Machine-Learning-Based Compressive Strength Prediction of Low-Carbon Concrete Incorporating Agro-Residue-Derived Pozzolanic Materials

Abstract

Rice husk ash (RHA) and sugarcane bagasse ash (SBA) are silica-rich materials generated in abundance during the cultivation and processing of rice and sugarcane. The use of these materials as supplementary cementitious materials (SCMs) in the preparation of Ordinary Portland Cement (OPC) concretes can help to reduce the carbon and cost footprint of OPC. In the present investigation, OPC-53 was partially replaced with RHA and SBA in the preparation of M30-grade concretes, with each additive replacing between 0 and 30% of the OPC by mass. Each of the ten concretes cast for this research were formed into 150 mm cubes (90 total specimens), and evaluated for workability and compressive strength at 7, 14 and 28 days of water curing. Results indicated that concretes containing a binary blend of 10% RHA and 10% SBA developed the highest strength (37.5 N/mm² at 28 days), a 28% increase over the plain OPC control concretes (29.3 N/mm² at 28 days), despite the fact that the workability of concretes decreased with the addition of RHA and increased with the addition of SBA. In addition to testing these concretes, five different machine learning models (multiple linear regression, support vector regression, random forest, extreme gradient boosting (XGBoost), and an artificial neural network) were trained on the results of these concretes in order to model the compressive strength of similar concretes with varying amounts of each SCM. Each of these models was trained and tested using leave-one-mix-out cross-validation to determine the generalization of each machine learning model, a method that avoids overestimating the skill of the models as is often done in other small sample sized studies. Results of the tests revealed that the artificial neural network models contained the best generalization of the tested models (R² = 0.90, RMSE = 1.34 N/mm², MAE = 0.94 N/mm², MAPE = 3.30%), outperforming the other four models tested. Additionally,SHAP analysis of the artificial neural network models indicated that the most important factors influencing the compressive strength of concretes were the age of curing, the amount of SBA added, and the amount of RHA added, each in that order. These results help to indicate that the use of RHA and SBA in binary combination is both technically feasible and potentially beneficial in relation to sustainability as a replacement for plain OPC concretes; furthermore, that there exist machine learning models with accurate predictions of the compressive strength of concretes based upon the parameters tested within this project.