Prediction of autogenous shrinkage in ultra-high-performance concrete (UHPC) using hybridized machine learning

Abstract

This study explores hybridized machine learning (ML) techniques to predict autogenous shrinkage (AS) in ultra-high-performance concrete (UHPC). The ensemble model, namely random forest (RF), extra tree regressor (ETR), light gradient boosting machine (LGBM), and extended gradient boosting (XGBoost), are adopted as the base algorithm. Further, a newly developed Sparrow Search Algorithm (SSA) is hybridized with XGBoost and proposed in the study for the prediction of shrinkage. The study adopts K-fold cross-validation to reduce the risk of overfitting. The results show that the hybridization of SSA-XGBoost outperforms all the algorithms with those without optimization, with the highest performance of R² of 0.91 and RMSE of 79.2 in the testing set. The model is subjected to five-fold cross-validating, ensuring the model is not overfitted. Regarding RMSE, the performance of other models like XGB, LGBM, ETR, and RF is restricted to 102.22,108.38,87.42 and 98.57, respectively. Further, the study incorporated the model explainability behavior and revealed that the curing relative humidity (CRH), steel fiber content (SFS), and sand are the highly influential features for predicting AS. The comprehensive assessment helps understand the parameters influencing AS, making it a helpful tool for researchers to make well-informed decisions.