Comparative analysis of selected machine learning techniques for predicting the pull-off strength of the surface layer of eco-friendly concrete

Abstract

With recent trends reducing the carbon footprint of concrete, more novel materials are designed. It's mostly done by replacing cement with admixtures that are wastes in industrial processes. There is a need to provide reliable and accurate models to estimate the properties of the material. In this case the selected ML algorithms such as ANN, RF and DT were used for estimating the pull-off strength of the surface layer of cement mortar containing granite powder, fly ash and ground granulated blast furnace slag. The focus was on the cement-sand ratio of 1:3, replacing up to 30 % of the binder. Ultrasonic pulse velocity and pull-off strength of the surface layer. The analyses were performed in comparative manner and proved the accuracy of the designed models. The error values (MAPE, NRMSE and MAE) of the most effective model is below 3,5 %, indicating an extremely high success rate in prediction. An R² ratio of 0.9436 confirms the very good fit of the model. Parametric tests were performed and SHAP analysis gave a better understanding of the models. The main conclusion of the study is to identify the possibility of replacing destructive testing with non-destructive testing supported by machine learning and material information to determine the pull-off strength of the subsurface layer at a selected depth for cement mortars containing waste materials. A particular advantage of the presented approach is the possibility of reducing the time to determine selected desired material parameters and the amount of testing required compared to the traditional approach.