Sei sulla pagina 1di 9

ACI MATERIALS JOURNAL TECHNICAL PAPER

Title no. 100-M20

Artificial Intelligence Model for Flowable Concrete Mixtures


Used in Underwater Construction and Repair
by Hassan El-Chabib, Moncef Nehdi, and Mohammed Sonebi

This study explores using artificial neural networks to predict the AWA and high-range water reducer (HRWR) directly affect,
rheological and mechanical properties of underwater concrete with different magnitudes, the stability of highly flowable
(UWC) mixtures and to evaluate the sensitivity of such properties UWC.6-8
to variations in mixture ingredients. Artificial neural networks
(ANN) mimic the structure and operation of biological neurons and Based on results of existing experimental studies, the
have the unique ability of self-learning, mapping, and functional effects of individual ingredients on the behavior of UWC
approximation. Details of the development of the proposed neural mixtures are generally well-known and defined. Some of
network model, its architecture, training, and validation are these ingredients have dual and sometimes conflicting effects
presented in this study. A database incorporating 175 UWC mixtures on the mixture’s performance, however, hence requiring the
from nine different studies was developed to train and test the ANN trial of several batches before achieving simultaneously
model. The data are arranged in a patterned format. Each pattern acceptable key mixture properties. In addition, some inter-
contains an input vector that includes quantity values of the mixture national standards have set guidelines and recommendations
variables influencing the behavior of UWC mixtures (that is, for the proportioning of concrete mixtures intended for casting
cement, silica fume, fly ash, slag, water, coarse and fine aggregates,
and chemical admixtures) and a corresponding output vector that
underwater. For example, the Japan Society of Civil Engineers
includes the rheological or mechanical property to be modeled. (JSCE)9 recommends that the w/b of concrete mixtures
Results show that the ANN model thus developed is not only capable designed for casting underwater reinforced concrete struc-
of accurately predicting the slump, slump-flow, washout resistance, tures be limited to less than 0.5 and 0.55 in seawater and
and compressive strength of underwater concrete mixtures used in the freshwater, respectively. The JSCE standard also requires
training process, but it can also effectively predict the aforemen- that specimens cast underwater should develop compressive
tioned properties for new mixtures designed within the practical strengths greater than 80% of similar specimens cast and
range of the input parameters used in the training process with an cured above water at any given age. Standards fall short,
absolute error of 4.6, 10.6, 10.6, and 4.4%, respectively. however, in defining the overall relationship between mixture
proportions and engineering properties.
Keywords: compressive strength; slump; underwater concrete.
Because of the conflicting effects of some of the mixture
parameters and the absence of a theoretical relationship
INTRODUCTION between mixture proportions and measured engineering
The growing need to simplify the complex and costly properties, the effect of mixture design on the rheological
operation of repairing aging underwater concrete (UWC) behavior and in-place concrete properties are often described
structures, such as marine piles and hydraulic canals and using regression analysis tools and traditional statistical
constructing new underwater structures such as cofferdam models.10 These methods are generally based on assumed
bases or bridge foundations, has lead to a growing interest in relationships and do not have a true predictive capability
developing a special type of high-performance concrete for outside the experimental domain used for their development.
underwater repair and construction. The objective of this study is to investigate the feasibility of
Concrete intended for underwater construction and place- using artificial neural networks as an effective tool to predict
ment should have the ability to flow readily into place, fill in the rheological behavior and mechanical performance of
complex and narrow formwork, consolidate itself with minimal UWC mixtures. Unlike traditional parametric methods, the
or no external vibration, and, most importantly, exhibit a ANN method does not need to assume a relationship between
good resistance to segregation and water dilution to develop mixture ingredients and concrete performance. This relation-
adequate in-place mechanical properties. The successful ship is rather generated from the experimental data provided
development of such concrete must ensure a good balance to build and train the model.
between deformability and stability, hence requiring a better
understanding of the effect of the mixture variables on its RESEARCH SIGNIFICANCE
performance. It often necessitates the manipulation of such
There is currently no rational and reliable method that can
variables to ensure acceptable rheological behavior, adequate
predict the performance of flowable UWC mixtures or quanti-
structural performance, and long-term durability. Previous
tatively relate their mixture proportions to expected rheological
research has indicated that sound and adequate in-place
properties of underwater-cast concrete are directly related to and mechanical characteristics. This study investigates the
the enhancement in washout (dilution) resistance.1-3 It was
argued that the incorporation of anti-washout admixtures ACI Materials Journal, V. 100, No. 2, March-April 2003.
MS No. 02-203 received June 5, 2002, and reviewed under Institute publication
(AWA) could significantly enhance the resistance of concrete policies. Copyright © 2003, American Concrete Institute. All rights reserved, including
to washout and segregation,4,5 and that the composition of the making of copies unless permission is obtained from the copyright proprietors.
Pertinent discussion will be published in the January-February 2004 ACI Materials
the binder, water-binder ratio (w/b), and the dosage of both Journal if received by October 1, 2003.

ACI Materials Journal/March-April 2003 165


feasibility of using artificial neural networks to create an
Hassan El-Chabib is a doctoral candidate in the Department of Civil and Environmental
Engineering, University of Western Ontario (UWO). His research interests include intelligent model that can be used to predict the properties of
self-consolidating concrete and use of artificial neural networks in modeling the UWC and assist in designing UWC mixtures with tailor-made
behavior of construction materials. properties without the need for a large number of trial mixtures.
ACI member Moncef Nehdi is Assistant Professor in the Department of Civil and
Environmental Engineering at UWO. He is a member of ACI Committees 225, Hydraulic ANN-BASED MODEL FOR UWC
Cements; 236, Material Science of Concrete; and E 803, Faculty Network Coordinating
Committee. He is a recipient of the ACI Young Member Award for Professional Artificial neural networks are powerful computational tools
Achievement. His research interests include modeling the behavior of cement- capable of self-organizing, pattern recognition, and functional
based materials, recycling by-products in construction, use of special concretes in approximation. The architecture of ANNs mimics that of
geotechnical engineering, and durability and repair of concrete infrastructure.
biological neurons and their operation essentially simulates
ACI member Mohammed Sonebi is a lecturer at the Advanced Concrete and the internal operation of the human brain. Because substantial
Masonry Centre, University of Paisley, Scotland, UK. He is a member of ACI literature has explained in detail the terminology of artificial
Committee 236, Material Science of Concrete. He received his MEng and PhD from
the Université de Sherbrooke, Canada. His research interests include self-compacting con- neural networks, this aspect is not addressed herein and
crete, underwater concrete, high-performance concrete, and rheology. interested readers are referred to the appropriate literature.11-13
The important aspects of building a successful neural network
model for UWC, however, are discussed as follows.
Recently, there has been a growing interest in using artificial
neural networks in engineering applications including the
modeling of concrete properties. For example, ANN was
used to predict carbonation depth in concrete structures14
and, to a limited extent, to model concrete properties as a
function of mixture composition.15,16 Although successfully
Fig. 1—Simplified model of artificial neuron. trained ANNs have proved to be very effective tools in iden-
tifying complex nonlinear relationships and in modeling
various applications14-17 in cement and concrete research, its
effectiveness depends mainly on the quality of the database
used for its training. The basic methodology for developing
an effective ANN model is to train a neural network to learn
the relationship between a set of inputs and corresponding
outputs. In other words, the network should be trained using
reliable experimental data that is comprehensive enough to
capture all essential aspects that might influence the input/
output relationship, and large enough to effectively cover the
practical range of all input parameters. A well-trained neural
network should have the capability of accurately predicting
an output when it is presented with a new set of inputs from
within the practical range of the training data.
In this study, a feed-forward back-propagation neural
network is used to predict the engineering properties of
underwater concrete mixtures and to evaluate their sensitivity
(a) to variations in mixture ingredients. Feed-forward ANNs are
the most commonly used neural networks in materials
modeling and they consist of an input layer, an output layer,
and a number of hidden layers. Each layer includes a number
of neurons (processing units) partially or fully connected to
neurons in the adjacent layers, and no connection is permitted
between neurons within the same layer. Each neuron in the
hidden and output layers receives multiple inputs from neurons
in the previous layer through connections with associated
strengths (weights). These weights represent the backbone of
the ANN model. Each neuron performs a simple computation
to calculate a net input (weighted sum) and transforms that net
input to an output using an assigned transfer function to serve as
an input to neurons in the subsequent layer or as a network output
if the neuron is in the output layer as shown in Fig. 1.
In general, three very important steps must be carefully
considered and addressed in building an effective ANN
(b) model: 1) selection of database; 2) network architecture; and
3) network training and testing.

Fig. 2—Values of HRWR in: (a) original mixtures collected Selection of database
to model washout property (before data screening); and (b) The selection of the database to train a neural network is
selected mixtures used to model washout property (after of paramount importance, and for a neural network to actually
data screening). capture the relationships between the parameters of UWC

166 ACI Materials Journal/March-April 2003


mixtures and its engineering properties, it must be trained on Figure 2(a) and (b) show an example of the screening process
large and comprehensive sets of reliable experimental data of HRWR dosages in the data sets used to model the washout
that contain the influential factors on underwater concrete resistance property. Table 1 includes the ranges, mean
properties. Basic parameters that affect the behavior of UWC values, and standard deviations of all input and output variables
mixtures such as ordinary portland cement, water, silica fume, in the final database. A total of 175 UWC mixtures have been
fly ash, slag, fine and coarse aggregates contents, and dosages selected from about 235 mixtures gathered from the literature
of anti-washout and HRWR admixtures were used in this because they met the criteria used in building the database.
study for training the ANN model to predict the properties The distribution of the selected database between training
of UWC mixtures such as slump, slump flow, washout resis- and testing is shown in Table 2.
tance, and 28-day compressive strength of cylinders cast
in and above water. Network architecture
An extensive literature review has identified a great deal As mentioned previously, a feed-forward back-propagation
of experimental data on UWC. Only data having mixture neural network is used in this study to predict the perfor-
ingredients with similar physical and chemical properties, mance of UWC mixtures. To date, there are no established
however, were considered in the training and testing process. rules to determine the architecture of such a network that
Moreover, a wide-spanning discontinuity of some of the produces best performance. Therefore, a trial-and-error
mixture variables can destabilize the training process and approach is usually adopted. It should be noted that it is possible
reduce the prediction capability of the network. It was decided to construct different network architectures with satisfactory
to disregard isolated data and reduce the range of variables. performance for the same application. A network architecture

Table 1—Range, average and standard deviation of measured input and output variables (data shown
is normalized)
Variables Training data Testing data
Range Average Standard deviation Range Average Standard deviation
Cement* 11.83 to 25.27 18.70 4.06 12.55 to 25.23 19.06 3.96
Water* 5.50 to 11.26 9.35 1.68 6.94 to 11.25 9.40 1.76
Fly ash* 0.00 to 5.00 1.22 1.95 0.00 to 5.00 0.96 1.80
Slag* 0.00 to 12.57 1.12 3.55 0.00 to 12.50 0.84 3.24
Silica fume* 0.00 to 5.00 0.97 1.13 0.00 to 2.40 1.08 0.98
* 26.70 to 39.57 30.45 1.96 29.46 to 35.21 30.53 1.59
Sand
Gravel* 25.24 to 50.22 37.90 4.63 29.94 to 44.89 37.83 4.81
AWA* 0.00 to 0.05 0.02 0.01 0.00 to 0.04 0.02 0.01
* 0.00 to 0.60 0.32 0.22 0.02 to 0.50 0.29 0.20
HRWR
Slump† 0.16 to 0.29 0.24 0.02 0.20 to 0.27 0.24 0.02
Slump flow† 0.20 to 0.67 0.46 0.09 0.28 to 0.58 0.43 0.09
‡ 0.01 to 0.16 0.07 0.05 0.01 to 0.13 0.06 0.03
Washout
fc′ (28 day)/air§ 0.30 to 0.67 0.52 0.08 0.38 to 0.62 0.51 0.08
fc′ (28 day)/water §
0.13 to 0.53 0.30 0.10 0.17 to 0.48 0.32 0.14
*
Percentage of total weight of mixture.

mm/1000.

% mass loss.
§
MPa/100.

Table 2—Database (distribution of data sets between training and testing)


Slump Slump flow Washout 28-day fc′ (air) 28-day fc′ (water)
Source of data Training Testing Training Testing Training Testing Training Testing Training Testing
18 8 1 8 1 6 1 7 0 6 1
Sonebi, M.
Sonebi and Khayat8 53 5 53 5 44 5 52 6 49 8
10 20 2 21 1 20 2 25 3 — —
Khayat et al.
Khayat, El Gattioui, and
13 2 0 0 13 2 13 2 — —
Nmai6
Khayat19 6 0 5 1 6 0 6 0 — —
20 3 0 0 0 3 1 3 0 — —
Khayat and Hester
21 24 3 23 4 24 3 — — — —
Khayat and Assaad
22 14 1 14 1 14 1 — — — —
Sonebi and Khayat
23 9 1 9 1 9 1 — — — —
Sonebi and Khayat
Total 150 15 133 14 139 16 106 11 55 9

ACI Materials Journal/March-April 2003 167


the outcomes of the network. Normalization was achieved
by simply dividing the experimental values of slump and
slump flow by 1000 and those of compressive strengths by
100. Experimental values of washout resistance were already
between 0 and 1 (percentage) and there was no need to
normalize them. Moreover, to simplify the training process and
to prevent the network from destabilizing, it is recommended
that the values of all input variables provided for training and
testing be neither too small nor too large. Therefore, all
values of mixture ingredients were normalized as percentage
of the mixture’s total weight.
Commercially available computer software24 was used to
build and train the feed-forward neural network shown in
Fig. 3. A special programming code was written to assign the
range of the input and output variables, the number of layers
Fig. 3—Architecture of neural network model. and number of units in each layer, the learning algorithm and
learning rate, the activation function for each processing unit,
and the minimum desired error as a stopping mechanism to
(Fig. 3) developed by Nehdi, El Chabib, and El-Naggar16 execute the training process. As mentioned previously, biases
was selected for the present study. It consists of an input layer, and weights or connection strengths between processing
an output layer, and two hidden layers. Trial-and-error tests units are modified through the training process and the final
demonstrated that this architecture offers better perfor- values represent the trained ANN model. Therefore, these
mance than a large number of other variations. The input weights and biases have to be initialized, and they are normally
layer contains nine processing units (neurons) representing assigned small random initial values. The network is presented
the influential parameters of UWC mixtures. The output layer with the variables in the input vector of the first training pattern
contains one unit representing the UWC property being and is asked to predict an appropriate output through the
modeled. The first hidden layer includes ten processing units computation and activation of processing units in the hidden
while the second hidden layer has only five units. A full and output layers. The predicted output is compared to the
connection between processing units in adjacent layers was normalized measured UWC property corresponding to the
adopted and a sigmoid function (logsig) was assigned as the same input vector and provided in the input/target training
transfer (activation) function for all processing units in the patterns. The error between the predicted output and the
hidden and output layers. normalized property of UWC is calculated and stored. The
network is then presented with the second training pattern,
Training process and so on, until all patterns available for training are covered.
Training a feed-forward back-propagation neural network The average sum-squared of all errors that the network
to predict the properties of UWC mixtures consists essentially stored after each training pattern is calculated and back-
of teaching the network the relationship between the mixture propagated through a learning mechanism to adjust the biases
ingredients and its performance and how individual parameters and connection strengths (weights) between the processing
influence the overall behavior of UWC mixtures. This learning units. Most learning algorithms are based on the generalized
process can take place in a supervised or unsupervised manner. delta rule; they update the network biases and weights in
Supervised learning is adopted in this study and consists of the direction of the steepest descent of the error to minimize
providing the network with sets of training patterns (inputs/ the network performance function (reduce the average
target) that include the measured output for each input vector sum-squared errors). The learning algorithm adopted in this
and the network is told what to learn. study is the Levenberg-Marquardt algorithm,24 however,
The training process is carried out by presenting the network which back-propagates the network errors based on the
with sets of experimental data in a patterned format. Each Jacobian matrix that contains the first derivatives of the
training pattern includes an input vector of nine elements network errors with respect to the weights and biases. A typical
representing the mixture ingredients (that is, cement, silica iteration of such algorithm can be written as
fume, fly ash, slag, water, sand, gravel, AWA, and HRWR
contents) and a corresponding output representing one UWC T –1 T
property (that is, slump, slump-flow, washout resistance, or w k + 1 = w k – [ J J + µI ] J e (1)
compressive strength). The exclusion of one or more properties
of UWC in some of the collected data has further reduced the where wk = a vector of current weights and biases; µ = a scalar
number of valid experimental training patterns that include (small value); J = the Jacobian matrix; JT = the transpose
all modeled properties of UWC, which makes it difficult to matrix of J; I = the identity matrix; and e = a vector of the
train the network to predict all properties simultaneously. network errors. This iterative process continues until the
Therefore, each property was modeled separately but using network stabilizes and converges to a desired minimum error
the same network architecture. This study will not focus on between the output predicted by the network and the provided
explaining the properties of UWC; unfamiliar readers are targets (normalized measured UWC property).
referred to the appropriate literature.7,8,10,18 It is important to note that during the training process, a
The sigmoid (logsig) function assigned as a transfer function neural network is trying to provide memory storage for the
for the output processing unit produces values between 0 and 1. data provided in the training patterns. In other words, the
Therefore, values of all UWC properties in the final database network is trying to memorize the values of UWC properties
had to be normalized between 0 and 1 to be compatible with provided in the training patterns that correspond to certain

168 ACI Materials Journal/March-April 2003


values of mixture ingredients. Therefore, it is possible to
continue the training to a point where the difference between
network prediction and normalized measured UWC property
for all training mixtures is virtually zero. In such a case,
however, the model will generally demonstrate poor
generalization capability and will not provide acceptable
predictions when it is presented with new mixtures excluded
from the training. This phenomenon is called over-fitting
(over-training), and it is usually avoided by dividing the
experimental data into training and validation sets; training
is stopped when the average sum-squared errors (ASSE) of the
validation sets starts to deviate from the ASSE of the training
sets. In this study, over-fitting was avoided via an early
stopping mechanism that consisted of providing an appropriate
desired minimum ASSE = 1E–4.
Fig. 4—ANN response in predicting slump values.
RESULTS AND DISCUSSION
The acceptance/rejection of an ANN model for UWC is
based on its ability to generalize its predictions to new mixtures
not previously used in the training and thus unfamiliar to the
model. Good generalization, however, greatly depends on
the final set of weights and biases and the degree of success
of the training process. In other words, before testing the
network model to determine its ability to predict the properties
of new UWC mixtures, the success of the training strategy
must first be evaluated and the quality of the model responses
to training cases must be analyzed.

Performance of ANN model using training data


The network model shown in Fig. 3 was trained to predict
the slump, slump flow, washout resistance, and 28-day
compressive strength of concrete mixtures designed for UWC
construction and repair. Because there was no clear trend in
neural network predictions for either over- or underestimating Fig. 5—ANN response in predicting slump-flow values.
the mixture properties, the reliability of its predictions was
evaluated using the average absolute error (AAE) given by

n
Y –Y
∑ ---------------------------------
1 meas pred
AAE = --- - (2)
n Y pred
i=1

where
Ymeas = measured value of concrete property from experi-
mental data;
Ypred = predicted value of concrete property by network; and
n = number of data points.
The network was trained to predict the slump values of
UWC mixtures using a total of 150 training patterns selected
from nine different sources as shown in Table 2. Each training
pattern contains an input vector of nine elements representing
the normalized ingredients of an UWC mixture and a target
representing the corresponding normalized measured slump. Fig. 6—ANN response in predicting washout values.
Satisfactory completion of the training process was verified
by calling the network model to predict the slump values of flow of UWC mixtures. The network was presented with
the mixtures used in the training process and the response is training patterns containing the normalized values of mixture
plotted in Fig. 4. Clearly, the network has successfully ingredients and the corresponding slump flow and its task
learned to map between mixture ingredients and associated was to predict the slump flow values for UWC mixtures used
slump values of the training patterns and its performance in in the training process. Figure 5 shows the relationship
predicting the slump of the training mixtures is satisfactory between the predicted and normalized measured slump flow
with an AAE = 0.7%. of the training data. Clearly, all data points are located along
A similar strategy was adopted in training the same network the equity line with an AAE = 1.9% and the model response
architecture using a total of 133 training patterns selected in predicting the slump flow of the training mixtures was
from seven different sources (Table 2) to predict the slump considered satisfactory.

ACI Materials Journal/March-April 2003 169


Table 3—Input variables for UWC mixtures used to
test predictions of ANN model (mixture components
shown are normalized as % total weight/m3)
Normalized values (% of mixture’s total weight/m3)
Mixture
No. C W FA Slag SF Sand Gravel AWA HRWR
M1 25.24 10.36 0.00 0.00 0.00 29.55 34.79 .0000 .0671
M2 22.70 10.32 0.00 0.00 2.48 29.55 34.69 .0176 .2643
M3 18.51 10.31 5.05 0.00 1.65 29.47 34.69 .0378 .3082
M4 12.55 10.31 0.00 12.55 0.00 29.61 34.74 .0176 .2207
M5 23.64 11.11 0.00 0.00 0.00 30.00 35.22 .0355 .0205
M6 21.23 11.15 0.00 0.00 2.35 29.91 35.12 .0354 .2378
M7 14.65 7.63 3.98 0.00 1.31 33.36 38.65 .0269 .6410
M8 14.61 7.61 4.00 0.00 1.33 30.49 41.58 .0010 .6416
Fig. 7—ANN response in predicting 28-day fc′ values (air-
M9 17.29 7.07 0.00 0.00 0.00 30.46 44.90 .0121 .2800
cast specimens).
M10 15.87 7.10 1.38 0.00 0.00 30.45 44.89 .0121 .3049
M11 17.21 6.94 0.00 0.00 1.49 30.05 43.64 .0187 .6489
M12 15.76 8.05 0.00 0.00 1.37 30.19 44.53 .0000 .0933
M13 21.66 11.07 0.00 0.00 1.87 35.21 29.94 .0235 .2410
M14 21.42 11.25 0.00 0.00 2.37 29.74 34.87 .0167 .3390
M15 23.61 11.10 0.00 0.00 0.00 29.96 35.18 .0354 .1385
M16 14.20 8.44 0.00 0.00 0.00 38.91 39.34 .0000 .0769
M17 18.82 6.98 0.00 0.00 0.00 30.23 43.90 .0000 .0658
M18 11.83 11.13 0.00 11.83 0.00 30.03 35.18 .0166 .2464
M19 14.66 7.59 4.00 0.00 1.31 30.47 41.58 .0529 .2697
M20 16.23 7.57 0.00 0.00 1.79 27.63 46.60 .0000 .1812
M21 13.07 7.18 3.57 0.00 1.21 37.51 36.93 .0125 .0850
M22 23.53 11.02 0.00 0.00 0.00 30.06 35.23 .0165 .1898

Fig. 8—ANN response in predicting 28-day fc′ values M23 12.55 10.30 0.00 12.55 0.00 29.59 34.72 .0376 .3026
(underwater-cast specimens). M24 21.24 11.15 0.00 0.00 2.35 39.92 35.13 .0165 .2069
M25 17.12 11.09 4.63 0.00 1.53 29.66 35.61 .0163 .3560
Similarly, the same network model was trained to predict Note: C = cement; W = water; FA = fly ash; SF = silica fume; AWA = anti-washout
admixture; HRWR = high-range water-reducing admixture.
the washout values of UWC mixtures using a total of 139
training patterns selected from nine different sources as
shown in Table 2. After completion of the training process, data and to perform well when it is presented with unfamiliar
the model was required to predict the washout values for the new data from within the range of the input variables used in
training mixtures and its response is plotted in Fig. 6. The the training. Therefore, the ability of the network model thus
difference between the network predictions and the measured developed to predict the fluidity and compressive strength of
washout values of the training mixtures was slightly larger new underwater concrete mixtures excluded from the training
than that of the slump and slump flow with an AAE = 13.5%. data must be validated. The model was presented with a total
The slightly lower performance of the network model in this of 15 mixtures, randomly selected from seven different studies
case is believed to be due to the numerically very small values as shown in Table 2 (proportions of all mixtures used to
of washout; even small prediction errors would lead to high validate the ANN model are listed in Table 3), and was required
AAE. It could also be due to the higher variability in the to predict the slump value associated with each mixture. At
experimental washout data itself, which is more delicate to this point, only an input vector of nine elements representing
obtain in the lab compared to slump and slump flow values. the normalized values of mixture ingredients was presented
Finally, the network model was trained on 106 training to the network and no knowledge of the actually measured
patterns from six different sources and 55 training patterns slump was provided. The predicted and normalized measured
from two different sources (Table 2) to predict the 28-day slump values are listed in Table 4 and the model response is
compressive strength of concrete specimens cast in the air shown in Fig. 4 where the predicted slump values are plotted
and underwater, respectively. The network predictions of against the normalized measured slump values. The testing
28-days fc′ (air) and 28-days fc′ (water) for the training points are slightly over or under the equity line with an AAE
mixtures are plotted in Fig. 7 and 8 with satisfactory perfor- = 4.6%, indicating that model predictions are appropriate.
mance and AAE of 1.2 and 1.0%, respectively. Similarly, the neural network model was tested on 14 and
16 new mixtures to validate its ability of generalizing its
Performance of ANN model using testing data predictions of the slump flow and washout, respectively.
As mentioned previously, the acceptance/rejection of a The network performance in predicting the associated slump
successfully trained neural network model is determined by flow and washout for each new mixture in the testing data
its ability to generalize its predictions beyond the training was satisfactory. The predicted slump flow and washout values

170 ACI Materials Journal/March-April 2003


Table 4—Measured and predicted values for Table 5—AAE of ANN model predictions for
properties of UWC mixtures used in testing training and testing data sets
ANN model Average absolute error, %
Underwater concrete
fc′ (28 days), MPa/100 properties Training Testing
Slump, Slump flow,
Mix- mm/1000 mm/1000 Washout, % Cast in air Cast in water Slump 0.7 4.6
ture
* †
No. Meas. Pred. Meas. Pred. Meas. Pred. Meas. Pred. Meas. Pred. Slump flow 1.9 10.6
M1 .235 .240 .400 .473 — — .461 .428 — — Washout 13.5 10.6
M2 .220 .214 .380 .341 .098 .103 .619 .625 .483 .491 28-days fc′ (air) 1.2 4.4

M3 .220 .222 .420 .429 .018 .013 .565 .543 .475 .447 28-days fc′ (water) 1.0 2.1

M4 .265 .265 .540 .507 .125 .134 .383 .400 .230 .229
M5 .200 .177 .350 .447 .052 .059 .447 .430 .286 .284
M6 .250 .250 .440 .432 .049 .050 .547 .551 .463 .462
M7 .253 .271 .500 .475 .060 .059 .617 .629 — —
M8 .273 .242 — — — — .442 .504 — —
M9 .230 .245 — — .043 .024 .405 .385 — —
M10 .225 .226 — — .030 .048 .567 .555 — —
M11 .230 .196 .415 .528 .028 .028 — — — —
M12 .220 .218 .380 .462 .115 .113 — — — —
M13 .270 .280 .585 .615 .082 .082 — — — —
M14 .270 .270 .550 .549 .048 .048 — — — —
M15 .250 .264 .500 .520 .062 .066 — — — —
M16 — — .280 .279 — — — — — —
M17 — — .330 .267 — — — — — — Fig. 9—Sensitivity of ANN model to HRWR in predicting
M18 — — — — 0.47 0.47 — — — — slump values.
M19 — — — — .036 .036 — — — —
M20 — — — — .090 .090 — — — —
M21 — — — — — — .547 .572 — —
M22 — — — — — — — — .317 .317
M23 — — — — — — — — .284 .284
M24 — — — — — — — — .377 .402
M25 — — — — — — — — .286 .291
*
Meas. = measured.

Pred. = predicted.

are plotted against the measured values in Fig. 5 and 6, respec-


tively. Again, it is shown that all testing points are located
within the cluster formed by the training data points and
slightly over or under the equity line with an AAE = 10.6%
in both cases. Fig. 10—Sensitivity of ANN model to AWA in predicting
Finally, to determine the ability of the network model to washout values.
predict the 28-day compressive strength of UWC mixtures
cast in the air or underwater, the model was presented with and maintaining all other mixture ingredients unchanged.
randomly selected 11 and 9 new mixtures, respectively. The Normalized ingredients of the base mixture include 21.2%
network predictions for the testing mixtures are plotted cement, 11.15% water, 2.35% silica fume, 30% sand,
against the corresponding normalized measured 28-day 35.12% gravel, 0.035 AWA, and 0.238% HRWR. The decision
compressive strength in Fig. 7 and 8, respectively. The network of changing the dosage of HRWR alone was based on the
response was accurate; all testing points in Fig. 7 and 8 are fact that its influence on the workability of concrete is well
located along the equity line with an AAE = 4.4% in the case known and documented. The eight new mixtures were
of fc′ (air) and 2.1% in the case of fc′ (water). AAEs for training presented to the ANN model to predict the associated slump
and testing sets are listed in Table 5. and the relationship between the HRWR dosage and predicted
Because the neural network model thus developed showed slump for the new mixtures is shown in Fig. 9. It is clear that
satisfactory performance and demonstrated its ability to predict by increasing the dosage of HRWR, the slump increased but
the properties of UWC mixtures designed within the practical reached a threshold value corresponding to the well-known
training range, it would be worth investigating whether the saturation dosage of HRWR.
model has captured the sensitivity of mixture properties to Similarly, six new mixtures were created from a single
individual ingredients. Therefore, a single mixture was mixture randomly selected from the testing data used in the
randomly selected from the testing data and used to create washout model, by only changing the dosage of AWA to
eight new mixtures by only changing the dosage of the HRWR investigate the ability of the ANN model to recognize the

ACI Materials Journal/March-April 2003 171


range of input variables by providing additional training sets
that cover the new range. In addition, the present model did
not incorporate durability characteristics of UWC and did
not account for the mixing, handling, and curing methods,
which may affect the behavior of UWC. This can, however,
be easily incorporated in the model provided that adequate
data for training the network on these aspects become
available in the future.

ACKNOWLEDGMENTS
This research was supported by the Natural Science and Engineering
Research Council of Canada (NSERC) through a grant awarded to M. Nehdi.
The authors acknowledge K. H. Khayat from Sherbrooke University in
whose lab most experimental data used to create the model was developed.

NOTATION
Fig. 11—Sensitivity of ANN model to HRWR in predicting AAE = average absolute error
ANN = artificial neural network
washout values.
AWA = anti-washout admixture
e = vector of neural network errors at output layer
influence of the AWA on the washout resistance property. fc′ = compressive strength of concrete (28 days)
I = identity matrix
Normalized ingredients of the base mixture include 21.6% J = Jacobian matrix
cement, 11.07% water, 1.87% silica fume, 30% sand, 35.2% JT = transpose of Jacobian matrix
gravel, 0.005 AWA, and 0.241% HRWR. The network was HRWR = high-range water-reducing admixture
called to predict the associated washout values of the new n = number of neurons in layer l – 1
mixtures and its predictions of washout versus the AWA Ujl = net input of neuron j in layer l
UWC = underwater concrete
dosage are illustrated in Fig. 10. It is indicated that by increasing Wk = vector of current weights and biases in neural network at time k
the dosage of AWA, the washout value decreased. The effect Wk + 1 = vector of current weights and biases in neural network at time k + 1
of the HRWR dosage on the washout resistance was investi- wjil = connection strength between neuron j in layer l and neuron i in
gated in a similar fashion and results are illustrated in Fig. 11. It previous layer
is shown that at low dosage, the HRWR has limited effect on Xl-1i = input coming from neuron i in layer l – 1
washout. A slight decrease of washout was observed as the Ylj = output of neuron j layer l
Ymeas = normalized measured underwater concrete property
HRWR dosage increased from 0.05 to 1% likely due to the Ypred = predicted underwater concrete property
improved homogeneity of the mixture. As the HRWR dosage θjl = threshold value
increased beyond this range, the vulnerability of concrete to µ = scalar in Levenberg-Marquardt learning algorithm
dilution increased and washout increased as a consequence.
The effect of the HRWR dosage on washout is, however, REFERENCES
much less significant than that of the AWA. 1. Yamaguchi, M.; Tsushida, T.; and Toyoizumi, H., “Development of
High-Viscosity Underwater Concrete for Marine Structures,” Marine
Concrete, International Conference on Concrete in Marine Environment,
CONCLUSIONS Concrete Society, 1986, pp. 235-245.
Based on this investigation, the following conclusions can 2. Nagataki, S., “Antiwashout Admixtures for Underwater Concrete,”
be warranted: RILEM Report No. 10, Application of Admixtures in Concrete, A. M.
Paillére, ed., E&FN Spon, 1995, pp. 77-87.
1. This study demonstrates the feasibility of using a feed- 3. Khayat, K. H., “Effect of Antiwashout Admixtures on Fresh Concrete
forward back-propagation neural network to predict the Properties,” ACI Materials Journal, V. 92, No. 2, Mar-Apr. 1995, pp. 164-171.
behavior of UWC mixtures with good accuracy. The model 4. Kawai, T., “Non-Dispersible Underwater Concrete Using Polymers,”
thus developed was not only able to predict the behavior of Proceedings, 5th International Congress on Polymers in Concrete,
Brighten, England, V. 5, Chapter 11, 1987, 6 pp.
mixtures used in the training process, but it also generalized 5. Ghio, V. A.; Monteiro, P. J. M.; and Gjorv, O. E., “Effects of Polysac-
its predictions beyond the training data to new mixtures charide Gums on Fresh Concrete Properties,” ACI Materials Journal, V. 91,
designed within the practical training range; No. 6, Nov.-Dec. 1994, pp. 602-606.
2. Predicting the properties of UWC mixtures as a function 6. Khayat, K. H.; El Gattioui, M.; and Nmai, C., “Effects of Silica Fume
and Fly Ash Replacement on Stability and Strength of Fluid Concrete
of mixture ingredients is difficult to achieve analytically, Containing Anti-Washout Admixtures,” Superplasticizers and other
whereas a successfully trained neural network model can Chemical Admixtures in Concrete, 5th International CANMET/ACI
predict such behavior instantly and accurately. Therefore, Conference, SP-173, V. M. Malhotra, ed., American Concrete Institute,
ANNs can provide a powerful alternative approach to the Farmington Hills, Mich., 1997, pp. 695-718.
7. Khayat, K. H., and Sonebi, M., “Effect of Mixture Composition on
available traditional statistical methods in modeling the Washout Resistance of Highly Flowable Underwater Concrete,” ACI Materials
properties of UWC mixtures; Journal, V. 98, No. 4, July-Aug. 2001, pp. 289-295.
3. The model developed herein had a demonstrated ability 8. Sonebi, M., and Khayat, K. H., “Effect of Mixture Composition on
to recognize and evaluate the effect of individual ingredients Relative Strength of Highly Flowable Underwater Concrete,” ACI Materials
Journal, V. 98, No. 3, May-June 2001, pp. 233-239.
on the properties of UWC mixtures. Thus, it could be used to 9. JSCE, “Recommendations for Design and Construction of Anti-Washout
develop UWC mixtures with tailor-made properties without Underwater Concrete,” Concrete Library of JSCE, V. 19, 1992, 89 pp.
need for a large number of trial batches as in the case of cur- 10. Khayat, K. H.; Sonebi, M.; Yahia, A.; and Skaggs, C. B., “Statistical
rent practice; and Models to Predict Flowability, Washout Resistance, and Strength of
Underwater Concrete,” Proceedings, RILEM International Conference
4. Although the prediction capability of the present model on Production Methods and Workability of Concrete, V. 3, Glasgow,
is limited to data located within the boundaries of the training England, 1996, pp. 463-481.
range, the model can easily be retrained to span a wider 11. Ghaboussi, J.; Garrett, J. H. Jr.; and Wu, X., “Knowledge-Based

172 ACI Materials Journal/March-April 2003


Modeling of Material Behavior with Neural Networks,” ASCE Journal of Conference on Computational Engineering Science, ICES98, Atlanta, Ga.,
Engineering Mechanics, V. 117, No. 1, 1991, pp. 132-153. Tech Science Press, 1998, pp. 551-557.
12. Haykin, S., Neural Networks: A Comprehensive Foundation, 18. Sonebi, M., “Development of High Performance, Self-Compacting
Prentice Hall, 2nd edition, Upper Saddle River, N.J., 1999, 842 pp. Concrete for Underwater Repair of Hydraulic Structures,” PhD thesis,
13. Rumelhart, D. E.; Hinton, G. E.; and Williams, R. J., “Learning Internal Université de Sherbrooke, Québec, Canada, 1997, 451 pp.
Representation by Error Propagation,” Rumelhart, D., McClelland, J. and the 19. Khayat, K. H., “Effects of Antiwashout Admixtures on Fresh Concrete,”
PDP Research Group, Parallel Distributed Processing, V. 1: Foundation, MIT ACI Materials Journal, V. 92, No. 2, Mar.-Apr. 1995, pp. 233-239.
Press, 1986, pp. 318-362. 20. Khayat, K. H., and Hester, H. T., “Evaluation of Concrete Mixtures
14. Buenfeld, N. R.; Hassanein, N. M.; and Jones, A. J., “An Artificial for Underwater Pile Repairs,” ASTM Cement, Concrete and Aggregates,
Neural Network for Predicting Carbonation Depth in Concrete Structures,” V. 13, No. 1, 1991, pp. 32-41.
Manuals and Reports on Engineering Practice, ASCE, 1998, pp. 77-117. 21. Khayat, K. H., and Assaad, J., “Relationship Between Washout
15. Oh, J. W.; Lee, I. W.; Kim, J. T.; and Lee, G. W., “Application of Resistance and Rheological Properties of High-Performance Underwater
Neural Networks for Proportioning of Concrete Mixes,” ACI Materials Concrete.” (submitted to ACI Materials Journal for publication)
Journal, V. 96, No. 1, Jan.-Feb. 1999, pp. 61-67. 22. Sonebi, M., and Khayat, K. H., “Effect of Free-Fall Height in Water
16. Nehdi, M.; El Chabib, H.; and El-Naggar, M. H., “Predicting the on Performance of Flowable Concrete,” ACI Materials Journal, V. 98, No. 1,
Performance of Self-Compacting Concrete Mixtures Using Artificial Neural Jan.-Feb. 2001, pp. 72-78.
Networks,” ACI Materials Journal, V. 98, No. 5, Sept.-Oct. 2001, pp. 394-401. 23. Sonebi, M., and Khayat, K. H., “Effect of Water Velocity on Perfor-
17. Haj-Ali, R. M.; Pecknold, D. A.; and Ghaboussi, J., “Micromechanics- mance of Self-Consolidating Underwater-Cast Concrete,” ACI Materials
Based Constitutive Damage Models for Composite Materials Using Artificial Journal, V. 96, No. 5, Sept.-Oct. 1999, pp. 519-528.
Neural Networks,” Modeling and Simulation Based Engineering, S. N. 24. Demuth, H., and Beal, M., Neural Network Tool Box for Use with
Atluri and P. E. O’Donoghue, eds., Proceedings of the International MATLAB Version 3.0, The Math Works Inc., 1998.

ACI Materials Journal/March-April 2003 173

Potrebbero piacerti anche