Intelligent prediction and understanding of self-shrinkage in ultra-high performance concrete based on machine learning

This study innovatively presents a machine learning framework utilizing advanced artificial intelligence technologies to predict the autogenous shrinkage behavior of ultra-high-performance concrete (UHPC). It provides an integrated approach encompassing data collection, preprocessing, model training, prediction, and result comparison to effectively illustrate the framework. Initially, missing data were interpolated using the modified Akima interpolation method, with visualized results provided for discussion. Subsequently, outlier detection and handling were performed using the Generalized Extreme Studentized Deviate ( GESD ) technique. At the same time, the analysis of feature contributions across different model predictions clarified the impact of dataset quality on model outcomes and demonstrated the necessity of outlier detection and handling. This study employed grid search to optimize the training of five models: RF, XGBoost, CatBoost, NGBoost, and LightGBM. The comparative analysis of the results revealed that XGBoost achieved the highest prediction accuracy, with R 2 values of 0.986 for the training set and 0.937 for the test set. Building on this, the four sets of key ratios of UHPC are quoted as input features to train to derive XGBoost-1, achieving R 2 values of 0.917 for the training set and 0.911 for the test set. This approach mitigated variances in raw material compositions, simplified data processing, and enhanced model prediction efficiency. Although its predictive accuracy is slightly lower than that of the initial model, it demonstrates better fitting and generalization performance. Finally, the analysis of correlations among parameters to assess feature importance is crucial for understanding how each parameter influences the auto-shrinkage of UHPC. In conclusion, this framework significantly enhances the accuracy and generalization capabilities of predictive models for UHPC auto-shrinkage and is applicable to other types of concrete materials. Fig. Model frame structure • A machine learning theoretical framework is proposed to predict the contraction performance of UHPC, and the framework is explained in an integrated and systematic manner from data collection to data cleaning to model optimization selection. • The partial correlation analysis reveals how the materials affect the auto-shrinkage of UHPC and provides theoretical guidance for practical engineering. • This framework significantly enhances the accuracy and generalization capabilities of predictive models for UHPC auto-shrinkage and is applicable to other types of concrete materials.