Data-driven compressive strength investigation and design suggestions for rubberized concrete

This study employs machine learning algorithms to explore the compressive strength of rubberized concrete. The investigation utilizes a dataset containing 228 synthesized samples and nine input features for training and testing six machine learning models. The SHapley Additive exPlanation (SHAP) algorithm is applied to elucidate the prediction processes of machine learning models and to identify the relationships between the nine input features and the compressive strength of rubberized concrete. The evaluation metrics comparison reveals that Support Vector Machine, Artificial Neural Networks, and eXtreme Gradient Boosting Trees models achieve the highest accuracy and generalization capabilities, with their Coefficient of Determination values exceeding 0.97 and Mean Absolute Percent Error consistently below 9 % on both training and testing datasets. The analysis of machine learning models indicates that the rubber and superplasticizer content, water-to-cement ratio, and curing time are crucial variables affecting compressive strength, while the fine aggregate content, silica fume-to-cement ratio, and cement usage have minimal impacts. Specifically, the water-to-cement ratio, rubber content, and fine aggregate usage exhibit negative correlations with the compressive strength of rubberized concrete, whereas longer curing times and increased superplasticizer content have positive effects on compressive strength. Furthermore, to mitigate the decrease in compressive strength resulting from rubber inclusion, it is recommended to limit the rubber content to below 50 kg/m³ and use a maximum size of rubber crumb greater than 4 mm. Besides, maintaining a water-to-cement ratio between 0.35 and 0.45, ensuring a coarse aggregate content exceeding 1100 kg/m³, and employing a silica fume-to-cement ratio above 10 % are also helpful measures. © 2025 Elsevier Ltd.