Predicting the behaviour of self-compacting concrete incorporating agro-industrial waste using experimental investigations and comparative machine learning modelling

Abstract

This study examines the effects of the silica fume (SF) and fly ash (FA) together as a binary replacement of cement on the behaviour of self-compacting concrete (SCC) using experimental investigation and application of various machine learning (ML) models developed for this study. The SF and FA were used in combination of 25/75, 50/50, 75/25 proportions with 5%, 10% and 15% cement replacement, respectively, at 3,7, 14- and 28-days curing. In first phase, unit weight, workability, compressive strength (CS) and dimensional stability of the sustainable SCC were investigated through experimental testing. The experimental results revealed that the fresh properties of SCC, at all replacement levels and combination of SF and FA, were remained within the range of the standard values. At the same replacement percentages, the combination of SF/FA of 50/50 was more effective than the other two combinations of SF/FA at the same dosage levels and at all the ages of SCC. The CS was improved by 17% at 15% replacement level when SF and FA were blended at the combination of 50/50. The second phase of this study differentiate this research from the past literature. The collection of modelling algorithms used in this study have never been applied to this kind of data to provide a comparative analysis over two different datasets. The ML modelling was performed using Neural Network Regression (NNR), Decision Forest Regression (DFR), Linear Regression (LR) and Bayesian regression (BR). Initially, the dataset achieved through the experimental investigations performed in this study was modelled. The models efficiently predicted the CS of sustainable SCC when compared with experimental results. Later, the dataset available in the literature was also used to verify the accuracy of the proposed ML models. Based on the findings, the best performing algorithms were LR and NNR, both with R2 = 0.9, that highlights the optimum and enhanced performance of the ML modelling over other numerical methods.