Sustainable prediction of concrete strength using rice husk ash and machine learning

Abstract

This paper evaluates variations in supplementary cementitious materials, water-cement ratios, curing times, aggregate sizes, and cement content, the predictive efficacy of five machine learning models – linear regression, random forest, support vector machine (SVM), k-nearest neighbours (KNN), and decision tree – on the compressive strength of concrete. With R2 values between 0.2767 and 0.7440, root mean squared error values ranging from 8.4935 to 17.0477, and mean absolute error values spanning 6.7134 to 14.1603, linear regression shown better accuracy. KNN did well with R2 values of 0.427 and 0.5209 in Cases 2 and 3, respectively, it performed badly in Case 4. under Case 2, the SVM obtained an excellent R2 of 0.4467. With a R2 of 0.8607, random forest performed best in Case 3, it failed in Case 4 nevertheless. Always underperforming, the decision tree showed negative R2 values and notable errors under all conditions. Hence improving model performance, the random forest and SVM obtained R2 values of 0.8607 and 0.8756, respectively. According to Shapley Additive explanations studies, ‘curing time days’ and ‘water-cement ratio’ greatly influence ‘compressive strength’. Emphasising the better predictive power of linear regression, the paper offers a thorough assessment of model efficiency and feature significance for improving concrete mixes.