Prediction of chloride concentration in concrete under multi-salt environment: Optimization of integrated algorithm based on MSCPO and interpretability analysis

Abstract

The accurate prediction of chloride concentration is vital for assessing reinforced concrete structure durability. However, diverse erosion media in engineering environments with varying ion concentrations present challenges for traditional prediction methods. This study conducted accelerated experiments to create a concrete chloride ion dataset in a multi-salt environment, with adjustments for abnormal data. The Crested Porcupine Optimizer (CPO) algorithm was enhanced with adaptive techniques, and the refined strategy’s effectiveness was verified through test function analysis. The Improved Mixture Self-Adaptation Crested Porcupine Optimizer (MSCPO) optimized hyperparameters for XGBoost, LightGBM, and Catboost models separately. The fitting, accuracy, and stability of each model in predicting concrete chloride concentration were quantitatively assessed. SHAP was used to explain the best-performing model, and its reliability was supported by microscopic observation results and literature. Results show that identifying and handling outliers enhance model performance. The proposed MSCPO excelled in hyperparameter search, with optimized ensemble models maintaining error within a reasonable range. XGBoost had the best performance, completing hyperparameter search in 45.52 s and achieving an R² of 92.86%. SHAP results aligned closely with experiments and supported existing literature conclusions.