Beruflich Dokumente
Kultur Dokumente
h i g h l i g h t s
a r t i c l e i n f o a b s t r a c t
Article history: This research is aimed to develop a four-layer multi-layer-perceptron (MLP) model for predicting the
Received 10 February 2020 compressive strength of ultra-high-performance concrete (UHPC), regardless of the combination of a
Received in revised form 10 April 2020 wide range of supplementary cementitious materials (SCM) or the maximum size of aggregate used.
Accepted 4 May 2020
UHPC is a high-tech type of concrete resulted of the mixture of several constituents. Therefore, the effect
of each component and their interactions on compressive strength is more difficult to understand than in
conventional concrete. A total of 210 own experimental campaign data added to 717 published work
Keywords:
throughout the world data were used for training purposes by using the R-code language. To analyze
UHPC
MLP
the relationships between the UHPC’s components and strength, the Olden algorithm was used. The
Compressive strength interpretation of both the statistical performance metrics and the results of Olden’s sensitivity analysis
Resilient back-propagation algorithm indicated that the proposed model was an efficient approach for predicting the compressive strength
Olden algorithm of UHPC. The trained MLP model can be used for forecasting the compressive strength for a given
Maximum size of aggregate UHPC mixture design in quick time without performing any trial.
SCM Ó 2020 Elsevier Ltd. All rights reserved.
1. Introduction years, great efforts have been made in achieving a less expensive
and more sustainable UHPC, by using supplementary cementitious
Over the last twenty years, remarkable advances have taken materials (SCM) as replacement for cement (partial), quartz pow-
place in the research on ultra-high-performance concrete (UHPC). der (total) and silica fume (partial or total). Therefore, some indus-
UHPC is a type high-tech concrete that is characterized by its trial by-products have been used as components of UHPC’s binder
ultra-high compressive strength, low permeability, and improved [10] such as fly ash (FA), ground granulated blast slag furnace
durability [1–8]. However, due to the absence of a coarse aggregate (GGBSF), recycled glass powder (GP), rice husk ash (RHA) and fluid
together with its need of achieving high packing density, UHPC catalytic cracking residue (FC3R) as well as other SCM such lime-
contents of binder used (usually cement and silica fume) are rela- stone powder (LP) and metakaolin (MK). Among others, RHA was
tively higher in relation to conventional concrete. The latter analyzed as SCM in UHPC with successful results in strength and
increases its cost and carbon footprint [1,2,9]. Therefore, notwith- durability due to its high silica content and the internal-curing pro-
standing the superb performance in strength and durability, UHPC cess of the concrete [11]. Blending limestone powder in UHPC’s
has not been widely applied in construction primarily because of binder enhanced the hydration process at the early-age, giving
its high costs and carbon footprint. According to that, in recent higher particle packing, density and enhancing mechanical proper-
ties [12]. Besides, nano-CaCO3 as a component of binder was also
⇑ Address: PhD Candidate of the Department of Civil Engineering at the investigated [13]. The conclusion of that work showed that com-
Polytechnic University of Madrid (UPM), Madrid, Spain. pressive strength increased a 17% compared to the UHPC control
E-mail address: j.abellang@alumnos.upm.es
https://doi.org/10.1016/j.conbuildmat.2020.119465
0950-0618/Ó 2020 Elsevier Ltd. All rights reserved.
2 J. Abellán-García / Construction and Building Materials 256 (2020) 119465
specimens without nano-CaCO3 [13]. On the other hand, Soliman & RHA, FC3R, MK and LP in addition to mineral powder such as
Tagnit-Hamou [14] conducted compressive tests on UHPC wherein quartz powder (QP). The model developed was also able to con-
fine glass powder (FGP), with a mean particle size (d50) of 3.8 mm sider form paste (non-aggregate concrete) to a maximum size of
was used as partial substitution of silica fume. They demonstrated aggregate of 19 mm.
that compressive strength of 235 and 220 MPa could be achieved
when using steam curing and replacing 30% and 50% of SF with
2. Database
FGP respectively. In another study by the same authors recycled
glass powder was used to take the place of quartz powder as well
2.1. Data collection
as partial substitution of quartz sand, and cement [15,16].
Other studies of UHPC were performed containing FA, SF and
FC3R, a by-product of the crude oil industry [17,18]. Abellan Datasets containing dosages and 28-days compressive strength
et al. [1] performed an UHPC mixture with a low silica fume and of UHPC were collected from several international symposiums on
cement content by using limestone powder and recycled glass high and ultra-high-performance concrete (including the interna-
powder as SCM. In other research the same authors demonstrated tional symposiums held in Kassel [25–28]), PhD thesis, and other
the capacity of FC3R when used as SCM [2]. These investigations published works. Only those dosages with information of the vir-
tual packing density or/and the mean particle size of all component
revealed the possibility of partial replacement of cement and silica
fume in UHPC mixes by other SCM, without significantly reducing were collected, totalizing 717 observations from the scientist liter-
ature. In addition, 210 experimental tests with different combina-
its compressive strength.
The key properties of UHPC with containing several SCMs tions of SCM were performed in various laboratories. As a result,
the present investigation has had a total of 927 observations for
require to be investigated experimentally due to the unclear com-
bination effects of different materials and dosages being utilized in the development of the mathematical model.
The values of compressive strength were transformed to cube of
the mixture. Besides, most of times laboratory tests are labor-
intensive, time-consuming, and expensive [5]. In order to decrease 100 mm side by means of transformation factors, owing to differ-
ent specimens shape and dimensions were observed in published
the experimental works and mix designs, probabilistic models can
be successfully applied to predict the compressive strength of con- investigations and experimental works [29–32]. The multiplication
transforming factors used are listed in Table 1.
crete. However, these models cannot be employed when the
approaching problem involves too many independent factors and On the one hand, the experimental campaign only dealt with
ASTM Type III HE cement and siliceous sand, in addition to not
the interactions amidst them are either too complex or unknown
to represent [5,6,19]. In the case like the concrete using several including quartz powder in the dosages. On the other hand, the
definitive database, enriched with other researches, included vari-
SCM the use of traditional techniques of approach fails to achieve
the expected accuracy and predictability, owing to its large amount ous types of cement, different possibilities of aggregate, from
pastes to concrete with coarse aggregate, and even considered
of components [19].
On the other hand, progresses in Artificial Neural Network dosages with quartz powder.
(ANN) have had an ever-expanding application for approaching
the real-world problems due to its immense ability to map non- 2.2. Dealing with the outliers
linear and unknown relationships between input and output data
pairs. Because of the latter, those mathematical procedures had Before performing regression analyses it was necessary to deal
demonstrated its effectiveness in answering difficult and complex with outliers, as they can greatly affect the resulting model [33].
engineering problems [20]. Hence, a descriptive statistical analysis was carried out on each
In the engineering field, ANN models have been employed in variable to identify outliers [34]. Unusual distributions (skewness),
applications like detection of structural damage, water resources data entry errors and outliers in the data were detected using
engineering, traffic engineering, structural system identification, bivariate boxplots. Bivariate boxplots are two-dimensional equiva-
material behavior modeling, concrete mix proportioning and con- lent of single variable boxplots that are employed to spot inconsis-
crete strength forecasting [21,22]. In fact, the application of ANN tent data and outliers by robust methods, drawing ellipses that lay
to forecast mechanical properties of pastes, mortars and concretes out possible disturbing observations [35]. Nevertheless, trusting in
had become one of the most fertile fields in the scientist literature the use of this methodology without a critical analysis of the data
of civil engineering production [23]. However, until now few inves- could also be a dangerous practice. Some of the suspicious points
tigations have been conducted on forecasting the performance of pointed out by bivariate boxplots could be showing the real behav-
UHPC using neural networks. Among those few researches, Ehsan ior of the data, while the rest of the observations could just be
Ghafari et al. [6], Taghaddos et al. [24], and Zhang and Zhao [5] clumped together very closely. Fig. 1 depicts the bivariate boxplot
investigations can be highlighted. Yet only one of these investiga- for the pair of variables compressive strength (MPa) and water con-
tions includes other supplementary cementitious material (i.e. fly tent (volume ratio). This graph clearly illustrates an example of the
ash) in addition to silica fume [5]. Furthermore, only one of the situation where important data would be lost if all the data marked
three considered the addition of quartz flour [6]. as outliers were to be eliminated. In Fig. 1 25% of the dosages con-
The main objective of this research work was to develop a four- taining MK (blue points of the graph) could be considered as out-
layer (i.e. one input, two hidden and one output) neural network liers (red points) when falling out of the ellipse. Having said that,
model for predicting the 28-days compressive strength of UHPC. it is important to highlight that several studies have demonstrated
The model was designed in such a way that it could consider any that the inclusion of MK in a concrete requires more water or
combination of a broad spectrum of SCM, viz., SF, FA, GGBSF, GP, superplasticizer to reach a determined workability, due to its reac-
Table 1
Transforming factors to turn the compressive strength results to 100 mm side cube specimen.
Type of specimen Cube Cube Cube Cube Cube Cyl. Cyl. Cyl. Cyl.
Dimensions (mm) 150 100 70 50 40 50 100 75 150 100 200 150 300
Factor to 100 mm cube 1.119 1.000 0.935 0.962 1.121 1.020 1.020 1.020 1.063
J. Abellán-García / Construction and Building Materials 256 (2020) 119465 3
where xnorm is the normalized value of the variable x, xmax and xmin
are the minimum and maximum values of variable x respectively.
Table 2 shows the range of variation of all variables considered
in the database (including both experimental and collected
dosages) after dealing with the outliers.
Fig. 1. Bivariate boxplot for the pair of variables Compressive strength and Water
content.
3. Experimental investigation
Fig. 3. Pie chart from data: a) Number of components of binder; b) Cement ASTM types reported; c) Maximum size of aggregate; and d) Type of coarse aggregate when used.
Fig. 4. Boxplot for the component of UHPC in the data base and some relationships such as water-to-binder ratio, water-to-total-powders ratio and virtual packing density.
particle, the spherical shape of the FA and SF particles, and the components of the concrete were determined by the A&Amod curve
small size of the latter. [42] according to Eq. (2).
3.2. Specimens Dq Dqmin
PðDÞ ¼ ð2Þ
Dqmax Dqmin
The experimentally concrete mixtures corresponded to 15
groups of component combinations. Once the combination of
SCM was selected, a three-factor Design of Experiments (DoE) with where D is the particle size, P(D) is the weight fraction of total solids
14 runs each was established and performed. The first factor of DoE that are smaller than D, Dmax and Dmin are the maximum and min-
corresponded to the amount of cement, the second to the amount imum particle sizes, respectively, and q is the Fuller exponent. The
of water and the third to the amount of superplasticizer. The other optimal value of q was determined in previous research [4].
J. Abellán-García / Construction and Building Materials 256 (2020) 119465 5
Fig. 5. Scatterplot matrix for every pair of variables in the database including the compressive strength of concrete (F28).
Table 2
Range of variation of all variables considered in this research.
C* SF* FA* GGBSF* RGP* RHA* FC3R* MK* LP* W* HRWR* QP* MSAy WB WP A* VPD CSà
xmax 0.503 0.155 0.250 0.247 0.202 0.177 0.091 0.161 0.190 0.364 0.063 0.235 19,000 0.322 0.322 0.694 0.877 214
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
Mean 0.246 0.063 0.016 0.008 0.027 0.004 0.005 0.002 0.025 0.202 0.026 0.026 1724 0.187 0.177 0.364 0.771 143
SD¥ 0.072 0.035 0.043 0.025 0.052 0.018 0.014 0.013 0.043 0.037 0.012 0.051 2498 0.037 0.034 0.130 0.044 23
Table 3
Physical and chemical compositions of cement and supplementary cementitious materials.
3.3. Items of investigation value to feed the database. The molds and test setup are shown in
Fig. 8.
After each dosage was designed, a 5-liter mortar mixer was uti-
lized to fabricate the concrete. Once the mixing was completed, the
UHPC was cast in molds using a vibrating table for one minute. The 4. Analytical investigation
prisms were demolded 24 h after casting and then cured in a mois-
ture room at 20 °C until the day of the test, without any heat treat- 4.1. Artificial neural networks
ment applied. For the determination of the compressive strength,
cubes of 50 mm were evaluated. A compression testing machine The simplest form of neural network architecture is the percep-
with a capacity of 3000 kN was utilized according to ASTM C109 tron, devised by Rosenblatt in 1958 [43]. Perceptron consists of one
[41]. Three samples were evaluated for checking 28-day compres- neuron with two inputs and one output [44] and is defined as a
sive strength. The average value of these three being the reported four-tuple entity (i.e., sensors that (i) receive inputs and (ii) multi-
6 J. Abellán-García / Construction and Building Materials 256 (2020) 119465
X
n
vj ¼ wij xi þ b ð3Þ
i¼1
where vj is the weighted sum of the jth neuron for the input col-
lected from the preceding layer with n neurons, wij is the weight
between the jth neuron and the ith neuron in the previous layer;
xi represents the output of the ith neuron in the preceding layer
and b is the bias [48].
Once the weighted sums of the input components are com-
puted, the activation function processes this value and ascertains
the neuron output. Some of the most utilized activation functions
are linear, ramp, tanh (hyperbolic tangent), sigmoid functions
and relu (rectified linear unit function) [49]. In this study the sig-
moid function was used as activation function for hidden layers
whereas linear function was used for output layer. The sigmoid
function takes the input, which can have any value between minus
Fig. 6. Particle size distribution of cement, sand and supplementary cementitious
and plus infinite and squashes the output in the range 0–1 [50].
materials.
The latter implies that the normalization of the data for the train-
ing of the network must be carried out.
ply them by weights, (iii) a function collecting all the weighted Regarding to the training process, Rumelhart et al. [51] devel-
data to produce a measurement on the impact of the observed phe- oped the most commonly used learning algorithm called back-
nomenon, and (iv) a constant threshold). Ascertaining these propagation (BP) for multi-layer perceptron. BP training algorithm
weights to yield a particular output is called ‘‘training”, which is is an iterative gradient designed to minimize the mean square
the procedure that allows the model to learn [45]. Fig. 9 shows error between the actual output of multi-layer feed forward neural
the schematic diagram of perceptron structure. network and the real values of response [52]. Latterly other learn-
For more complex applications, multi-layer perceptron (MLP) ing algorithms have been developed, such as the resilient back-
are used, which contain one input layer, one output layer, and propagation (Rprop) which has two main advantages over back-
one or more hidden layers as shown in Fig. 10. The multilayer per- propagation: First, training with Rprop is usually faster than train-
ceptron (feed-forward network) has been a commonly used neural ing with back-propagation. Second, Rprop remove the negative
network architecture [19,22,46]. influence of the size of the partial error derivative on the weight
Multi-layer perceptron models are composed of many highly step, as opposed to back-propagation which needs values for the
interconnected processing neurons working together [21]. Each learning rate (and frequently an optional momentum term). In
neuron is fully connected to the other through connection weights consequence only the sign of the derivative is taken into account
and receives an input signal from neurons connected to it. The suc- to indicate the direction of weight update [52,53]. For further infor-
cessive layers of neurons collect input from the previous layers; mation about Rprop refer to [52] and [53].
the outputs of neurons in each layer are inputs to neurons in the In this work, the multi-layer perceptron models were trained
next layer. To understand the latter it is necessary to analyze the using resilient backpropagation as learning algorithm.
schematic diagram of the perceptron showed in Fig. 1, where the
vector xi = (x1, x2,. . .,xn) is an input signal applied to the neuron. 4.2. Preparing training and test data set
As well, a bias is added to the neuron along with inputs. The vector
wi = (w1, w2,. . .,wn) is the weights for each input. The weighted sum The obtained data was randomized and divided into training
function computes the net input that approaches to a neuron [47]. and test datasets to facilitate training and testing of the multi-
The weighted sums of the input components are computed by layer perceptron. A total of 628 observations (75% of the data) were
using Eq. (3) as follows: used for training purposes and the remaining 209 (25% of the data)
Fig. 7. SEM of some supplementary cementitious materials used in research: (a) silica fume; (b) micro-limestone powder; (c) fly ash; (d) RHA; (e) GGBSF; (f) FC3R; (g) glass
powder; and (h) glass flour.
J. Abellán-García / Construction and Building Materials 256 (2020) 119465 7
Fig. 8. Compressive specimen moulds, cubes and test set-up according to ASTM C109 [41].
set would end up being very small for a neural network validation
(as in our case). In consequence, the validation scores might
change a lot depending on which data points were selected to
use for validation and which were selected for training: that is,
the validation scores might have a high variance regarding the val-
idation division. This would prevent from reliably evaluating of the
model [49].
The best practice in such situations is to utilize k-fold cross-
validation (see Fig. 11). It is based on the division of the available
observations into k partitions, instantiating k identical models,
and training each one on k – 1 partitions while checking the
remaining partition. The validation score for the model employed
is then the average of the k validation scores obtained [49]. In this
research k = 4 was considered.
Fig. 10. Schematic representation of the architecture of a two hidden layer multi- In this research a four-layer perceptron model (i.e. one input,
layer perceptron.
two hidden and one output layers) was developed using R version
3.5.2 (2018-12-20)[54] by using the neuralnet function [55]. The
selection of hidden layers and hidden layer neurons is a trial and
were used for testing of the trained models. Both subsets contained error process which frequently started by selecting a network with
all the possible components in the concrete. a minimum number of hidden layers and hidden neurons [19]. A
As well, the training data could be split into a training and a val- total of 225 different architectures were tested to achieve that
idation set to assess the neural network while keeping adjusting its optimal neural network architecture. In addition, for each of those
parameters (such as the number of neurons in the hidden layers). architectures 100 different models were computed by varying the
Nevertheless, since there are such few data points, the validation initial weight allocation of the neurons. Hence, a total of 22,500
8 J. Abellán-García / Construction and Building Materials 256 (2020) 119465
1:50
artificial networks were developed to predict compressive strength di
of UHPC using the 837 observations. The root mean squared error bij ¼ 1 1 ð6Þ
dj
(RMSE) was used as measure of accuracy.
Seventeen input signals grouped into two different classes were The residual packing density considered for each component is
used in the input layer: depicted in Table 4.
Constituents of UHPC expressed in volume ratio. This group
included cement (C), silica fume (SF), fly ash (FA), ground granu- 4.4. Model performance evaluation
lated blast slag furnace (GGBSF), recycled glass powder (GP), rice
husk ash (RHA), fluid catalytic residue (FC3R), metakaolin (MK), The multi-layer perceptron model was trained on the training
limestone powder (LP), water (W), high range water reducer super- data, through the k-fold validation, and then the chosen model
plasticizer (HRWR), quartz powder (QP) and total aggregate (A) goodness was checked on the test set. In the present work eight
which involved from micro sand to coarse aggregate when used. different statistical parameters have been used to measure the
Relationship between component and properties of UHPC, viz., model’s predictive accuracy, viz., mean absolute error (MAE), aver-
maximum size of aggregate (MSA), water to binder ratio (WB), age error (AE), mean absolute percentage error (MAPE), root mean
water to total powders ratio (WP), and virtual packing density squared error (RMSE), coefficient of efficiency (E), ratio of the root
(VPD) [56–58]. mean square error to the standard deviation of measured data
Virtual packing density VPD is not information frequently pro- (RSR), normalized mean bias error (NMBE) and coefficients of mul-
vided in the scientific papers. In most cases, it had been estimated tiple determination (R2), as indicated in Eqs. (7)–(14) respectively.
based on information provided in the articles such, mean particle 1X n
ai b
size (d50) of components and mixture proportions of concrete. Only MAE ¼ ai ð7Þ
n i¼1
mixtures with enough information to guesstimate the VPD had
been included in the data base. The VPD was guesstimated based
on the compressive packing model theory [56–58]. The general for- 1X n
AE ¼ ai b
ai ð8Þ
mulae that represents the virtual packing density of a granular mix n i¼1
containing n materials classes, when class i is dominant, is pre-
n
sented in Eq. (4). 100 X ai b ai
8 9 MAPE ¼ ð9Þ
< = n i¼1 jai j
bi
VPD ¼ Min h i h i i ¼ 1; ; n
P
:1 i1 1 b þ b b 1 1 y nP ;
a
j¼iþ1 1 b
b
ij i
yj vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
j¼1 i ij i b j uP
un 2
j j
ð4Þ u a b ai
ti¼1 i
where: RMSE ¼ ð10Þ
n
Table 4
Residual packing density considered for each component in the estimation of the VPD [56–58].
*Coarse aggregate.
Table 5
RMSE against number of neurons in the first and second hidden layers.
Table 6
ANN model performance measures.
the evaluation of results has been carried out on the following split UHPC whose largest aggregate is micro sand with maximum
of the dosages database: size aggregate in the range of 150–600 mm.
UHPC with maximum size aggregate in the range of 800–
UPHC pastes with no aggregate 1,200 mm.
J. Abellán-García / Construction and Building Materials 256 (2020) 119465 11
Fig. 16. Regression plot for UHPC with maximum size of aggregate in the range of
Fig. 14. Regression plot for all UHPC dosages. 150–600 mm.
Fig. 17. Regression plot for UHPC with maximum size of aggregate in the range of
800–1200 mm.
Fig. 15. Regression plot for UHPC pastes.
Fig. 18. Regression plot for UHPC with maximum size of aggregate in the range of Fig. 20. Regression plot for UHPC containing coarse aggregate.
1500–3000 mm.
Table 7
Coefficients of multiple determination (R2) for database splits.
In this paper, an accuracy four-layer perceptron was developed The authors declare that they have no known competing finan-
to predict the 28-day compressive strength of UHPC using different cial interests or personal relationships that could have appeared
combinations of supplementary cementitious materials viz., silica to influence the work reported in this paper.
fume, fly ash, ground granulated blast slag furnace, glass powder,
rice husk ash, fluid catalytic cracking residue, metakaolin, and lime- Acknowledgements
stone powder, in addition to quartz powder. Furthermore, different
range of maximum aggregate were considered in the trained four- Special thanks go to APOLO from EAFIT university (Medellín-
layer perceptron. And in order to avoid overfitting, k-fold validation Colombia) and OSIRIS from Escuela Colombiana de Ingenieria Julio
with four partitions was employed, thereby leading to higher confi- Garavito (Bogotá-Colombia) for the servers and computer related
dence of the model when predicting on new data. From the obtained support. Also, to Cementos Argos SA. for donating most of the
results of this investigation, the following conclusions are drawn: materials used in the research described herein and provide the
SEM images of the components. The supply of recycled glass from
1) The proposed model based on a wide range of experimental Cristaleria Peldar SA, FC3R from Ecopetrol SA, and GGBSF from Ger-
and previous work data can be very handy for forecasting dau SA for this research is highly appreciated. The writers would
the compressive strength of UHPC in quick time. It could also like to acknowledge the support and suggestions of Escuela
be helpful in the developing of UHPC as decision support Colombiana de Ingeniería Julio Garavito and Polytechnic University
tool, predicting the compressive strength of a particular of Madrid (UPM).
mix design. This procedure will considerably decrease the
effort, costs and time to design an UHPC dosage for a cus-
tomized strength without performing multiple trials. Data availability
2) The results of the combined use of the performance metrics,
which included MAE, AE, MAPE, RMSE, E, RSR, NMBE, R2, The data used during the present study and the R-code gener-
supplied an unbiased estimate which proved the adequacy ated are available from the corresponding author on reasonable
of the proposed four-layer perceptron model. request.
3) The outcomes of the Olden’s sensitivity CWA analysis sug-
gested consistency with the results of several international References
investigations. This reasserted the validity of the proposed
[1] J. Abellán, J. Fernández, N. Torres, A. Núñez, Statistical optimization of ultra-
model.
high-performance glass concrete, ACI Mater. J. 117 (2020) 243–254, https://
doi.org/10.14359/51720292.
The future work will concentrate on the development of regres- [2] Abellán-García, J., Núñez-López, A., Torres-Castellanos, N., Fernández-Gómez,
sion models for other properties of UHPC such as module of elastic- J., Effect of FC3R on the properties of ultra-high-performance concrete with
recycled glass Efecto del FC3R en las propiedades del concreto de ultra altas
ity, shrinkage, chloride penetration, and carbonization. Another prestaciones con vidrio reciclado, 86 (2019) 84–92. doi:10.15446/dyna.
direction for future research will be the development of regression v86n211.79596.
models for the properties derived from the addition of fibers to the [3] Abellan, J., Torres, N., Núñez, A., Fernández, J., Ultra high preformance fiber
reinforced concrete: state of the art, applications and possibilities into the latin
UHPC such as energy adsorption capacity and ductility. In addition american market, in: XXXVIII Jornadas Sudam. Ing. Estructural, Lima, Peru,
to multi-layer perceptron other artificial intelligence methodolo- 2018.
gies such as bootstrapping, random forest, recurrent networks, [4] Abellan, J., Torres, N., Núñez, A., Fernández, J., Influencia del exponente de
Fuller, la relación agua conglomerante y el contenido en policarboxilato en
among others could be investigated. The results of such studies concretos de muy altas prestaciones, in: IV Congr. Int. Ing. Civ., Havana, Cuba,
would directly benefit the investigation on UHPC. 2018.
[5] J. Zhang, Y. Zhao, Experimental investigation and prediction of compressive
strength of ultra-high performance concrete (UHPC) containing
CRediT authorship contribution statement supplementary cementitious materials, Hindawi Adv. Mater. Sci. Eng. 2017
(2017) 522–525, https://doi.org/10.1155/2017/4563164.
Joaquín Abellán-García: Conceptualization, Methodology, Soft- [6] E. Ghafari, M. Bandarabadi, H. Costa, E. Júlio, Prediction of fresh and hardened
state properties of UHPC: comparative study of statistical mixture design and
ware, Investigation, Data curation, Writing - original draft, Writing an artificial neural network model, J. Mater. Civ. Eng. 27 (2015) 04015017,
- review & editing. https://doi.org/10.1061/(ASCE)MT.1943-5533.0001270.
14 J. Abellán-García / Construction and Building Materials 256 (2020) 119465
[7] ACI. Committe 239, ACI – 239 Committee in Ultra-High Performance Concrete [35] Everitt, B., Hothorn, T., MVA: An Introduction to Applied Multivariate Analysis
2018. with R, (2015).
[8] J. Abellán-García, A. Núñez-López, N. Torres-Castellanos, J. Fernández-Gómez, [36] A. Arizzi, G. Cultrone, Comparing the pozzolanic activity of aerial lime mortars
Factorial design of reactive powder concrete containing electric arc slag made with metakaolin and fluid catalytic cracking catalyst residue: a
furnace and recycled glass powder, Dyna 87 (2020) 42–51, https://doi.org/ petrographic and physical-mechanical study, Constr. Build. Mater. 184
10.15446/dyna.v87n213.82655. (2018) 382–390, https://doi.org/10.1016/j.conbuildmat.2018.07.002.
[9] W. Meng, V.A. Samaranayake, K.H. Khayat, Factorial design and optimization of [37] Torres Castellanos, N., Estudio en estado fresco y endurecido de concretos
UHPC with lightweight sand, ACI Mater. J. (2018), https://doi.org/10.14359/ adicionados con catalizados de craqueo catalítico usado (FCC), Universidad
51700995. Nacional de Colombia, 2014.
[10] Abellán, J., Fernández, J., Torres, N., Núñez, A., Development of cost-efficient [38] Li, Z., Rangaraju, P.R., Development of UHPC Using Ternary Blends of Ultra-
UHPC with local materials in Colombia, in: Proc. Hipermat 2020 – 5th Int. Fine Class F Fly Ash , Meta-kaolin and Portland Cement, (2016) 1–12.
Symp. UHPC Nanotechnol. Constr. Mater., Kassel, Germany, 2020: pp. 97–98. doi:10.21838/uhpc.2016.64.
[11] Viet Thein An, V., Ludwig, H.-M., Proportioning Optimization of UHPC [39] Ferdosian, I., Camões, A., Ribeiro, M., ‘‘High-volume fly ash paste for developing
Containing Rice Husk Ash and Ground Granulated Blast-furnace Slag, in: M. ultra-high performance concrete (UHPC),” Cienc. e Tecnol. Dos Mater. 29
Schmidt, E. Fehling, C. Glotzbach, S. Fröhlich, S. Piotrowski (Eds.), 3rd Int. (2017) e157–e161. doi:10.1016/j.ctmat.2016.10.001.
Symp. UHPC Nanotechnol. Constr. Mater., Kassel Uni, Kassel, Germany, 2012: [40] Ghafari, E., Costa, H., Júlio, E., Portugal, A., Durães, L., Enhanced Durability of
pp. 197–205. Ultra High Performance Concrete by Incorporating Supplementary
[12] W. Li, Z. Huang, T. Zu, C. Shi, W.H. Duan, S.P. Shah, Influence of nanolimestone Cementitious Materials, Second Int. Conf. Microestructural-Related Durab.
on the hydration, mechanical strength, and autogenous shrinkage of ultrahigh- Cem. Compos. (2012) 11–13.
performance concrete, J. Mater. Civ. Eng. 28 (2016) 1–9, https://doi.org/ [41] ASTM, ‘‘Standard Test Method for Compressive Strength of Hydraulic Cement
10.1061/(ASCE)MT.1943-5533.0001327. Mortars (Using 2-in . or [ 50-mm ] Cube Specimens),” Am. Soc. Test. Mater. C-
[13] Huang, Z., Cao, F., ‘‘Effects of Nano-materials on the Performance of UHPC,” 材 109/C109M. (2010) 1–9. doi:10.1520/C0109.
料导报B:研究篇. 26 (2012) 136–141. [42] Funk, J.E., Dinger, D.R., ‘‘Predictive Process Control of Crowded Particulate
[14] N.A. Soliman, A. Tagnit-Hamou, Partial substitution of silica fume with fine Suspensions. Applied to Ceramic Manufacturing,” Springer Science, New York,
glass powder in UHPC: Filling the micro gap, Constr. Build. Mater. 139 (2017) 1994. doi:10.1007/978-1-4615-3118-0.
374–383, https://doi.org/10.1016/j.conbuildmat.2017.02.084. [43] F. Rosenblatt, The perceptron: a probabilistic model for information storage
[15] Tagnit-Hamou, A., Soliman, N., Omran, A., ‘‘Green Ultra - High - Performance and orgnization in the brain, Cornerr Aeronaut. Lab. 65 (1958) 386–408.
Glass Concrete,” First Int. Interact. Symp. UHPC – 2016. (2016). [44] Ghafari, E., Al., Optimization of UHPC by Adding Nanomaterials, in Proceedings
[16] N.A. Soliman, A. Tagnit-Hamou, Using glass sand as an alternative for quartz of Hipermat 2012, in: 3rd Int. Symp. UHPC Nanotechnol. Constr. Mater., Kassel
sand in UHPC, Constr. Build. Mater. 145 (2017) 243–252, https://doi.org/ Uni, Kassel, Alemania, 2012: pp. 71–78.
10.1016/j.conbuildmat.2017.03.187. [45] Estebon, M.D., Perceptrons : An Associative Learning Network by, Virginia
[17] Camacho, E., López, J.Á., Serna, P., ‘‘Definition of three levels of performance for Tech. (1997).
UHPFRC-VHPFRC with available materials, in Proceedings of Hipermat 2012,” [46] S. Gupta, Using Artificial Neural Network to Predict the Compressive Strength
in: M. Schmidt, E. Fehling, C. Glotzbach, S. Fröhlich, S. Piotrowski (Eds.), 3rd Int. of Concrete containing Nano-silica, Civ. Eng. Archit. 1 (2013) 96–102, https://
Symp. UHPC Nanotechnol. Constr. Mater., Kassel Uni, Kassel, Germany, 2012: doi.org/10.13189/cea.2013.010306.
pp. 249–256. [47] J.A. Anderson, Cognitive and psychological computation with neural models,
[18] Camacho Torregrosa, E., ‘‘Dosage optimization and bolted connections for IEEE Trans. Syst. Man. Cybern. 13 (1983) 799–816.
UHPFRC ties,” Polytechnic University of Valencia, Spain, 2013. [48] S.D. Bharathi, R. Manju, J. Premalatha, Prediction of compressive strength for
[19] V. Chandwani, V. Agrawal, R. Nagar, Modeling slump of ready mix concrete self- compacting concrete (SCC) using artificial intelligence and regression
using genetic algorithms assisted training of Artificial Neural Networks, Expert analysis, Int. J. ChemTech Res. 10 (2017) 263–275.
Syst. Appl. 42 (2015) 885–893, https://doi.org/10.1016/j.eswa.2014.08.048. [49] Chollet, F., Allaire, J.J., Deep Learning with R, Manning Publications Co., New
[20] A. Khashman, P. Akpinar, Science direct non-destructive prediction of concrete Jersey, 2018
compressive strength using neural networks prediction of concrete [50] Demir, F., Prediction of elastic modulus of normal and high strength concrete
compressive strength using neural networks, Procedia Comput. Sci. 108 by artificial neural networks, 22 (2008) 1428–1435. doi:10.1016/
(2017) 2358–2362, https://doi.org/10.1016/j.procs.2017.05.039. j.conbuildmat.2007.04.004.
[21] H. Adeli, Neural networks in civil engineering: 1989–2000, Comput. Civ. [51] Rumelhart, D., Hinton, G., Williams, R., Learning internal representations by
Infrastruct. Eng. 16 (2001) 126–142. error propagation, in: D. Rumelhart, J. McClelland (Eds.), Parallel Distrib.
[22] M. Aderaw, S. Muse, Z.C. Abiero, Artificial neural network based modelling Process. Explor. Microstruct. Cogn., 1986: pp. 318–362.
approach for strength prediction of concrete incorporating agricultural and [52] H.M. Mushgil, H.A. Alani, L.E. George, Comparison between resilient and
construction wastes, Constr. Build. Mater. 190 (2018) 517–525, https://doi. standard back propagation algorithms efficiency in pattern recognition, Int. J.
org/10.1016/j.conbuildmat.2018.09.097. Sci. Eng. Res. 6 (2015) 773–778.
[23] V. Chandwani, R. Nagar, Applications of artificial neural networks in modeling [53] Prasad, N., Singh, R., Lal, S.P., Comparison of back propagation and resilient
compressive strength of concrete: a state of the art review, Int. J. Curr. Eng. propagation algorithm for spam classification, Proc. Int. Conf. Comput. Intell.
Technol. 4 (2014) 2949–2956. Model. Simul. (2013) 29–34. doi:10.1109/CIMSim.2013.14.
[24] Taghaddos, H., Mahmoudzadeh, F., Pourmoghaddam, A., Shekarchizadeh, M., [54] R. Core Team R: A Language and Environment for Statistical Computing” 2018
Prediction of compressive strength behaviour in RPC with applying an Vienna, Austria.
Adaptive Network-Based Fuzzy Interface System, in: Proc. Int. Symp. Ultra [55] Günther, F., Fritsch, S., neuralnet: Training of Neural Networks, R J. 2 (2010)
High Perform. Concr., Kassel, Alemania, 2004. 30–38. doi:10.1109/SP.2010.25.
[25] M. Schmidt, E. Fehling, C. Geisenhanslükem (Eds.), Proceedings of the [56] F. de Larrard, T. Sedran, Optimization of ultra-high-performance concrete by
International Symposium on Ultra High Performance Concrete, University of the use of a packing model, Cem. Concr. Res. 24 (1994) 997–1009, https://doi.
Kassel, Germany, Kassel, 2004. org/10.1016/0008-8846(94)90022-1.
[26] Fehling, E., Schmidt, C., Stüwald, S., eds., Proceedings of the Second [57] De Larrard, F., Concrete mixture proportioning: a scientific approach., in: Mod.
International Symposium on Ultra High Performance Concrete, (2008). Concr. Technol. Ser., E&FN SPON, 1999.
[27] Schmidt, M., Fehling, E., Glotzbach, C., Fröhlich, S., Piotrowski, S., eds., [58] F. De Larrard, T. Sedran, Mixture-proportioning of high-performance concrete,
Proceedings of Hipermat 2012 3rd International Symposium on UHPC and Cem. Concr. Res. 32 (2002) 1699–1704, https://doi.org/10.1016/S0008-8846
Nanotechnology for High Performance Construction Materials, (2012). (02)00861-X.
[28] Fehling, E., Middendorf, B., Thiemicke, J., eds., Proceedings of Hipermat 2016 [59] D.N. Moriasi, J.G. Arnold, M.W. Van Liew, R.L. Bingner, R.D. Harmel, T.L. Veith,
4th International Symposium on UHPC and Nanotechnology for High Model evaluation guidelines for systematic quantification of accuracy in
Performance Construction Materials, University of Kassel, Germany, Kassel, watershed simulations, Am. Soc. Agric. Biol. Eng. 50 (2007) 885–900.
Germany, 2016. [60] Nash, E., Sutcliffe, V., RIVER FLOW FORECASTING THROUGH CONCEPTUAL
[29] Camacho Torregrosa, E., Dosage optimization and bolted connections for MODELS. PART I- A DISCUSSION OF PRINCIPLES *, J. Hydrol. 10 (1970) 282–290.
UHPFRC ties, 2013. [61] S. Srinivasulu, A. Jain, A comparative analysis of training methods for artificial
[30] Skazlic, M., Bjegovic, D., Serdar, M., Influence of test specimens geometry on neural network rainfall – runoff models, Appl. Soft Comput. 6 (2006) 295–306,
compressive strength of ultra-high performance concrete., in: Proc. 2nd Int. https://doi.org/10.1016/j.asoc.2005.02.002.
Symp. Ultra High Perform. Concr. Kassel, Ger., 2008: pp. 295–301. [62] J.D. Olden, M.K. Joy, R.G. Death, An accurate comparison of methods for
[31] Naaman, A.E., Wille, K., Some correlation between hihg packing density, ultra- quantifying variable importance in artificial neural networks using simulated
high performance, flow hability, and fiber reinforcement of a concrete matrix, data, Ecol. Modell. 178 (2004) 389–397, https://doi.org/10.1016/j.
in: Bac2010, Congr. Ibêrico Sobre Betão Auto-Compactável, 2010. ecolmodel.2004.03.013.
[32] Graybeal, B., Davis, M., Cylinder or Cube : Strength Testing of 80 to 200 Fiber- [63] E. Ghafari, H. Costa, E. Nuno, B. Santos, RSM-based model to predict the
Reinforced Concrete Cylinder or Cube : Strength Testing of 80 to 200 MPa (11 . performance of self-compacting UHPC reinforced with hybrid steel micro-
6 to 29 ksi) Ultra-High-Performance Fiber-Reinforced Concrete, ACI Mater. J. fibers, Constr. Build. Mater. 66 (2014) 375–383, https://doi.org/10.1016/
105 (2014) 603–609. j.conbuildmat.2014.05.064.
[33] A. Atkinson, M. Riani, Robust Diagnostic Regression Analysis, Springer, US, [64] S. Abbas, M.L. Nehdi, M.A. Saleem, Ultra-High Performance Concrete:
New York, 2000. Mechanical Performance, durability, Sustainability and Implementation
[34] W.K. Härdle, L. Simar, Applied Multivariate Statistical Analysis, Springer- Challenges, Int. J. Concr. Struct. Mater. 10 (2016) 271–295, https://doi.org/
Verlag GmbH, Berlin, 2012. 10.1007/s40069-016-0157-4.
J. Abellán-García / Construction and Building Materials 256 (2020) 119465 15
[65] Schmidt, C., Schmidt, M., ‘‘Whitetopping of Asphalt and Concrete Pavements [66] Zhang, J., Zhao, Y., Prediction of compressive strength of ultra-high
with thin layers of Ultra-High-Performance Concrete - Construction and performance concrete (UHPC) containing supplementary cementitious
economic efficiency,” in: Proc. Hipermat 2012 - 3rd Int. Symp. UHPC materials, Proc. - 2017 Int. Conf. Smart Grid Electr. Autom. ICSGEA 2017.
Nanotechnol. Constr. Mater., Kassel, Germany, 2012. 2017-Janua (2017) 522–525. doi:10.1109/ICSGEA.2017.150.