Evaluating the relevance of eggshell and glass powder for cement-based materials using machine learning and SHapley Additive exPlanations (SHAP) analysis

This study used machine learning methods to predict the water absorption (W-A) of cement-based material (CBM) containing eggshell and glass powder as sand and cement substitutes. A dataset from the laboratory experiments consisting of 234 points and seven input variables was used to develop models, including multilayer perceptron neural network (MLPNN), support vector ma-chine (SVM), adaptive boosting (AdaBoost), and extreme gradient boosting (XGBoost). Addi-tionally, a SHapley Additive exPlanations (SHAP) analysis was performed to investigate the relevance and interaction of raw components. When evaluating the prediction models for the W-A of CBM, it was found that the MLPNN and SVM models were moderately accurate (R2 = 0.74 and 0.78, respectively), while the AdaBoost and XGBoost models showed good agreement with the lab test results (R2 = 0.86 and 0.91, respectively). The SHAP approach revealed that while the cement quantity had a higher negative association with W-A of CBM, the quantities of eggshell powder, sand, and glass powder showed both favourable and detrimental correlations. Therefore, eggshell and glass powder must be used in optimal proportions of around 60 kg/m3 and 80 kg/m3, respectively, for maximum resistance to W-A. The AdaBoost and XGBoost models can potentially compute the W-A of CBMs by utilising various input parameter values, which may help decrease unnecessary test trials in labs. Furthermore, the SHAP investigation revealed the impact and relationship of the inputs on the W-A of CBMs, which might potentially assist researchers and the industry in determining the appropriate amount of raw materials during CBM production.