The properties of cementitious binders are controlled by their composition and structure at different scales. However, the complexity of their disordered, multi-scale structure makes it challenging to elucidate such linkages. In particular, due to a lack of physical models, predicting the strength development of concretes remains challenging. As an alternative route to physics-based models, machine learning offers a promising pathway to develop new predictive models for materials based on existing datasets. Here, we show that machine learning models can be used to reliably predict concrete’s strength development. This approach relies on the analysis of a large data set (>10,000 observations) of measured compressive strengths from actual (job-site) mixtures and their corresponding actual mixture proportions. The developed models successfully predict the 28-day strength of concretes based on the mere knowledge of the mixture proportions with an accuracy of ±4.5 MPa. We illustrate how we optimize various machine models in order to achieve a balance between accuracy and complexity.
Author: Boya Ouyang