Carbonation is a deleterious concrete durability problem which may alter concrete microstructure and yield initiation of corrosion in reinforcing steel bars. Previous studies focused on the use of Artificial Neural Networks (ANN) for the prediction of concrete carbonation depth and to minimize the need for destructive and elaborated civil engineering laboratory tests. This study aims to provide improved accuracy of simulation and prediction of carbonation with an ANN architecture including eighteen input parameters employing alternative Scaled Conjugate Gradient (SCG) function. After ensuring a promising value of the coefficient of correlation as high as 0.98, the influence of proposed input parameters on the progress of carbonation depth was studied. The results of this parametric analysis were observed to successfully comply with the conventional civil engineering experience. Hence, the employed ANN model can be used as an efficient tool to study in detail and to provide insights into the carbonation problem in concrete.
Carbonation problem in concrete structures results in the formation of calcium carbonate due to the neutralization reaction between calcium hydroxide and carbonic acid, which yielded by the ingress of CO2 gas from the atmosphere into the concrete microstructure. (
Determining the progress of concrete carbonation depth, which is known to depend on several factors, is critical for the evaluation of the performance of reinforced concrete structures throughout their service lives; since the initiation of carbonation-induced corrosion of reinforcing bars might cause severe damages in concrete (
Artificial neural networks (ANN) applications have been used effectively for several decades for simulation and prediction of varying case studies in both applied and social sciences (
This study, proposes an ANN model that considers eighteen environmental and concrete-related input parameters; and based on their selection, provides an estimation for the extent of the resultant concrete carbonation, as an efficient alternative to conventional laboratory experiments done by phenolphthalein indicator methods. Beside its advantageous use in carbonation depth determination, a viable intelligent prediction method has also the potential to enable researchers working in the field of concrete carbonation to gain further insights on the effect of each influencing parameters in a systematically way; without needing to carry out the “time, resources and labour-consuming” laboratory experiments. This study, which is a continuation of previous work (
The progress of carbonation front is known to be governed by cement composition, the composition of additives, concrete mix properties, curing conditions, environmental factors such as relative humidity and temperature, as well as the duration of carbonation attack (
Descriptive information on the input and output variables defined in the proposed model and the references indicating their relevance to concrete carbonation
Variables | Unit | Min | Max | Mean | Std. D | Related References |
---|---|---|---|---|---|---|
Cement CaO content | (%) | 45.7 | 68.02 | 62.19 | 3.63 | ( |
Cement SiO2content | (%) | 20.21 | 31.2 | 21.71 | 2.45 | ( |
Cement Fe2O3 content | (%) | 1 | 6.15 | 3.41 | 1.03 | ( |
Cement Al2O3 content | (%) | 3.1 | 9.2 | 4.71 | 1.28 | ( |
Fly Ash CaO content | (%) | 0 | 5.2 | 1.43 | 1.65 | ( |
Fly Ash SiO2content | (%) | 0 | 55.3 | 23.98 | 25.85 | ( |
Fly Ash Fe2O3 content | (%) | 0 | 13.92 | 4.75 | 5.55 | ( |
Fly Ash Al2O3 content | (%) | 0 | 30.08 | 12.79 | 13.78 | ( |
Total cement content | (kg/m3) | 0 | 280 | 50.19 | 83.38 | ( |
Total fly ash content | (kg/m3) | 67 | 486 | 300.08 | 98.31 | ( |
Water content | (kg/m3) | 102.12 | 220 | 167.87 | 27.19 | ( |
Water/binder ratio | (%) | 0.28 | 0.84 | 0.49 | 0.11 | ( |
Curing time | (days) | 1 | 90 | 31.16 | 22.22 | ( |
Curing relative humidity | (%) | 50 | 100 | 89.38 | 18.40 | ( |
External temperature | (oC) | 12 | 32.5 | 20.67 | 5.19 | ( |
External CO2content | (%) | 50 | 100 | 67.37 | 9.09 | ( |
External relative humidity | (%) | 0.03 | 100 | 20.03 | 24.72 | ( |
Carbonation duration | (days) | 3 | 2070 | 180.31 | 334.60 | ( |
Carbonation depth | (mm) | 0 | 64.03 | 16.36 | 14.84 |
In neural nets modelling, normalization of data is important to avoid early saturation within the hidden layer, which causes slower convergence (
In this study, feed-forward back propagation was used for two optimization functions. Results obtained with Levenberg-Marquardt Backpropagation (LM), which is commonly used in ANN studies of different fields due to yielding a lower number of iterations, were used for comparison with the results obtained from an alternative function, Scaled Conjugate Gradient (SCG) suggested in this study.
The network contains three (input-hidden-output) layers with one hidden layer. The input layer represents the sources node, which receives signals from the external environment. Within this layer, 18 nodes (see
Architecture of the proposed model.
Three different learning schemes (LS) were adopted with training/testing distributions, as LS1 being 40:60; LS2 being 50:50; and LS3 being 60:40 % of the sample set. Hence, combinations of two training algorithms with three learning schemes with a constant hidden neuron value were studied with each optimization function (see
Results for R, MSE and number of iterations with proposed learning schemes
Learning Scheme | Optimization Method | R | MSE | iter. |
---|---|---|---|---|
SCG | 0.964 | 0.0039 | 467 | |
LM | 0.957 | 0.0045 | 40 | |
SCG | 0.975 | 0.0026 | 789 | |
LM | 0.976 | 0.0026 | 44 | |
SCG | 0.983 | 0.0019 | 1581 | |
LM | 0.985 | 0.0016 | 41 |
Parametric analysis study was made to verify the robustness and generalization ability of the proposed ANN architecture, in addition, to provide insights into the effect of selected key input parameters on the evolution of carbonation depth. The individual effects of selected key input parameters were studied systematically by running the model with a varying parameter of interest while keeping the other independent parameters constant. For instance, while an external parameter such as temperature, CO2 content or relative humidity (inputs 15–17) were being varied in order to investigate their effects on carbonation progress, the other concrete related parameters (inputs 1–14) were kept constant. On the other hand, in order to study the effect of other interdependent parameters, the dependency of the related parameter was also considered in order to ensure the consistency of the findings. For instance, in the case of variation of cement content (kg/m3), the water content (kg/m3) was also varied accordingly, in order to be able to keep a
Results for correlation coefficient (R), mean squared error (MSE) and total network iteration for the two studied optimization functions are presented in
It is observed from
It was observed that the correlation coefficient values yielded with both LM and SCG functions were very close to each other within each individual case The results presented in
Training and testing subset network performance and the respective regression evolution for this best case for Scaled Conjugate Gradient (SCG) function are presented in
The best prediction case using SCG with 60:40 data distribution.
A partially comparable study was previously presented by Kellouche et al. (2017) (
After the determination of the best-case combination of the model yielding highest prediction accuracy for the use of the group of Scaled Conjugate Gradient Backpropagation, the model was further employed to study the progress of carbonation depth in concrete with the individual influence of each one of eight selected input parameters. The study focused only on the effects of CO2 content, external temperature and humidity, w/c ratio of concrete, fly ash inclusion, the cement content of concrete, as well as the contents of two well-known compounds of cement, that are CaO and SiO, with the intention of providing information primarily on the presented prominent parameters at this stage. Results obtained within this parametric sensitivity study are presented in
Effect of cement CaO content on carbonation.
Similarly,
Effect of cement CaO content on carbonation.
Cement types are mainly classified according to their chemical compositions. Both Silva et al. (2014) (
Effect of cement content on carbonation depth.
Effect of FA content on carbonation depth.
Effect of CO2 content on carbonation depth.
Effect of temperature on carbonation depth.
According to these obtained results, the carbonation values that are presented
Effect of relative humidity on carbonation.
Effect of water/binder ratio on carbonation.
In this study, primarily the influence of eight selected input parameters on the progress of concrete carbonation has been presented in order to investigate the feasibility and the efficiency of the proposed carbonation depth prediction model to be employed for further understanding the fundamentals of this concrete problem. Different than other previously suggested ANN models, the compositional information of both cement and potentially added fly ash have also been considered in this study, enabling to use the model by differentiating between different cement types that could be used in real cases. High overall prediction accuracy of the proposed model, as well as the compliance of the findings of generated parametric analyses with other reported independent experimental results in the literature, imply the possibility of successful employment of the proposed ANN model to investigate the progress of carbonation problem in concrete under numerous combinations of influencing parameters in an efficient way.
In this study, a novel ANN model including cement and fly ash compound composition as influencing input parameters was used by employing two different optimization functions. Relatively a large data set acquired systematically from the literature was used in order to ensure the representability and the reliability of the calculations carried out. An increase in training dataset from 40% to 60%, improves the network prediction performance both for the aspects of R and MSE values, with both models using Scaled Conjugate Gradient (SCG) and Levenberg-Marquardt (LM) functions. Significantly high R values, such as more than 0.98, together with significantly low MSE values obtained in this study for both SCG and LM functions, strongly suggest that the model with proposed architecture can be efficiently used for the prediction of carbonation depth in concrete. The insignificant difference between of LM’s and SCG’s accuracies, indicates the convenience of the use of alternative learning algorithms such as SCG in these fields of studies, since the level accuracy is a more critical efficiency criterion than the speed of convergence is for civil engineering applications.
Parametric studies carried out after the verification of the prediction efficiency of the proposed model provided noteworthy findings. Increasing cement’s CaO content of has been observed to yield an increasing tendency of carbonation depth of concrete. Moreover, concrete mixes with relatively higher CaO contents were observed to yield significantly higher carbonation depths even in the earlier ages. An increase in the cement’s SiO2 content has also been observed to an increase in concrete carbonation. This increase in carbonation depth was observed to start ceasing beyond a SiO2 content value of 24% in this study. Further investigations on binders’ compounds compositions have the potential to provide an improved understanding on the characteristics and the types of binders that should be sought after under critical carbonation exposure conditions. Carbonation depth in concrete was observed to be decreasing with increasing cement content for all ages. On the other hand, an increasing fly ash content was observed to increase the progress of carbonation. The effect of fly ash addition became more evident primarily at later ages such as 545 and 730 days.
The results obtained by the parametric analyses carried out with the proposed ANN model also showed that an increase in CO2 content within the range from 0.03% (i.e. natural value) up to 3% (i.e. hundred times more than the natural CO2 content), yielded a steady linear increase in carbonation depth of concrete. The actual depth of carbonation in millimeters yielded in this studied CO2 content range was observed to be relatively less remarkable when compared with the carbonation depth values yielded with during the parametric analyses of other inputs. Therefore, it is suggested that even though the presence of CO2 is essential for the occurrence of carbonation, the severity of the carbonation progress in concrete is primarily defined by a combination of other considered parameters’ critical values, rather than solely being due to an increased CO2 content. On the other hand, the variations in other external exposure conditions such as temperature, has been observed to yield a relatively much higher increase in the depth of carbonation in concrete. This increase in carbonation depth was observed to be with changing tendencies, when the range of temperature was extended to include higher values. It was also observed that an increase in the external relative humidity up to the value of 75%, yielded increased depths of concrete carbonation. However, when this value was exceeded a decrease in the progress of carbonation was experienced. It was also observed that concrete mixes having higher water/binder ratios were yielding an increased level of carbonation, as expected.
In addition to the obtained high general prediction accuracy, the high level of compliance of the findings obtained by the model on carbonation progress under the selected prominent parameters with the results reported by independent previous experimental studies in the literature, was interpreted as an indication of the proposed model’s potential success in investigating concrete carbonation cases with varying combinations of parameter values in a reliable way. Hence, the employed model has the potential to serve as an efficient tool to study carbonation problem in concrete, providing further insights on the fundamentals of this critical durability problem without using any elaborated, time and resources-consuming laboratory experiments
Further studies should be carried out in order to provide understanding on the remaining input parameters considered in the proposed model in this study. Also, the studies should be extended to include concrete parameters such as aggregate properties, loss on ignition of binders, as well as the cases of concrete mixes prepared with chemical admixture inclusions. Further improvement of the intelligent studies on concrete carbonation might be explored in the future studies by employing other promising methods such as deep learning and heuristic algorithms. Moreover, a detailed set of statistical studies that could be carried out in parallel with the presented ANN study, is expected to provide an understanding of interrelations of all parameters considered in an enhanced way.