Predicting the 28-day compressive strength by mix proportions: Insights from a large number of observations of industrially produced concrete

Abstract

Many previous studies explored machine learning (ML) models to predict the 28-day compressive strength (f28) by the mix proportions using laboratory data, but only a few exploited industrial data. Predicting industrial concrete performance by laboratory models could be questionable as an industrial environment involves much more uncertainties. This study has applied ML models on 12,107 observations of industrially produced concrete that are associated with a wide range of concrete applications of various strength and slump requirements. A systematic approach has been taken to develop seven ML models to predict f28, covering data visualization, model selection and assessment, model finalization, and variable importance analysis. Significant improvements have been achieved compared to previous studies using field concrete data. It is also found that the industrial data has a multimodal problem and is much noisier than laboratory data. The outcomes of this study can serve as a useful reference in predicting f28 and designing the mix to produce the desired concrete in an industrial environment.