Content uploaded by Ibrahim G Ogunsanya
Author content
All content in this area was uploaded by Ibrahim G Ogunsanya on Oct 17, 2024
Content may be subject to copyright.
Exploring machine learning to study and predict the chloride threshold
level for carbon steel reinforcement
Nicolas Maamary , Ibrahim G. Ogunsanya
*
University of Toronto, Department of Civil and Mineral Engineering, 35 St. George Street, Ontario, M5S 1A4, Toronto, Canada
ARTICLE INFO
Keywords:
Machine learning
Chloride threshold level
Reinforced concrete
Corrosion
Cement and supplementary cementitious
materials
ABSTRACT
Chloride-induced corrosion of steel reinforcing bar (rebar) is the primary cause of deterioration in reinforced
concrete structures, posing a signicant infrastructure challenge. The chloride threshold level (CTL) of rebar,
which represents the critical amount of chloride needed to initiate active corrosion, is crucial in corrosion and
service life prediction models. However, substantial uncertainties and a multitude of inuencing factors, along
with the absence of a universally accepted testing framework, hinder the achievement of a consistent CTL range
for service life models and complicate comparisons of published values. This study addresses these challenges by
developing multiple machine learning models to predict CTL, considering 21 carefully selected features. A
comprehensive database of 423 data points was compiled from an exhaustive literature review. Seven machine
learning models—linear regression, decision tree, random forest, K-nearest neighbors, support vector machine,
articial neural network, and an ensemble model—were developed and optimized. The ensemble model ach-
ieved superior prediction performance, with a mean absolute error of 0.218 % by weight of binder, root mean
square error of 0.321 %, and a coefcient of determination of 0.751 on unseen CTL data. Partial dependence
plots generated using the support vector machine model quantied the effect of each feature on CTL. The random
forest model identied SiO₂ binder content and exposed rebar area to chlorides as the most inuential factors.
The study also examined the impact of supplementary cementitious materials (SCMs), nding that only blast
furnace slag positively affected CTL.
1. Introduction
1.1. Chlorides induced corrosion
Chloride-induced corrosion of steel reinforcing bar (rebar) is the
leading cause and most costly form of deterioration of reinforced con-
crete structures [1]. Chlorides found in reinforced concrete stem from
two distinct sources: an external origin, such as marine environment and
de-icing salts, and an internal origin, such as chloride-based accelerator
compounds used to speed up cement hydration reactions [2] or seawater
used for mixing concrete to aid sustainability [3].
Upon accessing the structure, the chloride ions within cement-based
materials can be present as free or bound chlorides. Free chlorides refer
to the soluble chloride ions in the cement pore solution, while bound
chlorides refer to the chloride ions being captured by the cement hy-
drates (i.e., solid phases). Bound chlorides can either be physically
adsorbed or chemically bound to the cement hydration phases [4].
Chemical binding mainly results from the reaction between chlorides
and hydrated calcium aluminate phases (AFm) in the cement where the
formation of Friedel’s salt and Kuzel’s salt takes place through an ionic
interaction mechanism [5]. Physical binding results from the physical
adsorption of chloride ions onto the calcium silicate hydrate (C-S-H)
phase due to its high specic surface, causing them to be in the chem-
isorbed layer on hydrated calcium silicates, C-S-H interlayer spaces, and
C-S-H lattice [6]. Friedel’s salt is also acknowledged as a potential
physical binder of chlorides, adding another layer of consideration [7].
Additionally, the involvement of ettringite in chloride binding is a
subject of intense debate, and a similar level of uncertainty surrounds
the role of portlandite in chloride binding [7].
Many researchers report that only free chloride ions in the cement
pore solution are responsible for corrosion initiation [8]. However,
bound chloride may also play a role due to the release of some bound
chloride ions into the pore solution under some specic conditions [9].
The readily reversible nature of some chloride-based hydration reactions
has been evident in short-term laboratory experiments [10], indicating
that bound chloride can effectively buffer chloride ion activity in cement
* Corresponding author.
E-mail addresses: nicolas.maamary@mail.utoronto.ca (N. Maamary), ibrahim.ogunsanya@utoronto.ca (I.G. Ogunsanya).
Contents lists available at ScienceDirect
Cement and Concrete Composites
journal homepage: www.elsevier.com/locate/cemconcomp
https://doi.org/10.1016/j.cemconcomp.2024.105796
Received 18 July 2024; Received in revised form 14 September 2024; Accepted 9 October 2024
Cement and Concrete Composites 154 (2024) 105796
Available online 10 October 2024
0958-9465/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (
http://creativecommons.org/licenses/by-
nc-nd/4.0/ ).
pores. The cement pore solution has a high pH (~12.5–13.8), attributed
to the abundance of portlandite and the presence of alkali metal hy-
droxides [11]. In such an alkaline environment without chlorides, car-
bon steel undergoes passivation, resulting in remarkably low corrosion
rates [12]. When chlorides are present in sufcient quantity adjacent to
the rebar, its passive lm (i.e., iron oxide layer protecting against active
corrosion) is locally destroyed (i.e., depassivation), resulting in localized
corrosion known as pitting corrosion.
1.2. Chloride threshold level
Existing models for chloride-induced corrosion in reinforced con-
crete structures cover two successive stages: (i) corrosion initiation -
refers to the period during which aggressive chlorides penetrate the
concrete cover while the rebar is in its passive state, accumulating to a
point where they cause a local breakdown of the passive lm on the
rebar surface initiating active corrosion; and (ii) corrosion propagation -
characterized by continued active corrosion where initiated pits grow
[13]. Chloride threshold level (CTL), also known as critical chloride
level, conceptually indicates the quantity of chloride in the concrete
adjacent to the rebar surface that initiates active corrosion. Although
CTL was initially conceptualized as a practical engineering idea to
decouple the corrosion initiation–propagation phases [14], it has pro-
gressively evolved into a universally accepted parameter that holds
signicant inuence in both research and practical applications [15].
CTL is typically expressed in different forms: (i) as the total chloride
content relative to the weight of the cement or concrete they are present
in; (ii) as the free chloride content by weight of cement or concrete or
molar concentration in the cement pore solution [16]; and (iii) as the
ratio of chloride ion activity to the pH of the pore solution - i.e.
[Cl
−
]/[OH
−
] [17]. It was found, however, that the latter is not a reliable
index [18], especially because the inhibitive properties of the concrete
(e.g., chloride binding) cannot be expressed only by the OH
−
concen-
tration in the cement pore solution. In addition, as mentioned above,
bound chloride presents some corrosion risk, making total chloride by
weight of cement the most effective presentation of CTL, representing
the overall potential aggressive ion content relative to the total potential
inhibitor content [16].
CTL stands out as one of the pivotal input parameters in corrosion
initiation models, particularly in scenarios involving chloride ingress. In
service life modelling, CTL has been renowned to be one of the most
decisive input parameters [19]. Thus, a precise determination of CTL is
crucial as it enables the proper design and evaluation of the service life
for new reinforced concrete structures, and facilitates the prediction of
the residual service life for existing structures [20]. Limiting the chloride
content in concrete as outlined in codes and standards, has been a
fundamental durability design requirement for reinforced concrete
structures for many years [21]. For instance, RILEM [22] considers the
value of 0.4 % for the total chloride ions by weight of cement to be an
appropriate threshold for carbon steel rebar, and it has become a
prevalent practice, in both European countries and North America, to
restrict the acceptable chloride content to that value [23]. Similarly, to
account for uncertainties surrounding chloride-induced corrosion, con-
servative values like 0.2 % or 0.4 % of total chlorides by weight of
cement have been employed for CTL in predicting the corrosion-free life
[24,25].
Even though it is common in engineering practice to adopt conser-
vative approaches, assuming uniformly a lower CTL value will lead to
inaccurate calculations and severe underestimation of the structure’s
predicted service life. For example, Marksete [26], demonstrated,
through probabilistic service life calculations, that the service life ex-
tends from 50 to 100 years when the characteristic value of CTL in-
creases from 0.34 to 0.67 % by weight of cement. Also, as per the
LIFE-365 model [27], a 20 % increase in CTL results in a 28 % extension
in the time required for corrosion initiation [28]. Minor uncertainties in
CTL are strongly amplied in the predicted time to corrosion, leading to
signicant uncertainties in service life predictions [29].
1.3. Variability in CTL
Reporting CTL as a single value or a narrow range and its accurate
prediction is quite complex for several reasons, which lead to a large
scatter in literature data for carbon steel rebar in concrete or mortar,
ranging over more than two orders of magnitude [15]. These reasons are
grouped into 3 major categories: (i) stochastic nature of corrosion, (ii)
specimen properties and exposure conditions, and (iii) methodology of
testing.
Stochastic nature of corrosion: steel passive lm breakdown and
resulting pitting corrosion initiation is a stochastic complex phenome-
non [30]. Aleatory uncertainty also arises from the intrinsic variability
in measuring CTL that cannot be eradicated by conducting more precise
measurements. This is reected in the natural variability observed in the
results of two identical experiments [15].
Specimen properties and exposure conditions: CTL is inuenced by a
great number of parameters, namely [16,29,31–34]: (1) cementitious
composition (2) specimen (i.e., reinforced concrete/mortar); degree of
hydration; (3) specimen resistivity; (4) environmental exposure condi-
tions such as the specimen moisture content, the relative humidity inside
the pores, the temperature of the environment, and whether the
experiment is conducted in a laboratory controlled environment or in
the eld; (5) cation accompanying chloride (e.g., NaCl, MgCl
2
, CaCl
2
,
KCl); (6) pH of the cementitious pore solution; (7) water to binder ratio;
(8) rebar corrosion potential (E
corr
); (9) rebar surface condition (e.g.
polished, mill scaled/as-received, sandblasted or cleaned); (10) rebar
metallurgical characteristics (e.g. steel composition); (11) roughness of
rebar surface (e.g. ribbed or smooth); (12) oxygen availability at the
rebar surface; (13) exposed rebar surface area to chlorides; (14) char-
acteristics of the steel-concrete or mortar interface such as entrapped air
voids, cracks or gaps adjacent to the rebar surface, and bleed-water
zones below the rebar. Some of those parameters inuence each other,
making them interdependent. For instance, most of these parameters
impact the nature of the rebar passive layer, which in turn inuences the
electrochemical potential exhibited by the rebar when it is not
controlled externally. Also, the water/binder ratio and degree of hy-
dration affect the porosity of the specimen, and thus, the availability of
moisture and oxygen at the rebar. Another example would be that both
porosity and the moisture content are reected by the resistivity, which
is mainly inuenced by the water/binder ratio, binder type, and the
conditions of curing the concrete/mortar.
Methodology of testing: the methodology used to detect corrosion
initiation and obtain CTL greatly inuences CTL and can present a
higher dominant impact on the resulting CTL than the parameters under
test [16,20,29,31,33,35–38]. Some of these are (using continued
numbering system from previous paragraph): (15) electrochemical
technique used to detect rebar depassivation; (16) method of intro-
ducing chlorides to the specimen; (17) method of detecting chloride
content in the specimen; and (18) test setup used.
Regarding (15), CTL tests are carried out either in free corrosion
potential or under external polarization inuence (e.g., potentiostatic
control). For tests carried out in free corrosion potential, identifying
depassivation can be through (i) visual inspection of local rust; (ii)
monitoring the open circuit potential (OCP) of the rebar - depassivation
is considered to have occurred when a noticeable shift in OCP is
observed or when OCP drops below a specied value; (iii) monitoring
macrocell corrosion current between the rebar of interest and another
metal acting as a cathode or microcell corrosion current directly from
the rebar of interest obtained commonly using Linear polarization
resistance measurements (LPR) or Electrochemical impedance spec-
troscopy (EIS) - depassivation is considered to have occurred when a
signicant increase in current is observed or when measured current
exceeds a certain threshold. For tests carried out in potentiostatic po-
larization conditions, the current obtained from polarizing the rebar is
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
2
monitored until a sharp increase is observed which indicates depassi-
vation. It should also be noted that some researchers may use multiple
techniques simultaneously to detect depassivation such as monitoring
OCP and the corrosion current.
Regarding (16), chlorides could be mixed in, naturally diffused into
the specimen through pure diffusion or accelerated by capillary suction
under wetting and drying cycles, or alternatively accelerated through
the application of an electrical eld (chlorides migration). Regarding
(17), the most common methods used to measure chloride content that
resulted in depassivation and corrosion initiation are titration, spec-
trophotometric methods such as X-ray uorescence spectrometry (XRF),
or the use of ion-selective electrodes as well as pore solution expression.
Regarding (18), while continued effort is being made by some research
committees, such as ACI 222 committee, no standard test setup for
determining CTL has been established at this point. For this reason, the
corrosion cell arrangement varies between authors across the literature.
For instance, some of the setups that have been used follow the ASTM
G109 standard [39], accelerated chloride threshold setup [40], ETH
method, and Lollipop test setup [41], to name a few. This has caused
experimental uncertainties and obscurity and limited measurement
precision of CTL, contributing to the large variability of CTL data in the
literature.
Due to the 3 major categories discussed above, it remains tasking,
and nearly impossible, to compare the published values of CTL accu-
rately, or to adopt a consistent range of CTL values for engineering life
design models [20,37]. Consequently, CTL in codes of practice is dened
very conservatively [29]. Moreover, drawing denitive conclusions
about the primary factors inuencing corrosion initiation is challenging
due to signicant variations in experimental conditions across the
literature and the simultaneous variation of multiple factors, making it
difcult to isolate their individual impacts in various studies [42].
1.4. Existing CTL predictive models
In 2019, Shakouri and Trejo [35] used a hierarchical Bayesian
approach to construct a predictive model for CTL using 20 reported CTL
values from the literature. The developed model takes into consideration
the following inuencing factors as inputs: type and percent replace-
ment of SCM, water-to-binder ratio, and the surface condition and
roughness of the rebar. This model assumed a linear relationship be-
tween the input data and generated outputs, which is usually not the
case. A more advanced model was reported in 2021, where Zhu et al.
[43] combined articial neural network (ANN) with Kohonen
self-organized mapping (KSOM) for modelling CTL of rebar in reinforced
concrete from a sparse literature database. A total of 1840 CTL values
were gathered [44]: 1169 based on total chloride, 436 based on free
chloride, and 235 expressed as [Cl
−
]/[OH
−
]. The authors made several
assumptions that introduce major errors in the application of these data,
(i) they considered that all chloride in the concrete can be regarded as
free chloride and, thus, used reported CTL values based on total and free
chlorides interchangeably as total chloride, (ii) they used a simple
empirical equation that does not consider the interaction of the elec-
trolyte in the porous media within the solid hydrate phases of hardened
cement paste to convert CTL values expressed as [Cl
−
]/[OH
−
] into total
chlorides per weight of binder; and (iii) predicting missing data with
KSOM algorithm - the authors used 17 input parameters in their model
(rebar temperature in the concrete, rebar OCP at 0.5, 1, 5, and 8 ppm O
2
content, rebar passive lm breakdown potential, concrete pH, concrete
porosity, concrete water/binder, and cement oxides concentration for
SiO
2
, Al
2
O
3
, Fe
2
O
3
, CaO, MgO, Na
2
O, K
2
O, and SO
3
. The modelling
dataset comprised 1242 instances (a vector comprised of output and
input variables such as pH, temperature, etc.) corresponding to carbon
steel rebar in concrete or mortar experiments alone. However, 589 of the
used 1242 instances were incomplete (35 % of the input variables values
were missing), and were completed by training KSOM classication al-
gorithm to predict the missing values in the dataset. Following this, all
instances being complete were used to train a KSOM network again as a
means to reduce scattering and benet from dimensionality reduction,
and the weight input factors in each neuron of the KSOM map were fed
as inputs to the ANN regression model where the output of the network
was CTL. Finally, the trained KSOM-ANN was validated by comparing
predicted values with experimental ones in the evaluation database.
This approach, which involves predicting a large number of missing
values in the dataset using the rest of the data already present in it,
introduces an articial correlation in the training data, leading to
seemingly favorable model behavior that may not accurately reect
reality. To improve the reliability of the model, it is advisable to utilize a
high-condence dataset with minimal correlation between training in-
stances. This ensures that the model is trained on data that more
authentically represents the true complexities and variations within the
system.
The present work aims to narrow the gap in understanding the
impact of various inuencing factors on CTL and the selection of a
suitable CTL for a specic reinforced concrete structure. Until this point,
predictive CTL model has been limited to two publications, and the
adoption of machine learning to predict CTL has been limited to one
publication, which employed major assumptions leading to favorable
model/outcome, considered lesser inuencing factors than the present
work, and does not explore different algorithm to recommend the best
for future use. Therefore, multiple machine learning models are devel-
oped to predict CTL, considering its various inuencing factors. The
model’s prediction performance is examined to identify the best model
to be used for predicting CTL, eliminating the need for tailored CTL
determination experiments. In addition, a selected model is unfolded to
isolate each feature (input parameter) and quantitatively evaluate its
effect on CTL.
2. Methodology
2.1. Machine learning models and performance evaluation metrics
2.1.1. Machine learning model
The branch of machine learning used in this study is supervised
learning where the machine undergoes training using a large dataset
under a specic algorithm in order to develop a mathematical model
capable of forecasting outputs (labels) by analyzing the associated fea-
tures. The objective of supervised learning is to discover relationships
and mappings between features and labels [45]. Since CTL, the antici-
pated output, is a continuous variable that can take any value within a
numerical range, it is deemed as a regression problem, and regression
algorithms are adopted [46]. Six standalone models, linear regression,
K-nearest neighbors, support vector machine, decision tree, random
forest, articial neural networks, and one ensemble model (which
combines standalone models) were implemented. Although they have
been established in other elds, a brief summary of these models is
presented.
Linear regression (LR): This is the simplest model adopted. Multi-
variate linear regression establishes a linear relation between the
dependent variable (output) and one or more independent variables
(inputs). The coefcients for the independent variables are derived by
minimizing the mean squared error between the predicted and actual
values of the output [47]. Minimization is done using a closed-form
solution (i.e., no applied approximation/iteration).
K-Nearest Neighbors (KNN): This is an instance-based learning, stor-
ing all instances corresponding to the training dataset in n-dimensional
space. It does not build an internal model but rather predicts new data
points based on similarity measures with training data such as the
Euclidean distance function [48]. Employing a specied distance metric
to compute the distances between a given test instance and each
instance in the training set, these distances are sorted, and the K nearest
neighbors are identied (i.e., if distance ‘i’ ranks in the ith place, then its
corresponding instance is called the ith nearest neighbor). The
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
3
prediction output for the test instance is determined by averaging the
outputs of its K nearest neighbors, thus making predictions by consid-
ering the mean output of locally proximate instances in the training set
[49].
Support Vector Machines (SVM): This technique approximates the
nonlinear association between input variables and the output of a
dataset using an optimization strategy [50]. The regression version of
SVM is a regularized learning algorithm in reproducing Kernel Hilbert
spaces with what is called the
ε
-insensitive loss function [51]. Each
training sample is treated as a data point in a multidimensional feature
space and the loss function is used to establish a hyperplane (i.e.,
multidimensional plane), ensuring that the predicted response values for
each training sample deviate by at most
ε
from their observed values.
The hyperplane, along with their
ε
, forms an
ε
-insensitive tube, which
species the tolerance range within which prediction errors are ignored.
The optimization process involves minimizing the
ε
-insensitive tube
while encompassing most training samples. The hyperplane is then
represented in terms of a few support vectors which are training samples
lying outside the
ε
-insensitive tube boundary [52]. The outcome of the
training is a regression model capable of predicting response outputs for
new samples.
Decision tree (DT): This is a nonparametric technique that develops
decision rules based on the training dataset and presents the results in
the form of a dendrogram structure called decision tree [53]. Typically,
a decision tree consists of four components: root nodes, branches, in-
ternal nodes, and leaf nodes. The decision tree initiates at the root node
with outgoing branches connected to the internal nodes of the model,
which are systematically evaluated and lead to a series of leaf nodes
representing the model’s decision or prediction (numerical value) for a
specic combination of input features [54]. In the training phase, the
algorithm partitions/splits the input data at each node, generating split
function parameters and optimizing them in alignment with the training
set [55]. The initial step involves the decision tree making the best
possible split among all variables at the root node. As the splitting
process progresses, each subsequent node applies its split function to
new input data, continuing this recursive process until a leaf node is
reached. The present work used classication and regression tree
(CART) algorithm for the training process [56].
Random forest (RF): This grows and combines multiple uncorrelated
decision trees to create a forest using bagging and feature randomness
[57]. Bagging, short for bootstrap aggregating, is a parallelization
technique employed to coordinate multiple estimators, or decision trees
in this case, simultaneously. This involves generating random subsets of
the training dataset with replacement for each tree. Subsequently, each
estimator/tree is trained and makes predictions independently, and the
output from all trees is then averaged as the estimation of the random
forest [55].
Articial Neural Network (ANN): This is structured as a series of layers
containing multiple computational elements, termed neurons. This
design is inspired by the hierarchical organization of neurons in the
human brain. In an ANN, information is processed and propagated
through these layers which typically include an input layer, one or more
hidden layers, and an output layer, allowing for complex computations.
Neurons within the layers are connected by weighted links, and each
neuron processes input data and applies activation functions, contrib-
uting to the overall computation [58]. The base architecture of ANN
where each node in one layer connects to each node in the following
layer is called a multilayer perceptron (MLP) [48]. MLP employs the
"backpropagation" technique, identied as the most fundamental
building block in neural networks, to iteratively minimize the error by
adjusting the internal weight values during the model-building process
[59].
Ensemble methods: This involves combining predictions from multiple
trained models to enhance the overall performance and prediction ac-
curacy. The underlying idea is that combining diverse models can often
lead to better generalization and predictive accuracy compared to
individual models [60]. Bagging is one ensemble method that is used in
RF. Another ensemble method is “average voting” which combines
predictions from multiple regression models by simple averaging [61].
In contrast to bagging which uses random subset, the base models of
average voting are trained by employing the same training set. In this
work, the average-voting method with the combination of RF and SVM
models was used as the ensemble machine learning model.
2.1.2. Performance evaluation metrics
In order to quantitatively assess the prediction performance of the
machine learning models, three statistical parameters referred to as
performance evaluation metrics were used [55]. These parameters
fundamentally assess the cumulative error in CTL predicted by the
models compared to the actual values reported in the literature.
Mean Absolute Error: Equation (1) denes the mean absolute error
(MAE), which calculates the average of the absolute differences between
the actual and predicted values and assigns equal weight to all errors.
Smaller MAE values indicate more accurate prediction results.
MAE =1
n
n
i=1
|yi−
yi|(1)
Where n represents the number of data-records in the subset under
consideration, yi represents the actual (real) values and
yi represents the
predicted values.
Root Mean-Square Error: Equation (2) denes the root-mean-square
error (RMSE) used for assessing the disparity between values pre-
dicted by a model and the actual measurements. It shares similarities
with the MAE, but RMSE penalizes larger absolute values by assigning
them more weight than MAE. The magnitude of the deviation between
MAE and RMSE shows the extent of variance in the individual errors.
RMSE =
n
i=1
(yi−
yi)2
n
(2)
Coefcient of determination (R
2
): Equation (3) denes R
2
, which as-
sesses how well the predicted values from a model align with the actual
values. An optimal R
2
value is 1, signifying that the model can account
for the entire variability of the output.
R2=
nyi.
yi− (
yi)(yi)
n(yi2) − (yi)2
n(
yi2) − (
yi)2
2
(3)
In general, when the predicted values closely align with the actual
ones, the MAE and RMSE tend to approach zero. In contrast, an R
2
close
to 1 indicates a strong correspondence between the experimental and
predicted data.
2.1.3. Dataset division and model building steps
In machine learning, the original dataset is typically divided into
three subsets: the training set, the validation set, and the testing set [62].
The training set is the largest subset and is used to train the model. The
validation set is used to ne-tune the model’s hyperparameters (a
conguration setting of a model not learned from training data but set
before the training process) and to evaluate its performance during
training. The testing set is kept separate from both the training and
validation sets and is used to assess the model’s nal performance after
being trained on the validation and training set combined. The ideal
percentage of data allocated to the training set is generally found to be
between 40 % and 80 % [63]. Since the present work involved a
small-size dataset, an 80/20 split is used to divide labeled data into
subsets for modelling and testing. 80 % of the original dataset’s records
were randomly allocated for training and validating the machine
learning model. This modelling phase encompassed ne-tuning the
model hyperparameters and applying the 10-fold cross-validation
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
4
process (described later). At the end of this stage, the model is trained on
the full 80 % subset. The remaining 20 % of the dataset was reserved for
testing purposes. Despite the random partitioning, deliberate efforts
were made to ensure that the training and testing datasets accurately
mirrored the characteristics of the original dataset.
Overall, the building of each model involved three major steps: (i)
data processing, (ii) 10-fold cross-validation, and (iii) hyperparameters
tuning through grid search. Although these have been well explained in
the literature [64–66], a brief description is provided here.
Data preprocessing: This involves cleaning and transforming raw data
into a format appropriate for training [64]. For data cleaning, missing
values in the raw dataset are identied and handled appropriately (e.g.,
replaced by average feature value). In addition, instances with CTL
values considered as outliers are removed from the dataset. For data
transformation, many algorithms may not perform optimally when
dealing with input features that exhibit signicant differences in scales,
requiring feature scaling to transform numerical attributes to a consis-
tent scale that ensures their magnitudes do not disproportionately in-
uence the learning process [67]. Standardization, the chosen method
of feature scaling in this work, transforms numerical values of features in
a dataset into a standard normal distribution (by subtracting the mean
and scaling to unit variance) [68]. Categorical features, on the other
hand, are converted into a numerical format. For features with more
than 2 categories/classes, One-hot encoding is adopted. One-hot
encoding assigns a unique binary value for each unique category/class
in a categorical variable, creating a binary vector with all zero values
except the position corresponding to the class where a value of 1 is set
[69]. Ordinal encoding, which does not alter data dimensionality and is
substantially more memory efcient compared to one-hot encoding, is
used for binary categories in this study [70]. Even though it assigns
integer values based on the ordinal relationship between categories,
only 0 or 1 values will be created.
10-fold cross-validation: This evaluates a model’s performance by
dividing the modelling dataset into 10 subgroups, or folds. In each
iteration, one-fold is used as the validation data while the remaining
nine folds are used for training and this process is repeated 10 times,
with each fold serving as the validation set once [65]. This method
provides a more reliable estimate of a model’s performance compared to
a single train-test split. The resulting performance metrics from each
iteration are averaged, offering a more representative evaluation of the
model’s likely performance on unseen data [71].
Hyperparameters tuning through grid search: Hyperparameters are
conguration settings of a machine learning model that are not learned
from the training data but are set before the training process: e.g., the
activation function and optimizer types in ANN, and the kernel type in
SVM [72]. The performance of numerous machine learning techniques is
highly contingent on the specic congurations of hyperparameters
[73]. In this work, grid search is used for hyperparameter tuning which
involves searching for the best combination of hyperparameter values
that result in optimal model performance. Grid search exhaustively
considers all specied parameter combinations and uses cross-validation
to assess the performance of each combination of hyperparameters [68].
The performance metric used in this process is the validation set RMSE.
The hyperparameter combination resulting in the lowest RMSE is
selected.
To quantitatively evaluate the effect of each feature (inuencing
factors of CTL) on the output, partial dependence plots (PDPs) are
generated. PDPs are graphical tools used in machine learning to explain
the impact of one feature on the predicted outcome of a model [66].
After selecting the feature of interest, a grid of values to explore is
chosen. For each value in the grid, the feature of interest takes this value
in the entire dataset and all other features are kept constant, and the
model is used to make predictions across all instances that are then
averaged. Each PDP is created by plotting the values from the chosen
grid on the horizontal axis and the corresponding average predictions on
the vertical axis [74]. When dealing with categorical features, the grid
consists of all categories within the selected feature.
2.2. Choice of input parameters
The selected input parameters in this study are strategically designed
to encompass a comprehensive range of factors inuencing CTL. These
parameters are formulated to capture essential aspects based on quan-
tied information reported directly or indirectly from and across all the
studies included in the data collection process. For instance, while the
physical condition of the steel-concrete interface, particularly the
entrapped air void content, does impact CTL, there is a noticeable
scarcity of reported quantication of this effect. The challenge arises
from the difculty of non-destructively measuring the air void content at
the interface. The intricacy involved in quantifying the amount of
entrapped air voids and characterizing microscopic defects contributes
to the unquantiable nature of the interface condition. The same goes
for the oxygen availability on the rebar surface. Other inuencing fac-
tors seldom reported are specimen moisture content, relative humidity
inside the pores or the electrical resistivity of the specimen under
testing, and steel metallurgy. Table 1 presents the selected inputs, the
classes considered under categorical variables for training, and the
inuencing factors that are thought to be reected by each input with a
number referring to those discussed in Section 1.3. It is important to note
that some of these input parameters are interdependent. In fact, aiming
for input independence is advantageous for achieving increased pre-
diction accuracy, especially when there is ample information available
to articulate the distribution of data within the feature space. Never-
theless, it is essential to acknowledge that incorporating variables
interdependent on others, which encapsulate novel and more compre-
hensive information, can contribute to enhancing both the accuracy and
robustness of the model under development [75].
It is important to provide more details on some chosen features to aid
clarity on their categories and assumptions made. For curing duration
and temperature, 28 days curing duration was assumed for some
experimental work that did not report the specimen’s curing duration.
Furthermore, an average (room) temperature of 20 ◦C was assumed for
laboratory experiments that did not report testing temperature, while
the location’s mean temperature obtained from different weather
sources was used for eld experiments that did not report their tem-
perature. For binder oxide composition, most studies reported the oxide
composition of Portland cement or supplementary cementitious mate-
rials used. For some cases where it was not, an average oxide compo-
sition of similar binder material, based on literature data from the same
author or others in the same region, was used. Therefore, the oxide
composition used for blended cement mix represents the weighted
average of the cementitious components used in the binder mix based on
the % replacement of each.
For the exposure condition, 4 different classes presented in Fig. 1
were considered. A unique well-dened exposure condition for speci-
mens studied with mixed-in chlorides could not be established. Thus,
based on the principle of mean imputation [76], one of the simplest
methods for handling missing values in machine learning datasets where
missing values in a feature are replaced with the mean of the available
values in that feature, exposure condition for mixed-in chlorides tests
was set to ‘fully immersed in solution’. Also, eld exposure conditions
that do not fully match the dened classes were assigned to the closest
analogous class.
For rebar surface condition, rebar subjected to polishing with sand-
paper, sandblasting, or acid pickling to remove their mill scale before
testing is categorized as ‘mill scale removed’, while rebar tested in their
as-received surface conditions or degreased with acetone and/or ethyl
alcohol are categorized as ‘mill scale present’. For exposed rebar area, it
is computed based on the exposed length and circumference of the rebar
exposed to chloride ions, as highlighted in red in Fig. 2 for six cases
encountered in the literature (one mixed-in chloride and ve diffused
chloride). For instance, Case 4 highlights the importance of considering
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
5
the closest exposed rebar circumference portion to the specimen edge
facing chloride ingress. Even though chlorides will reach the non-red-
colored region of the rebar circumference, this will happen at a later
stage when compared to the red-colored region, due to the increased
distance the chloride ions need to travel before reaching the bar surface.
Therefore, CTL of the bar will be dependent on the diffusion of chlorides
to the red-colored rebar perimeter.
For corrosion mode setup, microcell corrosion is the term given to
the situation where active corrosion and the corresponding cathodic
half-cell reaction take place at adjacent parts of the same rebar. In
contrast, macrocell corrosion arises when an actively corroding bar is
coupled with another bar that remains passive due to its different
composition or exposure to a different environment [77]. Thus, in
microcell corrosion, the anode and cathode are spatially combined,
resulting in a uniform potential eld and little to no current owing
through the concrete. On the other hand, when the anode and cathode
are spatially separated, the corrosion mode is said to be macrocell or
non-uniform. In non-linear potentiostatic polarization experiments,
however, the rebar is polarized (typically using 100–500 mV) and held
to a certain potential during the experiment. The corrosion activity of
the polarized rebar is monitored by measuring the current exchange
(from the rebar to a counter electrode) that must occur in order to
maintain the potential level the bar was polarized to. For corrosion
detection technique, the methods used to obtain corrosion initiation
incorporates: (i) monitoring the rebar OCP; (ii) monitoring the microcell
or macrocell corrosion current; (iii) monitoring resulting polarization
current from experiments under potentiostatic polarization control; and
(iv) visually observing corrosion signs on the rebar. It is important to
note that ‘microcell corrosion current’ class incorporates corrosion
detection through LPR or EIS measurement with or without OCP
monitoring in parallel. Similarly, ‘macrocell corrosion current’ class
Table 1
Machine learning models input parameters with their corresponding classes,
units, and reected inuencing factors following section 1.3.
Categorical input
parameters
Classes Reected
inuencing factors
1) Rebar surface roughness ribbed surface 11
smooth surface
2) Rebar surface condition mill scale present 7
mill scale removed
3) Exposure condition fully immersed specimen in
chloride solution
4
ponding chloride solution on
specimen
wet and dry exposure to
chloride solution
partially immersed in
chloride solution
4) Testing environment laboratory environment 4
eld exposure
5) Rebar environment concrete specimen 14
mortar specimen
6) Method of chloride
introduction
diffusion (capillary suction
and/or natural diffusion)
16
mixed-in
migration (through potential
gradient)
7) Cation of the salt Na (Sodium) 5
others [Mg, magnesium; K,
potassium; Ca, calcium]
8) Corrosion modes setup microcell corrosion 18
non-linear potentiostatic
polarization
macrocell corrosion
9) Corrosion detection
technique
open circuit potential (OCP) 15
microcell corrosion current
(LPR, EIS)
non-linear potentiostatic
polarization current
gravimetric technique
(visual inspection)
macrocell corrosion current
Numerical input
parameters
Units Reected
inuencing
factors
10) Water/binder ratio unitless 7
11) Curing duration days 2
12) Temperature degree celsius 4
13) Rebar corrosion potential mV (saturated calomel
electrode, SCE)
8
14) Exposed rebar area mm
2
14
Binder oxide
composition
15) CaO % Oxide (equivalent
composition of the binder)
1,6
16) SiO
2
17)
Al
2
O
3
18)
Fe
2
O
3
19) SO
3
20) MgO
21)
Na
2
O
eq
Numerical Output
parameter
Units
CTL Total chlorides in % weight of binder
Fig. 1. Exposure condition classes considered for concrete or mortar specimens
under testing for CTL.
Fig. 2. Exposed rebar circumference and length to chlorides for various
experimental setups/procedures encountered in the literature. (1) mixed-in
chloride and (2–6) diffused chloride with different diffusion scenarios.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
6
incorporates corrosion detection through macrocell or galvanic current
measurement with or without OCP monitoring in parallel.
For rebar corrosion potential, it should be noted that without
external intervention, the rebar corrosion potential is not an absolute
property but rather a quantity that changes with the rebar environment
and during transitioning from passive to active corrosion. Thus, incor-
porating this parameter in this predictive model aims to reect only the
rebar corrosion potential held constant under non-linear potentiostatic
polarization-controlled experiments to induce passive corrosion poten-
tial without active corrosion initiation. All values are converted
accordingly to be expressed in mV (SCE). For tests carried out in free
corrosion potential, a dummy value corresponding to the average value
of ‘rebar corrosion potential’ feature is assigned such that there is no
need to create a partial feature dataset.
For CTL, this output parameter is expressed in % total chlorides per
weight of binder. For studies that only reported CTL expressed in terms
of free chlorides (7 % of the dataset), the values were converted to total
chlorides based on the average ratio of CTL (total chloride)/CTL (free
chloride) obtained from other tests that reported both quantities and
used similar binder.
2.3. Data collection
Data collection involved an exhaustive literature review and the
establishment of a comprehensive database focusing on CTL and its
related variables. The database creation adhered to specic criteria,
including (i) studies solely examining CTL in concrete and mortar sys-
tems, (ii) studies reporting CTL values as a percentage of total or free
chlorides by weight or volume of cement, cementitious material, or
concrete/mortar, (iii) studies reporting directly or indirectly a minimum
of 90 % of the variables of interest, and (iv) studies conducted on carbon
steel rebar without exposure to external stresses. Over 300 publications
from peer-reviewed journals and conference proceedings were
reviewed, yielding valuable data extracted from approximately 70 ref-
erences [10,20,21,28,31,37,40,41] [78–110] [111–147].
The initial database consisted of 423 data points (i.e., instances),
each representing experimented CTL and its associated features. It is
noteworthy to mention that no instances with identical feature attri-
butes existed; for experiments involving replicas, the average CTL was
utilized. Additionally, for each set of identical experiments (i.e., CTL
from replicate samples), the minimum, maximum, and standard devia-
tion were either directly extracted or computed manually. 92.9 % of the
data were CTL values based on total chloride, while others were free
chloride. Both total and free CTL were concurrently reported in 43.7 %
of the cases. Minimum and maximum CTL, along with CTL standard
deviation, were obtained for 47.8 %, 45.6 %, and 37 % of the dataset,
respectively. Any reported CTL expressed as mass per concrete/mortar
volume or percent weight of concrete/mortar was converted to percent
by weight of binder. From the initial CTL database analysis, 24 data
points were identied as outliers and subsequently removed. These
outliers were either instances where reported CTL exceeded 3 % of total
chlorides by weight of binder or where the values were deemed erro-
neous by the authors. This curation process resulted in a nal dataset
comprising 399 instances, which formed the basis for constructing the
predictive models.
2.4. Models building process
The overall training and testing process of the machine learning
models is described in Fig. 3. Each developed model underwent a
standardized process (described in Section 2.1.3) to ensure consistency
and reproducibility of the results. To start, a random state was specied
to shufe and split the dataset comprising 399 instances into modelling
and testing data, maintaining an 80:20 ratio respectively, and feature
scaling was applied to the modelling set to normalize the features. A
random state was applied throughout the process to ensure compara-
bility across the 7 models developed. Following this, hyperparameter
tuning was conducted via an exhaustive grid-search method, optimizing
the model performance based on a 10-fold cross-validation score. The
optimal combination of hyperparameters was then selected, the one
leading to the lowest average RMSE on the validation set. During this
stage, the training data consists of 90 % of the modelling dataset, while
the remaining 10 % is used for validation. Subsequently, the entire
modelling dataset was utilized to train and generate the best-performing
Fig. 3. Flowchart for the training and testing process of machine learning models.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
7
model for each machine learning algorithm using the identied optimal
hyperparameters. Following model training, predictions were made on
the testing data, employing the tted pipeline with feature scaling based
on the training data. The model’s performance was then evaluated using
three metrics: RMSE, MAE, and R
2
computed for the training and the
testing data.
Each machine learning algorithm used resulted in one nal opti-
mized model. A total of 7 models were then compared based on their
testing performance metrics. Finally, the algorithm associated with the
best-performing standalone model is chosen and is used to train and
build a model, with its optimal hyperparameters, using the entire scaled
dataset encompassing both modelling and testing data (i.e., no splits).
This model is employed to generate partial dependence plots, offering
insights into the relationship between all features and CTL across the
entire dataset. This comprehensive approach ensures robustness and
reliability in model development and evaluation.
3. Results and discussions
3.1. Features distribution and chloride binding
Figs. 4 and 5 display the cumulative distribution of the 9 categorical
and 12 numerical features considered in the developed complete data-
set. It can be noticed in Fig. 4 that some categorical features exhibit class
imbalance, where classes are not evenly distributed. For instance, a salt
other than NaCl has been used in only 5 % of the dataset, and gravi-
metric technique for corrosion detection has been employed in just 4 %
of the collected data. Class imbalance can affect the performance of
machine learning models, causing a bias towards the majority class.
Additionally, curing duration and temperature show a skewed cumula-
tive distribution shifted towards lower values, as seen in Fig. 5. For
instance, a curing period of 28 days or less was used in 89 % of the
collected data while only 11 % of the data involved a curing duration of
more than 28 days. Consequently, a trained machine learning model
may struggle to generalize well on instances with higher values for these
features.
Based on the reviewed studies that reported CTL as both free and
total chlorides, the average ratio of total CTL to free CTL was found to be
1.7, i.e., 41 % of chlorides are bound. Fig. 6 shows the cumulative dis-
tribution of this ratio, which ranges from 1.05 to 3, with 80 % of the data
exhibiting a ratio above 1.4. This emphasizes the importance of
considering total and free chlorides separately when modelling CTL. By
doing so, we achieve a more comprehensive understanding of the factors
affecting CTL, leading to more accurate predictions and better-informed
decisions in mitigating corrosion risks. Ignoring this distinction can
result in models that fail to account for critical variations in chloride
behavior, thus reducing their predictive reliability and effectiveness.
3.2. Machine learning models
Upon training 7 models (using 80 % of the database) to predict CTL
based on the 21 input parameters, Fig. 7 presents the model’s outcome,
where each model was evaluated on the testing dataset, and the pre-
dicted CTL values were compared to the experimental ones for both the
training and testing datasets. To assess the models’ performance, the
three statistical parameters (R
2
, MAE, and RMSE) were calculated ac-
cording to Equations (1)–(3), with results presented in Table 2. As shown
in Fig. 7 and Table 2, these models were able to predict CTL with vari-
able accuracy. In the testing data, RMSE ranged between 0.491 and
0.321 % weight (wt.) of binder, MAE ranged between 0.372 and 0.215 %
wt. of binder, and R
2
ranged between 0.415 and 0.751. Based on the
MAE when predicting unseen data, the best-performing standalone
model was the SVM followed by RF, KNN, ANN, DT, and LR. Similarly,
based on the RMSE and R
2
, RF and SVM rank 1st with negligible dif-
ferences, followed by KNN, ANN, DT, and LR. The LR performed the
poorest, and this was expected since the relation between the CTL and all
the inuencing parameters is beyond linearity. The ensemble model
achieved the most accurate outcomes of all models, possessing the
lowest testing MAE of 0.218 % wt. of binder, the lowest RMSE of 0.321
% wt. of binder, and the highest R
2
of 0.751, and most of its predictions
fall within the ±20 % bound lines.
Using data extracted from experiments with multiple replicas, the
mean of all CTL standard deviations observed was 0.194 % wt. of binder,
and the mean of the absolute difference between the average CTL and
the maximum and the minimum was found to be 0.234 and 0.245 % wt.
of binder respectively. These values are comparable to the MAE (0.218)
Fig. 4. Cumulative distribution of the categorical features considered in the developed complete dataset.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
8
and RMSE (0.321) obtained on the testing data for the ensemble model
which showcases its acceptable accuracy for use in predicting CTL. Fig. 8
shows the cumulative distribution of the absolute difference between
CTL average and CTL maximum and minimum, and CTL standard de-
viation for multiple replica experiments as well as CTL average in all the
established database. For 63 % of the database, CTL standard deviation
and the absolute difference between CTL average and the maximum or
the minimum fall below the value corresponding to the MAE of the
ensemble model. This also highlights the acceptable accuracy of this
model.
3.3. Partial dependence plots
3.3.1. PDP for numerical features
PDPs were developed using the SVM model, which proved to be the
best-performing standalone model. Plots corresponding to the 12
numerical features considered in this study are displayed in Figs. 9 and
10. In Fig. 9(a), CTL displays an increasing trend with increasing the
curing time up to ~25 days, after which it plateaus. In a study by Al-
Alaily and Hassan [148], which was not included in data used for the
models developed in this work [151], they showed that CTL was higher
in reinforced concrete samples with higher curing days. Curing for less
than 28 days results in a decreased pore solution pH, which can result in
a low quality (e.g. protectiveness) passive lm form on the embedded
rebar surface and a low chloride binding amount that allows more
chloride concentration at the rebar surface, and thus a lower CTL value
than expected [149]. In Fig. 9(b), a decrease in CTL is observed as the
water to binder ratio increases from 0.3 to 0.75. Concrete’s water to
binder ratio impacts porosity, particularly the pore structure, moisture
content, and oxygen concentration at the steel-concrete interface [20].
Angst et al. [34] in their review article of several experimented works
noted a consensus indicating that lowering the water to binder ratio
below 0.5 increases CTL up to approximately 30 %. However, ndings
regarding the impact of water to binder ratios on CTL above 0.5 were
conicting, with a more evident adverse effect observed. Cao et al.
[145], in their updated review considering Chinese experience, cor-
roborates these observations, with multiple studies indicating that an
increase in water to binder ratio leads to a decline in CTL within the
range of 0.30–0.60. In Fig. 9(c), CTL showed a decreasing trend from 20
to 60 ◦C. It has been shown that the total amount of bound chlorides
decreases with increasing temperature [150,151] because elevated
temperature intensies thermal vibration and the solubility of some
solid phases where chlorides are bound such as Friedel’s salt [152].
Since predicted CTL from the present work is expressed as total chloride,
encompassing bound and free chlorides in the cement pores, an
increase/decrease in binding affects predicted CTL. In addition, elevated
temperature increases the rate of corrosion reaction [153] and decreases
the pore solution pH at the steel-concrete interface [21], which in turn
facilitates rebar corrosion initiation and decreases CTL. The predicted
effect of potentiostatically controlled E
corr
(i.e., OCP) on CTL illustrated
in Fig. 9(d) aligns with observations from the literature. In a specic
Fig. 5. Cumulative distribution of the numerical features considered in the developed complete dataset.
Fig. 6. Cumulative distribution of the ratio of CTL expressed as total chlorides
and free chlorides based on the surveyed data where both quantities
were reported.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
9
environment, the OCP of a metal represents a mix potential resulting
from both cathodic and anodic half-cell reactions. For carbon steel rebar
in contact with the alkaline concrete pore solution, their OCP is
approximately −200 mV
SCE
, provided ample oxygen exists [32]. As
elucidated by Pourbaix [48], the thermodynamic switch from passive
iron oxide lm to iron ion corrosion products (i.e., depassivation)
strongly depends on the rebar’s potential. At continued increase of
chloride ions at the rebar surface the pitting potential, E
pit
, which exists
in the anodic range of a polarization curve, continues to shift towards
more negative values until OCP, meets or surpasses E
pit
, allowing pitting
corrosion to naturally (without external inducement) occur [16]. In a
potentiostatic controlled experiment, it was shown that no discernible
dependency on CTL is observed when OCP value is more positive than
−200 ±50 mV
SCE
, whereas a consistent increase in CTL was observed
for OCP value more negative than −200 ±50 mV (SCE) [32]. When OCP
is controlled externally, the applied polarization inuences oxygen
present at the steel-concrete interface (e.g. more positive polarization
increases oxygen reduction cathodic reaction and in turn increases OH
−
production) and alters the composition of the passive oxide layer on the
bar’s surface (by varying Fe(II) and Fe(III) content), leading to
Fig. 7. Training and testing experimental and predicted CTL by each developed model.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
10
variations in CTL [32].
Pit initiation often correlates with the existence of "weak" spots on
the metal surface, such as indentation, stress concentration, sulde in-
clusions, and grain boundaries. The quantity of such imperfections on a
specimen’s surface is theoretically directly proportional to its overall
exposed area, making larger specimens more likely to experience pitting
initiation compared to smaller ones [154]. This specimen size effect can
be explained by the heterogeneity of the rebar surface, particularly the
spatial variation in physical and chemical conditions of the
steel-concrete interface [136]. As the specimen size increases, the like-
lihood of conditions that enable corrosion to initiate at lower chloride
concentrations also rises [136], resulting in decreasing CTL [155]. The
effect of specimen size is observed in Fig. 9(e) where CTL decreases with
increasing rebar exposed surface area up to 7000 mm
2
, after which an
increase in CTL with increasing specimen exposed area is observed.
There are two possible explanations for this observation. The rst pos-
sibility is that the model may not have accurately captured the true
behavior beyond 7000 mm
2
since 93 % of the dataset used in their
development comprised rebar exposed areas below this value. The sec-
ond possibility is associated with the calculation of the corrosion current
density in experiments involving a large exposed rebar area, which
could result in lower values than the dened pass/fail limit for corrosion
initiation even when pitting corrosion has already been initiated. Since
the measured corrosion current is divided by a larger area in larger
specimens compared to smaller ones, while pitting corrosion remains
the target observation in both cases where the current density is ex-
pected to pass a certain dened marked to indicate pitting, a higher CTL
would be anticipated. This expectation arises from the continuation of
the experiment until the dened current threshold for corrosion initia-
tion is attained, allowing more chlorides to accumulate adjacent to the
rebar surface. This potential phenomenon becomes more pronounced as
the exposed rebar area increases, masking the size effect. For instance, if
the total exposed area is too large, the ratio of sum area anodic to sum
area cathodic of the rebar may be so small that the total corrosion cur-
rent density may not increase beyond 0.1 mA/cm
2
, which is considered
an indicator for signicant active corrosion initiation and continued
propagation [31].
The developed model predicts an increase in CTL with the increase in
CaO content as depicted in Fig. 9(f). An increase in CaO within the
binder utilized leads to a heightened concentration of Portlandite within
the concrete matrix. Consequently, the pH of the matrix rises. It should
be noted that the CaO content of Portland cement is typically >60 %,
which corresponds to the region (~60 %) in Fig. 9(f) where CTL begins
to level off, suggesting that the inclusion of SCM type and replacement
level that decreases CaO below 60 % will decrease CTL of rebar.
Furthermore, CaO and Al
2
O
3
phases of the binder bind chloride in
various forms of chloroaluminate hydrates such as Friedel’s salt [156,
157]. Therefore, higher pore solution pH and higher binding capacity
will result in an increase in CTL, which is expected with the increase in
CaO content.
Fig. 10(a) shows that there is a notable decline in CTL with
increasing SiO
2
content in the binder. This trend can be attributed to
several factors. Firstly, increasing SiO
2
in Portland cement, either from
its lower to higher allowable (17–25 %) range or beyond the maximum
allowable range, leads to a reduction in the pore solution pH [158] due
to fewer alkalis present in the pores and their uptake for the formation of
C-S-H with a lower Ca/Si ratio, resulting in lower CTL value. Also, the
increment of SiO
2
in Portland cement (e.g. through increased silica fume
replacement level) reduces the aluminate phases, thereby diminishing
the binder’s capacity to bind chloride and thus lowering CTL [159].
Similar to expected observations, Fig. 10(a) predicted a decreasing CTL
with increasing SiO
2
, up to 25 %. However, beyond an SiO
2
content of
~32 %, CTL is shown to increase. This contrasting behavior might be
due to two reasons: (i) it is possible that the expected decrease in CTL
with increasing SiO
2
only holds true for typical range used in Portland
cement, and so the impact of SiO
2
differs beyond the range – for
instance, >30 % SiO
2
may greatly rene concrete pores through
increased formation of C-S-H associated with physical adsorption of
chloride ions, thereby decreasing CTL [159], and (ii) it is possible that
the model’s accuracy could be limited beyond this 30 % threshold since
92 % of the dataset used in building the model consists of binders with
SiO
2
content below this value.
Fig. 10(b) shows a decrease in CTL with increasing Al
2
O
3
to 15 %,
and an increasing CTL value beyond 15 %. High alumina content in the
binder results in higher chemical binding capacity through the forma-
tion of Friedel’s salt [160], which explains increasing CTL at >15 %
Al
2
O
3
. However, the variation of CTL with Al
2
O
3
below 15 % cannot be
explained at this moment. Fig. 10(c) shows that Fe
2
O
3
does not have a
major inuence on CTL. Fe
2
O
3
is generally considered inert and there-
fore does not have a signicant impact on CTL [43]. Fig. 10(d) shows an
increase in CTL with increasing MgO content in the binder. The presence
of MgO content above 3 % in the binder is only associated with high
replacement level of blast furnace slag in the database used to build the
model. Thus, a higher chemical and physical chloride binding is ex-
pected with blast furnace slag addition [161], resulting in increasing
CTL. Furthermore, the additional M-S-H (magnesium silicate hydrate)
phases resulting from the use of slag-containing binders could improve
chloride binding capacity, and in turn, increase CTL of the rebar. Fig. 10
(e) shows no major dependence of CTL on the SO
3
content in the binder.
However, the decrease in CTL observed at high SO
3
content (>6.5 %)
can be explained by the negative inuence of SO
3
on the binding ca-
pacity of the binder [151]. Fig. 10(f) shows that CTL increases with
increasing equivalent Na
2
O binder content. The alkali content of cement
elevates the pH of the pore solution whose effect overshadows the
alkali-inhibiting effect of chloride binding [162,163].
Table 2
Developed models performance metrics (R
2
, MAE, and RMSE) calculated for
both training and testing data.
Model Training Testing
MAE RMSE R
2
MAE RMSE R
2
Linear Regression (LR) 0.340 0.463 0.364 0.372 0.491 0.415
Decision Tree (DT) 0.145 0.249 0.816 0.322 0.43 0.551
Support Vector Machine
(SVM)
0.0886 0.209 0.87 0.215 0.352 0.699
K Nearest Neighbors
(KNN)
0.0293 0.0928 0.974 0.274 0.399 0.614
Random Forest (RF) 0.124 0.171 0.913 0.261 0.345 0.711
Articial Neural
Networks (ANN)
0.152 0.209 0.871 0.292 0.413 0.587
Ensemble model (RF &
SVM)
0.103 0.173 0.911 0.218 0.321 0.751
Fig. 8. Cumulative distribution of CTL in the database, the distribution of the
difference between the average CTL and the minimum and maximum value,
and the corresponding standard deviation for multiple replica experiments.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
11
Fig. 9. PDPs for numerical features: curing duration, water to binder ratio, temperature, rebar corrosion potential, exposed rebar area, and CaO. Reference line
indicate ideal case where no change to CTL due to the numerical feature exist, while PDP indicates changes in CTL.
Fig. 10. PDPs for numerical features of binder oxide composition. Reference line indicate ideal case where no change to CTL due to the numerical feature exist, while
PDP indicates changes in CTL.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
12
3.3.2. PDP for categorical features
The PDP developed using the trained SVM model for the categorical
features considered in this study are shown in Figs. 11 and 12. In Fig. 11
(a), the dependence of CTL on rebar surface roughness indicates a
slightly higher value for ribbed bars compared to smooth ones. This is
likely due to two primary factors. Firstly, during the casting process of
concrete, the presence of more air voids at the surface of ribbed bars may
be considered to facilitate the preservation of additional oxygen for the
cathodic reaction needed for passivation purposes. Secondly, when
comparing bars of equal diameter, the ribbed bar has a larger specic
surface area than the smooth bar, a difference is further accentuated
when the ribbed bar has a greater diameter. This increased surface area
enables the ribbed bar to engage more effectively with alkaline sub-
stances present in the concrete pore solution, facilitating the formation
of a denser passivation lm [37]. Fig. 11(b) shows that a slightly higher
CTL is obtained when mill scale is removed on bars tested, which is
Fig. 11. PDPs for categorical variables of rebar surface roughness, rebar surface conditions, testing environment, rebar environment, and cation of the salt. Reference
line indicate ideal case where no change to CTL due to the categorical feature exist, while PDP indicates changes in CTL. Note that these plots should be interpreted
based on the intersection of PDP line with the vertical axis.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
13
supported by several studies. As reported in Ref. [20], the presence of a
mill scale layer on the surface of carbon steel rebar tends to promote
corrosion initiation, consequently reducing CTL when compared to in-
stances where the mill scale is removed. The mill scale presents a large
variability in its chemical and physical characteristics, making as
received rebar surfaces locally diverse and complex compared to uni-
form surfaces when it is removed [164]. Mahallati and Saremi [165]
demonstrated that the presence of mill scale possibly weakens the for-
mation and the protective characteristics of the passive lm layer on the
rebar surface.
Fig. 12. PDP for categorical variables of chloride introduction method, exposure condition, corrosion modes setup, and corrosion detection technique. Reference line
indicates ideal case where no change to CTL due to the categorical feature exists, while PDP indicates changes in CTL. Note that these plots should be interpreted
based on the intersection of PDP line with the vertical axis.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
14
Fig. 11(c) shows a higher CTL for laboratory experiments compared
to eld exposure. Contrary to eld experiments, concrete curing mostly
takes place under optimal conditions, in terms of relative humidity and
temperature [166], in laboratory experiments. In addition, eld exper-
iments are usually several orders of magnitude longer in exposure
duration (and sometimes without frequent monitoring to observe
corrosion initiation as with laboratory specimens) and susceptible to
other reinforced concrete deterioration modes such as atmospheric
carbonation. Thus, lower CTL is expected in eld experiments. Fig. 11(d)
reveals a slightly higher CTL for rebar in concrete specimens compared
to mortar specimens. There are two possible reasons for this observation.
First, the presence of coarse aggregates in concrete specimens modies
the steel-concrete interface. Second, in the process of measuring CTL
through titration of chloride in concrete specimens, a small amount of
concrete sample adjacent to the rebar is removed, and it could be likely
that it has a binder-to-coarse aggregate ratio higher than the bulk of the
sample. This would require the usage of a conversion factor lower than
the theoretical one to convert obtained CTL expressed as the total
amount of chloride present in the concrete into weight percentage of the
binder, without which, a CTL higher than the actual value will be
reported.
Although the literature provides strong evidence supporting the
observed inuence of rebar surface roughness, surface condition, testing
environment, and rebar environment on CLT, PDPs reveal that the
magnitude of these effects is relatively small when examined in the
context of experimental data scatter. Specically, the changes in CLT
attributed to these features are below the mean standard deviation
observed across all experiments (0.194 % wt. of binder) with multiple
replicas adopted for this modelling. This indicates that while these
factors are theoretically relevant and contribute to experimental results
and this model’s predictions, their practical, individual inuence on CLT
may be limited. In other words, the variation in CLT caused by these
features (i.e. rebar surface roughness, surface condition, testing envi-
ronment, and rebar environment) is within the normal range of exper-
imental variability and may not cause signicant deviations in CLT
beyond what would be expected from the stochastic nature of corrosion.
Therefore, although these features are accounted for in the model and
align with prior experimental studies, their isolated impact might be too
small to drive substantial (i.e. beyond reported value) changes in CLT in
practice.
Magnesium chloride and calcium chloride salts result in more bound
chlorides compared to sodium or potassium chlorides [167]. This would
have a positive effect on CTL expressed in this study as total chlorides by
% mass of binder. On the other hand, magnesium chloride and calcium
chloride reduce the pore solution pH more obviously than sodium
chloride and potassium chloride since the saturation limit of calcium
and magnesium hydroxide resulting from the use of former salts are
much lower than sodium and potassium hydroxide formed when the
latter salts [96], allowing them to greater impact corrosion initiation.
For instance, CaCl
2
is reported to have a greater corrosive effect than
NaCl [168]. These opposing effects can explain the negligible variation
observed in Fig. 11(e) for CTL between NaCl and other salts. However,
only 5 % of the dataset involves salts other than NaCl, which can deem
the dependence of CTL on the chloride cation type obtained from this
model inconclusive.
Fig. 12(a) shows that a lower CTL is obtained when chloride is
introduced in the specimen through migration compared to the diffusion
process. When chlorides migrate into the specimens under the inuence
of an electrical eld, the quantity of bound chlorides at the relatively
swift migration front will be considerably lower compared to chloride
penetration via diffusion because the relatively slow diffusion process
affords more time for chloride binding to take place [97]. The model also
predicted minimal or no effect on CTL between a diffusion process or
mixed-in chloride for chloride introduction to the rebar. However, CTL
is usually expected to be lower for experiments using mixed-in chloride
since mixing chloride during specimen preparation distributes chloride
more uniformly, which contrasts with the reality of non-uniform chlo-
ride concentration at the surface of a rebar. Also, using high chloride
concentration (for cases with mixed-in chloride) possibly decreases the
alkalinity of the concrete pore solution to a point where the rebar pas-
sive lm is less protective against corrosion [37].
Both water and oxygen are essential components for corrosion pro-
cess; therefore, the absence of either one impedes corrosion process.
Since the concrete moisture content regulates the availability of water
and oxygen at the rebar surface, when it is either fully saturated with
water or relatively dry, elevated chloride concentrations are necessary
to initiate corrosion. This is because, in fully saturated conditions, there
is limited oxygen available, and in dry conditions, there is insufcient
water, both of which are crucial for corrosion to occur [16]. In addition,
it has been observed experimentally that the most conducive conditions
for corrosion initiation in concrete are present when subjected to wet-
ting/drying cycles [16]. This was shown in the work of Sandberg and
Sorensen [124], and the unpublished data by Boschmann and Angst
where the CTL was found to be higher in continuously immersed con-
ditions compared to wetting and drying exposure as reported in
Ref. [34]. Furthermore, higher moisture contents in concrete decrease
CTL, likely due to a decrease in concrete resistivity (concrete with low
resistivity conducts electrical currents more easily, facilitating the
electrochemical reactions that lead to corrosion) [142]. On the other
hand, as depicted in Fig. 12(b), the model reveals contrary results.
Compared to a fully immersed condition, wet and dry exposure induces
an increase in CTL, partially immersed condition induces a less pro-
nounced increase in CTL while ponding solution on specimen results in
the lowest CTL. One possible explanation for this observation is that the
wet and dry cycles might not be fully saturating or drying out the con-
crete, creating less optimal conditions for corrosion compared to fully
immersed or ponded conditions. This is particularly relevant for testing
less permeable concrete, as the degree of wetness, especially near the
steel-concrete interface, is rarely measured. Inconsistent saturation
during these cycles could lead to uneven chloride ion penetration,
affecting the overall corrosion initiation process and resulting in higher
CTL values compared to consistently saturated conditions. Another
possible reason is the categorization of data for these exposure condi-
tions. While most experimental data t into the exposure categories
dened in the present work, some others possessed inherent difculty in
tting adopted categorization, yet did not have enough dataset to
become a standalone category. For instance, exposure condition for
mixed-in chloride tests was set to ‘fully immersed in solution’. In other
words, numerous experimental exposure conditions that did not align
precisely with the dened categories were assigned to the closest anal-
ogous class. This categorization may have introduced bias, particularly
given the relatively small size dataset available in the literature. Such
bias could prevent the model from accurately capturing the full effect of
some exposure conditions on CTL.
Fig. 12(c) reveals that macrocell corrosion test setup leads to lower
CTL compared to a microcell corrosion test setup and non-linear
potentiostatic polarization experiments which score the highest CTL. It
is important to bear in mind that this graph does not factor in the effect
of the applied potential value itself on CTL under potentiostatic tests.
When tests are carried out in free corrosion potential, there are greater
uncertainties in the electrochemical response of the passive layer. In
contrast, potentiostatic conditions create a more uniform system due to
the applied polarization [33]. The composition of rebar passive layer
depends on both the availability of oxygen and pH levels in cement pore
solution at the steel-concrete interface. At the same pH level, OCP is
inuenced by oxygen content and aging. However, when the metal is
polarized, the potential is determined by the externally applied polari-
zation, making the effects of oxygen and aging negligible. Regardless of
the oxygen concentration, when an external potential is applied, it be-
comes the primary factor determining the passive oxide layer properties
at a given pH [32], aiding a more homogeneous passive layer that resists
corrosion initiation and increasing CTL.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
15
Fig. 12(d) shows CTL is lower for a macrocell testing setup compared
to a microcell setup, which can be explained by cathode-to-anode (c/a)
ratio. Li et al. [98] observed an increasing CTL with decreasing c/a.
Transitioning from a c/a ratio of 0 (i.e., no separate physical cathode and
specimens are in microcell corrosion setup) to c/a ratios of 1, 2, and 4,
the CTL values were 0.713 %, 0.696 %, 0.65 %, and 0.562 % total
chlorides per weight of cement, respectively. This observation was
attributed to the higher adsorption of chlorides in the rebar’s passive
lm per unit area of the steel-concrete interface accompanying the in-
crease in c/a ratio. In a macrocell system with coupled cathodic bars,
there is an increased attraction of chlorides at the anode steel-concrete
interface, leading to a more severe localized chloride concentration
around the anodic rebar passive lm which subsequently reduces the
CTL. Fig. 12(d) and (e) show that slightly higher CTL is expected for
experiments measuring microcell corrosion current to detect corrosion
initiation instead of OCP measurement. Li et al. [89], observed the onset
of the sharp shift of corrosion current density coming later than the early
observed OCP drop indicating corrosion initiation (54 days average).
This was explained by recognizing that potential serves as a thermo-
dynamic measurement, whereas corrosion rate is a kinetic one. This time
difference would correspond to the period during which it takes time to
stabilize and attain maximum corrosion activity following the initiation
of a pit.
In Fig. 12(d), higher CTL was predicted under macrocell corrosion
test setup using microcell current detection technique compared to
macrocell current for corrosion initiation. When a macrocell corrosion
test setup is used, two types of corrosion current can be identied: the
macrocell corrosion current, which is measured across the cathode and
the anode coupled in series, and the microcell current, which is the
corrosion current density of anodic rebar obtained via LPR or EIS. Li
et al. [98] and Trejo et al. [132] observed that the macrocell current
density was almost twice that of the microcell current density at the
onset of corrosion. Similarly, Poursaee et al. [167] noted that macrocell
corrosion current densities are higher than microcell corrosion current
densities at low corrosion rates. Since macrocell currents are controlled
by the resistance to ionic current ow in the concrete between the
anodic and cathodic bars, this could be attributed to higher ionic
resistance of the binder in the early stages/age (i.e., when signicant
environmental differences exist between the top and bottom parts of the
samples) of samples with their cathodic and anodic rebar placed in the
same mortar specimen. In contrast, the anodic and cathodic half-cell
reactions occur at adjacent locations (on a microscopic scale) on the
rebar in microcell corrosion, requiring much shorter ionic current paths.
In a macrocell or microcell testing setup, using the gravimetric tech-
nique (mass loss criterion) to detect corrosion initiation will result in a
higher CTL compared to other detection techniques since the experi-
ments would have to be run for a longer period so that signicant
amount of corrosion can take place enough for mass loss to be captured.
However, this is not captured in the model output shown in Fig. 12(d)
and (e) due to the scarcity of data with experiments based on gravi-
metric method in our dataset (4%).
3.4. Input parameters relative importance and correlation
To further highlight the correlation between target CTL output and
the input inuencing parameters, two sets of quantitative coefcients,
correlation coefcients and feature importance coefcients, were
computed. Although RF was not the best-performing model, the model
allows generation of feature importance coefcients that at least quan-
titatively elucidate the relative impact of each input parameter on the
output. The correlation coefcient between each input feature and the
target variable CTL has been computed using the entire dataset after
ordinally encoding the categorical features (i.e., representing them with
ordered numerical values). The correlation coefcient measures the
strength and direction of a linear relationship between an input variable
or feature and the output. It ranges from −1 to 1, where 1 indicates a
perfect positive linear relationship (i.e., one variable increases, the other
variable also increases proportionally), and −1 indicates a perfect
negative linear relationship (i.e., one variable increases, the other var-
iable decreases proportionally). In addition, feature importance was
extracted from the trained RF model to highlight the relative importance
of each feature in predicting the target variable, where high feature
importance value indicates a feature that is most important to CTL. This
was computed based on the decrease in MSE each feature brings to the
trees in the forest. All features were normalized before training the RF
model which ensures that the feature importance rankings are not
biased by differences in the scale or units of the variables.
By examining Table 3, based on the correlation coefcient, the
corrosion detection technique and the proportion of Al
2
O
3
in the binder
have the strongest linear relationship with CTL. Furthermore, the two
most important features are the rebar exposed area and the proportion of
Al
2
O
3
in the binder. In fact, Al
2
O
3
plays a major role in chloride binding
and constitutes a major component in the formation of Fridel’s salt
[160]. In addition, exposed rebar area is shown to be another important
feature in CTL, which has also been experimentally proven to be a sig-
nicant parameter when studying CTL [136]. Therefore, binder
composition, specically Al
2
O
3
, exposed rebar area, and possibly
corrosion detection techniques are the major inuencing parameters in
any CTL study.
Data in Table 3 has been further presented in order of ranking in
Figs. 13 and 14 for easy comparison. As earlier mentioned, features
shown were normalized before training the model to eliminate bias, and
further normalization or standardization (i.e. feature scaling) to obtain
uniform scaling at this point introduces new problems. Firstly, it in-
troduces a bias between the numerical features - once scaled, the
changes shown in the PDPs for a given feature are no longer directly
interpretable in terms of their original units, making it harder to draw
meaningful conclusions about the feature’s true effect and distorts the
real-world meaning of the feature values. For instance, a small change in
a feature that naturally has a large range might seem comparable to a
change in a feature with a smaller range after scaling, even though their
true impacts in the original units differ signicantly). Secondly, it is
impractical to normalize categorical features since they are one-hot or
ordinally encoded and are unaffected by scaling - attempt at scaling
them with numerical features introduces inconsistency (i.e. an apples-
to-oranges comparison) as categorical features will retain their orig-
inal form. For instance, the PDP for numerical features may show
Table 3
Feature importance of the input factors considered in this study and their linear
correlation with CTL.
Input parameters Linear correlation
coefcient
Feature
importance
Rebar surface roughness 0.020 0.005
Rebar surface condition 0.031 0.003
Exposure condition 0.148 0.005
Testing environment 0.050 0.013
Rebar environment −0.217 0.005
Method of chloride
introduction
−0.027 0.007
Cation of the salt 0.008 0.004
Corrosion modes setup 0.069 0.006
Corrosion detection technique 0.225 0.005
Water to binder ratio 0.012 0.057
Curing duration −0.082 0.073
Temperature −0.084 0.058
Rebar corrosion potential −0.193 0.056
Exposed rebar area −0.076 0.117
CaO 0.208 0.087
SiO
2
−0.083 0.034
Al
2
O
3
−0.242 0.123
Fe
2
O
3
−0.092 0.102
SO
3
−0.075 0.051
MgO 0.005 0.051
Na
2
O
eq
−0.141 0.039
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
16
smoother, less pronounced variations after scaling, while categorical
features—unaffected by the transformation—may appear to have a
more signicant impact simply because they remain in their original,
unaltered state.
Overall, it is important to note that while feature importance ranking
provides an overview of the relative contributions of each variable in
predicting the output, it may not fully capture the practical relevance of
some variable in real-world applications, especially when it comes to
binder composition. For example, the variable ’SiO₂’ is treated as a
single entry, though it encompasses both the silica content from Port-
land cement and that from varying levels of pozzolanic admixtures used
as replacement; the latter (i.e. varying levels of SiO
2
) may impact CTL
differently. In order to have a better practical understanding on the ef-
fect of binder composition on CTL, Section 3.5 case study investigates
the direct effect of various cementitious blends on CTL.
3.5. Case study on the effect of supplementary cementitious materials on
CTL
SCMs, such as y Ash (FA), blast furnace slag (SLAG), silica fume
(SF), and limestone (L) are used as partial replacement for ordinary
Portland cement (OPC) for reducing waste and greenhouse-gas
emissions, as the cement industry is one of the largest CO2 emitters in
the world [169]. These SCMs differ in their impact on concrete micro-
structure and macroscopic behavior (e.g., alkalinity, porosity,
ion-binding capacity), thus affecting CTL differently. To investigate this,
an average oxide composition for OPC, FA, SLAG, SF, and L was obtained
from the gathered dataset and used for this purpose. Further details
regarding the average compositions used can be found in Appendix A.
The most common % replacements for each SCM found in the literature
were used to obtain the equivalent oxide composition of the multiple
cementitious blends investigated: 5–15 % SF replacement in 5 %
increment, 10–30 % L replacement in 10 % increment, 15–45 % SLAG
replacement in 15 % increment, and 15, 20, 35, and 45 % FA replace-
ment. The average CTL for each cementitious blend, including the
reference of 100 % OPC, was obtained using the trained SVM model. By
setting the oxide composition of each instance in the dataset to match
the blend being studied, the overall average CTL was calculated from all
predictions across the dataset. The percentage variation in CTL
compared to the 100 % OPC reference is presented in Fig. 15 for all
studied cementitious blends.
The partial substitution of SF for OPC decreases the quantity of
aluminate phases, which reduces the chemical binding of chlorides but
also renes the pore structure of the cement matrix, thus improving the
Fig. 13. Ranked feature importance of the input factors considered in this study.
Fig. 14. Ranked linear correlation of the input factors considered in this study with CTL.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
17
physical adsorption of chlorides. However, C-S-H formed in the presence
of SF may have weaker chloride sorption than those formed in OPC
alone [170]. Additionally, increasing SF substitution level decreases the
pH of the pore solution [171,172]. This explains the reason for the
decrease in CTL when SF is used, and is responsible for the increased
drop in CTL at 10%SF replacement from 5 %. However, SF also con-
tributes to the densication of the steel-concrete interface, which could
eventually limit the area of exposed rebar accessible to chloride causing
corrosion initiation and could counterbalance some effects from pH
decrease [173]. This may explain why the decrease is less pronounced at
15%SF replacement level. In other words, the steel-concrete interface
renement and a denser microstructure with high SF content not only
slows down chloride ingress but also reduces the oxygen content,
thereby lowering the rebar OCP. Overall, the use SF at 5–15 %
replacement level decreases CTL of carbon steel rebar.
The partial substitution of SLAG for OPC increases chloride binding
capacity in concrete by enhancing both physical and chemical chloride
binding [174–177]. This increase in the overall binding capacity is re-
ected by the model predicting a steady increase in CTL with up to 45 %
SLAG replacement. The improvement in CTL is less pronounced at 60 %
replacement, which could be attributed to the decrease in pH of the pore
solution. It has been reported that the presence of SLAG reduces the pH
of the pore solution [178]. This pH drop may counterbalance the
increased chloride binding capacity at higher slag replacement levels.
L is traditionally interground with OPC to act as a ller material and
an inoculant for hydration of OPC clinker, both improving permeability
of the concrete. Past research found the presence of L in concrete mixes
to reduce concrete’s chloride binding capacity, attributed to a lower
alumina content due to the reaction of L with constituents that bind
chlorides, most notably C
3
A and possibly C
4
AF [176,177]. As seen in
Fig. 13, there is a continuous decrease in CTL compared to a 100 % OPC
reference blend, with an increase in L replacement level from 10 % to 30
%.
The developed model predicts a steady decrease in CTL with an in-
crease in FA replacement going from 15 to 45 %. A 45 % replacement of
OCP with FA has the most negative effect on CTL among all studied
cementitious blends. Even though FA improves the overall chloride
binding capacity of the binder [179,180], it lowers the pH of the pore
liquid [181,182]. The reduction in pH appears to be more pronounced
than the improved chloride binding capacity [183], which explains the
negative effect observed on CTL.
4. Conclusion
Based on extensive data mining from the literature, 423 data records
of chloride threshold level (CTL) determination experiments were
collected. A database with 21 dened features was developed and used
to rigorously train and test optimized machine learning models for CTL
predictions. This study introduced six standalone models: linear
regression (LR), decision tree (DT), random forest (RF), K nearest
neighbors (KNN), support vector machine (SVM), articial neural
network (ANN), and an ensemble average-voting based machine
learning model. The accuracy of these models in predicting CTL was
compared and analyzed using three performance metrics. The best-
performing standalone model was selected to generate partial depen-
dence plots, which were used to study the effect of each variable on CTL
independently, and to examine the impact of the use of supplementary
cementitious materials as a substitute for ordinary Portland cement on
CTL. In addition, feature importance was obtained using the trained RF
model to highlight the most important factors affecting CTL.
The key ndings from the study are as follows:
1. Ensemble model performance: The ensemble machine learning
model demonstrated superior prediction performance compared to
any standalone model, achieving a validation MAE of 0.218 %
weight of binder, an RMSE of 0.321 % weight of binder, and a co-
efcient of determination of 0.751.
2. Partial dependence plots: SVM proved to be the best-performing
standalone model and was used to generate PDPs that can be used
as tools to estimate changes in CTL due to variations in the studied
features. These plots successfully quantied the effect of each feature
on CTL, with results generally aligning with existing literature
ndings. The analysis of the PDPs suggests that: (i) features such as
rebar surface roughness, rebar surface condition, and the rebar’s
environment (concrete or mortar) have a relatively minor inuence
on CTL (i.e. within experimental scatter); and (ii) to increase CTL and
ultimately enhance the resistance of reinforced concrete structures to
chloride-induced corrosion, it is recommended to improve concrete
curing with extended duration (up to 30 days), adopt the lowest
feasible water-to-binder ratio, and utilize cementitious blends with
high CaO content, Al₂O₃ content (>25 %), MgO content (>2.5 %),
SO₃ content (<6 %), and high alkali levels.
3. Feature importance: The RF model identied the Al
2
O
3
binder
content and exposed rebar area as the most inuential factors
affecting CTL. However, the sometimes-unbalanced data distribution
Fig. 15. Effect of different SCMs containing binders on CTL showing percent increase or decline from non-SCM containing binder. Note that 100 % OPC is the
reference (point zero), while other blends show deviation from this reference.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
18
in the collected literature might have inuenced the low importance
attributed to some of the other features.
4. Impact of supplementary cementitious materials: Among y ash,
limestone, silica fume, and blast furnace slag, only blast furnace slag
showed a positive impact on CTL, while y ash had the most negative
effect.
The study highlights promising prospects for employing machine
learning as a tool for CTL prediction, potentially eliminating the need
for specic CTL determination experiments. As well, it has shown the
effectiveness of machine learning in unfolding the effects of binder
composition, environmental factors, rebar characteristics, and other
parameters on CTL. Overall, the application of machine learning in
CTL prediction can be further improved in the future when the
database consists of standardized experiments for CTL determination
with uniform corrosion detection techniques, initiation criteria, and
experimental procedures.
CRediT authorship contribution statement
Nicolas Maamary: Writing – original draft, Software, Methodology,
Data curation, Conceptualization. Ibrahim G. Ogunsanya: Writing –
review & editing, Investigation, Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Acknowledgments
Financial support for this research is provided by the Natural Science
of Research Council of Canada (NSERC) and Canada Foundation for
Innovation (CFI). The authors also appreciate the support provided by
MITACS for the rst author.
6 Appendix A
Table A-1
Average oxide composition used in the case study on the effect of supplementary cementitious materials on CTL for ordinary Portland cement, y ash, blast furnace
slag, silica fume, limestone, and various cementitious blends.
Material Oxide composition (mass %)
CaO SiO
2
Al
2
O
3
Fe
2
O
3
SO
3
MgO Na
2
O
eq
Ordinary Portland cement (OPC) 62.49 20.74 5.50 3.16 2.74 1.87 0.85
Fly ash (FA) 7.65 49.47 24.97 8.60 0.82 1.77 1.77
blast furnace slag (SLAG) 40.08 36.25 13.03 0.72 1.56 6.30 0.65
Silica fume (SF) 0.70 95.05 0.40 0.35 0.20 0.27 0.47
Limestone (L) 51.58 4.50 0.53 0.31 0.19 0.69 0.26
85 % OPC +15 % FA 54.26 25.05 8.42 3.98 2.45 1.86 0.99
80 % OPC +20 % FA 48.78 27.92 10.37 4.52 2.26 1.85 1.08
65 % OPC +35 % FA 43.29 30.79 12.32 5.07 2.07 1.84 1.17
55 % OPC +45 % FA 37.81 33.67 14.26 5.61 1.87 1.83 1.26
85 % OPC +15 % SLAG 59.13 23.06 6.63 2.80 2.56 2.54 0.82
70 % OPC +30 % SLAG 55.77 25.39 7.76 2.43 2.39 3.20 0.79
55 % OPC +45 % SLAG 52.41 27.72 8.89 2.06 2.21 3.86 0.76
40 % OPC +60 % SLAG 49.05 30.04 10.02 1.70 2.03 4.53 0.73
95 % OPC +5 % SF 59.40 24.45 5.24 3.02 2.61 1.79 0.83
90 % OPC +10 % SF 56.31 28.17 4.99 2.88 2.49 1.71 0.81
85 % OPC +15 % SF 53.22 31.88 4.73 2.74 2.36 1.63 0.79
90 % OPC +10 % L 61.40 19.11 5.00 2.88 2.48 1.75 0.79
80 % OPC +20 % L 60.31 17.49 4.50 2.59 2.23 1.64 0.73
70 % OPC +30 % L 59.22 15.86 4.01 2.31 1.97 1.52 0.67
Data availability
Data will be made available on request.
References
[1] A.E.K. Jones, Development of an Holistic Approach to Ensure the Durability of
New Concrete Construction, British Cement Association, 1997.
[2] D. Bentz, F. Zunino, D. Lootens, Chemical vs. Physical acceleration of cement
hydration, Concr. Int. 38 (2016) 37–44.
[3] S.K. Kaushik, S. Islam, Suitability of sea water for mixing structural concrete
exposed to a marine environment, Cem. Concr. Compos. 17 (1995) 177–185,
https://doi.org/10.1016/0958-9465(95)00015-5.
[4] C. Shi, X. Hu, X. Wang, Z. Wu, G. de Schutter, Effects of chloride ion binding on
microstructure of cement pastes, J. Mater. Civ. Eng. 29 (2017) 4016183, https://
doi.org/10.1061/(ASCE)MT.1943-5533.0001707.
[5] J. Csizmadia, G. Bal´
azs, F. Tam´
as, Chloride ion binding capacity of
aluminoferrites, Cement Concr. Res. 31 (2001) 577–588, https://doi.org/
10.1016/S0008-8846(01)00458-6.
[6] V.S. Ramachandran, Possible states of chloride in the hydration of tricalcium
silicate in the presence of calcium chloride, Mat´
eriaux et Construction 4 (1971)
3–12. https://api.semanticscholar.org/CorpusID:198173664.
[7] M.V.A. Florea, H.J.H. Brouwers, Chloride binding related to hydration products:
Part I: ordinary Portland Cement, Cement Concr. Res. 42 (2012) 282–290,
https://doi.org/10.1016/j.cemconres.2011.09.016.
[8] A.K. Suryavanshi, J.D. Scantlebury, S.B. Lyon, Corrosion of Reinforcement Steel
Embedded in High Water-Cement Ratio Concrete Contaminated with Chloride,
1998.
[9] B. Reddy, G. Glass, P. Lim, N. Buenfeld, On the corrosion risk presented by
chloride bound in concrete, Cement & Concrete Composites - CEMENT
CONCRETE COMPOSITES 24 (2002) 1–5, https://doi.org/10.1016/S0958-9465
(01)00021-X.
[10] G.K. Glass, N.R. Buenfeld, PRESENTATION OF THE CHLORIDE FOR CORROSION
OF STEEL IN THRESHOLD CONCRETE LEVEL, 1997.
[11] C.L. Page, Mechanism of corrosion protection in reinforced concrete marine
structures, Nature 258 (1975) 514–515, https://doi.org/10.1038/258514a0.
[12] J. Van Muylder, M. Pourbaix, Atlas d’´
equilibres ´
electrochimiques `
a 25 C,
Gauthier-Villars & Cie, Paris, 1963, p. 378.
[13] K. Tuutti, Service Life of Structures with Regard to Corrosion of Embedded Steel,
ACI Symposium Publication vol. 65 (n.d.). https://doi.org/10.14359/6355.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
19
[14] D.A. Hausmann, Steel corrosion in concrete – HOW does it occur? Mater. Prot. 6
(11) (1967) 19–23. https://api.semanticscholar.org/CorpusID:137400102.
[15] U.M. Angst, O.B. Isgor, C.M. Hansson, A. Sagü´
es, M.R. Geiker, Beyond the
chloride threshold concept for predicting corrosion of steel in concrete, Appl.
Phys. Rev. 9 (2022), https://doi.org/10.1063/5.0076320.
[16] U. Angst, B. Elsener, C.K. Larsen, Ø. Vennesland, Critical chloride content in
reinforced concrete - a review, Cement Concr. Res. 39 (2009) 1122–1138,
https://doi.org/10.1016/j.cemconres.2009.08.006.
[17] V.K. Gouda, Corrosion and corrosion inhibition of reinforcing steel: I. Immersed
in alkaline solutions, Br. Corrosion J. 5 (1970) 198–203, https://doi.org/
10.1179/000705970798324450.
[18] C.L. Page, J. Havdahl, Electrochemical monitoring of corrosion of steel in
microsilica cement pastes, Mater. Struct. 18 (1985) 41–47, https://doi.org/
10.1007/BF02473363.
[19] J. Gulikers, Considerations on the reliability of service life predictions using a
probabilistic approach, J. Phys. IV France 136 (2006) 233–241, https://doi.org/
10.1051/jp4:2006136024.
[20] C. Chalhoub, R. François, M. Carcasses, Critical chloride threshold values as a
function of cement type and steel surface condition, Cement Concr. Res. 134
(2020), https://doi.org/10.1016/j.cemconres.2020.106086.
[21] A.M. Alhozaimy, M. Ahmed, R.R. Hussain, A. Al-Negheimish, Quantitative non-
linear effect of high ambient temperature on chloride threshold value for steel
reinforcement corrosion in concrete under extreme boundary conditions,
Materials 14 (2021), https://doi.org/10.3390/ma14247595.
[22] P. Schießl, Corrosion of steel in concrete : report of the Technical Committee 60-
CSC, RILEM (the International Union of Testing and Research Laboratories for
Materials and Structures) (1988). https://api.semanticscholar.org/CorpusID:
108191932.
[23] RILEM, Draft recommendation for repair strategies for concrete structures
damaged by reinforcement corrosion, Mater. Struct. 27 (1994) 415–436, https://
doi.org/10.1007/BF02473446.
[24] D. Hooton, M. Geiker, E. Bentz, Effects of curing on chloride ingress and
implications on service life, ACI Mater. J. 99 (2002) 201–206.
[25] Y. Liu, R.E. Weyers, Time to cracking for chloride-induced corrosion in reinforced
concrete, Spec. Publ. Roy. Soc. Chem. 183 (1996) 88–104.
[26] G. Markeset, Critical chloride content and its inuence on service life predictions,
Mater. Corros. (2009) 593–596, https://doi.org/10.1002/maco.200905288.
[27] M.A. Ehlen, M.D.A. Thomas, E. Bentz, Life-365 service life prediction Model
Version 2.0, Concr. Int. 31 (2009) 41–46.
[28] A. Kenny, A. Katz, Steel-concrete interface inuence on chloride threshold for
corrosion – empirical reinforcement to theory, Construct. Build. Mater. 244
(2020), https://doi.org/10.1016/j.conbuildmat.2020.118376.
[29] U.M. Angst, B. Elsener, Chloride threshold values for corrosion in concrete–a
myth?, in: Concrete Solutions: Proceedings of Concrete Solutions, 6th
International Conference on Concrete Repair, 2016, pp. 391–396.
[30] M. Dixit, A.K. Gupta, A review of different assessment methods of corrosion of
steel reinforcement in concrete, Iran. J. Sci. Technol, Trans. Civil. Eng. 46 (2022)
735–752, https://doi.org/10.1007/s40996-021-00644-5.
[31] C. Alonso, C. Andrade, M. Castellote, P. Castro, Chloride threshold values to
depassivate reinforcing bars embedded in a standardized OPC mortar, Cement
Concr. Res. 30 (2000) 1047–1055, https://doi.org/10.1016/S0008-8846(00)
00265-9.
[32] C. Alonso, M. Castellote, C. Andrade, Chloride threshold dependence of pitting
potential of reinforcements, n.d. www.elsevier.com/locate/electacta.
[33] M.C. Alonso, M. Sanchez, Analysis of the variability of chloride threshold values
in the literature, Mater. Corros. (2009) 631–637, https://doi.org/10.1002/
maco.200905296.
[34] U.M. Angst, M.R. Geiker, M.C. Alonso, R. Polder, O.B. Isgor, B. Elsener, H. Wong,
A. Michel, K. Hornbostel, C. Gehlen, R. François, M. Sanchez, M. Criado,
H. Sørensen, C. Hansson, R. Pillai, S. Mundra, J. Gulikers, M. Raupach,
J. Pacheco, A. Sagü´
es, The effect of the steel–concrete interface on chloride-
induced corrosion initiation in concrete: a critical review by RILEM TC 262-SCI,
Materials and Structures/Materiaux et Constructions 52 (2019), https://doi.org/
10.1617/s11527-019-1387-0.
[35] M. Shakouri, D. Trejo, Estimating the critical chloride threshold of reinforcing
steel in concrete using a hierarchical Bayesian model, Sustain Resilient Infrastruct
4 (2019) 152–172, https://doi.org/10.1080/23789689.2017.1405655.
[36] S. Rengaraju, R.G. Pillai, An accelerated chloride threshold test for uncoated steel
in highly resistive cementitious systems (hr-ACT test), Construct. Build. Mater.
305 (2021), https://doi.org/10.1016/j.conbuildmat.2021.124797.
[37] Y. Wang, C. Liu, Y. Wang, Q. Li, Z. Liu, Investigation on chloride threshold for
reinforced concrete by a test method combining ANDT and ACMT, Construct.
Build. Mater. 214 (2019) 158–168, https://doi.org/10.1016/j.
conbuildmat.2019.04.108.
[38] L. Bertolini, E. Redaelli, Depassivation of steel reinforcement in case of pitting
corrosion: detection techniques for laboratory studies, Mater. Corros. (2009)
608–616, https://doi.org/10.1002/maco.200905276.
[39] ASTM International, Standard Test Methods for Determining Effects of Chemical
Admixtures on Corrosion of Embedded Steel Reinforcement in Concrete Exposed
to Chloride Environments, ASTM G109-13, ASTM International, 2013.
[40] D. Trejo, P.J. Monteiro, Corrosion performance of conventional (ASTM A615) and
low-alloy (ASTM A706) reinforcing bars embedded in concrete and exposed to
chloride environments, Cement Concr. Res. 35 (2005) 562–571, https://doi.org/
10.1016/j.cemconres.2004.06.004.
[41] C. Boschmann K¨
athler, G. Ebell, S. Keßler, Y. Schiegg, C. Dauberschmidt, U.
M. Angst, A comparison of methods to assess the resistance of reinforcing steel
against chloride-induced corrosion in concrete—particular consideration of 12%
chromium steel, Mater. Corros. 73 (2022) 306–325, https://doi.org/10.1002/
maco.202112826.
[42] A.A. Ahmed, N.P. Vaddey, Reliability of chloride testing results in cementitious
systems containing admixed chlorides, Sustain Resilient Infrastruct 8 (2023)
209–221, https://doi.org/10.1080/23789689.2021.1917059.
[43] Y. Zhu, D.D. Macdonald, J. Qiu, M. Urquidi-Macdonald, Corrosion of rebar in
concrete. Part III: articial Neural Network analysis of chloride threshold data,
Corrosion Sci. 185 (2021) 109438, https://doi.org/10.1016/j.
corsci.2021.109438.
[44] Y. Zhu, D.D. Macdonald, J. Yang, J. Qiu, G.R. Engelhardt, Corrosion of rebar in
concrete. Part II: literature survey and statistical analysis of existing data on
chloride threshold, Corrosion Sci. 185 (2021) 109439, https://doi.org/10.1016/j.
corsci.2021.109439.
[45] D.S.J. Cunningham, P´
adraig and cord, supervised learning, in: P. Cord Matthieu,
Cunningham (Eds.), Machine Learning Techniques for Multimedia: Case Studies
on Organization and Retrieval, Springer Berlin Heidelberg, Berlin, Heidelberg,
2008, pp. 21–49, https://doi.org/10.1007/978-3-540-75171-7_2.
[46] V. Nasteski, An overview of the supervised machine learning methods, Horizons 4
(2017) 51–62. https://api.semanticscholar.org/CorpusID:171520859.
[47] J. Han, M. Kamber, J. Pei, 3 - data preprocessing, in: J. Han, M. Kamber, J. Pei
(Eds.), Data Mining, third ed.), third ed., Morgan Kaufmann, Boston, 2012,
pp. 83–124, https://doi.org/10.1016/B978-0-12-381479-1.00003-4.
[48] I. Sarker, Machine learning: algorithms, real-world applications and research
directions, SN Comput Sci 2 (2021), https://doi.org/10.1007/s42979-021-
00592-x.
[49] Y. Song, J. Liang, J. Lu, X. Zhao, An efcient instance selection algorithm for k
nearest neighbor regression, Neurocomputing 251 (2017) 26–34, https://doi.org/
10.1016/j.neucom.2017.04.018.
[50] A.J. Smola, B. Sch¨
olkopf, A tutorial on support vector regression, Stat. Comput.
14 (2004) 199–222, https://doi.org/10.1023/B:STCO.0000035301.49549.88.
[51] H. Tong, D.-R. Chen, L. Peng, Analysis of support vector machines regression,
Found. Comput. Math. 9 (2009) 243–257, https://doi.org/10.1007/s10208-008-
9026-0.
[52] F. Zhang, L.J. O’Donnell, Chapter 7 - support vector regression, in: A. Mechelli,
S. Vieira (Eds.), Mach Learn, Academic Press, 2020, pp. 123–140, https://doi.
org/10.1016/B978-0-12-815739-8.00007-9.
[53] Y.-Y. Song, Y. Lu, Decision tree methods: applications for classication and
prediction, Shanghai Arch Psychiatry 27 (2015) 130–135, https://doi.org/
10.11919/j.issn.1002-0829.215044.
[54] C. Qi, E. Yilmaz, Q. Chen, Background of machine learning, in: Machine Learning
Applications in Industrial Solid Ash, Elsevier, 2024, pp. 93–130, https://doi.org/
10.1016/B978-0-443-15524-6.00015-7.
[55] Y. Li, C. Zou, M. Berecibar, E. Nanini-Maury, J.C.-W. Chan, P. van den Bossche,
J. Van Mierlo, N. Omar, Random forest regression for online capacity estimation
of lithium-ion batteries, Appl. Energy 232 (2018) 197–210, https://doi.org/
10.1016/j.apenergy.2018.09.182.
[56] S.D. Jadhav, H. Channe, Efcient Recommendation System Using Decision Tree
Classier and Collaborative Filtering (2016). https://api.semanticscholar.org
/CorpusID:212463454.
[57] G. Biau, E. Scornet, A random forest guided tour, TEST 25 (2016) 197–227,
https://doi.org/10.1007/s11749-016-0481-7.
[58] I. Basheer, M.N. Hajmeer, Articial neural networks: fundamentals, computing,
design, and application, J. Microbiol. Methods 43 (2001) 3–31, https://doi.org/
10.1016/S0167-7012(00)00201-3.
[59] R. Hecht-Nielsen III, 3 - theory of the backpropagation neural Network**Based on
“nonindent” by robert hecht-nielsen, which appeared in proceedings of the
international joint conference on neural networks 1, 593–611, june 1989. © 1989
IEEE, in: H. Wechsler (Ed.), Neural Networks for Perception, Academic Press,
1992, pp. 65–93, https://doi.org/10.1016/B978-0-12-741252-8.50010-8.
[60] A. Subasi, Machine learning techniques, in: Practical Machine Learning for Data
Analysis Using Python, Elsevier, 2020, pp. 91–202, https://doi.org/10.1016/
b978-0-12-821379-7.00003-5.
[61] B. Erdebilli, B. Devrim-˙
Içtenbas¸, Ensemble voting regression based on machine
learning for predicting medical waste: a case from Turkey, Mathematics 10
(2022), https://doi.org/10.3390/math10142466.
[62] H.E.R. Lee, Seung B. Gui, Use of training, validation, and test sets for developing
automated classiers in quantitative ethnography, in: S.-E.A. Eagan Brendan,
Misfeldt (Eds.), Advances in Quantitative Ethnography, Springer International
Publishing, Cham, 2019, pp. 117–127.
[63] K.K. Dobbin, R.M. Simon, Optimally splitting cases for training and testing high
dimensional classiers, BMC Med. Genom. 4 (2011) 31, https://doi.org/10.1186/
1755-8794-4-31.
[64] H.F. García Salvador, Luengo, introduction, in: Data Preprocessing in Data
Mining, Springer International Publishing, Cham, 2015, pp. 1–17, https://doi.
org/10.1007/978-3-319-10247-4_1.
[65] S.M. Malakouti, M.B. Menhaj, A.A. Suratgar, The usage of 10-fold cross-validation
and grid search to enhance ML methods performance in solar farm power
generation prediction, Clean Eng Technol 15 (2023) 100664, https://doi.org/
10.1016/j.clet.2023.100664.
[66] J.H. Friedman, Greedy function approximation: a gradient boosting machine,
Ann. Stat. 29 (2001) 1189–1232. http://www.jstor.org/stable/2699986.
[67] D.U. Ozsahin, M. Taiwo Mustapha, A.S. Mubarak, Z. Said Ameen, B. Uzun, Impact
of feature scaling on machine learning models for the diagnosis of diabetes, in:
2022 International Conference on Articial Intelligence in Everything (AIE),
2022, pp. 87–94, https://doi.org/10.1109/AIE57029.2022.00024.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
20
[68] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
D. Cournapeau, M. Brucher, M. Perrot, ´
E. Duchesnay, Scikit-learn: machine
learning in Python, J. Mach. Learn. Res. 12 (2011) 2825–2830. http://jmlr.org
/papers/v12/pedregosa11a.html.
[69] S. Bagui, D. Nandi, S. Bagui, R.J. White, Machine learning and deep learning for
phishing email classication using one-hot encoding, J. Comput. Sci. 17 (2021)
610–623, https://doi.org/10.3844/jcssp.2021.610.623.
[70] K. Kunanbayev, I. Temirbek, A. Zollanvari, Complex encoding, in: 2021
International Joint Conference on Neural Networks (IJCNN), 2021, pp. 1–6,
https://doi.org/10.1109/IJCNN52387.2021.9534094.
[71] L.H. Refaeilzadeh Payam, Encyclopedia of database systems, in: M.T. LIU LING,
¨
OZSU (Eds.), Springer US, Boston, MA (2009) 532–538, https://doi.org/10.1007/
978-0-387-39940-9_565.
[72] G.I. Diaz, A. Fokoue-Nkoutche, G. Nannicini, H. Samulowitz, An effective
algorithm for hyperparameter optimization of neural networks, IBM J. Res. Dev.
61 (9) (2017) 1–9:11, https://doi.org/10.1147/JRD.2017.2709578.
[73] F. Hutter, H. Hoos, K. Leyton-Brown, An efcient approach for assessing
hyperparameter importance, in: E.P. Xing, T. Jebara (Eds.), Proceedings of the
31st International Conference on Machine Learning, PMLR, Bejing, China, 2014,
pp. 754–762, in: https://proceedings.mlr.press/v32/hutter14.html.
[74] B.M. Greenwell, Pdp: an R package for constructing partial dependence plots, R J
9 (2017) 421. https://api.semanticscholar.org/CorpusID:52254631.
[75] L.B. Coelho, D. Zhang, Y. Van Ingelgem, D. Steckelmacher, A. Now´
e, H. Terryn,
Reviewing machine learning of corrosion prediction in a data-oriented
perspective, npj Mater. Degrad. 6 (2022), https://doi.org/10.1038/s41529-022-
00218-4.
[76] T. Thomas, E. Rajabi, A systematic review of machine learning-based missing
value imputation techniques, Data Technol. Appl. 55 (2021) 558–585, https://
doi.org/10.1108/DTA-12-2020-0298.
[77] C.M. Hansson, A. Poursaee, A. Laurent, Macrocell and microcell corrosion of steel
in ordinary Portland cement and high performance concretes, Cement Concr. Res.
36 (2006) 2098–2102, https://doi.org/10.1016/j.cemconres.2006.07.005.
[78] H. Fakhri, K.L. Fishman, R. Ranade, A novel experimental method to determine
the critical chloride content in cement-based composites, Construct. Build. Mater.
263 (2020), https://doi.org/10.1016/j.conbuildmat.2020.120101.
[79] R.B. Polder, M. Van Put, W.H.A. Peelen, Accelerated testing for chloride threshold
of reinforcing steel in concrete, in: Fib Symposium, Fib. The International
Federation for Structural Concrete, 2018, pp. 2066–2073, https://doi.org/
10.1007/978-3-319-59471-2_236.
[80] G. Dev Vasudevan, D. Trejo, Assessing the critical chloride threshold of
conventional reinforcement embedded in alternate cementitious systems,
Construct. Build. Mater. 346 (2022), https://doi.org/10.1016/j.
conbuildmat.2022.128401.
[81] M. Babaee, A. Castel, Chloride diffusivity, chloride threshold, and corrosion
initiation in reinforced alkali-activated mortars: role of calcium, alkali, and
silicate content, Cement Concr. Res. 111 (2018) 56–71, https://doi.org/10.1016/
j.cemconres.2018.06.009.
[82] Y.A. Villagr´
an-Zaccardi, C. Andrade, Chloride ingress rate and threshold content,
as determined by the ‘Integral’ test method, in concrete with several w/c ratios in
saturated and unsaturated conditions, Dev. Built. Environ. 8 (2021), https://doi.
org/10.1016/j.dibe.2021.100062.
[83] Syed Ehtesham Hussain, Ahmad S. Al-Gahtani, Rasheeduzzafar, chloride
threshold for corrosion of reinforcement in concrete, ACI Mater. J. 93 (1996).
[84] C. Li, K. Xiao, Chloride threshold, modelling of corrosion rate and pore structure
of concrete with metakaolin addition, Construct. Build. Mater. 305 (2021),
https://doi.org/10.1016/j.conbuildmat.2021.124666.
[85] X. Hu, C.S. Poon, Chloride-related steel corrosion initiation in cement paste
prepared with the incorporation of blast-furnace slag, Cem. Concr. Compos. 126
(2022), https://doi.org/10.1016/j.cemconcomp.2021.104349.
[86] G.G. Cleme˜
na, Y.P. Virmani, Comparing the Chloride Resistances of Reinforcing
Bars Evaluating new, economical metallic reinforcement for its ability to
withstand high salt concentrations, Concr. Int. 26 (2004) 39–49.
[87] B. Elsener, H. B¨
ohni, Corrosion of steel in mortar studied by impedance
measurements, Mater. Sci. Forum 8 (1986) 363–372, https://doi.org/10.4028/
www.scientic.net/msf.8.363.
[88] M.J. Kim, K.Y. Ann, Corrosion risk of reinforced concrete structure arising from
internal and external chloride, Adv. Mater. Sci. Eng. 2018 (2018), https://doi.
org/10.1155/2018/7539349.
[89] C. Li, L. Jiang, S. Li, Effect of limestone powder addition on threshold chloride
concentration for steel corrosion in reinforced concrete, Cement Concr. Res. 131
(2020), https://doi.org/10.1016/j.cemconres.2020.106018.
[90] C. Li, L. Jiang, Effect of limestone powder addition on corrosion initiation time of
reinforced concrete, J. Build. Eng. 59 (2022), https://doi.org/10.1016/j.
jobe.2022.105132.
[91] Y. Wang, Z. Liu, Y. Wang, D. Wang, C. Yuan, R. Liu, Effect of recycled aggregate
and supplementary cementitious material on the chloride threshold for steel bar
corrosion in concrete, Construct. Build. Mater. 346 (2022), https://doi.org/
10.1016/j.conbuildmat.2022.128418.
[92] M. O’Reilly, O. Farshadfar, D. Darwin, Effect of supplementary cementitious
materials on chloride threshold and corrosion rate of reinforcement, ACI Mater. J.
116 (2019) 125–133, https://doi.org/10.14359/51710968.
[93] J.S. Reou, K.Y. Ann, Electrochemical assessment on the corrosion risk of steel
embedment in OPC concrete depending on the corrosion detection techniques,
Mater. Chem. Phys. 113 (2009) 78–84, https://doi.org/10.1016/j.
matchemphys.2008.07.063.
[94] F. Lollini, E. Redaelli, L. Bertolini, Electrochemical tests for the determination of
the critical chloride threshold of steel in concrete with blended cements, Key Eng.
Mater. 711 (2016) 60–67, https://doi.org/10.4028/www.scientic.net/
KEM.711.60.
[95] J. Xu, L. Jiang, W. Wang, Y. Jiang, Inuence of CaCl2 and NaCl from different
sources on chloride threshold value for the corrosion of steel reinforcement in
concrete, Construct. Build. Mater. 25 (2011) 663–669, https://doi.org/10.1016/j.
conbuildmat.2010.07.023.
[96] L. Jiang, R. Liu, L. Mo, J. Xu, H. Yang, Inuence of chloride salt type on critical
chloride content of reinforcement corrosion in concrete, Mag. Concr. Res. 65
(2013) 319–331, https://doi.org/10.1680/macr.12.00082.
[97] C. Boschmann K¨
athler, S.L. Poulsen, H.E. Sørensen, U.M. Angst, Investigations of
accelerated methods for determination of chloride threshold values for
reinforcement corrosion in concrete, Sustain Resilient Infrastruct 8 (2023)
197–208, https://doi.org/10.1080/23789689.2021.1905221.
[98] J. Li, J. Xiong, Z. Fan, M. Chen, L. Sun, C. Zhu, W. Liu, H. Zheng, W. Li,
Mechanistic study of macrocell effect on corrosion initiation and propagation of
reinforcement in submarine immersed tunnel, Cem. Concr. Compos. 136 (2023),
https://doi.org/10.1016/j.cemconcomp.2022.104890.
[99] G. Adil, C. Halmen, N.P. Vaddey, J. Pacheco, D. Trejo, Multi-laboratory validation
study of a CriticalChloride threshold test method, ACI Mater. J. 119 (2022),
https://doi.org/10.14359/51737195.
[100] Y. Gao, Y. Zheng, J. Zhang, J. Wang, X. Zhou, Y. Zhang, Randomness of critical
chloride concentration of reinforcement corrosion in reinforced concrete exural
members in a tidal environment, Ocean Eng. 172 (2019) 330–341, https://doi.
org/10.1016/j.oceaneng.2018.11.038.
[101] H. Fakhri, K.L. Fishman, R. Ranade, Rapid determination of critical chloride
content in cement-based composites, Construct. Build. Mater. 268 (2021),
https://doi.org/10.1016/j.conbuildmat.2020.121148.
[102] M. Castellote, C. Andrade, C. Alonso, Accelerated simultaneous determination of
the chloride depassivation threshold and of the non-stationary diffusion
coefcient values. https://doi.org/10.1016/S0010-938X(02)00060-4, 2002.
[103] P.V. Nygaard, M.R. Geiker, A method for measuring the chloride threshold level
required to initiate reinforcement corrosion in concrete, Materials and Structures/
Materiaux et Constructions 38 (2005) 489–494, https://doi.org/10.1617/14279.
[104] V.L. Jerˆ
onimo, G.R. Meira, L.C.P. da Silva Filho, Performance of self-compacting
concretes with wastes from heavy ceramic industry against corrosion by
chlorides, Construct. Build. Mater. 169 (2018) 900–910, https://doi.org/
10.1016/j.conbuildmat.2018.03.034.
[105] D. Izquierdo, C. Alonso, C. Andrade, M. Castellote, Potentiostatic determination of
chloride threshold values for rebar depassivation - experimental and statistical
study, Electrochim. Acta (2004) 2731–2739, https://doi.org/10.1016/j.
electacta.2004.01.034.
[106] V. Bouteiller, C. Cremona, V. Baroghel-Bouny, A. Maloula, Corrosion initiation of
reinforced concretes based on Portland or GGBS cements: chloride contents and
electrochemical characterizations versus time, Cement Concr. Res. 42 (2012)
1456–1467, https://doi.org/10.1016/j.cemconres.2012.07.004.
[107] P.B. Bamforth, The Derivation of Input Data for Modelling Chloride Ingress from
Eight-Year UK Coastal Exposure Trials, 1999.
[108] P. Castro-Borges, M. Balanc´
an-Zapata, A. L´
opez-Gonz´
alez, Analysis of tools to
evaluate chloride threshold for corrosion onset of reinforced concrete in tropical
marine environment of Yucat´
an, M´
exico, J. Chem. (2013), https://doi.org/
10.1155/2013/208619.
[109] U.M. Angst, B. Elsener, C.K. Larsen, Ø. Vennesland, Chloride induced
reinforcement corrosion: electrochemical monitoring of initiation stage and
chloride threshold values, Corrosion Sci. 53 (2011) 1451–1464, https://doi.org/
10.1016/j.corsci.2011.01.025.
[110] F. Lollini, E. Redaelli, L. Bertolini, Investigation on the effect of supplementary
cementitious materials on the critical chloride threshold of steel in concrete,
Mater. Struct. 49 (2016) 4147–4165, https://doi.org/10.1617/s11527-015-0778-
0.
[111] R. Liu, L. Jiang, G. Huang, Y. Zhu, X. Liu, H. Chu, C. Xiong, The effect of carbonate
and sulfate ions on chloride threshold level of reinforcement corrosion in mortar
with/without y ash, Construct. Build. Mater. 113 (2016) 90–95, https://doi.org/
10.1016/j.conbuildmat.2016.03.018.
[112] V. Garcia, R. François, M. Carcasses, P. Gegout, Potential measurement to
determine the chloride threshold concentration that initiates corrosion of
reinforcing steel bar in slag concretes, Materials and Structures/Materiaux et
Constructions 47 (2014) 1483–1499, https://doi.org/10.1617/s11527-013-0130-
5.
[113] R. Pillai, D. Trejo, Surface condition effects on critical chloride threshold of steel
reinforcement service life extension of post-tensioned concrete systems using
galvanic anode cathodic protection view project PhD work view project.
https://www.researchgate.net/publication/285640912, 2005.
[114] H. Yu, X. Shi, W.H. Hartt, B. Lu, Laboratory investigation of reinforcement
corrosion initiation and chloride threshold content for self-compacting concrete,
Cement Concr. Res. 40 (2010) 1507–1516, https://doi.org/10.1016/j.
cemconres.2010.06.004.
[115] T. Cheewaket, C. Jaturapitakkul, W. Chalee, Initial corrosion presented by
chloride threshold penetration of concrete up to 10 year-results under marine site,
Construct. Build. Mater. 37 (2012) 693–698, https://doi.org/10.1016/j.
conbuildmat.2012.07.061.
[116] J. Pacheco, R.B. Polder, Critical Chloride Concentrations in Reinforced Concrete
Specimens with Ordinary Portland and Blast Furnace Slag Cement, 2016.
[117] H.W. Song, V. Saraswathy, S. Muralidharan, K. Thangavel, Tolerance limit of
chloride for steel in blended cement mortar using the cyclic polarisation
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
21
technique, J. Appl. Electrochem. 38 (2008) 445–450, https://doi.org/10.1007/
s10800-007-9457-3.
[118] G.R. Meira, C. Andrade, E.O. Vilar, K.D. Nery, Analysis of chloride threshold from
laboratory and eld experiments in marine atmosphere zone, Construct. Build.
Mater. 55 (2014) 289–298, https://doi.org/10.1016/j.conbuildmat.2014.01.052.
[119] J. Xu, L. Jiang, W. Wang, Y. Xu, Y. Jiang, Chloride threshold value for
reinforcement corrosion in concrete with additions of silica fume or y ash, Mag.
Concr. Res. 63 (2011) 905–913, https://doi.org/10.1680/macr.10.00101.
[120] T.U. Mohammed, H. Hamada, Corrosion of steel bars in concrete with various
steel surface conditions fresh and hardened properties of micro-concrete used for
strengthening of RC members view project long-term performance of concrete
under marine environment in Bangladesh view project. https://www.researchg
ate.net/publication/280572043, 2006.
[121] B. Pradhan, B. Bhattacharjee, Rebar corrosion in chloride environment, Construct.
Build. Mater. 25 (2011) 2565–2575, https://doi.org/10.1016/j.
conbuildmat.2010.11.099.
[122] C.M. Hansson, B. Sorensen, The threshold concentration of in concrete for the
initiation of reinforcement corrosion chloride. www.astm.org, 1990.
[123] M. Manera, Ø. Vennesland, L. Bertolini, Chloride threshold for rebar corrosion in
concrete with addition of silica fume, Corrosion Sci. 50 (2008) 554–560, https://
doi.org/10.1016/j.corsci.2007.07.007.
[124] P. Sandberg, H. Sørensen, Factors affecting the chloride thresholds for uncracked
reinforced concrete exposed in a marine environment Part II: laboratory-and eld
exposure of corrosion cells, Concr. Sci. Eng. 1 (1999) 99–109.
[125] P.G. De Viedma, M. Castellote, C. Andrade, Comparison between several methods
for determining the depassivation threshold value for corrosion onset, J. Phys. IV :
JP (2006) 79–88, https://doi.org/10.1051/jp4:2006136009.
[126] P. Sandberg, Chioride Initiated Reinforcement Corrosion in Marine Concrete,
Lund University, 1998.
[127] P. Romano, P.S.D. Brito, L. Rodrigues, Monitoring of the degradation of concrete
structures in environments containing chloride ions, Construct. Build. Mater. 47
(2013) 827–832, https://doi.org/10.1016/j.conbuildmat.2013.05.042.
[128] B.H. Oh, S.Y. Jang, Y.S. Shin, Experimental Investigation of the Threshold
Chloride Concentration for Corrosion Initiation in Reinforced Concrete
Structures, 2003.
[129] A.N. S´
anchez, A.A. Sagü´
es, Chloride Threshold Dependence on Potential in
Reinforced Mortar, 2012.
[130] A.N. S´
anchez, A.A. Sagü´
es, Chloride Corrosion Threshold Dependence on Steel
Potential in Reinforced Concrete, 2014.
[131] S.A. Lee, K.P. Park, J. Kim, K.Y. Ann, Sensitivity analysis for binders in concrete
mix to the corrosion risk of steel embedment in chloride-bearing environments,
Construct. Build. Mater. 251 (2020), https://doi.org/10.1016/j.
conbuildmat.2020.118944.
[132] D. Trejo, N.P. Vaddey, C. Halmen, Standardizing test to quantify chloride
threshold of steel in concrete, ACI Mater. J. 118 (2021) 177–187, https://doi.org/
10.14359/51728283.
[133] K. Horiguchi, T. Yamaguchi, T. Maruya, K. Takewaka, Study on the method of
measuring the chloride threshold value of corrosion and estimation of the values
in durability design of concrete structures, J. Adv. Concr. Technol. 18 (2020)
571–587, https://doi.org/10.3151/JACT.18.571.
[134] D. Boubitsas, L. Tang, The inuence of reinforcement steel surface condition on
initiation of chloride induced corrosion, Materials and Structures/Materiaux et
Constructions 48 (2015) 2641–2658, https://doi.org/10.1617/s11527-014-0343-
2.
[135] D. Trejo, C. Tibbits, The Inuence of SCM Type and Quantity on the Critical
Chloride Corrosion Threshold, vol. 308, Special Publication, 2016, pp. 1–20.
[136] U.M. Angst, B. Elsener, The size effect in corrosion greatly inuences the
predicted life span of concrete infrastructures, Sci. Adv. 3 (2024) e1700751,
https://doi.org/10.1126/sciadv.1700751.
[137] T.B. Holder, The determination of chloride threshold concentrations using
different supplementary cementing materials in steel-reinforced concrete. https:
//api.semanticscholar.org/CorpusID:139528249, 1999.
[138] J. Park, J. Park, M. Jung, Variation in service life on RC structure according to
concrete binder type, Materials 13 (2020) 1–16, https://doi.org/10.3390/
ma13235430.
[139] J.S. Ryou, K.Y. Ann, Variation in the chloride threshold level for steel corrosion in
concrete arising from different chloride sources, Mag. Concr. Res. 60 (2008)
177–187, https://doi.org/10.1680/macr.2008.60.3.177.
[140] W. Morris, A. Vico, M. Vazquez, S.R. De Sanchez, Corrosion of reinforcing steel
evaluated by means of concrete resistivity measurements. www.elsevier.com/loc
ate/corsciCorrosionScience44, 2002.
[141] W. Zhu, R. François, Q. Fang, D. Zhang, Inuence of long-term chloride diffusion
in concrete and the resulting corrosion of reinforcement on the serviceability of
RC beams, Cem. Concr. Compos. 71 (2016) 144–152, https://doi.org/10.1016/j.
cemconcomp.2016.05.003.
[142] K.Y. Ann, H.W. Song, Chloride threshold level for corrosion of steel in concrete,
Corrosion Sci. 49 (2007) 4113–4133, https://doi.org/10.1016/j.
corsci.2007.05.007.
[143] J.M. Frederiksen, Chloride threshold values for service life design, in: C. Andrade,
J. Kropp (Eds.), Testing and Modelling the Chloride Ingress into Concrete, 2000,
pp. 397–414. www.aec-dk.com.
[144] M.A. Islam, B.P. Bergsma, C.M. Hansson, Chloride-induced corrosion behavior of
stainless steel and carbon steel reinforcing bars in sound and cracked concrete,
Corrosion 69 (2013) 303–312, https://doi.org/10.5006/0706E.
[145] Y. Cao, C. Gehlen, U. Angst, L. Wang, Z. Wang, Y. Yao, Critical chloride content in
reinforced concrete — an updated review considering Chinese experience,
Cement Concr. Res. 117 (2019) 58–68, https://doi.org/10.1016/j.
cemconres.2018.11.020.
[146] Z.H. Lu, P.Y. Lun, W. Li, Z. Luo, Y. Li, P. Liu, Empirical model of corrosion rate for
steel reinforced concrete structures in chloride-laden environments, Adv. Struct.
Eng. 22 (2019) 223–239, https://doi.org/10.1177/1369433218783313.
[147] N.P. Vaddey, D. Trejo, Optimizing test parameters for quantifying critical chloride
threshold, ACI Mater. J. 118 (2021) 53–66, https://doi.org/10.14359/51731547.
[148] H.S. Al-alaily, A.A.A. Hassan, A study on the effect of curing temperature and
duration on rebar corrosion, Mag. Concr. Res. 70 (2018) 260–270, https://doi.
org/10.1680/jmacr.17.00080.
[149] G. Glass, B. Reddy, N. Buenfeld, The participation of bound chloride in passive
lm breakdown on steel in concrete, Corrosion Science - CORROS SCI 42 (2000)
2013–2021, https://doi.org/10.1016/S0010-938X(00)00040-8.
[150] J. Larsson, The enrichment of chlorides in expressed concrete pore solution
submerged in saline solution, in: Proceedings of the Nordic Seminar on Field
Studies of Chloride Initiated Reinforcement Corrosion in Concrete, Lund
University of Technology, 1995, pp. 171–176. Report TVBM-3064.
[151] H. Zibara, Binding of External Chlorides by Cement Pastes, 2001.
[152] S. Caijun, Q. Yuan, F. He, X. Hu, Transport and interactions of chlorides in
cement-based materials. https://doi.org/10.1201/9780429466557, 2019.
[153] L. Cheng, I. Maruyama, Effect of relative humidity and temperature on steel
corrosion rate in chloride contaminated mortar, セメント・コンクリート論文集
75 (2022) 225–232.
[154] L. Li, A.A. Sagues, Chloride corrosion threshold of reinforcing steel in alkaline
solutions—effect of specimen size, Corrosion 60 (2004) 195–202, https://doi.org/
10.5006/1.3287721.
[155] U. Angst, A. Ronnquist, B. Elsener, C.K. Larsen, Ø. Vennesland, Probabilistic
considerations on the effect of specimen size on the critical chloride content in
reinforced concrete, Corrosion Sci. 53 (2011) 177–187, https://doi.org/10.1016/
j.corsci.2010.09.017.
[156] K.Y. Ann, T.-S. Kim, J.H. Kim, S.-H. Kim, The resistance of high alumina cement
against corrosion of steel in concrete, Construct. Build. Mater. 24 (2010)
1502–1510, https://doi.org/10.1016/j.conbuildmat.2010.01.022.
[157] G. Kakali, S. Tsivilis, E. Aggeli, M. Bati, Hydration products of C3A, C3S and
Portland cement in the presence of CaCO3, Cement Concr. Res. 30 (2000)
1073–1077, https://doi.org/10.1016/S0008-8846(00)00292-1.
[158] A. Vollpracht, B. Lothenbach, R. Snellings, J. Haufe, The pore solution of blended
cements: a review, Mater. Struct. 49 (2015), https://doi.org/10.1617/s11527-
015-0724-1.
[159] C.L. Page, Ø. Vennesland, Pore solution composition and chloride binding
capacity of silica-fume cement pastes, Mat´
eriaux et Construction 16 (1983)
19–25, https://doi.org/10.1007/BF02474863.
[160] R.K. Dhir, M.A.K. El-Mohr, T.D. Dyer, Developing chloride resisting concrete
using PFA, Cement Concr. Res. 27 (1997) 1633–1639, https://doi.org/10.1016/
S0008-8846(97)00146-4.
[161] S. Mahima, P. Moorthi, A. Bahurudeen, A. Gopinath, Inuence of chloride
threshold value in service life prediction of reinforced concrete structures,
Sådhanå 43 (2018) 1–19.
[162] S. Rasheeduzzafar, Ehtesham Hussain, S.S. Al-Saadoun, Effect of cement
composition on chloride binding and corrosion of reinforcing steel in concrete,
Cement Concr. Res. 21 (1991) 777–794, https://doi.org/10.1016/0008-8846(91)
90173-F.
[163] M.M.H.A.F. Dehwah, S.A. Austin, Effect of cement alkalinity on pore solution
chemistry and chloride-induced reinforcement corrosion, ACI Mater. J. 99 (2002),
https://doi.org/10.14359/11967.
[164] P. Ghods, O.B. Isgor, G.A. McRae, J. Li, G.P. Gu, Microscopic investigation of mill
scale and its proposed effect on the variability of chloride-induced depassivation
of carbon steel rebar, Corrosion Sci. 53 (2011) 946–954, https://doi.org/
10.1016/j.corsci.2010.11.025.
[165] E. Mahallati, M. Saremi, An assessment on the mill scale effects on the
electrochemical characteristics of steel bars in concrete under DC-polarization,
Cement Concr. Res. 36 (2006) 1324–1329, https://doi.org/10.1016/j.
cemconres.2006.03.015.
[166] R. Rodrigues, S. Gaboreau, J. Gance, I. Ignatiadis, S. Betelu, Reinforced concrete
structures: a review of corrosion mechanisms and advances in electrical methods
for corrosion monitoring, Construct. Build. Mater. 269 (2021) 121240, https://
doi.org/10.1016/j.conbuildmat.2020.121240.
[167] A. Poursaee, A. Laurent, C.M. Hansson, Corrosion of steel bars in OPC mortar
exposed to NaCl, MgCl2 and CaCl2: macro- and micro-cell corrosion perspective,
Cement Concr. Res. 40 (2010) 426–430, https://doi.org/10.1016/j.
cemconres.2009.09.029.
[168] C.M. Hansson, Th Frølund, J.B. Markussen, The effect of chloride cation type on
the corposion of steel in concrete by chloride salts, Cement Concr. Res. 15 (1985)
65–73, https://doi.org/10.1016/0008-8846(85)90009-2.
[169] R.M. Andrew, Global CO_2 emissions from cement production, Earth Syst. Sci.
Data 10 (2018) 195–217, https://doi.org/10.5194/essd-10-195-2018.
[170] C.K. Larsen, Chloride binding in concrete, Effect of surrounding environment and
concrete composition, Dr. Ing. Thesis 95 (1998) 337. NTNU, Trondheim.
[171] K. Byfors, Inuence of silica fume and yash on chloride diffusion and pH values
in cement paste, Cement Concr. Res. 17 (1987) 115–130, https://doi.org/
10.1016/0008-8846(87)90066-4.
[172] C.L. Page, Ø. Vennesland, Pore solution composition and chloride binding
capacity of silica-fume cement pastes, Mat´
eriaux et Construction 16 (1983)
19–25. https://api.semanticscholar.org/CorpusID:136964787.
[173] A. Scott, The inuence of binder type and cracking on reinforcing steel corrosion
in concrete. https://api.semanticscholar.org/CorpusID:136265318, 2004.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
22
[174] K. Kopecsk´
o, G. Bal´
azs, Concrete with improved chloride binding and chloride
resistivity by blended cements, Adv. Mater. Sci. Eng. 2017 (2017) 1–13, https://
doi.org/10.1155/2017/7940247.
[175] J.H. Potgieter, D.J. Delport, S. Verryn, S. Potgieter-Vermaak, Chloride-binding
effect of blast furnace slag in cement pastes containing added chlorides, South
African, J. Chem. 63 (2011) 108.
[176] C. Arya, N.R. Buenfeld, J.B. Newman, Factors inuencing chloride-binding in
concrete, Cement Concr. Res. 20 (1990) 291–300. https://api.semanticscholar.or
g/CorpusID:96303080.
[177] R. Luo, Y.B. Cai, C. Wang, X. Huang, Study of chloride binding and diffusion in
GGBS concrete, Cement Concr. Res. 33 (2003) 1–7. https://api.semanticscholar.
org/CorpusID:98518912.
[178] A. Cheng, R. Huang, J.-K. Wu, C.-H. Chen, Inuence of GGBS on durability and
corrosion behavior of reinforced concrete, Mater. Chem. Phys - MATER CHEM
PHYS 93 (2005) 404–411, https://doi.org/10.1016/j.matchemphys.2005.03.043.
[179] R.K. Dhir, M.R. Jones, Development of chloride-resisting concrete using y ash,
Fuel 78 (1999) 137–142, https://doi.org/10.1016/S0016-2361(98)00149-5.
[180] C. Arya, N.R. Buenfeld, J.B. Newman, Factors inuencing chloride-binding in
concrete, Cement Concr. Res. 20 (1990) 291–300, https://doi.org/10.1016/0008-
8846(90)90083-A.
[181] K. Byfors, Inuence of silica fume and yash on chloride diffusion and pH values
in cement paste, Cement Concr. Res. 17 (1987) 115–130, https://doi.org/
10.1016/0008-8846(87)90066-4.
[182] S. Diamond, Effects of two Danish yashes on alkali contents of pore solutions of
cement-yash pastes, Cement Concr. Res. 11 (1981) 383–394. https://api.
semanticscholar.org/CorpusID:94911289.
[183] M. Kawamura, O.A. Kayyali, M.N. Haque, Effects of a yash on pore solution
composition in calcium and sodium chloride-bearing mortars, Cement Concr. Res.
18 (1988) 763–773. https://api.semanticscholar.org/CorpusID:98006196.
N. Maamary and I.G. Ogunsanya
Cement and Concrete Composites 154 (2024) 105796
23
