Prediction and validation of compressive strength of metakaolin-based mortars using machine learning

Abstract

Metakaolin (MK)-based cement mortar plays a crucial role in the development of sustainable concrete structures due to its several environmental and performance benefits. It promotes sustainable concrete structures by improving durability, reducing environmental impacts, enhancing material efficiency, and supporting the circular economy in construction.In this research, a comparative study between eight ML classification techniques such as GB, CN2, NB, SVM, SGD, KNN, Tree and RF and one symbolic regression technique such as the RSM has been presented to estimate thecompressive strength of meta-kaolin-based mortarconsidering mixture components contents and its age. A total of 424 records were collected from literature for compressive strength for different mixing ratios of metakaolin-based mortarsat different ages and divided into training set (318 records = 75%) and validation set (106 records = 25%). At the end of the model protocol, SVM andKNN models showed an excellent accuracy of about 92%, while Tree and GB models showed very good accuracies of about 90%. Also, RF and CN2 models showed good accuracy level of about 76–88% and finally NB and SGD produced unacceptable accuracy of less than 60%. Both the correlation matrix and sensitivity analysis results indicated that Age, W/B, and MK/B are the most influential inputs with relative importance of 25% each, then B/S with relative importance of 15%, and SPand Fcem with relative importance of 7% each.Conversely, the RSM model with only two trees and four levels which increased up to four trees and eight levels produced an F value of 32.64, P values less than 0.0500, R² of 0.9422 and Adeq Precision of 31.678. This provides a robust framework for optimizing the mix design. The high R² indicates that the model explains 94.22% of the variance in the MK-based cement mortar compressive strength, making it highly reliable for predicting concrete performance.