Explainable machine learning framework for predicting concrete abrasion depth

This paper presents an approach for predicting concrete abrasion depth based on an advanced machine learning-based framework with explainability. The framework integrates multiple data pre-processing efforts, such as feature selection and anomaly detection, into machine learning. Different ensemble learning algorithms XGBoost, CatBoost, LightGBM, Extremely Randomized Trees, and Random Forest are compared using performance metrics, including mean absolute error, root mean squared error, coefficient of determination, mean absolute deviation, mean percentage deviation, and scatter index for regression analysis. Results show that machine learning models achieve high performance. Feature importance analysis and SHAP analysis are performed to assess the effects of factors that affect concrete abrasion depth. The time of testing-to-velocity, water-to-binder ratio, and aggregate compositions are the top three influencing factors. This study demonstrates the potential of machine learning in predicting concrete abrasion depth, providing valuable insights for the design and maintenance of hydraulic concrete structures. • Concrete abrasion depth is predicted using an explainable machine learning framework. • Ensemble learning methods were developed and comparatively evaluated. • Hyperparameter tuning and k-fold cross validation were incorporated to improve the models. • SHAP analysis was performed to evaluate the effects of various variables of the models. • Feature importance analysis was performed to identify important features.