Efficient mix design method for lightweight high strength concrete: A machine learning approach

Abstract

The application of lightweight high strength concrete (LWHSC) is becoming popular due to its superior mechanical properties and lighter weight nature. As a result, it offers advantages in terms of faster construction and cost effectiveness. However, achieving both high strength and lightweight is quite challenging as LWHSC possesses unique characteristics that require different mix design methods compared to conventional concrete. Additionally, the lack of appropriate mix design guidelines limits the usage of LWHSC. Therefore, appropriate tools are necessary to develop effective mix design methods for LWHSC, especially when aiming for high strength. In this study, four machine learning (ML) algorithms, namely support vector regressor (SVR), multilayer perceptron (MLP), gradient boosting regressor (GBR), and extreme gradient boosting (XGB), were applied to comprehensively analyse LWHSC mixes and provide effective prediction models to carry out mix design. To focus on the upper high strength margin of lightweight concrete (LWC), an experimental database consisting of 403 datasets with 28-day compressive strengths greater than 60 MPa and oven dry densities less than 2000 kg/m3 was developed from an extensive literature survey to predict the compressive and splitting tensile strength of LWHSC. The results demonstrated that all ML models predicted LWHSC strengths well with optimum hyperparameter combinations. Among them, GBR outperformed the other three models with an accuracy of less than 5% and 10% in predicting the average compressive and splitting tensile strengths, respectively. Furthermore, partial dependence plots (PDP) and individual conditional expectation (ICE) plots were provided to visualise the correlation between mix compositions and LWHSC strengths.