Abstract:

To develop an accurate prediction model for concrete compressive strength and compare the prediction performance of different machine learning methods， this study trained 17 machine learning models using a dataset consisting of 1 030 samples， which included eight features such as water-to-binder ratio， age， and aggregate-to-binder ratio. Six models with higher accuracy were selected from these. Subsequently， each of these six models was combined with five hyperparameter optimization algorithms， namely Bayesian optimization （BO）， sparrow search algorithm （SSA）， simulated annealing （SA）， sine cosine algorithm （SCA）， and particle swarm optimization （PSO）， resulting in 30 optimized hybrid models. The models with superior accuracy were further selected for external validation and SHapley Additive exPlanations （SHAP） interpretability analysis. The results show that the least squares boosting model optimized by SSA （LSBoost-SSA）， the eXtreme Gradient Boosting model optimized by BO （XGBoost-BO）， and the light gradient boosting machine model optimized by SSA （LGBM-SSA） exhibit the best performance； XGBoost-BO achieves the optimal performance in external validation， capable of controlling the prediction error of concrete strength within 5 MPa； SHAP analysis reveals that water-binder ratio， curing age， and cement content are the core factors influencing concrete compressive strength.

Key words: concrete, compressive strength, machine learning, hyperparameter optimization, prediction model