Machine Learning for Performance Prediction of Ultra-High Performance Concrete

Abstract

Ultra-high performance concrete (UHPC) has emerged as a revolutionary cementitious composite with exceptional mechanical properties and durability, yet its mixture design remains challenging due to complex nonlinear interactions among numerous constituents. Traditional experimental approaches are time-consuming, costly, and inefficient for optimizing UHPC formulations. Machine learning (ML) has recently gained significant attention as a powerful alternative for predicting UHPC performance and optimizing its mixture designs. This paper provides a comprehensive review of ML applications in UHPC, focusing on the prediction of compressive strength, flexural strength, workability, and durability properties. Various ML algorithms—including artificial neural networks (ANN), support vector regression (SVR), random forest (RF), extreme gradient boosting (XGBoost), CatBoost, and Bayesian neural networks—are critically examined. The review synthesizes findings from over 200 published studies and multiple publicly available datasets comprising more than 2,000 UHPC mix designs. Key challenges, including data quality, model interpretability, and external validation, are discussed. Future research directions, such as physics-informed neural networks, generative AI for data augmentation, and explainable AI frameworks, are proposed to advance the field toward reliable and interpretable UHPC design.