Predicting carbonation coefficient using Artificial neural networks and genetic programming
Introduction
Carbonation-induced corrosion and chloride-induced corrosion are considered the major causes of deterioration in reinforced concrete structures [1]. Therefore, concrete carbonation is a subject of great concern. Carbonation is a natural physicochemical process caused by the penetration of carbon dioxide from the surrounding environment into concrete through pores in the matrix, where the carbon dioxide reacts with hydrated cement. Calcium hydroxide (Ca (OH)2) in contact with carbon dioxide (CO2) forms calcium carbonate (CaCO3). This reaction neutralizes the natural protection of reinforcement steel provided by concrete, which in turn promotes the initiation of corrosion of the steel bars [2]; Sistonen et al., 2009; [3,4]. Conventionally, concrete carbonation depth at a given time under steady-state conditions can be reasonably estimated using Eq. (1). This well-known equation is based on Fick's second law of diffusion [1,3,5,6].where xc(t) is the carbonation depth at the time t [mm], k is coefficient of carbonation [mm/day0.5] and t is the duration of carbonation [days].
In this equation, the carbonation coefficient encompasses concrete quality (e.g. water/cement (w/c) ratio, binder type, and binder content) and the environmental conditions (e.g. temperature, relative humidity and concentration of carbon dioxide) [1].
It is generally agreed that carbonation does not occur in the same way in all mixes, nor does it occur in all circumstances. Different mixes, as well as the same mix exposed to different environments, will lead to distinct carbonation rates [7,[8], [49]]. The carbonation coefficient may significantly vary from one concrete element to another depending on the environmental exposure conditions and micro structural parameters, which are linked with the concrete composition and the type of materials used. The main factors affecting concrete carbonation are: the porous system of hardened concrete, which in turn depends on the w/c ratio, the type of binder, the amount of Ca(OH)2 and the degree of hydration; the relative humidity (for dissolution of Ca(OH)2); and the concentration of CO2 [5,9].
Developing an advanced carbonation prediction model that encompasses all the governing parameters for the carbonation coefficient is a difficult task, mainly because several parameters are complex to describe mathematically. Various theoretical and experimental studies have been carried out to estimate the concrete carbonation depth [[10], [11], [12], [13], [14], [15], [16]]. Most carbonation models are semi-empirical, i.e. their development starts from a theoretical basis, which is completed by fitting the required parameters to experimental results. In this sense, new models must be defined in order to describe the complex behaviour of the carbonation coefficient, which in turn will focus on structural health monitoring. The recent data driven modelling techniques of Artificial Neural Networks (ANNs) and Genetic Programming (GP) can be the most suitable way for this task owing to their learning ability from the available data and model free structure.
Earlier attempts were made to predict carbonation depth with relevant input parameters (Lu et al., 2014; [6,17]. For example, in Kwon and Song [18]; CO2 diffusion coefficients were estimated through a neural network algorithm, with three components in OPC mix design and relative humidity as inputs neurons. The results show a reasonable decrease in the diffusion coefficient with higher relative humidity and lower w/c ratio, with maximum differences of 6.3% between estimated and experimental data. A two-phase model was developed by Seung and Wang (2016) [45], in which the strength of slag concrete was predicted in phase 1 while, in phase 2, the results from the hydration model were used as input parameters for the carbonation reaction model with different curing conditions and slag contents. Further studies in this area suggest that ANNs can be an effective tool to predict the carbonation depth in concrete since they can capture the complex relationship amongst the parameters to predict the output (Luo et al., 2016; [19,20]. Several studies were conducted for determining the strength of concrete with different materials, testing methods, displaying good performance of the models [[21], [22], [43]]. Another technique that has been widely used to determine the strength of concrete is Genetic Programming [23]; Kulkarni and Londhe, 2018). Genetic algorithms were used to obtain the optimum concrete mix proportions, considering the exposure conditions of carbonation and design parameters and determining the intended diffusion coefficients [24].
In previous studies, there seems to be no consensus on the way the carbonation coefficient should be determined. This coefficient is fundamentally a durability indicator that comprises all the variables relating to the environmental severity and the characteristics of concrete itself [[12], [42]]. Artificial Neural Networks (ANNs) were usually used to predict the carbonation depth, but few works are available in the literature related with the use of Genetic Programming (GP) to predict the carbonation coefficient even though GP has been successfully used as a prediction/forecasting tool in many works [25]. The output in ANNs is in terms of weights and biases while that of GP in the form of equations, which can be easily used for replicating the study or to apply the results obtained in other contexts. The current research intends to develop models to predict the carbonation coefficient and compressive strength of concrete using Artificial neural networks (ANNs) and Genetic programming (GP) and compare their results. Additionally, a knowledge extraction exercise is carried out using the trained weights and biases of the ANNs, which serves two-fold purposes. First, the influence of each parameter on the carbonation coefficient can be determined, which allows evaluating the coherence of the models obtained. Secondly, the mathematical complexity of ANNs can be simplified, as a modelling tool, if the first purpose is accomplished. The influence of the various parameters adopted in the GP model is also discussed in the present study.
Some of the authors of this study developed multiple linear regression (MLR) models, as a first approach, to predict the carbonation coefficient and compressive strength of concrete [26]. In the previous study, the authors acknowledged the influence of the relative humidity of the exposure environment in the carbonation phenomenon; however, the MLR model proposed, due to its linear nature, was unable to model the nonlinear effect of the relative humidity in the concrete carbonation rate, and the sample analysed was divided into two samples, leading to two different models, one for cases in which the relative humidity is 70% or less, and the other for cases where the relative humidity is higher than 70%. In this sense, it was mentioned that a nonlinear model could also have been used, based on a polynomial or even an exponential function, among other options, to overcome the limitation of a linear model as MLR. The current work is thus an extension of the mentioned work, in which the carbonation coefficient and strength of concrete are predicted using Artificial Neural Networks (ANNs) and Genetic Programming (GP). The performance of the models developed using ANNs and GP are compared with models developed using MLR. The 28-day compressive strength of concrete was further predicted using the water-binder ratio, clinker ratio in the binder (%) and curing time (days) as input parameters [26].
Section snippets
Modeling techniques
In this study, the prediction of the carbonation coefficient and strength of concrete is performed using Artificial Neural Networks (ANNs) and Genetic Programming (GP). These approaches are described, in brief, in the next sections.
Data used in the study
The current work is an extension of the work earlier performed to predict the carbonation coefficient and strength of concrete using multiple linear regression (MLR) [26]. The data used in the present work are the same as used by Silva et al. [26]; which was collected from 17 previous studies, comprising concrete data with different binders and binder contents, water/binder ratios, compressive strengths, curing periods and slumps. Around 50% of the data are from concrete exposed to natural (in
Methodology adopted
The earlier attempt made by Silva et al. [26] to predict the carbonation coefficient, considered the variables or factors affecting the carbonation coefficient (kw) and the compressive strength (fc) of concrete. In the current work, the same variables are considered, identifying those with a higher influence to predict the output variables. The data mentioned in the previous section were used and a total of five models were developed using ANNs and GP. The details of the models with input,
Prediction of the carbonation coefficient
The performance of the models in SET 1, to predict the carbonation coefficient, is shown in Table 3. ANNs show a better performance in capturing the relationship between the carbonation coefficient and the variables included in the model, with a correlation coefficient of 0.940, as seen in the ANNCA1 model. ANNs display the output in terms of weights and biases; on the other hand, GP can provide a mathematical equation to predict a given output. Its codifying manner enables them to be easily
Conclusion
This study is a first approach towards modelling the phenomenon of concrete carbonation using data-driven techniques. Models using data-driven techniques, namely Artificial Neural Networks and Genetic Programming, were developed to predict the carbonation coefficient with the goal of making service life predictions. Five sets of models were developed and analysed, to predict the carbonation coefficient and to predict concrete strength, using ANNs and GP.
In the different models proposed, in sets
CRediT authorship contribution statement
S.N. Londhe: Conceptualization, Methodology, Software, Validation, Writing - review & editing. P.S. Kulkarni: Conceptualization, Methodology, Software, Validation, Writing - original draft. P.R. Dixit: Conceptualization, Methodology, Software, Validation, Writing - original draft. A. Silva: Formal analysis, Visualization, Writing - review & editing. R. Neves: Formal analysis, Visualization, Writing - review & editing. J. de Brito: Formal analysis, Visualization, Writing - review & editing.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
The authors gratefully acknowledge the support of the CERIS Research Institute, IST, University of Lisbon and the FCT (Foundation for Science and Technology) through the project BestMaintenance-LowerRisks (PTDC/ECI-CON/29286/2017).
References (46)
Time dependent reliability of structural systems subject to deterioration
Comput. Struct.
(1998)- et al.
Service life prediction of a concrete bridge structure subjected to carbonation
Construct. Build. Mater.
(2010) - et al.
Numerical analysis of the effect of coarse aggregate distribution on concrete carbonation
Construct. Build. Mater.
(2012) - et al.
Carbonation of concrete and its prediction
Cement Concr. Res.
(1987) - et al.
Statistical analysis of the carbonation coefficient in open air concrete structures
Construct. Build. Mater.
(2012) - et al.
The experimental investigation of concrete carbonation depth
Cement Concr. Res.
(2006) - et al.
Prediction model of carbonation depth for recycled aggregate concrete
Cement Concr. Compos.
(2018) - et al.
Model for practical carbonation depth prediction for high volume fly ash concrete and recycled aggregate concrete
Construct. Build. Mater.
(2019) - et al.
Analysis of carbonation behavior in concrete using neural network algorithm and carbonation modeling
Cement Concr. Res.
(2010) - et al.
CO2 uptake potential due to concrete carbonation: a case study
Case Stud. Construct. MAter,
(2017)
Neural networks for predicting properties of concretes with admixtures
Construct. Build. Mater.
Soft computing approach for real-time estimation of missing wave heights
Ocean Eng.
Statistical modeling of carbonation in reinforced concrete
Cement Concr. Compos.
Neural networks and M5 model trees in modeling water-level- discharge relationship
Neurocomputing
Assessing concrete carbonation resistance through air permeability measurements
Construct. Build. Mater.
Effect of carbonation on the rebound number and compressive strength of concrete
Cement Concr. Compos.
Modelling compressive strength of recycled aggregate concrete by artificial neural network, model tree and non-linear regression
Int. J. Sustain. Built Environ.
Evaluation of compressive strength development and carbonation depth of high volume slag-blended concrete
Construct. Build. Mater.
A method for the use of accelerated carbonation tests in durability design
Construct. Build. Mater.
Structural Concrete: Textbook on Behaviour, Design and Performance’, Fib Bulletin,1, Bulletin 51, V. 1
Properties of Concrete
Optimized neural network-based carbonation prediction model
The Air Permeability and Concrete Carbonation of Concrete in Structures
Cited by (23)
State-of-the-art review of geopolymer concrete carbonation: From impact analysis to model establishment
2024, Case Studies in Construction MaterialsPollutant concentration prediction from traffic data analysis for concrete durability studies in Madrid Calle 30 urban tunnels
2024, Tunnelling and Underground Space TechnologyPrediction and optimization model of sustainable concrete properties using machine learning, deep learning and swarm intelligence: A review
2023, Journal of Building EngineeringInhibition resistance and mechanism of migrating corrosion inhibitor on reinforced concrete under coupled carbonation and chloride attack
2023, Journal of Building EngineeringArtificial neural network model to estimate the long-term carbonation depth of concrete exposed to natural environments
2023, Journal of Building Engineering