 18 150 35 25 34.40 72 94 220 27 155 39 15 15.80
10 105 210 20 140 35 15 21.90 73 87 220 27 160 39 5 11.90
11 109 185 20 140 35 20 21.40 74 91 200 27 160 39 15 13.80
12 119 210 20 140 35 25 31.30 75 90 235 27 165 39 5 15.40
13 112 190 20 145 37 10 19.30 76 90 235 27 165 39 10 13.90
14 118 190 20 145 37 20 22.50 77 89 220 27 165 39 15 15.30
15 104 205 20 145 35 15 20.90
16 107 210 20 145 35 20 20.90
17 115 220 20 145 35 25 26.10
18 106 190 20 150 37 10 16.60 3.6. Slump and air content test
19 106 190 20 150 37 15 18.10
20 106 185 20 150 37 20 19.50
ASTM C 143-90a procedure was followed to determine
21 105 210 20 150 35 15 20.90
22 108 200 20 150 35 20 27.70 the slump of the fresh concrete immediately after comple-
23 112 190 20 150 35 25 28.70 tion of mixing. The air content in fresh concrete was
24 103 205 20 155 37 10 18.10 measured according to ASTM C 231-91b.
25 104 190 20 155 37 15 18.60
26 105 200 20 155 37 20 21.00
27 104 210 22 140 35 15 16.50
28 108 200 22 140 35 20 21.40 4. Application of genetic algorithm and result
29 116 210 22 140 35 25 21.90
30 103 200 22 145 35 15 19.30 4.1. Determination of the fitness functions
31 105 220 22 145 35 20 20.90
32 112 220 22 145 35 25 25.00
When a specific mix proportion for high-performance
33 102 210 22 150 35 15 23.00
34 107 220 22 150 35 20 20.90 concrete needs to be determined, compressive strength and
35 108 220 22 150 35 25 26.10 slump are very important characteristics in the process of
36 104 210 23 145 37 10 16.10 design. The fitness functions essential for genetic algorithm
37 105 210 23 145 37 15 15.10 program are determined in terms of factors influencing com-
38 107 200 23 145 37 20 21.90
pressive strength and slump. Therefore, the factors influenc-
39 103 205 23 150 37 10 14.60
40 103 190 23 150 37 15 14.60 ing compressive strength and slump must be determined.
41 104 200 23 150 37 20 18.60
42 97 230 23 155 39 5 18.50
43 103 200 23 155 39 10 18.80
44 102 210 23 155 39 15 18.30
45 103 200 23 155 37 15 14.60
46 100 190 23 155 37 20 13.70
47 94 230 23 160 39 5 16.80
48 98 230 23 160 39 10 13.80
49 98 230 23 160 39 15 19.80
50 94 235 23 165 39 5 19.80
51 97 230 23 165 39 10 18.00
52 100 215 23 165 39 15 17.10
53 103 200 25 145 37 10 17.60
54 103 190 25 145 37 15 13.40
55 104 190 25 145 37 20 21.00
56 101 200 25 150 37 10 12.70
57 102 205 25 150 37 15 13.70
58 103 200 25 150 37 20 14.60
59 95 200 25 155 39 5 14.80
60 97 210 25 155 39 10 18.30
61 99 230 25 155 39 15 19.80
62 99 200 25 155 37 10 12.20
Fig. 7. Mixing flowchart.
416 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420

Table 11 Table 13
Fitness functions The coefficients of determination (R2) (80 – 120 MPa)
Classification Fitness The number Strength fitness Slump fitness
functions of data function (%) function (%)
40 – 80 MPa Compressive fcV = 122.14  0.84 W/B  0.42 W + 77 90.5 74.2
strength 0.34 s/a  0.09 FA + 363.77 AE 70 90.5 74.2
Slump Slump =  463.21  3.05 W/B + 65 90.5 74.0
5.21 W + 0.54 s/a + 0.11 FA  60 90.5 74.0
6541.17 AE + 41.67 SP 55 90.3 74.0
80 – 120 MPa Compressive fcV = 170.37  1.87 W/B  0.51 W + 50 90.1 73.3
strength 1.20 s/a  0.49 SF
Slump Slump =  387.01 + 16.94 W/B + 0.58 W 
1.23 s/a  0.377 SF + 1.163 SP increase of the water content tends to reduce compressive
fcV: compressive strength (MPa), Slump: slump of concrete (mm), W/B: strength.
water to binder ratio (%), W: water content (kg/m3), s/a: fine aggregate In a laboratory experiment, with a constant W/B ratio of
ratio (%), FA: fly ash replacement ratio (%), SF: silica fume replacement
0.60, when the coarse/fine aggregate proportion and the
ratio (%), AE: air-entraining agent content (kg/m3), SP: superplasticizer
content (kg/m3). cement content of a concrete mixture were progressively
raised to increase the slump from 50 to 150 mm, a 12%
decrease in the average 7-day compressive strength was
4.2. Considering factors influencing compressive strength observed.
and slump for the determination of the independent variables The concrete without fly ash usually shows lower
strength at 1 and 3 days, but strength gains can be substan-
The factors influencing compressive strength and slump tial after about 7 days of curing. Highly active pozzolans are
must be considered to determine the independent variables capable of producing high-strength in concrete at both early
of the fitness functions in order to determine mix propor- and late ages, especially when a water-reducing agent has
tions by applying genetic algorithm to the fitness functions been used to reduce the water requirement. It has been well
composed of selected independent variables. The factors to known that silica fume has an effect on enhancing the
be considered in the fitness functions of compressive strength of concrete and has been widely used in producing
strength and slump are explained in the following sections. high-strength concrete.
For the most part, it is the W/B ratio that determines the
4.2.1. Considering factors on compressive strength porosity of the cement paste matrix at a given degree of
The factors affecting compressive strength are water to hydration; however, when air voids are incorporated into the
binder ratio (W/B, %), water content (W, kg/m3), fine system, either as a result of inadequate compaction or
aggregate ratio (s/a, %), replacement ratio of fly ash (FA, through the use of an air-entraining admixture, they also
%), replacement ratio of silica fume (SF, %), and content of have the effect of increasing the porosity and decreasing the
air-entraining agent (AE, %) [16]. strength of the system. It has been observed that the extent
When discussing effect of W/B on compressive strength, of the strength loss as a result of entrained air depends not
W/B ratio – porosity relation is indispensable. From the only on W/B ratio of concrete mixture but also on the
standpoint of strength, W/B ratio– porosity relation is un- cement content. In short, as a first approximation, the
doubtedly the most important factor because, independent of strength loss due to entrained air can be related to the
other factors, it affects the porosity of both the cement paste general level of concrete strength. At a given W/B ratio,
matrix and the transition zone between the matrix and the high-strength concretes suffer a significant strength loss
coarse aggregate. W/B ratio – strength relationship in con- with increasing amounts of entrained air, whereas low-
crete can easily be explained as the natural consequence of a strength concretes tend to suffer only a little strength loss
progressive weakening of the matrix caused by increasing
porosity with an increase of the W/B ratio. Therefore, any
Table 14
Significant levels for fitness function
Table 12
Independent Significant levels
The coefficients of determination (R2) (40 – 80 MPa)
variables
The number Strength fitness Slump fitness 40 – 80 MPa 80 – 120 MPa
of data function (%) function (%) Strength Slump Strength Slump
104 95.5 88.9 W/B .000 .102 .000 .000
100 95.5 88.9 W .000 .000 .000 .580
90 95.3 88.7 s/a .000 .553 .008 .832
80 95.1 88.5 FA,SF .001 .706 .000 .000
70 94.9 88.5 AE .000 .000 – –
60 94.7 87.8 SP – .000 – .002
 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420 417

Fig. 8. Relation between error and the number of iteration.

or may actually gain some strength as a result of air thawing cycles. A side effect from entrained air is the
entrainment. improved workability of concrete mixtures, particularly with
those containing less cement and water, rough textured
4.2.2. Considering factors affecting the slump aggregate, or lightweight aggregate. In addition, entrained
The factors affecting the slump are W/B (%), W (kg/m3), air reduces drastically concrete permeability, absorptiveness,
s/a (%), FA (%), SF (%), AE (%), and content of super- and shrinkage age that are practical properties as important
plasticizer (SP, kg/m3) [17]. as compressive strength.
In practice, predicting the effect of mix proportions on Superplasticizer, also called high-range water-reducing
workability requires care since of the three factors, W/B admixtures because they are able to reduce three to four
ratio, aggregate/cement ratio, and water content, only two times more water in a given concrete mixture than normal
are independent. For instance, if aggregate/cement ratio is water-reducing admixtures, can be used to increase the
reduced but W/B ratio is kept constant, water content slump for a given W/B ratio.
increases and consequently workability also increases. If,
on the other hand, water content is kept constant when 4.3. Multiple regression modeling and fitness function
aggregate/cement ratio is reduced, then the W/B ratio
decreases and as a result workability decreases somewhat The 181 sets of mixtures were used for finding fitness
but is not seriously affected. functions of compressive strength and slump by multiple
Fly ash consists essentially of small spherical particles of regression modeling. For a system with n inputs (independent
aluminosilicate glass that is produced on combustion of variables) and one output (dependent variable) y, the general
pulverized coal in thermal power plants. With fresh concrete least square (or linear regression) problem is to find out the
mixtures that show a tendency to bleed or segregate, it is unknown parameters b of the linear model as shown in Eq. (4).
well known that incorporation of finely divided particles
generally improves workability by reducing the size and y ¼ b0 þ b1 x1 þ b2 x2 þ . . . . . . þ bn1 xn1 þ bn xn ð4Þ
volume of voids. The finer a mineral admixture, the less of it Independent variables of the fitness functions, selected in
will be needed to enhance the cohesiveness and the work- consideration of influence on compressive strength and
ability of a freshly mixed concrete. The small size and the slump, were W/B (%), W (kg/m3), s/a (%), FA (%), SF
glassy texture of fly ash can possibly reduce the amount of (%), and AE (%) for dependent variable of compressive
water required for a given consistency. Compared to normal strength ( fcV) and W/B, W, s/a, FA, SF, AE, and SP (kg/m3)
Portland cement and typical fly ash, silica fume shows for dependent variable of the slump of concrete (slump, cm),
particle size distributions that are two orders of magnitude
finer. This is why on the one hand the material is highly Table 15
pozzolanic, but on the other hand it creates problems of Tested mixtures (40 – 80 MPa)
handling and increases appreciably the water requirement in Number fcV Slump W/B W s/a FA AE SP
concrete unless water-reducing admixtures are used. In (MPa) (mm) (%) (kg/m3) (%) (%) (kg/m3) (kg/m3)
conclusion, silica fume has an effect on decreasing the 1 41 205 45 180 47 10 0.040 2.00
workability of high-performance concrete. 2 57 230 35 170 42 10 0.063 4.86
The most important application of air-entraining admix- 3 66 195 30 170 46 10 0.074 8.50
ture is for concrete mixtures designed to resist freezing and 4 75 205 30 160 49 0 0.069 8.00
418 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420

Table 16
Tested mixtures (80 – 120 MPa)
Number fcV Slump W/B W s/a SF SP
(MPa) (mm) (%) (kg/m3) (%) (%) (kg/m3)
1 88 245 27 160 39 10 13.80
2 96 225 25 160 39 15 16.80
3 102 202 23 155 37 10 18.60
4 112 190 20 145 37 15 19.00

respectively. In this study, SPSS version 10 was used for

multiple regressions [18,19]. The modeling results (fitness
function) for compressive strength and slump are listed in
Table 11 together with the significant levels for fitness
function in Table 14. Generally, the coefficients of determi-
nation (R2), over 70%, are required to verify the compati-
bility of a regression model. In addition, the change of the Fig. 10. Relation between error and the number of initial individuals (80 –
coefficients of determination (R2) according to the change of 120 MPa).
the number of data needs to be observed to verify the
compatibility of the number of data. The compatibilities of determined. In this research, ranking selection based on
a regression model and the number of data were examined, normalized geometric distribution was used. The number of
as shown in Tables 12 and 13. The coefficients of determi- selected initial individuals was changed from 15 to 75
nation (R2) converged. Therefore, it was concluded that because the change of number of initial individuals is useful
those were compatible. for finding an optimum value; that is, too large or too small
number of initial individuals leads to the local solutions.
4.4. Application and verification of genetic algorithm Crossover was performed based on one-point, two-point,
and uniform crossover theory. Four mutation methods were
Based on the programs made by Houck et al. [29], four used in this study: boundary mutation, multi-nonuniform
genetic algorithm programs were developed for finding the mutation, nonuniform mutation, and uniform mutation.
design of high-performance concrete mixtures using MAT- Each input and output data of the program consist of
LAB version 5.1 [20,21]. In the first program—called compressive strength and slump. In the programs, after
GAST1—W/B, W, s/a, FA, and AE for a specific compres- creating initial population of strings composed of W/B, W,
sive strength in the range of 40 to 80 MPa were determined. s/a, FA, SF, AE, and SP, fitness increases through selection,
In GASLUMP1 program, using the results of GAST1, SP crossover, and mutation. When errors between input data
for a specific slump was determined. In the third program, and output data become minimized, fitness is satisfied
called GAST2, W/B, W, s/a, and SF for a specific com- (Table 14). Then, the repeating process is terminated and
pressive strength in the range of 80 to 120 MPa were optimal solution is approached. The fitness functions for
determined. In the fourth program, called GASLUMP2, compressive strength and slump listed in Table 11 were used
using the results of GAST2, SP for a specific slump was in GAST1, GASLUMP1, GAST2, and GASLUMP2 in
order. Fig. 8 shows that an error decreases when the number
of iteration increases.
To verify the accuracy and usefulness of the genetic
algorithm programs, eight sets of test mixtures listed in
Tables 15 and 16 were compared to the results obtained in
the genetic algorithm programs.

Table 17
The results from genetic algorithm and error (40 – 80 MPa)
Number W/B W s/a FA AE SP Average
(%) (kg/m3) (%) (%) (kg/m3) (kg/m3) error (%)
1 (error %) 44 179 46 10 0.045 2.30 5.40
(2.22) (0.56) (2.13) (0) (12.50) (15)
2 (error %) 35 171 42 10 0.065 5.56 3.03
(0) (0.59) (0) (0) (3.17) (14.40)
3 (error %) 31 164 46 10 0.070 8.15 2.73
(3.33) (3.53) (0) (0) (5.41) (4.12)
Fig. 9. Relation between error and the number of initial individuals 4 (error %) 30 160 53 0 0.075 9.17 5.25
(40 – 80 MPa). (0) (0) (8.16) (0) (8.70) (14.63)
 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420 419

Table 18 1. By applying the genetic algorithm for design of high-

The results from genetic algorithm and error (80 – 120 MPa)
performance concrete mixtures, the number of trial
Number W/B W s/a SF SP Average mixtures with desired properties in the field can be
(%) (kg/m3) (%) (%) (kg/m3) error (%)
reduced. This method is a new and alternative method to
1 (error %) 27 161 37 10 14.30 1.88 the current design method that has no exact design
(0) (0.63) (5.13) (0) (3.62)
criteria.
2 (error %) 24 163 38 15 16.24 2.35
(4) (1.88) (2.56) (0) (3.33) 2. Two small or large numbers of initial individuals made
3 (error %) 22 150 37 10 17.57 2.62 genetic algorithm find local solutions. The numbers of
(4.35) (3.23) (0) (0) (5.54) initial individuals for finding a global solution—the
4 (error %) 20 141 38 14 20.95 4.48 design of high-performance concrete mixtures with
(0) (2.76) (2.70) (6.67) (10.26)
specific compressive strength and slump—were 30
and 35.
3. It is believed that the errors of chemical admixtures are
After introducing specific compressive strength and not so large because the values of two chemical
slump, the genetic algorithm programs for the mix propor- admixtures are small and have decimal values. Therefore,
tioning of high-performance concrete were performed. To though the values change on a very small scale, the errors
find mix proportions with a minimum of error, the number of two chemical admixtures tend to have large values
of initial individuals to make a value with a minimum error when the errors are expressed in the unit of percentage.
in compressive strength and slum was changed from 15 to 4. The applicability of genetic algorithm for the mix design
75. The mix proportion with a minimum average error in of high-performance concrete was observed in this paper,
compressive strength and slump was obtained with a num- and a new method for design of high-performance
ber of initial individuals of 30 and 35. Figs. 9 and 10 show, concrete mixtures using genetic algorithm was provided.
respectively, the relation between the average of error in
compressive strength and slump and the number of initial
individuals. Acknowledgements
The average errors in compressive strength and slump
and comparative results varied with the mixtures from the The authors would like to express their gratitude to the
genetic algorithm programs, and the eight sets of test Korea Institute of Science and Technology Evaluation and
mixtures are listed in Tables 17 and 18. The range of errors Planning (KISTEP) for the financial support of the National
in W/B, W, s/a, F/A, S/F, AE, and SP are 0– 4%, 0 –4%, 0 – Research Laboratory (NRL).
8%, 0%, 0– 7%, 4 –13%, and 3 –10%, respectively. The
results show that the errors in W/B, W, s/a, F/A, and S/F are
smaller than those of AE and SP. The values of two References
chemical admixtures are small and have decimal values.
Therefore, though the values change on a very small scale, [1] A.M. Neville, P.C. Aitcin, High-performance concrete—An overview,
the errors of two chemical admixtures tend to have large Mat. Struct. 31 (3) (1998) 111 – 117.
values when the errors are expressed in the unit of percent- [2] P.C. Aitcin, A.M. Neville, High-performance concrete demystified,
Concr. Int. 1 (1993) 21 – 26.
age. Therefore, the errors of chemical admixture are con-
[3] P.A. Gutierrez, M.F. Canovas, High-performance concrete: require-
sidered to be not so large. ments for constituent materials and mix proportioning, ACI Mater. J.
5 – 6 (1996) 233 – 241.
[4] H.G. Russell, ACI defines high-performance concrete, Concr. Int. 21
5. Conclusions (2) (1999) 56 – 57.
[5] P. Zia, M.L. Leming, S.H. Ahmad, High-Performance Concretes, A
State-of-the Art Report, SHRP, National Research Council, Washing-
The object of this research is to propose a mix design ton, DC, 1991, SHRP-C/FR-91-103.
method for high-performance concrete using genetic algo- [6] P. Zia, M.L. Leming, S.H. Ahmad, J.J. Schemmel, R.P. Elliot, A.E.
rithm, which is a stochastic search technique based on the Naaman, Mechanical Behaviour of High-Performance Concretes,
mechanism of natural selection and natural genetics. In this SHRP, National Research Council, Washington, DC, 1993,
SHRP-C-361.
paper, genetic algorithm was developed to present a new
[7] S.W. Forster, High-performance concrete—Stretching the paradigm,
design method for high-performance concrete mixtures to Concr. Int. 16 (10) (1994) 33 – 34.
minimize the number of trial mixes and provides a [8] R.N. Swamy, High-performance and durability through design, Inter-
reasonable mix proportion. Experimental and analytic national Workshop on High-Performance Concrete, ACI, Farmington
investigations were carried out to develop genetic algo- Hills, MI, 1996, pp. 209 – 230, SP-159.
[9] P.K. Metha, P.C. Aitcin, Principles underlying the production of high-
rithm for mix proportioning of high-performance concrete
performance concrete, cement, concrete, and aggregates, ASTM 12
and to verify the proposed mix design. Based on the (2) (1990) 70 – 78.
results of this research, the following conclusions were [10] P.C. Aitcin, The art and science of high-performance concrete, in:
drawn. O.K. Metha (Ed.), Advances in Concrete Science and Technology,
420 C.-H. Lim et al. / Cement and Concrete Research 34 (2004) 409–420

Proceedings of the M. Collepardi Symposium, Rome vol. 10, 1997, [20] MATLAB version 5.1, The Math Works, 1997.
pp. 107 – 124. [21] J.S. Lim, Perfect Guide for MATLAB version 5.1, Ajin Press, South
[11] Federal Highway Administration, High-Performance Concrete Korea, 2000, pp. 14 – 126.
Bridges, Building Bridges for the 21st Century, U.S. Department of [22] J.H. Holland, Adaptation in Natural and Artificial Systems, University
Transportation, FHWA-SA-98-084. of Michigan Press, Ann Arbor, 1975.
[12] C.H. Goodspeed, S. Vanikar, R. Cook, High-performance concrete [23] L.G. Caldas, L.K. Norford, A design optimization tool based on a
defined for highway structures, Concr. Int. 18 (2) (1996 Feb.) 62 – 67. genetic algorithm, Autom. Constr. 11 (2) (2002) 173 – 184.
[13] K.E. Hassan, J.G. Cabrera, R.S. Maliehe, The effect of mineral ad- [24] P. Nanakorn, K. Meesomklin, An adaptive penalty function in genetic
mixtures on the properties of high-performance concrete, Cem. Concr. algorithms for structural design optimization, Comput. Struct. 79
Compos. 22 (4) (2000) 267 – 271. (29 – 30) (2001) 2527 – 2539.
[14] B.H. Bharatkumar, R. Narayanan, B.K. Raghuprasad, D.S. Rama- [25] J.C. Miles, G.M. Sisk, C.J. Moore, The conceptual design of commer-
chandramurthy, Mix proportioning of high-performance concrete, cial buildings using a genetic algorithm, Comput. Struct. 79 (17)
Cem. Concr. Compos. 23 (1) (2001) 71 – 80. (2001) 1583 – 1592.
[15] P.A. Gutierrez, M.F. Canovas, High-performance concrete: require- [26] E. Cantú-Paz, D.E. Goldberg, Efficient parallel genetic algorithms:
ments for constituent materials and mix proportioning, ACI Mater. theory and practice, Comput. Methods Appl. Mech. Eng. 186 (2 – 4)
J. (1996 May – June) 233 – 241. (2000) 221 – 238.
[16] P.K. Metha, Concrete-Structures, Properties, and Materials, 2nd ed., [27] K. Deb, An efficient constraint handling method for genetic algo-
Prentice-Hall, New Jersey, 1993, pp. 43 – 56. rithms, Comput. Methods Appl. Mech. Eng. 186 (2 – 4) (2000)
[17] A.M. Neville, Properties of Concrete, 3rd ed., Pitman, London, 311 – 338.
1981, pp. 203 – 207. [28] D. Goldberg, Genetic Algorithms in Search, Optimization and Ma-
[18] SPSS (Statistical Packages for Social Science) version 10, SPSS, chine Learning, Addison-Welsley, Reading, MA, 1989.
1999. [29] C.R. Houck, J. Joines, M. Kay, A genetic algorithm for function
[19] J.S. Choi, Manual for Statistical Analysis by SPSS (Statistical optimization: a MATLAB implementation, ACM Trans. Math. Softw.
Packages for Social Science) Version 10, Bogdoo Press, South (1996).
Korea, 2000, pp. 212 – 334.