Predicting the compressive strength of steelmaking slag concrete with machine learning - Considerations on developing a mix design tool

Steelmaking slag has been extensively studied as aggregate for cement-based composites. Because of the distinct properties of this residue, traditional mix design methods are not suitable to determine its target compressive strength, which hinders research studies and compromises its use on a large scale. In this context, the present work aims to develop machine learning (ML)-based models to predict the compressive strength of steel slag concretes from their mix proportion. For this purpose, a global data survey on steel slag concretes was carried out to create a benchmark dataset. Then, four ML-based models were trained and cross-validated using this dataset: Support Vector Regression (SVR), Artificial Neural Networks (ANN), Extreme Gradient Boost (XGBoost), and Gaussian Process Regression (GPR). Finally, new steel slag concrete specimens were built and tested to experi-mentally validate the adjusted models. The model that achieved the best performance using the literature dataset was the ANN, with a R2 of 0.79. However, the experimental validation was not satisfactory - the GPR, XGBoost and SVR models presented negative R2 values. These results brought light to some pivotal aspects that must be considered when using ML techniques: i) the size and homogeneity of the dataset; ii) the proper choice of input parameters; and iii) the use of cross-validation to adjust the models. Hence, although such techniques are promising and powerful, care must be taken on the generalization of their predictions, especially when the available data is limited.

Automated summary

This study developed machine learning models to predict the compressive strength of steel slag concretes. A benchmark dataset was created through a global data survey, and four ML-based models were trained and cross-validated. The experimental validation revealed some issues in using ML techniques to predict the compressive strength of steel slag concretes, especially when the available data is limited.