Advanced Machine Learning Approaches for Predicting Long-Term Concrete Shrinkage

Abstract

Long-term shrinkage strain is one of the most critical mechanical characteristics for assessing the long-term performance of concrete structures. This study aims to enhance machine learning methods for accurately predicting this phenomenon. A comprehensive and widely referenced dataset of concrete shrinkage, comprising 1869 shrinkage curves and 32318 individual shrinkage strain measurements from a diverse range of samples worldwide, is employed. The analysis uses the full database for model training, while curve-level evaluation emphasizes long-term total shrinkage. The advanced XGBoost algorithm, recognized as one of the most effective machine learning models for tabular data, is used to improve predictive performance. Short-term shrinkage strains measured at standard early ages of 1, 10 or 100 days are included as input features to efficiently predict long- to ultra-long-term shrinkage strains up to 30 years. To avoid data leakage, the dataset is split by shrinkage curves rather than individual measurement points for training and testing. This approach, achieving a testing R² score of up to 0.97, significantly outperforms conventional machine techniques that exclude early-age shrinkage data. The improved performance highlights the strong influence of early-age behavior on the long-term shrinkage response of concrete. The present results reflect the composition range represented in the curated database, and can be applied primarily to mixes without explicit slag/silica-fume/fly-ash tagging.