Machine learning based optimization for mix design of manufactured sand concrete

Abstract

Using manufactured sand (M-sand) as a replacement of natural sand is beneficial both environmentally and economically. However, apart from basic requirements like that for concrete using natural sand, mix design of manufactured sand concrete (MSC) needs to take more factors into consideration and requires more efficient optimization due to relative deficiency in testing data. This paper uses 86 instances of MSC with 6 features to predict compressive strength and chloride permeability coefficient (CPC) of MSC by employing four machine learning (ML) models (Back propagation (BP) neural network, random forest (RF), support vector regression (SVR) and eXtreme Gradient Boosting (XGBoost)). All four models have predicted the compressive strength and CPC of MSC accurately, with R² of test set ranging from 0.85 to 0.93 after hyperparameter optimization, with XGBoost models achieving the highest R² of 0.93 for both. Also, SHapley Additive exPlanations (SHAP) analysis indicates that cement content is the most predominant factor to affect compressive strength and CPC, followed by M-sand content and water/binder ratio (W/B ratio). Finally, CPC, compressive strength and unit cost are combined to construct a three-way fitness function and multi-objective optimization is performed using Non-Dominated Sorting Genetic Algorithm 2 (NSGA-II). Based on multi-algorithm comparison and cost-aware multi-objective XGBoost-NSGA-II optimization, the mix design method proposed is advantageous in terms of accuracy, reliability and production cost compared with ML models that employ a single model, predict a single property, and do not take in cost as a factor for mix design.