Development of Machine Learning Prediction Models as New Approaches for Foam Concrete Design and Implementation

Abstract Foam concrete is a type of light-weight cement mortar containing air bubbles possessing qualities such as improved flowability, heat insulation, fire resistance, and soundproofing. It is produced by chemical or mechanical foaming processes. It was brought in as far back as the dawn of the nineteenth century, yet remains not so popular worldwide due to a lack of knowledge and technological limitations. Machine learning (ML)-based predictive models offer a potential approach to accurately predicting foam concrete’s qualities and thereby its design and use in construction. In this current research, using support vector regression (SVR), decision tree (DT), and gradient boosting regression (GBR) to predict foam concrete’s 28-day compressive strength (Comp-28d) is investigated. This research utilizes metaheuristic optimization and ensemble learning techniques capable of enhancing the predictive capabilities of conventional ML models. Sobol and LIME were utilized for sensitivity analysis and interpretation of the model. In experiments, the DTR model was found to be the best among all of its base models in terms of high R 2 of 0.997, extremely low root mean square error (RMSE) of 0.023, and a Theil’s UI of 0.023. It was further optimized using the Kepler optimization algorithm (KOA) to generate a hybrid decision tree regression + Kepler optimization algorithm (DTKO). DTKO possesses superior predictability in terms of an R 2 of 0.998 (increased by 0.1% over DTR) and minimizing its RMSE to 0.017, an error reduction of 26.1%. Successful utilization of such computational methodology offers practical answers for use in construction as well as for opening areas for potential future research in knowledge-based designing of materials.