ACI MATERIALS JOURNAL TECHNICAL PAPER

Title No. 117-M22

Statistical Optimization of Ultra-High-Performance Glass

Concrete
by Joaquín Abellán, Jaime Fernández, Nancy Torres, and Andrés Núñez

This paper presents the experimental results of research carried out density, high amounts of ultrafine SF (25 to 30% of cement
involving the compressive strength and slump flow of ultra-high- weight) are used to fill the voids between cement particles
performance concrete (UHPC) made with cementitious blends of in UHPC. This significantly increases the concrete cost and
recycled glass flour, recycled glass powder, micro limestone powder, decreases its workability. All these drawbacks restrict the
silica fume, and portland cement. The adopted second-order poly-
wider use of UHPC in the construction market.7,19
nomic regression model provided an accurate correlation between
In recent years, several researchers have focused on
the considered variables and the obtained responses. A numerical
optimization was then performed to obtain an eco-friendly mixture eco-friendly and cost-efficient UHPC. Li et al.20 proved
with the proper flow, highest compressive strength, and minimum that incorporating limestone powder in UHPC improved
content of cement. The use of 603 kg/m3 of cement in the mixture the hydration process at early age, leading to denser particle
can be considered as the most appropriate amount to be employed packing and improving mechanical properties. Huang and
in UHPC mixtures, fulfilling the limit values of compressive Cao21 used nano-CaCO3 and reported a 17% increase in
strength and spread flow. compressive strength compared to the control UHPC spec-
imens. As awareness for protecting the environment has
Keywords: compressive strength; optimization; recycled glass concrete;
risen, the possibility of converting solid waste into concrete
response surface methodology (RSM); sustainability; ultra-high-perfor-
mance concrete (UHPC). components has drawn more interest.22 As a result, some
waste materials such as recycled glass powder are also
INTRODUCTION included in the production of UHPC to reduce its cost and
In recent years, enormous developments in concrete tech- improve its environmental attributes. In this field, Vaitkev-
nology have achieved the development of new concrete icius et al.23 investigated the effect of glass powder on the
types, such as ultra-high-performance concrete (UHPC).1 microstructure of UHPC. Their results revealed that glass
UHPC is defined as a new high-tech material, with superior powder increases the dissolution rate of portland cement
mechanical properties such as a compressive strength of over submitted to heat treatment, thus the hydration process is
150 MPa,2-4 and improved durability due to its extremely accelerated. Soliman and Tagnit-Hamou1 investigated the
low porosity.4-6 A typical UHPC mixture contains portland partial substitution of silica fume with fine glass powder
cement (CEM), silica fume (SF), quartz powder (QP), silica (FGP) in UHPC. The results demonstrated that compressive
sand (SS) with a maximum size of 600 μm, high-range water strength values of 235 and 220 MPa after 2 days of steam
reducers (HRWRs), and sometimes steel fiber.7-9 The fiber curing could be obtained, respectively, when replacing 30
inclusion in UHPC improves the material’s ductility, as well and 50% of SF with FGP with a mean particle size (d50)
as its tensile and flexural capacity.10,11 Currently, UHPC is of 3.8 µm. However, the amount of cement used exceeded
used in the construction of footbridges, precast deck panel 800 kg/m3. In another study, Tagnit-Hamou et al.7 and
bridge joints, special prestressed and precast concrete Soliman and Tagnit-Hamou11 used recycled glass powder
elements, concrete repair, marine platforms, precast walls, to replace quartz sand, cement, and quartz powder particles.
urban furniture, overlay on damaged pavements and indus- Aiming for a cost-effective and more sustainable UHPC,
trial floors, and other architectural applications.11-15 the objective of this study is to effectively design and produce
Today, the sustainability of the construction sector is a UHPC with a low cement content, a maximum content of
priority for the scientific community.16 For this reason, the silica fume of 100 kg/m3, and avoiding the use of QP. The
production of cement-based materials must consider not design of the concrete mixtures was based in achieving
only good mechanical and durability properties, but also great compressive strength and rheological properties with
environmentally friendly, ecological, and socioeconomic a minimum amount of cement through a three-factor Design
benefits.7,17 A typical UHPC design8 has a cement content of Experiments (DoE). To ensure a densely compacted
that varies from 800 to 1000 kg/m3. This high cement cementitious matrix, the modified Andreasen & Andersen
content not only affects production costs and consumes particle packing model (A&Amod)24 was used. QP was totally
natural resources; it also negatively affects the environment
ACI Materials Journal, V. 117, No. 1, January 2020.
because the production of 1 ton of portland cement releases MS No. M-2019-044, doi: 10.14359/51720292, received February 1, 2019, and
approximately 1 ton of CO2.7,18 QP has an immediate and reviewed under Institute publication policies. Copyright © 2020, American Concrete
Institute. All rights reserved, including the making of copies unless permission is
long-term harmful effect on human health as it is classified obtained from the copyright proprietors. Pertinent discussion including author’s
closure, if any, will be published ten months from this journal’s date if the discussion
as a carcinogen agent.7,11 To achieve the optimum packing is received within four months of the paper’s print publication.

ACI Materials Journal/January 2020 243

replaced by recycled glass flour. Moreover, supplementary Table 1—Chemical characterization of cement, silica
cementitious materials such as micro limestone powder and fume, micro limestone powder, and recycled glass
recycled glass powder were used as partial replacements of Micro
cement and silica fume in the concrete. Chemical Silica limestone Recycled Silica
With the concept of DoE, a set of well-selected experiments analysis, % Cement fume powder glass sand
were used. The aim of this so-called design is to optimize a Specific
process or system by performing each experiment and draw gravity, 3.16 2.20 2.73 2.55 2.65
conclusions regarding the behavior of the studied object g/cm3
from the results of the experiments. Considering the cost Loss of
2.58 0.60 42.21 1.00 0.20
of a single experiment, minimizing the amount performed ignition, %
is always something to consider. With DoE, the number of SiO2, % 19.42 92.29 0.90 72.89 99.80
tests to be performed is kept as low as possible and the most
Al2O3, % 4.00 0.59 0.10 1.67 0.14
informative combination of the factors is chosen. Hence,
DoE is an effective and economical solution.25 To summa- CaO, % 64.42 3.89 55.51 9.73 0.17
rize, the advantages of using an experimental design method MgO, % 1.52 0.26 0.70 2.08 0.01
include: 1) the construction of an empirical model with the SO3, % 1.93 0.07 0.10 0.01 —
relevant parameters and their corresponding response; 2) the
Na2O, % 0.19 0.31 0.03 12.54 —
reduction of the number of experiments; 3) the assessment of
the interaction between parameters; and 4) the determination K2O, % 0.39 0.54 0.00 0.76 0.05
of the optimal response within the experimental region data.2,26 TiO2, % 0.38 0.01 0.00 0.04 —
On the other hand, Response Surface Methodology (RSM) Mn3O4, % 0.05 0.01 0.01 0.01 —
is a powerful experimental design technique for the modeling
Fe2O3, % 3.61 0.24 0.05 0.81 0.04
and analysis of problems in which a response of interest is
influenced by several variables.2,27,28 This method has been
widely used for experimental process optimization in the with by-products of local industry such as recycled glass
UHPC research. Ghafari et al.2 used a central composite used as partial substitution for cement and silica fume. Waste
response surface to predict the performance of self-consol- glass is a non-biodegradable material, so reusing it means
idating UHPC reinforced with hybrid steel fibers on prop- less material to be stockpiled in landfills. The methodology,
erties such as slump flow, V-funnel flow time, and flexural based on the particle packing theory and statistical design
strength. Mosaberpanah and Eren29 used RSM to create approaches, results in an optimized eco-friendly and cost-
models that included results from compressive strength effective dosage of ultra-high-performance glass concrete.
tests at the age of 7, 14, and 28 days, and 28-day splitting
tensile and modulus of rupture tests for UHPC. In other EXPERIMENTAL INVESTIGATION
research,30 they investigated the effect of QP, quartz sand Materials
(QS), and different water curing temperatures on mechan- The materials used to manufacture the concrete were
ical properties of UHPC, also through an analysis by DoE. locally available in Colombia. ASM Type HE cement was
Sayed Ahmed and Sennah13 and Sayed Ahmed et al.31 devel- used. The HE cement had a specific gravity of 3.15, and
oped an RSM model to optimize different UHPC mixtures mean particle diameter (d50) of 8 μm. The SF used in the
to reach the desired strength and rheology for joint precast mixture proportioning complied with ASTM C1240 specifi-
deck applications. Van and Ludwig32 proposed an analytical cations and had a specific gravity of 2.20 and d50 of 0.15 μm.
model based on RSM to optimize the replacement of cement The UHPC was also designed with a SS of a specific gravity
(partial) and silica fume (total) through a three-component of 2.65, maximum particle size (dmax) of 600 μm and d50 of
binder containing rice husk ash (RHA) and ground-granu- 165 μm. Recycled glass flour (GF) with a specific gravity of
lated blast-furnace slag (GGBFS). 2.55 and d50 of 28 μm was used as a filler material. Micro
In this study, a multi-objective simultaneous optimiza- limestone powder (MLP) with a specific gravity of 2.73 and
tion technique was adopted, also aiming at determining d50 of 2 μm and FGP with a specific gravity of 2.55 and d50
the optimum settings of the mixture variables that result of 7 μm were used as supplementary cementitious materials.
in a proper flow with the maximum compressive strength Glass powder and glass flour were obtained by grounding
(exceeding 150 MPa) with the lowest content of cement locally available recycled glass with a jet mill to different
possible. Finally, the accuracy of the model was checked degrees of fineness, by applying different grinding speeds.
through comparison with selected experimental work results. Table 1 shows the chemical composition of the materials
used in this study. Figure 1 provides the particle size distri-
RESEARCH SIGNIFICANCE bution (PSD) of the cement, SF, MLP, GP, GF, and SS. A
UHPC research has been moving forward amazingly polycarboxylate (PCE)-based HRWR with a specific gravity
quickly in recent years, but the environmental and cost of 1.07 and solid content of 40% was used as a HRWR.
issues are still present. Optimizing the mixture design is
a challenge when looking for a minimum compressive Specimens
strength of 150 MPa and proper slump flow, while dealing The mixture design involved several steps. The amount
of cement, water, and HRWR was established by the DoE.

244 ACI Materials Journal/January 2020

Fig. 1—Particle size distribution of used materials, target, and mixture curve.
The proportion of the other components was guided by the
A&Amod, using a q value of 0.264, according to Eq. (1)

P ( D) = �
(D q q
− � Dmin ) (1)
(Dq
max −D q
min )
where D is the particle size; P(D) is the weight fraction of
total solids that are smaller than D; Dmax and Dmin are the
maximum and minimum particle sizes, respectively; and q
is the Fuller exponent. The value of q was determined in
previous research.33
A Central Composite Design (CCD) is an efficient exper-
imental design used in response surface methodology to
construct a second-order polynomic model for each response.
The CCD has several advantages, such as the ability to esti-
mate the quadratic effect of each response, the analysis of the
response surface with a relatively small number of experi-
mental runs, the determination of the interrelations between Fig. 2—Central composite design for three factors (A, B,
factors, and the ability to locate the optimal response.2,34 The and C) at two levels.
second-order model is widely used in this methodology for
the cube; and c center points with all levels set to coded level
the following reasons: 1) it works on a wide variety of func-
0, which are in the center of the cube. A graphic of a three-di-
tional forms, so it will often perform as an approximation to
mensional CCD for k = 3 independent variables is shown in
the true response surface; 2) it is easy to estimate its param-
Fig. 2. A design with three factors, four center points, eight
eters—β in Eq. (2)—in the second-order model; and 3) there
factorial points, and six axial points (α = ±1.78885) with a
is considerable practical experience indicating that second-
total of 18 set points, was used in this research.
order models work well in solving real response surface prob-
The variables of the experiments were coded according
lems. A CCD consists of a full or fractional factorial design
to Eq. (2)
with center points that are enlarged with star points, which
increase the variable space and allow the estimation of the
quadratic terms. The CCD structure for a k factor three-level Xj =
(Z j − Z0 j )� (2)
experiment design requires 2k + 2k + c design points, where ∆j
k is the number of studied variables, with 2k factorial points
where Xj is the coded value of the independent variable; Zj
representing all combinations of coded values x = ±1, which
is the real value of the independent variable; Z0j is the real
are in the corners of a cube; plus 2k axial points at a distance
value of the independent variable at the center point; and Δj
± α from the origin, which are in the center of each face of
is the step change value.

ACI Materials Journal/January 2020 245

 A CCD was used to estimate the response calculated some tests were immediately conducted, still on fresh state,
according to Eq. (3) in the form of a second-degree polyno- to determine the static slump flow diameter in accordance
mial equation35 to the ASTM C1437 specification.36 In the slump flow test,
a truncated conic mold was placed on a smooth plate, filled
Y = β0 + ∑ βi xi + ∑ βii xi2 + ∑ βij xi x j (3) with UHPC and lifted upward. The spread diameter of the
mortar was measured in four perpendicular directions, and
where Y is the predicted response variable; β0 is the intercep- the average of the diameters was reported as the spread flow
tion; βi are linear coefficients; βii are quadratic coefficients; of the concrete (Øm) in mm, according to Eq. (4)
βij are coefficients of the interaction; and xi and xj represent
the chosen independent variables. The variables defined for 1 4
Øm = ∑Øi (4)
this experiment and their control levels are shown in Table 2. 4 i =1
To summarize, a central composite design with a total of
18 set point combinations, with a randomly chosen order, After performing the workability test, the UHPC was cast
according to a CCD configuration of three independent vari- in molds and compacted on a vibrating table. The prisms
ables, was conducted. The corresponding mixture propor- were demolded approximately 24 hours after casting and
tions of this DoE, adjusted according to the A&Amod curve, then cured in a moisture room at 20°C until the day of the
are presented in Table 3. test, without any heat treatment applied. For the determina-
tion of the compressive strength, 50 mm cubes were tested.
Items of investigation A compression testing machine with a capacity of 3000 kN
Once the dosage was designed, a 5 L mortar mixer was used was used, following ASTM C109.37 Three samples were
to produce the concrete. After the mixing was completed, tested for each of the following ages: 24 hours, 7 days, and
28 days.
Table 2—Variables and range of variation To summarize, three independent variables were chosen:
Range of variation
cement content in kg/m3, coded as Factor A; water-binder
ratio (w/b), coded as Factor B; and dosage of HRWR, coded
Variables Coded –1.789 –1 0 1 1.789
as Factor C. To determine the proportion of the rest of the
Cement,
A 580.279 600.000 625.000 650.000 669.721 concrete components, the A&Amod curve was used. Four
kg/m3 responses were also considered, including compressive
w/b B 0.156 0.16 0.165 0.17 0.174 strength with standard curing conditions at 24 hours (R1),
HRWR, % 7 days (R7), 28 days (R28), and spread flow (Øm).
C 1.842 2 2.2 2.4 2.556
vol.

Table 3—Proportion of mixture components expressed as function of weight of cement

Particle
Micro limestone packing
Run Cement Silica fume Glass flour Glass powder powder HRWR Sand w/b density
1 1 0.167 0.313 0.313 0.225 0.036 1.552 0.160 0.793
2 1 0.154 0.246 0.246 0.172 0.033 1.510 0.160 0.790
3 1 0.167 0.313 0.313 0.225 0.036 1.499 0.170 0.790
4 1 0.154 0.246 0.246 0.172 0.033 1.463 0.170 0.788
5 1 0.167 0.313 0.313 0.225 0.043 1.548 0.160 0.792
6 1 0.154 0.246 0.246 0.172 0.040 1.506 0.160 0.790
7 1 0.167 0.313 0.313 0.182 0.043 1.553 0.170 0.791
8 1 0.154 0.246 0.246 0.135 0.040 1.459 0.170 0.786
9 1 0.160 0.270 0.270 0.189 0.038 1.539 0.165 0.792
10 1 0.160 0.270 0.270 0.189 0.038 1.539 0.165 0.792
11 1 0.172 0.323 0.323 0.231 0.041 1.625 0.165 0.796
12 1 0.149 0.224 0.224 0.161 0.035 1.434 0.165 0.787
13 1 0.160 0.270 0.270 0.189 0.038 1.583 0.156 0.794
14 1 0.160 0.270 0.270 0.189 0.038 1.495 0.174 0.791
15 1 0.160 0.270 0.270 0.19 0.032 1.543 0.165 0.792
16 1 0.160 0.270 0.270 0.189 0.044 1.535 0.165 0.792
17 1 0.160 0.270 0.270 0.189 0.038 1.539 0.165 0.792
18 1 0.167 0.313 0.313 0.225 0.038 1.539 0.165 0.792

246 ACI Materials Journal/January 2020

Table 4—Set point combinations and Table 5—Results for developed regression models
corresponding experimental responses Lack Model Shapiro
Ø m, R1, R7, R28, Responses R2 Adj-R2 RMSE F-Stat of fit P-value P-value
Run A B C mm MPa MPa MPa Øm 0.943 0.926 3.59 53.92 4.22 <0.0001 0.389
1 –1 –1 –1 221.25 27.01 104.72 139.84 R1 0.933 0.874 1.99 15.58 2.82 0.0002 0.168
2 1 –1 –1 203.00 40.84 77.56 100.68 R7 0.956 0.917 2.04 24.60 2.89 <0.0001 0.072
3 –1 1 –1 265.00 46.27 105.38 134.84 R28 0.925 0.894 1.24 22.52 1.60 <0.0001 0.172
4 1 1 –1 256.50 56.18 118.14 134.62
5 –1 –1 1 245.75 45.93 121.55 148.72 statistical computing R.38 The results are given in Table 5.
The significance of the model was determined by using both
6 1 –1 1 219.00 45.81 99.93 127.29
F-test and P-value. The P-values lower than 0.0001 were
7 –1 1 1 274.00 24.20 83.28 141.05 obtained for all the models, implying their significance.
8 1 1 1 264.25 9.90 90.66 156.42 The analysis of results for full regression models
9 0 0 0 247.75 47.72 79,99 143.51 included the determination coefficient (R2), the adjusted
coefficient of multiple determinations (Adj-R2), the root-
10 0 0 0 240.25 47.81 96,93 123.65
mean-square error (RMSE), the F Statistic value, the lack-
11 –1.789 0 0 239.00 27.78 101.70 140.00 of-fit p-value, the model P-value, and the Shapiro-Wilk
12 1.789 0 0 261.00 50.67 107.60 148.00 P-value given in Table 5. The models presented a high
13 0 –1.789 0 203.50 46.84 97.44 129.76 determination coefficient (R2) showing 94%, 93%, 96%,
and 92% of the variability in the responses Øm, R1, R7,
14 0 1.789 0 278.75 40.63 110.24 162.64
and R28, respectively. This indicates the goodness of fit
15 0 0 –1.789 233.50 52.49 113.93 153.86 of the model as well as its high statistical significance.
16 0 0 1.789 250.75 18.24 94.63 154.54 It shows the high correlation between the experimental
17 0 0 0 239.25 39.39 98.54 142.92 and predicted values. These high values of the coefficient
validate the adequacy of the model used along the design
18 0 0 0 231.00 37.38 76.73 152.84
space. Also, the adjusted R2 values were very close to the
R2, which shows that no unnecessary terms were added
EXPERIMENTAL RESULTS AND DISCUSSION
to the model. The adequacy of the model was verified
Model fitting and validation
by performing a lack-of-fit test, being the desired result
Table 4 presents the arrangement of all the set point combi-
an insignificant lack-of-fit. All the P-values obtained by
nations and the corresponding experimental response values
the ANOVA implied that the lack-of-fit was insignificant
obtained. R, a programming language and environment used
compared to the pure error. In addition, the Shapiro-Wilk
in statistical computing,38 as well as in experiment design
test showed P-values higher than 0.05 in all cases.
and analysis, was used to plan the experiment.
For each response (R1, R7, R28, and Øm) a second-order
Parametric study using response surface methods
polynomial model based on Eq. (3) was created from the
(RSM)
CCD data. The process of modeling involved two main
Once the model was validated, the effect of each param-
steps: calculating an appropriate model, and verifying its
eter on the response was investigated. The ANOVA was used
efficiency. A backward stepwise process was used to achieve
to assess the effect of individual parameters as well as the
an accurate model for each response. It starts with the full
interaction of variables on the response. The significance
second-order polynomial model, Eq. (3), used to estimate
of each parameter was judged by determining if a P-value
the relationship between the variables and the responses
was below the limit value (usually 0.05), which indicates
based on experimental results from the CCD. Then, the
that the terms are significant and that their contribution
process is followed by removing the variable with the largest
improves the model. The coefficient of the model (estima-
P-value. The procedure continues until only those variables
tion parameter) and its corresponding P-values are presented
which are significant (P-value < 0.05) remain in the model.
in Table 6. Model terms with a P-value higher than 0.05
After removing each term, the fitting process is repeated
were not considered as statistically significant, so they were
until all the insignificant terms have been removed from the
removed. RSM can be graphically illustrated by presenting
model. After the model has been generated, the next step is
the response as a function of the factors. Figures 3, 4, 5, and
to evaluate its efficiency by performing a residual analysis.
6 show the response surface plots in the restrained region
The latter involves statistical calculation such as residual
in function of Factors B (w/b) and C (HRWR content in
standard deviation, as well as residual plots, in which the
volume) while the cement content was fixed at 625 kg/m3
adequacy of the selected model can be graphically evalu-
(A = 0). This plot provides information on the effect of the
ated. If the test demonstrates the adequacy of the model, the
two independent components on average spread (Øm) and
response surface counter can be plotted.28
compressive strength with standard curing conditions at
To investigate the significance of the model, an analysis
24 hours (R1), 7 days (R7), and 28 days (R28).
of variance (ANOVA) was performed using the language for

ACI Materials Journal/January 2020 247

Table 6—ANOVA results for fitted numerical models
Øm R1 R7 R28
Model terms Coeff. P-value Coeff. P-value Coeff. P-value Coeff. P-value
Intercept 250.509 <0.0001 35.658 <0.0001 101.650 <0.0001 157.198 <0.0001
A –3.123 0.0148 1.852 0.0340 3.693 0.0009 2.401 0.0001
B 13.680 <0.0001 –4.493 0.0002 –1.946 0.0307 –0.611 0.0441
C 5.332 0.0004 –4.321 0.0003 –6.931 <0.0001 0.744 0.0144
A:B — — — — 2.571 0.0328 — —
A:C — — –3.761 0.0044 3.836 0.0045 4.197 <0.0001
B:C — — –3.783 0.0042 — — — —
A:B:C — — — — 3.118 0.0137 — —
A 2
— — –1.455 0.0502 –3.049 0.0021 –1.654 0.0016
B2
5.961 <0.0001 1.348 0.0715 7.339 <0.0001 — —
C2
— — –2.064 0.0190 — — –2.263 0.0001

Spread flow the formation of ettringite.41 Factor A has a positive effect

The effect of variation of Factors B and C on average on 1-day compressive strength according to results shown
spread (Øm) for a fixed value of Factor A is shown in Fig. 3. in Table 6.
According to ANOVA results, presented in Table 6, a linear
relation was obtained between the spread length and the Seven-day compressive strength
water to binder ratio, which is the most significative factor. Figure 5 shows the 3-D plot of the effect of Factors B
According to EFNARC,39 a spread flow that varies from 240 and C on the 7-day compressive strength for a fixed value
to 260 mm is considered adequate for a plain self-consol- of Factor A. According to ANOVA results presented in
idating concrete (SCC) mixture. The reduction of cement Table 6, Factor A had a positive effect on 7-day compres-
(d50 = 8 µm) led to a higher amount of glass powder (d50 = sive strength of UHPC. However, Factors B and C had
7 µm) due to the adjusting of the mixture through the A&Amod an adverse effect on this response. Moreover, similar to
curve. As the glass powder has a very low water absorption 1-day compressive strength, a nonlinear relationship was
rate,7,11 with more glass particles and less cement content obtained between the 7-day compressive strength and the
(factor A), there is a higher value of spread flow. In addi- three factors studied.
tion, the alkaline content of the mixture increases with the
partial replacement of cement by recycled glass due to the Twenty-eight-day compressive strength
high Na2O content of the latter (refer to Table 1). Therefore, The effect of the studied variables on 28-day compres-
the low cement content has the lower shear strength of the sive strength is shown in Fig. 6 as a 3-D plot. Factors A and
paste, providing the higher flow due to the higher alkalinity C had a positive effect on 28-day compressive strength of
of the liquid phase.40 UHPC. However, Factor B had an adverse effect on 28-day
As expected, the w/b (Factor B) and HRWR content compressive strength.
(Factor C) have a positive effect in the spread flow value. The effect of HRWR content on 28-day compressive strength
The w/b has the most notable influence in this field; however, is given in Fig. 6. It shows that, contrary to what happens with
its superior limit mentioned previously, and its implications early strength, by increasing the HRWR content, the 28-day
on compressive strength, forced a limitation of its value. compressive strength rate increases as well.
Once again, a nonlinear relationship was obtained
One-day compressive strength between the 28-day compressive strength and the three
The effect of Factors B and C on 1-day compressive factors studied.
strength for a fixed value of factor A is shown as the three-di-
mensional (3-D) plot in Fig. 4. According to ANOVA results, Model checking
presented in Table 6, a nonlinear relationship was obtained Generally, the model adequacy is investigated by exam-
between the 1-day compressive strength and the three factors. ination of residuals. For the model to be adequate, the pattern
As shown in Fig. 4, the early compressive strength increases of a residuals plot should be structureless.
with a decrease of w/b (Factor B) and HRWR content The normal probability plot of the residuals of each RSM
(Factor C). The negative effect of the increase of water in (Fig. 7) shows that the residuals lie reasonably close to a
the compressive strength is a well-known effect. Even the straight line, implying that errors are distributed normally,
negative effect of polycarboxylate on early strength develop- supporting the idea that the terms included in the model are
ment has been demonstrated by several researchers.41,42 The significant.
polycarboxylate-based ether HRWR slows down the hydra- A graph of the predicted response values versus the actual
tion of silicates (especially the alite phase), as well as affects response values is shown in Fig. 8. It helps to detect a value,

248 ACI Materials Journal/January 2020

Fig. 3—Response surface 3-D plots indicating interaction Fig. 5—Response surface 3-D plots indicating interaction
effects of w/b (B) and HRWR content (C) on spread flow. The effects of w/b (B) and HRWR content (C) on 7-day compres-
cement content was fixed at 625 kg/m3 (A = 0). sive strength. The cement content was fixed at 625 kg/m3
(A = 0).

Fig. 4—Response surface 3-D plots indicating interaction Fig. 6—Response surface 3-D plots indicating interac-
effects of w/b (B) and HRWR content (C) on 1-day compres- tion effects of w/b (B) and HRWR content (C) on 28-day
sive strength. The cement content was fixed at 625 kg/m3 compressive strength. The cement content was fixed at
(A = 0). 625 kg/m3 (A = 0).
or group of values, that are not easily predicted by the model. satisfy the requirements for each of the responses as much
The figure also shows any abnormality in the response. as possible, without compromising any of the requirements.2
Optimization can be achieved by either maximizing or
MULTI-OBJECTIVE OPTIMIZATION OF minimizing the value of the response or by trying to set the
ECO-FRIENDLY UHPC response on a specific target. To optimize under a multi-ob-
Objective and methods jective criteria, the global desirability function developed by
A multi-objective optimization method was performed Derringer and Suich,43 Eq. (5) was used
with the purpose of finding the optimum values of the
independent variables that could lead to the best response. 1

( ) = ∏dirn 
1 n ∑ ri
After establishing the regression model between the mixture D = d1r1 × d 2r2 × d3r3 ×…× d nrn ∑ ri
(5)
 i =1 
design variables and responses, all independent variables
were varied simultaneously and independently to opti- where n is the number of responses included in the optimiza-
mize the objective functions. This optimal solution aims to tion; and ri is the relative importance of each individual func-
tion (di). The importance (ri) varies from 1 (least important)

ACI Materials Journal/January 2020 249

Fig. 7—Normal plot of residuals for each response model: (a) Øm; (b) R1; (c) R7; and (d) R28.
to 5 (most important). Individual desirability functions (di) In cases of minimization, maximization and in range of
range between 0, for a completely undesired response, and 1, individual responses, the desirability can be defined by the
for a fully desired response. For a value of D close to 1, the formulas given in Eq. (6), (7), and (8), respectively
combination of the different criteria is globally optimal, so
the response values are close to the target values. 1 Yi ≤ L
If any of the responses or factors fall outside their desir-  wti

ability range, the overall function becomes zero. In a  U − Yi 

d =   L < Yi < U (6)
numerical optimization approach, the objective would be  U − L 
to apply either mixture design variables or responses. There 0 Yi ≥ U

are several ways of calculating the desirability of a func-
tion depending on the desired result. The range of desired 0 Yi ≤ L
responses can be one of the following: maximum, minimum,  wti
 Y − L 
target, or in range. d =  i  L < Yi < U (7)
The mixture design variables were included in the optimi-  U − L 
zation, with their design range being maximum or minimum. 1 Yi ≥ U

In this study, the desired responses were defined as maximum
and in range.

250 ACI Materials Journal/January 2020

Fig. 8—Predicted versus actual values for each response model: (a) Øm; (b) R1; (c) R7; and (d) R28.
Table 7—Optimization of individual responses for self-consolidating ultra-high-performance glass concrete
mixture with high 28-day compressive strength (Criteria I) and minimum cement content (Criteria II)
Criteria I Criteria II
Responses and variables Lower Upper Goal Importance Goal Importance
Øm, mm 240 260 In range 5 In range 5
R28, MPa 150 165 Maximum 5 Maximum 5
Cement, kg/m 3
600 650 — — Minimum 5

0 Yi ≤ L ability for each objective can change using different weights.

 The weight value varies between 0.1 and 10. The values
d = 1 L < Yi < U (8) higher than 1 give more emphasis to the objective, and those
0 Y ≥ U
 i lower than 1 have the opposite effect. With a weight of 1,
where L is a lower limit; U is the upper limit of a response; di varies in a linear manner. In this study, the optimization
and wti is the weight of a response. The shape of the desir- process was carried out in two consecutive steps, in which

ACI Materials Journal/January 2020 251

Table 8—Optimum mixtures for Criteria I and Criteria II
A B C
Mixture Cement, kg/m3 w/b HRWR, % vol. Desirability
Coded 1.789 –1.193 1.634
Criteria I 0.978
Real 669.72 0.159 0.025
Coded –0.894 0 –0.745
Criteria II 0.691
Real 602.639 0.165 0.020

both responses and variables were defined as a specific Table 9—Predicted responses by model versus
objective by assigning an importance degree. The objec- experimental measurement
tives of all the optimizations for each response are presented Øm, mm R28
in Table 7. Criteria I has been proposed for selecting the
Experi- Experi-
optimum mixture design variables to obtain a self-com-
Mixture mental Model RDL mental Model RDL
pacting mixture according to the EFNARC criteria39 with
the maximum 28-day compressive strength. In this scope, Criteria I 250 250.54 0.22% 165.8 164.43 –0.83%
the slump flow was defined as an “in range” objective, while Criteria
252 250.04 –0.78% 152.45 155.25 1.80%
28-day compressive strength was defined as a “maximum” II
objective. At the end of the multi-objective optimization
process, one optimal solution that satisfied the specified 1. An optimal mixture was designed to reach a 150 MPa
constraints was obtained. Criteria II adds to Criteria I the 28-day compressive strength with only 603 kg/m3 of cement
variable of minimum content of cement, maintaining a and 100 kg/m3 of silica fume, providing an ultra-high-perfor-
28-day compressive strength of over 150 MPa. Both opti- mance concrete which incorporates by-products in its dosage
mized mixtures are given in Table 8. such as recycled glass powder and recycled glass flour.
2. The regression model, using RSM, provides a thor-
Validation of proposed model ough examination of eco-friendly and cost-effective UHPC
The efficiency of the designed models was confirmed by properties over the selected range of cement content, w/b,
performing the experiment under the optimal conditions and and HRWR content. The lack-of-fit test results combined
comparing predicted-to-measured values obtained with the with high values of coefficients of multiple determinations
two mixtures presented in Table 8. The arrangement and (R2) demonstrated the accuracy of the second-order model
results of the confirmation tests are shown in Table 9. The to predict the required performance of UHPC in relation
absolute relative deviation, RDL (%), was used as a mean to compressive strength at 1, 7, and 28 days, as well as the
to measure accuracy for validation. Results showed that the spread flow value. The ANOVA results also verified that the
experimental values were similar to those predicted by the inclusion of all model parameters was statistically signifi-
proposed model. cant based on a very low P-value.
It is important to note that the mixture designed under 3. A numerical optimization of several responses was
Criterion II and tested experimentally resulted in a reduc- also efficiently accomplished to obtain an eco-friendly
tion of cement content of approximately 25% and of silica mixture with the maximum compressive strength and
fume by over 50%, compared to the typical dose of UHPC minimum cement content. The obtained results showed
proposed by Richard and Cheyrezy.8 These partial substitu- that the optimal values of design variables were 603 kg/m3
tions for less-expensive materials such as recycled glass and of cement, reaching compressive strength of 152 MPa at
limestone powders leads to a reduction in the final cost of the 28 days. Allowing a higher amount of cement, the numer-
mixture of approximately 20%. ical optimization showed an experimentally validated
compressive strength result of 165.8 MPa with 667 kg/m3
CONCLUSIONS of cement.
In this study, a multi-objective simultaneous optimi- 4. The results predicted by the model, along with those
zation technique was adopted to obtain a cost-effective obtained with the experiments, drive to the conclusion that
and eco-friendly ultra-high-performance concrete using the statistical model can be used to predict the properties of
by-products as filler, such as recycled glass flour, and as new mixtures with good accuracy.
cementitious supplementary materials, such as recycled
glass powder. Also, micro limestone powder was used as AUTHOR BIOS
Joaquín Abellán is a PhD Candidate in the Department of Civil Engi-
partial replacement of silica fume. The new material exhib- neering at the Polytechnic University of Madrid (UPM), Madrid, Spain.
ited excellent workability and rheological properties. In this His research interests include mathematical optimization of eco-friendly
model, a response surface method (RSM) was incorporated. ultra-high-performance concrete and seismic behavior of high strain-hard-
ening cementitious composites.
Five performance parameters—spread flow; compressive
strength at 1, 7, and 28 days; and cement content—were Jaime Fernández is a Professor in the Department of Civil Engineering at
analyzed and used in the optimization of the mixture design. UPM, where he received his BS and PhD in civil engineering. His research
interests include structural analysis, construction engineering, and civil
Based on the results of this experimental investigation, the engineering materials.
following conclusions are drawn:

252 ACI Materials Journal/January 2020

Nancy Torres is a Professor of civil engineering at Escuela Colombiana de tional Symposium on Ultra High Performance Concrete, Kassel University,
Ingeniería Julio Garavito, Bogotá, Colombia. She received her BS in civil Kassel, Germany, 2004.
engineering from Universidad Francisco de Paula, Santander, Colombia, 16. Russell, E.; Lee, J.; and Clift, R., “Can the SDGs Provide a Basis for
and her MS in structural engineering and her PhD in engineering from the Supply Chain Decisions in the Construction Sector?” Sustainability, V. 10,
Universidad Nacional de Colombia, Bogotá, Colombia. No. 3, 2018, 19 pp. doi: 10.3390/su1003062910.3390/su10030629
17. Ferdosian, I.; Camões, A.; and Ribeiro, M., “High-Volume Fly
Andrés Núñez is a Civil Engineer at Cementos Argos SA, Medellín, Ash Paste for Developing Ultra-High Performance Concrete (UHPC),”
Colombia. He received his BS in civil engineering from the Universidad Ciência e Tecnologia dos Materiais, V. 29, No. 1, 2017, pp. e157-e161. doi:
del Quindío, Armenia, Colombia, and his MS and PhD in civil engineering 10.1016/j.ctmat.2016.10.001
from the Polytechnic University of Valencia, Valencia, Spain. 18. Ghafari, E.; Costa, H.; and Júlio, E., “Statistical Mixture Design
Approach for Eco- Efficient UHPC,” Cement and Concrete Composites,
V. 55, 2015, pp. 17-25. doi: 10.1016/j.cemconcomp.2014.07.016
ACKNOWLEDGMENTS 19. Meng, W.; Samaranayake, V. A.; and Khayat, K. H., “Factorial Design
Special thanks go to Cementos Argos S.A. for the donation of most and Optimization of Ultra-High-Performance Concrete with Lightweight
of the materials used in the research described herein. The supply of Sand,” ACI Materials Journal, V. 115, No. 1, Jan. 2018, pp. 129-138. doi:
recycled glass from Cristalería Peldar S.A. and electric arc slag furnace 10.14359/51700995
form GERDAU S.A. for this research is highly appreciated. The writers 20. Li, W.; Huang, Z.; Zu, T.; Shi, C.; Duan, W. H.; and Shah, S. P.,
would also like to acknowledge the support and suggestions from “Influence of Nanolimestone on the Hydration, Mechanical Strength, and
Escuela Colombiana de Ingeniería Julio Garavito and Polytechnic Autogenous Shrinkage of Ultrahigh-Performance Concrete,” Journal of
University of Madrid. Materials in Civil Engineering, ASCE, V. 28, No. 1, 2016, pp. 1-9. doi:
10.1061/(ASCE)MT.1943-5533.0001327
REFERENCES 21. Huang, Z., and Cao, F., “Effects of Nano-Materials on the Perfor-
1. Soliman, N. A., and Tagnit-Hamou, A., “Partial Substitution of Silica mance of UHPC,” Materials Review, V. 26, No. 9, 2012, pp. 136-141.
Fume with Fine Glass Powder in UHPC: Filling the Micro Gap,” Construc- 22. Yu, R.; Tang, P.; Spiesz, P.; and Brouwers, H. J. H., “A Study of
tion and Building Materials, V. 139, 2017, pp. 374-383. doi: 10.1016/j. Multiple Effects of Nano-Silica and Hybrid Fibres on the Properties of
conbuildmat.2017.02.084 Ultra-High Performance Fibre Reinforced Concrete (UHPFRC) Incorpo-
2. Ghafari, E.; Costa, H.; and Júlio, E., “RSM-Based Model to Predict rating Waste Bottom Ash (WBA),” Construction and Building Materials,
the Performance of Self-Compacting UHPC Reinforced with Hybrid V. 60, 2014, pp. 98-110. doi: 10.1016/j.conbuildmat.2014.02.059
Steel Micro-Fibers,” Construction and Building Materials, V. 66, 2014, 23. Vaitkevičius, V.; Šerelis, E.; and Hilbig, H., “The Effect of Glass
pp. 375-383. doi: 10.1016/j.conbuildmat.2014.05.064 Powder on the Microstructure of Ultra High Performance Concrete,”
3. Schmidt, C., and Schmidt, M., “Whitetopping of Asphalt and Concrete Construction and Building Materials, V. 68, 2014, pp. 102-109. doi:
Pavements with Thin Layers of Ultra-High-Performance Concrete - 10.1016/j.conbuildmat.2014.05.101
Construction and Economic Efficiency,” Proceedings of Hipermat 2012 - 24. Funk, J. E., and Dinger, D., Predictive Process Control of Crowded
3rd International Symposium on UHPC and Nanotechnology for Construc- Particulate Suspensions: Applied to Ceramic Manufacturing, Springer
tion Materials, Kassel, Germany, 2012, pp. 921-928. Science+Business Media, New York, 1994, 786 pp.
4. Abbas, S.; Nehdi, M.; and Saleem, M. A., “Ultra-High Performance 25. Eriksson, L.; Johansson, E.; Kettaneh-Wold, N.; Wikström, C.; and
Concrete: Mechanical Performance, Durability, Sustainability and Imple- Wold, S., Design of Experiments: Principles and Applications, Umetrics
mentation Challenges,” International Journal of Concrete Structures and AB, Umeå, Sweden, 2000, 425 pp.
Materials, V. 10, No. 3, 2016, pp. 271-295. doi: 10.1007/s40069-016-0157-4 26. Upasani, R. S., and Banga, A. K., “Response Surface Methodology
5. Soliman, N. A., and Tagnit-Hamou, A., “Using Particle Packing to Investigate the Iontophoretic Delivery of Tacrine Hydrochloride,” Phar-
and Statistical Approach to Optimize Eco-Efficient Ultra-High-Perfor- maceutical Research, V. 21, No. 12, 2004, pp. 2293-2299. doi: 10.1007/
mance Concrete,” ACI Materials Journal, V. 114, No. 6, Nov.-Dec. 2017, s11095-004-7682-6
pp. 847-858. doi: 10.14359/51701001 27. Lenth, R. V., “Response-Surface Methods in R, Using RSM,”
6. Jammes, F.-X.; Cespedes, X.; and Resplendino, J., “Design of Offshore Journal of Statistical Software, V. 32, No. 7, 2012, pp. 1-17. doi: 10.18637/
Wind Turbines,” RILEM-Fib-AFGC Int. Symposium on Ultra-High Perfor- jss.v032.i07
mance Fibre-Reinforced Concrete UHPFRC 2013, 2013, pp. 443-452. 28. Montgomery, D. C., Design and Analysis of Experiments, John Wiley
7. Tagnit-Hamou, A.; Soliman, N. A.; and Omran, A., “Green & Sons, Inc, New York, 2005.
Ultra-High-Performance Glass Concrete,” First International Interactive 29. Mosaberpanah, M. A., and Eren, O., “Statistical Models for
Symposium on UHPC, Des Moines, IA, July 18-20, 2016, 10 pp. Mechanical Properties of UHPC Using Response Surface Methodology,”
8. Richard, P., and Cheyrezy, M., “Composition of Reactive Powder Computers and Concrete, V. 19, No. 6, 2017, pp. 667-675. doi: 10.12989/
Concretes,” Cement and Concrete Research, V. 25, No. 7, 1995, pp. 1501- cac.2017.19.6.667
1511. doi: 10.1016/0008-8846(95)00144-2 30. Mosaberpanah, M. A., and Eren, O., “Effect of Quartz Powder, Quartz
9. De Larrard, F., and Sedran, T., “Mixture-Proportioning of High-Per- Sand and Water Curing Regimes on Mechanical Properties of UHPC Using
formance Concrete,” Cement and Concrete Research, V. 32, No. 11, Response Surface Modeling,” Advances in Concrete Construction, V. 5,
pp. 2002, pp. 1699-1704. doi: 10.1016/S0008-8846(02)00861-X10.1016/ No. 5, 2017, pp. 481-492. doi: 10.12989/acc.2017.5.5.481
S0008-8846(02)00861-X 31. Sayed Ahmed, M. S.; Sennah, K.; and Monsif, M. Y., “Mechanical
10. Kou, S. C., and Xing, F., “The Effect of Recycled Glass Powder and Behaviour of Ultra-High Performance Concrete Obtained with Different
Reject Fly Ash on the Mechanical Properties of Fibre-Reinforced Ultrahigh Concrete Constituents and Mix Designs,” Resilient Infrastructure, London,
Performance Concrete,” Advances in Matererials Science and Engineering, UK, 2016, 10 pp.
V. 2012, 2012, 8 pp. doi: 10.1155/2012/26324310.1155/2012/263243 32. Van, V. T. A., and Ludwig, H.-M., “Proportioning Optimization of
11. Soliman, N. A., and Tagnit-Hamou, A., “Using Glass Sand as an UHPC Containing Rice Husk Ash and Ground Granulated Blast-furnace
Alternative for Quartz Sand in UHPC,” Construction and Building Mate- Slag,” Proceedings of Hipermat 2012 - 3rd International Symposium on
rials, V. 145, 2017, pp. 243-252. doi: 10.1016/j.conbuildmat.2017.03.187 UHPC and Nanotechnology for Construction Materials. Kassel University,
12. Kalny, M.; Kvasnicka, V.; and Komanec, J., “First Practical Applica- Kassel, Germany, 2012, pp. 197-205.
tions of UHPC in the Czech Republic,” Proceedings of Hipermat 2016 - 4th 33. Abellan, J.; Torres, N.; Núñez, A.; and Fernández, J., “Influencia del
International Symposium on UHPC and Nanotechnology for Construc- Exponente de Fuller, la Relación Agua Conglomerante y el Contenido en
tion Materials, E. Fehling, B. Middendorf, and J. Thiemicke, eds., Kassel, Policarboxilato en Concretos de Muy Altas Prestaciones,” IV Congreso
Germany, 2016, pp. 147-148. Internacional de Ingenieria Civil. Havana, Cuba, 2018.
13. Sayed Ahmad, M., and Sennah, K., “Development of Ultra-High 34. Ghafari, E.; Costa, H.; and Júlio, E., “Critical Review on Eco-Effi-
Performance Concrete Jointed Precast Decks and Concrete Piles in Inte- cient Ultra High Performance Concrete Enhanced with Nano- Materials,”
gral Abutment Bridges,” The First International Symposium on Jointless & Construction and Building Materials, V. 101, 2015, pp. 201-208. doi:
Sustainable Bridges, Fuzhou, Fujian, China, 2016. 10.1016/j.conbuildmat.2015.10.066
14. Haber, Z. B.; Munoz, J. F.; and Graybeal, B. A., “Field Testing of an 35. Kumar, R., and Tiwari, O. P., “Experimental Investigation of
Ultra-High Performance Concrete Overlay,” FHWA-HRT-17-096, Federal Mechanical Characterization and Drilling of Fabricated GFRP Compos-
Highway Administration, McLean, VA, 2017, 57 pp. ites Reinforced with Al2O3 Micro Particles,” International Journal of
15. Acker, P., and Behloul, M., “Ductal Technology: A Large Spectrum Advance Research, Ideas and Innovations Technology, V. 4, No. 4, 2018,
of Properties, a Wide Range of Applications,” Proceedings of the Interna- pp. 191-199.

ACI Materials Journal/January 2020 253

 36. ASTM C1437-13, “Standard Test Method for Flow of Hydraulic tion and Building Materials, V. 51, 2014, pp. 484-491. doi: 10.1016/j.
Cement Mortar,” ASTM International, West Conshohocken, PA, 2013, conbuildmat.2013.10.004
2 pp. 41. Puertas, F.; Santos, H.; Palacios, M.; and Martínez-Ramírez, S., “Poly-
37. ASTM C109/C109M-11b, “Standard Test Method for Compressive carboxylate Superplasticiser Admixtures: Effect on Hydration, Microstruc-
Strength of Hydraulic Cement Mortars (Using 2-in. or [ 50-mm ] Cube ture and Rheological Behaviour in Cement Pastes,” Advances in Cement
Specimens),” ASTM International, West Conshohocken, PA, 2011, 10 pp. Research, V. 17, No. 2, 2005, pp. 77-89. doi: 10.1680/adcr.2005.17.2.77
38. R Core Team, “R: A Language and Environment for Statistical 42. Kubens, S., Interaction of Cement and Admixtures and Its Influence
Computing,” Vienna, Austria, 2018, www.r-project.org. (last accessed Dec. on Rheological Properties, Cuvillier Verlag, Göttingen, Germany, 2010.
13, 2019) 43. Derringer, G., and Suich, R., “Simultaneous Optimization of Several
39. The European Project Group, “The European Guidelines for Response Variables,” Journal of Quality Technology, V. 12, No. 4, 1980,
Self-Compacting Concrete,” 2005, 63 pp. pp. 214-219. doi: 10.1080/00224065.1980.11980968
40. Pedrajas, C.; Rahhal, V.; and Talero, R., “Determination of Char-
acteristic Rheological Parameters in Portland Cement Pastes,” Construc-

254 ACI Materials Journal/January 2020