A convolutional neural network-based deep learning approach for predicting surface chloride concentration of concrete in marine tidal zones

Abstract

Chloride-induced corrosion is a major threat to the durability of reinforced concrete (RC) structures. This is especially critical in marine tidal zones, where surface chloride concentration (Cs) plays a key role in predicting chloride ingress using Fick's second law. However, traditional assessment methods are time-consuming and impractical, necessitating advanced predictive models. This study developed a deep learning-based framework utilizing a convolutional neural network (CNN) trained on 284 samples with 11 critical features related to material composition and environmental conditions. The CNN's performance was benchmarked against four machine learning (ML) models: stepwise linear regression (SLR), support vector machine (SVM), Gaussian process regression (GPR), and random forest (RF). Results demonstrated CNN's superiority, achieving a coefficient of determination (R²) = 0.849 and a lower root mean square error (RMSE) = 0.18%, outperforming conventional models. Shapley additive explanation (SHAP) analysis revealed exposure time, water content, and fine aggregate as the most critical factors influencing Cs predictions. The findings highlighted the importance of material composition and environmental exposure in optimizing concrete mix designs to mitigate chloride ingress in tidal zones. This research can enhance durability assessment, proactive maintenance strategies, and service life estimation of RC structures in harsh marine environments. Furthermore, it can contribute to the sustainable development goals (SDGs) by promoting resilient infrastructure, sustainable construction practices, and improved climate adaptation strategies. By integrating deep learning in durability assessments, this study can provide a scalable, efficient solution for optimizing maintenance planning and reducing premature failures in coastal RC structures, ultimately extending their service life.

Keywords: Convolutional neural network; Deep learning; Gaussian process regression; Surface chloride concentration; Sustainable development goals; Tidal zones.

Fig. 1

Different exposure zones of marine structure.

Fig. 2

Visual diagram for study workflow.

Fig. 3

Samples of variable distribution alongside the Gaussian distribution curve.

Fig. 3

Samples of variable distribution alongside the Gaussian distribution curve.

Fig. 4

Scatter plots illustrating the relationships between each individual input variable and Cs. The R reflects the strength of the linear association and is used here for exploratory analysis to identify potentially influential variables prior to model development.

Fig. 5

Pearson correlation coefficient (PCC) matrix of the chosen 11 input parameters and surface chloride concentration (Cs).

Fig. 6

The architecture of the proposed one-dimensional convolutional neural network (1D-CNN) model.

Fig. 7

Training and validation loss curves.

Fig. 8

Correlation between the forecasted and experimental surface chloride concentration (Cs) results during both the training and testing stages. (a) Training phase. (b) Testing phase. (c) Training phase. (d) Testing phase. (e) Training phase. (f) Testing phase. (g) Training phase. (h) Testing phase. (i) Training phase. (j) Testing phase.

Fig. 8

Correlation between the forecasted and experimental surface chloride concentration (Cs) results during both the training and testing stages. (a) Training phase. (b) Testing phase. (c) Training phase. (d) Testing phase. (e) Training phase. (f) Testing phase. (g) Training phase. (h) Testing phase. (i) Training phase. (j) Testing phase.

Fig. 9

Variation of experimental and predicted results and their error range for predicting surface chloride concentration (Cs) in concrete. (a) SVM model. (b) GPR model. (c) SLR model. (d) RF model. (e) CNN model.

Fig. 9

Variation of experimental and predicted results and their error range for predicting surface chloride concentration (Cs) in concrete. (a) SVM model. (b) GPR model. (c) SLR model. (d) RF model. (e) CNN model.

Fig. 10

Density distribution plot of experimental and projected values for five machine learning (ML) models.

Fig. 11

Graphical user interface (GUI) of surface chloride concentration (Cs) prediction through the proposed machine learning (ML) models based on the database: (a) convolutional neural network (CNN), (b) support vector machine (SVM), (c) Gaussian process regression (GPR); (d) stepwise linear regression (SLR), and (e) random forest (RF).

Fig. 12

Graphical user interface (GUI) of surface chloride concentration (Cs) prediction through CNN model based on water to cement ratio: (a) 0.3; (b) 0.4; (c) 0.5; and (d) 0.6.

Fig. 13

Feature importance analysis. (a) Mean absolute SHAP values. (b) Shapley additive explanations (SHAP) global explanation.

Fig. 14

Partial dependence plots (PDPs) showing the impact of key factors on surface chloride concentration (Cs) evolution, predicted by machine learning (ML) models.