Interpretable hybrid learning model for estimating the compressive strength of sustainable pumice based concrete

Designing concrete mixes that are both strong and environmentally sustainable is a growing priority, especially when incorporating supplementary materials like pumice to reduce cement usage. This study investigates the effectiveness of Machine Learning (ML) models in predicting the 28 day Compressive Strength (CS) of pumice based concrete. A curated dataset of 304 concrete mix designs was compiled from existing literature, and seven widely used ML algorithms, including artificial neural network, decision tree, and boosting methods, were developed and compared. The novelty of this work lies in three aspects; first, it extends hybrid ensemble learning to pumice-based sustainable concrete, which has been rarely explored in ML research; second, it integrates multiple base learners (ANN, RF, GBRT, CatBoost, LGBM, and GPR) through Grey Wolf Optimizer (GWO) weighting to improve predictive robustness; and third, it incorporates SHAP analysis to provide interpretable, engineering-relevant insights into mix design parameters, bridging a gap that is often overlooked in prior studies. The framework achieved both high predictive accuracy and robust generalization. In addition, a hybrid ensemble model was created using the Grey Wolf Optimizer to combine the strengths of the individual models. This hybrid approach achieved the highest predictive accuracy, reaching an R² of 0.954 and RMSE of 5.00 MPa on unseen test data, and the A20-index of this model had the best performance after the Light Gradient Boosting Machine (LGBM) model, with a value of 88.25%. Beyond numerical accuracy, the hybrid model demonstrates practical significance by helping engineers understand the role of key factors such as cement content, CA/FA ratio, superplasticizer dosage, and W/P ratio, thereby guiding sustainable and performance based mix design. Code is available at: https://github.com/MLcode-Concrete/Concrete_strength .