Interpretable Machine Learning Modeling for Surface Chloride Concentration Prediction in Marine Concrete

The chloride concentration (Cs) on the concrete surface is an indispensable metric for designing resilience and estimating the lifespan of concrete structures in aquatic settings. Consequently, due to Chloride action, many reinforced concrete constructions cannot reach their intended or planned life span and experience early degradation. This study utilizes three independent machine learning techniques, Extreme Gradient Boosting (XGBoost), Multi Expression Programming (MEP), and Decision Tree (DT), to forecast the concrete’s surface chloride concentration (Cs). To achieve this objective, a comprehensive database consisting of 642 observations of Cs exposure data in the marine field, including the applicable mixture quantity of constraints, conditions of the environment, and exposure time, has been created through a thorough investigation of relevant literature. This database covers tidal, splash, and water-logged zones in numerous locations across the globe. The database trains three independent machine learning models: XGBoost, MEP, and DT. These models are then employed to compare their prediction performances with each other and with previous modeling techniques. Diverse statistical criteria evaluated the model's accuracy and suitability. During the validation process, the XGBoost, MEP, and DT models demonstrated enhanced accuracy with coefficients of correlation (R) 0.98, 0.97, 0.95 for training and 0.97 for validation. Additionally, these models' mean absolute errors (MAE) were correspondingly 0.020, 0.021, and 0.09. The SHapley Additive exPlanations (SHAP) technique has found the chloride binder content as a crucial component in determining the concentration of Cs. The SHAP method provides a thorough understanding of the links between various input variables and their effects on the concrete’s surface concentration of chloride, both on a global and local scale. The results indicate that by including more diverse datasets and considering new variables, the predicted accuracy of standard models may be improved. The XGBoost machine learning model, trained on a vast database, can effectively include 14 key characteristics that pose challenges for conventional models. It exhibits superior predicted accuracy and enhanced time efficiency compared to conventional models. However, to lessen the problem of overfitting, it is advisable to use a more extensive dataset, including synthetic or genuine experimental data.