ACI MATERIALS JOURNAL TECHNICAL PAPER

Title No. 118-M51

Artificial Neural Network Model for Strength Prediction of

Ultra-High-Performance Concrete
by Joaquín Abellán-García

Ultra-high-performance concrete (UHPC) is the outcome of the specimens without nano-CaCO3. The results of the partial
mixture of several constituents, leading to a highly complex mate- substitution of silica fume with fine glass powder (FGP),
rial, which makes it more difficult to understand the effect of each with a mean particle size (d50) of 3.8 µm, in UHPC were
component and their interactions on compressive strength. This analyzed.17 They showed that compressive strength of 235
research goal is developing an artificial neural network (ANN)
and 220 MPa could be achieved when using steam curing
approach to predict the compressive strength of UHPC, being able
and replacing 30% and 50% of SF with FGP, respectively.
to incorporate supplementary cementitious materials (SCMs), and
even different situations in relation to the aggregate: from pastes In another study, recycled glass powder was used to replace
to incorporation of coarse aggregate. The one-hidden-layer ANN quartz sand, cement, and quartz powder particles.18,19 RHA
model was trained with 927 data by using the R-code language. was also analyzed as an SCM in UHPC with successful
The data was produced by collecting data from 210 experiments results in strength and durability due to the internal curing
combined with 717 dosages from previous research. The Olden process of the concrete.20 These investigations demonstrated
algorithm was used to analyze the relationships between the the possibility of partial replacement of cement and silica
UHPC’s components and strength. The results indicated that the fume in UHPC by using other SCMs, while maintaining its
ANN is an efficient model for predicting the compressive strength of outstanding mechanical and durability features.
UHPC, regardless of the SCM used or maximum size of aggregate Nevertheless, the key properties of UHPC with several
considered.
SCMs need to be checked experimentally because of the
Keywords: activation function; artificial neural network (ANN); compres- unclear combination effects of different materials and
sive strength; resilient backpropagation algorithm (Rprop); supplementary dosages used in the mixture. Laboratory tests are often
cementitious material (SCM); ultra-high-performance concrete (UHPC). costly, time-consuming, and labor-intensive.3 To reduce the
number of experimental works and mixture designs, proba-
INTRODUCTION bilistic models can be successfully employed to forecast the
Ultra-high-performance concrete (UHPC) is a new compressive strength of concrete. However, these approaches
high-tech type of concrete characterized by its ultra-high cannot be applied when the modeling problem involves too
compressive strength (over 150 MPa), low permeability, many independent variables and the interactions among
and improved durability.1-7 However, due to the absence of them are either unknown or too complex to perform.3,4,21 In
a coarse aggregate, together with its high packing density, the case of the UHPC, due to its large number of compo-
UHPC contents of cement, quartz sand, quartz powder, and nents, the use of traditional techniques of regression fails to
silica fume are relatively higher in relation to current concrete, achieve the expected precision and predictability.21
increasing the cost of UHPC and resulting in a significant Developments in artificial neural networks (ANNs) have
environmental impact.8-10 Hence, considering those disad- resulted in an increasingly wider use of those methods for
vantages that restrict its wider use, some industrial by-prod- modeling real-world problems because of its immense
ucts such as fly ash (FA), ground-granulated blast-furnace ability to capture inter-relationships among input and output
slag (GBFS), recycled glass powder (GP), rice husk ash data pairs that are unknown, nonlinear, or too difficult to
(RHA), and fluid catalytic cracking residue (FCC), as well formulate, demonstrating its effectiveness in solving diffi-
as other supplementary cementitious materials (SCMs) and cult and complex engineering problems.22 Among others,
mineral fillers such limestone powder (LP) and metakaolin ANN models have been employed in applications such as
(MK), have been used as replacements for cement (partial), water resources engineering, traffic engineering, detection
quartz powder (total), and silica fume (partial/total).11,12 of structural damage, structural system identification, mate-
Among others, studies of UHPC containing FA, SF, and rial behavior modeling, concrete mixture proportioning, and
FCC—a by-product of the crude oil industry—were carried concrete strength prediction.23,24 Furthermore, one of the
out.13,14 Their conclusion showed the possibility of a partial most fertile fields in the scientific literature of civil engi-
replacement of cement and silica fume. Blending limestone neering production is the application of ANN to predict
powder in a UHPC binder improved the hydration process
at the early age, producing denser particle packing and ACI Materials Journal, V. 118, No. 4, July 2021.
MS No. M-2019-494.R2, doi: 10.14359/51732710, received December 15, 2020,
improving mechanical properties.15 The use of nano-CaCO3 and reviewed under Institute publication policies. Copyright ©  2021, American
Concrete Institute. All rights reserved, including the making of copies unless
as a component of binder was also studied.16 Compressive permission is obtained from the copyright proprietors. Pertinent discussion including
strength increased 17% compared to the UHPC control author’s closure, if any, will be published ten months from this journal’s date if the
discussion is received within four months of the paper’s print publication.

ACI Materials Journal/July 2021 3

Table 1—Transforming factors to turn compressive strength results to 100 mm cube specimen
Type of specimen Cube Cube Cube Cube Cube Cylinder Cylinder Cylinder Cylinder
Dimensions, mm 150 100 70 50 40 50 x 100 75 x 150 100 x 200 150 x 300
Factor to 100 mm cube 1.119 1.000 0.935 0.962 1.121 1.020 1.020 1.020 1.063

mechanical properties of pastes, mortars, and concretes.25

However, little research has been conducted on predicting
the performance of UHPC using ANN, among which inves-
tigations by Ghafari et al.,4 Taghaddos et al.,26 and Zhang
et al.3 can be highlighted.
The main objective of this research work is to develop an
ANN model for predicting the 28-day compressive strength
UHPC. The model is designed in such a way that it can
consider any combination of a broad spectrum of SCMs,
including silica fume, fly ash, ground-granulated blast-
furnace slag, recycled glass powder, rice husk ash, fluid
catalytic cracking residue, and metakaolin, besides mineral
powder such as limestone powder and quartz powder (QP).
The developed model was also able to consider form paste Fig. 1—Bivariate boxplot for pair of variables compressive
(non-aggregate concrete) to a maximum size of aggregate strength and water content.
of 19 mm. Although the experimental campaign only dealt with
ASTM Type III high-early-strength (HE) cement and sili-
RESEARCH SIGNIFICANCE ceous sand, in addition to not including quartz powder in
Notwithstanding the excellent performance in strength the dosages, the definitive database, enriched with other
and durability, UHPC has not been widely used in construc- research, contemplates various types of cement, different
tion primarily due to its high costs and carbon footprint. possibilities of aggregate, from pastes to concrete with
In recent years, great efforts have been made in achieving coarse aggregate, and even the use of quartz powder.
a less expensive and more sustainable UHPC, using SCMs
as partial replacement for cement. However, given the Treating outliers
relatively high possibility of SCMs and combinations, the Dealing with outliers is required before performing
number of trials in material development research are often regression analyses, as they can greatly affect the resulting
high, leading to considerable experimental costs. This study model.34 Therefore, a descriptive statistical analysis was
provides an accurate model for predicting the compressive performed on each variable to detect outliers.35 Data entry
strength of UHPC, even when dealing with SCMs, which errors, outliers, and unusual distributions (skewness) in the
could help reduce the number of tests required. data were identified using bivariate boxplots and Cook’s
distance. Bivariate boxplots are two-dimensional analogs
DATABASE of single-variable boxplots that are used to spot inconsistent
Data collection data and outliers by robust methods, drawing ellipses that lay
Data sets from the 28-day compressive strength of UHPC out possible disturbing observations.36 However, relying on
with different combinations of the aforementioned SCMs the use of this technique without a critical analysis of the data
were carried out in various laboratories, totaling 210  tests, could also be a dangerous practice. Some of the suspicious
which were supplemented with 717 tests extracted points identified by bivariate boxplots could be telling the
from several international symposiums on high- and real behavior of the data, while the rest of the observations
ultra-high-performance concrete (including Kassel 2004, could just be clumped together very closely. Figure 1 shows
2008, 2012, and 201627-30), PhD theses, and other papers the bivariate boxplot for the pair of variables’ compressive
published in international journals. It is important to high- strength (expressed in MPa) and water content (expressed in
light that only those experiments with information on the volume ratio). This plot clearly illustrates an example of the
virtual packing density and/or the mean particle size of all situation where important data would be lost if all the data
components were considered for the database. Besides, in all marked as outliers were to be eliminated. In Fig. 1, 25% of
cases, compressive strength referred to no heat curing and no the dosages containing MK (blue points of the graph) could
vacuum mixed procedure. be considered as outliers (red points) when falling out of
Due to the fact that different specimens’ shapes and the ellipse. It is important to highlight that several studies
dimensions were observed in published research and exper- have demonstrated that the inclusion of MK in concrete
imental campaigns, the values of compressive strength were requires more water or high-range water-reducing admixture
transformed to that of a cube with 100 mm sides by means (HRWRA) to reach a determined workability due to its reac-
of transformation factors.14,31-33 The multiplication trans- tive alumina-silicate content and its amorphous shape.37-40
forming factors used are listed in Table 1. Besides, the polycarboxylate content in UHPC is restricted
because its excess makes the mixtures sticky,41 preventing

4 ACI Materials Journal/July 2021

the exit of trapped air, which has a significant negative effect This is an iterative way of analyzing the influence of that
on the strength of the concrete. Therefore, eliminating the observation.42
dosages with metakaolin by having a higher water content At the end of this process, 130 observations were removed
outside the ellipse would not be appropriate. Other similar from the database, leaving 797 for training and validation of
situations, such as those corresponding to the pairs of RHA the ANN models.
content and water-binder ratio (w/b), and FCC content and Figures 2, 3, and 4 depict some of the descriptive statistics
water-total powders ratio, among others, have also been taken of the database after removing the outliers.
into account.
On the other hand, Cook’s distance is used in regression Data normalization
analysis to find influential outliers in a set of predictor vari- Once the outliers were eliminated from the database, the
ables.42 It is estimated by removing the i-th observation next step consisted of data normalization. The input data
from the model and recalculating the regression. All the and output data generally comprise different identities,
values in the regression model are then observed whether either having no or minimum similarities. Preprocessing or
changes have been pointed out after the removal of the point. normalization of data eliminates the possibility of neural
network bias towards the different identities and scales down
all the input and output data. When sigmoid function was
used as an activation function, linear scaling in the range
[0, 1] was used, having a function as depicted in Eq. (1). On
the other hand, when hyperbolic tangent was used as acti-
vation function, data was scaled into the range [–1, 1] as
depicted in Eq. (2)

x − xmin
xnorm − sigmoid = (1)
xmax − xmin

x − xmin
xnorm − tanh = 2 −1 (2)
xmax − xmin
where xnorm is the normalized value of the variable x; and xmax
and xmin are the minimum and maximum values of variable
x, respectively.
Table 2 shows the range of variation of all variables
Fig. 2—Histogram of 28-day compressive strength reached considered in this research.
by dosages in database.

Fig. 3—Pie chart from data: (a) number of components of binder; (b) ASTM cement types reported; (c) maximum size of aggre-
gate; and (d) type of coarse aggregate, when used.

ACI Materials Journal/July 2021 5

Fig. 4—Boxplot for component of UHPC in database and some relationships such as w/b, w/p, and VPD.
Table 2—Range of variation of all variables considered in this research
C* SF* FA* GBFS* RGP* RHA* FC3R* MK* LP* W* HRWR* QP* MSA† WB WP A* VPD CS‡
xmax 0.503 0.155 0.250 0.247 0.202 0.177 0.091 0.161 0.190 0.361 0.063 0.235 19,000 0.322 0.322 0.676 0.877 213
xmin 0.102 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.103 0.026 0.000 0.000 0.124 0.107 0.000 0.581 100
Average 0.245 0.064 0.016 0.009 0.028 0.004 0.005 0.003 0.024 0.201 0.026 0.026 1724 0.187 0.177 0.364 0.771 153
SD 0.072 0.035 0.042 0.024 0.053 0.017 0.015 0.014 0.043 0.038 0.012 0.051 2498 0.037 0.034 0.130 0.044 24
*
Expressed in m3/m3.
†
Expressed in microns.
‡
Compressive strength expressed in MPa.

EXPERIMENTAL INVESTIGATION ball mill for 90 minutes. After treatment, RHA had a d50 of
Cube specimens with 50 mm sides (1.968 in.) of 210 7.5 μm (0.0030 in.) and a specific gravity of 2.16. FCC, a
separate UHPC dosages were used to measure the 28-day by-product of the crude industry, with a specific gravity of
compressive strength following ASTM C109.43 2.76 and d50 of 9.5 μm (0.0037 in.), was also used in some
experimental works. GP and GF were obtained by grounding
Materials locally available recycled glass with a jet mill to different
The materials used to elaborate the UHPC were locally degrees of fineness, by applying different grinding speeds.
available in Colombia. ASTM Type III HE cement was used. The GF with d50 of 28 μm (0.0011 in.) and GP with d50 of
The HE cement had a specific gravity of 3.15 and a mean 7  μm (0.0003 in.), both with a specific density of 2.55, were
particle diameter (d50) of 8 μm (0.0003 in.). The SF used taken into account. Commercially available MK employed
in the mixture proportioning complied with ASTM C1240 in some of the dosages had a specific gravity of 2.66 and d50
specifications and had a specific gravity of 2.20 and d50 of of 10 μm (0.0004 in.). Finally, two sizes of commercialized
0.15 μm (0.0006 in.). The UHPC was also manufactured limestone powder with a specific gravity of 2.73 were used.
with a silica sand of a specific gravity of 2.65, maximum LP with d50 of 15 μm (0.0006 in.) and MLP with d50 of 2  μm
particle size (dmax) of 600 μm (0.0240 in.), and d50 of (0.0001 in.) were considered. Table 3 shows the chemical
165  μm (0.0065 in.). Besides cement, silica fume, and silica composition of the materials used in this study. Figure 5
sand, the following SCMs were used as components of the provides the particle size distribution (PSD) of the cement,
binder: locally available FA, GBFS, GP, recycled glass flour SF, FA, RHA, GBFS, GP, GF, FCC, MK, LP, MLP, and SS.
(GF), RHA, FCC, MK, LP, and micro limestone powder A polycarboxylate (PCE)-based HRWRA with a specific
(MLP). Locally available FA with a specific gravity of 2.32 gravity of 1.07 and solid content of 40% was used.
and d50 of 30 μm (0.0012 in.) was considered as part of The supplementary cementitious materials used were
the binder for some mixtures. The GBFS considered in the analyzed by scanning electron microscopy (SEM). Some
experimental campaign had a specific gravity of 2.95 and of these analyses are depicted in Fig. 6. Results signaled
d50 of 2.4 μm (0.0009 in.). RHA used in this research was the lack of porosity of the two products related to recycled
produced by combustion with a peak temperature of approx- glass, the high porosity of the RHA particle, and the spher-
imately 700°C. The ash obtained was ground in a vibrating ical shape of the FA and SF particles, besides the small size
of the latter.

6 ACI Materials Journal/July 2021

Table 3—Physical and chemical compositions of cement and supplementary cementitious materials
C SF FA GBFS GP GF RHA FC3R MK LP MLP SS
Specific gravity 3.16 2.20 2.32 2.95 2.55 2.55 2.16 2.76 2.66 2.73 2.73 2.65
Loss on ignition, % 2.58 0.60 11.54 0.00 0.80 1.00 4.82 10.61 2.21 42.41 42.21 0.20
SiO2, % 19.42 92.29 50.09 36.93 75.47 72.89 88.59 39.61 52.17 0.90 0.90 99.80
Al2O3, % 4.00 0.59 22.26 16.45 1.09 1.67 0.31 42.47 39.11 0.10 0.10 0.14
CaO, % 64.42 3.89 2.19 33.78 9.02 9.73 0.74 2.85 0.78 55.51 55.51 0.17
MgO, % 1.52 0.26 0.53 3.91 1.97 2.08 0.66 0.07 0.19 0.70 0.70 0.01
SO3, % 1.93 0.07 0.03 2.51 0.00 0.01 0.26 0.62 0.22 0.10 0.10 —
Na2O, % 0.19 0.31 0.31 0.39 11.65 12.54 0.26 0.61 0.24 0.03 0.03 —
K2O, % 0.39 0.54 0.99 0.40 0.75 0.76 2.46 0.06 0.20 0.01 0.00 0.05
TiO2, % 0.38 0.01 1.05 0.57 0.04 0.04 0.02 0.67 1.48 0.00 0.00 —
MnO, % 0.05 0.01 0.01 3.41 0.01 0.01 0.21 0.00 0.00 0.00 0.01 —
Fe2O3, % 3.61 0.24 9.33 1.24 0.79 0.81 0.29 0.69 1.07 0.05 0.05 0.04

 Note: — indicates items not measured.

Fig. 5—Particle size distribution of cement, sand, and supplementary cementitious materials.
Specimens Items of investigation
The experimentally run dosages correspond to 15 groups Once the dosage was designed, a 5 L mortar mixer
of component combinations. For each of these combinations, was used to produce the concrete. After the mixing was
a three-factor design of experiment (DoE) with 14 runs each completed, the UHPC was cast in molds and compacted on
was proposed. The first factor corresponded to the amount of a vibrating table. The prisms were demolded approximately
cement, the second to the amount of water, and the third to 24 hours after casting and then cured in a moisture room at
the amount of HRWRA. The other components were ruled 20°C until the day of the test, without any heat treatment
by the A&Amod curve44 according to Eq. (3) applied. For the determination of the compressive strength,
50 mm (1.968 in.) cubes were tested. A compression testing

P ( D) =
(D q q
− � Dmin ) � (3)
machine with a capacity of 3000 kN was used, following
ASTM C109.43 Three samples were tested for determining
(D q
max −D q
min ) the compressive strength at 28 days.
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 ANALYTICAL INVESTIGATION
maximum and minimum particle sizes, respectively; and q is Artificial neural networks
the Fuller exponent. The optimal value of q was determined ANN models are composed of many highly intercon-
in previous research.2 nected processing elements (neurons) working together.23

ACI Materials Journal/July 2021 7

Fig. 6—SEM of some supplementary cementitious materials used in research: (a) SF; (b) MLP; (c) FA; (d) RHA; (e) GBFS;
(f)  FCC; (g) GP; and (h) GF.

Fig. 7—Schematic diagram of artificial neurons structure.

Each node is fully connected to the other through connection n

weights and receives an input signal from neurons linked v j = ∑wij xi + b (4)
i =1
to it. The successive layers of nodes receive input from the
previous layers; the outputs of nodes in each layer are inputs where vj is the net input of the j-th neuron for the signals
to nodes in the next layer. The simplest form of ANN archi- received from the preceding layer with n neurons; wij is the
tecture is the perceptron, devised by Rosenblatt45 in 1958, weight between the j-th neuron and the i-th neuron in the
which consists of one neuron with two inputs and one output. preceding layer; xi is the output of the i-th neuron in the
The schematic diagram of the artificial neuron model is preceding layer; and b is a constant also known as bias.47
composed of some interconnected sets of artificial neurons, The activation function processes the input obtained from
which are closely linked to each other the by connection the sum function and determines the neuron output. Some
weights and obtained input signs from neurons that linked to activation functions examples include step, linear, ramp,
it, as illustrated in Fig. 7, where the vector xi = (x1, x2,...,xn) tanh (hyperbolic tangent), sigmoid functions, and relu (recti-
represents an input signal applied to the neuron. Besides, a fied linear unit function).48 In this work, the performance of
bias is also added to the neuron along with inputs. On the hyperbolic tangent and sigmoid functions as activation func-
other hand, the vector wi = (w1, w2,...,wn) represents the tions were evaluated. Hyperbolic tangent function maps the
weights for each input. The weighted sum function calcu- inputs between –1 and +1,21 while the sigmoid function takes
lates the net input (vj) that approaches to a neuron,46 as the input, which can have any value between minus and plus
depicted in Eq. (4) infinite, and squashes the output in the range 0 to 1.49 The latter
has an implication that the way in which the normalization

8 ACI Materials Journal/July 2021

Fig. 8—K-fold validation.48 (Note: T is training set; V is validation set.)
of the data for the training of the network should be carried models, and training each one on k – 1 partitions while eval-
out differs depending on the activation function chosen. uating the remaining partition. The validation score for the
On the other hand, adjusting the ANN weights to produce model used is then the average of the k validation scores
a particular output is called “training,” which is the mecha- obtained.48 In this research, k = 10 was considered.
nism that allows the network to learn.50 Rumelhart et al.51
developed the most commonly used learning algorithm ANN architecture design
called backpropagation (BP) for ANN. BP training algo- In this research, a one-hidden layer ANN model was
rithm is an iterative gradient designed to minimize the mean developed using R version 3.5.2 (2018-12-20)54 by using the
square error between the actual output of multilayer feed- neuralnet55 function. For each activation function, 16 archi-
forward neural network and the desired output.52 In recent tectures were tested. Besides, for each of those architectures,
years, other learning algorithms have been developed, such 100 different models were computed by varying the initial
as the resilient backpropagation (Rprop). Rprop has two weight allocation of the neurons. Criteria for selecting the
main advantages over the BP algorithm. First, training with initial weight allocation and the number of neurons in the
Rprop is often faster than training with BP. Second, Rprop hidden layer was based on minimizing the root mean square
eliminates the harmful influence of the size of the partial error (RMSE) between real and predicted values during the
error derivative on the weight step, as opposed to the BP k-fold validation48 of training. Therefore, a total of 3200
algorithm, which needs values for the learning rate (and artificial networks were constructed to predict compressive
usually an optional momentum term). Consequently, only strength of UHPC using the 797 aforementioned observa-
the sign of the derivative is considered to indicate the direc- tions. The networks were composed of input neurons (input
tion of weight update.52,53 For further information about layer), computational neurons (hidden layer), and an output
Rprop, refer to Mushgil et al.52 and/or Prasad et al.53 layer. The selection of the input variables was performed
In this study, the ANN models were trained using resilient by means of an analysis of variance. Therefore, among
backpropagation as learning algorithm. 179 properties, 17 input signals grouped into two different
classes were used:
Preparing training and test data set 1. Components of UHPC expressed in volume ratio. This
To facilitate training and testing of the ANN models, the group included cement (C); silica fume (SF); FA; GBFS;
collected data was randomized and split into training and GP; RHA; FCC; MK; LP; water (W); HRWRA; QP; and
test data sets. Five hundred and ninety-seven observations total aggregate (A), which considered a range from micro
were used for training purposes, and the remaining 200 were sand to coarse aggregate when used.
used for testing of the trained models. Both subsets must 2. Properties of UHPC (relationship between compo-
contain all the possible components in the mixture. nents). This group included maximum size of aggregate
Likewise, to evaluate the network while continuing to (MSA), water-binder ratio (w/b), water-total powders ratio
adjust its parameters (such as the number of neurons in the (w/p), and virtual packing density (VPD).56-58
hidden layer), the training data could be split into a training VPD is not information that is usually provided in the arti-
set and a validation set. However, due to the fact that there cles. In most cases, it had been estimated based on informa-
are so few data points, the validation set would end up tion provided in the articles such as mean particle size (d50) and
being very small for a neural network validation (in this dosage. Only mixtures with enough information to estimate the
case, approximately 162 examples). Hence, the validation VPD have been considered in the database. The VPD was calcu-
scores might change a lot depending on which data points lated according to the compressive packing model theory.56-58
were chosen to use for validation and which were chosen
for training: that is, the validation scores might have a high Model performance evaluation
variance regarding the validation split. This would prevent The ANN model was trained on the training data, through
reliable evaluation of the model.48 the k-fold validation, and then the selected model accu-
The best practice in such situations is to apply k-fold racy was evaluated on the test set. The model’s predictive
cross-validation (refer to Fig. 8). It consists of splitting accuracy was measured through computing mean absolute
the available data into k partitions, instantiating k identical percentage error (MAPE), root mean square error (RSME),

ACI Materials Journal/July 2021 9

Fig. 9—Variations of RMSE against number of hidden
neurons and activation function in k-fold validation training.
normalized mean bias error (NMBE), and coefficient of
multiple determination (R2), using Eq. (5) to (8), respectively

100 n ai − â i
MAPE = ∑ (5)
n i =1 â i Fig. 10—Proposed ANN architecture.
wherein the prediction is the mean of all samples.61 R2 statis-
∑ i =1 ( ai − â i )
n

RMSE = (6) tics are dependent on the linear relationships between the
n observed and predicted values and may sometimes give
biased results when this relationship is not linear or when the
values contain many outliers. A R2 value close to unity indi-
∑ i =1 ( ai − â i )
1 n
cates perfect association between the observed and predicted
NMBE ( %) = n × 100 (7) values. A combined use of the performance metrics narrated
ai
previously can provide an unbiased estimate for prediction
ability of the neural network models.
∑ i =1 ( ai − â i )
n 2

R2 = 1 − (8) RESULTS AND DISCUSSION

∑ i =1 ( â i )
n 2

ANN architecture
where a is the target or experimental value; ā represents the Selecting the model architecture is the first important step
mean observed target; â is the predicted value or model’s in developing an ANN model that best suits the problem at
output; and n is the total number of observations in the hand. Figure 9 depicts the effect of neuron number in hidden
current data. layer and activation function on the performance of different
MAPE is a normalized statistic that provides an effective architectures during the training using k-fold validation. It is
way of comparing the residual error for each data point with important to highlight that Fig. 9 only shows the best perfor-
respect to the observed or target value.21 A lower value of mance initial weight allocation of the neurons for each archi-
MAPE indicates good prediction performance of the model; tecture and activation function combination. According to
RMSE is one of the commonly used error-index statistics.59 this plot, k-fold validation RMSE stopped improving signifi-
RMSE compares the observed values to the predicted values cantly after four neurons in the hidden layer when dealing
and computes the square root of the average residual error, with the hyperbolic tangent activation function, reaching a
indicating error in the units (or squared units) of the constit- minimum RMSE value of 10.82 MPa. On the other hand, the
uent of interest, which aids in analysis of the results. RMSE hidden layer with seven neurons gave the minimum k-fold
values of zero indicate a perfect fit. However, RMSE gives validation RMSE value when using the sigmoid activation
more weightage to large errors21; the NMBE provides infor- function, reaching a value of 10.69 MPa. Consequently,
mation on the mean bias in the predictions from a model. the network architecture shown in Fig. 10, which uses the
A positive NMBE indicates under-prediction and a negative sigmoid activation function, was selected.
NMBE indicates over-prediction of the model.60 The coef- The weights that define the ANN model are presented in
ficient of determination (R2) compares the accuracy of the Table 4. It is important to note that to use the network in
model with the accuracy of a superficial benchmark model future research, it is necessary to normalize the data entering

10 ACI Materials Journal/July 2021

Table 4—Values of ANN weights (refer to Fig. 10)
H1 H2 H3 H4 H5 H6 H7 O1
B1 –0.7317 1.8198 –2.9871 0.3998 –3.0001 –3.0987 –11.7121 B2 0.9876
I1 –1.9959 –4.6757 –2.8102 0.1791 10.1511 –1.1212 –4.2666 H1 –3.3711
I2 –3.6117 1.1302 1.5449 –2.1723 –7.7673 –0.6401 3.8677 H2 –0.5765
I3 –3.5141 2.8134 –1.1638 –1.0176 3.9214 –0.5612 –2.6761 H3 –5.1612
I4 –0.1510 –1.6022 0.3176 4.4168 2.7345 0.1891 –5.1010 H4 –0.6673
I5 2.2141 –1.4110 –0.3681 0.4679 9.5341 0.1916 –7.2502 H5 –0.5234
I6 –1.2082 –5.6107 0.8896 1.1314 1.8629 –0.0226 –11.7821 H6 –0.9649
I7 2.4554 0.7294 0.6481 –0.9417 –0.3235 0.3028 –2.7391 H7 –0.5351
I8 –0.5671 –0.0442 –0.8573 –2.3601 2.4953 –0.2411 –5.7012
I9 –1.5391 3.1214 –2.1750 –0.3476 –14.7221 –0.8743 17.0340
I10 0.1451 –3.3746 2.2488 0.1701 –3.3444 –0.9642 8.5653
I11 1.0902 0.3624 –0.4215 –1.1121 –7.3939 –0.7169 7.5193
I12 0.7312 –2.4810 –1.1588 3.3616 –20.4854 0.4679 1.1041
I13 –0.9385 –6.8422 1.0987 3.1362 0.9391 –0.7713 –3.8823
I14 3.9424 11.1925 0.6876 –2.5395 –8.9491 –0.1905 –5.0774
I15 –2.0765 –8.2185 1.4952 –2.4481 9.7872 –0.9702 8.4455
I16 –1.1771 –3.0601 –1.0142 –1.8633 –4.6713 –1.6241 12.2734
I17 –6.4649 –1.8099 –2.3832 –0.9480 7.6759 –1.9820 –0.8218

Table 5—ANN model performance measures

Model performance
Data subset MAE AE MAPE RMSE E RSR NMBE R2
Train 8.324 5.318 5.830% 10.051 0.809 0.437 3.737% 0.866
Test 8.054 3.977 5.738% 8.924 0.808 0.438 2.779% 0.846

the network according to Eq. (1) by using the range of values be the fact that these researchers worked only with their
described in Table 2. own experimental works. The high amount of data used in
this research, together with the use of k-fold validation, led
Performance evaluation of trained model to a lower R2 value of 0.840. Nevertheless, the proposed
The results in terms of the performance statistics of the model allows one to predict the compressive strength of a
selected model are presented in Table 5. It is observed in wider spectrum of UHPC: form pastes to coarse aggregate
Table 5 that performance evaluation indicators have a concrete, and consider combinations of several SCMs.
similar value in train and test set. The latter indicates the Furthermore, Fig. 11 clearly exhibits the strong coherence
proper carrying out of the k-fold validation used as a proce- between the predicted compressive strength values and the
dure to avoid the overfitting of the ANN model48 during the real ones.
training. Besides, the proposed ANN model predicted the Some of the factors that introduced statistical noise into
28-day compressive strength on the test data set with RMSE, the system, preventing a more precise model, could be:
MAE, and AE values of 9.106, 8.054, and 3.977 MPa, 1) the use of different types of cement; 2) in the data obtained
respectively—that is, the differences between predicted and from scientific literature, the lack of knowledge of whether
experimental compressive strength values were relatively the specimens have been vibrated or not after pouring into
small for UHPC. E and RSR values of 0.808 and 0.438 also the molds; 3) the different nature of aggregates used in
indicated the accuracy of the proposed model. MAPE shows the database; and 4) the different technology of HRWRAs
that the predicted compressive strength deviated on average considered in the database.
by 5.854% from target. NMBE statistics of 2.835% pointed
out a little under-prediction of the model, which also could Connection Weight Approach (CWA)
be observed in Fig. 11. For the other index values (MAE), ANNs have been usually known as “black box” because
on the other hand, it can be noticed that the coefficients of they are believed to present little explanatory insight into the
multiple determination presented (R2) were less than those contributions of the input signals in the prediction process.
obtained by other authors consulted, who obtained correla- However, Olden et al.62 presented the Connection Weight
tion coefficients between 0.96 and 0.98 when predicting Approach (CWA) to assess the importance of the variables
UHPC’s properties using ANN.3,4 One of the reasons could of each model. This method estimates importance as the

ACI Materials Journal/July 2021 11

summed product of the raw input-hidden and hidden-output leads to an internal curing process in the concrete.20 There-
connection weights between each input and output node. An after, other input signals indicated as having positive effect
advantage is the relative contributions of each connection on compressive strength were FA, LP, QP, MK, A, SF, RHA,
weight are kept in both magnitude and sign. For example, and HRWRA, although with a more limited importance. GP
connection weights that change sign (for example, positive to seemed to have a limited negative effect on the compressive
negative) between the input-hidden to hidden-output layers strength of UHPC. It could be explained by the fact that GP
would have a canceling effect. The CWA method was used tends to have a non-regular shape, even needle shape, after
in this study to evaluate the importance of every variable the milling process (refer to Fig. 6(g)), which could decrease
in compressive strength. These analyses are of great help compressive strength of concrete. As expected, the CWA
both to understand the influence of each of the components analysis results seem to indicate that the w/b had the most
of the UHPC on 28-day compressive strength, as well as to negative importance for the strength prediction, followed by
be able to select the components that best fit the properties the water content (W). Finally, previous researches by the
required in future research. The results are shown in Fig. 12. author6,66 confirmed the positive effect of HRWRA and the
In this plot, it can be observed that the most important posi- negative effect of FCC on compressive strength, according
tive inputs belong to VPD. This would agree with the studies to the results presented by the CWA analysis.
that justify the ultra-high compressive strength of UHPC due
to its ultra-dense structure.63-65 It is also important to denote FURTHER RESEARCH
that the effect of SF could be masked into the effect of VPD Efforts should be made to find out the possible ways to
because silica fume content, due to its lower size (refer to decrease the cost and duration of the experimental campaign
Fig. 6(a)), is responsible for increasing VPD by filling the when developing UHPC using supplementary cementitious
gap between the other particles.10 As such, cement content materials, which would allow a faster spread of the use of
was presented to have the second positive importance in these materials. For this purpose, predictive mathemat-
compressive strength. Besides, CWA analysis indicated that ical models can help, allowing pre-selection of available
RHA also had a positive effect on compressive strength that component combinations and reducing the number of trials
could be related to its high siliceous content (as indicated to be performed. This research provided an accurate one-
in Table 3) and its high porosity (refer to Fig. 6(d)), which hidden-layer ANN model for predicting the 28-day compres-
sive strength of UHPC. Areas for future research include
the development of regression models for other properties
of UHPC such as module of elasticity, chloride penetra-
tion, shrinkage, and carbonization, even for the properties
derived from the addition of fibers to the UHPC such as
toughness and ductility. Those mathematical models could
include other artificial intelligence procedures to improve
the correlation coefficient and accuracy of predictions.
Such mathematical procedures could include bootstrapping,
random forest, and recurrent networks, among others. The
results of such studies would directly benefit the research
about UHPC.

CONCLUSIONS
This study provides an accurate ANN model for predic-
tion of the 28-day compressive strength of UHPC using
different combinations of supplementary cementitious
materials and fillers, such as silica fume, fly ash, ground-
granulated blast-furnace slag, glass powder, rice husk ash,
fluid catalytic cracking residue, metakaolin, limestone
Fig. 11—Regression plot for UHPC pastes.

Fig. 12—Importance of input variables for 28-day compressive strength.

12 ACI Materials Journal/July 2021

powder, and quartz powder. Furthermore, different ranges of Performance Concrete with Recycled Glass,” Dyna (Bilbao), V. 86,
No.  211, 2019, pp. 84-93. doi: 10.15446/dyna.v86n211.79596
maximum aggregate were considered in the development of 7. Abellán-García, J.; Núñez-López, A.; Torres-Castellanos, N.; and
the ANN to be able to predict the compressive strength of Fernández-Gómez, J., “Factorial Design of Reactive Concrete Powder
pastes, mortars, and concrete type UHPC. To avoid over- Containing Electric Arc Slag Furnace and Recycled Glass Powder,”
Dyna (Bilbao), V. 87, No. 213, 2020, pp. 42-51. doi: 10.15446/dyna.
fitting, k-fold validation with 10 partitions was used, which v87n213.82655
lead to higher confidence avoiding the overfitting when new 8. Abellán-García, J., “Four-Layer Perceptron Approach for Strength
data are used. Based on the results of this investigation, the Prediction of UHPC,” Construction and Building Materials, V. 256, 2020,
p. 119465. doi: 10.1016/j.conbuildmat.2020.119465
following conclusions are drawn: 9. Abellán, J.; Fernández, J.; Torres, N.; and Núñez, A., “Development
1. The proposed model is able to predict the compres- of Cost-Efficient UHPC with Local Materials in Colombia,” Proceedings
sive strength of new data regardless of the supplementary of Hipermat 2020 - 5th International Symposium on UHPC and Nanotech-
nology for Construction Materials, Kassel, Germany, 2020, pp. 97-98.
cementitious materials combination and maximum size of 10. Meng, W.; Samaranayake, V.; and Khayat, K. H., “Factorial Design
aggregate. It could be helpful in the developing of UHPC and Optimization of Ultra-High-Performance Concrete with Lightweight
by reducing costs and time in the experimental campaign Sand,” ACI Materials Journal, V. 115, No. 1, Jan. 2018, pp. 129-138. doi:
10.14359/51700995
by pre-selecting the combinations of available SCMs, which 11. Abellán-García, J.; Santofimo-Vargas, M. A.; and Torres-Castellanos,
have better results in the model at lower cost. N., “Analysis of Metakaolin as Partial Substitution of Ordinary Portland
2. The results of the combined use of the performance Cement in Reactive Powder Concrete,” Advances in Civil Engineering
Materials, V. 9, No. 1, 2020, pp. 368-386. doi: 10.1520/ACEM20190224
metrics, which included MAPE, RMSE, NMBE, and 12. Abellán-García, J.; Fernández-Gómez, J.; and Torres-Castellanos,
R2, provided an unbiased estimate which confirmed the N., “Properties Prediction of Environmentally Friendly Ultra-High-
adequacy of the proposed model. Performance Concrete Using Artificial Neural Networks,” European
Journal of Environmental and Civil Engineering, 2020, pp. 1-25. doi:
3. The results of the sensitivity CWA analysis showed 10.1080/19648189.2020.1762749
consistency with the results of several international inves- 13. Camacho, E.; López, J. Á.; and Serna, P., “Definition of Three Levels
tigations. This reaffirmed the validity of the model from a of Performance for UHPFRC-VHPFRC with Available Materials,” 3rd
International Symposium on UHPC and Nanotechnology for Construc-
scientific point of view. tion Materials, M. Schmidt, E. Fehling, C. Glotzbach, S. Fröhlich, and S.
Piotrowski, eds., Kassel University, Kassel, Germany, 2012, pp. 249-256.
AUTHOR BIOS 14. Camacho Torregrosa, E. E., “Dosage Optimization and Bolted
Joaquín Abellán-García is a PhD candidate of the Department of Civil Connections for UHPFRC Ties,” PhD thesis, Universitat Politècnica de
Engineering at the Polytechnic University of Madrid (UPM), Madrid, València, València, Spain, 2013, 276 pp.
Spain. He is also a Professor of civil engineering at Escuela Colombiana 15. Li, W.; Huang, Z.; Zu, T.; Shi, C.; Duan, W. H.; and Shah, S. P.,
de Ingeniería Julio Garavito, Bogotá, Colombia. His research interests “Influence of Nanolimestone on the Hydration, Mechanical Strength, and
include ultra-high-performance concrete, artificial intelligence applied Autogenous Shrinkage of Ultrahigh-Performance Concrete,” Journal of
to civil engineering, mathematical optimization of concrete, and seismic Materials in Civil Engineering, ASCE, V. 28, No. 1, 2016, pp. 1-9. doi:
behavior of high-strain-hardening cementitious composites. 10.1061/(ASCE)MT.1943-5533.0001327
16. Huang, Z., and Cao, F., “Effects of Nano-Materials on the Perfor-
mance of UHPC,” Materials Review B, Research, V. 26, No. 9, 2012,
ACKNOWLEDGMENTS pp.  136-141.
Special thanks go to APOLO from EAFIT University, Medellín, 17. Soliman, N. A., and Tagnit-Hamou, A., “Partial Substitution of Silica
Colombia, and OSIRIS from Escuela Colombiana de Ingeniería Julio Gara- Fume with Fine Glass Powder in UHPC: Filling the Micro Gap,” Construc-
vito for the servers and computer-related support. Also, to Cementos Argos tion and Building Materials, V. 139, 2017, pp. 374-383. doi: 10.1016/j.
SA for donating most of the materials used in the research described herein conbuildmat.2017.02.084
and providing the SEM images of the components. The supply of recycled 18. Tagnit-Hamou, A.; Soliman, N.; and Omran, A., “Green
glass from Cristalería Peldar SA, FCC from Ecopetrol SA, and GBFS from Ultra-High-Performance Glass Concrete,” First International Interactive
Gerdau SA for this research are highly appreciated. The author would also Symposium on UHPC, 2016.
like to acknowledge the support and suggestions of Escuela Colombiana de 19. Soliman, N. A., and Tagnit-Hamou, A., “Using Glass Sand as an
Ingeniería Julio Garavito and the Polytechnic University of Madrid (UPM). Alternative for Quartz Sand in UHPC,” Construction and Building Mate-
rials, V. 145, 2017, pp. 243-252. doi: 10.1016/j.conbuildmat.2017.03.187
20. Viet Thein An, V., and Ludwig, H.-M., “Proportioning Optimization
REFERENCES of UHPC Containing Rice Husk Ash and Ground Granulated Blast-Furnace
1. Abellán, J.; Torres, N.; Núñez, A.; and Fernández, J., “Ultra High
Slag,” 3rd International Symposium on UHPC and Nanotechnology for
Performance Fiber Reinforced Concrete: State of the Art, Applications and
Construction Materials, M. Schmidt, E. Fehling, C. Glotzbach, S. Fröh-
Possibilities into the Latin American Market,” XXXVIII Jornadas Sudamer-
lich, and S. Piotrowski, eds., Kassel University, Kassel, Germany, 2012,
icanas de Ingeniería Estructural, Lima, Peru, 2018.
pp. 197-205.
2. Abellán, J.; Torres, N.; Núñez, A.; and Fernández, J., “Influencia del
21. Chandwani, V.; Agrawal, V.; and Nagar, R., “Modeling Slump of
Exponente de Fuller, la Relación Agua Conglomerante y el Contenido en
Ready Mix Concrete Using Genetic Algorithms Assisted Training of Arti-
Policarboxilato en Concretos de Muy Altas Prestaciones,” IV Congreso
ficial Neural Networks,” Expert Systems with Applications, V. 42, No. 2,
Internacional de Ingeniería Civil, Havana, Cuba, 2018.
2015, pp. 885-893. doi: 10.1016/j.eswa.2014.08.048
3. Zhang, J.; Zhao, Y.; and Li, H., “Experimental Investigation and
22. Khashman, A., and Akpinar, P., “Non-Destructive Prediction of
Prediction of Compressive Strength of Ultra-High Performance Concrete
Concrete Compressive Strength Using Neural Networks,” Procedia
(UHPC) Containing Supplementary Cementitious Materials,” Advances
Computer Science, V. 108, 2017, pp. 2358-2362. doi: 10.1016/j.
in Materials Science and Engineering, V. 2017, 2017, pp. 522-525. doi:
procs.2017.05.039
10.1155/2017/4563164
23. Adeli, H., “Neural Networks in Civil Engineering: 1989 − 2000,”
4. Ghafari, E.; Bandarabadi, M.; Costa, H.; and Júlio, E., “Prediction
Computer-Aided Civil and Infrastructure Engineering, V. 16, No. 2, 2001,
of Fresh and Hardened State Properties of UHPC: Comparative Study
pp. 126-142. doi: 10.1111/0885-9507.00219
of Statistical Mixture Design and an Artificial Neural Network Model,”
24. Getahun, M. A.; Shitote, S. M.; and Abiero Gariy, Z. C., “Artificial
Journal of Materials in Civil Engineering, ASCE, V. 27, No. 11, 2015,
Neural Network Based Modelling Approach for Strength Prediction of
p.  04015017. doi: 10.1061/(ASCE)MT.1943-5533.0001270
Concrete Incorporating Agricultural and Construction Wastes,” Construc-
5. ACI Committee 239, “Ultra-High-Performance Concrete: An
tion and Building Materials, V. 190, 2018, pp. 517-525. doi: 10.1016/j.
Emerging Technology Report (ACI 239R-18),” American Concrete Insti-
conbuildmat.2018.09.097
tute, Farmington Hills, MI, 2018, 21 pp.
25. Chandwani, V., and Nagar, R., “Applications of Artificial Neural
6. Abellán-García, J.; Núñez-López, A.; Torres-Castellanos, N.; and
Networks in Modeling Compressive Strength of Concrete: A State of the
Fernández-Gómez, J., “Effect of FC3R on the Properties of Ultra-High-
Art Review,” International Journal of Current Engineering and Tech-
nology, V. 4, No. 4, 2014, pp. 2949-2956.

ACI Materials Journal/July 2021 13

 26. Taghaddos, H.; Mahmoudzadeh, F.; Pourmoghaddam, A.; and 47. Bharathi, S. D.; Manju, R.; and Premalatha, J., “Prediction of
Shekarchizadeh, M., “Prediction of Compressive Strength Behaviour in Compressive Strength for Self- Compacting Concrete (SCC) Using Artifi-
RPC with Applying an Adaptive Network-Based Fuzzy Interface System,” cial Intelligence and Regression Analysis,” International Journal of Chem-
Proceedings of the International Symposium on Ultra High Performance tech Research, V. 10, No. 8, 2017, pp. 263-275.
Concrete, Kassel, Germany, 2004. 48. Chollet, F., and Allaire, J. J., Deep Learning with R, Manning Publi-
27. Schmidt, M.; Fehling, E.; and Geisenhanslükem, C., eds., Proceed- cations Co., Shelter Island, NY, 2018.
ings of the International Symposium on Ultra High Performance Concrete, 49. Demir, F., “Prediction of Elastic Modulus of Normal and High
University of Kassel, Kassel, Germany, 2004. Strength Concrete by Artificial Neural Networks,” Construction and
28. Fehling, E.; Schmidt, C.; and Stüwald, S., eds., Proceedings of the Building Materials, V. 22, No. 7, 2008, pp. 1428-1435. doi: 10.1016/j.
Second International Symposium on Ultra High Performance Concrete, conbuildmat.2007.04.004
University of Kassel, Kassel, Germany, 2008. 50. Estebon, M. D., “Perceptrons: An Associative Learning Network,”
29. Schmidt, M.; Fehling, E.; Glotzbach, C.; Fröhlich, S.; and Piotrowski, Virginia Tech, Blacksburg, VA, 1997.
S., eds., Proceedings of Hipermat 2012 3rd International Symposium on 51. Rumelhart, D.; Hinton, G.; and Williams, R., “Learning Internal
UHPC and Nanotechnology for High Performance Construction Materials, Representations by Error Propagation,” Parallel Distributed Processing:
University of Kassel, Kassel, Germany, 2012. Explorations in the Microstructures of Cognition, D. Rumelhart and J.
30. Fehling, E.; Middendorf, B.; and Thiemicke, J., eds., Proceedings McClelland, eds., 1986, pp. 318-362.
of Hipermat 2016 4th International Symposium on UHPC and Nano- 52. Mushgil, H. M.; Alani, H. A.; and George, L. E., “Comparison
technology for High Performance Construction Materials, University of between Resilient and Standard Back Propagation Algorithms Efficiency in
Kassel, Kassel, Germany, 2016. Pattern Recognition,” International Journal of Scientific and Engineering
31. Skazlic, M.; Bjegovic, D.; and Serdar, M., “Influence of Test Spec- Research, V. 6, No. 3, 2015, pp. 773-778.
imens Geometry on Compressive Strength of Ultra-High Performance 53. Prasad, N.; Singh, R.; and Lal, S. P., “Comparison of Back Propaga-
Concrete,” Proceedings of the 2nd International Symposium on Ultra High tion and Resilient Propagation Algorithm for Spam Classification,” Proceed-
Performance Concrete, Kassel, Germany, 2008, pp. 295-301. ings of International Conference on Computational Intelligence, Modelling
32. Naaman, A. E., and Wille, K., “Some Correlation between High and Simulation, 2013, pp. 29-34. doi: 10.1109/CIMSim.2013.1410.1109/
Packing Density, Ultra-High Performance, Flow Ability, and Fiber Rein- CIMSim.2013.14
forcement of a Concrete Matrix,” Bac2010, Congresso Ibêrico Sobre Betão 54. R Core Team, R: A Language and Environment for Statistical
Auto-Compactável, 2010. Computing, Vienna, Austria, 2018, www.r-project.org. (last accessed
33. Graybeal, B., and Davis, M., “Cylinder or Cube: Strength Testing June 8, 2021)
of 80 to 200 MPa (11.6 to 29 ksi) Ultra-High-Performance Fiber- 55. Günther, F., and Fritsch, S., “neuralnet: Training of Neural Networks,”
Reinforced Concrete,” ACI Materials Journal, V. 105, No. 6, Nov.-Dec. The R Journal, V. 2, No. 1, 2010, pp. 30-38. doi: 10.32614/RJ-2010-006
2008, pp.  603-609. 56. de Larrard, F., and Sedran, T., “Optimization of Ultra-High-
34. Atkinson, A., and Riani, M., Robust Diagnostic Regression Analysis, Performance Concrete by the Use of a Packing Model,” Cement
Springer, New York, 2000. and Concrete Research, V. 24, No. 6, 1994, pp. 997-1009. doi:
35. Härdle, W. K., and Simar, L., Applied Multivariate Statistical Analysis, 10.1016/0008-8846(94)90022-1
Springer-Verlag GmbH, Berlin, Germany, 2012. 57. de Larrard, F., “Concrete Mixture Proportioning: A Scientific
36. Everitt, B., and Hothorn, T., An Introduction to Applied Multivariate Approach,” Modern Concrete Technology Series, E&FN Spon, London,
Analysis with R, Springer, New York, 2011, 288 pp. UK, 1999.
37. Arizzi, A., and Cultrone, G., “Comparing the Pozzolanic Activity of 58. de Larrard, F., and Sedran, T., “Mixture-Proportioning of High-
Aerial Lime Mortars Made with Metakaolin and Fluid Catalytic Cracking Performance Concrete,” Cement and Concrete Research, V. 32, No. 11,
Catalyst Residue: A Petrographic and Physical-Mechanical Study,” 2002, pp. 1699-1704. doi: 10.1016/S0008-8846(02)00861-X
Construction and Building Materials, V. 184, 2018, pp. 382-390. doi: 59. Moriasi, D. N.; Arnold, J. G.; Van Liew, M. W.; Bingner, R. L.;
10.1016/j.conbuildmat.2018.07.002 Harmel, R. D.; and Veith, T. L., “Model Evaluation Guidelines for System-
38. Torres Castellanos, N., “Estudio en Estado Fresco y Endure- atic Quantification of Accuracy in Watershed Simulations,” Transactions
cido de Concretos Adicionados con Catalizados de Craqueo Catalítico of the ASABE, V. 50, No. 3, 2007, pp. 885-900. doi: 10.13031/2013.23153
Usado (FCC),” PhD thesis, Universidad Nacional de Colombia, Bogotá, 60. Srinivasulu, S., and Jain, A., “A Comparative Analysis of Training
Colombia, 2014. Methods for Artificial Neural Network Rainfall–Runoff Models,”
39. Li, Z., and Rangaraju, P. R., “Development of UHPC Using Ternary Applied Soft Computing, V. 6, No. 3, 2006, pp. 295-306. doi: 10.1016/j.
Blends of Ultra-Fine Class F Fly Ash, Metakaolin and Portland Cement,” asoc.2005.02.002
First International Interactive Symposium on UHPC, 2016, pp. 1-12. doi: 61. Gupta, S., “Using Artificial Neural Network to Predict the Compres-
10.21838/uhpc.2016.6410.21838/uhpc.2016.64 sive Strength of Concrete containing Nano-silica,” Civil Engineering and
40. Ferdosian, I.; Camões, A.; and Ribeiro, M., “High-Volume Fly Architecture, V. 1, No. 3, 2013, pp. 96-102. doi: 10.13189/cea.2013.010306
Ash Paste for Developing Ultra-High Performance Concrete (UHPC),” 62. Olden, J. D.; Joy, M. K.; and Death, R. G., “An Accurate Compar-
Ciência e Tecnologia dos Materiais, V. 29, No. 1, 2017, pp. e157-e161. doi: ison of Methods for Quantifying Variable Importance in Artificial Neural
10.1016/j.ctmat.2016.10.001 Networks Using Simulated Data,” Ecological Modelling, V. 178, No. 3-4,
41. Ghafari, E.; Costa, H.; Júlio, E.; Portugal, A.; and Durães, L., 2004, pp. 389-397. doi: 10.1016/j.ecolmodel.2004.03.013
“Enhanced Durability of Ultra High Performance Concrete by Incorpo- 63. Ghafari, E.; Costa, H.; and Júlio, E., “RSM-Based Model to Predict
rating Supplementary Cementitious Materials,” Second International the Performance of Self-Compacting UHPC Reinforced with Hybrid
Conference on Microstructure Related Durability of Cementitious Compos- Steel Micro-Fibers,” Construction and Building Materials, V. 66, 2014,
ites, Amsterdam, the Netherlands, 2012, pp. 11-13. pp.  375-383. doi: 10.1016/j.conbuildmat.2014.05.064
42. Cook, R. D., “Detection of Influential Observation in Linear 64. Abbas, S.; Nehdi, M. L.; and Saleem, M. A., “Ultra-High Perfor-
Regression,” Technometrics, V. 19, No. 1, 1977, pp. 15-18. doi: mance Concrete: Mechanical Performance, Durability, Sustainability
10.1080/00401706.1977.10489493 and Implementation Challenges,” International Journal of Concrete
43. ASTM C109/C109M-11b, “Standard Test Method for Compressive Structures and Materials, V. 10, No. 3, 2016, pp. 271-295. doi: 10.1007/
Strength of Hydraulic Cement Mortars (Using 2-in. or [50-mm] Cube Spec- s40069-016-0157-4
imens),” ASTM International, West Conshohocken, PA, 2011, 9 pp. 65. Schmidt, C., and Schmidt, M., “‘Whitetopping’ of Asphalt and
44. Funk, J. E., and Dinger, D. R., Predictive Process Control Concrete Pavements with Thin Layers of Ultra-High-Performance Concrete -
of Crowded Particulate Suspensions. Applied to Ceramic Construction and Economic Efficiency,” 3rd International Symposium on
Manufacturing, Springer Science, New York, 1994. doi: UHPC and Nanotechnology for Construction Materials, M. Schmidt, E.
10.1007/978-1-4615-3118-010.1007/978-1-4615-3118-0 Fehling, C. Glotzbach, S. Fröhlich, and S. Piotrowski, eds., Kassel Univer-
45. Rosenblatt, F., “The Perceptron: A Probabilistic Model for Informa- sity, Kassel, Germany, 2012.
tion Storage and Organization in the Brain,” Psychological Review, V. 65, 66. Abellán, J.; Fernández, J.; Torres, N.; and Núñez, A., “Statistical
No. 6, 1958, pp. 386-408. doi: 10.1037/h0042519 Optimization of Ultra-High-Performance Glass Concrete,” ACI Materials
46. Anderson, J. A., “Cognitive and Psychological Computation with Journal, V. 117, No. 1, Jan. 2020, pp. 243-254. doi: 10.14359/51720292
Neural Models,” IEEE Transactions on Systems, Man, and Cybernetics,
V.  13, No. 5, 1983, pp. 799-815. doi: 10.1109/TSMC.1983.6313074

14 ACI Materials Journal/July 2021