When analyzing chloride migration in concrete, one can enhance its structural durability and reduce corrosion risks. Employing machine learning to forecast chloride migration coefficients emerges as a cost-effective alternative to labor-intensive experimentation.
Study: Enhancing Concrete Durability: Insights from Advanced Machine Learning. Image credit: vajaraphol/Shutterstock
In a recent paper published in the journal Scientific Reports, researchers addressed a significant research gap by utilizing a comprehensive dataset and advanced AI techniques to create a reliable model for predicting chloride resistance in various concrete compositions.
Background
Concrete composition significantly impacts its mechanical and durability characteristics. Researchers face challenges in understanding how different compositions affect durability, especially with the emergence of new concrete types. Chloride penetration poses a significant threat to reinforced concrete structures, and this threat leads to the corrosion of steel reinforcement or steel fibers. This corrosion causes economic losses and environmental impacts. Several parameters influence the chloride resistance of reinforced concrete, with concrete composition playing a major role. Previous studies emphasize that concrete composition directly affects freshness, mechanical, and durability characteristics.
Concrete composition encompasses water-to-binder ratio, water content, cement type, cement content, aggregate type, aggregate content, fillers, mineral additives, nanomaterials, and chemical admixtures. Achieving a concrete mixture with low permeability is crucial to preventing chloride ion penetration and internal damage. Supplementary cementitious materials (SCMs), fillers, and nanomaterials can be used to create a chloride-resistant mixture. However, as new concrete types have been introduced, standards recommend experimental tests to evaluate their performance against chloride attacks. These tests, such as salt ponding, bulk diffusion, and rapid chloride permeability tests, are time-consuming and require significant resources.
Experimental tests consider various parameters to determine chloride resistance, with the chloride diffusion coefficient playing a critical role. Different factors affect this coefficient, including concrete composition, chloride exposure time, curing conditions, and exposure location. Multiple methods exist to determine the chloride diffusion coefficient, ranging from traditional bulk diffusion tests to accelerated methods. Despite their accuracy, these tests are often impractical for individual concrete mixtures due to time and cost constraints.
Machine learning models and analysis
Artificial intelligence (AI) has gained popularity for predicting cementitious composites' characteristics, including chloride resistance. The current study comprehensively evaluates the use of regression and classification algorithms in machine learning to predict non-steady-state migration coefficients (DNSSM) without limitations on concrete type or missing input data. This is archived in three steps: data cleaning, data visualization, and model training.
Data Cleaning: The study utilized experimental datasets obtained from 24 research papers, comprising 1073 datasets. Some datasets had missing input features, including fresh density, compressive strength test age, and concrete compressive strength. To address this, a feed-forward backpropagation network and the Levenberg-Marquardt training function were applied to predict missing data, with different hidden layers for each dataset group based on the missing parameter type.
The accuracy of the artificial neural network (ANN) prediction confirmed its high accuracy. This approach successfully resolved the missing dataset issue, allowing for the utilization of complete datasets in the machine learning process. The study also involves data preprocessing, which includes removing outliers and ensuring the dataset's quality.
Data Visualization: Data visualization techniques, such as distplots, heatmaps, joint plots, and kernel density estimation (KDE), is used to explore the relationships between variables in the dataset. The pair-plot offers a comprehensive view of variable relationships within the dataset, with scatter plots and KDE functions revealing associations.
Notably, it suggests that higher water content may lead to concrete samples with higher fresh density and better chloride resistance. It also highlights the significance of fine aggregates and fresh density in achieving better chloride resistance. Moreover, it underscores the interplay of various factors, such as the interaction between cement, SCMs, and coarse aggregate on chloride resistance.
The heatmap plot illustrates Pearson correlation coefficients, revealing linear associations between continuous variables. It shows that water content positively correlates with cement but negatively with slag content. Coarse aggregate content is negatively correlated with cement and positively with fresh density, which is crucial for predicting DNSSM.
The KDE joint-plot provides probability density estimates for independent variables. It indicates that chloride-resistant concrete mixtures often have specific ranges for cement, water content, fine aggregate, coarse aggregate, and fresh density.
Model Training: Researchers employed supervised learning techniques to construct a unified DNSSM prediction model. Supervised learning encompasses two primary models, namely regression and classification. Various regression models are applied, including Simple Linear Regression, Ridge, Lasso, Elastic Net, Random Forest, K-Nearest Neighbors (KNN), Decision Tree, Support Vector Machine (SVM), and XGBoost.
Linear regression establishes linear relationships between a response variable and predictor variables. SVM creates a linear decision space for classification, even when the data are not linearly separable. KNN performs non-parametric classification based on data proximity. Logistic regression extends linear regression for classification by using likelihood thresholds. Decision trees hierarchically represent outcome spaces based on queries.
Results and Analysis: The results showcased that XGBoost and SVR, as regression models, demonstrated remarkable predictive accuracy, with R-square scores surpassing 0.90. Among the classification models, LightGBM and XGBoost exhibited the highest accuracy. Shapley additive explanations (SHAP) analysis unveiled significant influences between the variables. SHAP analysis highlights the interplay between cement-water, slag-superplasticizer, fly ash-coarse aggregate, silica fume-fine aggregate, compressive strength-fine aggregate, fresh density-coarse aggregate, and fresh density-compressive strength in shaping the DNSSM predictive model.
Conclusion
In summary, researchers developed a comprehensive AI model that can predict the chloride diffusion coefficient (DNSSM) for various concrete types. The study includes a novel dataset-cleaning technique to account for missing data and compares classification and regression algorithms. Additionally, the study seeks to identify the most influential parameters controlling the prediction accuracy of models.
Machine Learning Identifies Seismic Precursors, Advancing Earthquake Forecasting Capabilities