Prediction of concrete porosity using machine learning
Chong Cao
Abstract
Porosity is an important indicator of the durability performance of concrete. The objective of this study is to apply machine learning methods to empirically predict the porosity of high-performance concrete containing supplementary cementitious materials. The assembled database for concrete porosity consists of 240 data records, featuring 74 unique concrete mixture designs. The compositional features of concrete include water/cement ratio, fly ash, slag, aggregate content, superplasticizers and curing conditions. The numerical results suggest that gradient boosting trees outperform random forests in terms of their prediction accuracy. XGBoost achieves the best performance with additional regularization over model complexity to prevent overfitting. Compared with the conventional chemo-mechanical model for predicting concrete porosity, the proposed data-driven approach not only overcomes the difficulty in estimating the time-dependent degree of hydration, but also achieves a higher prediction accuracy of R2 = 0.9770, MAPE = 2.97%, and RMSE = 0.431 for porosity (%). The predictor importance plot shows that curing days, water/binder ratio, and aggregate content are the most important predictors of concrete porosity.