Proc. Int. Symp.

,,Brittle Matrix Composites 10"

A.MBrandt, J. Olek, MA. Glinicki, C.K. Y.Leung, eds.
Warsaw, October 15-17, 2012
IFTR and Woodhead Pub!., Warsaw 2012

DESIGN OF UHPC USING ARTIFICIAL NEURAL NETWORKS

E. GHAFARI (ll, M. BANDARABADI (2), H. COSTA <3l, E. JULIO <4l

(1) ICIST & Dep. of Civil Engineering, University of Coimbra, Portugal
(2) CISUC, Dep. of Informatics Engineering, University of Coimbra, Portugal
(3) ICIST & Polytechnic Institute of Coimbra - ISEC, Coimbra, Portugal
(4) ICIST& DECivil IST, Technical University of Lisbon, Portugal

ABSTRACT

Ultra-high performance concrete (UHPC) results from the mixture of several constituents giving rise to a highly
complex material in hardened state. The higher number of constituents in relation to current concrete, together
with a higher number of possible combinations and relative proportioning, makes the behavior of this type of
concrete more difficult to predict. Until now, most of the proposed mixture design methods are based on a trial
and error procedure, which is expensive and work intensive. Moreover, these methods are not efficient in
predicting neither the consistency in fresh state nor the strength in hardened state, and do not consider the effect
of curing on the latter.
The main objective of the research study herein described is to build an analytical model, based on artificial
neural networks (ANN), to predict the required performance of UHPC. Specifically, back-propagation neural
networks (BPNN) are adopted to model the relation between the input and the output parameters of UHPC, for
two different curing conditions, including heat treatment and water storage. In order to train the neural network,
a total set of 53 different mixtures were designed. It is concluded that the developed model can be used as a
reliable method to predict the performance ofUHPC.

Keywords
Ultra-high performance concrete, mixture design, artificial neural networks

INTRODUCTION

Ultra-high performance concrete (UHPC) results from the mixture of several constituents
giving rise to a highly complex material in hardened state. The higher number of constituents
in relation to current concrete, together with a higher number of possible combinations and
relative proportioning, makes the behavior of this type of concrete more difficult to predict.
Dewar et al. [I] presented a model for predicting the packing density of multi-sized
aggregate from the grading of the aggregate. However, it has been proved that applying the
hypothetical models developed for aggregate particles is not efficient. The same conclusion
was drawn relatively to model the cementitious materials due to differences in packing
behavior [2]. De Larrard [3] has also presented a packing model called "compressible
packing model", to determine the packing density achieved by a granular mixture. Fennis et
al. [4] explained that with small particles, the packing density is influenced by the interaction
caused by surface forces, such as van der Waals forces, electrical charges and steric forces;
therefore, for very small particles, particle packing predictions by CPM deviate considerably
from experiments data and, thus, the current CPM model cannot accurately predict the
behavior of the mixtures containing micro- and nano-sized particle. Fennis et al.[4] presented
62 E. GHAFARI, M. BANDARABADI, H. COSTA and E. JULIO

a new method based on particle density of powder using centrifugal procedure. They
indicated that by combining cement and fillers in a smart way, for instance by replacing
cement with fly ash, the packing density of a concrete mixture can be increased and the water
demand of concrete reduced. Park et al. [5] performed a fundamental study in order to
optimize mix proportion of UHPC by considering different variables, such as water to
powder ratio and different percentage of supplementary cementitious materials. Schmidt et al.
[6] presented a method for modeling and for calculation of high density packing for cement
and fillers in UHPC, which was optimized for compressive strength higher than 200 N/mm2 •
Recently, Wille et al. [7] presented a simple mixture design of UHPC according to
paste rheologhical properties, in which high spread values of paste, obtained from mini-slump
measurement, imply high packing density and, thus, high compressive strength.
In spite of all of the aforementioned studies, which aimed to optimize the mixture
proportion and confirming high-strength mixes of ultra-high performance concrete, most of
the proposed mixture design methods have involved an extensive series of tests. Therefore,
these procedures are expensive and work intensive, and results are often only applicable to a
limit range of materials. Moreover, these methods are not efficient in predicting neither the
consistency in fresh state nor the strength in hardened state, and do not consider the effect of
curing on the latter.
The main objective of the research study herein described is to build an analytical
model, based on artificial neural networks (ANN), to predict the required performance of
UHPC.

EXPRIMENTAL PROGRAM

The UHPC mixtures were prepared with the following main constituents: ordinary Portland
cement type I 52.5 R; silica fume (SF); quartz flour (P600), used as a micro filler (particle
sizes less than 10 µm); siliceous sand, with maximum aggregate size of 0.6 mm;
polycarboxylate ether based superplasticizers (SP) and; micro steel fibers, with 10 mm length
and 0.15 mm diameter.
In order to investigate the effect of curing, half of the specimens were cured in water, at
20 °C, and the rest were cured at 90 °C and 95 RH, for 48 hours, including one hour ramp up
and down and, finally, cooled in ambient air. After the mixing was completed, tests were
conducted immediately, still on fresh state, to measure the consistency of the UHPC. The
consistency of the UHPC was measured according to a flow table test, according to ASTM
Cl437 [8]. In this test, the mini slump cone is filled, and then removed, allowing the concrete
to flow outward. Once the concrete reaches a steady state, the flow table is dropped 20 times
in approximately 20 seconds. Again, the concrete is allowed to settle, and then its average
diameter is recorded. The axial compressive strength was performed at 28 days of age, using
cubic (40x40x40 mm3) specimens.

ARTIFICIAL NEURAL NETWORK ARCHITECTURE

Artificial neural networks (ANNs) are data processing paradigm which is inspired by
biological nervous systems. They are composed of large number of highly interconnected
processing elements (neurons) working together. Rosenblatt [9] devised a machine called the
perceptron that operated in the same way as the human mind. Percepton is the most popular
neural network used in modeling that has an input layer and an output layer as well as a
number of hidden layers which are fully connected to each other, [10].
Design of UHPC using artificial neural networks 63

Fig. 1 shows a typical diagram for an artificial neurons structure which is composed of
five major parts: inputs, weights, sum function, activation function and outputs. The Xi=(X1,
X2 ... Xn) represent number of input link which are applied to each neuron, Wi= (W 1, W 2...
Wn) represent the weight for each of the inputs. Sum function computes the net input that
approaches to a neuron, [ 11]. The weighted sums of the input components are calculated by
using Eq. (1) as follows;

(1)

where ne~ is the weighted sum of the /11 neuron for the input received from the preceding
layer with n neurons, Wii is the weight between the jth neuron and the ith neuron in the
preceding layer; Xi is the output of the ith neuron in the preceding layer and b is a fix value as
internal addition, [12].
Activation function is defined as a function that processes the net input obtained from
sum function and determines the neuron output. There are different kinds of activation
function, such as Tansig, Logsig, and Purelin, among others [13]; however, the most popular
activation function is a binary sigmoid function that is applied in this study. As it is shown in
Fig 1, the sigmoid function takes the input, which can have any value between minus and
plus infinite, and squashes the output into the range 0 to 1[14]. The output of the /h neuron
ou~ is computed by using Eq. (2) as follows:

(2)

Where K is a constant used to control the slope of the semi linear region.

l»g-sigmoid
Activation function

Input

Fig 1. A schematic diagram of artificial neurons structure

TRAINING AND TESTING OF ANN MODEL

The fully connected perceptron which is constitute of an input and output layer of nodes
is designed with an associated weight assigned to it to give an output as a result of a given
64 E. GHAFARl, M. BANDARABADI, H. COSTA and E. JULlO

input. Adjusting these weights to produce a particular output is called ''training'' of the
network which is the mechanism that allows the network to learn [15].
Among various types of ANNs learning algorithms, the multi-layer perceptron (MLP)
is the most popular one, in which one or more layers called "hidden layers" exist between the
input and the output layers. Multiple layers of neurons with nonlinear transfer functions allow
the network to learn nonlinear relationships between input and output vectors [14]. Rumelhart
et al. [16] derived a learning algorithm for perceptron networks with constituted hidden units.
Their learning algorithm is called back-propagation (BP) that is the most commonly used
learning algorithm.
The process of this learning algorithm consists of two main steps: in the first step, after
generating a forward flow of signal from the input layer to the output layer, the result of the
output layer, outj, is compared with the target (experimental) value, and the error of each
output neuron is then determined from the difference between the computed values and the
target (experimental) values and the resulted error will be propagated backward the network's
weight to minimize the overall error. The second stage involves the readjustment of the
weights and biases in the hidden and output layers to reduce the difference between the
computed and desired outputs. The adjustment of the weights is performed using a
"generalized delta rule" through the gradient descent on the error. The training process is
performed iteratively until the average sum-squared errors between the actual and the desired
outputs are minimized. Fig. 2 is shown a typical architecture of a multilayer perceptron neural
network with back propagation, an input layer, a hidden layer and an output layer [12].
In order to simplify the learning process of the back-propagation neural network and to
reduce the time required for training, the learning algorithm adopted to train the network
model in this study is the Levenberg-Marquardt algorithm (LMBP). It is often the fastest
available back-propagation algorithm, and is highly recommended as a first-choice
supervised algorithm, although it requires more memory than other algorithms [ 17]. In this
study a multilayer feedforward neural network (LMBP) with back propagation training
algorithm was employed.

!n1nn
faye-r

Fig 2. The architecture of a multilayer perceptron neural network with back propagation
Design ofUHPC using artificial neural networks 65

RES UL TS AND DISCUSSION

A neural network model was conducted, trained and tested using these available test data. A
data set including 212 data samples for each type of curing, obtained from experimental tests
were used to construct a desired artificial neural networks. One of the common problems
during the training set is overfitting that may occur when a network has memorized the
training set but has not learned to generalize to new inputs. It gives a relatively small value of
error on the training set but a much larger error when new data is presented to the network
[14].
Early stopping feature is a widely used and effective way to overcome overfitting
problem, in which, the data sets is divided into three subsets, including training set, validation
set and testing sets. In this procedure, the error on the validation set is monitored during the
training process and as soon as the validation error increases for a specified number of
iterations, the training is stopped and the weights and biases at the minimum of the validation
error are returned [ 14].
In this study, 80% of the total patterns were used for training and the other 20% were
used for validation and testing the performance of the network. It should be mentioned that
the division of the data was done randomly between the two sets. The ranges of input
parameters are shown in Table 1. The range of data set for compressive strength of ultra-high
performance concrete at 28 days and slump flow varies from 98 to 220 MPa and from 170 to
240 mm respectively.
The structure of network used in this study has seven neurons in input layer (variables);
one hidden layer with fifteen neurons was used in the architecture because of its minimum
root mean square error (RMSE) values for training and testing sets, and three neurons in
output layer as shown in Fig 3. The inputs are cement, sand, silica fumes, quartz flour, water,
superplasticizer and micro steel fiber and the outputs are compressive strength ofUHPC at 28
days in both condition of steam curing (FsT) and water storage (Fwy) and also slump flow
(SL).
There is no exact method to determine the optimal number of neurons in the hidden
layer neither in the number of hidden layers. Therefore, a total series of 25 architecture
network with different number of neurons in hidden layer were studied. In this study, the best
network configuration was selected based on two main criteria; the minimum RMSE error
and the highest regression coefficient (R2).

Table 1. The ranges of in~ut parameters

Cement 620 1070

Sand 560 1245
Silica fume 105 305
Quartz flour 0 300
Water 160 220
Superplasticizer 28 40
Steel fiber 0 155

Fig 4 shows the relation of the RMSE error in both training set and testing set different
number of hidden units after 100 iterations. As it can be seen, hidden layer with fifteen
neurons gave the minimum RMSE values for training and testing sets.
The prediction performances of training and test sets of the proposed ANN model are
presented in Fig 4-6. A comparison of ANN model predictions and experimental data value
obtained for Fsy, FWT and SL versus predicted compressive strengths by ANN model with R2
66 E. GHAFARI, M. BANDARABADI, H. COSTA and E. JULIO

coefficients. It can be seen that predicted and measured values, for all three outputs are well
correlated. The obtained correlation coefficients for FsT, FWT and SL in training sets are R 2=
0.99, R2= 0.98 and R2= 0.96 respectively.
Hiddens

Ce1nent

Sand
--
Silica Fmue
-- FsT

--
Quartz Flour FwT

"'Nater SL

Superplacticizen-

Steel Fiber
-
Fig. 3. The structure of ANN model

0.12

-9-TrainingSets ....-Tcsting Sets

0.1

0.08

~
~ 0.06

0.04

0.02

10 15 20 25 30
NumbcrofNodes in Hidden Layer
Fig. 4. The RMSE error vs. number of hidden units in both training and testing set
High values of correlation coefficient implying indicated training efficiency of the
proposed ANN model in learning the relation between different variables and targets.
Design of UHPC using artificial neural networks 67

220

200 R2 :=0.98

180

:lj 160

:E
zz
14<1
<

..... 120

+Training Set •Test Set

80
140 160 um 80 100 120 140 160 180 200 220
ExprimentalData Expriment.alDat.a

Fig. 5. The prediction performance of ANN model for FsT: (a) training set (b) testing set

R'=0.98
.
...""..
.. ·.
+Training Set +Testing Set

140
ExprimentalData "'
ExprimentalData

Fig. 6. The prediction performance of ANN model for FwT: (a) training set (b) testing set

240

R'=0.96 R"=0.97

,_

....... ..
•i J "'
..
~

"'
16'J
•Training Set •Training Set
1..

140 '································································································································ 140'

170 200 116 220 170 180 190 200 210 220
ExprimentalData ExprimentalData

Fig. 7. The prediction performance of ANN model for SL: (a) training set (b) testing set

Recalling the early stopping procedure, the performance of ANN model was studied in
the test sets. Fig Sb, 6b and 7b show the results of testing phase in which regression
coefficients of 0.98, 0.98 and 0.97 were achieved between the experimental and ANN model
outputs. The results imply the capability of the proposed ANN model to generalize between
input and the output variables by providing reasonable prediction when presented with the
new data sets.
68 E. GHAFARI, M. BANDARABADI, H. COSTA and E. JULIO

The correlation coefficient and RMS values in training set and test sets for all three
outputs are presented in Table 2. All of the statistical values proved that the proposed ANN
model can be used as an efficient tool to predict the strength and workability of UHPC.

Table 2. Evaluation of statistic parameters of ANN Model

Name of Statistic parameters of ANN Model
Function Training Set Testing Set
RMS R2 RMS R2
0.015 0.99 0.025 0.98
0.025 0.98 0.020 0.98
0.021 0.96 0.026 0.97

CONCLUSION

In this research, the viability of using an ANN model to predict both the compressive strength
and slump flow of UHPC was studied. From results, it can be concluded that the proposed
ANN model can be used as an efficient method for predicting the performance ofUHPC. The
architecture of ANN model was optimized with fifteen neurons in hidden layer, in which the
high value of correlation coefficient along with minimum value of RMSE were achieved. It
has also been found that the ANNs method can save more computational effort in comparison
with conventional method.

ACKNOWLEDGEMENTS

The authors thank the financial support of the Portuguese Science and Technology
Foundation (FCT) for the project PTDC/ECM/098497/2008 entitled "Intelligent Super Skin
(ISS)".

REFERENCES

l. Dewar, J.D., Computer Modelling of Concrete Mixtures. in E&FN Spon. 1999.

London
2. Kwan A.K.H., M.C.F., "Effects of various shape parameters on packing of aggregate
particles", Magazine of Concrete Research, 2001. vol. 53, 2, 91-100
3. De Larrard, F., Concrete Mixture Proportioning: A Scientific Approach. in E&FN
Spon, 1999 London
4. Sonja A.A.M. Fennis, J.C.W., Joop A. den Uijl The use of particle packing models to
design ecological concrete HERON, vol. 54, 2009
5. Jung Park, S.t.K., Koh Kyung Taek, Kim Sung wook. Influence of the ingredients on
the compressive strenght of UHPC as a funamental study to optimize the mixing
proportion. in International Symposium on Ultra High Performance Concrete
(UHPC). 2008. Kassel
6. Teichmann, T., M.S. Influence of the packing density of fine particles on structure,
strength and durability of UHPC in International Symposium on Ultra High
Performance Concrete September 13-15, 2004 Kassel, Germany
Design ofUHPC using artificial neural networks 69

7. Kay Wille, A.E.N., Parra-Montesinos, G.J., Ultra-High Performance Concrete with

Compressive Strength Exceeding 150 MPa (22 ksi): A Simpler Way. ACI Materials
Journal,January-February2011, 108, 1
8. Practice, A.S.f.T.a.M.S., ASTM Cl437 Standard Test Method for Flow of Hydraulic
Cement Mortar, Philadelphia, PA, 2001
9. Rosenblatt, F., The Perceptron: A Probabilistic Model for Information Storage and
Organization in the Brain,. Cornell Aeronautical Laboratory, Psychological Review,
1958,65,6,386-408
10. Jou, I.C., S.-S. You, L.-W. Chang, Analysis of hidden nodes for multi-layer
perceptron neural networks. Pattern Recognition, 1994, 27, 6, 859-864
11. Cybern, A.J., Cognitive and psychological computation with neural models. IEEE
Trans System, 1983, 5, 799-814
12. Liu SW, H.J., Sung JC, Lee CC., Detection of cracks using neural networks and
computational mechanics. Comput Meth Appl Mech Eng 2002. 191, 2831-2845
13. Hagan MT, D.H., Beale M. , Neural network design. Singapore: Thompson Asia
Private Limited, 1996
14. Mark Hudson Beale, M.T.H., Howard B.Demuth, Neural Network Toolbox User's
Guide. March 2012,version 7.0.3-The MathWorks Inc., USA
15. Estebon, M.D., Perceptrons: An Associative Leaming Network. Virginia Tech, Spring
1997, cs 3604
16. Rumelhart DE H.G., William RJ Leaming internal representation by error propa-
gation .. Proceeding parallel distributed processing foundation, 1986. 1, Cambridge-
MIT Press
17. Fischer, A. P.K. Shukla, A Levenberg-Marquardt algorithm for unconstrained
multicriteria optimization. Operations Research Letters, 2008, 36, 5, 643-646