Machine Learning Prediction of Self-Compacting Concrete Workability: Integrating Multiscale Rheological Relationships between Paste and Concrete Derived from Paste Rheological Threshold Theory into the Dataset

Abstract

Precise prediction of self-compacting concrete (SCC) workability is vital for efficient production and robust structural performance. This study integrates paste rheological threshold theory with machine learning (ML) to enhance the accuracy of SCC slump flow (SF) and V-Funnel (VF) predictions. A multiscale dataset comprising 313 experimental sets was constructed, featuring both SCC mix proportion data and paste experimental parameters as inputs, which capture critical factors not represented by mix proportions alone. Eight distinct ML models were trained and evaluated, and the results confirmed that models using the multiscale dataset achieved higher predictive accuracy than those trained on mix proportion data alone. The best-performing model was then refined via the Optuna framework for hyperparameter optimization, achieving R² values of 0.96 for SF and 0.91 for VF. Shapley Additive Explanations were employed to elucidate the contributions of each input variable, demonstrating that paste workability significantly enhances both predictive capability and model interpretability. These findings establish a convenient and intelligent approach for predicting SCC workability, thereby promoting its practical engineering application.