Response surface regression and machine learning models to predict the porosity and compressive strength of pervious concrete based on mix design parameters

Abstract

This study investigates the influence of aggregate size, aggregate-to-cement ratio, and compaction effort on pervious concrete's porosity and compressive strength. It proposes using response surface methodology and machine learning techniques to predict porosity and compressive strength. Fifteen mix designs (three aggregate sizes and five aggregate-to-cement ratios) with seven compaction energy levels were employed. Experimental data analysis using various regression models revealed that the quadratic model provided the best fit for predicting both porosity and compressive strength. Machine learning models were employed to predict porosity and compressive strength more accurately. Among the models investigated, the artificial neural network achieved superior performance across all datasets (training, testing, and validation). This suggests the artificial neural network model can effectively capture the complex relationships between input and response variables. Sensitivity analysis using SHAP (SHapley Additive exPlanations) revealed that compaction energy significantly impacts both porosity and compressive, while aggregate size has the most negligible influence.