Optimized machine-learning methods for predicting the long-term viscoelastic behavior of heterogeneous concrete mixtures

Long-term creep compliance is one of the most important mechanical properties for evaluating the long-term behavior of concrete structures. This paper aims to optimize machine-learning models to predict this viscoelastic property. The most relevant dataset available in the literature is considered, cleaned, and preprocessed to optimize the outcome. The advanced XGBoost model, which is to be the most effective shallow machine-learning model for modeling tabular datasets, is employed in this study to maximize model accuracy. Short-term creep compliances of a given sample at typical ages are used as input features to model the long-term creep compliance of concrete. This approach outperforms standard machine-learning approaches that do not include short-term creep as an input feature. Indeed, the short-term behavior of concrete strongly influents its long-term one. The optimized machine model presented herein is accurate and useful for practical applications. It uses input features that are easy to obtain to predict long-term creep compliance up to several decades, which is difficult and expensive to measure.