Safety prediction of bearing capacity of rectangular concrete four-pile caps using XGBoost-PSO regression

Abstract The design of the reinforced concrete structures using machine-learning techniques corresponds to the symmetrical distribution of the results, where the application of a safety coefficient includes sliding for all values (lower and upper bands). This paper presents a practical and comprehensive implementation of machine learning based on the Extreme gradient boosting model for the safety prediction of the bearing capacity of four concrete pile caps. To target the upper band only for the safety factor, the suggested model seeks to accurately under-predict the strength capacity to accommodate safety issues in pile caps design using a customized asymmetric loss function optimized by the particle swarm optimization algorithm. For the development of the Extreme gradient boosting regression model, a widely used dataset consisting of 107 four-pile cap tests is used for training and testing the model. Using five-fold cross-validation for model performance evaluation and the grid search method to tune the hyper-parameters, the developed Extreme gradient boosting regressor outperforms several recently reported mechanical models and different design codes (Concrete Reinforcing Steel Institute Design 2002; American Concrete Institute 318-05, and Canadian Standards Association A23.3). The results exhibited very acceptable performance metrics (coefficient of variation = 4.2 and a mean ratio of the bearing capacity values (test sample/model) equal to 1.09). The obtained results prove the ability of the proposed model to accurately and safely predict the bearing capacity, which allows it to be incorporated into any design software engineering for complete safety calculation.