Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Prediction of Concrete Compressive Strength Using Artificial Intelligence

Methods
To cite this article: H N Muliauwan et al 2020 J. Phys.: Conf. Ser. 1625 012018

View the article online for updates and enhancements.

This content was downloaded from IP address 158.46.144.19 on 31/10/2020 at 22:27

2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

Prediction of Concrete Compressive Strength Using Artificial

Intelligence Methods

H N Muliauwan, D Prayogo, G Gaby and K Harsono

Department of Civil Engineering, Petra Christian University, Surabaya - Indonesia
Corresponding author: [email protected]

Abstract. Concrete is one of the most used materials in buildings today; yet, predicting the
accurate concrete compressive strength remains challenging because of the highly complex
relationship between its mixture. An accurate method of predicting concrete compressive
strength can provide a significant advantage to the construction material industry, particularly
within the concrete material industry. Many methods can be used to build the prediction model
of concrete compressive strength. However, the traditional methods have so many
shortcomings, including expensive experimental costs and the inability to formulate an
accurate complex relationship between the components of a concrete mixture with the
compressive strength. To overcome this issue, this study applies multiple artificial intelligence
(AI) methods to find the most accurate input and output relationships within concrete mixtures.
The three types of AI methods that will be used in this study are artificial neural networks
(ANN), support vector machine (SVM), and linear regression (LR). This study uses 1030 data
samples from concrete compressive strength tests obtained from University of California,
Irvine, to demonstrate the use of AI prediction models. The obtained results of the simulation
show that these artificial intelligence methods can build predictive models without conducting
any expensive experiments in the laboratory with good accuracy.
Keywords: artificial intelligence, artificial neural network, concrete compressive strength,
prediction, support vector machine

1. Introduction
Concrete material in recent years has become the most used material in buildings because it has
proven to be able to show stable high strength characteristics. In addition to the four basic ingredients
of conventional concrete, namely cement, coarse aggregate, fine aggregate, and water, a number of
other cement composition materials have been used, such as fly ash, blast furnace slag, and chemical
mixtures such as superplasticizer [1]. The use of additional cement composition materials also
produces economic benefits because Portland cement is the most expensive component of concrete
mixtures. On the other hand, the use of additives in concrete is popular because it can increase the
workability, durability, and strength of the concrete. These new additions to the concrete mixture
introduce new dimensions to the modeling of the concrete compressive strength, which results in
increased complexity. However, traditional modeling methods that are usually used to predict concrete
behavior are less able to produce accurate prediction results.
Generally, strength tests are carried out 7–28 days after the concrete casting process. The 28-day
waiting period required to carry out such tests can cause the next construction process to be slightly
delayed [2-3]. However, ignoring testing will limit the quality control on a large and complicated
construction scale. Therefore, fast and reliable predictability of concrete compressive strength is very
important for quality control, even in the early design stages [4]. If adjusting the proportion of the

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

mixture can be done immediately when the strength of the concrete is not compatible with the
specifications, there will be savings in terms of time and construction costs. The early prediction of
concrete strength is extremely important in estimating the time needed to open the concrete formwork,
project scheduling, and quality control [4].
Because the relationship between components and the character of concrete is nonlinear, the
modeling using a mathematical approach becomes very complicated. The empirical equation presented
in the current standard code for estimating compressive strength is based on testing concrete without
additional cement composition material. Understanding the relationship between the concrete
composition and its strength is very important for optimizing a concrete mixture [5-6]. Significant
amounts of research have done by using experimental tests that usually require a lot of time and
money. For this reason, a new modeling system is needed that is not dependent on experimentation but
still can predict the concrete compressive strength with good accuracy.
Recently, artificial intelligence (AI) methods are increasingly being used in solving classification
and regression problems because AI methods have proven to be more accurate than conventional
methods [7-12]. This research develops AI techniques to accurately predict the concrete compressive
strength from various components. Experimental data were taken from a machine learning repository
held at the University of California, Irvine (UCI), that was collected by Yeh [1] to predict the
compressive strength at HPC. AI modeling is carried out in Clementine 12.0 using three predictive
techniques, namely artificial neural networks (ANN), support vector machines (SVM), and linear
regression (LR). This study uses two model construction, namely single model and ensemble model.
In ensemble models, a combination of classifiers can compensate for errors made by the individual
classifiers on different parts of the input space. The strategy used in ensemble systems is to create
many classifiers and combine their outputs such that the combination improves upon the performance
of single classifiers in isolation.

2. Artificial Intelligence Methods

2.1 Artificial Neural Network (ANN)

ANN is a computational model that tries to simulate the structure and functional aspects of biological
neural networks. Various ANN applications can be categorized as classification models or regression
models. The use of ANN specifically to predict the compressive strength of concrete has been
intensively studied [1, 6]. The researchers also explored the use of ANN to build concrete compressive
strength models that were more accurate than the regression model. The most used ANN model is the
multilayer perceptron (MLP) model. In the MLP model, the input layer contains a set of sensory input
nodes, one or more hidden layers that functionate to compute, and the output layer contains one
computational node that represents the concrete compressive strength. The most commonly used and
effective learning algorithm for training the MLP model is a back-propagation (BP) algorithm. The
activation process of each neuron can be seen in Equations (1) and (2).
𝑛𝑒𝑡𝑘 = ∑𝑤𝑘𝑗𝑜𝑗 (1)
𝑦𝑘 = 𝑓 (𝑛𝑒𝑡𝑘) (2)
Where 𝑛𝑒𝑡𝑘 is the activation of neurons to k, j is the set of neurons in the previous layer. 𝑤𝑘𝑗 is the
weight of the connection between neurons k and neurons j, 𝑜𝑗 is the output of neurons j, and 𝑦𝑘
represents the output that is usually calculated in sigmoid and logistical transfer functions. Illustration
of ANN structure can be seen through Figure 1.

2
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

Figure 1. Illustration of ANN structure.

2.2 Support Vector Machine (SVM)

SVM was first introduced by Vapnik (1995) [13]. SVM has been used in many civil engineering
applications, and in recent years, it has often been used to predict concrete compressive strength [14-
16]. In this study, support vector regression (ε-SVR), which is a variation of SVM, is used to build an
input-output model from concrete. SVM uses an objective function that allows the function estimation
process to occur, as illustrated in Figure 2. When nonlinear space occurs, the kernel radial-based
function (RBF) is selected as a kernel function in SVM because it can provide better results than the
other kernels.

Figure 2. Illustration of hyperplane separation and determination of support vector by SVM.

The following model underlies the functional relationship between one or more independent
variables with the response variable:
𝑦(x) = wT ϕ(x)+b (3)
where x Є R, y Є R, and ϕ(x): R is the process of mapping to higher dimensional feature space.
n
𝑵
In SVM for regression analysis, a data set of {Xk, Yk}𝒌=𝟏 , objective functions can be formulated as
follows:
𝟏 𝟐
Min. Jp (w,e) = wT wϕ + c∑𝑵
𝒌=𝟏 𝒆 𝒌 (4)
𝟐

s.t 𝑦k = wT ϕ(x)+b+ek, k = 1,…,N (5)

where ek Є R is the error variable; c indicates the regularization constant.

2.3 Linear Regression (LR)

LR, an extension of the simple regression model, determines the relationship between a numerical
response variable and two or more explanatory variables [17]. It is used for the modeling mechanical
properties of construction materials. This study investigated the applicability of LR. The
computational problem addressed by LR is fitting a hyperplane to an n-dimensional space where n is

3
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

the number of independent variables. For a system with n inputs(or independent variables), X’s, and
one output (or dependent variable), Y, the general least square problem is to determine unknown
parameters of the linear regression model as illustrated in Figure 3.

Figure 3. Illustration of linear regression.

The general formula for LR models is shown in Equation 6.

Y = β0 + β1 X1+ β2 X2 + . . . + βn Xn + ε (6)
In the proposed model, βi is a regression coefficient (i= 1,2,3,…,n), X’s values represent concrete
attributes, ε is an error term, and Y is concrete compressive strength. Regression analysis estimates the
unbiased values of the regression coefficients βi against the training data set.

3. Experimental Setting
Experimental data were obtained from a machine learning repository held at the University of
California, Irvine (UCI) that was collected by Yeh [1]. A total of 1030 concrete samples evaluated by
various university research laboratories were used to test the prediction models of each AI method. All
tests were carried out on a 15 cm cylindrical concrete specimen using standard procedures. Table 1
shows the experimental data set of nine HPC variables used in this study.

Table 1. Variables that affect the compressive strength of concrete and its descriptive statistics.
Standard
Variables Unit Min Mean Max
deviation
X1: Cement kg/m3 102.0 281.17 540.0 104.51
X2: Blast-furnace slag kg/m3 11.0 107.28 359.4 61.88
X3: Fly ash kg/m3 24.5 83.86 200.1 39.99
X4: Water kg/m3 121.8 181.57 247.0 24.35
X5: Superplasticizer kg/m3 1.7 8.49 32.2 4.04
X6: Coarse aggregate kg/m3 801.0 972.92 1,145.0 77.75
X7: Fine aggregate kg/m3 594.0 773.58 992.6 80.18
X8: Age of testing Day 1.0 45.66 365.0 63.17
Y: HPC compressive strength MPa 2.3 35.82 82.6 16.71

Table 2 shows the performance measurement model used to estimate the accuracy of each
predictive method. This accuracy model is calculated based on how accurate it is between actual data
and the predicted results of the output variable. These four performance measurement models include
root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE)
and coefficient of correlation (R). RMSE measures the average deviation of each actual data point and
the predicted results. MAE calculates the average error using the absolute difference from the actual
data and the predicted results. The strength of the linear relationship between the two variables is

4
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

measured by R. Although it is similar to MAE which uses absolute difference in accuracy calculations,
MAPE has the advantage that it is not affected by the unit and the size of the predicted and actual
values, so that it can be more efficient in determining the relative differences between models. Here,
the value of R = -1 represents a perfect negative correlation, while the value of R = 1 indicates a
perfect positive correlation. The best model results are shown by the highest R values, and the lowest
RMSE, MAE, and MAPE values.

Table 2. Indicators of prediction model accuracy.

Performance measurement Mathematical formula

R=
n y  y' − ( y)( y')
Correlation coefficient (R)
n(  y 2 ) − (  y ) 2 n(  y ' 2 ) − (  y ' ) 2

1 n
 ( y − y' )2
Root mean squared error
RMSE =
(RMSE) n i =1
n
y − y'

Mean absolute percentage 1
MAPE =
error (MAPE) n i =1 y
n
 y − y'
1
Mean absolute error (MAE) MAE =
n i =1

The AI model of the Clementine 12.0 software is used to form an accurate prediction model.
Although this platform is relatively easy to use for the purposes of AI methods, a precise flow chart is
still needed to form an accurate prediction model. The five steps of forming the AI model are
described as follows.
1. Data input: This step is the first step where data collection will be taken.
2. Training and testing: Input data will be divided into 2 data groups, namely, training and testing.
Training is used to create prediction models that fit the data while testing is used to test prediction
models that are built. In this study, 70% of the data will be used for training, and 30% of the data
will be used for testing.
3. The learning process of AI prediction models with training data.
4. The process of testing AI prediction models with data testing.
5. Prediction results: The four accuracy indicators will be used to measure the performance of the
predicted results obtained for each model.
Figure 4 explains the flow diagram to building the AI prediction using a single model.

Figure 4. Flow diagram of the formation of AI prediction using a single model.

The simplest method of combining multiple classifiers is voting. In the cases of prediction, the
outputs of the individual classifiers are pooled. Then, the class which receives the largest number of
votes is selected as the final classification decision. In general, the numerical output can be determined
by different combinations of probability estimates [17]. By combining two or three different individual
classifiers, this study obtained four different ensemble classifiers. Ensemble of two different classifiers

5
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

included ANN+SVM, ANN+LR, and SVM+LR. Ensemble of three different classifiers included
ANN+SVM+LR. Figure 5 explains the flow diagram to build the AI prediction using ensembles
model.

Figure 5. Flow diagram of the formation of AI prediction using ensemble model.

4. Experimental Results
This study compares the ability of two construction models, namely single models and ensemble
models in predicting concrete compressive strength from 1030 concrete samples. Of the 1030 samples,
723 were used for training, and 307 were used for testing. ANN, SVM, and LR use the default
parameters of Clementine 12.0. As described previously, R, RMSE, MAPE, and MAE are used to find
out how accurate each prediction method is. The results of the analysis of the four indicators of each
method of testing data can be seen in Table 3.

Table 3. The results of the performance evaluation model of the prediction on testing data

Methods Testing Result

R RMSE MAPE MAE
(MPa) (%) (MPa)
ANN 0.909 7.176 19.93 5.550
SVM 0.779 10.914 32.59 8.687
LR 0.768 10.975 31.69 8.634
ANN+SVM+LR 0.855 9.039 26.99 7.103
ANN+SVM 0.883 8.360 25.29 6.622
ANN+LR 0.876 8.319 24.35 6.504
SVM+LR 0.774 10.862 32.32 8.607

From Table 3, it can be seen that ANN has succeeded in surpassing all existing methods in all four
indicators of accuracy. ANN gets an R-value of 0.909, higher than the other methods and is close to 1.
Meanwhile, ANN produces a lower error rate compared to all existing methods. ANN received 7,176
MPa, 19.93%, 5,550 MPa for RMSE, MAPE, and MAE, respectively. Meanwhile, other methods
generate RMSE, MAPE, and MAE numbers that are far adrift with ANN. Furthermore, Figures 6 and
7 show the predicted results of the single model construction of ANN, SVM, and LR.

6
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

(a) (b)
Figure 6. Comparison of prediction and actual results in testing data using (a) ANN method, and (b)
SVM method.

Figure 7. Comparison of prediction and actual results in testing data using LR method.

Figures 8 and 9 show the predicted results of the ensemble model, namely ANN + SVM + LR,
ANN + SVM, ANN + LR, and SVM + LR.

(a) (b)
Figure 8. Comparison of prediction and actual results in testing data using ensemble method (a)
ANN+SVM+LR and (b) ANN+SVM.

7
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

(a) (b)
Figure 9. Comparison of prediction and actual results in testing data using ensemble method (a) ANN
+LR and (b) SVM+LR.
As seen in the figures above, if the prediction results get closer to the diagonal line, the prediction
results can be said to have higher accuracy. From Figure 6 to Figure 9, it can be concluded that the
ANN prediction results are more accurate than the other methods because the predicted results from
each point of ANN are closer to the diagonal line. To support the results of this study, we compare it
with the results of the previous study that has been done by Prof. Jui-Sheng Chou. Table 4 show the
predicted results of the best individual and ensemble models research by Prof. Jui-Sheng Chou.

Table 4. The results of the performance evaluation model of the prediction on testing data [17].
Dataset Predictive technique MAE (MPa) RMSE (MPa) MAPE (%)
SVM 3.75 5.59 12.03
Dataset 1
CART+SVM+LR
(Taiwan) 3.52 5.08 11.97
(stacking)
Dataset 2 MLP 4.00 5.40 6.81
(Chile) CART (bagging) 3.82 5.13 6.50
Dataset 3 MLP 4.28 5.81 10.51
(Hong Kong) MLP+SVM+LR (stacking) 4.70 6.02 11.75
SVM 1.31 1.73 2.95
Dataset 4
CART+SVM+LR
(South Korea) 1.09 1.51 2.11
(stacking)
Dataset 5 CART 5.42 7.11 12.55
(Iran) MLP+CART+LR (voting) 4.06 5.61 9.87

Specifically, the ensemble learning based models that per- formed the best, in terms of SI
values, over the five datasets were stacking-based CART+SVM+LR for Datasets 1 and 4,
bagging-based CART for Dataset 2, stacking-based MLP+SVM+LR for Dataset 3, and voting-
based MLP+CART+LR for Dataset 5[17]. Based on the latest research this method has a good
accuracy approaching the real results.

5. Conclusion
This study presents a comparative study between two construction models, namely the single model
and ensemble models, in predicting the concrete compressive strength based on 1030 concrete
samples. These samples were used to form a database that was used to create a prediction model as
well as to test the accuracy of the prediction model formed. Four accuracy indicators are used in
evaluating the performance of each method, including R, RMSE, MAPE, and MAE. Experiments were
carried out using Clementine 12.0 software. The finding showed that ANN achieved the best accuracy

8
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

for all performance measures because it can provide the most optimal performance when viewed from
the four indicators. This research succeeded in proving that AI methods are able to predict concrete
compressive strength without conducting experiments in the laboratory with good accuracy.

6. References
[1] Yeh I-C 1998 Modeling of strength of high performance concrete using artificial neural
networks Cement and Concrete Research, 28(12) 1797-1808.
[2] Prayogo D 2018 Metaheuristic-based machine learning system for prediction of compressive
strength based on concrete mixture properties and early-age strength test results Civil
Engineering Dimension 20(1) 21-29.
[3] Prayogo D, Cheng M-Y, Widjaja J, Ongkowijoyo H, and Prayogo H 2017 Prediction of
concrete compressive strength from early age test result using an advanced metaheuristic-
based machine learning technique Proceedings of the 34th International Symposium on
Automation and Robotics in Construction 856-863.
[4] Chou J S, Chiu C K, Farfoura M, and Al-Taharwa I 2011 Optimizing the prediction accuracy of
concrete compressive strength based on a comparison of data-mining techniques Journal of
Computing in Civil Engineering 25(3) 242-253.
[5] Cheng M-Y, Firdausi P M, and Prayogo D 2014 High-performance concrete compressive
strength prediction using Genetic Weighted Pyramid Operation Tree (GWPOT) Engineering
Applications of Artificial Intelligence 29 104-113.
[6] Yeh I-C 1999 Design of high-performance concrete mixture using neural networks and
nonlinear programming Journal of Computing in Civil Engineering 13(1) 36-42.
[7] Cheng M-Y, Prayogo D, and Wu Y-W 2019 Prediction of permanent deformation in asphalt
pavements using a novel symbiotic organisms search–least squares support vector regression
Neural Computing and Applications 31(10) 6261-6273.
[8] Cheng M-Y, Prayogo D, Ju Y-H, Wu Y-W, and Sutanto S 2016 Optimizing mixture properties
of biodiesel production using genetic algorithm-based evolutionary support vector machine
International Journal of Green Energy 13(15) 1599-1607.
[9] Cheng M-Y, Prayogo D, and Wu Y-W 2015 A self-tuning least squares support vector machine
for estimating the pavement rutting behavior of asphalt mixtures Soft Computing 23(17)
7755-7768.
[10] Hoang N-D, Liao K-W, and Tran X-L 2018 Estimation of scour depth at bridges with complex
pier foundations using support vector regression integrated with feature selection Journal of
Civil Structural Health Monitoring 8(3) 431-442.
[11] Prayogo D and Susanto Y T T 2018 Optimizing the prediction accuracy of friction capacity of
driven piles in cohesive soil using a novel self-tuning least squares support vector machine
Advances in Civil Engineering 2018.
[12] Tran T-H and Hoang N-D 2016 Predicting colonization growth of algae on mortar surface with
artificial neural network Journal of Computing in Civil Engineering 30(6) 04016030.
[13] Vapnik V N 1995 The nature of statistical learning theory: Springer-Verlag New York, Inc.
[14] Cheng M-Y, Prayogo D, and Wu Y-W 2014 Novel genetic algorithm-based evolutionary
support vector machine for optimizing high-performance concrete mixture Journal of
Computing in Civil Engineering 28(4) 06014003.
[15] Pham A-D, Hoang N-D, and Nguyen Q-T 2015 Predicting compressive strength of high-
performance concrete using metaheuristic-optimized least squares support vector regression
Journal of Computing in Civil Engineering 30(3) 06015002.
[16] Prayogo D, Wong F T, & Tjandra D 2018 Prediction of high-performance concrete strength
using a hybrid artificial intelligence approach MATEC Web of Conferences 203 06006.
[17] Chou J-S, Tsai C-F, Pham A-D, and Lu Y-H 2014 Machine learning in concrete strength
simulations: multi-nations Construction and Building Materials 73 771-780.

9
2nd International Conference on Sustainable Infrastructure IOP Publishing
Journal of Physics: Conference Series 1625 (2020) 012018 doi:10.1088/1742-6596/1625/1/012018

Acknowledgments
Authors wishing to acknowledge the present research is financially supported by The Ministry of
Research, Technology and Higher Education of the Republic of Indonesia under the Penelitian Dasar
Unggulan Perguruan Tinggi (PDUPT) 2019 Research Grant Scheme.

10