Cement & Concrete Composites 50 (2014) 73–81

Contents lists available at ScienceDirect

Cement & Concrete Composites

journal homepage: www.elsevier.com/locate/cemconcomp

Statistical modelling of carbonation in reinforced concrete

A. Silva a, R. Neves b, J. de Brito c,⇑
a
IST, Technical University of Lisbon, Av. Rovisco Pais, 1049-001 Lisbon, Portugal
b
ESTBarreiro, Polytechnic Institute of Setúbal, R. Américo da Silva Marinho, 2839-001 Lisbon, Portugal
c
Department of Civil Engineering and Architecture, IST, Technical University of Lisbon, Av. Rovisco Pais, 1049-001 Lisbon, Portugal

a r t i c l e i n f o a b s t r a c t

Article history: As time passes the properties of concrete change as a result of its interaction with the environment and
Received 23 May 2013 durability is affected. Reinforcement corrosion is singled out in various studies as being mainly respon-
Received in revised form 21 November 2013 sible for reinforced concrete degradation. Concrete alkalinity protects the reinforcement bars from corro-
Accepted 2 December 2013
sion but the carbonation phenomenon significantly contributes to the destruction of their passive
Available online 11 December 2013
coating, thus favouring the corrosion onset. Therefore, concrete carbonation is considered an important
problem both in Civil Engineering and in Materials Science. This study’s main objective is to try to quan-
Keywords:
tify the contribution of potential conditioning factors to concrete carbonation’s rate. This study addresses
Reinforced concrete
Durability
the statistical modelling of the concrete carbonation phenomenon, using a large number of results (964
Statistical modelling case studies), collected in the literature. A computational method (multiple linear regression analysis) is
Multiple linear regression used to define a mathematical model that can estimate the carbonation coefficient as a function of a set of
conditioning factors. These models allow the estimation of the carbonation coefficient, and consequently
the carbonation as a function of the variables considered statistically significant in explaining this phe-
nomenon. Two distinct models are proposed to suit predictions to two environmental exposure relative
humidity ranges.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction [60] reports that concrete carbonation is one of the main phenom-
ena to initiate the process of reinforcement corrosion. Carbonation
Reinforced structures are subject to deterioration over time. The is characterized by a physical–chemical process in which a series of
degradation of their mechanical characteristics jeopardizes their chemical reactions occur in the presence of carbon dioxide (CO2),
functional capacity. This degradation may occur due to in-service which fosters the reduction of pH in concrete. CO2 penetrates con-
conditions, exposure to aggressive environmental agents, and crete predominantly through a diffusion mechanism. This penetra-
inadequate use or maintenance conditions [9]. tion and carbonation reaction occurs gradually, leading to a
Neville [44] notes that only rarely is concrete degradation due carbonated layer (limited by the so-called carbonation front) that
to a single cause. The concrete degradation mechanisms can be increases in thickness over time [6].
physical, chemical or mechanical, and the chemical and physical It is generally agreed that carbonation does not occur in the
phenomena may be synergetic [49]. The cause of deterioration of same way in all mixes, nor does it occur in all circumstances;
reinforced concrete that deserves most attention is reinforcement different mixes will exhibit distinct carbonation and the same
corrosion [11,56,41]. In fact, it is one of the most important patho- mix exposed to different environments will not show the same
logical manifestations to affect reinforced concrete structures, and carbonation [40]. This study’s main objective is to try to under-
is difficult to intervene or repair [37]. stand the conditioning factors so as to explain concrete carbon-
The alkalinity of concrete protects the reinforcement from cor- ation. For this purpose a large number of related studies were
rosion until chemical or physical changes occur that enable exter- analysed, 17 of which were selected (because of their more
nal aggressive agents to act [28]. According to Hussain and Ishida complete data), amounting to 964 case studies. The statistical
[22], two main agents initiate reinforcement corrosion by destroy- modelling of the concrete carbonation phenomenon was
ing its passive coating: carbonation and chloride ingress. Tuutti performed using multiple linear regression. Mathematical
models were proposed for estimating the carbonation coefficient,
and consequently the carbonation as a function of the
⇑ Corresponding author. Tel.: +351 21 8419709; fax: +351 21 8497650. variables considered statistically significant in explaining this
E-mail addresses: [email protected] (A. Silva), [email protected] phenomenon.
(R. Neves), [email protected] (J. de Brito).

0958-9465/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.cemconcomp.2013.12.001
74 A. Silva et al. / Cement & Concrete Composites 50 (2014) 73–81

2. Background a statistical technique, innovative in the field of carbonation

modelling.
When reinforced concrete was first developed and used the
main characteristic of concrete to be controlled was its compres- 3. Statistical modelling
sive strength, which for a long time was considered the safest de-
sign specification [20], while durability was relegated to a second 3.1. Assumptions of the statistical model and selection of the variables
level. However, concrete does not have an infinite service life and
its durability will fundamentally depend more on its composition The statistical modelling of the concrete carbonation phenome-
than on its mechanical strength [43]. non was performed using multiple linear regression. Regression
The durability of reinforced concrete has been widely studied in analysis is one of the most widely used statistical techniques for
recent years. Andrade and Dal Molin [2] report that research re- studying the behaviour of a dependent variable as a function of
lated to the service life prediction of reinforced concrete structures other variables responsible for that behaviour, called independent
in terms of carbonation-induced corrosion is very active, using variables [35]. In multiple linear regression analysis the relation-
mathematical models. However, developing these models may ship between the dependent variable and the independent vari-
pose some difficulties, such as the number and scatter of the fac- ables is generically given by Eq. (1):
tors intervening in the carbonation process, lack of information
y ¼ B0 þ B1  x1 þ B2  x2 þ ::: þ Bk  xk þ e ð1Þ
and the difficulty and length of time implicit in the validation of
the models [1]. where y represents the dependent variable, B0, B1, . . ., Bk the regres-
A prediction model must be reliable, take into account all rele- sion coefficients, x1, x2, . . ., xk the independent variables and e the
vant factors of the mechanism to be modelled and have a sensitiv- random errors of the model.
ity that suits the precision of the input parameters [40]. Modelling In the definition of the multiple linear regression models, the
carbonation is a delicate issue since although several factors carbonation coefficient is the dependent variable. The independent
should be taken into account users expect research to deliver sim- variables analysed in this study can be grouped as:
ple and user-friendly models [42].
In their critical analysis of the various carbonation prediction  Factors inherent to concrete – type of cement (clinker percent-
models, Carmona and Helene [7] observed that there are models age, type I (practically inert) percentage of additions, percent-
that still need developing since they present theoretical inconsis- age of pozzolanic additions, percentage of latent hydraulic
tencies, and that the values obtained by different models may be additions); binder content; clinker content; water/binder ratio;
significantly distinct. Yazigi [65] refers that some models are enor- water/clinker ratio; 28-day mechanical strength; admixture
mously complex to apply, e.g. define some input parameters, mak- content; type of admixtures; slump.
ing it difficult to confidently estimate the service life of structures  Curing and moulding conditions – relative humidity and tem-
in-service. In fact there are very complex models that do not have perature of the curing environment; curing extent; compaction
an analytical solution [4,33]. type.
Most of the carbonation prediction models [60,47,3,19] generi-  Exposure (environmental) condition – relative humidity and
cally assume that the carbonation depth can be estimated through temperature of the exposure environment; carbon dioxide con-
the product
pffiffi of a carbonation coefficient by the square root of time tent; exposure class; protection from the action of rain; expo-
ðh ¼ k  t Þ. Kropp et al. [31] says that, notwithstanding the defi- sure to action of salts.
ciencies identified in the theoretical approach behind the model,
it does provide satisfactory results in stable environmental condi- Five of the studied variables are categorical (non-numerical)
tions, i.e. in laboratory. Papadakis et al. [45] report that the equa- and need to be codified. The variable compaction was codified in
tion yields scattered results when applied to concrete with no two categories: it takes the value 1 for normal compaction and
protection from the action of climatic agents, since the changes the value 1 for self-compacting concrete. As for the variable protec-
in relative humidity cause variations in the carbonation depth. tion against the action of rain, it was codified as follows: if concrete
On the other hand, Wierig [62] concludes that the classical model is protected from the rain the variable takes the value 1; if not it
is sufficiently accurate for natural atmospheric conditions, as long is equal to 1. The same occurs with the variable exposure to salts: if
as the concrete is protected from the action of rain. the concrete is exposed to salts it takes the value 1; if not it is
Nevertheless, several authors argue that the model is suitable to equal to 1. Finally, for the variable type of admixtures added to
estimate the evolution of the carbonation depth over time under the mix, four categories are considered: 0 if no admixture has been
natural conditions [14,38,40,59]. In this case the influence of added; 1 for plasticizers or superplasticizers; 2 for air-entraining
humidity on the carbonation rate may be considered in the carbon- admixtures; and 3 if more than one type of admixture has been
ation coefficient. added.
In the various existing models there is not a consensus on the The fifth categorical variable is exposure class. The categories for
way the carbonation coefficient is determined. This coefficient is this variable were established based on the exposure classes re-
fundamentally a durability indicator that comprises all the vari- lated to reinforcement corrosion due to concrete carbonation, de-
ables relating to the environmental severity and the characteristics fined in EN 206-1 [12]. This standard defines four classes for
of the concrete itself [10,38]. The quantification of this coefficient carbonation-induced corrosion: XC1, XC2, XC3 and XC4. The XC1
is usually hard since it depends on many factors, making the mod- environmental class stands for permanently dry (e.g., buildings’
elling of carbonation depth evolution a complex task. interior) or permanently wet (e.g., totally immersed) concrete.
Most carbonation models are semi-empirical, i.e. their develop- The XC2 environmental class stands for concrete with long periods
ment starts from a theoretical basis (e.g. Fick’s first law) and is in contact with water (e.g., rainwater drainage systems). Environ-
completed by fitting the required parameters to experimental re- ments with moderate humidity (e.g., concrete in open air
sults ([13,48,52,54]). However, the empirical part of these models structures sheltered from rain) correspond to class XC3 and with
is based in the results of one study or a limited number of studies. dry–wet cycles (e.g., concrete in open air structures not sheltered
The main purpose of this research is to propose a simple model from rain) correspond to class XC4. The value 1 is assigned to con-
for concrete carbonation, founded in a large number of data and in crete under XC1 environmental conditions, 2 to concrete in XC3
 A. Silva et al. / Cement & Concrete Composites 50 (2014) 73–81 75

environmental conditions and 3 to concrete exposed to XC2 and 4. Results

XC4 environmental conditions. SPSS (Statistical Package for the So-
cial Sciences) was used to build the model. There are various meth- Once the data have been post-treated the statistical modelling
ods to select the variables to be included in the model [57]. The can begin. First, the case studies (14.3% of the whole sample) that
Stepwise method was used here, which enabled only the statisti- were found to be influential points or outliers were removed. In
cally significant independent variables to be included in the model. this study all the points whose standard deviation values relative
Multicollinearity effects, which occur when the independent vari- to the average were higher than 3 were considered influential; in
ables are correlated and jeopardize the whole regression analysis practice this means that all the points whose real values of the car-
[50], are thus eliminated. In this selection method the predictors bonation coefficients are very different from the values predicted
of the dependent variable are entered in the model by ranking of by the model were removed.
relevance.

3.2. Data used in the statistical modelling 4.1. Influence of relative humidity on concrete carbonation

This study’s objective was to statistically model the concrete The presence of CO2 in concrete is a necessary condition for car-
carbonation phenomenon, based on previous related studies. bonation but is not sufficient on its own. In fact, the relative
These studies were analysed and data was collected from 17 of humidity of the exposure environment plays a very important role
them [46,25,26,27,34,52,5,51,32,30,17,8,58,21,53,24,39]. Each of in the carbonation phenomenon. Bakker [3] states that the diffu-
these studies analyses various mixes, with different characteris- sion of CO2 in a liquid medium is around 104 times slower than
tics and varying curing and environmental exposure conditions. in a gaseous medium, and therefore in saturated concrete the pen-
A sample of 964 case studies was obtained. The data sample com- etration of CO2 is practically nil; on the other hand, there is no
prises concrete with different binders and binder contents, water- aqueous medium, or electrolyte, in a dry concrete for Ca(OH)2
binder ratios ranging from 0.24 to 1.0, compressive strengths and gaseous CO2 to dissolve and for carbonation itself to occur.
ranging from 9 to 128 MPa, curing periods from 7 to 91 days, Therefore the humidity of concrete has antagonist effects on con-
slumps ranging from 20 to 300 mm. Around 50% of the data set crete carbonation. A study by Roy et al. [55] concluded that when
are from concrete exposed to natural (in service) conditions while the humidity increases from 52% to 75% there is a significant in-
the other 50% are from concrete exposed to accelerated carbon- crease in the carbonation depth, but it decreases for humidity be-
ation conditions. Thus CO2 concentration ranged from 0.02% to tween 75% and 84% and increases again for humidity higher than
50%. 92%. Helene [18] reports that the greatest carbonation rates occur
One of the commonest problems in data collection is missing at relative humidity values between 60% and 85%. Wolf and Dal
values, i.e. when some information relative to the phenomenon Molin [64] obtained the greatest carbonation depths for a relative
being studied is omitted [36]. In the overall sample of 964 humidity of 70%.
cases some data values relating to the variables (described in The state of the art information suggests that the carbonation
the previous section) that are to be modelled are missing. Since rate increases with the relative humidity up to a given point and
the case studies were collected from different works it is then tends to zero as concrete nears saturation. The model pro-
natural that there is information in one study but missing posed here is linear and therefore prevents a function (a straight
in another. The missing values may skew the statistical model- line) from changing course. In general terms this means that the
ling; therefore they cannot simply be ignored and must be use of this variable in a linear function may compromise the mod-
post-treated. el’s performance. If a second order function were to be used, it
There are various statistical techniques that can solve the would be possible to obtain models with results closer experimen-
problem of missing values. One approach is to perform the tal values but also more complex.
statistical modelling with the complete data only or eliminate It is therefore considered extremely relevant to split the total
the variables whose percentage of missing values is too high. sample into two, in terms of the relative humidity of the exposure
This approach is the simplest and most direct way of dealing environment. Analysis of the relationship between carbonation
with missing values but it has several disadvantages, the most depth and relative humidity shows that there is a great scatter of
obvious being that in many situations the sample is reduced to results for the sample; however, the maximum values of carbon-
an inadequate size [15]. However, there are procedures that ation depth are obtained for a relative humidity of approximately
handle the input of missing data from the sample. Of these 70%. Therefore two distinct models are obtained: the first one with
the simple imputation and multiple imputation methods stand a sample comprising the cases in which the relative humidity is
out. 70% or less (616 case studies); the second one with the cases where
In this study each missing value within the sample is analysed. the relative humidity is higher than 70% (210 cases). The variables
In some cases likely values are imputed to the missing data based under study are characterized in Tables 1 and 2, considering the
on the other characteristics of concrete and of the data in the study two samples after removing the influential points.
from which the values were taken. When there is not enough infor- In the model proposed for the first sample it is found that the
mation to draw any conclusion, the widely used mean substitution variables considered by ranking of relevance are: the CO2 content
method (simple imputation method) was chosen [61,63] because it of the exposure environment; the exposure class; and the concrete
is easy to implement. In this method the missing predictor values 28-day compressive strength. Table 3 presents a summary of the
are replaced by the mean value of the variable under analysis, so model, showing the coefficients used in the analysis of the statisti-
that the complete data of the variable can be used to determine cal validity of the regression model. The analysis of the results
the value by default. shows that 71.2% of the variability of carbonation coefficient is ex-
It was not considered appropriate to use more complex meth- plained by the three variables used, while the other 28.8% is due to
ods in this study, which, even though they might be statistically other causes that were not analysed in this study. It is therefore
more acceptable, do not give any guarantee that the output values concluded that the model is statistically significant.
are more realistic or adequate. Many of these methods lead to a Fig. 1 presents the correlation matrix between the variables in-
significant loss of the generalization capacity of the model, some- cluded in the model. The explanatory variables do not have a
thing that is to be avoided. strong inter-correlation, and this was expected since otherwise
76 A. Silva et al. / Cement & Concrete Composites 50 (2014) 73–81

Table 1 Table 3
Characterization of the variables under study for the sample with relative humidity of Summary of the multiple linear regression model for the sample with relative
the environment of 70% or less, after removing the outliers. humidity of the environment of 70% or less.

Numerical values Average Range of Modeld R R2 R2adjusted Square of the mean square error
value results
1 0.825a 0.681 0.680 8.862
Binder content (kg/m3) 376.7 [200–600] 2 0.838b 0.703 0.702 8.856
Clinker content (kg/m3) 295.8 [66–503.4] 3 0.844c 0.712 0.711 8.428
Water/binder ratio 0.5 [0.24–1]
a
Water/clinker ratio 0.7 [0.35–2.5] Independent variables: (Constant), CO2.
b
Independent variables: (Constant), CO2, Exposure class.
Curing conditions c
Independent variables: (Constant), CO2, Exposure class, 28-day compressive
Relative humidity (%) 92.8 [60–100]
strength.
Temperature (°C) 20.5 [20–30] d
Dependent variable: Carbonation coefficient.
Percentage of admixtures by cement weight 0.4 [0–2.5]
28-day Compressive strength (MPa) 50.3 [8.8–80.7]
Slump (mm) 114.9 [20–300]
Exposure conditions higher than the critical value of F (which is 2.60). Therefore, the
Relative humidity (%) 65.3 [50–70] model is shown to have a statistically significant explanatory capa-
Temperature (°C) 20.8 [20–40]
bility. Furthermore, the significance value (0.000) is lower than 5%
Carbon dioxide content 20.8 [0,02–50]
(p-value allowed), thus proving that the independent variables
Type of cement
considered are statistically significant in explaining the dependent
Clinker (percentage) 79.5 [20–100]
Latent hydraulic additions (percentage) 8.4 [0–80] variable. Therefore it is important to know the coefficients of the
Pozzolanic additions (percentage) 6.1 [0–39,3] multiple linear regression. Table 5 presents the linear regression
Practically inert additions (percentage) 3.3 [0–50] coefficients (B) of the final model. It is found that for all indepen-
Number of curing days 17.2 [7–91] dent variables the significance value is lower than the p-value
Categorical variables Mode Range of (5%), which demonstrates that they are all able to explain the
results
dependent variable.
Compaction 1 [1 to 1]
Protection from the action of rain 1 [1 to 1]
It is then possible to define expression (2) to determine the car-
Exposure to salts 1 [1 to 1] bonation coefficient when the relative humidity of the exposure
Type of admixture added to the mix 1 [0–3] environment is 70% or less:
Exposure class 1 [1–3]
kd ¼ 0556  c  3602  X  0148f c þ 18; 734 ð2Þ
kd – carbonation coefficient for RH 6 70% (mm/year0.5).
Table 2 c – CO2 content (%).
Characterization of the variables under study for the sample with relative humidity of
X – exposure class (1- XC1; 2- XC3; 3- XC4; exposure class XC2
the environment higher than 70%, after removing the outliers.
corresponds to situations where the relative humidity of the
Numerical values Average Range of exposure environment is above 70%; in Eq. (2) X is replaced
value results
by number 3 for exposure class XC4, and class XC2 is never con-
Binder content (kg/m3) 342.8 [230–580.7] sidered, as it makes no physical sense);
Clinker content (kg/m3) 281.7 [141.1–529.2]
fc – 28-day compressive strength (MPa).
Water/binder ratio 0.5 [0.26–0.84]
Water/clinker ratio 0.7 [0.28–1.2]
In the model proposed for the second sample (RH > 70%) it is
Curing conditions
Relative humidity (%) 91.2 [65–100] found that the variables considered by ranking of relevance
Temperature (°C) 21.0 [20–23] are: the CO2 content of the exposure environment; the clinker
Percentage of admixtures by cement 1.0 [0–3] content; and the concrete 28-day compressive strength. Table 6
weight presents the summary of the model, showing that 91.7% of the
28-day Compressive strength (MPa) 44.5 [17.5–127.5]
Slump (mm) 117.7 [30–270]
variability of the carbonation coefficient is explained by the three
variables considered, while the other 8.3% is due to other causes
Exposure conditions
Relative humidity (%) 80.2 [72–90]
not analysed here. It is thus found that the model is statistically
Temperature (°C) 17.1 [9–28.9] significant and that there is a very strong correlation between
Carbon dioxide content 1.1 [0.03–5] its variables.
Type of cement Fig. 2 presents the correlation matrix between the variables in-
Clinker (percentage) 82.3 [54.3–95] cluded in the model. Table 7 presents an analysis of variance of the
Latent hydraulic additions (percentage) 0 [0–0] global model (ANOVA table), where it is found that the value of F
Pozzolanic additions (percentage) 12.5 [0–39.3]
(Fisher–Snedecor test) of the model is much higher than the critical
Practically inert additions (percentage) 1.6 [0–6]
Number of curing days 21.8 [7–28] value of F (which is 2.60). Therefore, the model has a statistically
significant explanatory capability. Furthermore, the significance
Categorical variables Mode Range of
results value (0.000) is lower than 5% (p-value allowed), thus proving that
Compaction 1 [1 to 1] the independent variables considered are statistically significant in
Protection from the action of rain 1 [1 to 1] explaining the dependent variable. Therefore it is important to
Exposure to salts 1 [1 to 1] know the coefficients of the multiple linear regression. Table 8 pre-
Type of admixture added to the mix 1 [0–3]
Exposure class 3 [1–3]
sents the linear regression coefficients (B) of the final model. It is
found that for all independent variables the significance value is
lower than the p-value (5%), which demonstrates that they are
all able to explain the dependent variable.
there would be a problem of multicollinearity (described above). It is then possible to define expression (3) to determine the car-
Table 4 shows an analysis of variance of the global model (ANOVA bonation coefficient when the relative humidity of the exposure
table). The F value (Fisher–Snedecor test) of the model is much environment is higher than 70%:
 A. Silva et al. / Cement & Concrete Composites 50 (2014) 73–81 77

R2 = 0.681 R2 = 0.348 R2 = 0.340

Carbonation
coefficient

R2 = 0.594 R2 = 0.311
Exposure
2
R = 0.681 environment
CO2

28-day com-
R2 = 0.348 R2 = 0.594 pressive R2 = 0.262
strength

R2 = 0.340 R2 = 0.311 R2 = 0.262

Fig. 1. Matrix of correlation between the variables included in the model (sample with relative humidity of the exposure environment of 70% or less).

Table 4 Table 6
ANOVA table of the model for the sample with relative humidity of the environment Summary of the multiple linear regression model for the sample with relative
of 70% or less. humidity of the environment higher than 70%.

Modelb Variation – sum Freedom Mean F Sig. Modeld R R2 R2adjusted Square of the mean square error
of the squared degree squared a
deviations (FD) deviation 1 0.917 0.841 0.840 3.174
2 0.956b 0.913 0.913 2.347
3 3 0.958c 0.917 0.916 2.297
Regression 107569.149 3 35856.383 504.854 0.000a
a
Residual 43466.243 612 71.023 Independent variables: (Constant), CO2.
b
Total 151035.391 615 Independent variables: (Constant), CO2, Clinker content.
c
Independent variables: (Constant), CO2, Clinker content, 28-day compressive
a
Independent variables: (Constant), CO2, Exposure class, 28-day compressive strength.
strength. d
Dependent variable: Carbonation coefficient.
b
Dependent variable: Carbonation coefficient.

kw ¼ 3:355 c  0:019 C  0:042 f c þ 10:83 ð3Þ

The other variables (described in Section 3.1) were excluded
kw – carbonation coefficient for RH > 70% (mm/year0.5). from the model because they were not statistically significant,
c – CO2 content (%). for a 5% significance level, to explain the carbonation coefficient.
C – clinker content (kg/m3). However, it is found that three of the excluded variables are statis-
fc – 28-day compressive strength (MPa). tically significant to explain the 28-day compressive strength,

Table 5
Table relative to the regression coefficients of the model for the sample with relative humidity of the environment of 70% or less.

Modela Linear regression coefficients Standardized coefficients t Significance B

B Standard deviation associated to the regression coefficients Beta
3
(Constant) 18.734 1.785 10.493 0.000
CO2 0.556 0.024 0.834 23.383 0.000
Exposure class 3.602 0.482 0.198 7.473 0.000
28-Day compressive strength 0.148 0.033 0.154 4.464 0.000
a
Dependent variable: Carbonation coefficient.
78 A. Silva et al. / Cement & Concrete Composites 50 (2014) 73–81

R2 = 0.841 R2 = 0.103 R2 = 0.324

Carbonation
coefficient

Exposure
environment R2 = 0.149
R2 = 0.003
CO2
R2 = 0.841

R2 = 0.103 R2 = 0.003 R2 = 0.442

Clinker con-
tent (kg/m3)

R2 = 0.324 R2 = 0.149 R2 = 0.442

28-day com-
pressive
strength

Fig. 2. Matrix of correlation between the variables included in the model (sample with relative humidity of the exposure environment higher than 70%).

tion of these explanatory variables. As a matter of fact, for this

Table 7 sample it is found that 75.1% of the variability of the 28-day com-
ANOVA table of the model for the sample with relative humidity of the environment
pressive strength is explained by these variables.
higher than 70%.

Modeld Variation – sum of Freedom Mean F Sig. fc ¼ 104:18w=b  0:652d  0:115C f þ 76:311 ð4Þ
the squared degree squared
deviations (FD) deviation fc – 28-day compressive strength (MPa).
w/b – water/binder ratio ().
3
Regression 12083.396 3 4027.799 763.111 0.000c Cf – clinker ratio in the binder (%).
Residual 1087.295 206 5.278 d – curing time (days).
Total 13170.690 209
c
Independent variables: (Constant), CO2, Clinker content, 28-day compressive Once the carbonation coefficient has been defined, it is possible
strength. to rewrite the carbonation depth expression as in (5), as a function
d
Dependent variable: Carbonation coefficient. of the relative humidity of the exposure environment.
 pffiffi
h ¼ kd ptffiffi RH < 70%
ð5Þ
h ¼ kw t RH > 70%
which are, by ranking of relevance: the water/binder ratio; the h – carbonation depth (mm).
curing time; and the content of clinker in the binder. Expression kd, kw – carbonation coefficients (mm/year0.5).
(4) can be defined to estimate the compressive strength as a func- t – CO2 exposure time (years).

Table 8
Table relative to the regression coefficients of the model for the sample with relative humidity of the environment higher than 70%.

Modela Linear regression coefficients Standardized coefficients t Significance

B Standard deviation associated to the regression coefficients Beta
8
(Constant) 10.830 0.571 18.975 0.000
CO2 3.355 0.088 0.686 38.292 0.000
Clinker content 0.019 0.003 0.208 7.422 0.000
28-day Compressive strength 0.042 0.013 0.096 3.159 0.002
a
Dependent variable: Carbonation coefficient.
 A. Silva et al. / Cement & Concrete Composites 50 (2014) 73–81 79

obtained from the sample with relative humidity higher than

60 70% leads to a determination coefficient (R2) of 0.917.
Data Table 9 presents some statistical indicators that can relate the
Predicted kd (mm/year0,5)

50 observed values of the carbonation coefficient with those predicted

Equality line
by the models proposed, considering the separation in terms of the
40 95% prediction
relative humidity of the exposure environment. For both models
limit
the average value of the ratio between the observed and the pre-
30 dicted value was found to be close to 1, which indicates that the
values are relatively close in most case studies. As the determina-
20 tion coefficient suggests, the model for relative humidity higher
than 70% yields results closer to reality.
10 Various studies argue that the concrete carbonation rate funda-
mentally depends on the CO2 diffusivity, which in turn depends on
0 the concrete’s porous structure and on its eventual filling by other
0 10 20 30 40
substances, particularly water. This study focused on the variables
Observed kd (mm/year0,5) that are presumed to have influence on the process, and the ones
unarguably found to be statistically relevant were the CO2 content
Fig. 3. Relationship between the carbonation coefficient observed and that
predicted by the model (sample with relative humidity of the exposure environ- within the exposure environment and the 28-day compressive
ment of 70% or less). strength. For the sample with relative humidity of 70% or less, it
was found that the exposure class is also significant. On the other
hand, for the sample with relative humidity higher than 70% it was
found that the clinker content is the third explanatory variable. The
30
remaining variables were not included the model since, for a given
significance level (5%), they were regarded as not statistically
25 significant.
Predicted kw (mm/year0,5)

According to the results obtained, the most important factor in

20 concrete carbonation is the concentration of carbon dioxide. In
terms of influence on the carbonation rate in environments with
Data RH lower than 70% the next factor is concrete protection from
15
the action of water. In fact, even if the average RH of the environ-
Equality line ment is not high, provided concrete is in contact with water the
10
diffusibility of CO2 is low and consequently the carbonation rate
falls. In environments with RH higher than 70% this variable is less
5 95% prediction important since the humidity of the environment is itself enough
limit
to slow down the carbonation process.
0 Compressive strength is an explanatory variable in both groups
0 5 10 15 20 25 studied. This property somehow expresses the influence of the
Observed kw (mm/year0,5) concrete’s porous structure on the diffusion of carbon dioxide,
and thus on carbonation. Besides being generally associated with
Fig. 4. Relationship between the carbonation coefficient observed and that more compact concrete mixes, high compressive strength is also
predicted by the model (sample with relative humidity of the exposure environ- associated with increased clinker content. The clinker content in
ment higher than 70%).
concrete’s composition is a determining factor in the amount of
calcium hydroxide present, since calcium hydroxide results from
the hydration of dicalcium and tricalcium silicates, with emphasis
5. Discussion of results
on the latter. Since carbonation is the transformation of calcium
hydroxide into calcium carbonate, the more calcium hydroxide
First, it was found that the models proposed are statistically sig-
there is to react, the more carbon dioxide will have to be consumed
nificant. Figs. 3 and 4 present the relationship between the ob-
to carbonate concrete, thereby decreasing the carbonation rate.
served values and those predicted by the models, as a function of
In the group of concrete mixes in environments with HR > 70%,
the relative humidity of the exposure environment for two
the clinker content becomes especially relevant since in this case
humidity ranges. The model obtained from the sample with rela-
the concrete’s porous structure, represented in the model by the
tive humidity of 70% or less leads to a determination coefficient
compressive strength, proves to be less important. In fact, regard-
(R2) of 0.712, which indicates strong correlation between the
less of whether the porous structure is more or less open, the
values obtained and predicted. On the other hand, the model

Table 9
Statistical indicators relative to the determination of the carbonation coefficient using the two models proposed.

Statistical markers Cases in which the relative humidity of the exposure Cases in which the relative humidity of the exposure environment
environment is 70% or less is higher than 70%
Observed values (OV) Predicted values (PV) Observed values (OV) Predicted values (PV)
Average 17.051 17.035 7.352 7.397
Standard deviation 15.671 13.217 7.938 7.597
I.C. at 95% ±1.238 ±1.044 ±1.074 ±1.028
Average OV/PV 1.052 0.994
Cases where OV/PV > 1.3 25.8% 14.3%
Cases where OV/PV < 0.7 33.8% 27.1%
Cases where 0.8 6 OV/PV 6 1.3 40.4% 58.6%
80 A. Silva et al. / Cement & Concrete Composites 50 (2014) 73–81

diffusion of CO2 is influenced by it being filled with water. There- acknowledged that in some cases the relationship between the
fore, the most important factors are the driving force of the diffu- dependent variable, such as the CO2 content [23] and the compres-
sion process (the CO2 content) and the availability of carbonation- sive strength [29], and the explanatory variable is nonlinear, it is
prone material. The latter is represented in the model by the clin- believed that this is offset by the fact that other variables are also
ker content. considered in the model, rather than the single effect of a given
The association of each model’s explanatory variables with the variable being modelled. Alternatively, a nonlinear model could
dependent variable is in agreement with the theoretical assump- have been used, based on a polynomial or even an exponential
tions that explain the influence of each of them. The carbon dioxide function. However, for that analysis to be viable it would have been
content is directly linked to the carbonation coefficient, i.e. the car- necessary to know before-hand the expression that best describes
bonation rate increases with increasing carbon dioxide content. the influence of each variable on the concrete’s carbonation phe-
The exposure class, the compressive strength and the clinker con- nomenon. That expression could have been defined through an
tent are inversely associated with the carbonation coefficient, i.e. iterative process, by trying various equations until one is achieved
the carbonation rate declines as these variables rise. The variable that is fully capable of appropriately expressing the modelled
exposure class is categorical and it is higher when the concrete is process.
less protected from the water action or the environmental humid- The method proposed is simple to apply and understand but is
ity cycles it is exposed to. not without uncertainty. This does not hinder its application since
it can be used as a tool to approximate reality, and provide data for
subsequent research and development [16]. Namely, the proposed
6. Conclusions model can be useful within a mix design process, helping to narrow
the range of variables such as target compressive strength or clin-
This work is a first approach to the statistical modelling of the ker content as it is believed to be also suitable to estimate the ser-
phenomenon of concrete carbonation. Mathematical models have vice life in existing structures, where there is no experimental
been tentatively established to be applied to the estimation of assessment of carbonation depths and there is information on con-
the carbonation coefficient and then of the carbonation depth, with crete properties and construction (curing) procedures, provided
the goal of making service life predictions. The statistical modelling that is associated to reliability concepts. Furthermore, in future
was based on data collected from 17 studies, leading to a sample studies, and through the acquisition of new data, it will be possible
comprising 964 case studies. to extend the study to other cases. Ideally, it would be interesting
The total sample was divided into two groups. The first con- to perform an experimental programme of a sufficiently wide
sisted of those cases where the relative humidity of the exposure scope to confirm our results or even calibrate the mathematical
environment was 70% or less. For this sample it was concluded that models proposed.
the statistically significant samples are the CO2 content of the
exposure environment, the exposure class (parameter X) and the Acknowledgements
28-day compressive strength. The second sample consisted of the
cases where the relative humidity of the exposure environment The authors gratefully acknowledge the support received from
was higher than 70%. In that group the significant variables are the ICIST Research Institute, IST, Technical University of Lisbon
the CO2 content of the exposure environment, the clinker content and of the FCT (Foundation for Science and Technology).
and the 28-day compressive strength.
The concrete mixes analysed had very distinct characteristics. References
Besides the mixes’ characteristics and exposure conditions, infor-
mation related to the carbonation depth and exposure time was [1] Andrade JJO. Contribution to the service life prediction of reinforced concrete
also collected. Knowledge of these two last parameters enabled structures affected by reinforcement corrosion: Initiation by chlorides. PhD
Thesis in Engineering, Engineering School, Federal University of Rio Grande do
the carbonation
pffiffi coefficient to be determined based on the relation- Sul, Porto Alegre, Brazil; 2001.
ship h ¼ k  t. Despite the limitations, often pointed out, this is the [2] Andrade JJO, Dal Molin DC. Considerations on the models employed in service
model most often used. life prediction of reinforced concrete structures: depassivation by chloride
ions. Engenharia Civil 2003;18:31–44 [in Portuguese].
The values obtained in this study are believed to be very plau-
[3] Bakker R. Initiation period. Corrosion of steel in concrete. Chapman & Hall;
sible. In the first place, the complexity of the phenomenon to be 1988. p. 22–5.
modelled causes any estimates based on simplifications to become [4] Balen K, Gemert D. Modelling lime mortar carbonation. Mater Struct
less accurate. Carbonation depends on many factors that may be 1994;27(7):393–8.
[5] Baroghel-Bouny V, Arnaud S, Gawsewitch J, Belin P, Olivier G, Bissonnette B,
intercorrelated in one way or another. It has been proven statisti- Ounoughi K. Vieillissement des bétons en milieu naturel: Une expérimentation
cally that the variables water/binder ratio and quantity of clinker pour le XXI siècle: IV. Résultats issus des prélèvements effectués sur les corps
in the binder are directly related to the concrete compressive d’épreuve de différents sites aux premières échéances de mesure, Bulletin des
Laboratoires des Ponts et Chaussées 249, LCPC; 2004. p. 49–100
strength. Therefore, there are several intercorrelated variables in [6] Carmona TG. Models of prediction of the depassivation of reinforcement in
the statistical modelling of carbonation, which may imply that a concrete structures subjected to carbonation. Master’s Dissertation in
relatively complex model must be built. Second, this study is based Engineering, Polytechnic School, University of São Paulo, São Paulo, Brazil;
2005 [in Portuguese].
on data collected from various works that dealt with a great diver- [7] Carmona TG, Helene PRL. Models of prediction of the depassivation of
sity of concrete characteristics and exposure conditions, besides reinforcement in concrete structures subjected to carbonation. Technical
other factors related to the conditions in which the test samples bulletin PCC/425 of the Polytechnic School of the University of São Paulo, São
Paulo, Brazil; 2006 [in Portuguese].
were made and the way the results were reported, which also [8] Chang CF, Chen J-W. The experimental investigation of concrete carbonation
influence the values obtained. Therefore, a large scatter on the re- depth. Cem Concr Res 2006;36(9):1760–7.
sults of the proposed model was observed (Figs. 3 and 4). However, [9] Ciampoli M. Time dependent reliability of structural systems subject to
deterioration. Comput Struct 1998;67(1–3):29–35.
based on the large number of data and range of variables in which
[10] Coutinho JS. Improvement of the durability of concrete by treatment of the
it is founded, it is thought that the proposed model is robust with moulding. PhD Thesis in Civil Engineering, Faculty of Engineering, University
regard to the possibility of its use in various scenarios. of Porto, Portugal; 1998 [in Portuguese].
Furthermore, the models proposed assume that the relationship [11] Du YG, Chan AHC, Clark LA. Finite element analysis of the effects of radial
expansion of corroded reinforcement. Comput Struct 2006;84(13–14):917–29.
between the variables and the carbonation coefficient is linear, [12] EN 206-1: 2004 Concrete. Part 1: specification, performance, production and
using linear functions in data approximation. Even though it is conformity. CEN; 2004.
 A. Silva et al. / Cement & Concrete Composites 50 (2014) 73–81 81

[13] FIB task group 5.6 model code for service life design. Bulletin 34, FIB; 2006. [39] Neves R, Santos JV, Monteiro A. Study on the influence of compacting on
[14] Galan I, Andrade C. Comparison of carbonation models. In: 3rd International concrete’s resistance to carbonation. In: 6th Portuguese-Mozambican
RILEM PhD student workshop on modelling the durability of reinforced Congress of Engineering, CD-ROM, Maputo, Mozambique; 2011 [in
concrete. Portugal: Guimarães; 2009. p. 41–49. Portuguese].
[15] Hair JF, Black WC, Babin B, Anderson RE, Tatham RL. Multivariate data analysis, [40] Neves R. The air permeability and concrete carbonation of concrete in
Prentice-Hall Publishers. 6th ed. New Jersey, USA: Englewood Cliffs; 2007. structures. PhD Thesis in Civil Engineering, Instituto Superior Técnico,
[16] Hed G. Service life planning of building components. In: 8th DBMC Technical University of Lisbon, Portugal; 2012 [in Portuguese].
international conference on durability of building materials and [41] Neves R, Branco FA, de Brito J. A method for the use of accelerated carbonation
components. Canada: Vancouver; 1999. p. 1543–51. tests in durability design. Constr Build Mater 2012;36:585–91.
[17] Heirman G, Boel V, Audenaert K, Schutter G, D’Hemricourt J, Desmet B, [42] Neves R, Branco FA, de Brito J. Field assessment of the relationship between
Vantomme J. Chloride penetration and carbonation in self-compacting natural and accelerated concrete carbonation resistance. Cem Concr Compos
concrete. In: International RILEM-JCI seminar on concrete durability and 2013;41:9–15.
service life planning: curing, crack control, performance in harsh [43] Neville A. Why we have concrete durability problems. SP-100-3, concrete
environments. Dead Sea; 2006. p. 13–23. durability, Katharine and Bryant Mather International Conference 1987,
[18] Helene PRL. Contribution to the study of corrosion of concrete reinforcement. American Concrete Institute, Farmington Hills, USA, p. 21–30.
PhD Thesis in Engineering, Polytechnic School, University of São Paulo, São [44] Neville AM. Properties of concrete. 4th ed. UK: Pearson Education Limited
Paulo, Brazil; 1993 [in Portuguese]. Editors; 1997.
[19] Helene PRL. Service life of concrete structures. Porto Alegre, Brazil: IV Iberian [45] Papadakis V, Vayenas C, Fardis M. A reaction engineering approach to the
American Congress of Pathology of Constructions; 1997, vol. 1, p. 1–30. problem of concrete carbonation. AIChE J 1989;35(10):1639–50.
[20] Ho D, Lewis R. The specification of concrete for reinforcement protection: [46] Papadakis V, Vayenas C, Fardis M. Physical and chemical characteristics
performance criteria and compliance by strength. Cem Concr Res affecting the durability of concrete. ACI Mater J 1991;88(2):186–96.
1988;18(4):584–94. [47] Parrott L. Discussion on chapter 1: mechanisms and their mathematical
[21] Hoppe Filho J. Effects of the addition of hydrated lime on the permeability of descriptions. In: 4th RILEM, international symposium on creep and shrinkage
oxygen and capillary absorption of concrete with high contents of mineral of concrete: mathematical modeling. USA: Evanston; 1986. p. 459–68.
additions. Master’s Dissertation in Civil Engineering, Federal University of [48] Parrott L. Design for avoiding damage due to carbonation-induced corrosion.
Santa Maria, Santa Maria, Brazil; 2008 [in Portuguese]. TC104/WG1/Panel 1, CEN; 1992.
[22] Hussain RR, Ishida T. Development of numerical model for FEM computation [49] Pauletti C. Comparative analysis of the procedures for accelerated carbonation
of oxygen transport through porous media coupled with micro-cell corrosion tests. Master’s Dissertation in Civil Engineering, Federal University of Rio
model of steel in concrete structures. Comput Struct 2010;88(9–10):639–47. Grande do Sul, Porto Alegre, Brazil; 2004 [in Portuguese].
[23] Hyvert N, Sellier A, Duprat F, Rougeau P, Francisco P. Dependency of C–S–H [50] Pires JCM, Martins FG, Sousa SIV, Alvim-Ferraz MCM, Pereira MC. Selection and
carbonation rate on CO2 pressure to explain transition from accelerated tests validation of parameters in multiple linear and principal component
to natural carbonation. Cem Concr Res 2010;40:1582–9. regressions. Environ Model Software 2008;23(1):50–5.
[24] Imamoto K, Shimozawa K, Nagayama M, Yamasaki J, Nimura S. Air [51] Possan E, Andrade J, Molin D. Carbonation in urban environment: Correlation
permeability of concrete cover and its relation with carbonation progress between accelerated and non-accelerated tests. VIII Latin–American Congress
under long-term exposure test in Japan. In: RILEM conference concrete in of Pathology of Construction, vol. I, chapter VI, pp. 17–24, Asunción, Paraguay;
aggressive aqueous environments – performance, testing, and modeling. 2005 [in Portuguese].
Toulouse; 2009. p. 508–14. [52] Ribeiro AB, Machado A, Gonçalves A, Salta M. A contribution to the
[25] Jiang L, Lin B, Caib Y. A model for predicting carbonation of high-volume fly ash development of performance-related design methods. In: RILEM workshop
concrete. Cem Concr Res 2000;30(5):699–702. on life prediction and aging management of concrete structures. Paris; 2003. p.
[26] Jones M, Dhir R, Newlands M, Abbas A. A study of the CEN test method for 223–35.
measurement of the carbonation depth of hardened concrete. Mater Struct [53] Ribeiro S, Ribeiro AB, Gonçalves A. Estimated and real concrete resistances to
2000;33(2):135–42. carbonation. Application of the LNEC E465 specification, National Meeting
[27] Jones M, Newlands M, Abbas A, Dhir R. Comparison of 2 year carbonation Structural Concrete 2008, CD-ROM, Guimarães, Portugal; 2008 [in
depths of common cement concretes using the modified draft CEN test. Mater Portuguese].
Struct 2001;34(7):396–403. [54] RILEM TC 60-CSC Corrosion of steel in concrete. RILEM; 1988.
[28] Kazmierczak CS. Performance for the analysis of the efficiency of skins applied [55] Roy SK, Poh KB, Northwood DO. Durability of concrete – accelerated
on reinforced concrete structures with the objective of carbonation protection. carbonation and weathering studies. Build Environ 1999;34(5):597–606.
PhD Thesis in Engineering, Polytechnic School, University of São Paulo, São [56] Saetta A, Schrefler B, Vitaliani R. 2-D model for carbonation and moisture/heat
Paulo, Brazil; 1995 [in Portuguese]. flows in porous materials. Cem Concr Res 1995;25(8):1703–12.
[29] Khan M, Lynsdale C. Strength, permeability, and carbonation of high- [57] Samal AR, Mohanty MK, Fifarek RH. Backward elimination procedure for a
performance concrete. Cem Concr Res 2002;32:123–31. predictive model of gold concentration. J Geochem Explorat 2008;97(2–
[30] Khunthongkeaw J, Tangtermsirikula S, Leelawat T. A study on carbonation 3):69–82.
depth prediction for fly ash concrete. Constr Build Mater 2006;20(9):744–53. [58] Silva V. Action of carbonation in in-service reinforced concrete beams, built at
[31] Kropp J, Hilsdorf H, Grube H, Andrade C, Nilsson LO. Transport mechanisms natural and reduced scale. PhD Thesis in Structural Engineering, University of
and definitions. Performance criteria for concrete durability. RILEM; 1995. p. São Paulo, São Paulo, Brazil; 2007 [in Portuguese].
4–14. [59] Smolczyk H. Discussions to M. Hamada’s paper: Neutralization (carbonation)
[32] Lana E. Carbonation of the cover layer of prototypes of concrete with fly ash of concrete and corrosion of reinforcing steel. In: 5th international symposium
and lime. Master’s Dissertation in Civil Engineering, Federal University of on cement chemistry, Tokyo, Japan; 1968. p. 369–83.
Santa Maria, Santa Maria; 2005 [in Portuguese]. [60] Tuutti K. Corrosion of steel in concrete. Stockholm, Sweden: Swedish Cement
[33] Liang MT, Lin SM. Mathematical modeling and applications for concrete and Concrete Research Institute, Stockholm; 1982.
carbonation. J Mar Sci Technol 2003;11(1):20–33. [61] Wayman JC. Multiple imputation for missing data: what is it and how can I use
[34] Lo Y, Lee H. Curing effects on carbonation of concrete using a phenolphthalein it? Annual Meeting of the American Educational Research Association,
indicator and Fourier-transform infrared spectroscopy. Build Environ Chicago; 2003. p. 1–16.
2002;37(5):507–14. [62] Wierig H. Long-time studies on the carbonation of concrete under normal
[35] Mata J. Interpretation of concrete dam behaviour with artificial neural outdoor exposure. RILEM seminar on the durability of concrete structures
network and multiple linear regression models. Eng Struct under normal outdoor exposure. Hannover; 1984. p. 239–49.
2011;33(3):903–10. [63] Wilson FBAA. Treatment of missing values in a data set. Graduation
[36] McKnight PE, McKnight KM, Sidani S, Figueredo AJ. Missing data: a gentle Dissertation in Statistics, Faculty of Sciences, Eduardo Mondlane University,
introduction (methodology in the social sciences). 1st ed. New York, Maputo, Mozambique; 2010 [in Portuguese].
USA: Guilford Press; 2007. [64] Wolf J, Dal Molin DCC. Carbonation of mortars and concrete. Intervening
[37] Mehta PK. Durability of concrete – fifty years of progress? In: 2nd factors. Symposium on Pathology of Buildings: Prevention and Recovery,
International conference on durability of concrete. Montreal, Canada; 1991. Federal University of Rio Grande do Sul, Porto Alegre, Brazil; 1989. p. 118–33
p. 1–31. [in Portuguese].
[38] Monteiro I, Branco FA, de Brito J, Neves R. Statistical analysis of the [65] Yazigi R. Evaluation of carbonation in reinforced concrete viaducts. Master’s
carbonation coefficient in open air concrete structures. Constr Build Mater Dissertation in Aeronautical Infrastructure Engineering, Aeronautical
2012;29:263–9. Technology institute, São Paulo, Brazil; 2008 [in Portuguese].