Predicting compressive strength of sustainable concrete using advanced AI models: DLNN, RF, and MARS

Abstract

Recycled aggregate is becoming a sustainable construction resource that minimizes the carbon footprint in concrete structures. To predict the behavior of environmentally friendly (EnF) concrete in sustainable construction, it is necessary to predict the compressive strength using recycled materials accurately. The current research highlights the performance of the Deep Learning Neural Network (DLNN), Random Forests (RFs), and Multivariate Adaptive Regression Splines (MARS) are evaluated and extensive analysis of data segmentation was performed by splitting the dataset used in this study into 75–25% as well as 80–20% training/testing scenarios using Extreme Gradient Boosting (XG Boost), a quantitative measurement of the effect of data segmentation on model efficiency. The combination of AI models with Extreme Gradient Boosting (XG Boost) was employed to ascertain the governing variables on the CS prediction. Numerous statistical models developed were used to compare the effectiveness of these given models showing the best performance of the DLNN model based on the least RMSE (2.93). The results found that more variables should be added to the prediction problem for better prediction accuracy and the data split of 80–20% was the best choice. Based on the high accuracy of models, the results demonstrated that over the other established models, the DLNN model surpasses them in the analysis of concrete behavior and is useful for future applications.