Prediction of Interface Rheological Properties of Pumpable Concrete Using AutoGluon with Comparative Analysis of Model-Agnostic Interpretability Methods

The flow characteristics of pumpable concrete can be described effectively using interface rheological properties, i.e., yield stress (YS) and plastic viscosity (PV). Developing accurate predictive models for these properties is crucial to reducing reliance on trial-and-error laboratory testing. Despite the availability of various machine learning (ML) modeling techniques, model selection and hyperparameter optimization complicate the workflow. In addition, the black-box nature of these models hinders their adoption in real-world engineering applications. This study addresses these challenges by proposing an interpretable, automated machine learning (AutoML) framework that streamlines the ML process through automated model selection and optimization, while simultaneously enhancing transparency through model-agnostic interpretability methods. Three traditional ML models—artificial neural networks (ANNs), extreme gradient boosting (XGBoost), and random forest (RF)—were implemented, and the built-in automated optimization strategy of AutoGluon was used for performance and computational time comparison. The results demonstrate that AutoGluon achieved predictive accuracy comparable to that of the best manually tuned models, while offering the additional advantages of data preprocessing. To enhance transparency, three model-agnostic methods—SHapley Additive exPlanations (SHAP), local interpretable model-agnostic explanations (LIME), and permutation feature importance—were applied and compared to provide a better understanding. Global and aggregated local feature importance, together with SHAP dependency plots, highlighted the key influencing parameters and revealed the nonlinear behaviors of input features with respect to the outputs. This explainable AutoGluon framework offers a practical approach for modeling and understanding the interface rheological properties of pumpable concrete, reducing reliance on empirical methods and extensive laboratory testing.