Concrete strength and durability prediction through deep learning and artificial neural networks

Abstract

The mechanical and durability characteristics of concrete are crucial for designing and evaluating concrete structures throughout their entire operational lifespan. The main objective of this research is to use the deep learning (DL) method along with an artificial neural network (ANN) to predict the chloride migration coefficient and concrete compressive strength. An expansive experimental database of nearly 1100 data points was gathered from existing scientific literature. Four forecast models were created, utilizing between 10 and 12 input features. The ANN was used to address the missing data gaps in the literature. A comprehensive pre-processing approach was then implemented to identify outliers and encode data attributes. The use of mean absolute error (MAE) as an evaluation metric for regression tasks and the employment of a confusion matrix for classification tasks were found to produce accurate results. Additionally, both the compressive strength and chloride migration coefficient exhibit a high level of accuracy, above 0.85, in both regression and classification tasks. Moreover, a user-friendly web application was successfully developed in the present study using the Python programming language, improving the ability to integrate smoothly with the user’s device.

Graphical abstract

Keywords

chloride migration coefficient / compressive strength / concrete / artificial neural network / deep learning

1 Introduction

Despite the recent advancements in construction materials, concrete is still the most commonly used building material worldwide. The demand for this material is expected to reach 18 billion tons by 2050, primarily due to the growth in global population and urbanization [1,2]. To improve the design of concrete structures, it is crucial to have a clear understanding of the performance of these structures by accurately evaluating their mechanical specifications. Concerning the mechanical and durability specifications of concrete, the compressive strength and resistance to chloride ingress are two key parameters that directly affect the safety and functionality of structures during their lifespan [3,4]. The usual methods for assessing concrete compressive strength and chloride diffusion are physical tests on cured concrete samples at specific conditions and ages [5,6]. External factors such as specimen preparation and test environment often affect these tests. Also, these tests frequently involve the destruction of the specimen. Therefore, scientists suggest practical regression procedures to estimate chloride resistance and compressive strength of concrete [5–7]. However, these methods have a disadvantage because they can produce inaccurate results due to the nonlinear relationship between the ingredients used to create concrete and the resulting strength and durability characteristics. Nevertheless, there are alternative methods for predicting the behavior of concrete, including numerical simulation techniques. However, achieving a proper agreement between the modeling and actual conditions can be challenging due to concrete’s random and nonlinear nature [8–10].

Thanks to the increase in computational power and the rise of big data analytics, artificial intelligence (AI) has experienced significant growth over recent years [11]. AI refers to the ability of a machine to perform tasks that typically require human intelligence, such as learning, planning, reasoning, and creativity [12]. Many AI applications include personalized shopping tools, smartphones, educational evaluation devices, and self-driving cars. Along with advancements in AI technology, machine learning (ML) techniques have also been improved in solving real-world problems such as statistics, robotics, bioinformatics, and construction materials. Interestingly, ML has also been implemented in structural engineering fields like seismic performance assessment [13], modeling of shear [14], tension [15], and compression strengths [16], vibration suppression [17], structural system identification [18], etc. Compared to common experimental models and regression approaches, ML has numerous advantages like cost-effectiveness, high responsiveness, and better performance speed, featuring distinct algorithms that learn from data and replicate output results with higher accuracy [19–21]. The algorithms in ML allow machines to get the required knowledge by applying procedures based on a sufficient number of data samples and the theory of probability, as opposed to using symbolic approaches [22–25]. As a powerful subset of ML algorithms, deep learning (DL) is used in diagnosis, classification, and solving manipulated features in complicated ML lineups [26]. DL was initially inspired by developments in neuroscience and, thus, associated with selecting, organizing, and interpreting information similar to the nervous system. DL techniques are capable of dealing with networks that can learn from raw data, extracting optimum input directly from unstructured and unsupervised data with no user interference. Accordingly, DL supports both feature-output relations and feature extraction procedures [27,28]. In this respect, artificial neural network (ANN) has been proven to be an efficient tool in handling complex engineering problems [29].

Several studies have explored the application of AI in specific fields of expertise. For example, Deng et al. [30] efficiently estimated the compressive strength of concrete containing recycled aggregates (RA). The type and replacement percentage of RA were considered as input parameters. Using soft computing techniques, Kaloop et al. [31] predicted the resilient modulus of RA concrete (RAC). Chou et al. [14] used SVM to estimate the shear strength of deep reinforced concrete beams. Chou et al. [32] examined the compressive strength of high-performance concrete (HPC) using ML by considering the main mix ingredients and concrete age as input. Gholampour et al. [33] investigated the compressive and tensile strengths and the Young’s modulus of RAC using a decision tree-based algorithm. Several other studies [34–36] have also estimated HPC using either of the data mining ML models with concrete mix ingredients and curing age as input variables, reporting satisfactory results in predicting compressive strength. Dantas et al. [37] implemented ANN to study the strength of concrete made with demolition wastes. They noted that ANN has the potential to predict compressive strength at different early and old ages. Duan et al. [38] assessed different mechanical properties of RAC by various inputs using a back-propagation neural network (BPNN), reporting the viability of this approach in predicting the RAC specifications. Naderpour et al. [39] studied RAC via BPNN with six input parameters, evaluating the effect of each variable on the compressive strength. It was concluded that water absorption mainly affects compressive strength compared with other variables. Topçu and Saridemir [40] used an AI fuzzy logic model and compared RAC’s compressive and tensile strengths at different ages with those of the BPNN approach. The evaluation metrics proved the robust performance of BPNN. Nikoo et al. [41] reported the ANN model’s higher accuracy, competency, and flexibility in estimating concrete compressive strength. Applying ANN versus multiple regression analyses to predict the compressive strength of HPC, Chithra et al. [42] stated that ANN reflects lower mean absolute percentage error (MAPE) and root-mean-square deviation (RMSD) compared to the regression models. The better efficacy and widespread use of ANNs in various fields have been widely validated in scholarly literature compared to alternative ML techniques [43–45]. Guo et al. [46] proposed a deep collocation method (DCM) addressing the plate bending problem using BPNN algorithms in DL based on a novel deep neural network (DNN). They could solve the complications in traditional mesh-based approaches by approximating the Kirchhoff plate deflections with different geometries. In an alternative study, DCM was employed for heterogeneous porous media [47] and heat transfer analysis of Sisko fluids [48]. Samaniego et al. [49] applied DNNs in the approximation of the solution of Partial Differential Equations. In a comprehensive study, Zhuang et al. [50] proposed a deep autoencoder-based energy method in the vibration analysis of plates with various geometries, load conditions, and boundary conditions featuring unsupervised learning methodology based upon randomly generated points within the investigated domain. Guo et al. [51] presented a DL model for the heat transfer analysis of functionally graded materials with exponential material variations. The proposed method was used to predict flux and temperature distributions as an alternative model in transient dynamic analysis.

1.1 Research significance

Despite the abundance of experimental research on concrete strength and durability parameters, there is little investigation into the application of AI in these parameters. The present study proposes a novel DL-based approach and ANN methodology in predicting the chloride migration coefficient and compressive strength of concrete through two distinct regression and classification tasks using a data set of over 1000 points from the extant literature. The literature’s missing data gap was filled using the ANN method to ensure a comprehensive set of variables for the prediction process. The accuracy of ANN relies mainly on the data set, which consists of around 1100 data points gathered from literature sources. Afterwards, 12 and 10 distinct input parameters were employed to predict the chloride migration coefficient and compressive strength, respectively, including different mix ingredients and other concrete mechanical and age parameters. The comprehensiveness and diversity of the input data set contributed to the increased robustness and reliability of the results. This investigation offers a deeper and more extensive comprehension of the correlation between input characteristics and target parameters by considering various prediction tasks. In the end, a user-friendly web application was developed utilizing the Python programming language, which gives a more remarkable ability to integrate with the user’s device. This way, the user can conveniently enter the input parameters and obtain the prediction results for both the chloride migration coefficient and compressive strength of concrete.

2 Methodology

2.1 Data set collection

ML method heavily relies on data sets, which makes high-quality training data a crucial factor for successful training. A well-prepared data set leads to more accurate trained models, better convergence, unbiased results, and fair outcomes. As part of the investigation discussed in this paper, a significant amount of experimental database has been acquired. The data set, which is available as a supplementary file that is attached to this paper, consists of approximately 1100 data points. These data points have been sourced from the literature and are listed in Tab.1. In total, the data set has been classified into 12 input variables. These variables include water, cement, slag, fly ash, silica fume, fine aggregate, coarse aggregate, superplasticizer, fresh density, migration test age, compressive strength test age, and compressive strength measures. The purpose of utilizing these variables is to predict the non-steady-state chloride migration coefficient (D_nssm). This parameter is one of the significant durability properties of concrete, the estimation of which was conducted in this study based on a substantial size of data set, ensuring the reliability of estimation. In this respect, a reduced set of 10 input data features was used in the present model to create a realistic scenario. The features included water, cement, slag, fly ash, silica fume, fine aggregate, coarse aggregate, superplasticizer, fresh density, and compressive strength test age. This approach allowed for an accurate prediction of the compressive strength based on mixed ingredients before the estimation of D_nssm. However, it is essential to be mindful of the duplication or redundancy of data during the training of ML models. Since the same data points are repeatedly used in the investigated set of data, the model will overfit these data points and may perform poorly during testing.

2.2 Outliner removal

The data set’s characteristics are a crucial component of any neural network, significantly impacting its performance and accuracy. Since experimental investigations are typically affected by external disturbing factors, unreasonable data points almost always exist in the output. These inhomogeneities are handled by different approaches, such as the customary data deletion, interpolation, or replacement of estimated data points to represent the missing ones. Outliers refer to those points that diverge meaningfully from the rest of the observation data set. They arise due to measurement variability and misconceptions in data point filling. The most common causes of outliers on a data set include: 1) human errors, including those produced while collecting, recording, or entering data; 2) instrumental errors prompted by faulty measurement setup; 3) errors occurring in data planning or instrument execution; 4) dummy outliers; 5) data set manipulation or accidental mutation; 6) sampling error that arises from mixing data with wrong origins; and 7) natural outliers. In this respect, outliers can be of two types uni-variate and multivariate. Uni-variate outliers are found when the distribution of a single variable is sought. Also, multi-variate outliers are those in an n-dimensional space. The majority of ML algorithms fall short in the existence of outliers. Therefore, it is necessary to identify and eliminate outliers because they can affect regression assumptions. Thus, outliers can be treated in different ways, such as trimming, capping, discretization, or by treating them as missing values. One of the simplest methods for detecting outliers is using box plots. A box plot is a graphical display describing the data distributions. Box plots use the median and the lower and upper quartiles. In this study, the identification and management of outliers were accomplished by utilizing a box plot. Subsequently, an ANN with input, output, and several hidden layers was employed to compute and replace the missing data values present in four distinct categories, investigated earlier by Ref. [75]. Outliers seriously disturb the standard deviation and mean of the data set, thus giving statistically erroneous results. Fig.1 illustrates the configuration of the implemented ANN structure. Tab.2 also displays the statistical data, including the count, minimum, mean, maximum, and standard deviation of the input variables after removing outliners.

2.3 Data normalization

Normalizing data are an essential preprocessing step in neural networks, involving scaling and transforming input data to have a consistent range and distribution. The normalization process brings several benefits to the performance and convergence of neural networks. In addition, it mitigates the influence of scale, enhances convergence, enables efficient optimization, handles outliers, and promotes generalization and robustness. In this regard, neural networks are sensitive to the scale of input features. Some weights and biases may dominate the learning process when features have significantly different scales. Consequently, the network may converge slowly or get trapped in suboptimal solutions. Normalizing the data ensures that all features contribute equally by bringing them to a similar scale. The proposed modeling approach benefits the utilization of a min−max scaling function to eliminate the dominance of any particular input feature on the target prediction output. As a result, the rate of convergence is enhanced drastically. Accordingly, this technique facilitates the transformation of features into a predetermined range in terms of minimum and maximum values.

2.4 Feature encoding

This research used data encoding to prepare the data set for classification purposes. Subsequently, the resulting values of D_nssm were categorized into five groups low, moderate, high, very high, and extremely high, according to the NT BUILD 492 [71] recommendations, as reported in Tab.3. Afterwards, each category was labeled from 0 to 4, respectively. Furthermore, a classification task was carried out in this study to predict the compressive strength interval. Tab.4 illustrates the classification of data points into three distinct groups low, moderate, and high. The corresponding strengths were labeled as 0 to 2, respectively.

2.5 Data set splitting

Splitting data into train, validation, and test sets is a crucial step in ML model development. It helps to prevent over-fitting, evaluate model performance, and ensure that the model generalizes well to new unseen data. Accordingly, the training set is the sample of data used to fit the model, i.e., the actual subset of the data set used to train the model by estimating the weights and biases in the case of a neural network. The model observes and learns from this data and optimize its parameters. Also, the appropriate model or the degree of the polynomial is chosen by minimizing the error on the validation set. Furthermore, test set is the sample of data used to provide an unbiased evaluation of the final model fit on the training data set. It is applied once the model is thoroughly trained with training and validation sets. Therefore, test sets are employed to replicate the type of situation to be encountered once the model is deployed for real-time use. Before the initiation of the prediction procedure in this study, the data set was split into two subsets of train and test, with 20% of the data allocated for testing and the remaining 80% designated for training purposes. Thus, 20% of the available data was assigned for model accuracy and performance evaluation. This percentage was utilized across all four models, that is, the prediction of D_nssm and compressive strength, each with regression and classification tasks. It is worth noticing that, because of the splitting of data into training and testing sets, it is recommended to apply the min–max scaling function only to the training set. This is to avoid any potential data leakage from the test data set.

2.6 Data visualization

Data visualization is useful for data cleaning, exploring data, detecting outliers and unusual groups, identifying trends and clusters, etc. Some instances of data visualization plots in spotting outliers involve box and whisker plots, scatter plots, histograms, distribution plots, and QQ plots. Heatmaps offer a graphical depiction of the possible relationships among variables, thus facilitating the detection of trends and interrelationships. In this regard, a heatmap exhibits the magnitude and orientation of the correlations by using a system of color-coding to represent different values. When interpreting a heatmap, a correlation value close to 1.0 indicates that a particular feature is highly correlated with the target variable. Such a condition exhibits the data leakage issue, in which information from the features inadvertently leaks into the target label. Therefore, it is worthwhile to focus on identifying the predictor of the target. Fig.2 demonstrates the heatmap of the input variables in estimating the compressive strength values in this study. Data mining is essential for extracting valuable insights from extensive databases, allowing for the discovery of hidden patterns and relationships. Data mining techniques are crucial in this study for preprocessing the gathered data set and determining key variables to forecast concrete compressive strength and D_nssm. ANNs are utilized to estimate values that are missing in the data set. Ensuring the completeness of the data set and boosting the accuracy of predictive models are particularly crucial. ANNs can be employed to estimate missing data points.

Moreover, feature engineering encompasses the process of modifying existing features to enhance the efficacy of prediction models. This study employs feature engineering techniques to extract pertinent information from raw data and improve the prediction capability of the model.

Statistically, some variables in the data set may have better correlations with the goal variables, compressive strength, and D_nssm. Both the heatmap and joint plot were employed. Ref. [75] explored the importance of features and the use of SHapley Additive exPlanations to determine the effect of each feature on the target variable.

Jointplots are a combination of scatter plots and histograms that help display the joint distribution of two variables and their individual distributions. This kind of illustration helps analyze the correlation between variables and identify possible patterns or groupings. Using heatmaps and jointplots together with distplots, we can gain a comprehensive understanding of the data set, which can reveal significant insights and allow predictions about compressive strength based on mix proportions. The representation of jointplot of D_nssm versus different input variables is shown in Fig.3.

2.7 Deep learning

Utilizing neurons as the functional units within the human body can be a promising approach to predicting output values based on their input counterparts. Therefore, weight, bias, and activation functions are essential components of this approach. In this investigation, several activation functions were examined, and ultimately, the rectified linear unit was chosen due to its favorable impact on performance. Hence, a mutually exclusive data set was employed, wherein each data point was associated with a singular output. As such, the Softmax function was adopted in the last layer of the ANN within the classification task. In the context of neural networks, a primary neuron operates by performing a weighted multiplication of each input value, followed by adding a bias term. The resulting output is then subject to an activation function, typically situated at the final layer of the network, as expressed in Eqs. (1) and (2):

(1)\mathit{\text{z}}=wx+b,

(2)C=\frac{1}{2n}{\left(\left(\sum _x^{}y(x)-a^L(x)\right)\right)}^2,

where x, w, b, y, L, a, and C represent input value, weight, bias, true value, last layer count, last layer value, and quadratic cost function, respectively. Equation (2) bears a resemblance to the mean square error commonly applied in ML. When this criteria square values, all errors get positive measures, and thus, more significant errors are penalized. As a consequence, this equation has the potential to enhance outcomes as well as efficiency. According to Eq. (2), the variable C depends on input values, bias, weight, and true values. Therefore, because a network has only one neuron, using the derivative can facilitate determining the optimal solution. However, in the case of complicated networks and models, the application of gradient descent should be a suitable alternative. To determine the optimal learning rate in this context, utilizing the adaptive gradient descent optimizer may be an appropriate approach. Employing this optimization technique starts with larger steps and gradually reduces them to determine the point at which the slope approaches zero, representing the best possible outcome. Furthermore, using the cross entropy as the loss function for classification is promising. In this study, Tensorflow was employed as an end-to-end open-source library for the DL method. The efficacy of this approach relies on carefully selecting neurons and hidden layer quantities, as these factors can directly impact the outcomes. Due to the extensive data set employed in this investigation, batches were defined to introduce data in all models. It is to be noted that the numbers of hidden layers in regression and classification tasks to predict compressive strength were 12 and 8, respectively. The corresponding values were 9 and 8 in the estimation of D_nssm. Also, the numbers of neurons used in hidden layers were 24 and 22 in regression and classification tasks to predict compressive strength. Likewise, the numbers of neurons were 25 and 18 for estimating D_nssm. Notably, a decrease in the batch size results in an increase in the training time while simultaneously reducing the potential for overfitting. Therefore, batch size in regression and classification tasks in predicting the compressive strength were 16 and 8, respectively. The equivalent values were 32 and 8 in the estimation of the D_nssm.

The study focused specifically on the crucial element of hyperparameter tuning and selection in the ANN framework. The arrangement of hidden layers and the number of neurons within them were carefully fine-tuned to improve the model’s performance. In addition, dropout regularization was implemented to address overfitting and enhance the network’s ability to generalize and to address concerns regarding overfitting during model evaluation. This comprehensive approach to hyperparameter optimization guarantees that the ANN attains its utmost capability in capturing the fundamental patterns present in the data, while also preserving resilience against noise and unpredictability. The goal is to achieve a compromise between model complexity and generalizability by carefully examining various configurations and regularization procedures. This will enable a fair evaluation and comparison across diverse experimental conditions. In general, carefully considering hyperparameters not only improves the accuracy of the ANN’s predictions but also promotes a better understanding of its behavior and effectiveness in real-world scenarios.

2.8 Evaluation metrics

The cost function can exhibit the degree of deviation between the anticipated value and the actual one. This metric can be utilized to monitor the model’s efficacy throughout the training process. It gradually decreases within every training epoch until a state of convergence is reached. In this study, the gradient descent was chosen as suitable for minimising a given cost function. Furthermore, a confusion matrix was generated for classification, which can subsequently be utilized to derive additional evaluation metrics. To assess the performance of a classification model, confusion matrices with size N × N are applied, in which N represents the target class number. These matrices compare the observed values with the predictions produced by the ML model. Accordingly, a confusion matrix summarizes the number of correct and incorrect predictions made by a classifier. It is used to measure and evaluate a classification model’s performance by calculating performance metrics such as accuracy, precision, recall, and F1 score. In this regard, metrics including precision, recall, and F1 score can effectively evaluate the performance of a model by computing the ratio of negative and positive predictions, as expressed in Eqs. (3)–(5):

(3)\text{precision}=\frac{TP}{TP+FP},

(4)\text{recall}=\frac{TP}{TP+FN},

(5)F1\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\frac{2\times \text{precision}\times \text{recall}}{\text{precision}+\text{recall}},

where TP denotes true positive values, while both observed and predicted target values are positive; FP stands for false positives when the actual is negative, but the prediction is positive; and FN is false negatives, when the actual is positive, but the prediction is negative. It is to be noted that a good model is the one with high TP and TN rates, as well as low FP and FN rates. Moreover, precision is a correctness criterion attained in true predictions. It is important to note that the number of positive predictions is relatively high compared to the total number of positive predictions. Precision is a metric that measures the number of correctly identified positive target values divided by the overall count of predicted positive values. It is important that precision is high, ideally reaching 1. In DL, precision is a valuable metric when false positives are more important than false negatives. Recall is another metric that measures correctly predicted actual observations, or in other words, the number of positive class observations that are correctly identified as positive. Recall is a valid evaluator when the objective is to capture as many positive values as possible.

In addition, Recall is defined as the ratio of the total number of correctly classified positive classes divided by the total number of positive classes. The value of recall has to be high, where it is preferably 1. It is an advantageous metric when FNs are higher than FPs. Similarly, the F1 score, which is a harmonic mean of recall and precision, lies between 0 and 1. It should be noted that this type of mean is implemented because, contrary to simple averages, it is not susceptible to extremely large target values. In fact, the F1 score keeps a balance between recall and precision regarding the classifier. Given that either the precision or recall is low, the F1 will also be low. This metric reflects both trends in a single value. It is thus a harmonic mean of recall and precision. Likewise, the F1 score should be high, where it is ideally 1. Following each training epoch on the training data set, it is possible to compute the loss of test data and monitor the model’s efficiency on both training and testing data sets. Nevertheless, the test data does not influence the weights and bias of the ANN. Sketching a plot within this range can facilitate the determination of whether the model is subject to overfitting or not. Thus, a significant disparity in the loss between test and train sets at the tail of the plot could indicate the presence of overfitting. The findings of this investigation show no evidence of overfitting in any of the models under consideration. Moreover, the overall flowchart of the modeling process from data selection to model deployment is demonstrated in Fig.4.

3 Results and discussion

3.1 Compressive strength results

The comparison of training and validation loss in the concrete compressive strength versus different epochs obtained from the regression task in the modeling procedure is carried out in Fig.5. The train and test losses show remarkable agreement, with the two curves overlapping at different modeling epochs. This observation suggests that the proposed prediction model is not overfitting and can be utilized comfortably. Furthermore, Fig.6 compares the proposed model estimations with the actual compressive strength results. As it is seen, there is an admissible accord between the actual values and those of the recommended model regarding compressive strength. The mean absolute error (MAE) in the regression process was 3.843 in predicting the compressive strength, with a mean value of 46.75.

As for the classification task, the evaluation metrics for the estimation of the compressive strength are demonstrated in Fig.7. Accordingly, the model’s effectiveness throughout the training process can be examined using the measures for different feature encoding. Fig.7 sketches the confusion matrix results along with precision, recall, and F1 score estimates. As can be seen, these values lie in acceptable ranges. The accuracy of the modeling through the classification task was 0.88.

3.2 Chloride migration coefficient results

In like manner, the results of the comparison between test and training losses in D_nssm in concrete versus different epochs obtained from the regression task in the modeling procedure are reported in Fig.8. Additionally, the proposed model estimations with the actual D_nssm results are shown in Fig.9. As such, an acceptable agreement between the actual values and those of the suggested model is seen in terms of migration coefficient. The MAE in the regression process was 1.123 in predicting the D_nssm with a mean value of 7.24.

Similar to the compressive strength prediction in the classification task, the evaluation metrics for the estimation of the D_nssm are summarized in Fig.10. Accordingly, the precision, recall, and F1 score estimates for the five categories of feature encoding are demonstrated in Fig.10. It is seen that these values are in suitable ranges. The accuracy of the modeling through the classification task was 0.86.

Tab.5 displays a comparative comparison of different ML methods, such as Elastic Net, Lasso, Linear Regression, Ridge Regression, Random Forest, K-Nearest Neighbors, Decision Tree, SVM, and XGBoost. The performance of these algorithms in predicting D_nssm was assessed by comparing them to the suggested ANN model. The findings suggest that although all algorithms show different levels of predictive abilities, the ANN model outperforms standard ML methods in terms of accuracy for the D_nssm prediction test. This discovery highlights the efficacy of utilizing ANN structures to capture intricate connections within the data set, resulting in more precise forecasts.

4 Model deployment

The authors developed a user-friendly web-based application so that anyone with the given input parameters could easily estimate the compressive and D_nssm of a specific type of concrete. A representative schematic of the application is shown in Fig.11. Thus, the integrity of the results delivered by the application solely relies on the ANNs, which is, per se, dependent on the data set containing approximately 1100 data points collected from the top-notch scholarly literature.

5 Conclusions

The mechanical and durability characteristics of concrete involving compressive strength and D_nssm were estimated using DL and ANN methods considering the two regression and classification tasks. A comprehensive experimental database of over 1000 points gathered from previous investigations was used. The prediction model relied on 12 input parameters based on mix design ingredients and concrete age variables. The model estimated compressive strength and D_nssm. Based on the results obtained, the following observations can be made.

1) The missing data gap in the literature was filled through the ANN approach programmed in the modeling process.

2) The comprehensiveness and diversity of the model input data set, including different mix ingredients from water, cement, and aggregates to silica fume, slag, and fly ash, contributed to the increased robustness and reliability of the results.

3) An inclusive pre-processing procedure was accomplished to detect outliers and encode data features.

4) The selection of MAE for the regression task and confusion matrix for the classification task as evaluation metrics delivered precise results.

5) Compressive strength and D_nssm exhibited an accuracy of over 0.85 in both regression and classification tasks.

6) A user-friendly web application was developed utilizing the Python programming language, which gives a greater ability to integrate with the user’s device. This way, any user can conveniently enter the input parameters and obtain the prediction results for both the D_nssm and the compressive strength of a given concrete.

References

[1] Feng D C, Liu Z T, Wang X D, Chen Y, Chang J Q, Wei D F, Jiang Z M. Machine learning-based compressive strength prediction for concrete: An adaptive boosting approach. Construction & Building Materials, 2020, 230: 117000

[2] Alqahtani F K, Zafar I. Characterization of processed lightweight aggregate and its effect on physical properties of concrete. Construction & Building Materials, 2020, 230: 116992

[3] Mansoori A, Mohtasham Moein M, Mohseni E. Effect of micro silica on fiber-reinforced self-compacting composites containing ceramic waste. Journal of Composite Materials, 2020, 55(1): 95–107

[4] Tahmouresi B, Nemati P, Asadi M A, Saradar A, Moein M M. Mechanical strength and microstructure of engineered cementitious composites: A new configuration for direct tensile strength, experimental and numerical analysis. Construction & Building Materials, 2021, 269: 121361

[5] Zain M F M, Abd S M. Multiple regression model for compressive strength prediction of high performance concrete. Journal of Applied Science, 2009, 9(1): 155–160

[6] Bharatkumar B H, Narayanan R, Raghuprasad B K, Ramachandramurthy D S. Mix proportioning of high performance concrete. Cement and Concrete Composites, 2001, 23: 71–80

[7] Bhanja S, Sengupta B. Investigations on the compressive strength of silica fume concrete using statistical methods. Cement and Concrete Research, 2002, 32(9): 1391–1394

[8] Oliver J, Huespe A E, Samaniego E, Chaves E W V. Continuum approach to the numerical simulation of material failure in concrete. International Journal for Numerical and Analytical Methods in Geomechanics, 2004, 28(7–8): 609–632

[9] Feng D C, Xie S C, Deng W N, Ding Z D. Probabilistic failure analysis of reinforced concrete beam–column sub-assemblage under column removal scenario. Engineering Failure Analysis, 2019, 100: 381–392

[10] Feng D C, Wang Z, Wu G. Progressive collapse performance analysis of precast reinforced concrete structures. Structural Design of Tall and Special Buildings, 2019, 28(5): e1588

[11] Haenlein M, Kaplan A. A brief history of artificial intelligence: On the past, present, and future of artificial intelligence. California Management Review, 2019, 61(4): 5–14

[12] Moein M M, Saradar A, Rahmati K, Mousavinejad S H G, Bristow J, Aramali V, Karakouzian M. Predictive models for concrete properties using machine learning and deep learning approaches: A review. Journal of Building Engineering, 2023, 63: 105444

[13] Liu Q, Sun P, Fu X, Zhang J, Yang H, Gao H, Li Y. Comparative analysis of BP neural network and RBF neural network in seismic performance evaluation of pier columns. Mechanical Systems and Signal Processing, 2020, 141: 106707

[14] Chou J S, Ngo N T, Pham A D. Shear strength prediction in reinforced concrete deep beams using nature-inspired metaheuristic support vector regression. Journal of Computing in Civil Engineering, 2016, 30(1): 08215001

[15] Behnood A, Verian K P, Gharehveran M M. Evaluation of the splitting tensile strength in plain and steel fiber-reinforced concrete based on the compressive strength. Construction & Building Materials, 2015, 98: 519–529

[16] Han B, Wu Y, Liu L. Prediction and uncertainty quantification of compressive strength of high-strength concrete using optimized machine learning algorithms. Structural Concrete, 2022, 23(6): 3772–3785

[17] Abdeljaber O, Avci O, Inman D J. Active vibration control of flexible cantilever plates using piezoelectric materials and artificial neural networks. Journal of Sound and Vibration, 2016, 363: 33–53

[18] Jiang X, Mahadevan S, Adeli H. Bayesian wavelet packet denoising for structural system identification. Structural Control & Health Monitoring, 2007, 14(2): 333–356

[19] HosseinzadehMDehestaniMHosseinzadehA. Exploring elastic properties of fly ash recycled aggregate concrete: Insights from multiscale modeling and machine learning. Structures, 59: 105720

[20] Huang J C, Ko K M, Shu M H, Hsu B M. Application and comparison of several machine learning algorithms and their integration models in regression problems. Neural Computing & Applications, 2020, 32(10): 5461–5469

[21] Hosseinzadeh M, Dehestani M, Hosseinzadeh A. Prediction of mechanical properties of recycled aggregate fly ash concrete employing machine learning algorithms. Journal of Building Engineering, 2023, 76: 107006

[22] MüllerA CGuidoS. Introduction to Machine Learning with Python: A Guide for Data Scientists. Sebastopol: O’Reilly Media, 2016

[23] Rosental C. Certifying knowledge: The sociology of a logical theorem in artificial intelligence. American Sociological Review, 2003, 68(4): 623–644

[24] Langley P. The changing science of machine learning. Machine Learning, 2011, 82(3): 275–279

[25] Dietterich T G. Ensemble methods in machine learning. Lecture Notes in Computer Science, 2000, 1857: 1–15

[26] Avci O, Abdeljaber O, Kiranyaz S, Hussein M, Gabbouj M, Inman D J. A review of vibration-based damage detection in civil structures: from traditional methods to Machine Learning and Deep Learning applications. Mechanical Systems and Signal Processing, 2021, 147: 107077

[27] Zhang C, Zhang X. Multi-target domain-based hierarchical dynamic instance segmentation method for steel defects detection. Neural Computing & Applications, 2023, 35(10): 7389–7406

[28] Schmidhuber J. Deep Learning in neural networks: An overview. Neural Networks, 2015, 61: 85–117

[29] Chu X H, Wang J Y, Li S X, Chai Y J, Guo Y Q. Empirical study on meta-feature characterization for multi-objective optimization problems. Neural Computing & Applications, 2022, 34(19): 16255–16273

[30] Deng F, He Y, Zhou S, Yu Y, Cheng H, Wu X. Compressive strength prediction of recycled concrete based on deep learning. Construction & Building Materials, 2018, 175: 562–569

[31] Kaloop M R, Gabr A R, El-Badawy S M, Arisha A, Shwally S, Hu J W. Predicting resilient modulus of recycled concrete and clay masonry blends for pavement applications using soft computing techniques. Frontiers of Structural and Civil Engineering, 2019, 13(6): 1379–1392

[32] Chou J S, Chiu C K, Farfoura M, Al-Taharwa I. Optimizing the prediction accuracy of concrete compressive strength based on a comparison of data-mining techniques. Journal of Computing in Civil Engineering, 2011, 25(3): 242–253

[33] Gholampour A, Mansouri I, Kisi O, Ozbakkaloglu T. Evaluation of mechanical properties of concretes containing coarse recycled concrete aggregates using multivariate adaptive regression splines (MARS), M5 model tree (M5Tree), and least squares support vector regression (LSSVR) models. Neural Computing & Applications, 2020, 32(1): 295–308

[34] Deepa C, SathiyaKumari K, Pream S V. Prediction of the compressive strength of high performance concrete mix using tree based modeling. International Journal of Computer Applications, 2010, 6: 18–24

[35] Behnood A, Behnood V, Modiri Gharehveran M, Alyamac K E. Prediction of the compressive strength of normal and high-performance concretes using M5P model tree algorithm. Construction & Building Materials, 2017, 142: 199–207

[36] Ayaz Y, Kocamaz A F, Karakoç M B. Modeling of compressive strength and UPV of high-volume mineral-admixtured concrete using rule-based M5 rule and tree model M5P classifiers. Construction & Building Materials, 2015, 94: 235–240

[37] Dantas A T A, Batista Leite M, de Jesus Nagahama K. Prediction of compressive strength of concrete containing construction and demolition waste using artificial neural networks. Construction & Building Materials, 2013, 38: 717–722

[38] Duan Z H, Kou S C, Poon C S. Using artificial neural networks for predicting the elastic modulus of recycled aggregate concrete. Construction & Building Materials, 2013, 44: 524–532

[39] Naderpour H, Poursaeidi O, Ahmadi M. Shear resistance prediction of concrete beams reinforced by FRP bars using artificial neural networks. Measurement, 2018, 126: 299–308

[40] Topçu I B, Saridemir M. Prediction of mechanical properties of recycled aggregate concretes containing silica fume using artificial neural networks and fuzzy logic. Computational Materials Science, 2008, 42(1): 74–82

[41] Nikoo M, TorabianMoghadam F, Sadowski L. Prediction of concrete compressive strength by evolutionary artificial neural networks. Advances in Materials Science and Engineering, 2015, 2015: 849126

[42] Chithra S, Kumar S R R S, Chinnaraju K, Alfin Ashmita F. A comparative study on the compressive strength prediction models for High Performance Concrete containing nano silica and copper slag using regression analysis and Artificial Neural Networks. Construction & Building Materials, 2016, 114: 528–535

[43] Ranjbar I, Toufigh V, Boroushaki M. A combination of deep learning and genetic algorithm for predicting the compressive strength of high-performance concrete. Structural Concrete, 2022, 23(4): 2405–2418

[44] Kina C, Turk K, Tanyildizi H. Estimation of strengths of hybrid FR-SCC by using deep-learning and support vector regression models. Structural Concrete, 2022, 23(5): 3313–3330

[45] el Asri Y, Benaicha M, Zaher M, Hafidi Alaoui A. Prediction of the compressive strength of self-compacting concrete using artificial neural networks based on rheological parameters. Structural Concrete, 2022, 23(6): 3864–3876

[46] GuoHZhuangXRabczukT. 2021. A deep collocation method for the bending analysis of Kirchhoff plate. Computers Materials & Continua, 59(2): 433–456

[47] Guo H, Zhuang X, Chen P, Alajlan N, Rabczuk T. Stochastic deep collocation method based on neural architecture search and transfer learning for heterogeneous porous media. Engineering with Computers, 2022, 38(6): 5173–5198

[48] Guo H, Zhuang X, Alajlan N, Rabczuk T. Physics-informed deep learning for melting heat transfer analysis with model-based transfer learning. Computers & Mathematics with Applications, 2023, 143: 303–317

[49] Samaniego E, Anitescu C, Goswami S, Nguyen-Thanh V M, Guo H, Hamdia K, Zhuang X, Rabczuk T. An energy approach to the solution of partial differential equations in computational mechanics via machine learning: Concepts, implementation and applications. Computer Methods in Applied Mechanics and Engineering, 2020, 362: 112790

[50] Zhuang X, Guo H, Alajlan N, Zhu H, Rabczuk T. Deep autoencoder based energy method for the bending, vibration, and buckling analysis of Kirchhoff plates with transfer learning. European Journal of Mechanics. A, Solids, 2021, 87: 104225

[51] Guo H, Zhuang X, Fu X, Zhu Y, Rabczuk T. Physics-informed deep learning for three-dimensional transient heat transfer analysis of functionally graded materials. Computational Mechanics, 2023, 72(3): 513–524

[52] Hoang N D, Chen C T, Liao K W. Prediction of chloride diffusion in cement mortar using multi-gene genetic programming and multivariate adaptive regression splines. Measurement, 2017, 112: 141–149

[53] Gao W, Chen X, Chen D. Genetic programming approach for predicting service life of tunnel structures subject to chloride-induced corrosion. Journal of Advanced Research, 2019, 20: 141–152

[54] Ahmad A, Farooq F, Ostrowski K A, Śliwa-Wieczorek K, Czarnecki S. Application of novel machine learning techniques for predicting the surface chloride concentration in concrete containing waste material. Materials, 2021, 14(9): 2297

[55] Liu K H, Zheng J K, Pacheco-Torgal F, Zhao X Y. Innovative modeling framework of chloride resistance of recycled aggregate concrete using ensemble-machine-learning methods. Construction & Building Materials, 2022, 337: 127613

[56] ASTM-C1202. Standard Test Method for Electrical Indication of Concrete’s Ability to Resist Chloride Ion Penetration. Philadelphia: American Society Testing and Materials, 2012

[57] Jin L, Dong T, Fan T, Duan J, Yu H, Jiao P, Zhang W. Prediction of the chloride diffusivity of recycled aggregate concrete using artificial neural network. Materials Today Communications, 2022, 32: 104137

[58] Alabdullah A A, Iqbal M, Zahid M, Khan K, Amin M N, Jalal F E. Prediction of rapid chloride penetration resistance of metakaolin based high strength concrete using light GBM and XGBoost models by incorporating SHAP analysis. Construction & Building Materials, 2022, 345: 128296

[59] Amin M N, Raheel M, Iqbal M, Khan K, Qadir M G, Jalal F E, Alabdullah A A, Ajwad A, Al-Faiad M A, Abu-Arab A M. Prediction of rapid chloride penetration resistance to assess the influence of affecting variables on metakaolin-based concrete using gene expression programming. Materials, 2022, 15(19): 6959

[60] YaoLRenLGongG. Evaluation of chloride diffusion in concrete using PSO-BP and BP neural network. In: Proceedings of IOP Conference Series (Earth and Environmental Science). Zhuhai: IOP Publishing, 2021, 687(1): 012037

[61] Cai R, Han T, Liao W, Huang J, Li D, Kumar A, Ma H. Prediction of surface chloride concentration of marine concrete using ensemble machine learning. Cement and Concrete Research, 2020, 136: 106164

[62] Tran A T, Le T H, Nguyen H M. Forecast of surface chloride concentration of concrete utilizing ensemble decision tree boosted. Journal of Science & Transport Technology, 2022, 2(1): 44–56

[63] Tran V Q. Machine learning approach for investigating chloride diffusion coefficient of concrete containing supplementary cementitious materials. Construction & Building Materials, 2022, 328: 127103

[64] Guo Z, Guo R, Lin S. Multi-factor fuzzy prediction model of concrete surface chloride concentration with trained samples expanded by random forest algorithm. Marine Structures, 2022, 86: 103311

[65] Guo Z, Guo R, Yao G. Multi-factor model to predict surface chloride concentration of concrete based on fuzzy logic system. Case Studies in Construction Materials, 2022, 17: e01305

[66] Taffese W Z, Espinosa-Leal L. A machine learning method for predicting the chloride migration coefficient of concrete. Construction & Building Materials, 2022, 348: 128566

[67] Taffese W Z, Espinosa-Leal L. Prediction of chloride resistance level of concrete using machine learning for durability and service life assessment of building structures. Journal of Building Engineering, 2022, 60: 105146

[68] Tran V Q, Giap V L, Vu D P, George R C, Ho L S. Application of machine learning technique for predicting and evaluating chloride ingress in concrete. Frontiers of Structural and Civil Engineering, 2022, 16(9): 1153–1169

[69] Van RossumG. Python Programming Language. In: USENIX Annual Technical Conference. Santa Clara: CA, 2007, 1–36

[70] GilatA. MATLAB: An introduction with Applications. Hoboken: John Wiley & Sons, 2004

[71] Nordtest. NT Build 492. Concrete, mortar and cement-based repair materials: Chloride migration coefficient from non-steady-state migration experiments. Oslo: NORDTEST, 1999

[72] Costa A, Appleton J. Chloride penetration into concrete in marine environment—Part I: Main parameters affecting chloride penetration. Materials and Structures, 1999, 32(4): 252–259

[73] Thomas M D, Bamforth P B. Modelling chloride diffusion in concrete—Effect of fly ash and slag. Cement and Concrete Research, 1999, 29(4): 487–495

[74] Hou H B, Zhang G Z. Assessment on chloride contaminated resistance of concrete with non-steady-state migration method. Journal of Wuhan University of Technology—Materials Science Edition, 2004, 19(4): 6–8

[75] Hosseinzadeh M, Mousavi S S, Hosseinzadeh A, Dehestani M. An efficient machine learning approach for predicting concrete chloride resistance using a comprehensive dataset. Scientific Reports, 2023, 13(1): 15024

RIGHTS & PERMISSIONS

Higher Education Press