International Journal of GEOMATE, July, 2017, Vol.13, Issue 35, pp.

87-92
Geotec., Const. Mat. & Env., ISSN:2186-2990, Japan, DOI: http://dx.doi.org/10.21660/2017.35.6683

MODELLING OF CARBONATION OF REINFORCED CONCRETE

STRUCTURES IN INTRAMUROS, MANILA
USING ARTIFICIAL NEURAL NETWORK

*Richard M. De Jesus1, Joshua A. M. Collado2, Jemison L. Go3, Mike A. Rosanto4, and John L. Tan5
1
Faculty of Civil Engineering in De La Salle University (Manila), Philippines
2-5
Undergraduate Student of Civil Engineering in De La Salle University (Manila), Philippines

*Corresponding Author, Received: 15 June 2016, Revised: 29 Nov. 2016, Accepted: 25 Dec. 2016

ABSTRACT: Corrosion is a perennial problem in reinforced concrete structures, and is a serious concern
due to the deterioration that it causes to reinforced concrete members. Though regarded as having a minor
influence to corrosion compared to chloride-induced corrosion, carbonation is becoming a serious threat due
to continuous development of cities like Manila. Expectedly, as Manila continues to develop, carbon
emission shoots up to alarming proportions, calling out for studies to investigate and mitigate its effect to
human health and structures. Artificial Neural Network (ANN) is known for establishing relationships among
parameters with unknown dependency towards another variable, similar to the case of carbonation’s
dependency with age, temperature, relative humidity, and moisture content. Utilizing field-gathered
secondary data as training and testing parameter for back propagation algorithm, an ANN model is proposed.
Prediction of carbonation depth using ANN Model C421 showed reliable results. Validation of performance
of Model C421 was further checked by comparing its prediction with a different set of field-gathered
secondary data and results confirmed good agreement between prediction and measured values.

Keywords: Carbonation, Artificial Neural Network (ANN), model C421, reinforced concrete

1. INTRODUCTION was presented in the study of [4]. According to this

study, the level of CO2 has reached 90.78 million
Recent years had shown Philippines’ economy metric tons in 2013. By the beginning of 2015,
rising fast. As of first quarter of 2016, Philippine CO2 emissions has reached 97.91 million metric
economy grew close to 7% and considered to be tons [5].
the highest among 11 Asian countries as reported With this alarming rate of increasing CO2
in [1]. The same news agency, on May 16 2016, levels in Metro Manila, it is not only human health
reported that Philippines is the top performing that is at risk, but the health of reinforced concrete
country in Southeast Asia based on the report of (RC) structures as well. CO2 penetration in RC
Oxford Business Group (OBG) [1]. structures, known as carbonation, poses a great
According to the Philippine Statistics Authority threat to the structural health as its risk to corrode
(PSA), Philippines’ gross domestic product (GDP), reinforcing steel bars (rebar) is high. Structures in
for the 1st quarter of 2016, was 6.9 percent. It is urban/cities are most vulnerable to carbonation due
recorded to be higher than that of China’s 6.7% to high levels of CO2. And to raise the concern
and Vietnam’s 5.7% [2]. farther, only limited studies were done in the
The country’s economic rise is warmly country that investigates the effect of CO2 in RC
welcomed by Filipinos. But it calls for the country structures, and how fast it contributes to the
to also grow accordingly with such economic deterioration of structures. This deterioration
growth. An indicator that the country is lagging happens within RC members are hard to detect
behind with this growth, at some areas, is the from the outside which makes it threat even more
failure to anticipate overwhelming increase in alarming.
motor vehicles and its effects, in the past years. As
reported by [3], passenger car sale grew from a 2. THEORETICAL BACKGROUND
little over 20,000 in 2005 to over 63,000 in 2015
while that of commercial vehicles grew from over 2.1 Carbonation in Reinforced Concrete
10,000 to over 30,000 in the same period.
A quantification of effects of increasing car Carbonation is a chemical process that involves
volume, in terms of carbon dioxide (CO2) the reaction of some chemicals in concrete,
emission, specifically the cement hydration elements calcium

87
 International Journal of GEOMATE, July, 2017, Vol.13, Issue 35, pp.87-92

hydroxide (Ca(OH)2), and calcium-silicate- lowers the rate of carbonation according to [8]-[9]
hydrates (C-S-H), with CO2 from the atmosphere due to presence of high moisture in concrete pores.
that penetrates into concrete. The chemical Since diffusivity of CO2 is lower in water than in
reaction produces calcium carbonate (CaCO3) air, it is anticipated that CO2 diffusion in concrete
which can potentially lower the passivity of steel pores lowers with increasing humidity. Moreover,
bars, hence, contributing to corrosion. This [8]-[9] claimed that thin layer of water covering
chemical process can be represented by the the pores tend to block the intrusion of CO2.
following chemical equation:
2.2 Artificial Neural Network
CO2 + Ca(OH)2  CaCO3 + H2O (1)
Artificial Neural Network (ANN) is an
The study of [6] stated that carbonation effective tool for prediction of behavior or
involves the dissolution of CO2 with water in the phenomenon. Unlike other tools, it requires no
pores of concrete, before reacting with Ca(OH)2, assumption as to how independent variables are
to form carbonic acid (CaCO3). A quantification affected by a single or multiple independent
of how much of CO2 that penetrates into concrete variables. The only basic requirement it requires is
is attributed to carbonation is explained by [7]. It sufficient amount of data. ANN can be seen as a
claimed that only half (50%) of CO2 reacts with black box where input data are being fed to
Ca(OH)2. The other half reacts with C-S-H. produce an output data [12]. ANN is efficient and
The rate of how fast carbonation progresses in effective in modelling systems influenced by
concrete is not constant [8]-[9]. This is attributed multiple variables with unestablished
to continuous hydration of cement particles in interrelationships especially with incomplete data
concrete and with continuous reaction of CO2 with [12]. Reference [13] described ANN as made up of
Ca(OH)2. With less Ca(OH)2 available, less large number of highly interconnected processing
reaction with CO2 occurs. The continuous elements called neurons which work in unison to
hydration also densifies the microstructure of solve a problem.
concrete which lowers the carbonation diffusion Before ANN can establish multi-variable
rate further. interrelationships and come up with a prediction, it
Several factors were identified to influence needs to learn the system first which can be
carbonation. Major factors were water to cement achieved by feeding input data with set initial
ratio, degree of hydration, CO2 concentration of conditions. Reference [12] added that ANN
the surrounding air, moisture content, temperature, requires no functional relationship among
alkali content, and presence of damaged zones and variables since the process is data driven implying
cracks [10]. that it can adopt to the training data and captures
This study, however, will only consider age, the relationship between the input and output
temperature, relative humidity, and moisture variables.
content in modelling carbonation through artificial
neural network (ANN) due to availability of data. 3. NEURAL NETWORK MODELLING,
Obviously, age is of primary importance in DATA, AND RESULTS
terms of influence to carbonation. With lengthy
exposure to CO2, and extended duration of 3.1 Data and its preparation for ANN modelling
diffusion of CO2 in concrete, carbonation depth
increases. It is assumed in this study that Fick’s A secondary data, based from [14] was utilized
law of diffusion which relates carbonation depth for this study for the training of the neural network
with the product of carbonation coefficient and model. The validation of model was conducted
square root of time, is valid. using another secondary data based on the study of
As temperature increases, the diffusivity of [4]. Both studies conducted carbonation test to
gaseous CO2 increases [8]. The diffusivity is seen structures within the same vicinity – Intramuros,
as the rate of CO2 to react and fully blend with Manila.
cement hydrating elements. The study further Normalization of data was performed for these
added that the rise in diffusivity can be attributed two secondary data. Normalization is to address
to increased molecular activity. According to [11], the wide gap in terms of magnitude between field-
the rise in temperature translates to higher measured carbonation data and that of the input
solubility of chemical compounds. This is seen as variables. The process was done by simply getting
a process that potentially facilitates the carbonation the quotient between the data and the maximum
process faster. value for that particular parameter. This process
Relative humidity is seen to contribute to was implemented to all parameters – age,
carbonation in an opposite manner compared to temperature, relative humidity, moisture content,
that of temperature. Higher relative humidity and carbonation depth. After normalization, all

88
 International Journal of GEOMATE, July, 2017, Vol.13, Issue 35, pp.87-92

data range from 0 to 1. top three (3). Thus, it was decided that Model
C421 will represent the modelling of the
carbonation depth using ANN.
Prediction of carbonation depth against
3.2 Choosing the ANN architecture measured values is shown on Fig. 2 to 7. These
show slight deviation between prediction and
A sample ANN architecture is shown in Figure measured values showing that ANN closely
1. The first 2 brown circles (a.k.a. neurons) are the established the relationship between carbonation
inputs and could vary from 1 to N depending on depth and the carbonation parameters (age,
the number of variables considered to influence an temperature, relative humidity, and moisture
output (represented by red circle). The two blue content).
circle in between represents the neurons in the
hidden layer. Table 1. Comparison of six ANN architecture

The arrows connecting the neurons have ANN error training R testing R
corresponding weights which change as ANN Model
“learns” the relationship between input and output.
C411 0.034 0.811 0.967
This sample ANN architecture represents a feed-
forward back propagation network which was C421 0.011 0.942 0.969
adopted in this study. C431 0.010 0.949 0.922
C441 0.015 0.912 0.975
C451 0.014 0.923 0.989
C4101 0.016 0.931 0.959

The measured carbonation depths as presented in

Fig. 2 to 7 were taken from field-measured
carbonation depth.
Figure 1. A sample neural network

The neural network architecture adopted in this

study considered 4 inputs – age, temperature,
relative humidity, and moisture content; while
output was set at 1 – carbonation depth. One
hidden layer was also considered in this study as
more hidden layers tend to make the training and
learning with ANN more difficult and complex.
The number of neurons in the hidden layer,
however, was varied from 1 to 10 based on the
recommendation of [15]. Choosing the number of
neurons, for the final ANN model, in the hidden
layer was based on the following criteria: (1) the
training R (Pearson’s coefficient) must be 0.80 or Figure 2. C411 prediction v. actual
higher; (2) the testing R must be 0.90 or higher;
and (3) error must be lower than 5% (or 0.05);
Tansig (or log sigmoid) transfer (or activation)
function was adopted in the study with Levenberg-
Marquadt (LM) as the training algorithm. The rest
of the training parameters used were the default
values in MATLAB – the software used to execute
ANN process.
Comparison of these 3 criteria for 6 considered
ANN architecture was summarized in Table 1. Six
ANN architectures, of varying number of neurons
in the hidden layer (n =1, 2, 3, 4, 5, & 10), were
compared based on the above criteria. Based on all
3 criteria, all architectures met the requirements.
Ranking of ANN architectures based on the 3 Figure. 3 C421 prediction v. actual
criteria showed that Model C421 was always in the

89
 International Journal of GEOMATE, July, 2017, Vol.13, Issue 35, pp.87-92

Comparison between measured carbonation depths humidity, m.c. is moisture content, and depth is the
and the prediction of Model C421, as shown in Fig. carbonation depth
3 showed minimal variation. ANN model C421 is
considered to have learned the carbonation
behavior via Artificial Neural Network (ANN).
How carbonation depth varies with age,
temperature,
relative humidity, and moisture content is closely
predicted by C421 model.

Figure 6. C451 prediction v. actual

Figure 4. C431 prediction v. actual

Figure 7. C4101 prediction v. actual

3.3 Validation of ANN Model (Model C421)

To further evaluate the performance of model

C421, it is tested against the carbonation data
gathered from the study of [4]. Raw data from this
study were also normalized, similar to the process
Figure 5. C441 prediction v. actual adopted for [14]. ANN model C421 were adopted
and its optimized weights (shown in Table 3) were
Since all models showed close values to set as the initial weights for training the network
measured ones, choosing C421 on the basis of using MATLAB. Its biases, shown in Table 2,
testing R criteria is supported. were also adopted.

Table 2. Initial Biases for C421 Model During validation, C421’s training R is 0.9831
while its testing R is 1.0. The results of validation
can be seen in Fig. 8. This result confirmed that
Input to hidden Hidden to ouput Model C421 approximates the carbonation
-0.003 0.395 prediction with high reliability.
-3.012

Table 3. Initial Weights for C421 Validation

nodes Age Temp R.H. m.c. Depth

1 0.6 0.72 0.078 -1.07 3.30
2 -0.2 -8.97 -1.61 5.00 3.25
Note: nodes refer to hidden nodes, R.H is relative

90
 International Journal of GEOMATE, July, 2017, Vol.13, Issue 35, pp.87-92

[2] Remitio, R. (2016, May 19). PH is fastest-

growing economy in Asia, expands by 6.9% in
Q1. CNN Philippines. Retrieved on June 11,
2016 from http://cnnphilippines.com/news/
2016/05/19/philippines-fastest-growing-
economy-asia-gdp-q1.html
[3] Punongbayan, J. C., & Mandrilla, K. (2015,
August 15). Carmaggeddon: Why are there so
many cars in Metro Manila?. Rappler.
Retrieved on June 11, 2016 from
http://www.rappler.com/ views/imho/102701-
carmageddon-metro-manila
Figure 8. C421 validation [4] Collado, J., Go, J.L., Rosanto, M.A., Tan, J.L.,
and De Jesus, R.M. (2015). Employing
CONCLUSION Artificial Neural Network in Assessing
Various Factors Affecting Carbonation in old
The country’s remarkable economic growth structures. Thesis. De La Salle University,
has been applauded locally and internationally. But Manila.
some of the environmental effects of a rising [5] Philippine Carbon Dioxide Emissions. (2013).
economy are treated with less attention. Among Retrieved on June 11, 2016 from https://
these is the alarming increase in CO2 levels in the ycharts.com/indicators/philippines_carbon_di
environment and the potential hazard it poses to oxide_emissions.
reinforced concrete structures due to carbonation. [6] Burkan Isgor, O., and Razanqapur, A. G.
There are few studies conducted locally to (2004). Finite element modeling of couple
investigate, and analyze the effect of CO2 heat transfer, moisture transport and
intrusion in RC structures. This study was among carbonation processes in concrete structures.
the initial steps to understand the carbonation Cement and Concrete Composites, 26(1), 57-
phenomenon. Here, CO2 intrusion in concrete, 73.
through carbonation depth, is related to age, [7] Park, D. C. (2008). Carbonation of concrete in
temperature, relative humidity, and moisture relation to CO2 permeability and degradation
content using feed-forward backpropagation of coatings. Construction and Building
artificial neural network. Materials, 22(11), 2260-2268.
ANN model C421 were trained using a [8] Talukdar, S., Banthia, N., & Grace, J. R.
secondary data and found close prediction values (2012). Carbonation in concrete infrastructure
with measured carbonation depth. This was further in the context of global climate change – Part
validated by comparing the prediction values with 1: Experimental results and model
measured values from another secondary data. development. Cement and Concrete
Results showed that model C421 were able to Composites, 34(8), 924-930.
validate the reliability and accuracy of the model [9] Yoon, I. S., Copuroglu, O., & Park, K. B.
as depicted by close values between prediction and (2007). Effect of global climatic change on
measured carbonation depths. carbonation progress of concrete. Atmospheric
Environment, 41(34), 7274-728.
ACKNOWLEDGEMENTS [10] Houst, Y. F. & Wittman, F. H. (2002). Depth
profiles of carbonates formed during natural
We thank Engr. Ronaldo S. Gallardo, and his carbonation. Cement and Concrete Research,
former student thesis advisees, for allowing the use 32 (12), 1923-1930.
of data from their study in the development (& [11] Kolio, A., Pakkala, T. A., Lahdensivu, J., &
training) of the ANN model in this paper. Kiviste, M. (2014). Durability demands
related to carbonation induced corrosion for
REFERENCES Finnish concrete buildings in changing climate.
Engineering Structures, 62-63, 42-52.
[1] Jiao, C. (2016, May 16). PH is best economy [12] Oreta, A. W. C. (2004). Simulating size effect
in Southeast Asia – Oxford Business Group. on shear strength of RC beams without
CNN Philippines. Retrieved on June 11, 2016 stirrups using neural networks. Engineering
from Structures, 26(5), 681-691.
http://cnnphilippines.com/business/2016/05/13 [13] Lu, C., & Liu, R. (2009). Predicting
/ oxford-business-group-economy-philippines- carbonation depth of prestressed concrete
southeast-asia.html under different stress states using artificial

91
 International Journal of GEOMATE, July, 2017, Vol.13, Issue 35, pp.87-92

neural network. Advances in Artificial Neural

Systems,2009,1-8.
[14] Bata, K. J., Beltran, B., Sacaguing, A., &
Copyright © Int. J. of GEOMATE. All rights
Gallardo, R. (2013). Carbonation profile of
reserved, including the making of copies unless
old and new reinforced concrete structures in
permission is obtained from the copyright
proprietors.
[15] Hect-Nielson, R. (1989). Theory of back
propagation neural networks. Proceedings of
the International Joint Conference on Neural
Networks, 593-611.

92