Hybrid Physics-Informed Machine Learning via residual fitting for the prediction of creep and shrinkage in concrete

Abstract Recent works about data-driven analysis of concrete creep and shrinkage have enhanced structural analysis, providing more efficient results compared to standard models. Moreover, recent studies have shown strong efficiency when physics laws are coupled with machine learning (ML), an approach called physics-informed machine learning (PIML). In this paper, the purpose is to implement Hybrid PIML via residual fitting to enhance standard ML models for predicting creep and shrinkage in concrete structures. While purely data-driven ML models offer high accuracy, they frequently lack physical grounding and extrapolative reliability. Conversely, traditional models provide theoretical consistency but struggle to capture multidimensional material behaviors. Motivated by this gap, the hybrid approach leverages established physical laws, Prony series for viscoelasticity and the ACI 209 shrinkage model, as baseline predictors. Advanced ML algorithms (XGBoost, Support Vector Machines, and Artificial Neural Networks) are then utilized to capture and correct the residual errors. The analysis relies on the Northwestern University database, providing an extensive historical record from the 1950s up to 2023. Briefly, the Hybrid PIML XGBoost Prony achieved $$R^2$$ = 0.97, RMSE of 14.34 MPa, and MAE of 5.34 MPa for creep. For shrinkage, HPIML SVM-OU ACI achieved $$R^2$$ = 0.94, RMSE of 63.32 MPa, and MAE of 31.77 MPa. Both models outperformed ACI 209, B3, and B4 models, which achieved $$R^2$$ of 0.41, 0.43, and 0.43, respectively. These findings exhibit that HPIML models are capable of reaching better results than standard and conventional ML models.