Predicting properties of concretes and modelling the ready mixed concrete

delivery system with artificial neural networks

By -Diwate.Somanath.N1 Guided by Dr. Paresh.chandra.Deka2
1,2

National Institute Of Technology Karnataka, India

1) Introduction

Concrete is the most widely used structural material in the world for
construction. Traditionally, concrete has been fabricated from a few well-defined
components: cement, water, fine aggregate, coarse aggregate, etc. In concrete mix
design and quality control, the strength of concrete is the most important property.
And many properties of concrete such as elastic modulus, water-tightness or
impermeability, resistance to weathering agents, etc. are directly related to the
strength. The strengths of concrete include compressive strength, tensile strength,
flexural strength, shear strength, bond strength, and so on. The compressive
strength of concrete is many times greater than other types of strength, and a
majority of concrete elements are designed to take advantage of the higher
compressive strength of the material. Most often an ultimate target in the mixture
design is the 28-day compressive strength. The 28-day compressive strength is
determined based on a standard tri-axial compression test and is accepted. The
compressive strength of concrete is mostly used criterion introducing concrete.
However, testing for Strength of concrete specimens is a complicated and timeconsuming task. More importantly, it is too late to make improvement if the test result
does not satisfy the required strength, since the test is usually performed at the 28th
day after the placement of concrete at the construction site. Therefore, strength
prediction before the placement of concrete is highly desirable. This study presents
the effort in applying the neural network technique for predicting the compressive
strength of concrete based on concrete mix proportions.
Ready mixed concrete (RMC) is an essential material in contemporary
construction and engineering projects and thus it is imperative that the process of
acquiring and handling RMC is managed with the utmost efficiency and accuracy.
Further, RMC must be delivered in a workable state and, given that it has a limited
shelf-life of roughly one and a half hours, any mismanagement of the processes of
delivery and use of this vulnerable material could result in this shelf-life being
exceeded.

The design of a concrete mix involves the choice of constituent proportions

that will result in certain target properties. Most mix design methods recognise slump
in the fresh state and strength in the hardened state as being the key properties of a
concrete mix. For example, in the British method, the starting point is the selection,
based on established empirical relationships, of the two independent parameters
water cement ratio and water content that will yield a target strength and target
slump, respectively. The method then goes on to describe the proportioning of the
various constituents, but ends with guidelines for modifying those proportions based
on trial mixes, thus implying that the empirical relationships have to be updated.
Furthermore, any mix design that involves mineral additives and chemical
admixtures will require more comprehensive empirical relationships.

2) Data collection and modelling

The data for the ready mixed concretes (RMC) were collected from batching
plants. The input parameters were the weights per unit concrete volume of: i).
cement; ii). water; iii). fine aggregate; iv)coarse aggregate; and v)fly ash vi)admixture
i.e. a retarding admixture with some water reducing properties. The output
parameters were standard 28-day cube strength. When the data were transformed
into non-dimensional ratios, the input ratios were: i. Cement /water; ii. Total/water
where total refers to the weight of all the mix constituents; iii. Coarse aggregate/fine
aggregate; and iv. Retarder/cement; the outputs remained slump and strength.

For modelling of the ready mixed concrete delivery system the data to be
collected Month of operation Type of operation Truck volume (m 3) Total operation
volume (m) Average inter arrival time (minutes) Number of loads in operation
Number of accepted loads No. of rejected loads

3) Literature review
3.1
Neural networks for predicting properties of concretes with admixtures.

W.P.S. Dias, S.P.Pooliyadda Department of Civil Engineering, University of

Moratuwa, Moratuwa, Sri Lanka, 20 December 2000

Back propagation neural networks were used to predict the strength and
slump of ready mixed concrete and high strength concrete, in which chemical
admixtures and/or mineral additives were used. Although various data transforms
were tried, it was found that models based on raw data gave the best results. When
non-dimensional ratios were used, arranging the ratios such that their changes
resulted in corresponding changes in the output e.g. increases in ratios to cause
increases in output values improved network performance.

Objectives
The main objective was to explore the feasibility of using neural networks for
predicting the strength and slump of concrete. Two types of concretes were
considered, namely: i). Normal strength ready mixed plant mixes where only
retarders with moderate water reducing properties were used; and ii. High strength
concrete mixes that used silica fume and plasticisers. The data were obtained from
industry and literature surveys, respectively, as opposed to laboratory results from a
single source; thus the modelling was a genuine test of the generalising capability of
the neural networks. Furthermore, both types of mixes involved admixtures and/or
additives over and above the basic concrete constituents, thus testing the neural
networks in applications which are not very well covered by conventional mix design
methods and, hence, where they would be particularly useful.

Data collection and modelling

The data for the ready mixed concretes (RMC) were collected from three
batching plants. Of these, 93 records were used for training and 44 for testing. The
input parameters were the weights per unit concrete volume of i. Cement; ii. Water;
iii. Fine aggregate;
iv. Coarse aggregate; and v. retarder i.e. a retarding admixture with some water
reducing properties.. The output parameters were slump and standard 28-day cube
strength. When the data were transformed into non-dimensional ratios, the input
ratios were: i. Cement/water; ii. Total/water where total refers to the weight of all the
mix constituents. iii. coarse aggregate/fine aggregate; and iv. Retarder/cement; the
outputs remained slump and strength.

The training of the networks was carried out with various numbers of nodes in
the single middle layer and various target accuracies. An upper bound to the number
of middle layers was arrived at from the condition that the total number of
parameters or weights of the network.

Conclusions

The following conclusions can be drawn from this work:

1. Neural networks trained with raw data on concrete mix constituent contents make
better predictions of strength and slump than those trained using non dimensional
ratios.
2. If non-dimensional ratios are used, it is better to define the ratios such that
increases (rather than decreases) in the ratios correspond to increases in strength or
vice versa.
3. The neural network models, which can easily incorporate additional model
parameters, result in less scatter in predicted values than those given by the multiple
regression models.
4. The modelling of slump requires less parameters in neural network models, but
the accuracy of prediction is markedly less than for strength prediction.

5. The modelling of concrete strength using non dimensional input parameters can
result in discrepancies with domain knowledge in concrete technology, because
undue weightage can be assigned to other ratios that change incidentally with the
binder/water ratio, which is the primary ratio that governs strength.
6. In order to perform rational and realistic sensitivity analyses for concrete mix
inputs, the sum of input weights per unit volume must be constrained to equal the
wet density of concrete. Such sensitivity analyses can only be done on networks
trained with raw data, rather than with non-dimensional ratios.
7. These sensitivity analyses demonstrated that the neural network modelling had
picked up the well known major dependencies of concrete strength, i.e. that
binder/water ratio was the key factor influencing concrete strength and that increases
in cement replacement with silica fume increased strength. The modelling had also
picked up second order effects, such as the increase of strength with increasing
coarse aggregate/fine aggregate ratio and the decrease of strength even at constant
binder/water ratio at excessive cementitious paste contents.

3.2 Modeling the ready mixed concrete delivery system with neural networks

L. Darren Graham a, Doug R. Forbes b, Simon D. Smith a,

a School of Engineering and Electronics, The University of Edinburgh, Edinburgh,
EH9 3JN, United Kingdom,
b Gleeson Construction Services, Haredon House, North Cheam, Sutton, Surrey,
SM3 9BS, United Kingdom;30 August 2005

The ready mixed concrete delivery system is a common construction process

in a very wide range of construction projects. The ability of the planners and
estimators of such projects to accurately determine the level of resources needed,
and to estimate the output of an efficient and effective operation is highly important

and thus modelling of the process can be useful. This paper presents a Neural
Network methodology to the modelling problem and outlines the two main
architectures employed: a feed-forward network and an Elman network. Many
combinations of layers, training algorithms, number of neurons, activation functions
and format of data were considered and the results were validated using an
independent validation data set with five goodness-of-fit tests. The results indicate
that two- and three-layer feed forward networks provide the best estimates of
concrete placing productivity and that the Elman network, not previously considered
in this type of study, was less successful.

Project data:

The data used in model development was taken from observation of the
RMCDS on three projects to construct wastewater treatment plants in the United
Kingdom. Each project involved unique clients and contractors. These observations
provided 212 examples of the RMCDS. NN model development involves using: a
training data set to provide examples to the NN from which the NN can learn; a
testing data set to gauge to ability of the trained NN to recreate accurately past
examples of the RMCDS not used in the training process. The model development
data set was split using a ratio of 75%/25% for training and testing, respectively. This
provided 159 training cases and 53 testing cases. An additional data set was
available for use in model validation. This data was collected by a contractor on a
project
to construct a major highway viaduct in the United Kingdom. This project involved a
different client, different contractors and operating conditions from those projects that
the model development data was taken. These observations provided 39 examples
of the RMCDS. The variance in client, contractors, operating conditions and project
type has produced a diverse set of project data, and any ability of a NN model to
accurately recreate the validation data, which is different from that on which the
model has been developed, will indicate the generic capability of NNs to model the
RMCDS.

Data transformation

The final step in preparing the data for use in modelling was to transform
(scale) the data into the range between _1 and +1, to improve the density of the data
over the problem domain. To allow a comparison between using scaled and raw data
in the training of neural networks, a set of raw data will also be used in neural
network model development. The scaled and raw data sets, each consisting of four
variables and 251 examples of past concreting operations are now ready for use in
neural network model development.

Conclusions and future work

The following conclusions may be drawn from this study:
(a) The standard NN development practice of scaling (transforming) raw data was
found to be produce NN models which were of poor quality (had a high test MSE).
This is most likely due to the application of only one type of transfer (scaling) function
in this research project. The scaling of data is problem-dependent and a survey of
transfer functions may improve the performance of NN models based upon scaled
data. Such a survey could be coupled with a sensitivity analysis to determine the
effect of transfer function on the performance of NNs and this is an area for future
work in this research area. In this research project, only raw data
Models were considered for use in the model validation process.
(b) Feed-forward NN models were found to be able to produce productivity estimates
of the RMCDS to a high degree of accuracy. This finding suggests that feed forward
NN models may be capable of modelling other similar construction processes, which
as yet have not been studied.

(c) The Elman NN had not been considered before in construction process
modelling, and was found to be capable of producing estimates of the RMCDS with a
promising degree of accuracy (13% error).
(d) Feed-forward NN models were found to be more capable of estimating the
RMCDS process productivity Elman models. This finding is somewhat surprising
given that the 3-layer Elman NN model produced the smallest MSE during training
and testing, a result which
should have translated to the ability to accurately estimate the process productivity,
but this did not happen as successfully as it could have. The reason for this could be
that the Elman architecture is more complex than a feed-forward one, as there are
more interconnecting neurons, etc., and this complexity possibly makes an Elman
NN more sensitive to outliers in the data and missing process variables.

Future research could be undertaken to determine some practical guidelines

for use in NN modelling of construction processes. This could include information
relating to: the effect of the size of the training set on model performance; the
validation methods available, and their effect on a models acceptability; and the use
of factorial analysis in analyzing a NN model. Additionally it may prove useful to
improve the validation process by determining how sensitive the model is to
weighting the importance of the goodness-of-fit tests that are used to select the
optimum NN model. This could be the subject of future research in this area

3.3)Application of Artificial Neural Network to Predict Compressive

Strength of High Strength Concrete
J. Noorzaei*, Universiti Putra Malaysia, MALAYSIA
S. J. S. Hakim, Universiti Putra Malaysia, MALAYSIA
M. S. Jaafar, Universiti Putra Malaysia, MALAYSIA

In this paper, a method to predict 28-day high compressive strength of

concrete by using multi layer feed forward neural networks (MFNNs) is proposed.
ANN model is constructed, trained and tested using the available data. Total of 368
different data High Strength Concrete (HSC)mix-designs collected from the technical
literature. The data used to predict the compressive strength in ANN are consisted
with eight input parameters that cover the cement, water, coarse aggregate, fine
aggregate, silica fume, super plasticizer, fly ash and granulated grated blast furnace
slag. For the training phase, different combination of layers, number neurons,
learning rate, momentum and activation functions were considered. The result shows
that the 8-10-6-1 architecture was the best possible architecture. The Relative
Percentage Error (RPE) for the training set was 7.02% and the testing set was
12.64%. The neural network (NN) models give high prediction accuracy, and the
research results conform to some rules of mix proportion of concrete. These
demonstrate that using NNs to predict concrete strength is practical and beneficial.

PROBLEM PRESENTATION
High strength concrete is complicated mixture, the influence parameters of the
materials are included cement, aggregate, water, mineral and chemical admixture,
etc. For example, cements have different types, chemical compounds, fineness and
strength. Coarse aggregates have natural, crushed and uncrushed; fine aggregates
maybe have different size, quality and mines. Admixtures used in high strength
concrete also have different types of chemical compounds. Moreover, the methods
of mixing, testing age, loading rate during test and transporting also can affect the
durability and strength of concrete. Thats way to product a high strength of concrete
became more complicated because of too much parameters and characteristic need
to be considered. Therefore, to find out the major parameters and the most important
characteristic, which influence the high strength of concrete, is an important point
discuss in this research. In addition, focusing on the potentials and suitability of ANN
applies in this application.

DATA SET FOR NEURAL NETWORK TRAINING AND TESTING

368 different high strength concrete mix designs experimental data are
collected from several different sources from the technical literature. Test data were
assembled for high strength concrete containing cement, coarse aggregate, fine
aggregate, water, fly ash, silica fume, granulated graded blast furnace slag, and

super plasticizer. These data was gathered for compressive strength of high strength
concrete at 28 days and the range of compressive strength is from 40 to 140MPa. A
neural network model was conducted, trained and testing using these available test
data. Outputs of 368 specimens, 256 were used as training examples, and 112 were
used as testing examples. The dividing of the data was carried out randomly
between the two sets.
CONCLUSION
I) The 8-10-6-1 architecture network is better than other architecture networks where
there are eight neurons in the input layer, ten neurons in the first hidden layer, six
neurons in the second hidden layer and one neuron in the output layer. II)
Calculation of mean percentage relative error for training and testing set data, shows
that artificial neural network is predicted the compressive high strength concrete with
error of 7.02% and 12.64%, respectively, that are acceptable in concrete technology.
III) The results prove that the artificial neural networks can work sufficiently for
predicting of compressive high strength of concrete and are more accurate than the
model use the regression analysis and conventional methods.
IV) From the results obtained it can be concluded that the artificial neural networks
are more suitable in modelling of complex problems and save a lot of computational
effort compared to conventional methods significantly. The use of these networks will
help in solving more complex problems.

3.4) Comparison Of Concrete Strength Prediction Techniques With Artificial

Neural Network Approach
Meltem zturan1, Birgl Kutlu1, Turan zturan2

Prediction of concrete strength is an important issue in ready-mixed concrete

industry, especially, in proportioning new mixtures and for the quality assurance of
the concrete produced. In this paper, it is aimed to illustrate that the artificial neural
networks can be used for predicting the 28-day strength of low to medium strength
concretes. The compositional, fresh concrete and early strength data obtained from
different batching plants of a ready-mixed concrete company have been defined in

terms of ten independent variables that are grouped in five different system models
to which neural network and multiple linear regression models have been applied.
The accuracies of prediction by artificial neural network and multiple linear
regression models as well as by Abrams law are compared on the basis of the
coefficient of determination. It appears that the best results are obtained by the
artificial neural network models using data for fresh concrete and early strength
simultaneously.

Research Significance
The objective of this paper is to illustrate that the ANNs can be used to predict the
28-day compressive strength of ready-mixed concrete through a comparative
analysis by ANN, MLR and Abrams prediction models. Data are gathered directly
from a ready-mixed concrete company for low to moderate strength concrete
mixtures produced over a period of six months at its different batching plants using
the same type of constituent materials. Various sets of combinations of available
data are considered for prediction modeling. Supervised learning models have been
utilized in which certain output nodes are trained to respond to certain input patterns
and the changes in connection weights due to learning have caused those same
nodes to respond to more general classes of patterns.

Modeling of the Prediction of 28-day Compressive Strength of Concrete

System Models
Strength of concrete is mostly defined by its components with regard to their
mass contents in unit volume of concrete and individual characteristics. Additional
factors related to the batching, mixing, placing and curing procedures affect the
strength development of concrete. Although experimental data available in the
literature provide a great deal of information, some important details might be lacking
in many situations, making the strength prediction a highly uncertain task. Therefore,
in this study the following input variables obtained through different stages of the
ready-mixed concrete production are considered to ultimately affect the 28-day
compressive strength of the concrete, fC28 (MPa).

Data Sets
Experimental data collected from five different batching plants of a readymixed concrete company for a period of six months for a specified strength concrete,
produced with the same type of materials, are used to train and test the strength
prediction models without normalization. The data set of 111 records indicate that the
workability range from stiff to plastic and even fluid consistencies, depending on the
casting and compacting conditions at the site. Statistical analysis of the 28- day 150
mm cube compressive strength indicates a mean value of 26.2 MPa and standard
deviation of 3.0 MPa for 111 test results. Statistical evaluation of concrete strength
based on EN206-1 conformity criteria for the mean strength and the minimum
individual value reveals a characteristic strength of 21.8 MPa and 23 MPa,
respectively. This evaluation indicates a C16/20 quality for the concrete. Moreover, a
coefficient of variation of 11.5 % indicates a moderate degree of quality control.

Network Parameters
For each system model described before, the Generalized Feed Forward
Artificial Neural Network model is applied for two different numbers of hidden layers
(HL = 1, 2) at six different numbers of first hidden layer neurons as stated before for
the data set. Thus, 12 different ANN models have been used for each system model.
For all of those ANN models, the following network parameters are taken the same:
Learning rule: Momentum (momentum factor = 0.5)
Stopping criteria: Mean Square Error (minimum MSE = 0.005)
Learning rate: 0.1
Activation function: Linear Sigmoid
Initial weight: Randomized

Training Results
In this study, 12 ANN models are applied to each of the five system models, using
Neuro Solutions ANN software package. Thus, a total of 60 strength prediction

models have been tried. ANN models performances are measured by the coefficient
of determination (R2)

3.4.3) Conclusions

This study is aimed at finding the best system model for the prediction of the
28-day compressive strength of concrete through a comparative analysis by ANN,
MLR and Abrams prediction models. A total of 65 sets of predictions, which result
from the application of 12 ANN models and one MLR model to five system models,
are produced. Abrams law is also considered for predicting the strength
on w/c basis. All of those results are compared using the coefficients of
determination calculated for the models with highest R2 values for each system
model to be more confident about the results. Based on the findings of this study it
can be concluded that:
1. MLR models are better in predicting the strength of concrete than ANN models for
M1 and M2 which include only the constituent materials and fresh concrete data.
However, the inclusion of early strength data in M4 and M5 results in better
prediction of strength by the ANN models. Both ANN and MLR models, on the other
hand, show almost the same performance for M3 which included only the early
strength data.
2. Although the inclusion of the early strength data, fC7, increases the speed of
learning process and the accuracy of prediction, best results for each of the prediction models are obtained when all of the input variables defined in this study have
been considered (M5).
3. Accuracy of the ANN models varies with the number of hidden layers and the
number of neurons in the hidden layers. For all system models better accuracies are
obtained with two hidden layers but no specified empirical criteria common to all

system models is found for determining the number of neurons in the first hidden
layer and implicitly in the second hidden layer.
4. Abrams law, predicting the strength based on w/c, gives considerably low
coefficient of determination which may be due to the effect of plasticizing admixture
on the microstructure of the concrete. Thus, it may state that an empirical equation
for determining strength must include terms representing concrete composition.
5. Plot of compressive strengths predicted by ANN models as function of w/c shows
a trend similar to the one obtained from the real data indicating that the ANN
prediction makes sense also from a physical stand point.

4) Methodology