Combining machine learning methods with feature selection for predicting concrete basic creep

Conducting long-term concrete creep tests with multiple influencing variables is both time-consuming and costly. The advantage of using machine learning (ML) algorithms to predict the mechanical properties of concrete has attracted increasing attention. To address these issues, this study adopted three ML models to establish more accurate prediction models. These models were developed and verified with data from the updated Infrastructure Technology Institute of Northwestern University (NU-ITI) database. Moreover, the maximum information correlation (MIC) of big data, light gradient boosting machine (LGBM), and prediction risk-based feature selection for support vector regression (PRIFER) methods were adopted to analyze the relationships among the concrete creep influence parameters and obtain the feature importance. Feature selection methods were employed to improve the efficiency of the model. The findings show that ML models can accurately and effectively predict concrete basic creep and are superior to code models. The LGBM has strong robustness for predicting creep compliance. The most relevant and important parameters for predicting creep are the concrete compressive strength and the load over time, respectively. Furthermore, the prediction performances of all the models improved after feature selection. Finally, the applicability of the parameterized gene expression programming (GEP) model was verified using five sets of test data outside the database.