A optimum prediction model of chloride ion diffusion coefficient of machine-made sand concrete based on different machine learning methods

Abstract

The chloride ion diffusion coefficient (CIDC) is an important index to evaluate the durability performance of machined sand concrete. In this study, a database of CIDC is established for 82 sets of experimental data for 4 influence parameters. Meanwhile, different models with 4 input variables (Water to binder ratio, fly ash ratio, relative humidity and temperature) and 1 output variable (CIDC) were constructed to predict CIDC using extreme gradient boosting algorithm (XGBoost), backpropagation algorithm (BP), support vector regression algorithm (SVR), and bp algorithm optimized by Bayesian formula. Meanwhile, the SHapley Additive exPlanations (SHAP) method in python is introduced to characterize the established models to clarify the principles behind the model prediction. Based on the findings, it appears that the Bayesian-BP model is the one that performs the best in terms of making predictions. Furthermore, based on the prediction of the Bayesian-BP model, the percentage contribution of various variable factors in the model is evaluated by the SHAP analysis method, and the corresponding variables are characterized. It is found that the obtained characteristic variables with the CIDC have the same trend as the reality that verifies the reliability of the proposed optimum model. Moreover, CIDC was predicted by controlling variables, and flexible adjustments to CIDC were achieved based on this.